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  * Hardware event to monitor via a performance monitoring counter:
 107  */
 108 struct perf_counter_hw_event {
 109         /*
 110          * The MSB of the config word signifies if the rest contains cpu
 111          * specific (raw) counter configuration data, if unset, the next
 112          * 7 bits are an event type and the rest of the bits are the event
 113          * identifier.
 114          */
 115         __u64                   config;
 116
 117         __u64                   irq_period;
 118         __u64                   record_type;
 119         __u64                   read_format;
 120
 121         __u64                   disabled       :  1, /* off by default        */
 122                                 nmi            :  1, /* NMI sampling          */
 123                                 inherit        :  1, /* children inherit it   */
 124                                 pinned         :  1, /* must always be on PMU */
 125                                 exclusive      :  1, /* only group on PMU     */
 126                                 exclude_user   :  1, /* don't count user      */
 127                                 exclude_kernel :  1, /* ditto kernel          */
 128                                 exclude_hv     :  1, /* ditto hypervisor      */
 129                                 exclude_idle   :  1, /* don't count when idle */
 130                                 include_tid    :  1, /* include the tid */
 131
 132                                 __reserved_1   : 54;
 133
 134         __u32                   extra_config_len;
 135         __u32                   __reserved_4;
 136
 137         __u64                   __reserved_2;
 138         __u64                   __reserved_3;
 139 };
 140
 141 /*
 142  * Ioctls that can be done on a perf counter fd:
 143  */
 144 #define PERF_COUNTER_IOC_ENABLE         _IO('$', 0)
 145 #define PERF_COUNTER_IOC_DISABLE        _IO('$', 1)
 146
 147 /*
 148  * Structure of the page that can be mapped via mmap
 149  */
 150 struct perf_counter_mmap_page {
 151         __u32   version;                /* version number of this structure */
 152         __u32   compat_version;         /* lowest version this is compat with */
 153         __u32   lock;                   /* seqlock for synchronization */
 154         __u32   index;                  /* hardware counter identifier */
 155         __s64   offset;                 /* add to hardware counter value */
 156
 157         __u32   data_head;              /* head in the data section */
 158 };
 159
 160 struct perf_event_header {
 161         __u32   type;
 162         __u32   size;
 163 };
 164
 165 enum perf_event_type {
 166         PERF_EVENT_IP           = 0,
 167         PERF_EVENT_GROUP        = 1,
 168
 169         __PERF_EVENT_TID        = 0x100,
 170 };
 171
 172 #ifdef __KERNEL__
 173 /*
 174  * Kernel-internal data types and definitions:
 175  */
 176
 177 #ifdef CONFIG_PERF_COUNTERS
 178 # include <asm/perf_counter.h>
 179 #endif
 180
 181 #include <linux/list.h>
 182 #include <linux/mutex.h>
 183 #include <linux/rculist.h>
 184 #include <linux/rcupdate.h>
 185 #include <linux/spinlock.h>
 186 #include <linux/hrtimer.h>
 187 #include <asm/atomic.h>
 188
 189 struct task_struct;
 190
 191 static inline u64 perf_event_raw(struct perf_counter_hw_event *hw_event)
 192 {
 193         return hw_event->config & PERF_COUNTER_RAW_MASK;
 194 }
 195
 196 static inline u64 perf_event_config(struct perf_counter_hw_event *hw_event)
 197 {
 198         return hw_event->config & PERF_COUNTER_CONFIG_MASK;
 199 }
 200
 201 static inline u64 perf_event_type(struct perf_counter_hw_event *hw_event)
 202 {
 203         return (hw_event->config & PERF_COUNTER_TYPE_MASK) >>
 204                 PERF_COUNTER_TYPE_SHIFT;
 205 }
 206
 207 static inline u64 perf_event_id(struct perf_counter_hw_event *hw_event)
 208 {
 209         return hw_event->config & PERF_COUNTER_EVENT_MASK;
 210 }
 211
 212 /**
 213  * struct hw_perf_counter - performance counter hardware details:
 214  */
 215 struct hw_perf_counter {
 216 #ifdef CONFIG_PERF_COUNTERS
 217         union {
 218                 struct { /* hardware */
 219                         u64                             config;
 220                         unsigned long                   config_base;
 221                         unsigned long                   counter_base;
 222                         int                             nmi;
 223                         unsigned int                    idx;
 224                 };
 225                 union { /* software */
 226                         atomic64_t                      count;
 227                         struct hrtimer                  hrtimer;
 228                 };
 229         };
 230         atomic64_t                      prev_count;
 231         u64                             irq_period;
 232         atomic64_t                      period_left;
 233 #endif
 234 };
 235
 236 struct perf_counter;
 237
 238 /**
 239  * struct hw_perf_counter_ops - performance counter hw ops
 240  */
 241 struct hw_perf_counter_ops {
 242         int (*enable)                   (struct perf_counter *counter);
 243         void (*disable)                 (struct perf_counter *counter);
 244         void (*read)                    (struct perf_counter *counter);
 245 };
 246
 247 /**
 248  * enum perf_counter_active_state - the states of a counter
 249  */
 250 enum perf_counter_active_state {
 251         PERF_COUNTER_STATE_ERROR        = -2,
 252         PERF_COUNTER_STATE_OFF          = -1,
 253         PERF_COUNTER_STATE_INACTIVE     =  0,
 254         PERF_COUNTER_STATE_ACTIVE       =  1,
 255 };
 256
 257 struct file;
 258
 259 struct perf_mmap_data {
 260         struct rcu_head                 rcu_head;
 261         int                             nr_pages;
 262         atomic_t                        wakeup;
 263         atomic_t                        head;
 264         struct perf_counter_mmap_page   *user_page;
 265         void                            *data_pages[0];
 266 };
 267
 268 /**
 269  * struct perf_counter - performance counter kernel representation:
 270  */
 271 struct perf_counter {
 272 #ifdef CONFIG_PERF_COUNTERS
 273         struct list_head                list_entry;
 274         struct list_head                event_entry;
 275         struct list_head                sibling_list;
 276         int                             nr_siblings;
 277         struct perf_counter             *group_leader;
 278         const struct hw_perf_counter_ops *hw_ops;
 279
 280         enum perf_counter_active_state  state;
 281         enum perf_counter_active_state  prev_state;
 282         atomic64_t                      count;
 283
 284         struct perf_counter_hw_event    hw_event;
 285         struct hw_perf_counter          hw;
 286
 287         struct perf_counter_context     *ctx;
 288         struct task_struct              *task;
 289         struct file                     *filp;
 290
 291         struct perf_counter             *parent;
 292         struct list_head                child_list;
 293
 294         /*
 295          * Protect attach/detach and child_list:
 296          */
 297         struct mutex                    mutex;
 298
 299         int                             oncpu;
 300         int                             cpu;
 301
 302         /* mmap bits */
 303         struct mutex                    mmap_mutex;
 304         atomic_t                        mmap_count;
 305         struct perf_mmap_data           *data;
 306
 307         /* poll related */
 308         wait_queue_head_t               waitq;
 309         /* optional: for NMIs */
 310         int                             wakeup_pending;
 311
 312         void (*destroy)(struct perf_counter *);
 313         struct rcu_head                 rcu_head;
 314 #endif
 315 };
 316
 317 /**
 318  * struct perf_counter_context - counter context structure
 319  *
 320  * Used as a container for task counters and CPU counters as well:
 321  */
 322 struct perf_counter_context {
 323 #ifdef CONFIG_PERF_COUNTERS
 324         /*
 325          * Protect the states of the counters in the list,
 326          * nr_active, and the list:
 327          */
 328         spinlock_t              lock;
 329         /*
 330          * Protect the list of counters.  Locking either mutex or lock
 331          * is sufficient to ensure the list doesn't change; to change
 332          * the list you need to lock both the mutex and the spinlock.
 333          */
 334         struct mutex            mutex;
 335
 336         struct list_head        counter_list;
 337         struct list_head        event_list;
 338         int                     nr_counters;
 339         int                     nr_active;
 340         int                     is_active;
 341         struct task_struct      *task;
 342 #endif
 343 };
 344
 345 /**
 346  * struct perf_counter_cpu_context - per cpu counter context structure
 347  */
 348 struct perf_cpu_context {
 349         struct perf_counter_context     ctx;
 350         struct perf_counter_context     *task_ctx;
 351         int                             active_oncpu;
 352         int                             max_pertask;
 353         int                             exclusive;
 354
 355         /*
 356          * Recursion avoidance:
 357          *
 358          * task, softirq, irq, nmi context
 359          */
 360         int                     recursion[4];
 361 };
 362
 363 /*
 364  * Set by architecture code:
 365  */
 366 extern int perf_max_counters;
 367
 368 #ifdef CONFIG_PERF_COUNTERS
 369 extern const struct hw_perf_counter_ops *
 370 hw_perf_counter_init(struct perf_counter *counter);
 371
 372 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
 373 extern void perf_counter_task_sched_out(struct task_struct *task, int cpu);
 374 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
 375 extern void perf_counter_init_task(struct task_struct *child);
 376 extern void perf_counter_exit_task(struct task_struct *child);
 377 extern void perf_counter_notify(struct pt_regs *regs);
 378 extern void perf_counter_print_debug(void);
 379 extern void perf_counter_unthrottle(void);
 380 extern u64 hw_perf_save_disable(void);
 381 extern void hw_perf_restore(u64 ctrl);
 382 extern int perf_counter_task_disable(void);
 383 extern int perf_counter_task_enable(void);
 384 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
 385                struct perf_cpu_context *cpuctx,
 386                struct perf_counter_context *ctx, int cpu);
 387 extern void perf_counter_update_userpage(struct perf_counter *counter);
 388
 389 extern void perf_counter_output(struct perf_counter *counter,
 390                                 int nmi, struct pt_regs *regs);
 391 /*
 392  * Return 1 for a software counter, 0 for a hardware counter
 393  */
 394 static inline int is_software_counter(struct perf_counter *counter)
 395 {
 396         return !perf_event_raw(&counter->hw_event) &&
 397                 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
 398 }
 399
 400 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *);
 401
 402 #else
 403 static inline void
 404 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
 405 static inline void
 406 perf_counter_task_sched_out(struct task_struct *task, int cpu)          { }
 407 static inline void
 408 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
 409 static inline void perf_counter_init_task(struct task_struct *child)    { }
 410 static inline void perf_counter_exit_task(struct task_struct *child)    { }
 411 static inline void perf_counter_notify(struct pt_regs *regs)            { }
 412 static inline void perf_counter_print_debug(void)                       { }
 413 static inline void perf_counter_unthrottle(void)                        { }
 414 static inline void hw_perf_restore(u64 ctrl)                            { }
 415 static inline u64 hw_perf_save_disable(void)                  { return 0; }
 416 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
 417 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
 418
 419 static inline void perf_swcounter_event(u32 event, u64 nr,
 420                                         int nmi, struct pt_regs *regs)  { }
 421 #endif
 422
 423 #endif /* __KERNEL__ */
 424 #endif /* _LINUX_PERF_COUNTER_H */