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