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. */
66 unsigned char ptype:1;
67 unsigned char format_allocated:1;
68 char name[0]; /* Contains name'\0'format'\0' */
71 static struct hlist_head marker_table[MARKER_TABLE_SIZE];
74 * __mark_empty_function - Empty probe callback
75 * @probe_private: probe private data
76 * @call_private: call site private data
78 * @...: variable argument list
80 * Empty callback provided as a probe to the markers. By providing this to a
81 * disabled marker, we make sure the execution flow is always valid even
82 * though the function pointer change and the marker enabling are two distinct
83 * operations that modifies the execution flow of preemptible code.
85 void __mark_empty_function(void *probe_private, void *call_private,
86 const char *fmt, va_list *args)
89 EXPORT_SYMBOL_GPL(__mark_empty_function);
92 * marker_probe_cb Callback that prepares the variable argument list for probes.
93 * @mdata: pointer of type struct marker
94 * @call_private: caller site private data
95 * @...: Variable argument list.
97 * Since we do not use "typical" pointer based RCU in the 1 argument case, we
98 * need to put a full smp_rmb() in this branch. This is why we do not use
99 * rcu_dereference() for the pointer read.
101 void marker_probe_cb(const struct marker *mdata, void *call_private, ...)
107 * rcu_read_lock_sched does two things : disabling preemption to make
108 * sure the teardown of the callbacks can be done correctly when they
109 * are in modules and they insure RCU read coherency.
111 rcu_read_lock_sched();
112 ptype = mdata->ptype;
113 if (likely(!ptype)) {
114 marker_probe_func *func;
115 /* Must read the ptype before ptr. They are not data dependant,
116 * so we put an explicit smp_rmb() here. */
118 func = mdata->single.func;
119 /* Must read the ptr before private data. They are not data
120 * dependant, so we put an explicit smp_rmb() here. */
122 va_start(args, call_private);
123 func(mdata->single.probe_private, call_private, mdata->format,
127 struct marker_probe_closure *multi;
130 * Read mdata->ptype before mdata->multi.
133 multi = mdata->multi;
135 * multi points to an array, therefore accessing the array
136 * depends on reading multi. However, even in this case,
137 * we must insure that the pointer is read _before_ the array
138 * data. Same as rcu_dereference, but we need a full smp_rmb()
139 * in the fast path, so put the explicit barrier here.
141 smp_read_barrier_depends();
142 for (i = 0; multi[i].func; i++) {
143 va_start(args, call_private);
144 multi[i].func(multi[i].probe_private, call_private,
145 mdata->format, &args);
149 rcu_read_unlock_sched();
151 EXPORT_SYMBOL_GPL(marker_probe_cb);
154 * marker_probe_cb Callback that does not prepare the variable argument list.
155 * @mdata: pointer of type struct marker
156 * @call_private: caller site private data
157 * @...: Variable argument list.
159 * Should be connected to markers "MARK_NOARGS".
161 void marker_probe_cb_noarg(const struct marker *mdata, void *call_private, ...)
163 va_list args; /* not initialized */
166 rcu_read_lock_sched();
167 ptype = mdata->ptype;
168 if (likely(!ptype)) {
169 marker_probe_func *func;
170 /* Must read the ptype before ptr. They are not data dependant,
171 * so we put an explicit smp_rmb() here. */
173 func = mdata->single.func;
174 /* Must read the ptr before private data. They are not data
175 * dependant, so we put an explicit smp_rmb() here. */
177 func(mdata->single.probe_private, call_private, mdata->format,
180 struct marker_probe_closure *multi;
183 * Read mdata->ptype before mdata->multi.
186 multi = mdata->multi;
188 * multi points to an array, therefore accessing the array
189 * depends on reading multi. However, even in this case,
190 * we must insure that the pointer is read _before_ the array
191 * data. Same as rcu_dereference, but we need a full smp_rmb()
192 * in the fast path, so put the explicit barrier here.
194 smp_read_barrier_depends();
195 for (i = 0; multi[i].func; i++)
196 multi[i].func(multi[i].probe_private, call_private,
197 mdata->format, &args);
199 rcu_read_unlock_sched();
201 EXPORT_SYMBOL_GPL(marker_probe_cb_noarg);
203 static void free_old_closure(struct rcu_head *head)
205 struct marker_entry *entry = container_of(head,
206 struct marker_entry, rcu);
207 kfree(entry->oldptr);
208 /* Make sure we free the data before setting the pending flag to 0 */
210 entry->rcu_pending = 0;
213 static void debug_print_probes(struct marker_entry *entry)
221 printk(KERN_DEBUG "Single probe : %p %p\n",
223 entry->single.probe_private);
225 for (i = 0; entry->multi[i].func; i++)
226 printk(KERN_DEBUG "Multi probe %d : %p %p\n", i,
227 entry->multi[i].func,
228 entry->multi[i].probe_private);
232 static struct marker_probe_closure *
233 marker_entry_add_probe(struct marker_entry *entry,
234 marker_probe_func *probe, void *probe_private)
237 struct marker_probe_closure *old, *new;
241 debug_print_probes(entry);
244 if (entry->single.func == probe &&
245 entry->single.probe_private == probe_private)
246 return ERR_PTR(-EBUSY);
247 if (entry->single.func == __mark_empty_function) {
249 entry->single.func = probe;
250 entry->single.probe_private = probe_private;
253 debug_print_probes(entry);
261 /* (N -> N+1), (N != 0, 1) probes */
262 for (nr_probes = 0; old[nr_probes].func; nr_probes++)
263 if (old[nr_probes].func == probe
264 && old[nr_probes].probe_private
266 return ERR_PTR(-EBUSY);
268 /* + 2 : one for new probe, one for NULL func */
269 new = kzalloc((nr_probes + 2) * sizeof(struct marker_probe_closure),
272 return ERR_PTR(-ENOMEM);
274 new[0] = entry->single;
277 nr_probes * sizeof(struct marker_probe_closure));
278 new[nr_probes].func = probe;
279 new[nr_probes].probe_private = probe_private;
280 entry->refcount = nr_probes + 1;
283 debug_print_probes(entry);
287 static struct marker_probe_closure *
288 marker_entry_remove_probe(struct marker_entry *entry,
289 marker_probe_func *probe, void *probe_private)
291 int nr_probes = 0, nr_del = 0, i;
292 struct marker_probe_closure *old, *new;
296 debug_print_probes(entry);
298 /* 0 -> N is an error */
299 WARN_ON(entry->single.func == __mark_empty_function);
301 WARN_ON(probe && entry->single.func != probe);
302 WARN_ON(entry->single.probe_private != probe_private);
303 entry->single.func = __mark_empty_function;
306 debug_print_probes(entry);
309 /* (N -> M), (N > 1, M >= 0) probes */
310 for (nr_probes = 0; old[nr_probes].func; nr_probes++) {
311 if ((!probe || old[nr_probes].func == probe)
312 && old[nr_probes].probe_private
318 if (nr_probes - nr_del == 0) {
319 /* N -> 0, (N > 1) */
320 entry->single.func = __mark_empty_function;
323 } else if (nr_probes - nr_del == 1) {
324 /* N -> 1, (N > 1) */
325 for (i = 0; old[i].func; i++)
326 if ((probe && old[i].func != probe) ||
327 old[i].probe_private != probe_private)
328 entry->single = old[i];
333 /* N -> M, (N > 1, M > 1) */
335 new = kzalloc((nr_probes - nr_del + 1)
336 * sizeof(struct marker_probe_closure), GFP_KERNEL);
338 return ERR_PTR(-ENOMEM);
339 for (i = 0; old[i].func; i++)
340 if ((probe && old[i].func != probe) ||
341 old[i].probe_private != probe_private)
343 entry->refcount = nr_probes - nr_del;
347 debug_print_probes(entry);
352 * Get marker if the marker is present in the marker hash table.
353 * Must be called with markers_mutex held.
354 * Returns NULL if not present.
356 static struct marker_entry *get_marker(const char *name)
358 struct hlist_head *head;
359 struct hlist_node *node;
360 struct marker_entry *e;
361 u32 hash = jhash(name, strlen(name), 0);
363 head = &marker_table[hash & ((1 << MARKER_HASH_BITS)-1)];
364 hlist_for_each_entry(e, node, head, hlist) {
365 if (!strcmp(name, e->name))
372 * Add the marker to the marker hash table. Must be called with markers_mutex
375 static struct marker_entry *add_marker(const char *name, const char *format)
377 struct hlist_head *head;
378 struct hlist_node *node;
379 struct marker_entry *e;
380 size_t name_len = strlen(name) + 1;
381 size_t format_len = 0;
382 u32 hash = jhash(name, name_len-1, 0);
385 format_len = strlen(format) + 1;
386 head = &marker_table[hash & ((1 << MARKER_HASH_BITS)-1)];
387 hlist_for_each_entry(e, node, head, hlist) {
388 if (!strcmp(name, e->name)) {
390 "Marker %s busy\n", name);
391 return ERR_PTR(-EBUSY); /* Already there */
395 * Using kmalloc here to allocate a variable length element. Could
396 * cause some memory fragmentation if overused.
398 e = kmalloc(sizeof(struct marker_entry) + name_len + format_len,
401 return ERR_PTR(-ENOMEM);
402 memcpy(&e->name[0], name, name_len);
404 e->format = &e->name[name_len];
405 memcpy(e->format, format, format_len);
406 if (strcmp(e->format, MARK_NOARGS) == 0)
407 e->call = marker_probe_cb_noarg;
409 e->call = marker_probe_cb;
410 trace_mark(core_marker_format, "name %s format %s",
414 e->call = marker_probe_cb;
416 e->single.func = __mark_empty_function;
417 e->single.probe_private = NULL;
420 e->format_allocated = 0;
423 hlist_add_head(&e->hlist, head);
428 * Remove the marker from the marker hash table. Must be called with mutex_lock
431 static int remove_marker(const char *name)
433 struct hlist_head *head;
434 struct hlist_node *node;
435 struct marker_entry *e;
437 size_t len = strlen(name) + 1;
438 u32 hash = jhash(name, len-1, 0);
440 head = &marker_table[hash & ((1 << MARKER_HASH_BITS)-1)];
441 hlist_for_each_entry(e, node, head, hlist) {
442 if (!strcmp(name, e->name)) {
449 if (e->single.func != __mark_empty_function)
451 hlist_del(&e->hlist);
452 if (e->format_allocated)
454 /* Make sure the call_rcu has been executed */
462 * Set the mark_entry format to the format found in the element.
464 static int marker_set_format(struct marker_entry *entry, const char *format)
466 entry->format = kstrdup(format, GFP_KERNEL);
469 entry->format_allocated = 1;
471 trace_mark(core_marker_format, "name %s format %s",
472 entry->name, entry->format);
477 * Sets the probe callback corresponding to one marker.
479 static int set_marker(struct marker_entry *entry, struct marker *elem,
483 WARN_ON(strcmp(entry->name, elem->name) != 0);
486 if (strcmp(entry->format, elem->format) != 0) {
488 "Format mismatch for probe %s "
489 "(%s), marker (%s)\n",
496 ret = marker_set_format(entry, elem->format);
502 * probe_cb setup (statically known) is done here. It is
503 * asynchronous with the rest of execution, therefore we only
504 * pass from a "safe" callback (with argument) to an "unsafe"
505 * callback (does not set arguments).
507 elem->call = entry->call;
510 * We only update the single probe private data when the ptr is
511 * set to a _non_ single probe! (0 -> 1 and N -> 1, N != 1)
513 WARN_ON(elem->single.func != __mark_empty_function
514 && elem->single.probe_private != entry->single.probe_private
516 elem->single.probe_private = entry->single.probe_private;
518 * Make sure the private data is valid when we update the
522 elem->single.func = entry->single.func;
524 * We also make sure that the new probe callbacks array is consistent
525 * before setting a pointer to it.
527 rcu_assign_pointer(elem->multi, entry->multi);
529 * Update the function or multi probe array pointer before setting the
533 elem->ptype = entry->ptype;
534 elem->state = active;
540 * Disable a marker and its probe callback.
541 * Note: only waiting an RCU period after setting elem->call to the empty
542 * function insures that the original callback is not used anymore. This insured
543 * by rcu_read_lock_sched around the call site.
545 static void disable_marker(struct marker *elem)
547 /* leave "call" as is. It is known statically. */
549 elem->single.func = __mark_empty_function;
550 /* Update the function before setting the ptype */
552 elem->ptype = 0; /* single probe */
554 * Leave the private data and id there, because removal is racy and
555 * should be done only after an RCU period. These are never used until
556 * the next initialization anyway.
561 * marker_update_probe_range - Update a probe range
562 * @begin: beginning of the range
563 * @end: end of the range
565 * Updates the probe callback corresponding to a range of markers.
567 void marker_update_probe_range(struct marker *begin,
571 struct marker_entry *mark_entry;
573 mutex_lock(&markers_mutex);
574 for (iter = begin; iter < end; iter++) {
575 mark_entry = get_marker(iter->name);
577 set_marker(mark_entry, iter, !!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);
639 ret = marker_set_format(entry, format);
640 else if (strcmp(entry->format, format))
647 * If we detect that a call_rcu is pending for this marker,
648 * make sure it's executed now.
650 if (entry->rcu_pending)
652 old = marker_entry_add_probe(entry, probe, probe_private);
657 mutex_unlock(&markers_mutex);
658 marker_update_probes(); /* may update entry */
659 mutex_lock(&markers_mutex);
660 entry = get_marker(name);
662 if (entry->rcu_pending)
665 entry->rcu_pending = 1;
666 /* write rcu_pending before calling the RCU callback */
668 call_rcu_sched(&entry->rcu, free_old_closure);
670 mutex_unlock(&markers_mutex);
673 EXPORT_SYMBOL_GPL(marker_probe_register);
676 * marker_probe_unregister - Disconnect a probe from a marker
678 * @probe: probe function pointer
679 * @probe_private: probe private data
681 * Returns the private data given to marker_probe_register, or an ERR_PTR().
682 * We do not need to call a synchronize_sched to make sure the probes have
683 * finished running before doing a module unload, because the module unload
684 * itself uses stop_machine(), which insures that every preempt disabled section
687 int marker_probe_unregister(const char *name,
688 marker_probe_func *probe, void *probe_private)
690 struct marker_entry *entry;
691 struct marker_probe_closure *old;
694 mutex_lock(&markers_mutex);
695 entry = get_marker(name);
698 if (entry->rcu_pending)
700 old = marker_entry_remove_probe(entry, probe, probe_private);
701 mutex_unlock(&markers_mutex);
702 marker_update_probes(); /* may update entry */
703 mutex_lock(&markers_mutex);
704 entry = get_marker(name);
707 if (entry->rcu_pending)
710 entry->rcu_pending = 1;
711 /* write rcu_pending before calling the RCU callback */
713 call_rcu_sched(&entry->rcu, free_old_closure);
714 remove_marker(name); /* Ignore busy error message */
717 mutex_unlock(&markers_mutex);
720 EXPORT_SYMBOL_GPL(marker_probe_unregister);
722 static struct marker_entry *
723 get_marker_from_private_data(marker_probe_func *probe, void *probe_private)
725 struct marker_entry *entry;
727 struct hlist_head *head;
728 struct hlist_node *node;
730 for (i = 0; i < MARKER_TABLE_SIZE; i++) {
731 head = &marker_table[i];
732 hlist_for_each_entry(entry, node, head, hlist) {
734 if (entry->single.func == probe
735 && entry->single.probe_private
739 struct marker_probe_closure *closure;
740 closure = entry->multi;
741 for (i = 0; closure[i].func; i++) {
742 if (closure[i].func == probe &&
743 closure[i].probe_private
754 * marker_probe_unregister_private_data - Disconnect a probe from a marker
755 * @probe: probe function
756 * @probe_private: probe private data
758 * Unregister a probe by providing the registered private data.
759 * Only removes the first marker found in hash table.
760 * Return 0 on success or error value.
761 * We do not need to call a synchronize_sched to make sure the probes have
762 * finished running before doing a module unload, because the module unload
763 * itself uses stop_machine(), which insures that every preempt disabled section
766 int marker_probe_unregister_private_data(marker_probe_func *probe,
769 struct marker_entry *entry;
771 struct marker_probe_closure *old;
773 mutex_lock(&markers_mutex);
774 entry = get_marker_from_private_data(probe, probe_private);
779 if (entry->rcu_pending)
781 old = marker_entry_remove_probe(entry, NULL, probe_private);
782 mutex_unlock(&markers_mutex);
783 marker_update_probes(); /* may update entry */
784 mutex_lock(&markers_mutex);
785 entry = get_marker_from_private_data(probe, probe_private);
787 if (entry->rcu_pending)
790 entry->rcu_pending = 1;
791 /* write rcu_pending before calling the RCU callback */
793 call_rcu_sched(&entry->rcu, free_old_closure);
794 remove_marker(entry->name); /* Ignore busy error message */
796 mutex_unlock(&markers_mutex);
799 EXPORT_SYMBOL_GPL(marker_probe_unregister_private_data);
802 * marker_get_private_data - Get a marker's probe private data
804 * @probe: probe to match
805 * @num: get the nth matching probe's private data
807 * Returns the nth private data pointer (starting from 0) matching, or an
809 * Returns the private data pointer, or an ERR_PTR.
810 * The private data pointer should _only_ be dereferenced if the caller is the
811 * owner of the data, or its content could vanish. This is mostly used to
812 * confirm that a caller is the owner of a registered probe.
814 void *marker_get_private_data(const char *name, marker_probe_func *probe,
817 struct hlist_head *head;
818 struct hlist_node *node;
819 struct marker_entry *e;
820 size_t name_len = strlen(name) + 1;
821 u32 hash = jhash(name, name_len-1, 0);
824 head = &marker_table[hash & ((1 << MARKER_HASH_BITS)-1)];
825 hlist_for_each_entry(e, node, head, hlist) {
826 if (!strcmp(name, e->name)) {
828 if (num == 0 && e->single.func == probe)
829 return e->single.probe_private;
833 struct marker_probe_closure *closure;
836 for (i = 0; closure[i].func; i++) {
837 if (closure[i].func != probe)
840 return closure[i].probe_private;
845 return ERR_PTR(-ENOENT);
847 EXPORT_SYMBOL_GPL(marker_get_private_data);