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_register_prio_may_exist(struct tracepoint *tp, void *probe, void *data,
  46                                          int prio);
  47 extern int
  48 tracepoint_probe_unregister(struct tracepoint *tp, void *probe, void *data);
  49 static inline int
  50 tracepoint_probe_register_may_exist(struct tracepoint *tp, void *probe,
  51                                     void *data)
  52 {
  53         return tracepoint_probe_register_prio_may_exist(tp, probe, data,
  54                                                         TRACEPOINT_DEFAULT_PRIO);
  55 }
  56 extern void
  57 for_each_kernel_tracepoint(void (*fct)(struct tracepoint *tp, void *priv),
  58                 void *priv);
  59
  60 #ifdef CONFIG_MODULES
  61 struct tp_module {
  62         struct list_head list;
  63         struct module *mod;
  64 };
  65
  66 bool trace_module_has_bad_taint(struct module *mod);
  67 extern int register_tracepoint_module_notifier(struct notifier_block *nb);
  68 extern int unregister_tracepoint_module_notifier(struct notifier_block *nb);
  69 #else
  70 static inline bool trace_module_has_bad_taint(struct module *mod)
  71 {
  72         return false;
  73 }
  74 static inline
  75 int register_tracepoint_module_notifier(struct notifier_block *nb)
  76 {
  77         return 0;
  78 }
  79 static inline
  80 int unregister_tracepoint_module_notifier(struct notifier_block *nb)
  81 {
  82         return 0;
  83 }
  84 #endif /* CONFIG_MODULES */
  85
  86 /*
  87  * tracepoint_synchronize_unregister must be called between the last tracepoint
  88  * probe unregistration and the end of module exit to make sure there is no
  89  * caller executing a probe when it is freed.
  90  */
  91 #ifdef CONFIG_TRACEPOINTS
  92 static inline void tracepoint_synchronize_unregister(void)
  93 {
  94         synchronize_srcu(&tracepoint_srcu);
  95         synchronize_sched();
  96 }
  97 #else
  98 static inline void tracepoint_synchronize_unregister(void)
  99 { }
 100 #endif
 101
 102 #ifdef CONFIG_HAVE_SYSCALL_TRACEPOINTS
 103 extern int syscall_regfunc(void);
 104 extern void syscall_unregfunc(void);
 105 #endif /* CONFIG_HAVE_SYSCALL_TRACEPOINTS */
 106
 107 #define PARAMS(args...) args
 108
 109 #define TRACE_DEFINE_ENUM(x)
 110 #define TRACE_DEFINE_SIZEOF(x)
 111
 112 #ifdef CONFIG_HAVE_ARCH_PREL32_RELOCATIONS
 113 static inline struct tracepoint *tracepoint_ptr_deref(tracepoint_ptr_t *p)
 114 {
 115         return offset_to_ptr(p);
 116 }
 117
 118 #define __TRACEPOINT_ENTRY(name)                                        \
 119         asm("   .section \"__tracepoints_ptrs\", \"a\"          \n"     \
 120             "   .balign 4                                       \n"     \
 121             "   .long   __tracepoint_" #name " - .              \n"     \
 122             "   .previous                                       \n")
 123 #else
 124 static inline struct tracepoint *tracepoint_ptr_deref(tracepoint_ptr_t *p)
 125 {
 126         return *p;
 127 }
 128
 129 #define __TRACEPOINT_ENTRY(name)                                         \
 130         static tracepoint_ptr_t __tracepoint_ptr_##name __used           \
 131         __attribute__((section("__tracepoints_ptrs"))) =                 \
 132                 &__tracepoint_##name
 133 #endif
 134
 135 #endif /* _LINUX_TRACEPOINT_H */
 136
 137 /*
 138  * Note: we keep the TRACE_EVENT and DECLARE_TRACE outside the include
 139  *  file ifdef protection.
 140  *  This is due to the way trace events work. If a file includes two
 141  *  trace event headers under one "CREATE_TRACE_POINTS" the first include
 142  *  will override the TRACE_EVENT and break the second include.
 143  */
 144
 145 #ifndef DECLARE_TRACE
 146
 147 #define TP_PROTO(args...)       args
 148 #define TP_ARGS(args...)        args
 149 #define TP_CONDITION(args...)   args
 150
 151 /*
 152  * Individual subsystem my have a separate configuration to
 153  * enable their tracepoints. By default, this file will create
 154  * the tracepoints if CONFIG_TRACEPOINT is defined. If a subsystem
 155  * wants to be able to disable its tracepoints from being created
 156  * it can define NOTRACE before including the tracepoint headers.
 157  */
 158 #if defined(CONFIG_TRACEPOINTS) && !defined(NOTRACE)
 159 #define TRACEPOINTS_ENABLED
 160 #endif
 161
 162 #ifdef TRACEPOINTS_ENABLED
 163
 164 /*
 165  * it_func[0] is never NULL because there is at least one element in the array
 166  * when the array itself is non NULL.
 167  *
 168  * Note, the proto and args passed in includes "__data" as the first parameter.
 169  * The reason for this is to handle the "void" prototype. If a tracepoint
 170  * has a "void" prototype, then it is invalid to declare a function
 171  * as "(void *, void)". The DECLARE_TRACE_NOARGS() will pass in just
 172  * "void *data", where as the DECLARE_TRACE() will pass in "void *data, proto".
 173  */
 174 #define __DO_TRACE(tp, proto, args, cond, rcuidle)                      \
 175         do {                                                            \
 176                 struct tracepoint_func *it_func_ptr;                    \
 177                 void *it_func;                                          \
 178                 void *__data;                                           \
 179                 int __maybe_unused __idx = 0;                           \
 180                                                                         \
 181                 if (!(cond))                                            \
 182                         return;                                         \
 183                                                                         \
 184                 /* srcu can't be used from NMI */                       \
 185                 WARN_ON_ONCE(rcuidle && in_nmi());                      \
 186                                                                         \
 187                 /* keep srcu and sched-rcu usage consistent */          \
 188                 preempt_disable_notrace();                              \
 189                                                                         \
 190                 /*                                                      \
 191                  * For rcuidle callers, use srcu since sched-rcu        \
 192                  * doesn't work from the idle path.                     \
 193                  */                                                     \
 194                 if (rcuidle) {                                          \
 195                         __idx = srcu_read_lock_notrace(&tracepoint_srcu);\
 196                         rcu_irq_enter_irqson();                         \
 197                 }                                                       \
 198                                                                         \
 199                 it_func_ptr = rcu_dereference_raw((tp)->funcs);         \
 200                                                                         \
 201                 if (it_func_ptr) {                                      \
 202                         do {                                            \
 203                                 it_func = (it_func_ptr)->func;          \
 204                                 __data = (it_func_ptr)->data;           \
 205                                 ((void(*)(proto))(it_func))(args);      \
 206                         } while ((++it_func_ptr)->func);                \
 207                 }                                                       \
 208                                                                         \
 209                 if (rcuidle) {                                          \
 210                         rcu_irq_exit_irqson();                          \
 211                         srcu_read_unlock_notrace(&tracepoint_srcu, __idx);\
 212                 }                                                       \
 213                                                                         \
 214                 preempt_enable_notrace();                               \
 215         } while (0)
 216
 217 #ifndef MODULE
 218 #define __DECLARE_TRACE_RCU(name, proto, args, cond, data_proto, data_args) \
 219         static inline void trace_##name##_rcuidle(proto)                \
 220         {                                                               \
 221                 if (static_key_false(&__tracepoint_##name.key))         \
 222                         __DO_TRACE(&__tracepoint_##name,                \
 223                                 TP_PROTO(data_proto),                   \
 224                                 TP_ARGS(data_args),                     \
 225                                 TP_CONDITION(cond), 1);                 \
 226         }
 227 #else
 228 #define __DECLARE_TRACE_RCU(name, proto, args, cond, data_proto, data_args)
 229 #endif
 230
 231 /*
 232  * Make sure the alignment of the structure in the __tracepoints section will
 233  * not add unwanted padding between the beginning of the section and the
 234  * structure. Force alignment to the same alignment as the section start.
 235  *
 236  * When lockdep is enabled, we make sure to always do the RCU portions of
 237  * the tracepoint code, regardless of whether tracing is on. However,
 238  * don't check if the condition is false, due to interaction with idle
 239  * instrumentation. This lets us find RCU issues triggered with tracepoints
 240  * even when this tracepoint is off. This code has no purpose other than
 241  * poking RCU a bit.
 242  */
 243 #define __DECLARE_TRACE(name, proto, args, cond, data_proto, data_args) \
 244         extern struct tracepoint __tracepoint_##name;                   \
 245         static inline void trace_##name(proto)                          \
 246         {                                                               \
 247                 if (static_key_false(&__tracepoint_##name.key))         \
 248                         __DO_TRACE(&__tracepoint_##name,                \
 249                                 TP_PROTO(data_proto),                   \
 250                                 TP_ARGS(data_args),                     \
 251                                 TP_CONDITION(cond), 0);                 \
 252                 if (IS_ENABLED(CONFIG_LOCKDEP) && (cond)) {             \
 253                         rcu_read_lock_sched_notrace();                  \
 254                         rcu_dereference_sched(__tracepoint_##name.funcs);\
 255                         rcu_read_unlock_sched_notrace();                \
 256                 }                                                       \
 257         }                                                               \
 258         __DECLARE_TRACE_RCU(name, PARAMS(proto), PARAMS(args),          \
 259                 PARAMS(cond), PARAMS(data_proto), PARAMS(data_args))    \
 260         static inline int                                               \
 261         register_trace_##name(void (*probe)(data_proto), void *data)    \
 262         {                                                               \
 263                 return tracepoint_probe_register(&__tracepoint_##name,  \
 264                                                 (void *)probe, data);   \
 265         }                                                               \
 266         static inline int                                               \
 267         register_trace_prio_##name(void (*probe)(data_proto), void *data,\
 268                                    int prio)                            \
 269         {                                                               \
 270                 return tracepoint_probe_register_prio(&__tracepoint_##name, \
 271                                               (void *)probe, data, prio); \
 272         }                                                               \
 273         static inline int                                               \
 274         unregister_trace_##name(void (*probe)(data_proto), void *data)  \
 275         {                                                               \
 276                 return tracepoint_probe_unregister(&__tracepoint_##name,\
 277                                                 (void *)probe, data);   \
 278         }                                                               \
 279         static inline void                                              \
 280         check_trace_callback_type_##name(void (*cb)(data_proto))        \
 281         {                                                               \
 282         }                                                               \
 283         static inline bool                                              \
 284         trace_##name##_enabled(void)                                    \
 285         {                                                               \
 286                 return static_key_false(&__tracepoint_##name.key);      \
 287         }
 288
 289 /*
 290  * We have no guarantee that gcc and the linker won't up-align the tracepoint
 291  * structures, so we create an array of pointers that will be used for iteration
 292  * on the tracepoints.
 293  */
 294 #define DEFINE_TRACE_FN(name, reg, unreg)                                \
 295         static const char __tpstrtab_##name[]                            \
 296         __attribute__((section("__tracepoints_strings"))) = #name;       \
 297         struct tracepoint __tracepoint_##name                            \
 298         __attribute__((section("__tracepoints"), used)) =                \
 299                 { __tpstrtab_##name, STATIC_KEY_INIT_FALSE, reg, unreg, NULL };\
 300         __TRACEPOINT_ENTRY(name);
 301
 302 #define DEFINE_TRACE(name)                                              \
 303         DEFINE_TRACE_FN(name, NULL, NULL);
 304
 305 #define EXPORT_TRACEPOINT_SYMBOL_GPL(name)                              \
 306         EXPORT_SYMBOL_GPL(__tracepoint_##name)
 307 #define EXPORT_TRACEPOINT_SYMBOL(name)                                  \
 308         EXPORT_SYMBOL(__tracepoint_##name)
 309
 310 #else /* !TRACEPOINTS_ENABLED */
 311 #define __DECLARE_TRACE(name, proto, args, cond, data_proto, data_args) \
 312         static inline void trace_##name(proto)                          \
 313         { }                                                             \
 314         static inline void trace_##name##_rcuidle(proto)                \
 315         { }                                                             \
 316         static inline int                                               \
 317         register_trace_##name(void (*probe)(data_proto),                \
 318                               void *data)                               \
 319         {                                                               \
 320                 return -ENOSYS;                                         \
 321         }                                                               \
 322         static inline int                                               \
 323         unregister_trace_##name(void (*probe)(data_proto),              \
 324                                 void *data)                             \
 325         {                                                               \
 326                 return -ENOSYS;                                         \
 327         }                                                               \
 328         static inline void check_trace_callback_type_##name(void (*cb)(data_proto)) \
 329         {                                                               \
 330         }                                                               \
 331         static inline bool                                              \
 332         trace_##name##_enabled(void)                                    \
 333         {                                                               \
 334                 return false;                                           \
 335         }
 336
 337 #define DEFINE_TRACE_FN(name, reg, unreg)
 338 #define DEFINE_TRACE(name)
 339 #define EXPORT_TRACEPOINT_SYMBOL_GPL(name)
 340 #define EXPORT_TRACEPOINT_SYMBOL(name)
 341
 342 #endif /* TRACEPOINTS_ENABLED */
 343
 344 #ifdef CONFIG_TRACING
 345 /**
 346  * tracepoint_string - register constant persistent string to trace system
 347  * @str - a constant persistent string that will be referenced in tracepoints
 348  *
 349  * If constant strings are being used in tracepoints, it is faster and
 350  * more efficient to just save the pointer to the string and reference
 351  * that with a printf "%s" instead of saving the string in the ring buffer
 352  * and wasting space and time.
 353  *
 354  * The problem with the above approach is that userspace tools that read
 355  * the binary output of the trace buffers do not have access to the string.
 356  * Instead they just show the address of the string which is not very
 357  * useful to users.
 358  *
 359  * With tracepoint_string(), the string will be registered to the tracing
 360  * system and exported to userspace via the debugfs/tracing/printk_formats
 361  * file that maps the string address to the string text. This way userspace
 362  * tools that read the binary buffers have a way to map the pointers to
 363  * the ASCII strings they represent.
 364  *
 365  * The @str used must be a constant string and persistent as it would not
 366  * make sense to show a string that no longer exists. But it is still fine
 367  * to be used with modules, because when modules are unloaded, if they
 368  * had tracepoints, the ring buffers are cleared too. As long as the string
 369  * does not change during the life of the module, it is fine to use
 370  * tracepoint_string() within a module.
 371  */
 372 #define tracepoint_string(str)                                          \
 373         ({                                                              \
 374                 static const char *___tp_str __tracepoint_string = str; \
 375                 ___tp_str;                                              \
 376         })
 377 #define __tracepoint_string     __attribute__((section("__tracepoint_str"), used))
 378 #else
 379 /*
 380  * tracepoint_string() is used to save the string address for userspace
 381  * tracing tools. When tracing isn't configured, there's no need to save
 382  * anything.
 383  */
 384 # define tracepoint_string(str) str
 385 # define __tracepoint_string
 386 #endif
 387
 388 /*
 389  * The need for the DECLARE_TRACE_NOARGS() is to handle the prototype
 390  * (void). "void" is a special value in a function prototype and can
 391  * not be combined with other arguments. Since the DECLARE_TRACE()
 392  * macro adds a data element at the beginning of the prototype,
 393  * we need a way to differentiate "(void *data, proto)" from
 394  * "(void *data, void)". The second prototype is invalid.
 395  *
 396  * DECLARE_TRACE_NOARGS() passes "void" as the tracepoint prototype
 397  * and "void *__data" as the callback prototype.
 398  *
 399  * DECLARE_TRACE() passes "proto" as the tracepoint protoype and
 400  * "void *__data, proto" as the callback prototype.
 401  */
 402 #define DECLARE_TRACE_NOARGS(name)                                      \
 403         __DECLARE_TRACE(name, void, ,                                   \
 404                         cpu_online(raw_smp_processor_id()),             \
 405                         void *__data, __data)
 406
 407 #define DECLARE_TRACE(name, proto, args)                                \
 408         __DECLARE_TRACE(name, PARAMS(proto), PARAMS(args),              \
 409                         cpu_online(raw_smp_processor_id()),             \
 410                         PARAMS(void *__data, proto),                    \
 411                         PARAMS(__data, args))
 412
 413 #define DECLARE_TRACE_CONDITION(name, proto, args, cond)                \
 414         __DECLARE_TRACE(name, PARAMS(proto), PARAMS(args),              \
 415                         cpu_online(raw_smp_processor_id()) && (PARAMS(cond)), \
 416                         PARAMS(void *__data, proto),                    \
 417                         PARAMS(__data, args))
 418
 419 #define TRACE_EVENT_FLAGS(event, flag)
 420
 421 #define TRACE_EVENT_PERF_PERM(event, expr...)
 422
 423 #endif /* DECLARE_TRACE */
 424
 425 #ifndef TRACE_EVENT
 426 /*
 427  * For use with the TRACE_EVENT macro:
 428  *
 429  * We define a tracepoint, its arguments, its printk format
 430  * and its 'fast binary record' layout.
 431  *
 432  * Firstly, name your tracepoint via TRACE_EVENT(name : the
 433  * 'subsystem_event' notation is fine.
 434  *
 435  * Think about this whole construct as the
 436  * 'trace_sched_switch() function' from now on.
 437  *
 438  *
 439  *  TRACE_EVENT(sched_switch,
 440  *
 441  *      *
 442  *      * A function has a regular function arguments
 443  *      * prototype, declare it via TP_PROTO():
 444  *      *
 445  *
 446  *      TP_PROTO(struct rq *rq, struct task_struct *prev,
 447  *               struct task_struct *next),
 448  *
 449  *      *
 450  *      * Define the call signature of the 'function'.
 451  *      * (Design sidenote: we use this instead of a
 452  *      *  TP_PROTO1/TP_PROTO2/TP_PROTO3 ugliness.)
 453  *      *
 454  *
 455  *      TP_ARGS(rq, prev, next),
 456  *
 457  *      *
 458  *      * Fast binary tracing: define the trace record via
 459  *      * TP_STRUCT__entry(). You can think about it like a
 460  *      * regular C structure local variable definition.
 461  *      *
 462  *      * This is how the trace record is structured and will
 463  *      * be saved into the ring buffer. These are the fields
 464  *      * that will be exposed to user-space in
 465  *      * /sys/kernel/debug/tracing/events/<*>/format.
 466  *      *
 467  *      * The declared 'local variable' is called '__entry'
 468  *      *
 469  *      * __field(pid_t, prev_prid) is equivalent to a standard declariton:
 470  *      *
 471  *      *       pid_t   prev_pid;
 472  *      *
 473  *      * __array(char, prev_comm, TASK_COMM_LEN) is equivalent to:
 474  *      *
 475  *      *       char    prev_comm[TASK_COMM_LEN];
 476  *      *
 477  *
 478  *      TP_STRUCT__entry(
 479  *              __array(        char,   prev_comm,      TASK_COMM_LEN   )
 480  *              __field(        pid_t,  prev_pid                        )
 481  *              __field(        int,    prev_prio                       )
 482  *              __array(        char,   next_comm,      TASK_COMM_LEN   )
 483  *              __field(        pid_t,  next_pid                        )
 484  *              __field(        int,    next_prio                       )
 485  *      ),
 486  *
 487  *      *
 488  *      * Assign the entry into the trace record, by embedding
 489  *      * a full C statement block into TP_fast_assign(). You
 490  *      * can refer to the trace record as '__entry' -
 491  *      * otherwise you can put arbitrary C code in here.
 492  *      *
 493  *      * Note: this C code will execute every time a trace event
 494  *      * happens, on an active tracepoint.
 495  *      *
 496  *
 497  *      TP_fast_assign(
 498  *              memcpy(__entry->next_comm, next->comm, TASK_COMM_LEN);
 499  *              __entry->prev_pid       = prev->pid;
 500  *              __entry->prev_prio      = prev->prio;
 501  *              memcpy(__entry->prev_comm, prev->comm, TASK_COMM_LEN);
 502  *              __entry->next_pid       = next->pid;
 503  *              __entry->next_prio      = next->prio;
 504  *      ),
 505  *
 506  *      *
 507  *      * Formatted output of a trace record via TP_printk().
 508  *      * This is how the tracepoint will appear under ftrace
 509  *      * plugins that make use of this tracepoint.
 510  *      *
 511  *      * (raw-binary tracing wont actually perform this step.)
 512  *      *
 513  *
 514  *      TP_printk("task %s:%d [%d] ==> %s:%d [%d]",
 515  *              __entry->prev_comm, __entry->prev_pid, __entry->prev_prio,
 516  *              __entry->next_comm, __entry->next_pid, __entry->next_prio),
 517  *
 518  * );
 519  *
 520  * This macro construct is thus used for the regular printk format
 521  * tracing setup, it is used to construct a function pointer based
 522  * tracepoint callback (this is used by programmatic plugins and
 523  * can also by used by generic instrumentation like SystemTap), and
 524  * it is also used to expose a structured trace record in
 525  * /sys/kernel/debug/tracing/events/.
 526  *
 527  * A set of (un)registration functions can be passed to the variant
 528  * TRACE_EVENT_FN to perform any (un)registration work.
 529  */
 530
 531 #define DECLARE_EVENT_CLASS(name, proto, args, tstruct, assign, print)
 532 #define DEFINE_EVENT(template, name, proto, args)               \
 533         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 534 #define DEFINE_EVENT_FN(template, name, proto, args, reg, unreg)\
 535         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 536 #define DEFINE_EVENT_PRINT(template, name, proto, args, print)  \
 537         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 538 #define DEFINE_EVENT_CONDITION(template, name, proto,           \
 539                                args, cond)                      \
 540         DECLARE_TRACE_CONDITION(name, PARAMS(proto),            \
 541                                 PARAMS(args), PARAMS(cond))
 542
 543 #define TRACE_EVENT(name, proto, args, struct, assign, print)   \
 544         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 545 #define TRACE_EVENT_FN(name, proto, args, struct,               \
 546                 assign, print, reg, unreg)                      \
 547         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 548 #define TRACE_EVENT_FN_COND(name, proto, args, cond, struct,            \
 549                 assign, print, reg, unreg)                      \
 550         DECLARE_TRACE_CONDITION(name, PARAMS(proto),    \
 551                         PARAMS(args), PARAMS(cond))
 552 #define TRACE_EVENT_CONDITION(name, proto, args, cond,          \
 553                               struct, assign, print)            \
 554         DECLARE_TRACE_CONDITION(name, PARAMS(proto),            \
 555                                 PARAMS(args), PARAMS(cond))
 556
 557 #define TRACE_EVENT_FLAGS(event, flag)
 558
 559 #define TRACE_EVENT_PERF_PERM(event, expr...)
 560
 561 #endif /* ifdef TRACE_EVENT (see note above) */