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
 142                                 __reserved_1   : 54;
 143
 144         __u32                   extra_config_len;
 145         __u32                   __reserved_4;
 146
 147         __u64                   __reserved_2;
 148         __u64                   __reserved_3;
 149 };
 150
 151 /*
 152  * Ioctls that can be done on a perf counter fd:
 153  */
 154 #define PERF_COUNTER_IOC_ENABLE         _IO('$', 0)
 155 #define PERF_COUNTER_IOC_DISABLE        _IO('$', 1)
 156
 157 /*
 158  * Structure of the page that can be mapped via mmap
 159  */
 160 struct perf_counter_mmap_page {
 161         __u32   version;                /* version number of this structure */
 162         __u32   compat_version;         /* lowest version this is compat with */
 163         __u32   lock;                   /* seqlock for synchronization */
 164         __u32   index;                  /* hardware counter identifier */
 165         __s64   offset;                 /* add to hardware counter value */
 166
 167         __u32   data_head;              /* head in the data section */
 168 };
 169
 170 struct perf_event_header {
 171         __u32   type;
 172         __u32   size;
 173 };
 174
 175 enum perf_event_type {
 176         PERF_EVENT_IP           = 0,
 177         PERF_EVENT_GROUP        = 1,
 178
 179         __PERF_EVENT_TID        = 0x100,
 180 };
 181
 182 #ifdef __KERNEL__
 183 /*
 184  * Kernel-internal data types and definitions:
 185  */
 186
 187 #ifdef CONFIG_PERF_COUNTERS
 188 # include <asm/perf_counter.h>
 189 #endif
 190
 191 #include <linux/list.h>
 192 #include <linux/mutex.h>
 193 #include <linux/rculist.h>
 194 #include <linux/rcupdate.h>
 195 #include <linux/spinlock.h>
 196 #include <linux/hrtimer.h>
 197 #include <asm/atomic.h>
 198
 199 struct task_struct;
 200
 201 static inline u64 perf_event_raw(struct perf_counter_hw_event *hw_event)
 202 {
 203         return hw_event->config & PERF_COUNTER_RAW_MASK;
 204 }
 205
 206 static inline u64 perf_event_config(struct perf_counter_hw_event *hw_event)
 207 {
 208         return hw_event->config & PERF_COUNTER_CONFIG_MASK;
 209 }
 210
 211 static inline u64 perf_event_type(struct perf_counter_hw_event *hw_event)
 212 {
 213         return (hw_event->config & PERF_COUNTER_TYPE_MASK) >>
 214                 PERF_COUNTER_TYPE_SHIFT;
 215 }
 216
 217 static inline u64 perf_event_id(struct perf_counter_hw_event *hw_event)
 218 {
 219         return hw_event->config & PERF_COUNTER_EVENT_MASK;
 220 }
 221
 222 /**
 223  * struct hw_perf_counter - performance counter hardware details:
 224  */
 225 struct hw_perf_counter {
 226 #ifdef CONFIG_PERF_COUNTERS
 227         union {
 228                 struct { /* hardware */
 229                         u64                             config;
 230                         unsigned long                   config_base;
 231                         unsigned long                   counter_base;
 232                         int                             nmi;
 233                         unsigned int                    idx;
 234                 };
 235                 union { /* software */
 236                         atomic64_t                      count;
 237                         struct hrtimer                  hrtimer;
 238                 };
 239         };
 240         atomic64_t                      prev_count;
 241         u64                             irq_period;
 242         atomic64_t                      period_left;
 243 #endif
 244 };
 245
 246 struct perf_counter;
 247
 248 /**
 249  * struct hw_perf_counter_ops - performance counter hw ops
 250  */
 251 struct hw_perf_counter_ops {
 252         int (*enable)                   (struct perf_counter *counter);
 253         void (*disable)                 (struct perf_counter *counter);
 254         void (*read)                    (struct perf_counter *counter);
 255 };
 256
 257 /**
 258  * enum perf_counter_active_state - the states of a counter
 259  */
 260 enum perf_counter_active_state {
 261         PERF_COUNTER_STATE_ERROR        = -2,
 262         PERF_COUNTER_STATE_OFF          = -1,
 263         PERF_COUNTER_STATE_INACTIVE     =  0,
 264         PERF_COUNTER_STATE_ACTIVE       =  1,
 265 };
 266
 267 struct file;
 268
 269 struct perf_mmap_data {
 270         struct rcu_head                 rcu_head;
 271         int                             nr_pages;
 272         atomic_t                        wakeup;
 273         atomic_t                        head;
 274         struct perf_counter_mmap_page   *user_page;
 275         void                            *data_pages[0];
 276 };
 277
 278 struct perf_wakeup_entry {
 279         struct perf_wakeup_entry *next;
 280 };
 281
 282 /**
 283  * struct perf_counter - performance counter kernel representation:
 284  */
 285 struct perf_counter {
 286 #ifdef CONFIG_PERF_COUNTERS
 287         struct list_head                list_entry;
 288         struct list_head                event_entry;
 289         struct list_head                sibling_list;
 290         int                             nr_siblings;
 291         struct perf_counter             *group_leader;
 292         const struct hw_perf_counter_ops *hw_ops;
 293
 294         enum perf_counter_active_state  state;
 295         enum perf_counter_active_state  prev_state;
 296         atomic64_t                      count;
 297
 298         /*
 299          * These are the total time in nanoseconds that the counter
 300          * has been enabled (i.e. eligible to run, and the task has
 301          * been scheduled in, if this is a per-task counter)
 302          * and running (scheduled onto the CPU), respectively.
 303          *
 304          * They are computed from tstamp_enabled, tstamp_running and
 305          * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
 306          */
 307         u64                             total_time_enabled;
 308         u64                             total_time_running;
 309
 310         /*
 311          * These are timestamps used for computing total_time_enabled
 312          * and total_time_running when the counter is in INACTIVE or
 313          * ACTIVE state, measured in nanoseconds from an arbitrary point
 314          * in time.
 315          * tstamp_enabled: the notional time when the counter was enabled
 316          * tstamp_running: the notional time when the counter was scheduled on
 317          * tstamp_stopped: in INACTIVE state, the notional time when the
 318          *      counter was scheduled off.
 319          */
 320         u64                             tstamp_enabled;
 321         u64                             tstamp_running;
 322         u64                             tstamp_stopped;
 323
 324         struct perf_counter_hw_event    hw_event;
 325         struct hw_perf_counter          hw;
 326
 327         struct perf_counter_context     *ctx;
 328         struct task_struct              *task;
 329         struct file                     *filp;
 330
 331         struct perf_counter             *parent;
 332         struct list_head                child_list;
 333
 334         /*
 335          * These accumulate total time (in nanoseconds) that children
 336          * counters have been enabled and running, respectively.
 337          */
 338         atomic64_t                      child_total_time_enabled;
 339         atomic64_t                      child_total_time_running;
 340
 341         /*
 342          * Protect attach/detach and child_list:
 343          */
 344         struct mutex                    mutex;
 345
 346         int                             oncpu;
 347         int                             cpu;
 348
 349         /* mmap bits */
 350         struct mutex                    mmap_mutex;
 351         atomic_t                        mmap_count;
 352         struct perf_mmap_data           *data;
 353
 354         /* poll related */
 355         wait_queue_head_t               waitq;
 356         /* optional: for NMIs */
 357         struct perf_wakeup_entry        wakeup;
 358
 359         void (*destroy)(struct perf_counter *);
 360         struct rcu_head                 rcu_head;
 361 #endif
 362 };
 363
 364 /**
 365  * struct perf_counter_context - counter context structure
 366  *
 367  * Used as a container for task counters and CPU counters as well:
 368  */
 369 struct perf_counter_context {
 370 #ifdef CONFIG_PERF_COUNTERS
 371         /*
 372          * Protect the states of the counters in the list,
 373          * nr_active, and the list:
 374          */
 375         spinlock_t              lock;
 376         /*
 377          * Protect the list of counters.  Locking either mutex or lock
 378          * is sufficient to ensure the list doesn't change; to change
 379          * the list you need to lock both the mutex and the spinlock.
 380          */
 381         struct mutex            mutex;
 382
 383         struct list_head        counter_list;
 384         struct list_head        event_list;
 385         int                     nr_counters;
 386         int                     nr_active;
 387         int                     is_active;
 388         struct task_struct      *task;
 389
 390         /*
 391          * time_now is the current time in nanoseconds since an arbitrary
 392          * point in the past.  For per-task counters, this is based on the
 393          * task clock, and for per-cpu counters it is based on the cpu clock.
 394          * time_lost is an offset from the task/cpu clock, used to make it
 395          * appear that time only passes while the context is scheduled in.
 396          */
 397         u64                     time_now;
 398         u64                     time_lost;
 399 #endif
 400 };
 401
 402 /**
 403  * struct perf_counter_cpu_context - per cpu counter context structure
 404  */
 405 struct perf_cpu_context {
 406         struct perf_counter_context     ctx;
 407         struct perf_counter_context     *task_ctx;
 408         int                             active_oncpu;
 409         int                             max_pertask;
 410         int                             exclusive;
 411
 412         /*
 413          * Recursion avoidance:
 414          *
 415          * task, softirq, irq, nmi context
 416          */
 417         int                     recursion[4];
 418 };
 419
 420 /*
 421  * Set by architecture code:
 422  */
 423 extern int perf_max_counters;
 424
 425 #ifdef CONFIG_PERF_COUNTERS
 426 extern const struct hw_perf_counter_ops *
 427 hw_perf_counter_init(struct perf_counter *counter);
 428
 429 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
 430 extern void perf_counter_task_sched_out(struct task_struct *task, int cpu);
 431 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
 432 extern void perf_counter_init_task(struct task_struct *child);
 433 extern void perf_counter_exit_task(struct task_struct *child);
 434 extern void perf_counter_do_pending(void);
 435 extern void perf_counter_print_debug(void);
 436 extern void perf_counter_unthrottle(void);
 437 extern u64 hw_perf_save_disable(void);
 438 extern void hw_perf_restore(u64 ctrl);
 439 extern int perf_counter_task_disable(void);
 440 extern int perf_counter_task_enable(void);
 441 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
 442                struct perf_cpu_context *cpuctx,
 443                struct perf_counter_context *ctx, int cpu);
 444 extern void perf_counter_update_userpage(struct perf_counter *counter);
 445
 446 extern void perf_counter_output(struct perf_counter *counter,
 447                                 int nmi, struct pt_regs *regs);
 448 /*
 449  * Return 1 for a software counter, 0 for a hardware counter
 450  */
 451 static inline int is_software_counter(struct perf_counter *counter)
 452 {
 453         return !perf_event_raw(&counter->hw_event) &&
 454                 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
 455 }
 456
 457 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *);
 458
 459 #else
 460 static inline void
 461 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
 462 static inline void
 463 perf_counter_task_sched_out(struct task_struct *task, int cpu)          { }
 464 static inline void
 465 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
 466 static inline void perf_counter_init_task(struct task_struct *child)    { }
 467 static inline void perf_counter_exit_task(struct task_struct *child)    { }
 468 static inline void perf_counter_do_pending(void)                        { }
 469 static inline void perf_counter_print_debug(void)                       { }
 470 static inline void perf_counter_unthrottle(void)                        { }
 471 static inline void hw_perf_restore(u64 ctrl)                            { }
 472 static inline u64 hw_perf_save_disable(void)                  { return 0; }
 473 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
 474 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
 475
 476 static inline void
 477 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs)  { }
 478
 479 #endif
 480
 481 #endif /* __KERNEL__ */
 482 #endif /* _LINUX_PERF_COUNTER_H */