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 <[email protected]>
  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                         rcu_irq_enter_irqson();                         \
 164                 }                                                       \
 165                                                                         \
 166                 it_func_ptr = rcu_dereference_raw((tp)->funcs);         \
 167                                                                         \
 168                 if (it_func_ptr) {                                      \
 169                         do {                                            \
 170                                 it_func = (it_func_ptr)->func;          \
 171                                 __data = (it_func_ptr)->data;           \
 172                                 ((void(*)(proto))(it_func))(args);      \
 173                         } while ((++it_func_ptr)->func);                \
 174                 }                                                       \
 175                                                                         \
 176                 if (rcuidle) {                                          \
 177                         rcu_irq_exit_irqson();                          \
 178                         srcu_read_unlock_notrace(&tracepoint_srcu, idx);\
 179                 }                                                       \
 180                                                                         \
 181                 preempt_enable_notrace();                               \
 182         } while (0)
 183
 184 #ifndef MODULE
 185 #define __DECLARE_TRACE_RCU(name, proto, args, cond, data_proto, data_args) \
 186         static inline void trace_##name##_rcuidle(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), 1);                 \
 193         }
 194 #else
 195 #define __DECLARE_TRACE_RCU(name, proto, args, cond, data_proto, data_args)
 196 #endif
 197
 198 /*
 199  * Make sure the alignment of the structure in the __tracepoints section will
 200  * not add unwanted padding between the beginning of the section and the
 201  * structure. Force alignment to the same alignment as the section start.
 202  *
 203  * When lockdep is enabled, we make sure to always do the RCU portions of
 204  * the tracepoint code, regardless of whether tracing is on. However,
 205  * don't check if the condition is false, due to interaction with idle
 206  * instrumentation. This lets us find RCU issues triggered with tracepoints
 207  * even when this tracepoint is off. This code has no purpose other than
 208  * poking RCU a bit.
 209  */
 210 #define __DECLARE_TRACE(name, proto, args, cond, data_proto, data_args) \
 211         extern struct tracepoint __tracepoint_##name;                   \
 212         static inline void trace_##name(proto)                          \
 213         {                                                               \
 214                 if (static_key_false(&__tracepoint_##name.key))         \
 215                         __DO_TRACE(&__tracepoint_##name,                \
 216                                 TP_PROTO(data_proto),                   \
 217                                 TP_ARGS(data_args),                     \
 218                                 TP_CONDITION(cond), 0);                 \
 219                 if (IS_ENABLED(CONFIG_LOCKDEP) && (cond)) {             \
 220                         rcu_read_lock_sched_notrace();                  \
 221                         rcu_dereference_sched(__tracepoint_##name.funcs);\
 222                         rcu_read_unlock_sched_notrace();                \
 223                 }                                                       \
 224         }                                                               \
 225         __DECLARE_TRACE_RCU(name, PARAMS(proto), PARAMS(args),          \
 226                 PARAMS(cond), PARAMS(data_proto), PARAMS(data_args))    \
 227         static inline int                                               \
 228         register_trace_##name(void (*probe)(data_proto), void *data)    \
 229         {                                                               \
 230                 return tracepoint_probe_register(&__tracepoint_##name,  \
 231                                                 (void *)probe, data);   \
 232         }                                                               \
 233         static inline int                                               \
 234         register_trace_prio_##name(void (*probe)(data_proto), void *data,\
 235                                    int prio)                            \
 236         {                                                               \
 237                 return tracepoint_probe_register_prio(&__tracepoint_##name, \
 238                                               (void *)probe, data, prio); \
 239         }                                                               \
 240         static inline int                                               \
 241         unregister_trace_##name(void (*probe)(data_proto), void *data)  \
 242         {                                                               \
 243                 return tracepoint_probe_unregister(&__tracepoint_##name,\
 244                                                 (void *)probe, data);   \
 245         }                                                               \
 246         static inline void                                              \
 247         check_trace_callback_type_##name(void (*cb)(data_proto))        \
 248         {                                                               \
 249         }                                                               \
 250         static inline bool                                              \
 251         trace_##name##_enabled(void)                                    \
 252         {                                                               \
 253                 return static_key_false(&__tracepoint_##name.key);      \
 254         }
 255
 256 #ifdef CONFIG_HAVE_ARCH_PREL32_RELOCATIONS
 257 #define __TRACEPOINT_ENTRY(name)                                        \
 258         asm("   .section \"__tracepoints_ptrs\", \"a\"          \n"     \
 259             "   .balign 4                                       \n"     \
 260             "   .long   __tracepoint_" #name " - .              \n"     \
 261             "   .previous                                       \n")
 262 #else
 263 #define __TRACEPOINT_ENTRY(name)                                         \
 264         static struct tracepoint * const __tracepoint_ptr_##name __used  \
 265         __attribute__((section("__tracepoints_ptrs"))) =                 \
 266                 &__tracepoint_##name
 267 #endif
 268
 269 /*
 270  * We have no guarantee that gcc and the linker won't up-align the tracepoint
 271  * structures, so we create an array of pointers that will be used for iteration
 272  * on the tracepoints.
 273  */
 274 #define DEFINE_TRACE_FN(name, reg, unreg)                                \
 275         static const char __tpstrtab_##name[]                            \
 276         __attribute__((section("__tracepoints_strings"))) = #name;       \
 277         struct tracepoint __tracepoint_##name                            \
 278         __attribute__((section("__tracepoints"), used)) =                \
 279                 { __tpstrtab_##name, STATIC_KEY_INIT_FALSE, reg, unreg, NULL };\
 280         __TRACEPOINT_ENTRY(name);
 281
 282 #define DEFINE_TRACE(name)                                              \
 283         DEFINE_TRACE_FN(name, NULL, NULL);
 284
 285 #define EXPORT_TRACEPOINT_SYMBOL_GPL(name)                              \
 286         EXPORT_SYMBOL_GPL(__tracepoint_##name)
 287 #define EXPORT_TRACEPOINT_SYMBOL(name)                                  \
 288         EXPORT_SYMBOL(__tracepoint_##name)
 289
 290 #else /* !TRACEPOINTS_ENABLED */
 291 #define __DECLARE_TRACE(name, proto, args, cond, data_proto, data_args) \
 292         static inline void trace_##name(proto)                          \
 293         { }                                                             \
 294         static inline void trace_##name##_rcuidle(proto)                \
 295         { }                                                             \
 296         static inline int                                               \
 297         register_trace_##name(void (*probe)(data_proto),                \
 298                               void *data)                               \
 299         {                                                               \
 300                 return -ENOSYS;                                         \
 301         }                                                               \
 302         static inline int                                               \
 303         unregister_trace_##name(void (*probe)(data_proto),              \
 304                                 void *data)                             \
 305         {                                                               \
 306                 return -ENOSYS;                                         \
 307         }                                                               \
 308         static inline void check_trace_callback_type_##name(void (*cb)(data_proto)) \
 309         {                                                               \
 310         }                                                               \
 311         static inline bool                                              \
 312         trace_##name##_enabled(void)                                    \
 313         {                                                               \
 314                 return false;                                           \
 315         }
 316
 317 #define DEFINE_TRACE_FN(name, reg, unreg)
 318 #define DEFINE_TRACE(name)
 319 #define EXPORT_TRACEPOINT_SYMBOL_GPL(name)
 320 #define EXPORT_TRACEPOINT_SYMBOL(name)
 321
 322 #endif /* TRACEPOINTS_ENABLED */
 323
 324 #ifdef CONFIG_TRACING
 325 /**
 326  * tracepoint_string - register constant persistent string to trace system
 327  * @str - a constant persistent string that will be referenced in tracepoints
 328  *
 329  * If constant strings are being used in tracepoints, it is faster and
 330  * more efficient to just save the pointer to the string and reference
 331  * that with a printf "%s" instead of saving the string in the ring buffer
 332  * and wasting space and time.
 333  *
 334  * The problem with the above approach is that userspace tools that read
 335  * the binary output of the trace buffers do not have access to the string.
 336  * Instead they just show the address of the string which is not very
 337  * useful to users.
 338  *
 339  * With tracepoint_string(), the string will be registered to the tracing
 340  * system and exported to userspace via the debugfs/tracing/printk_formats
 341  * file that maps the string address to the string text. This way userspace
 342  * tools that read the binary buffers have a way to map the pointers to
 343  * the ASCII strings they represent.
 344  *
 345  * The @str used must be a constant string and persistent as it would not
 346  * make sense to show a string that no longer exists. But it is still fine
 347  * to be used with modules, because when modules are unloaded, if they
 348  * had tracepoints, the ring buffers are cleared too. As long as the string
 349  * does not change during the life of the module, it is fine to use
 350  * tracepoint_string() within a module.
 351  */
 352 #define tracepoint_string(str)                                          \
 353         ({                                                              \
 354                 static const char *___tp_str __tracepoint_string = str; \
 355                 ___tp_str;                                              \
 356         })
 357 #define __tracepoint_string     __attribute__((section("__tracepoint_str")))
 358 #else
 359 /*
 360  * tracepoint_string() is used to save the string address for userspace
 361  * tracing tools. When tracing isn't configured, there's no need to save
 362  * anything.
 363  */
 364 # define tracepoint_string(str) str
 365 # define __tracepoint_string
 366 #endif
 367
 368 /*
 369  * The need for the DECLARE_TRACE_NOARGS() is to handle the prototype
 370  * (void). "void" is a special value in a function prototype and can
 371  * not be combined with other arguments. Since the DECLARE_TRACE()
 372  * macro adds a data element at the beginning of the prototype,
 373  * we need a way to differentiate "(void *data, proto)" from
 374  * "(void *data, void)". The second prototype is invalid.
 375  *
 376  * DECLARE_TRACE_NOARGS() passes "void" as the tracepoint prototype
 377  * and "void *__data" as the callback prototype.
 378  *
 379  * DECLARE_TRACE() passes "proto" as the tracepoint protoype and
 380  * "void *__data, proto" as the callback prototype.
 381  */
 382 #define DECLARE_TRACE_NOARGS(name)                                      \
 383         __DECLARE_TRACE(name, void, ,                                   \
 384                         cpu_online(raw_smp_processor_id()),             \
 385                         void *__data, __data)
 386
 387 #define DECLARE_TRACE(name, proto, args)                                \
 388         __DECLARE_TRACE(name, PARAMS(proto), PARAMS(args),              \
 389                         cpu_online(raw_smp_processor_id()),             \
 390                         PARAMS(void *__data, proto),                    \
 391                         PARAMS(__data, args))
 392
 393 #define DECLARE_TRACE_CONDITION(name, proto, args, cond)                \
 394         __DECLARE_TRACE(name, PARAMS(proto), PARAMS(args),              \
 395                         cpu_online(raw_smp_processor_id()) && (PARAMS(cond)), \
 396                         PARAMS(void *__data, proto),                    \
 397                         PARAMS(__data, args))
 398
 399 #define TRACE_EVENT_FLAGS(event, flag)
 400
 401 #define TRACE_EVENT_PERF_PERM(event, expr...)
 402
 403 #endif /* DECLARE_TRACE */
 404
 405 #ifndef TRACE_EVENT
 406 /*
 407  * For use with the TRACE_EVENT macro:
 408  *
 409  * We define a tracepoint, its arguments, its printk format
 410  * and its 'fast binary record' layout.
 411  *
 412  * Firstly, name your tracepoint via TRACE_EVENT(name : the
 413  * 'subsystem_event' notation is fine.
 414  *
 415  * Think about this whole construct as the
 416  * 'trace_sched_switch() function' from now on.
 417  *
 418  *
 419  *  TRACE_EVENT(sched_switch,
 420  *
 421  *      *
 422  *      * A function has a regular function arguments
 423  *      * prototype, declare it via TP_PROTO():
 424  *      *
 425  *
 426  *      TP_PROTO(struct rq *rq, struct task_struct *prev,
 427  *               struct task_struct *next),
 428  *
 429  *      *
 430  *      * Define the call signature of the 'function'.
 431  *      * (Design sidenote: we use this instead of a
 432  *      *  TP_PROTO1/TP_PROTO2/TP_PROTO3 ugliness.)
 433  *      *
 434  *
 435  *      TP_ARGS(rq, prev, next),
 436  *
 437  *      *
 438  *      * Fast binary tracing: define the trace record via
 439  *      * TP_STRUCT__entry(). You can think about it like a
 440  *      * regular C structure local variable definition.
 441  *      *
 442  *      * This is how the trace record is structured and will
 443  *      * be saved into the ring buffer. These are the fields
 444  *      * that will be exposed to user-space in
 445  *      * /sys/kernel/debug/tracing/events/<*>/format.
 446  *      *
 447  *      * The declared 'local variable' is called '__entry'
 448  *      *
 449  *      * __field(pid_t, prev_prid) is equivalent to a standard declariton:
 450  *      *
 451  *      *       pid_t   prev_pid;
 452  *      *
 453  *      * __array(char, prev_comm, TASK_COMM_LEN) is equivalent to:
 454  *      *
 455  *      *       char    prev_comm[TASK_COMM_LEN];
 456  *      *
 457  *
 458  *      TP_STRUCT__entry(
 459  *              __array(        char,   prev_comm,      TASK_COMM_LEN   )
 460  *              __field(        pid_t,  prev_pid                        )
 461  *              __field(        int,    prev_prio                       )
 462  *              __array(        char,   next_comm,      TASK_COMM_LEN   )
 463  *              __field(        pid_t,  next_pid                        )
 464  *              __field(        int,    next_prio                       )
 465  *      ),
 466  *
 467  *      *
 468  *      * Assign the entry into the trace record, by embedding
 469  *      * a full C statement block into TP_fast_assign(). You
 470  *      * can refer to the trace record as '__entry' -
 471  *      * otherwise you can put arbitrary C code in here.
 472  *      *
 473  *      * Note: this C code will execute every time a trace event
 474  *      * happens, on an active tracepoint.
 475  *      *
 476  *
 477  *      TP_fast_assign(
 478  *              memcpy(__entry->next_comm, next->comm, TASK_COMM_LEN);
 479  *              __entry->prev_pid       = prev->pid;
 480  *              __entry->prev_prio      = prev->prio;
 481  *              memcpy(__entry->prev_comm, prev->comm, TASK_COMM_LEN);
 482  *              __entry->next_pid       = next->pid;
 483  *              __entry->next_prio      = next->prio;
 484  *      ),
 485  *
 486  *      *
 487  *      * Formatted output of a trace record via TP_printk().
 488  *      * This is how the tracepoint will appear under ftrace
 489  *      * plugins that make use of this tracepoint.
 490  *      *
 491  *      * (raw-binary tracing wont actually perform this step.)
 492  *      *
 493  *
 494  *      TP_printk("task %s:%d [%d] ==> %s:%d [%d]",
 495  *              __entry->prev_comm, __entry->prev_pid, __entry->prev_prio,
 496  *              __entry->next_comm, __entry->next_pid, __entry->next_prio),
 497  *
 498  * );
 499  *
 500  * This macro construct is thus used for the regular printk format
 501  * tracing setup, it is used to construct a function pointer based
 502  * tracepoint callback (this is used by programmatic plugins and
 503  * can also by used by generic instrumentation like SystemTap), and
 504  * it is also used to expose a structured trace record in
 505  * /sys/kernel/debug/tracing/events/.
 506  *
 507  * A set of (un)registration functions can be passed to the variant
 508  * TRACE_EVENT_FN to perform any (un)registration work.
 509  */
 510
 511 #define DECLARE_EVENT_CLASS(name, proto, args, tstruct, assign, print)
 512 #define DEFINE_EVENT(template, name, proto, args)               \
 513         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 514 #define DEFINE_EVENT_FN(template, name, proto, args, reg, unreg)\
 515         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 516 #define DEFINE_EVENT_PRINT(template, name, proto, args, print)  \
 517         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 518 #define DEFINE_EVENT_CONDITION(template, name, proto,           \
 519                                args, cond)                      \
 520         DECLARE_TRACE_CONDITION(name, PARAMS(proto),            \
 521                                 PARAMS(args), PARAMS(cond))
 522
 523 #define TRACE_EVENT(name, proto, args, struct, assign, print)   \
 524         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 525 #define TRACE_EVENT_FN(name, proto, args, struct,               \
 526                 assign, print, reg, unreg)                      \
 527         DECLARE_TRACE(name, PARAMS(proto), PARAMS(args))
 528 #define TRACE_EVENT_FN_COND(name, proto, args, cond, struct,            \
 529                 assign, print, reg, unreg)                      \
 530         DECLARE_TRACE_CONDITION(name, PARAMS(proto),    \
 531                         PARAMS(args), PARAMS(cond))
 532 #define TRACE_EVENT_CONDITION(name, proto, args, cond,          \
 533                               struct, assign, print)            \
 534         DECLARE_TRACE_CONDITION(name, PARAMS(proto),            \
 535                                 PARAMS(args), PARAMS(cond))
 536
 537 #define TRACE_EVENT_FLAGS(event, flag)
 538
 539 #define TRACE_EVENT_PERF_PERM(event, expr...)
 540
 541 #endif /* ifdef TRACE_EVENT (see note above) */