tracepoints: use modules notifiers
[linux-2.6.git] / 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
28 extern struct tracepoint __start___tracepoints[];
29 extern struct tracepoint __stop___tracepoints[];
30
31 /* Set to 1 to enable tracepoint debug output */
32 static const int tracepoint_debug;
33
34 /*
35  * tracepoints_mutex nests inside module_mutex. Tracepoints mutex protects the
36  * builtin and module tracepoints and the hash table.
37  */
38 static DEFINE_MUTEX(tracepoints_mutex);
39
40 /*
41  * Tracepoint hash table, containing the active tracepoints.
42  * Protected by tracepoints_mutex.
43  */
44 #define TRACEPOINT_HASH_BITS 6
45 #define TRACEPOINT_TABLE_SIZE (1 << TRACEPOINT_HASH_BITS)
46 static struct hlist_head tracepoint_table[TRACEPOINT_TABLE_SIZE];
47
48 /*
49  * Note about RCU :
50  * It is used to to delay the free of multiple probes array until a quiescent
51  * state is reached.
52  * Tracepoint entries modifications are protected by the tracepoints_mutex.
53  */
54 struct tracepoint_entry {
55         struct hlist_node hlist;
56         void **funcs;
57         int refcount;   /* Number of times armed. 0 if disarmed. */
58         char name[0];
59 };
60
61 struct tp_probes {
62         union {
63                 struct rcu_head rcu;
64                 struct list_head list;
65         } u;
66         void *probes[0];
67 };
68
69 static inline void *allocate_probes(int count)
70 {
71         struct tp_probes *p  = kmalloc(count * sizeof(void *)
72                         + sizeof(struct tp_probes), GFP_KERNEL);
73         return p == NULL ? NULL : p->probes;
74 }
75
76 static void rcu_free_old_probes(struct rcu_head *head)
77 {
78         kfree(container_of(head, struct tp_probes, u.rcu));
79 }
80
81 static inline void release_probes(void *old)
82 {
83         if (old) {
84                 struct tp_probes *tp_probes = container_of(old,
85                         struct tp_probes, probes[0]);
86                 call_rcu_sched(&tp_probes->u.rcu, rcu_free_old_probes);
87         }
88 }
89
90 static void debug_print_probes(struct tracepoint_entry *entry)
91 {
92         int i;
93
94         if (!tracepoint_debug || !entry->funcs)
95                 return;
96
97         for (i = 0; entry->funcs[i]; i++)
98                 printk(KERN_DEBUG "Probe %d : %p\n", i, entry->funcs[i]);
99 }
100
101 static void *
102 tracepoint_entry_add_probe(struct tracepoint_entry *entry, void *probe)
103 {
104         int nr_probes = 0;
105         void **old, **new;
106
107         WARN_ON(!probe);
108
109         debug_print_probes(entry);
110         old = entry->funcs;
111         if (old) {
112                 /* (N -> N+1), (N != 0, 1) probes */
113                 for (nr_probes = 0; old[nr_probes]; nr_probes++)
114                         if (old[nr_probes] == probe)
115                                 return ERR_PTR(-EEXIST);
116         }
117         /* + 2 : one for new probe, one for NULL func */
118         new = allocate_probes(nr_probes + 2);
119         if (new == NULL)
120                 return ERR_PTR(-ENOMEM);
121         if (old)
122                 memcpy(new, old, nr_probes * sizeof(void *));
123         new[nr_probes] = probe;
124         new[nr_probes + 1] = NULL;
125         entry->refcount = nr_probes + 1;
126         entry->funcs = new;
127         debug_print_probes(entry);
128         return old;
129 }
130
131 static void *
132 tracepoint_entry_remove_probe(struct tracepoint_entry *entry, void *probe)
133 {
134         int nr_probes = 0, nr_del = 0, i;
135         void **old, **new;
136
137         old = entry->funcs;
138
139         if (!old)
140                 return ERR_PTR(-ENOENT);
141
142         debug_print_probes(entry);
143         /* (N -> M), (N > 1, M >= 0) probes */
144         for (nr_probes = 0; old[nr_probes]; nr_probes++) {
145                 if ((!probe || old[nr_probes] == probe))
146                         nr_del++;
147         }
148
149         if (nr_probes - nr_del == 0) {
150                 /* N -> 0, (N > 1) */
151                 entry->funcs = NULL;
152                 entry->refcount = 0;
153                 debug_print_probes(entry);
154                 return old;
155         } else {
156                 int j = 0;
157                 /* N -> M, (N > 1, M > 0) */
158                 /* + 1 for NULL */
159                 new = allocate_probes(nr_probes - nr_del + 1);
160                 if (new == NULL)
161                         return ERR_PTR(-ENOMEM);
162                 for (i = 0; old[i]; i++)
163                         if ((probe && old[i] != probe))
164                                 new[j++] = old[i];
165                 new[nr_probes - nr_del] = NULL;
166                 entry->refcount = nr_probes - nr_del;
167                 entry->funcs = new;
168         }
169         debug_print_probes(entry);
170         return old;
171 }
172
173 /*
174  * Get tracepoint if the tracepoint is present in the tracepoint hash table.
175  * Must be called with tracepoints_mutex held.
176  * Returns NULL if not present.
177  */
178 static struct tracepoint_entry *get_tracepoint(const char *name)
179 {
180         struct hlist_head *head;
181         struct hlist_node *node;
182         struct tracepoint_entry *e;
183         u32 hash = jhash(name, strlen(name), 0);
184
185         head = &tracepoint_table[hash & (TRACEPOINT_TABLE_SIZE - 1)];
186         hlist_for_each_entry(e, node, head, hlist) {
187                 if (!strcmp(name, e->name))
188                         return e;
189         }
190         return NULL;
191 }
192
193 /*
194  * Add the tracepoint to the tracepoint hash table. Must be called with
195  * tracepoints_mutex held.
196  */
197 static struct tracepoint_entry *add_tracepoint(const char *name)
198 {
199         struct hlist_head *head;
200         struct hlist_node *node;
201         struct tracepoint_entry *e;
202         size_t name_len = strlen(name) + 1;
203         u32 hash = jhash(name, name_len-1, 0);
204
205         head = &tracepoint_table[hash & (TRACEPOINT_TABLE_SIZE - 1)];
206         hlist_for_each_entry(e, node, head, hlist) {
207                 if (!strcmp(name, e->name)) {
208                         printk(KERN_NOTICE
209                                 "tracepoint %s busy\n", name);
210                         return ERR_PTR(-EEXIST);        /* Already there */
211                 }
212         }
213         /*
214          * Using kmalloc here to allocate a variable length element. Could
215          * cause some memory fragmentation if overused.
216          */
217         e = kmalloc(sizeof(struct tracepoint_entry) + name_len, GFP_KERNEL);
218         if (!e)
219                 return ERR_PTR(-ENOMEM);
220         memcpy(&e->name[0], name, name_len);
221         e->funcs = NULL;
222         e->refcount = 0;
223         hlist_add_head(&e->hlist, head);
224         return e;
225 }
226
227 /*
228  * Remove the tracepoint from the tracepoint hash table. Must be called with
229  * mutex_lock held.
230  */
231 static inline void remove_tracepoint(struct tracepoint_entry *e)
232 {
233         hlist_del(&e->hlist);
234         kfree(e);
235 }
236
237 /*
238  * Sets the probe callback corresponding to one tracepoint.
239  */
240 static void set_tracepoint(struct tracepoint_entry **entry,
241         struct tracepoint *elem, int active)
242 {
243         WARN_ON(strcmp((*entry)->name, elem->name) != 0);
244
245         /*
246          * rcu_assign_pointer has a smp_wmb() which makes sure that the new
247          * probe callbacks array is consistent before setting a pointer to it.
248          * This array is referenced by __DO_TRACE from
249          * include/linux/tracepoints.h. A matching smp_read_barrier_depends()
250          * is used.
251          */
252         rcu_assign_pointer(elem->funcs, (*entry)->funcs);
253         elem->state = active;
254 }
255
256 /*
257  * Disable a tracepoint and its probe callback.
258  * Note: only waiting an RCU period after setting elem->call to the empty
259  * function insures that the original callback is not used anymore. This insured
260  * by preempt_disable around the call site.
261  */
262 static void disable_tracepoint(struct tracepoint *elem)
263 {
264         elem->state = 0;
265         rcu_assign_pointer(elem->funcs, NULL);
266 }
267
268 /**
269  * tracepoint_update_probe_range - Update a probe range
270  * @begin: beginning of the range
271  * @end: end of the range
272  *
273  * Updates the probe callback corresponding to a range of tracepoints.
274  */
275 void tracepoint_update_probe_range(struct tracepoint *begin,
276         struct tracepoint *end)
277 {
278         struct tracepoint *iter;
279         struct tracepoint_entry *mark_entry;
280
281         mutex_lock(&tracepoints_mutex);
282         for (iter = begin; iter < end; iter++) {
283                 mark_entry = get_tracepoint(iter->name);
284                 if (mark_entry) {
285                         set_tracepoint(&mark_entry, iter,
286                                         !!mark_entry->refcount);
287                 } else {
288                         disable_tracepoint(iter);
289                 }
290         }
291         mutex_unlock(&tracepoints_mutex);
292 }
293
294 /*
295  * Update probes, removing the faulty probes.
296  */
297 static void tracepoint_update_probes(void)
298 {
299         /* Core kernel tracepoints */
300         tracepoint_update_probe_range(__start___tracepoints,
301                 __stop___tracepoints);
302         /* tracepoints in modules. */
303         module_update_tracepoints();
304 }
305
306 static void *tracepoint_add_probe(const char *name, void *probe)
307 {
308         struct tracepoint_entry *entry;
309         void *old;
310
311         entry = get_tracepoint(name);
312         if (!entry) {
313                 entry = add_tracepoint(name);
314                 if (IS_ERR(entry))
315                         return entry;
316         }
317         old = tracepoint_entry_add_probe(entry, probe);
318         if (IS_ERR(old) && !entry->refcount)
319                 remove_tracepoint(entry);
320         return old;
321 }
322
323 /**
324  * tracepoint_probe_register -  Connect a probe to a tracepoint
325  * @name: tracepoint name
326  * @probe: probe handler
327  *
328  * Returns 0 if ok, error value on error.
329  * The probe address must at least be aligned on the architecture pointer size.
330  */
331 int tracepoint_probe_register(const char *name, void *probe)
332 {
333         void *old;
334
335         mutex_lock(&tracepoints_mutex);
336         old = tracepoint_add_probe(name, probe);
337         mutex_unlock(&tracepoints_mutex);
338         if (IS_ERR(old))
339                 return PTR_ERR(old);
340
341         tracepoint_update_probes();             /* may update entry */
342         release_probes(old);
343         return 0;
344 }
345 EXPORT_SYMBOL_GPL(tracepoint_probe_register);
346
347 static void *tracepoint_remove_probe(const char *name, void *probe)
348 {
349         struct tracepoint_entry *entry;
350         void *old;
351
352         entry = get_tracepoint(name);
353         if (!entry)
354                 return ERR_PTR(-ENOENT);
355         old = tracepoint_entry_remove_probe(entry, probe);
356         if (IS_ERR(old))
357                 return old;
358         if (!entry->refcount)
359                 remove_tracepoint(entry);
360         return old;
361 }
362
363 /**
364  * tracepoint_probe_unregister -  Disconnect a probe from a tracepoint
365  * @name: tracepoint name
366  * @probe: probe function pointer
367  *
368  * We do not need to call a synchronize_sched to make sure the probes have
369  * finished running before doing a module unload, because the module unload
370  * itself uses stop_machine(), which insures that every preempt disabled section
371  * have finished.
372  */
373 int tracepoint_probe_unregister(const char *name, void *probe)
374 {
375         void *old;
376
377         mutex_lock(&tracepoints_mutex);
378         old = tracepoint_remove_probe(name, probe);
379         mutex_unlock(&tracepoints_mutex);
380         if (IS_ERR(old))
381                 return PTR_ERR(old);
382
383         tracepoint_update_probes();             /* may update entry */
384         release_probes(old);
385         return 0;
386 }
387 EXPORT_SYMBOL_GPL(tracepoint_probe_unregister);
388
389 static LIST_HEAD(old_probes);
390 static int need_update;
391
392 static void tracepoint_add_old_probes(void *old)
393 {
394         need_update = 1;
395         if (old) {
396                 struct tp_probes *tp_probes = container_of(old,
397                         struct tp_probes, probes[0]);
398                 list_add(&tp_probes->u.list, &old_probes);
399         }
400 }
401
402 /**
403  * tracepoint_probe_register_noupdate -  register a probe but not connect
404  * @name: tracepoint name
405  * @probe: probe handler
406  *
407  * caller must call tracepoint_probe_update_all()
408  */
409 int tracepoint_probe_register_noupdate(const char *name, void *probe)
410 {
411         void *old;
412
413         mutex_lock(&tracepoints_mutex);
414         old = tracepoint_add_probe(name, probe);
415         if (IS_ERR(old)) {
416                 mutex_unlock(&tracepoints_mutex);
417                 return PTR_ERR(old);
418         }
419         tracepoint_add_old_probes(old);
420         mutex_unlock(&tracepoints_mutex);
421         return 0;
422 }
423 EXPORT_SYMBOL_GPL(tracepoint_probe_register_noupdate);
424
425 /**
426  * tracepoint_probe_unregister_noupdate -  remove a probe but not disconnect
427  * @name: tracepoint name
428  * @probe: probe function pointer
429  *
430  * caller must call tracepoint_probe_update_all()
431  */
432 int tracepoint_probe_unregister_noupdate(const char *name, void *probe)
433 {
434         void *old;
435
436         mutex_lock(&tracepoints_mutex);
437         old = tracepoint_remove_probe(name, probe);
438         if (IS_ERR(old)) {
439                 mutex_unlock(&tracepoints_mutex);
440                 return PTR_ERR(old);
441         }
442         tracepoint_add_old_probes(old);
443         mutex_unlock(&tracepoints_mutex);
444         return 0;
445 }
446 EXPORT_SYMBOL_GPL(tracepoint_probe_unregister_noupdate);
447
448 /**
449  * tracepoint_probe_update_all -  update tracepoints
450  */
451 void tracepoint_probe_update_all(void)
452 {
453         LIST_HEAD(release_probes);
454         struct tp_probes *pos, *next;
455
456         mutex_lock(&tracepoints_mutex);
457         if (!need_update) {
458                 mutex_unlock(&tracepoints_mutex);
459                 return;
460         }
461         if (!list_empty(&old_probes))
462                 list_replace_init(&old_probes, &release_probes);
463         need_update = 0;
464         mutex_unlock(&tracepoints_mutex);
465
466         tracepoint_update_probes();
467         list_for_each_entry_safe(pos, next, &release_probes, u.list) {
468                 list_del(&pos->u.list);
469                 call_rcu_sched(&pos->u.rcu, rcu_free_old_probes);
470         }
471 }
472 EXPORT_SYMBOL_GPL(tracepoint_probe_update_all);
473
474 /**
475  * tracepoint_get_iter_range - Get a next tracepoint iterator given a range.
476  * @tracepoint: current tracepoints (in), next tracepoint (out)
477  * @begin: beginning of the range
478  * @end: end of the range
479  *
480  * Returns whether a next tracepoint has been found (1) or not (0).
481  * Will return the first tracepoint in the range if the input tracepoint is
482  * NULL.
483  */
484 int tracepoint_get_iter_range(struct tracepoint **tracepoint,
485         struct tracepoint *begin, struct tracepoint *end)
486 {
487         if (!*tracepoint && begin != end) {
488                 *tracepoint = begin;
489                 return 1;
490         }
491         if (*tracepoint >= begin && *tracepoint < end)
492                 return 1;
493         return 0;
494 }
495 EXPORT_SYMBOL_GPL(tracepoint_get_iter_range);
496
497 static void tracepoint_get_iter(struct tracepoint_iter *iter)
498 {
499         int found = 0;
500
501         /* Core kernel tracepoints */
502         if (!iter->module) {
503                 found = tracepoint_get_iter_range(&iter->tracepoint,
504                                 __start___tracepoints, __stop___tracepoints);
505                 if (found)
506                         goto end;
507         }
508         /* tracepoints in modules. */
509         found = module_get_iter_tracepoints(iter);
510 end:
511         if (!found)
512                 tracepoint_iter_reset(iter);
513 }
514
515 void tracepoint_iter_start(struct tracepoint_iter *iter)
516 {
517         tracepoint_get_iter(iter);
518 }
519 EXPORT_SYMBOL_GPL(tracepoint_iter_start);
520
521 void tracepoint_iter_next(struct tracepoint_iter *iter)
522 {
523         iter->tracepoint++;
524         /*
525          * iter->tracepoint may be invalid because we blindly incremented it.
526          * Make sure it is valid by marshalling on the tracepoints, getting the
527          * tracepoints from following modules if necessary.
528          */
529         tracepoint_get_iter(iter);
530 }
531 EXPORT_SYMBOL_GPL(tracepoint_iter_next);
532
533 void tracepoint_iter_stop(struct tracepoint_iter *iter)
534 {
535 }
536 EXPORT_SYMBOL_GPL(tracepoint_iter_stop);
537
538 void tracepoint_iter_reset(struct tracepoint_iter *iter)
539 {
540         iter->module = NULL;
541         iter->tracepoint = NULL;
542 }
543 EXPORT_SYMBOL_GPL(tracepoint_iter_reset);
544
545 int tracepoint_module_notify(struct notifier_block *self,
546                              unsigned long val, void *data)
547 {
548         struct module *mod = data;
549
550         switch (val) {
551         case MODULE_STATE_COMING:
552                 tracepoint_update_probe_range(mod->tracepoints,
553                         mod->tracepoints + mod->num_tracepoints);
554                 break;
555         case MODULE_STATE_GOING:
556                 tracepoint_update_probe_range(mod->tracepoints,
557                         mod->tracepoints + mod->num_tracepoints);
558                 break;
559         }
560         return 0;
561 }
562
563 struct notifier_block tracepoint_module_nb = {
564         .notifier_call = tracepoint_module_notify,
565         .priority = 0,
566 };
567
568 static int init_tracepoints(void)
569 {
570         return register_module_notifier(&tracepoint_module_nb);
571 }
572 __initcall(init_tracepoints);