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.rst.
   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/srcu.h>
  19 #include <linux/errno.h>
  20 #include <linux/types.h>
  21 #include <linux/cpumask.h>
  22 #include <linux/rcupdate.h>
  23 #include <linux/tracepoint-defs.h>
  24
  25 struct module;
  26 struct tracepoint;
  27 struct notifier_block;
  28
  29 struct trace_eval_map {
  30         const char              *system;
  31         const char              *eval_string;
  32         unsigned long           eval_value;
  33 };
  34
  35 #define TRACEPOINT_DEFAULT_PRIO 10
  36
  37 extern struct srcu_struct tracepoint_srcu;
  38
  39 extern int
  40 tracepoint_probe_register(struct tracepoint *tp, void *probe, void *data);
  41 extern int
  42 tracepoint_probe_register_prio(struct tracepoint *tp, void *probe, void *data,
  43                                int prio);
  44 extern int
  45 tracepoint_probe_unregister(struct tracepoint *tp, void *probe, void *data);
  46 extern void
  47 for_each_kernel_tracepoint(void (*fct)(struct tracepoint *tp, void *priv),
  48                 void *priv);
  49
  50 #ifdef CONFIG_MODULES
  51 struct tp_module {
  52         struct list_head list;
  53         struct module *mod;
  54 };
  55
  56 bool trace_module_has_bad_taint(struct module *mod);
  57 extern int register_tracepoint_module_notifier(struct notifier_block *nb);
  58 extern int unregister_tracepoint_module_notifier(struct notifier_block *nb);
  59 #else
  60 static inline bool trace_module_has_bad_taint(struct module *mod)
  61 {
  62         return false;
  63 }
  64 static inline
  65 int register_tracepoint_module_notifier(struct notifier_block *nb)
  66 {
  67         return 0;
  68 }
  69 static inline
  70 int unregister_tracepoint_module_notifier(struct notifier_block *nb)
  71 {
  72         return 0;
  73 }
  74 #endif /* CONFIG_MODULES */
  75
  76 /*
  77  * tracepoint_synchronize_unregister must be called between the last tracepoint
  78  * probe unregistration and the end of module exit to make sure there is no
  79  * caller executing a probe when it is freed.
  80  */
  81 #ifdef CONFIG_TRACEPOINTS
  82 static inline void tracepoint_synchronize_unregister(void)
  83 {
  84         synchronize_srcu(&tracepoint_srcu);
  85         synchronize_sched();
  86 }
  87 #else
  88 static inline void tracepoint_synchronize_unregister(void)
  89 { }
  90 #endif
  91
  92 #ifdef CONFIG_HAVE_SYSCALL_TRACEPOINTS
  93 extern int syscall_regfunc(void);
  94 extern void syscall_unregfunc(void);
  95 #endif /* CONFIG_HAVE_SYSCALL_TRACEPOINTS */
  96
  97 #define PARAMS(args...) args
  98
  99 #define TRACE_DEFINE_ENUM(x)
 100 #define TRACE_DEFINE_SIZEOF(x)
 101
 102 #endif /* _LINUX_TRACEPOINT_H */
 103
 104 /*
 105  * Note: we keep the TRACE_EVENT and DECLARE_TRACE outside the include
 106  *  file ifdef protection.
 107  *  This is due to the way trace events work. If a file includes two
 108  *  trace event headers under one "CREATE_TRACE_POINTS" the first include
 109  *  will override the TRACE_EVENT and break the second include.
 110  */
 111
 112 #ifndef DECLARE_TRACE
 113
 114 #define TP_PROTO(args...)       args
 115 #define TP_ARGS(args...)        args
 116 #define TP_CONDITION(args...)   args
 117
 118 /*
 119  * Individual subsystem my have a separate configuration to
 120  * enable their tracepoints. By default, this file will create
 121  * the tracepoints if CONFIG_TRACEPOINT is defined. If a subsystem
 122  * wants to be able to disable its tracepoints from being created
 123  * it can define NOTRACE before including the tracepoint headers.
 124  */
 125 #if defined(CONFIG_TRACEPOINTS) && !defined(NOTRACE)
 126 #define TRACEPOINTS_ENABLED
 127 #endif
 128
 129 #ifdef TRACEPOINTS_ENABLED
 130
 131 /*
 132  * it_func[0] is never NULL because there is at least one element in the array
 133  * when the array itself is non NULL.
 134  *
 135  * Note, the proto and args passed in includes "__data" as the first parameter.
 136  * The reason for this is to handle the "void" prototype. If a tracepoint
 137  * has a "void" prototype, then it is invalid to declare a function
 138  * as "(void *, void)". The DECLARE_TRACE_NOARGS() will pass in just
 139  * "void *data", where as the DECLARE_TRACE() will pass in "void *data, proto".
 140  */
 141 #define __DO_TRACE(tp, proto, args, cond, rcuidle)                      \
 142         do {                                                            \
 143                 struct tracepoint_func *it_func_ptr;                    \
 144                 void *it_func;                                          \
 145                 void *__data;                                           \
 146                 int __maybe_unused idx = 0;                             \
 147                                                                         \
 148                 if (!(cond))                                            \
 149                         return;                                         \
 150                                                                         \
 151                 /* srcu can't be used from NMI */                       \
 152                 WARN_ON_ONCE(rcuidle && in_nmi());                      \
 153                                                                         \
 154                 /* keep srcu and sched-rcu usage consistent */          \
 155                 preempt_disable_notrace();                              \
 156                                                                         \
 157                 /*                                                      \
 158                  * For rcuidle callers, use srcu since sched-rcu        \
 159                  * doesn't work from the idle path.                     \
 160                  */                                                     \
 161                 if (rcuidle)                                            \
 162                         idx = srcu_read_lock_notrace(&tracepoint_srcu); \
 163                                                                         \
 164                 it_func_ptr = rcu_dereference_raw((tp)->funcs);         \
 165                                                                         \
 166                 if (it_func_ptr) {                                      \
 167                         do {                                            \
 168                                 it_func = (it_func_ptr)->func;          \
 169                                 __data = (it_func_ptr)->data;           \
 170                                 ((void(*)(proto))(it_func))(args);      \
 171                         } while ((++it_func_ptr)->func);                \
 172                 }                                                       \
 173                                                                         \
 174                 if (rcuidle)                                            \
 175                         srcu_read_unlock_notrace(&tracepoint_srcu, idx);\
 176                                                                         \
 177                 preempt_enable_notrace();                               \
 178         } while (0)
 179
 180 #ifndef MODULE
 181 #define __DECLARE_TRACE_RCU(name, proto, args, cond, data_proto, data_args) \
 182         static inline void trace_##name##_rcuidle(proto)                \
 183         {                                                               \
 184                 if (static_key_false(&__tracepoint_##name.key))         \
 185                         __DO_TRACE(&__tracepoint_##name,                \
 186                                 TP_PROTO(data_proto),                   \
 187                                 TP_ARGS(data_args),                     \
 188                                 TP_CONDITION(cond), 1);                 \
 189         }
 190 #else
 191 #define __DECLARE_TRACE_RCU(name, proto, args, cond, data_proto, data_args)
 192 #endif
 193
 194 /*
 195  * Make sure the alignment of the structure in the __tracepoints section will
 196  * not add unwanted padding between the beginning of the section and the
 197  * structure. Force alignment to the same alignment as the section start.
 198  *
 199  * When lockdep is enabled, we make sure to always do the RCU portions of
 200  * the tracepoint code, regardless of whether tracing is on. However,
 201  * don't check if the condition is false, due to interaction with idle
 202  * instrumentation. This lets us find RCU issues triggered with tracepoints
 203  * even when this tracepoint is off. This code has no purpose other than
 204  * poking RCU a bit.
 205  */
 206 #define __DECLARE_TRACE(name, proto, args, cond, data_proto, data_args) \
 207         extern struct tracepoint __tracepoint_##name;                   \
 208         static inline void trace_##name(proto)                          \
 209         {                                                               \
 210                 if (static_key_false(&__tracepoint_##name.key))         \
 211                         __DO_TRACE(&__tracepoint_##name,                \
 212                                 TP_PROTO(data_proto),                   \
 213                                 TP_ARGS(data_args),                     \
 214                                 TP_CONDITION(cond), 0);                 \
 215                 if (IS_ENABLED(CONFIG_LOCKDEP) && (cond)) {             \
 216                         rcu_read_lock_sched_notrace();                  \
 217                         rcu_dereference_sched(__tracepoint_##name.funcs);\
 218                         rcu_read_unlock_sched_notrace();                \
 219                 }                                                       \
 220         }                                                               \
 221         __DECLARE_TRACE_RCU(name, PARAMS(proto), PARAMS(args),          \
 222                 PARAMS(cond), PARAMS(data_proto), PARAMS(data_args))    \
 223         static inline int                                               \
 224         register_trace_##name(void (*probe)(data_proto), void *data)    \
 225         {                                                               \
 226                 return tracepoint_probe_register(&__tracepoint_##name,  \
 227                                                 (void *)probe, data);   \
 228         }                                                               \
 229         static inline int                                               \
 230         register_trace_prio_##name(void (*probe)(data_proto), void *data,\
 231                                    int prio)                            \
 232         {                                                               \
 233                 return tracepoint_probe_register_prio(&__tracepoint_##name, \
 234                                               (void *)probe, data, prio); \
 235         }                                                               \
 236         static inline int                                               \
 237         unregister_trace_##name(void (*probe)(data_proto), void *data)  \
 238         {                                                               \
 239                 return tracepoint_probe_unregister(&__tracepoint_##name,\
 240                                                 (void *)probe, data);   \
 241         }                                                               \
 242         static inline void                                              \
 243         check_trace_callback_type_##name(void (*cb)(data_proto))        \
 244         {                                                               \
 245         }                                                               \
 246         static inline bool                                              \
 247         trace_##name##_enabled(void)                                    \
 248         {                                                               \
 249                 return static_key_false(&__tracepoint_##name.key);      \
 250         }
 251
 252 /*
 253  * We have no guarantee that gcc and the linker won't up-align the tracepoint
 254  * structures, so we create an array of pointers that will be used for iteration
 255  * on the tracepoints.
 256  */
 257 #define DEFINE_TRACE_FN(name, reg, unreg)                                \
 258         static const char __tpstrtab_##name[]                            \
 259         __attribute__((section("__tracepoints_strings"))) = #name;       \
 260         struct tracepoint __tracepoint_##name                            \
 261         __attribute__((section("__tracepoints"))) =                      \
 262                 { __tpstrtab_##name, STATIC_KEY_INIT_FALSE, reg, unreg, NULL };\
 263         static struct tracepoint * const __tracepoint_ptr_##name __used  \
 264         __attribute__((section("__tracepoints_ptrs"))) =                 \
 265                 &__tracepoint_##name;
 266
 267 #define DEFINE_TRACE(name)                                              \
 268         DEFINE_TRACE_FN(name, NULL, NULL);
 269
 270 #define EXPORT_TRACEPOINT_SYMBOL_GPL(name)                              \
 271         EXPORT_SYMBOL_GPL(__tracepoint_##name)
 272 #define EXPORT_TRACEPOINT_SYMBOL(name)                                  \
 273         EXPORT_SYMBOL(__tracepoint_##name)
 274
 275 #else /* !TRACEPOINTS_ENABLED */
 276 #define __DECLARE_TRACE(name, proto, args, cond, data_proto, data_args) \
 277         static inline void trace_##name(proto)                          \
 278         { }                                                             \
 279         static inline void trace_##name##_rcuidle(proto)                \
 280         { }                                                             \
 281         static inline int                                               \
 282         register_trace_##name(void (*probe)(data_proto),                \
 283                               void *data)                               \
 284         {                                                               \
 285                 return -ENOSYS;                                         \
 286         }                                                               \
 287         static inline int                                               \
 288         unregister_trace_##name(void (*probe)(data_proto),              \
 289                                 void *data)                             \
 290         {                                                               \
 291                 return -ENOSYS;                                         \
 292         }                                                               \
 293         static inline void check_trace_callback_type_##name(void (*cb)(data_proto)) \
 294         {                                                               \
 295         }                                                               \
 296         static inline bool                                              \
 297         trace_##name##_enabled(void)                                    \
 298         {                                                               \
 299                 return false;                                           \
 300         }
 301
 302 #define DEFINE_TRACE_FN(name, reg, unreg)
 303 #define DEFINE_TRACE(name)
 304 #define EXPORT_TRACEPOINT_SYMBOL_GPL(name)
 305 #define EXPORT_TRACEPOINT_SYMBOL(name)
 306
 307 #endif /* TRACEPOINTS_ENABLED */
 308
 309 #ifdef CONFIG_TRACING
 310 /**
 311  * tracepoint_string - register constant persistent string to trace system
 312  * @str - a constant persistent string that will be referenced in tracepoints
 313  *
 314  * If constant strings are being used in tracepoints, it is faster and
 315  * more efficient to just save the pointer to the string and reference
 316  * that with a printf "%s" instead of saving the string in the ring buffer
 317  * and wasting space and time.
 318  *
 319  * The problem with the above approach is that userspace tools that read
 320  * the binary output of the trace buffers do not have access to the string.
 321  * Instead they just show the address of the string which is not very
 322  * useful to users.
 323  *
 324  * With tracepoint_string(), the string will be registered to the tracing
 325  * system and exported to userspace via the debugfs/tracing/printk_formats
 326  * file that maps the string address to the string text. This way userspace
 327  * tools that read the binary buffers have a way to map the pointers to
 328  * the ASCII strings they represent.
 329  *
 330  * The @str used must be a constant string and persistent as it would not
 331  * make sense to show a string that no longer exists. But it is still fine
 332  * to be used with modules, because when modules are unloaded, if they
 333  * had tracepoints, the ring buffers are cleared too. As long as the string
 334  * does not change during the life of the module, it is fine to use
 335  * tracepoint_string() within a module.
 336  */
 337 #define tracepoint_string(str)                                          \
 338         ({                                                              \
 339                 static const char *___tp_str __tracepoint_string = str; \
 340                 ___tp_str;                                              \
 341         })
 342 #define __tracepoint_string     __attribute__((section("__tracepoint_str")))
 343 #else
 344 /*
 345  * tracepoint_string() is used to save the string address for userspace
 346  * tracing tools. When tracing isn't configured, there's no need to save
 347  * anything.
 348  */
 349 # define tracepoint_string(str) str
 350 # define __tracepoint_string
 351 #endif
 352
 353 /*
 354  * The need for the DECLARE_TRACE_NOARGS() is to handle the prototype
 355  * (void). "void" is a special value in a function prototype and can
 356  * not be combined with other arguments. Since the DECLARE_TRACE()
 357  * macro adds a data element at the beginning of the prototype,
 358  * we need a way to differentiate "(void *data, proto)" from
 359  * "(void *data, void)". The second prototype is invalid.
 360  *
 361  * DECLARE_TRACE_NOARGS() passes "void" as the tracepoint prototype
 362  * and "void *__data" as the callback prototype.
 363  *
 364  * DECLARE_TRACE() passes "proto" as the tracepoint protoype and
 365  * "void *__data, proto" as the callback prototype.
 366  */
 367 #define DECLARE_TRACE_NOARGS(name)                                      \
 368         __DECLARE_TRACE(name, void, ,                                   \
 369                         cpu_online(raw_smp_processor_id()),             \
 370                         void *__data, __data)
 371
 372 #define DECLARE_TRACE(name, proto, args)                                \
 373         __DECLARE_TRACE(name, PARAMS(proto), PARAMS(args),              \
 374                         cpu_online(raw_smp_processor_id()),             \
 375                         PARAMS(void *__data, proto),                    \
 376                         PARAMS(__data, args))
 377
 378 #define DECLARE_TRACE_CONDITION(name, proto, args, cond)                \
 379         __DECLARE_TRACE(name, PARAMS(proto), PARAMS(args),              \
 380                         cpu_online(raw_smp_processor_id()) && (PARAMS(cond)), \
 381                         PARAMS(void *__data, proto),                    \
 382                         PARAMS(__data, args))
 383
 384 #define TRACE_EVENT_FLAGS(event, flag)
 385
 386 #define TRACE_EVENT_PERF_PERM(event, expr...)
 387
 388 #endif /* DECLARE_TRACE */
 389
 390 #ifndef TRACE_EVENT
 391 /*
 392  * For use with the TRACE_EVENT macro:
 393  *
 394  * We define a tracepoint, its arguments, its printk format
 395  * and its 'fast binary record' layout.
 396  *
 397  * Firstly, name your tracepoint via TRACE_EVENT(name : the
 398  * 'subsystem_event' notation is fine.
 399  *
 400  * Think about this whole construct as the
 401  * 'trace_sched_switch() function' from now on.
 402  *
 403  *
 404  *  TRACE_EVENT(sched_switch,
 405  *
 406  *      *
 407  *      * A function has a regular function arguments
 408  *      * prototype, declare it via TP_PROTO():
 409  *      *
 410  *
 411  *      TP_PROTO(struct rq *rq, struct task_struct *prev,
 412  *               struct task_struct *next),
 413  *
 414  *      *
 415  *      * Define the call signature of the 'function'.
 416  *      * (Design sidenote: we use this instead of a
 417  *      *  TP_PROTO1/TP_PROTO2/TP_PROTO3 ugliness.)
 418  *      *
 419  *
 420  *      TP_ARGS(rq, prev, next),
 421  *
 422  *      *
 423  *      * Fast binary tracing: define the trace record via
 424  *      * TP_STRUCT__entry(). You can think about it like a
 425  *      * regular C structure local variable definition.
 426  *      *
 427  *      * This is how the trace record is structured and will
 428  *      * be saved into the ring buffer. These are the fields
 429  *      * that will be exposed to user-space in
 430  *      * /sys/kernel/debug/tracing/events/<*>/format.
 431  *      *
 432  *      * The declared 'local variable' is called '__entry'
 433  *      *
 434  *      * __field(pid_t, prev_prid) is equivalent to a standard declariton:
 435  *      *
 436  *      *       pid_t   prev_pid;
 437  *      *
 438  *      * __array(char, prev_comm, TASK_COMM_LEN) is equivalent to:
 439  *      *
 440  *      *       char    prev_comm[TASK_COMM_LEN];
 441  *      *
 442  *
 443  *      TP_STRUCT__entry(
 444  *              __array(        char,   prev_comm,      TASK_COMM_LEN   )
 445  *              __field(        pid_t,  prev_pid                        )
 446  *              __field(        int,    prev_prio                       )
 447  *              __array(        char,   next_comm,      TASK_COMM_LEN   )
 448  *              __field(        pid_t,  next_pid                        )
 449  *              __field(        int,    next_prio                       )
 450  *      ),
 451  *
 452  *      *
 453  *      * Assign the entry into the trace record, by embedding
 454  *      * a full C statement block into TP_fast_assign(). You
 455  *      * can refer to the trace record as '__entry' -
 456  *      * otherwise you can put arbitrary C code in here.
 457  *      *
 458  *      * Note: this C code will execute every time a trace event
 459  *      * happens, on an active tracepoint.
 460  *      *
 461  *
 462  *      TP_fast_assign(
 463  *              memcpy(__entry->next_comm, next->comm, TASK_COMM_LEN);
 464  *              __entry->prev_pid       = prev->pid;
 465  *              __entry->prev_prio      = prev->prio;
 466  *              memcpy(__entry->prev_comm, prev->comm, TASK_COMM_LEN);
 467  *              __entry->next_pid       = next->pid;
 468  *              __entry->next_prio      = next->prio;
 469  *      ),
 470  *
 471  *      *
 472  *      * Formatted output of a trace record via TP_printk().
 473  *      * This is how the tracepoint will appear under ftrace
 474  *      * plugins that make use of this tracepoint.
 475  *      *
 476  *      * (raw-binary tracing wont actually perform this step.)
 477  *      *
 478  *
 479  *      TP_printk("task %s:%d [%d] ==> %s:%d [%d]",
 480  *              __entry->prev_comm, __entry->prev_pid, __entry->prev_prio,
 481  *              __entry->next_comm, __entry->next_pid, __entry->next_prio),
 482  *
 483  * );
 484  *
 485  * This macro construct is thus used for the regular printk format
 486  * tracing setup, it is used to construct a function pointer based
 487  * tracepoint callback (this is used by programmatic plugins and
 488  * can also by used by generic instrumentation like SystemTap), and
 489  * it is also used to expose a structured trace record in
 490  * /sys/kernel/debug/tracing/events/.
 491  *
 492  * A set of (un)registration functions can be passed to the variant
 493  * TRACE_EVENT_FN to perform any (un)registration work.
 494  */
 495
 496 #define DECLARE_EVENT_CLASS(name, proto, args, tstruct, assign, print)
 497 #define DEFINE_EVENT(template, name, proto, args)               \
 498         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 499 #define DEFINE_EVENT_FN(template, name, proto, args, reg, unreg)\
 500         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 501 #define DEFINE_EVENT_PRINT(template, name, proto, args, print)  \
 502         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 503 #define DEFINE_EVENT_CONDITION(template, name, proto,           \
 504                                args, cond)                      \
 505         DECLARE_TRACE_CONDITION(name, PARAMS(proto),            \
 506                                 PARAMS(args), PARAMS(cond))
 507
 508 #define TRACE_EVENT(name, proto, args, struct, assign, print)   \
 509         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 510 #define TRACE_EVENT_FN(name, proto, args, struct,               \
 511                 assign, print, reg, unreg)                      \
 512         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 513 #define TRACE_EVENT_FN_COND(name, proto, args, cond, struct,            \
 514                 assign, print, reg, unreg)                      \
 515         DECLARE_TRACE_CONDITION(name, PARAMS(proto),    \
 516                         PARAMS(args), PARAMS(cond))
 517 #define TRACE_EVENT_CONDITION(name, proto, args, cond,          \
 518                               struct, assign, print)            \
 519         DECLARE_TRACE_CONDITION(name, PARAMS(proto),            \
 520                                 PARAMS(args), PARAMS(cond))
 521
 522 #define TRACE_EVENT_FLAGS(event, flag)
 523
 524 #define TRACE_EVENT_PERF_PERM(event, expr...)
 525
 526 #endif /* ifdef TRACE_EVENT (see note above) */