include/linux/tracepoint.h

   1 #ifndef _LINUX_TRACEPOINT_H
   2 #define _LINUX_TRACEPOINT_H
   3
   4 /*
   5  * Kernel Tracepoint API.
   6  *
   7  * See Documentation/trace/tracepoints.txt.
   8  *
   9  * Copyright (C) 2008-2014 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
  10  *
  11  * Heavily inspired from the Linux Kernel Markers.
  12  *
  13  * This file is released under the GPLv2.
  14  * See the file COPYING for more details.
  15  */
  16
  17 #include <linux/smp.h>
  18 #include <linux/errno.h>
  19 #include <linux/types.h>
  20 #include <linux/cpumask.h>
  21 #include <linux/rcupdate.h>
  22 #include <linux/static_key.h>
  23
  24 struct module;
  25 struct tracepoint;
  26 struct notifier_block;
  27
  28 struct tracepoint_func {
  29         void *func;
  30         void *data;
  31         int prio;
  32 };
  33
  34 struct tracepoint {
  35         const char *name;               /* Tracepoint name */
  36         struct static_key key;
  37         void (*regfunc)(void);
  38         void (*unregfunc)(void);
  39         struct tracepoint_func __rcu *funcs;
  40 };
  41
  42 struct trace_enum_map {
  43         const char              *system;
  44         const char              *enum_string;
  45         unsigned long           enum_value;
  46 };
  47
  48 #define TRACEPOINT_DEFAULT_PRIO 10
  49
  50 extern int
  51 tracepoint_probe_register(struct tracepoint *tp, void *probe, void *data);
  52 extern int
  53 tracepoint_probe_register_prio(struct tracepoint *tp, void *probe, void *data,
  54                                int prio);
  55 extern int
  56 tracepoint_probe_unregister(struct tracepoint *tp, void *probe, void *data);
  57 extern void
  58 for_each_kernel_tracepoint(void (*fct)(struct tracepoint *tp, void *priv),
  59                 void *priv);
  60
  61 #ifdef CONFIG_MODULES
  62 struct tp_module {
  63         struct list_head list;
  64         struct module *mod;
  65 };
  66
  67 bool trace_module_has_bad_taint(struct module *mod);
  68 extern int register_tracepoint_module_notifier(struct notifier_block *nb);
  69 extern int unregister_tracepoint_module_notifier(struct notifier_block *nb);
  70 #else
  71 static inline bool trace_module_has_bad_taint(struct module *mod)
  72 {
  73         return false;
  74 }
  75 static inline
  76 int register_tracepoint_module_notifier(struct notifier_block *nb)
  77 {
  78         return 0;
  79 }
  80 static inline
  81 int unregister_tracepoint_module_notifier(struct notifier_block *nb)
  82 {
  83         return 0;
  84 }
  85 #endif /* CONFIG_MODULES */
  86
  87 /*
  88  * tracepoint_synchronize_unregister must be called between the last tracepoint
  89  * probe unregistration and the end of module exit to make sure there is no
  90  * caller executing a probe when it is freed.
  91  */
  92 static inline void tracepoint_synchronize_unregister(void)
  93 {
  94         synchronize_sched();
  95 }
  96
  97 #ifdef CONFIG_HAVE_SYSCALL_TRACEPOINTS
  98 extern void syscall_regfunc(void);
  99 extern void syscall_unregfunc(void);
 100 #endif /* CONFIG_HAVE_SYSCALL_TRACEPOINTS */
 101
 102 #define PARAMS(args...) args
 103
 104 #define TRACE_DEFINE_ENUM(x)
 105
 106 #endif /* _LINUX_TRACEPOINT_H */
 107
 108 /*
 109  * Note: we keep the TRACE_EVENT and DECLARE_TRACE outside the include
 110  *  file ifdef protection.
 111  *  This is due to the way trace events work. If a file includes two
 112  *  trace event headers under one "CREATE_TRACE_POINTS" the first include
 113  *  will override the TRACE_EVENT and break the second include.
 114  */
 115
 116 #ifndef DECLARE_TRACE
 117
 118 #define TP_PROTO(args...)       args
 119 #define TP_ARGS(args...)        args
 120 #define TP_CONDITION(args...)   args
 121
 122 /*
 123  * Individual subsystem my have a separate configuration to
 124  * enable their tracepoints. By default, this file will create
 125  * the tracepoints if CONFIG_TRACEPOINT is defined. If a subsystem
 126  * wants to be able to disable its tracepoints from being created
 127  * it can define NOTRACE before including the tracepoint headers.
 128  */
 129 #if defined(CONFIG_TRACEPOINTS) && !defined(NOTRACE)
 130 #define TRACEPOINTS_ENABLED
 131 #endif
 132
 133 #ifdef TRACEPOINTS_ENABLED
 134
 135 /*
 136  * it_func[0] is never NULL because there is at least one element in the array
 137  * when the array itself is non NULL.
 138  *
 139  * Note, the proto and args passed in includes "__data" as the first parameter.
 140  * The reason for this is to handle the "void" prototype. If a tracepoint
 141  * has a "void" prototype, then it is invalid to declare a function
 142  * as "(void *, void)". The DECLARE_TRACE_NOARGS() will pass in just
 143  * "void *data", where as the DECLARE_TRACE() will pass in "void *data, proto".
 144  */
 145 #define __DO_TRACE(tp, proto, args, cond, prercu, postrcu)              \
 146         do {                                                            \
 147                 struct tracepoint_func *it_func_ptr;                    \
 148                 void *it_func;                                          \
 149                 void *__data;                                           \
 150                                                                         \
 151                 if (!cpu_online(raw_smp_processor_id()))                \
 152                         return;                                         \
 153                                                                         \
 154                 if (!(cond))                                            \
 155                         return;                                         \
 156                 prercu;                                                 \
 157                 rcu_read_lock_sched_notrace();                          \
 158                 it_func_ptr = rcu_dereference_sched((tp)->funcs);       \
 159                 if (it_func_ptr) {                                      \
 160                         do {                                            \
 161                                 it_func = (it_func_ptr)->func;          \
 162                                 __data = (it_func_ptr)->data;           \
 163                                 ((void(*)(proto))(it_func))(args);      \
 164                         } while ((++it_func_ptr)->func);                \
 165                 }                                                       \
 166                 rcu_read_unlock_sched_notrace();                        \
 167                 postrcu;                                                \
 168         } while (0)
 169
 170 #ifndef MODULE
 171 #define __DECLARE_TRACE_RCU(name, proto, args, cond, data_proto, data_args)     \
 172         static inline void trace_##name##_rcuidle(proto)                \
 173         {                                                               \
 174                 if (static_key_false(&__tracepoint_##name.key))         \
 175                         __DO_TRACE(&__tracepoint_##name,                \
 176                                 TP_PROTO(data_proto),                   \
 177                                 TP_ARGS(data_args),                     \
 178                                 TP_CONDITION(cond),                     \
 179                                 rcu_irq_enter(),                        \
 180                                 rcu_irq_exit());                        \
 181         }
 182 #else
 183 #define __DECLARE_TRACE_RCU(name, proto, args, cond, data_proto, data_args)
 184 #endif
 185
 186 /*
 187  * Make sure the alignment of the structure in the __tracepoints section will
 188  * not add unwanted padding between the beginning of the section and the
 189  * structure. Force alignment to the same alignment as the section start.
 190  *
 191  * When lockdep is enabled, we make sure to always do the RCU portions of
 192  * the tracepoint code, regardless of whether tracing is on. However,
 193  * don't check if the condition is false, due to interaction with idle
 194  * instrumentation. This lets us find RCU issues triggered with tracepoints
 195  * even when this tracepoint is off. This code has no purpose other than
 196  * poking RCU a bit.
 197  */
 198 #define __DECLARE_TRACE(name, proto, args, cond, data_proto, data_args) \
 199         extern struct tracepoint __tracepoint_##name;                   \
 200         static inline void trace_##name(proto)                          \
 201         {                                                               \
 202                 if (static_key_false(&__tracepoint_##name.key))         \
 203                         __DO_TRACE(&__tracepoint_##name,                \
 204                                 TP_PROTO(data_proto),                   \
 205                                 TP_ARGS(data_args),                     \
 206                                 TP_CONDITION(cond),,);                  \
 207                 if (IS_ENABLED(CONFIG_LOCKDEP) && (cond)) {             \
 208                         rcu_read_lock_sched_notrace();                  \
 209                         rcu_dereference_sched(__tracepoint_##name.funcs);\
 210                         rcu_read_unlock_sched_notrace();                \
 211                 }                                                       \
 212         }                                                               \
 213         __DECLARE_TRACE_RCU(name, PARAMS(proto), PARAMS(args),          \
 214                 PARAMS(cond), PARAMS(data_proto), PARAMS(data_args))    \
 215         static inline int                                               \
 216         register_trace_##name(void (*probe)(data_proto), void *data)    \
 217         {                                                               \
 218                 return tracepoint_probe_register(&__tracepoint_##name,  \
 219                                                 (void *)probe, data);   \
 220         }                                                               \
 221         static inline int                                               \
 222         register_trace_prio_##name(void (*probe)(data_proto), void *data,\
 223                                    int prio)                            \
 224         {                                                               \
 225                 return tracepoint_probe_register_prio(&__tracepoint_##name, \
 226                                               (void *)probe, data, prio); \
 227         }                                                               \
 228         static inline int                                               \
 229         unregister_trace_##name(void (*probe)(data_proto), void *data)  \
 230         {                                                               \
 231                 return tracepoint_probe_unregister(&__tracepoint_##name,\
 232                                                 (void *)probe, data);   \
 233         }                                                               \
 234         static inline void                                              \
 235         check_trace_callback_type_##name(void (*cb)(data_proto))        \
 236         {                                                               \
 237         }                                                               \
 238         static inline bool                                              \
 239         trace_##name##_enabled(void)                                    \
 240         {                                                               \
 241                 return static_key_false(&__tracepoint_##name.key);      \
 242         }
 243
 244 /*
 245  * We have no guarantee that gcc and the linker won't up-align the tracepoint
 246  * structures, so we create an array of pointers that will be used for iteration
 247  * on the tracepoints.
 248  */
 249 #define DEFINE_TRACE_FN(name, reg, unreg)                                \
 250         static const char __tpstrtab_##name[]                            \
 251         __attribute__((section("__tracepoints_strings"))) = #name;       \
 252         struct tracepoint __tracepoint_##name                            \
 253         __attribute__((section("__tracepoints"))) =                      \
 254                 { __tpstrtab_##name, STATIC_KEY_INIT_FALSE, reg, unreg, NULL };\
 255         static struct tracepoint * const __tracepoint_ptr_##name __used  \
 256         __attribute__((section("__tracepoints_ptrs"))) =                 \
 257                 &__tracepoint_##name;
 258
 259 #define DEFINE_TRACE(name)                                              \
 260         DEFINE_TRACE_FN(name, NULL, NULL);
 261
 262 #define EXPORT_TRACEPOINT_SYMBOL_GPL(name)                              \
 263         EXPORT_SYMBOL_GPL(__tracepoint_##name)
 264 #define EXPORT_TRACEPOINT_SYMBOL(name)                                  \
 265         EXPORT_SYMBOL(__tracepoint_##name)
 266
 267 #else /* !TRACEPOINTS_ENABLED */
 268 #define __DECLARE_TRACE(name, proto, args, cond, data_proto, data_args) \
 269         static inline void trace_##name(proto)                          \
 270         { }                                                             \
 271         static inline void trace_##name##_rcuidle(proto)                \
 272         { }                                                             \
 273         static inline int                                               \
 274         register_trace_##name(void (*probe)(data_proto),                \
 275                               void *data)                               \
 276         {                                                               \
 277                 return -ENOSYS;                                         \
 278         }                                                               \
 279         static inline int                                               \
 280         unregister_trace_##name(void (*probe)(data_proto),              \
 281                                 void *data)                             \
 282         {                                                               \
 283                 return -ENOSYS;                                         \
 284         }                                                               \
 285         static inline void check_trace_callback_type_##name(void (*cb)(data_proto)) \
 286         {                                                               \
 287         }                                                               \
 288         static inline bool                                              \
 289         trace_##name##_enabled(void)                                    \
 290         {                                                               \
 291                 return false;                                           \
 292         }
 293
 294 #define DEFINE_TRACE_FN(name, reg, unreg)
 295 #define DEFINE_TRACE(name)
 296 #define EXPORT_TRACEPOINT_SYMBOL_GPL(name)
 297 #define EXPORT_TRACEPOINT_SYMBOL(name)
 298
 299 #endif /* TRACEPOINTS_ENABLED */
 300
 301 #ifdef CONFIG_TRACING
 302 /**
 303  * tracepoint_string - register constant persistent string to trace system
 304  * @str - a constant persistent string that will be referenced in tracepoints
 305  *
 306  * If constant strings are being used in tracepoints, it is faster and
 307  * more efficient to just save the pointer to the string and reference
 308  * that with a printf "%s" instead of saving the string in the ring buffer
 309  * and wasting space and time.
 310  *
 311  * The problem with the above approach is that userspace tools that read
 312  * the binary output of the trace buffers do not have access to the string.
 313  * Instead they just show the address of the string which is not very
 314  * useful to users.
 315  *
 316  * With tracepoint_string(), the string will be registered to the tracing
 317  * system and exported to userspace via the debugfs/tracing/printk_formats
 318  * file that maps the string address to the string text. This way userspace
 319  * tools that read the binary buffers have a way to map the pointers to
 320  * the ASCII strings they represent.
 321  *
 322  * The @str used must be a constant string and persistent as it would not
 323  * make sense to show a string that no longer exists. But it is still fine
 324  * to be used with modules, because when modules are unloaded, if they
 325  * had tracepoints, the ring buffers are cleared too. As long as the string
 326  * does not change during the life of the module, it is fine to use
 327  * tracepoint_string() within a module.
 328  */
 329 #define tracepoint_string(str)                                          \
 330         ({                                                              \
 331                 static const char *___tp_str __tracepoint_string = str; \
 332                 ___tp_str;                                              \
 333         })
 334 #define __tracepoint_string     __attribute__((section("__tracepoint_str")))
 335 #else
 336 /*
 337  * tracepoint_string() is used to save the string address for userspace
 338  * tracing tools. When tracing isn't configured, there's no need to save
 339  * anything.
 340  */
 341 # define tracepoint_string(str) str
 342 # define __tracepoint_string
 343 #endif
 344
 345 /*
 346  * The need for the DECLARE_TRACE_NOARGS() is to handle the prototype
 347  * (void). "void" is a special value in a function prototype and can
 348  * not be combined with other arguments. Since the DECLARE_TRACE()
 349  * macro adds a data element at the beginning of the prototype,
 350  * we need a way to differentiate "(void *data, proto)" from
 351  * "(void *data, void)". The second prototype is invalid.
 352  *
 353  * DECLARE_TRACE_NOARGS() passes "void" as the tracepoint prototype
 354  * and "void *__data" as the callback prototype.
 355  *
 356  * DECLARE_TRACE() passes "proto" as the tracepoint protoype and
 357  * "void *__data, proto" as the callback prototype.
 358  */
 359 #define DECLARE_TRACE_NOARGS(name)                                      \
 360                 __DECLARE_TRACE(name, void, , 1, void *__data, __data)
 361
 362 #define DECLARE_TRACE(name, proto, args)                                \
 363                 __DECLARE_TRACE(name, PARAMS(proto), PARAMS(args), 1,   \
 364                                 PARAMS(void *__data, proto),            \
 365                                 PARAMS(__data, args))
 366
 367 #define DECLARE_TRACE_CONDITION(name, proto, args, cond)                \
 368         __DECLARE_TRACE(name, PARAMS(proto), PARAMS(args), PARAMS(cond), \
 369                         PARAMS(void *__data, proto),                    \
 370                         PARAMS(__data, args))
 371
 372 #define TRACE_EVENT_FLAGS(event, flag)
 373
 374 #define TRACE_EVENT_PERF_PERM(event, expr...)
 375
 376 #endif /* DECLARE_TRACE */
 377
 378 #ifndef TRACE_EVENT
 379 /*
 380  * For use with the TRACE_EVENT macro:
 381  *
 382  * We define a tracepoint, its arguments, its printk format
 383  * and its 'fast binary record' layout.
 384  *
 385  * Firstly, name your tracepoint via TRACE_EVENT(name : the
 386  * 'subsystem_event' notation is fine.
 387  *
 388  * Think about this whole construct as the
 389  * 'trace_sched_switch() function' from now on.
 390  *
 391  *
 392  *  TRACE_EVENT(sched_switch,
 393  *
 394  *      *
 395  *      * A function has a regular function arguments
 396  *      * prototype, declare it via TP_PROTO():
 397  *      *
 398  *
 399  *      TP_PROTO(struct rq *rq, struct task_struct *prev,
 400  *               struct task_struct *next),
 401  *
 402  *      *
 403  *      * Define the call signature of the 'function'.
 404  *      * (Design sidenote: we use this instead of a
 405  *      *  TP_PROTO1/TP_PROTO2/TP_PROTO3 ugliness.)
 406  *      *
 407  *
 408  *      TP_ARGS(rq, prev, next),
 409  *
 410  *      *
 411  *      * Fast binary tracing: define the trace record via
 412  *      * TP_STRUCT__entry(). You can think about it like a
 413  *      * regular C structure local variable definition.
 414  *      *
 415  *      * This is how the trace record is structured and will
 416  *      * be saved into the ring buffer. These are the fields
 417  *      * that will be exposed to user-space in
 418  *      * /sys/kernel/debug/tracing/events/<*>/format.
 419  *      *
 420  *      * The declared 'local variable' is called '__entry'
 421  *      *
 422  *      * __field(pid_t, prev_prid) is equivalent to a standard declariton:
 423  *      *
 424  *      *       pid_t   prev_pid;
 425  *      *
 426  *      * __array(char, prev_comm, TASK_COMM_LEN) is equivalent to:
 427  *      *
 428  *      *       char    prev_comm[TASK_COMM_LEN];
 429  *      *
 430  *
 431  *      TP_STRUCT__entry(
 432  *              __array(        char,   prev_comm,      TASK_COMM_LEN   )
 433  *              __field(        pid_t,  prev_pid                        )
 434  *              __field(        int,    prev_prio                       )
 435  *              __array(        char,   next_comm,      TASK_COMM_LEN   )
 436  *              __field(        pid_t,  next_pid                        )
 437  *              __field(        int,    next_prio                       )
 438  *      ),
 439  *
 440  *      *
 441  *      * Assign the entry into the trace record, by embedding
 442  *      * a full C statement block into TP_fast_assign(). You
 443  *      * can refer to the trace record as '__entry' -
 444  *      * otherwise you can put arbitrary C code in here.
 445  *      *
 446  *      * Note: this C code will execute every time a trace event
 447  *      * happens, on an active tracepoint.
 448  *      *
 449  *
 450  *      TP_fast_assign(
 451  *              memcpy(__entry->next_comm, next->comm, TASK_COMM_LEN);
 452  *              __entry->prev_pid       = prev->pid;
 453  *              __entry->prev_prio      = prev->prio;
 454  *              memcpy(__entry->prev_comm, prev->comm, TASK_COMM_LEN);
 455  *              __entry->next_pid       = next->pid;
 456  *              __entry->next_prio      = next->prio;
 457  *      ),
 458  *
 459  *      *
 460  *      * Formatted output of a trace record via TP_printk().
 461  *      * This is how the tracepoint will appear under ftrace
 462  *      * plugins that make use of this tracepoint.
 463  *      *
 464  *      * (raw-binary tracing wont actually perform this step.)
 465  *      *
 466  *
 467  *      TP_printk("task %s:%d [%d] ==> %s:%d [%d]",
 468  *              __entry->prev_comm, __entry->prev_pid, __entry->prev_prio,
 469  *              __entry->next_comm, __entry->next_pid, __entry->next_prio),
 470  *
 471  * );
 472  *
 473  * This macro construct is thus used for the regular printk format
 474  * tracing setup, it is used to construct a function pointer based
 475  * tracepoint callback (this is used by programmatic plugins and
 476  * can also by used by generic instrumentation like SystemTap), and
 477  * it is also used to expose a structured trace record in
 478  * /sys/kernel/debug/tracing/events/.
 479  *
 480  * A set of (un)registration functions can be passed to the variant
 481  * TRACE_EVENT_FN to perform any (un)registration work.
 482  */
 483
 484 #define DECLARE_EVENT_CLASS(name, proto, args, tstruct, assign, print)
 485 #define DEFINE_EVENT(template, name, proto, args)               \
 486         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 487 #define DEFINE_EVENT_FN(template, name, proto, args, reg, unreg)\
 488         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 489 #define DEFINE_EVENT_PRINT(template, name, proto, args, print)  \
 490         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 491 #define DEFINE_EVENT_CONDITION(template, name, proto,           \
 492                                args, cond)                      \
 493         DECLARE_TRACE_CONDITION(name, PARAMS(proto),            \
 494                                 PARAMS(args), PARAMS(cond))
 495
 496 #define TRACE_EVENT(name, proto, args, struct, assign, print)   \
 497         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 498 #define TRACE_EVENT_FN(name, proto, args, struct,               \
 499                 assign, print, reg, unreg)                      \
 500         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 501 #define TRACE_EVENT_CONDITION(name, proto, args, cond,          \
 502                               struct, assign, print)            \
 503         DECLARE_TRACE_CONDITION(name, PARAMS(proto),            \
 504                                 PARAMS(args), PARAMS(cond))
 505
 506 #define TRACE_EVENT_FLAGS(event, flag)
 507
 508 #define TRACE_EVENT_PERF_PERM(event, expr...)
 509
 510 #endif /* ifdef TRACE_EVENT (see note above) */