2 * Copyright (C) 2007 Mathieu Desnoyers
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.
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.
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.
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/marker.h>
25 #include <linux/err.h>
26 #include <linux/slab.h>
28 extern struct marker __start___markers[];
29 extern struct marker __stop___markers[];
31 /* Set to 1 to enable marker debug output */
32 static const int marker_debug;
35 * markers_mutex nests inside module_mutex. Markers mutex protects the builtin
36 * and module markers and the hash table.
38 static DEFINE_MUTEX(markers_mutex);
41 * Marker hash table, containing the active markers.
42 * Protected by module_mutex.
44 #define MARKER_HASH_BITS 6
45 #define MARKER_TABLE_SIZE (1 << MARKER_HASH_BITS)
49 * It is used to make sure every handler has finished using its private data
50 * between two consecutive operation (add or remove) on a given marker. It is
51 * also used to delay the free of multiple probes array until a quiescent state
53 * marker entries modifications are protected by the markers_mutex.
56 struct hlist_node hlist;
59 void (*call)(const struct marker *mdata, void *call_private, ...);
60 struct marker_probe_closure single;
61 struct marker_probe_closure *multi;
62 int refcount; /* Number of times armed. 0 if disarmed. */
63 unsigned char ptype:1;
64 char name[0]; /* Contains name'\0'format'\0' */
67 static struct hlist_head marker_table[MARKER_TABLE_SIZE];
70 * __mark_empty_function - Empty probe callback
71 * @probe_private: probe private data
72 * @call_private: call site private data
74 * @...: variable argument list
76 * Empty callback provided as a probe to the markers. By providing this to a
77 * disabled marker, we make sure the execution flow is always valid even
78 * though the function pointer change and the marker enabling are two distinct
79 * operations that modifies the execution flow of preemptible code.
81 void __mark_empty_function(void *probe_private, void *call_private,
82 const char *fmt, va_list *args)
85 EXPORT_SYMBOL_GPL(__mark_empty_function);
88 * marker_probe_cb Callback that prepares the variable argument list for probes.
89 * @mdata: pointer of type struct marker
90 * @call_private: caller site private data
91 * @...: Variable argument list.
93 * Since we do not use "typical" pointer based RCU in the 1 argument case, we
94 * need to put a full smp_rmb() in this branch. This is why we do not use
95 * rcu_dereference() for the pointer read.
97 void marker_probe_cb(const struct marker *mdata, void *call_private, ...)
103 * rcu_read_lock_sched does two things : disabling preemption to make
104 * sure the teardown of the callbacks can be done correctly when they
105 * are in modules and they insure RCU read coherency.
107 rcu_read_lock_sched();
108 ptype = mdata->ptype;
109 if (likely(!ptype)) {
110 marker_probe_func *func;
111 /* Must read the ptype before ptr. They are not data dependant,
112 * so we put an explicit smp_rmb() here. */
114 func = mdata->single.func;
115 /* Must read the ptr before private data. They are not data
116 * dependant, so we put an explicit smp_rmb() here. */
118 va_start(args, call_private);
119 func(mdata->single.probe_private, call_private, mdata->format,
123 struct marker_probe_closure *multi;
126 * Read mdata->ptype before mdata->multi.
129 multi = mdata->multi;
131 * multi points to an array, therefore accessing the array
132 * depends on reading multi. However, even in this case,
133 * we must insure that the pointer is read _before_ the array
134 * data. Same as rcu_dereference, but we need a full smp_rmb()
135 * in the fast path, so put the explicit barrier here.
137 smp_read_barrier_depends();
138 for (i = 0; multi[i].func; i++) {
139 va_start(args, call_private);
140 multi[i].func(multi[i].probe_private, call_private,
141 mdata->format, &args);
145 rcu_read_unlock_sched();
147 EXPORT_SYMBOL_GPL(marker_probe_cb);
150 * marker_probe_cb Callback that does not prepare the variable argument list.
151 * @mdata: pointer of type struct marker
152 * @call_private: caller site private data
153 * @...: Variable argument list.
155 * Should be connected to markers "MARK_NOARGS".
157 void marker_probe_cb_noarg(const struct marker *mdata, void *call_private, ...)
159 va_list args; /* not initialized */
162 rcu_read_lock_sched();
163 ptype = mdata->ptype;
164 if (likely(!ptype)) {
165 marker_probe_func *func;
166 /* Must read the ptype before ptr. They are not data dependant,
167 * so we put an explicit smp_rmb() here. */
169 func = mdata->single.func;
170 /* Must read the ptr before private data. They are not data
171 * dependant, so we put an explicit smp_rmb() here. */
173 func(mdata->single.probe_private, call_private, mdata->format,
176 struct marker_probe_closure *multi;
179 * Read mdata->ptype before mdata->multi.
182 multi = mdata->multi;
184 * multi points to an array, therefore accessing the array
185 * depends on reading multi. However, even in this case,
186 * we must insure that the pointer is read _before_ the array
187 * data. Same as rcu_dereference, but we need a full smp_rmb()
188 * in the fast path, so put the explicit barrier here.
190 smp_read_barrier_depends();
191 for (i = 0; multi[i].func; i++)
192 multi[i].func(multi[i].probe_private, call_private,
193 mdata->format, &args);
195 rcu_read_unlock_sched();
197 EXPORT_SYMBOL_GPL(marker_probe_cb_noarg);
199 static void debug_print_probes(struct marker_entry *entry)
207 printk(KERN_DEBUG "Single probe : %p %p\n",
209 entry->single.probe_private);
211 for (i = 0; entry->multi[i].func; i++)
212 printk(KERN_DEBUG "Multi probe %d : %p %p\n", i,
213 entry->multi[i].func,
214 entry->multi[i].probe_private);
218 static struct marker_probe_closure *
219 marker_entry_add_probe(struct marker_entry *entry,
220 marker_probe_func *probe, void *probe_private)
223 struct marker_probe_closure *old, *new;
227 debug_print_probes(entry);
230 if (entry->single.func == probe &&
231 entry->single.probe_private == probe_private)
232 return ERR_PTR(-EBUSY);
233 if (entry->single.func == __mark_empty_function) {
235 entry->single.func = probe;
236 entry->single.probe_private = probe_private;
239 debug_print_probes(entry);
247 /* (N -> N+1), (N != 0, 1) probes */
248 for (nr_probes = 0; old[nr_probes].func; nr_probes++)
249 if (old[nr_probes].func == probe
250 && old[nr_probes].probe_private
252 return ERR_PTR(-EBUSY);
254 /* + 2 : one for new probe, one for NULL func */
255 new = kzalloc((nr_probes + 2) * sizeof(struct marker_probe_closure),
258 return ERR_PTR(-ENOMEM);
260 new[0] = entry->single;
263 nr_probes * sizeof(struct marker_probe_closure));
264 new[nr_probes].func = probe;
265 new[nr_probes].probe_private = probe_private;
266 entry->refcount = nr_probes + 1;
269 debug_print_probes(entry);
273 static struct marker_probe_closure *
274 marker_entry_remove_probe(struct marker_entry *entry,
275 marker_probe_func *probe, void *probe_private)
277 int nr_probes = 0, nr_del = 0, i;
278 struct marker_probe_closure *old, *new;
282 debug_print_probes(entry);
284 /* 0 -> N is an error */
285 WARN_ON(entry->single.func == __mark_empty_function);
287 WARN_ON(probe && entry->single.func != probe);
288 WARN_ON(entry->single.probe_private != probe_private);
289 entry->single.func = __mark_empty_function;
292 debug_print_probes(entry);
295 /* (N -> M), (N > 1, M >= 0) probes */
296 for (nr_probes = 0; old[nr_probes].func; nr_probes++) {
297 if ((!probe || old[nr_probes].func == probe)
298 && old[nr_probes].probe_private
304 if (nr_probes - nr_del == 0) {
305 /* N -> 0, (N > 1) */
306 entry->single.func = __mark_empty_function;
309 } else if (nr_probes - nr_del == 1) {
310 /* N -> 1, (N > 1) */
311 for (i = 0; old[i].func; i++)
312 if ((probe && old[i].func != probe) ||
313 old[i].probe_private != probe_private)
314 entry->single = old[i];
319 /* N -> M, (N > 1, M > 1) */
321 new = kzalloc((nr_probes - nr_del + 1)
322 * sizeof(struct marker_probe_closure), GFP_KERNEL);
324 return ERR_PTR(-ENOMEM);
325 for (i = 0; old[i].func; i++)
326 if ((probe && old[i].func != probe) ||
327 old[i].probe_private != probe_private)
329 entry->refcount = nr_probes - nr_del;
333 debug_print_probes(entry);
338 * Get marker if the marker is present in the marker hash table.
339 * Must be called with markers_mutex held.
340 * Returns NULL if not present.
342 static struct marker_entry *get_marker(const char *name)
344 struct hlist_head *head;
345 struct hlist_node *node;
346 struct marker_entry *e;
347 u32 hash = jhash(name, strlen(name), 0);
349 head = &marker_table[hash & ((1 << MARKER_HASH_BITS)-1)];
350 hlist_for_each_entry(e, node, head, hlist) {
351 if (!strcmp(name, e->name))
358 * Add the marker to the marker hash table. Must be called with markers_mutex
361 static struct marker_entry *add_marker(const char *name, const char *format)
363 struct hlist_head *head;
364 struct hlist_node *node;
365 struct marker_entry *e;
366 size_t name_len = strlen(name) + 1;
367 size_t format_len = 0;
368 u32 hash = jhash(name, name_len-1, 0);
371 format_len = strlen(format) + 1;
372 head = &marker_table[hash & ((1 << MARKER_HASH_BITS)-1)];
373 hlist_for_each_entry(e, node, head, hlist) {
374 if (!strcmp(name, e->name)) {
376 "Marker %s busy\n", name);
377 return ERR_PTR(-EBUSY); /* Already there */
381 * Using kmalloc here to allocate a variable length element. Could
382 * cause some memory fragmentation if overused.
384 e = kmalloc(sizeof(struct marker_entry) + name_len + format_len,
387 return ERR_PTR(-ENOMEM);
388 memcpy(&e->name[0], name, name_len);
390 e->format = &e->name[name_len];
391 memcpy(e->format, format, format_len);
392 if (strcmp(e->format, MARK_NOARGS) == 0)
393 e->call = marker_probe_cb_noarg;
395 e->call = marker_probe_cb;
396 trace_mark(core_marker_format, "name %s format %s",
400 e->call = marker_probe_cb;
402 e->single.func = __mark_empty_function;
403 e->single.probe_private = NULL;
407 hlist_add_head(&e->hlist, head);
412 * Remove the marker from the marker hash table. Must be called with mutex_lock
415 static int remove_marker(const char *name)
417 struct hlist_head *head;
418 struct hlist_node *node;
419 struct marker_entry *e;
421 size_t len = strlen(name) + 1;
422 u32 hash = jhash(name, len-1, 0);
424 head = &marker_table[hash & ((1 << MARKER_HASH_BITS)-1)];
425 hlist_for_each_entry(e, node, head, hlist) {
426 if (!strcmp(name, e->name)) {
433 if (e->single.func != __mark_empty_function)
435 hlist_del(&e->hlist);
441 * Set the mark_entry format to the format found in the element.
443 static int marker_set_format(struct marker_entry **entry, const char *format)
445 struct marker_entry *e;
446 size_t name_len = strlen((*entry)->name) + 1;
447 size_t format_len = strlen(format) + 1;
450 e = kmalloc(sizeof(struct marker_entry) + name_len + format_len,
454 memcpy(&e->name[0], (*entry)->name, name_len);
455 e->format = &e->name[name_len];
456 memcpy(e->format, format, format_len);
457 if (strcmp(e->format, MARK_NOARGS) == 0)
458 e->call = marker_probe_cb_noarg;
460 e->call = marker_probe_cb;
461 e->single = (*entry)->single;
462 e->multi = (*entry)->multi;
463 e->ptype = (*entry)->ptype;
464 e->refcount = (*entry)->refcount;
465 hlist_add_before(&e->hlist, &(*entry)->hlist);
466 hlist_del(&(*entry)->hlist);
469 trace_mark(core_marker_format, "name %s format %s",
475 * Sets the probe callback corresponding to one marker.
477 static int set_marker(struct marker_entry **entry, struct marker *elem,
481 WARN_ON(strcmp((*entry)->name, elem->name) != 0);
483 if ((*entry)->format) {
484 if (strcmp((*entry)->format, elem->format) != 0) {
486 "Format mismatch for probe %s "
487 "(%s), marker (%s)\n",
494 ret = marker_set_format(entry, elem->format);
500 * probe_cb setup (statically known) is done here. It is
501 * asynchronous with the rest of execution, therefore we only
502 * pass from a "safe" callback (with argument) to an "unsafe"
503 * callback (does not set arguments).
505 elem->call = (*entry)->call;
508 * We only update the single probe private data when the ptr is
509 * set to a _non_ single probe! (0 -> 1 and N -> 1, N != 1)
511 WARN_ON(elem->single.func != __mark_empty_function
512 && elem->single.probe_private
513 != (*entry)->single.probe_private &&
515 elem->single.probe_private = (*entry)->single.probe_private;
517 * Make sure the private data is valid when we update the
521 elem->single.func = (*entry)->single.func;
523 * We also make sure that the new probe callbacks array is consistent
524 * before setting a pointer to it.
526 rcu_assign_pointer(elem->multi, (*entry)->multi);
528 * Update the function or multi probe array pointer before setting the
532 elem->ptype = (*entry)->ptype;
533 elem->state = active;
539 * Disable a marker and its probe callback.
540 * Note: only waiting an RCU period after setting elem->call to the empty
541 * function insures that the original callback is not used anymore. This insured
542 * by rcu_read_lock_sched around the call site.
544 static void disable_marker(struct marker *elem)
546 /* leave "call" as is. It is known statically. */
548 elem->single.func = __mark_empty_function;
549 /* Update the function before setting the ptype */
551 elem->ptype = 0; /* single probe */
553 * Leave the private data and id there, because removal is racy and
554 * should be done only after an RCU period. These are never used until
555 * the next initialization anyway.
560 * marker_update_probe_range - Update a probe range
561 * @begin: beginning of the range
562 * @end: end of the range
564 * Updates the probe callback corresponding to a range of markers.
566 void marker_update_probe_range(struct marker *begin,
570 struct marker_entry *mark_entry;
572 mutex_lock(&markers_mutex);
573 for (iter = begin; iter < end; iter++) {
574 mark_entry = get_marker(iter->name);
576 set_marker(&mark_entry, iter,
577 !!mark_entry->refcount);
579 * ignore error, continue
582 disable_marker(iter);
585 mutex_unlock(&markers_mutex);
589 * Update probes, removing the faulty probes.
591 * Internal callback only changed before the first probe is connected to it.
592 * Single probe private data can only be changed on 0 -> 1 and 2 -> 1
593 * transitions. All other transitions will leave the old private data valid.
594 * This makes the non-atomicity of the callback/private data updates valid.
596 * "special case" updates :
601 * Other updates all behave the same, just like the 2 -> 3 or 3 -> 2 updates.
602 * Site effect : marker_set_format may delete the marker entry (creating a
605 static void marker_update_probes(void)
607 /* Core kernel markers */
608 marker_update_probe_range(__start___markers, __stop___markers);
609 /* Markers in modules. */
610 module_update_markers();
614 * marker_probe_register - Connect a probe to a marker
616 * @format: format string
617 * @probe: probe handler
618 * @probe_private: probe private data
620 * private data must be a valid allocated memory address, or NULL.
621 * Returns 0 if ok, error value on error.
622 * The probe address must at least be aligned on the architecture pointer size.
624 int marker_probe_register(const char *name, const char *format,
625 marker_probe_func *probe, void *probe_private)
627 struct marker_entry *entry;
629 struct marker_probe_closure *old;
631 mutex_lock(&markers_mutex);
632 entry = get_marker(name);
634 entry = add_marker(name, format);
636 ret = PTR_ERR(entry);
640 old = marker_entry_add_probe(entry, probe, probe_private);
645 mutex_unlock(&markers_mutex);
646 marker_update_probes(); /* may update entry */
649 mutex_lock(&markers_mutex);
650 entry = get_marker(name);
653 mutex_unlock(&markers_mutex);
656 EXPORT_SYMBOL_GPL(marker_probe_register);
659 * marker_probe_unregister - Disconnect a probe from a marker
661 * @probe: probe function pointer
662 * @probe_private: probe private data
664 * Returns the private data given to marker_probe_register, or an ERR_PTR().
665 * We do not need to call a synchronize_sched to make sure the probes have
666 * finished running before doing a module unload, because the module unload
667 * itself uses stop_machine(), which insures that every preempt disabled section
670 int marker_probe_unregister(const char *name,
671 marker_probe_func *probe, void *probe_private)
673 struct marker_entry *entry;
674 struct marker_probe_closure *old;
677 mutex_lock(&markers_mutex);
678 entry = get_marker(name);
681 old = marker_entry_remove_probe(entry, probe, probe_private);
682 mutex_unlock(&markers_mutex);
683 marker_update_probes(); /* may update entry */
686 mutex_lock(&markers_mutex);
687 entry = get_marker(name);
690 remove_marker(name); /* Ignore busy error message */
693 mutex_unlock(&markers_mutex);
696 EXPORT_SYMBOL_GPL(marker_probe_unregister);
698 static struct marker_entry *
699 get_marker_from_private_data(marker_probe_func *probe, void *probe_private)
701 struct marker_entry *entry;
703 struct hlist_head *head;
704 struct hlist_node *node;
706 for (i = 0; i < MARKER_TABLE_SIZE; i++) {
707 head = &marker_table[i];
708 hlist_for_each_entry(entry, node, head, hlist) {
710 if (entry->single.func == probe
711 && entry->single.probe_private
715 struct marker_probe_closure *closure;
716 closure = entry->multi;
717 for (i = 0; closure[i].func; i++) {
718 if (closure[i].func == probe &&
719 closure[i].probe_private
730 * marker_probe_unregister_private_data - Disconnect a probe from a marker
731 * @probe: probe function
732 * @probe_private: probe private data
734 * Unregister a probe by providing the registered private data.
735 * Only removes the first marker found in hash table.
736 * Return 0 on success or error value.
737 * We do not need to call a synchronize_sched to make sure the probes have
738 * finished running before doing a module unload, because the module unload
739 * itself uses stop_machine(), which insures that every preempt disabled section
742 int marker_probe_unregister_private_data(marker_probe_func *probe,
745 struct marker_entry *entry;
747 struct marker_probe_closure *old;
749 mutex_lock(&markers_mutex);
750 entry = get_marker_from_private_data(probe, probe_private);
755 old = marker_entry_remove_probe(entry, NULL, probe_private);
756 mutex_unlock(&markers_mutex);
757 marker_update_probes(); /* may update entry */
760 mutex_lock(&markers_mutex);
761 entry = get_marker_from_private_data(probe, probe_private);
763 remove_marker(entry->name); /* Ignore busy error message */
765 mutex_unlock(&markers_mutex);
768 EXPORT_SYMBOL_GPL(marker_probe_unregister_private_data);
771 * marker_get_private_data - Get a marker's probe private data
773 * @probe: probe to match
774 * @num: get the nth matching probe's private data
776 * Returns the nth private data pointer (starting from 0) matching, or an
778 * Returns the private data pointer, or an ERR_PTR.
779 * The private data pointer should _only_ be dereferenced if the caller is the
780 * owner of the data, or its content could vanish. This is mostly used to
781 * confirm that a caller is the owner of a registered probe.
783 void *marker_get_private_data(const char *name, marker_probe_func *probe,
786 struct hlist_head *head;
787 struct hlist_node *node;
788 struct marker_entry *e;
789 size_t name_len = strlen(name) + 1;
790 u32 hash = jhash(name, name_len-1, 0);
793 head = &marker_table[hash & ((1 << MARKER_HASH_BITS)-1)];
794 hlist_for_each_entry(e, node, head, hlist) {
795 if (!strcmp(name, e->name)) {
797 if (num == 0 && e->single.func == probe)
798 return e->single.probe_private;
802 struct marker_probe_closure *closure;
805 for (i = 0; closure[i].func; i++) {
806 if (closure[i].func != probe)
809 return closure[i].probe_private;
814 return ERR_PTR(-ENOENT);
816 EXPORT_SYMBOL_GPL(marker_get_private_data);