kernel/tracepoint.c

   1 /*
   2  * Copyright (C) 2008 Mathieu Desnoyers
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License as published by
   6  * the Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write to the Free Software
  16  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  17  */
  18 #include <linux/module.h>
  19 #include <linux/mutex.h>
  20 #include <linux/types.h>
  21 #include <linux/jhash.h>
  22 #include <linux/list.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/tracepoint.h>
  25 #include <linux/err.h>
  26 #include <linux/slab.h>
  27 #include <linux/sched.h>
  28 #include <linux/jump_label.h>
  29
  30 extern struct tracepoint __start___tracepoints[];
  31 extern struct tracepoint __stop___tracepoints[];
  32
  33 /* Set to 1 to enable tracepoint debug output */
  34 static const int tracepoint_debug;
  35
  36 /*
  37  * tracepoints_mutex nests inside module_mutex. Tracepoints mutex protects the
  38  * builtin and module tracepoints and the hash table.
  39  */
  40 static DEFINE_MUTEX(tracepoints_mutex);
  41
  42 /*
  43  * Tracepoint hash table, containing the active tracepoints.
  44  * Protected by tracepoints_mutex.
  45  */
  46 #define TRACEPOINT_HASH_BITS 6
  47 #define TRACEPOINT_TABLE_SIZE (1 << TRACEPOINT_HASH_BITS)
  48 static struct hlist_head tracepoint_table[TRACEPOINT_TABLE_SIZE];
  49
  50 /*
  51  * Note about RCU :
  52  * It is used to delay the free of multiple probes array until a quiescent
  53  * state is reached.
  54  * Tracepoint entries modifications are protected by the tracepoints_mutex.
  55  */
  56 struct tracepoint_entry {
  57         struct hlist_node hlist;
  58         struct tracepoint_func *funcs;
  59         int refcount;   /* Number of times armed. 0 if disarmed. */
  60         char name[0];
  61 };
  62
  63 struct tp_probes {
  64         union {
  65                 struct rcu_head rcu;
  66                 struct list_head list;
  67         } u;
  68         struct tracepoint_func probes[0];
  69 };
  70
  71 static inline void *allocate_probes(int count)
  72 {
  73         struct tp_probes *p  = kmalloc(count * sizeof(struct tracepoint_func)
  74                         + sizeof(struct tp_probes), GFP_KERNEL);
  75         return p == NULL ? NULL : p->probes;
  76 }
  77
  78 static void rcu_free_old_probes(struct rcu_head *head)
  79 {
  80         kfree(container_of(head, struct tp_probes, u.rcu));
  81 }
  82
  83 static inline void release_probes(struct tracepoint_func *old)
  84 {
  85         if (old) {
  86                 struct tp_probes *tp_probes = container_of(old,
  87                         struct tp_probes, probes[0]);
  88                 call_rcu_sched(&tp_probes->u.rcu, rcu_free_old_probes);
  89         }
  90 }
  91
  92 static void debug_print_probes(struct tracepoint_entry *entry)
  93 {
  94         int i;
  95
  96         if (!tracepoint_debug || !entry->funcs)
  97                 return;
  98
  99         for (i = 0; entry->funcs[i].func; i++)
 100                 printk(KERN_DEBUG "Probe %d : %p\n", i, entry->funcs[i].func);
 101 }
 102
 103 static struct tracepoint_func *
 104 tracepoint_entry_add_probe(struct tracepoint_entry *entry,
 105                            void *probe, void *data)
 106 {
 107         int nr_probes = 0;
 108         struct tracepoint_func *old, *new;
 109
 110         WARN_ON(!probe);
 111
 112         debug_print_probes(entry);
 113         old = entry->funcs;
 114         if (old) {
 115                 /* (N -> N+1), (N != 0, 1) probes */
 116                 for (nr_probes = 0; old[nr_probes].func; nr_probes++)
 117                         if (old[nr_probes].func == probe &&
 118                             old[nr_probes].data == data)
 119                                 return ERR_PTR(-EEXIST);
 120         }
 121         /* + 2 : one for new probe, one for NULL func */
 122         new = allocate_probes(nr_probes + 2);
 123         if (new == NULL)
 124                 return ERR_PTR(-ENOMEM);
 125         if (old)
 126                 memcpy(new, old, nr_probes * sizeof(struct tracepoint_func));
 127         new[nr_probes].func = probe;
 128         new[nr_probes].data = data;
 129         new[nr_probes + 1].func = NULL;
 130         entry->refcount = nr_probes + 1;
 131         entry->funcs = new;
 132         debug_print_probes(entry);
 133         return old;
 134 }
 135
 136 static void *
 137 tracepoint_entry_remove_probe(struct tracepoint_entry *entry,
 138                               void *probe, void *data)
 139 {
 140         int nr_probes = 0, nr_del = 0, i;
 141         struct tracepoint_func *old, *new;
 142
 143         old = entry->funcs;
 144
 145         if (!old)
 146                 return ERR_PTR(-ENOENT);
 147
 148         debug_print_probes(entry);
 149         /* (N -> M), (N > 1, M >= 0) probes */
 150         for (nr_probes = 0; old[nr_probes].func; nr_probes++) {
 151                 if (!probe ||
 152                     (old[nr_probes].func == probe &&
 153                      old[nr_probes].data == data))
 154                         nr_del++;
 155         }
 156
 157         if (nr_probes - nr_del == 0) {
 158                 /* N -> 0, (N > 1) */
 159                 entry->funcs = NULL;
 160                 entry->refcount = 0;
 161                 debug_print_probes(entry);
 162                 return old;
 163         } else {
 164                 int j = 0;
 165                 /* N -> M, (N > 1, M > 0) */
 166                 /* + 1 for NULL */
 167                 new = allocate_probes(nr_probes - nr_del + 1);
 168                 if (new == NULL)
 169                         return ERR_PTR(-ENOMEM);
 170                 for (i = 0; old[i].func; i++)
 171                         if (probe &&
 172                             (old[i].func != probe || old[i].data != data))
 173                                 new[j++] = old[i];
 174                 new[nr_probes - nr_del].func = NULL;
 175                 entry->refcount = nr_probes - nr_del;
 176                 entry->funcs = new;
 177         }
 178         debug_print_probes(entry);
 179         return old;
 180 }
 181
 182 /*
 183  * Get tracepoint if the tracepoint is present in the tracepoint hash table.
 184  * Must be called with tracepoints_mutex held.
 185  * Returns NULL if not present.
 186  */
 187 static struct tracepoint_entry *get_tracepoint(const char *name)
 188 {
 189         struct hlist_head *head;
 190         struct hlist_node *node;
 191         struct tracepoint_entry *e;
 192         u32 hash = jhash(name, strlen(name), 0);
 193
 194         head = &tracepoint_table[hash & (TRACEPOINT_TABLE_SIZE - 1)];
 195         hlist_for_each_entry(e, node, head, hlist) {
 196                 if (!strcmp(name, e->name))
 197                         return e;
 198         }
 199         return NULL;
 200 }
 201
 202 /*
 203  * Add the tracepoint to the tracepoint hash table. Must be called with
 204  * tracepoints_mutex held.
 205  */
 206 static struct tracepoint_entry *add_tracepoint(const char *name)
 207 {
 208         struct hlist_head *head;
 209         struct hlist_node *node;
 210         struct tracepoint_entry *e;
 211         size_t name_len = strlen(name) + 1;
 212         u32 hash = jhash(name, name_len-1, 0);
 213
 214         head = &tracepoint_table[hash & (TRACEPOINT_TABLE_SIZE - 1)];
 215         hlist_for_each_entry(e, node, head, hlist) {
 216                 if (!strcmp(name, e->name)) {
 217                         printk(KERN_NOTICE
 218                                 "tracepoint %s busy\n", name);
 219                         return ERR_PTR(-EEXIST);        /* Already there */
 220                 }
 221         }
 222         /*
 223          * Using kmalloc here to allocate a variable length element. Could
 224          * cause some memory fragmentation if overused.
 225          */
 226         e = kmalloc(sizeof(struct tracepoint_entry) + name_len, GFP_KERNEL);
 227         if (!e)
 228                 return ERR_PTR(-ENOMEM);
 229         memcpy(&e->name[0], name, name_len);
 230         e->funcs = NULL;
 231         e->refcount = 0;
 232         hlist_add_head(&e->hlist, head);
 233         return e;
 234 }
 235
 236 /*
 237  * Remove the tracepoint from the tracepoint hash table. Must be called with
 238  * mutex_lock held.
 239  */
 240 static inline void remove_tracepoint(struct tracepoint_entry *e)
 241 {
 242         hlist_del(&e->hlist);
 243         kfree(e);
 244 }
 245
 246 /*
 247  * Sets the probe callback corresponding to one tracepoint.
 248  */
 249 static void set_tracepoint(struct tracepoint_entry **entry,
 250         struct tracepoint *elem, int active)
 251 {
 252         WARN_ON(strcmp((*entry)->name, elem->name) != 0);
 253
 254         if (elem->regfunc && !elem->state && active)
 255                 elem->regfunc();
 256         else if (elem->unregfunc && elem->state && !active)
 257                 elem->unregfunc();
 258
 259         /*
 260          * rcu_assign_pointer has a smp_wmb() which makes sure that the new
 261          * probe callbacks array is consistent before setting a pointer to it.
 262          * This array is referenced by __DO_TRACE from
 263          * include/linux/tracepoints.h. A matching smp_read_barrier_depends()
 264          * is used.
 265          */
 266         rcu_assign_pointer(elem->funcs, (*entry)->funcs);
 267         if (!elem->state && active) {
 268                 jump_label_enable(&elem->state);
 269                 elem->state = active;
 270         } else if (elem->state && !active) {
 271                 jump_label_disable(&elem->state);
 272                 elem->state = active;
 273         }
 274 }
 275
 276 /*
 277  * Disable a tracepoint and its probe callback.
 278  * Note: only waiting an RCU period after setting elem->call to the empty
 279  * function insures that the original callback is not used anymore. This insured
 280  * by preempt_disable around the call site.
 281  */
 282 static void disable_tracepoint(struct tracepoint *elem)
 283 {
 284         if (elem->unregfunc && elem->state)
 285                 elem->unregfunc();
 286
 287         if (elem->state) {
 288                 jump_label_disable(&elem->state);
 289                 elem->state = 0;
 290         }
 291         rcu_assign_pointer(elem->funcs, NULL);
 292 }
 293
 294 /**
 295  * tracepoint_update_probe_range - Update a probe range
 296  * @begin: beginning of the range
 297  * @end: end of the range
 298  *
 299  * Updates the probe callback corresponding to a range of tracepoints.
 300  */
 301 void
 302 tracepoint_update_probe_range(struct tracepoint *begin, struct tracepoint *end)
 303 {
 304         struct tracepoint *iter;
 305         struct tracepoint_entry *mark_entry;
 306
 307         if (!begin)
 308                 return;
 309
 310         mutex_lock(&tracepoints_mutex);
 311         for (iter = begin; iter < end; iter++) {
 312                 mark_entry = get_tracepoint(iter->name);
 313                 if (mark_entry) {
 314                         set_tracepoint(&mark_entry, iter,
 315                                         !!mark_entry->refcount);
 316                 } else {
 317                         disable_tracepoint(iter);
 318                 }
 319         }
 320         mutex_unlock(&tracepoints_mutex);
 321 }
 322
 323 /*
 324  * Update probes, removing the faulty probes.
 325  */
 326 static void tracepoint_update_probes(void)
 327 {
 328         /* Core kernel tracepoints */
 329         tracepoint_update_probe_range(__start___tracepoints,
 330                 __stop___tracepoints);
 331         /* tracepoints in modules. */
 332         module_update_tracepoints();
 333 }
 334
 335 static struct tracepoint_func *
 336 tracepoint_add_probe(const char *name, void *probe, void *data)
 337 {
 338         struct tracepoint_entry *entry;
 339         struct tracepoint_func *old;
 340
 341         entry = get_tracepoint(name);
 342         if (!entry) {
 343                 entry = add_tracepoint(name);
 344                 if (IS_ERR(entry))
 345                         return (struct tracepoint_func *)entry;
 346         }
 347         old = tracepoint_entry_add_probe(entry, probe, data);
 348         if (IS_ERR(old) && !entry->refcount)
 349                 remove_tracepoint(entry);
 350         return old;
 351 }
 352
 353 /**
 354  * tracepoint_probe_register -  Connect a probe to a tracepoint
 355  * @name: tracepoint name
 356  * @probe: probe handler
 357  *
 358  * Returns 0 if ok, error value on error.
 359  * The probe address must at least be aligned on the architecture pointer size.
 360  */
 361 int tracepoint_probe_register(const char *name, void *probe, void *data)
 362 {
 363         struct tracepoint_func *old;
 364
 365         mutex_lock(&tracepoints_mutex);
 366         old = tracepoint_add_probe(name, probe, data);
 367         mutex_unlock(&tracepoints_mutex);
 368         if (IS_ERR(old))
 369                 return PTR_ERR(old);
 370
 371         tracepoint_update_probes();             /* may update entry */
 372         release_probes(old);
 373         return 0;
 374 }
 375 EXPORT_SYMBOL_GPL(tracepoint_probe_register);
 376
 377 static struct tracepoint_func *
 378 tracepoint_remove_probe(const char *name, void *probe, void *data)
 379 {
 380         struct tracepoint_entry *entry;
 381         struct tracepoint_func *old;
 382
 383         entry = get_tracepoint(name);
 384         if (!entry)
 385                 return ERR_PTR(-ENOENT);
 386         old = tracepoint_entry_remove_probe(entry, probe, data);
 387         if (IS_ERR(old))
 388                 return old;
 389         if (!entry->refcount)
 390                 remove_tracepoint(entry);
 391         return old;
 392 }
 393
 394 /**
 395  * tracepoint_probe_unregister -  Disconnect a probe from a tracepoint
 396  * @name: tracepoint name
 397  * @probe: probe function pointer
 398  *
 399  * We do not need to call a synchronize_sched to make sure the probes have
 400  * finished running before doing a module unload, because the module unload
 401  * itself uses stop_machine(), which insures that every preempt disabled section
 402  * have finished.
 403  */
 404 int tracepoint_probe_unregister(const char *name, void *probe, void *data)
 405 {
 406         struct tracepoint_func *old;
 407
 408         mutex_lock(&tracepoints_mutex);
 409         old = tracepoint_remove_probe(name, probe, data);
 410         mutex_unlock(&tracepoints_mutex);
 411         if (IS_ERR(old))
 412                 return PTR_ERR(old);
 413
 414         tracepoint_update_probes();             /* may update entry */
 415         release_probes(old);
 416         return 0;
 417 }
 418 EXPORT_SYMBOL_GPL(tracepoint_probe_unregister);
 419
 420 static LIST_HEAD(old_probes);
 421 static int need_update;
 422
 423 static void tracepoint_add_old_probes(void *old)
 424 {
 425         need_update = 1;
 426         if (old) {
 427                 struct tp_probes *tp_probes = container_of(old,
 428                         struct tp_probes, probes[0]);
 429                 list_add(&tp_probes->u.list, &old_probes);
 430         }
 431 }
 432
 433 /**
 434  * tracepoint_probe_register_noupdate -  register a probe but not connect
 435  * @name: tracepoint name
 436  * @probe: probe handler
 437  *
 438  * caller must call tracepoint_probe_update_all()
 439  */
 440 int tracepoint_probe_register_noupdate(const char *name, void *probe,
 441                                        void *data)
 442 {
 443         struct tracepoint_func *old;
 444
 445         mutex_lock(&tracepoints_mutex);
 446         old = tracepoint_add_probe(name, probe, data);
 447         if (IS_ERR(old)) {
 448                 mutex_unlock(&tracepoints_mutex);
 449                 return PTR_ERR(old);
 450         }
 451         tracepoint_add_old_probes(old);
 452         mutex_unlock(&tracepoints_mutex);
 453         return 0;
 454 }
 455 EXPORT_SYMBOL_GPL(tracepoint_probe_register_noupdate);
 456
 457 /**
 458  * tracepoint_probe_unregister_noupdate -  remove a probe but not disconnect
 459  * @name: tracepoint name
 460  * @probe: probe function pointer
 461  *
 462  * caller must call tracepoint_probe_update_all()
 463  */
 464 int tracepoint_probe_unregister_noupdate(const char *name, void *probe,
 465                                          void *data)
 466 {
 467         struct tracepoint_func *old;
 468
 469         mutex_lock(&tracepoints_mutex);
 470         old = tracepoint_remove_probe(name, probe, data);
 471         if (IS_ERR(old)) {
 472                 mutex_unlock(&tracepoints_mutex);
 473                 return PTR_ERR(old);
 474         }
 475         tracepoint_add_old_probes(old);
 476         mutex_unlock(&tracepoints_mutex);
 477         return 0;
 478 }
 479 EXPORT_SYMBOL_GPL(tracepoint_probe_unregister_noupdate);
 480
 481 /**
 482  * tracepoint_probe_update_all -  update tracepoints
 483  */
 484 void tracepoint_probe_update_all(void)
 485 {
 486         LIST_HEAD(release_probes);
 487         struct tp_probes *pos, *next;
 488
 489         mutex_lock(&tracepoints_mutex);
 490         if (!need_update) {
 491                 mutex_unlock(&tracepoints_mutex);
 492                 return;
 493         }
 494         if (!list_empty(&old_probes))
 495                 list_replace_init(&old_probes, &release_probes);
 496         need_update = 0;
 497         mutex_unlock(&tracepoints_mutex);
 498
 499         tracepoint_update_probes();
 500         list_for_each_entry_safe(pos, next, &release_probes, u.list) {
 501                 list_del(&pos->u.list);
 502                 call_rcu_sched(&pos->u.rcu, rcu_free_old_probes);
 503         }
 504 }
 505 EXPORT_SYMBOL_GPL(tracepoint_probe_update_all);
 506
 507 /**
 508  * tracepoint_get_iter_range - Get a next tracepoint iterator given a range.
 509  * @tracepoint: current tracepoints (in), next tracepoint (out)
 510  * @begin: beginning of the range
 511  * @end: end of the range
 512  *
 513  * Returns whether a next tracepoint has been found (1) or not (0).
 514  * Will return the first tracepoint in the range if the input tracepoint is
 515  * NULL.
 516  */
 517 int tracepoint_get_iter_range(struct tracepoint **tracepoint,
 518         struct tracepoint *begin, struct tracepoint *end)
 519 {
 520         if (!*tracepoint && begin != end) {
 521                 *tracepoint = begin;
 522                 return 1;
 523         }
 524         if (*tracepoint >= begin && *tracepoint < end)
 525                 return 1;
 526         return 0;
 527 }
 528 EXPORT_SYMBOL_GPL(tracepoint_get_iter_range);
 529
 530 static void tracepoint_get_iter(struct tracepoint_iter *iter)
 531 {
 532         int found = 0;
 533
 534         /* Core kernel tracepoints */
 535         if (!iter->module) {
 536                 found = tracepoint_get_iter_range(&iter->tracepoint,
 537                                 __start___tracepoints, __stop___tracepoints);
 538                 if (found)
 539                         goto end;
 540         }
 541         /* tracepoints in modules. */
 542         found = module_get_iter_tracepoints(iter);
 543 end:
 544         if (!found)
 545                 tracepoint_iter_reset(iter);
 546 }
 547
 548 void tracepoint_iter_start(struct tracepoint_iter *iter)
 549 {
 550         tracepoint_get_iter(iter);
 551 }
 552 EXPORT_SYMBOL_GPL(tracepoint_iter_start);
 553
 554 void tracepoint_iter_next(struct tracepoint_iter *iter)
 555 {
 556         iter->tracepoint++;
 557         /*
 558          * iter->tracepoint may be invalid because we blindly incremented it.
 559          * Make sure it is valid by marshalling on the tracepoints, getting the
 560          * tracepoints from following modules if necessary.
 561          */
 562         tracepoint_get_iter(iter);
 563 }
 564 EXPORT_SYMBOL_GPL(tracepoint_iter_next);
 565
 566 void tracepoint_iter_stop(struct tracepoint_iter *iter)
 567 {
 568 }
 569 EXPORT_SYMBOL_GPL(tracepoint_iter_stop);
 570
 571 void tracepoint_iter_reset(struct tracepoint_iter *iter)
 572 {
 573         iter->module = NULL;
 574         iter->tracepoint = NULL;
 575 }
 576 EXPORT_SYMBOL_GPL(tracepoint_iter_reset);
 577
 578 #ifdef CONFIG_MODULES
 579
 580 int tracepoint_module_notify(struct notifier_block *self,
 581                              unsigned long val, void *data)
 582 {
 583         struct module *mod = data;
 584
 585         switch (val) {
 586         case MODULE_STATE_COMING:
 587         case MODULE_STATE_GOING:
 588                 tracepoint_update_probe_range(mod->tracepoints,
 589                         mod->tracepoints + mod->num_tracepoints);
 590                 break;
 591         }
 592         return 0;
 593 }
 594
 595 struct notifier_block tracepoint_module_nb = {
 596         .notifier_call = tracepoint_module_notify,
 597         .priority = 0,
 598 };
 599
 600 static int init_tracepoints(void)
 601 {
 602         return register_module_notifier(&tracepoint_module_nb);
 603 }
 604 __initcall(init_tracepoints);
 605
 606 #endif /* CONFIG_MODULES */
 607
 608 #ifdef CONFIG_HAVE_SYSCALL_TRACEPOINTS
 609
 610 /* NB: reg/unreg are called while guarded with the tracepoints_mutex */
 611 static int sys_tracepoint_refcount;
 612
 613 void syscall_regfunc(void)
 614 {
 615         unsigned long flags;
 616         struct task_struct *g, *t;
 617
 618         if (!sys_tracepoint_refcount) {
 619                 read_lock_irqsave(&tasklist_lock, flags);
 620                 do_each_thread(g, t) {
 621                         /* Skip kernel threads. */
 622                         if (t->mm)
 623                                 set_tsk_thread_flag(t, TIF_SYSCALL_TRACEPOINT);
 624                 } while_each_thread(g, t);
 625                 read_unlock_irqrestore(&tasklist_lock, flags);
 626         }
 627         sys_tracepoint_refcount++;
 628 }
 629
 630 void syscall_unregfunc(void)
 631 {
 632         unsigned long flags;
 633         struct task_struct *g, *t;
 634
 635         sys_tracepoint_refcount--;
 636         if (!sys_tracepoint_refcount) {
 637                 read_lock_irqsave(&tasklist_lock, flags);
 638                 do_each_thread(g, t) {
 639                         clear_tsk_thread_flag(t, TIF_SYSCALL_TRACEPOINT);
 640                 } while_each_thread(g, t);
 641                 read_unlock_irqrestore(&tasklist_lock, flags);
 642         }
 643 }
 644 #endif