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
 131                                 __reserved_1   : 55;
 132
 133         __u32                   extra_config_len;
 134         __u32                   __reserved_4;
 135
 136         __u64                   __reserved_2;
 137         __u64                   __reserved_3;
 138 };
 139
 140 /*
 141  * Ioctls that can be done on a perf counter fd:
 142  */
 143 #define PERF_COUNTER_IOC_ENABLE         _IO('$', 0)
 144 #define PERF_COUNTER_IOC_DISABLE        _IO('$', 1)
 145
 146 /*
 147  * Structure of the page that can be mapped via mmap
 148  */
 149 struct perf_counter_mmap_page {
 150         __u32   version;                /* version number of this structure */
 151         __u32   compat_version;         /* lowest version this is compat with */
 152         __u32   lock;                   /* seqlock for synchronization */
 153         __u32   index;                  /* hardware counter identifier */
 154         __s64   offset;                 /* add to hardware counter value */
 155
 156         __u32   data_head;              /* head in the data section */
 157 };
 158
 159 #ifdef __KERNEL__
 160 /*
 161  * Kernel-internal data types and definitions:
 162  */
 163
 164 #ifdef CONFIG_PERF_COUNTERS
 165 # include <asm/perf_counter.h>
 166 #endif
 167
 168 #include <linux/list.h>
 169 #include <linux/mutex.h>
 170 #include <linux/rculist.h>
 171 #include <linux/rcupdate.h>
 172 #include <linux/spinlock.h>
 173 #include <linux/hrtimer.h>
 174 #include <asm/atomic.h>
 175
 176 struct task_struct;
 177
 178 static inline u64 perf_event_raw(struct perf_counter_hw_event *hw_event)
 179 {
 180         return hw_event->config & PERF_COUNTER_RAW_MASK;
 181 }
 182
 183 static inline u64 perf_event_config(struct perf_counter_hw_event *hw_event)
 184 {
 185         return hw_event->config & PERF_COUNTER_CONFIG_MASK;
 186 }
 187
 188 static inline u64 perf_event_type(struct perf_counter_hw_event *hw_event)
 189 {
 190         return (hw_event->config & PERF_COUNTER_TYPE_MASK) >>
 191                 PERF_COUNTER_TYPE_SHIFT;
 192 }
 193
 194 static inline u64 perf_event_id(struct perf_counter_hw_event *hw_event)
 195 {
 196         return hw_event->config & PERF_COUNTER_EVENT_MASK;
 197 }
 198
 199 /**
 200  * struct hw_perf_counter - performance counter hardware details:
 201  */
 202 struct hw_perf_counter {
 203 #ifdef CONFIG_PERF_COUNTERS
 204         union {
 205                 struct { /* hardware */
 206                         u64                             config;
 207                         unsigned long                   config_base;
 208                         unsigned long                   counter_base;
 209                         int                             nmi;
 210                         unsigned int                    idx;
 211                 };
 212                 union { /* software */
 213                         atomic64_t                      count;
 214                         struct hrtimer                  hrtimer;
 215                 };
 216         };
 217         atomic64_t                      prev_count;
 218         u64                             irq_period;
 219         atomic64_t                      period_left;
 220 #endif
 221 };
 222
 223 struct perf_counter;
 224
 225 /**
 226  * struct hw_perf_counter_ops - performance counter hw ops
 227  */
 228 struct hw_perf_counter_ops {
 229         int (*enable)                   (struct perf_counter *counter);
 230         void (*disable)                 (struct perf_counter *counter);
 231         void (*read)                    (struct perf_counter *counter);
 232 };
 233
 234 /**
 235  * enum perf_counter_active_state - the states of a counter
 236  */
 237 enum perf_counter_active_state {
 238         PERF_COUNTER_STATE_ERROR        = -2,
 239         PERF_COUNTER_STATE_OFF          = -1,
 240         PERF_COUNTER_STATE_INACTIVE     =  0,
 241         PERF_COUNTER_STATE_ACTIVE       =  1,
 242 };
 243
 244 struct file;
 245
 246 struct perf_mmap_data {
 247         struct rcu_head                 rcu_head;
 248         int                             nr_pages;
 249         atomic_t                        head;
 250         struct perf_counter_mmap_page   *user_page;
 251         void                            *data_pages[0];
 252 };
 253
 254 /**
 255  * struct perf_counter - performance counter kernel representation:
 256  */
 257 struct perf_counter {
 258 #ifdef CONFIG_PERF_COUNTERS
 259         struct list_head                list_entry;
 260         struct list_head                event_entry;
 261         struct list_head                sibling_list;
 262         struct perf_counter             *group_leader;
 263         const struct hw_perf_counter_ops *hw_ops;
 264
 265         enum perf_counter_active_state  state;
 266         enum perf_counter_active_state  prev_state;
 267         atomic64_t                      count;
 268
 269         struct perf_counter_hw_event    hw_event;
 270         struct hw_perf_counter          hw;
 271
 272         struct perf_counter_context     *ctx;
 273         struct task_struct              *task;
 274         struct file                     *filp;
 275
 276         struct perf_counter             *parent;
 277         struct list_head                child_list;
 278
 279         /*
 280          * Protect attach/detach and child_list:
 281          */
 282         struct mutex                    mutex;
 283
 284         int                             oncpu;
 285         int                             cpu;
 286
 287         /* mmap bits */
 288         struct mutex                    mmap_mutex;
 289         atomic_t                        mmap_count;
 290         struct perf_mmap_data           *data;
 291
 292         /* poll related */
 293         wait_queue_head_t               waitq;
 294         /* optional: for NMIs */
 295         int                             wakeup_pending;
 296
 297         void (*destroy)(struct perf_counter *);
 298         struct rcu_head                 rcu_head;
 299 #endif
 300 };
 301
 302 /**
 303  * struct perf_counter_context - counter context structure
 304  *
 305  * Used as a container for task counters and CPU counters as well:
 306  */
 307 struct perf_counter_context {
 308 #ifdef CONFIG_PERF_COUNTERS
 309         /*
 310          * Protect the states of the counters in the list,
 311          * nr_active, and the list:
 312          */
 313         spinlock_t              lock;
 314         /*
 315          * Protect the list of counters.  Locking either mutex or lock
 316          * is sufficient to ensure the list doesn't change; to change
 317          * the list you need to lock both the mutex and the spinlock.
 318          */
 319         struct mutex            mutex;
 320
 321         struct list_head        counter_list;
 322         struct list_head        event_list;
 323         int                     nr_counters;
 324         int                     nr_active;
 325         int                     is_active;
 326         struct task_struct      *task;
 327 #endif
 328 };
 329
 330 /**
 331  * struct perf_counter_cpu_context - per cpu counter context structure
 332  */
 333 struct perf_cpu_context {
 334         struct perf_counter_context     ctx;
 335         struct perf_counter_context     *task_ctx;
 336         int                             active_oncpu;
 337         int                             max_pertask;
 338         int                             exclusive;
 339
 340         /*
 341          * Recursion avoidance:
 342          *
 343          * task, softirq, irq, nmi context
 344          */
 345         int                     recursion[4];
 346 };
 347
 348 /*
 349  * Set by architecture code:
 350  */
 351 extern int perf_max_counters;
 352
 353 #ifdef CONFIG_PERF_COUNTERS
 354 extern const struct hw_perf_counter_ops *
 355 hw_perf_counter_init(struct perf_counter *counter);
 356
 357 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
 358 extern void perf_counter_task_sched_out(struct task_struct *task, int cpu);
 359 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
 360 extern void perf_counter_init_task(struct task_struct *child);
 361 extern void perf_counter_exit_task(struct task_struct *child);
 362 extern void perf_counter_notify(struct pt_regs *regs);
 363 extern void perf_counter_print_debug(void);
 364 extern void perf_counter_unthrottle(void);
 365 extern u64 hw_perf_save_disable(void);
 366 extern void hw_perf_restore(u64 ctrl);
 367 extern int perf_counter_task_disable(void);
 368 extern int perf_counter_task_enable(void);
 369 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
 370                struct perf_cpu_context *cpuctx,
 371                struct perf_counter_context *ctx, int cpu);
 372 extern void perf_counter_update_userpage(struct perf_counter *counter);
 373
 374 extern void perf_counter_output(struct perf_counter *counter,
 375                                 int nmi, struct pt_regs *regs);
 376 /*
 377  * Return 1 for a software counter, 0 for a hardware counter
 378  */
 379 static inline int is_software_counter(struct perf_counter *counter)
 380 {
 381         return !perf_event_raw(&counter->hw_event) &&
 382                 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
 383 }
 384
 385 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *);
 386
 387 #else
 388 static inline void
 389 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
 390 static inline void
 391 perf_counter_task_sched_out(struct task_struct *task, int cpu)          { }
 392 static inline void
 393 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
 394 static inline void perf_counter_init_task(struct task_struct *child)    { }
 395 static inline void perf_counter_exit_task(struct task_struct *child)    { }
 396 static inline void perf_counter_notify(struct pt_regs *regs)            { }
 397 static inline void perf_counter_print_debug(void)                       { }
 398 static inline void perf_counter_unthrottle(void)                        { }
 399 static inline void hw_perf_restore(u64 ctrl)                            { }
 400 static inline u64 hw_perf_save_disable(void)                  { return 0; }
 401 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
 402 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
 403
 404 static inline void perf_swcounter_event(u32 event, u64 nr,
 405                                         int nmi, struct pt_regs *regs)  { }
 406 #endif
 407
 408 #endif /* __KERNEL__ */
 409 #endif /* _LINUX_PERF_COUNTER_H */