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. */
65 unsigned char rcu_pending:1;
66 unsigned char ptype:1;
67 char name[0]; /* Contains name'\0'format'\0' */
70 static struct hlist_head marker_table[MARKER_TABLE_SIZE];
73 * __mark_empty_function - Empty probe callback
74 * @probe_private: probe private data
75 * @call_private: call site private data
77 * @...: variable argument list
79 * Empty callback provided as a probe to the markers. By providing this to a
80 * disabled marker, we make sure the execution flow is always valid even
81 * though the function pointer change and the marker enabling are two distinct
82 * operations that modifies the execution flow of preemptible code.
84 void __mark_empty_function(void *probe_private, void *call_private,
85 const char *fmt, va_list *args)
88 EXPORT_SYMBOL_GPL(__mark_empty_function);
91 * marker_probe_cb Callback that prepares the variable argument list for probes.
92 * @mdata: pointer of type struct marker
93 * @call_private: caller site private data
94 * @...: Variable argument list.
96 * Since we do not use "typical" pointer based RCU in the 1 argument case, we
97 * need to put a full smp_rmb() in this branch. This is why we do not use
98 * rcu_dereference() for the pointer read.
100 void marker_probe_cb(const struct marker *mdata, void *call_private, ...)
106 * preempt_disable does two things : disabling preemption to make sure
107 * the teardown of the callbacks can be done correctly when they are in
108 * modules and they insure RCU read coherency.
111 ptype = mdata->ptype;
112 if (likely(!ptype)) {
113 marker_probe_func *func;
114 /* Must read the ptype before ptr. They are not data dependant,
115 * so we put an explicit smp_rmb() here. */
117 func = mdata->single.func;
118 /* Must read the ptr before private data. They are not data
119 * dependant, so we put an explicit smp_rmb() here. */
121 va_start(args, call_private);
122 func(mdata->single.probe_private, call_private, mdata->format,
126 struct marker_probe_closure *multi;
129 * multi points to an array, therefore accessing the array
130 * depends on reading multi. However, even in this case,
131 * we must insure that the pointer is read _before_ the array
132 * data. Same as rcu_dereference, but we need a full smp_rmb()
133 * in the fast path, so put the explicit barrier here.
135 smp_read_barrier_depends();
136 multi = mdata->multi;
137 for (i = 0; multi[i].func; i++) {
138 va_start(args, call_private);
139 multi[i].func(multi[i].probe_private, call_private,
140 mdata->format, &args);
146 EXPORT_SYMBOL_GPL(marker_probe_cb);
149 * marker_probe_cb Callback that does not prepare the variable argument list.
150 * @mdata: pointer of type struct marker
151 * @call_private: caller site private data
152 * @...: Variable argument list.
154 * Should be connected to markers "MARK_NOARGS".
156 void marker_probe_cb_noarg(const struct marker *mdata, void *call_private, ...)
158 va_list args; /* not initialized */
162 ptype = mdata->ptype;
163 if (likely(!ptype)) {
164 marker_probe_func *func;
165 /* Must read the ptype before ptr. They are not data dependant,
166 * so we put an explicit smp_rmb() here. */
168 func = mdata->single.func;
169 /* Must read the ptr before private data. They are not data
170 * dependant, so we put an explicit smp_rmb() here. */
172 func(mdata->single.probe_private, call_private, mdata->format,
175 struct marker_probe_closure *multi;
178 * multi points to an array, therefore accessing the array
179 * depends on reading multi. However, even in this case,
180 * we must insure that the pointer is read _before_ the array
181 * data. Same as rcu_dereference, but we need a full smp_rmb()
182 * in the fast path, so put the explicit barrier here.
184 smp_read_barrier_depends();
185 multi = mdata->multi;
186 for (i = 0; multi[i].func; i++)
187 multi[i].func(multi[i].probe_private, call_private,
188 mdata->format, &args);
192 EXPORT_SYMBOL_GPL(marker_probe_cb_noarg);
194 static void free_old_closure(struct rcu_head *head)
196 struct marker_entry *entry = container_of(head,
197 struct marker_entry, rcu);
198 kfree(entry->oldptr);
199 /* Make sure we free the data before setting the pending flag to 0 */
201 entry->rcu_pending = 0;
204 static void debug_print_probes(struct marker_entry *entry)
212 printk(KERN_DEBUG "Single probe : %p %p\n",
214 entry->single.probe_private);
216 for (i = 0; entry->multi[i].func; i++)
217 printk(KERN_DEBUG "Multi probe %d : %p %p\n", i,
218 entry->multi[i].func,
219 entry->multi[i].probe_private);
223 static struct marker_probe_closure *
224 marker_entry_add_probe(struct marker_entry *entry,
225 marker_probe_func *probe, void *probe_private)
228 struct marker_probe_closure *old, *new;
232 debug_print_probes(entry);
235 if (entry->single.func == probe &&
236 entry->single.probe_private == probe_private)
237 return ERR_PTR(-EBUSY);
238 if (entry->single.func == __mark_empty_function) {
240 entry->single.func = probe;
241 entry->single.probe_private = probe_private;
244 debug_print_probes(entry);
252 /* (N -> N+1), (N != 0, 1) probes */
253 for (nr_probes = 0; old[nr_probes].func; nr_probes++)
254 if (old[nr_probes].func == probe
255 && old[nr_probes].probe_private
257 return ERR_PTR(-EBUSY);
259 /* + 2 : one for new probe, one for NULL func */
260 new = kzalloc((nr_probes + 2) * sizeof(struct marker_probe_closure),
263 return ERR_PTR(-ENOMEM);
265 new[0] = entry->single;
268 nr_probes * sizeof(struct marker_probe_closure));
269 new[nr_probes].func = probe;
270 new[nr_probes].probe_private = probe_private;
271 entry->refcount = nr_probes + 1;
274 debug_print_probes(entry);
278 static struct marker_probe_closure *
279 marker_entry_remove_probe(struct marker_entry *entry,
280 marker_probe_func *probe, void *probe_private)
282 int nr_probes = 0, nr_del = 0, i;
283 struct marker_probe_closure *old, *new;
287 debug_print_probes(entry);
289 /* 0 -> N is an error */
290 WARN_ON(entry->single.func == __mark_empty_function);
292 WARN_ON(probe && entry->single.func != probe);
293 WARN_ON(entry->single.probe_private != probe_private);
294 entry->single.func = __mark_empty_function;
297 debug_print_probes(entry);
300 /* (N -> M), (N > 1, M >= 0) probes */
301 for (nr_probes = 0; old[nr_probes].func; nr_probes++) {
302 if ((!probe || old[nr_probes].func == probe)
303 && old[nr_probes].probe_private
309 if (nr_probes - nr_del == 0) {
310 /* N -> 0, (N > 1) */
311 entry->single.func = __mark_empty_function;
314 } else if (nr_probes - nr_del == 1) {
315 /* N -> 1, (N > 1) */
316 for (i = 0; old[i].func; i++)
317 if ((probe && old[i].func != probe) ||
318 old[i].probe_private != probe_private)
319 entry->single = old[i];
324 /* N -> M, (N > 1, M > 1) */
326 new = kzalloc((nr_probes - nr_del + 1)
327 * sizeof(struct marker_probe_closure), GFP_KERNEL);
329 return ERR_PTR(-ENOMEM);
330 for (i = 0; old[i].func; i++)
331 if ((probe && old[i].func != probe) ||
332 old[i].probe_private != probe_private)
334 entry->refcount = nr_probes - nr_del;
338 debug_print_probes(entry);
343 * Get marker if the marker is present in the marker hash table.
344 * Must be called with markers_mutex held.
345 * Returns NULL if not present.
347 static struct marker_entry *get_marker(const char *name)
349 struct hlist_head *head;
350 struct hlist_node *node;
351 struct marker_entry *e;
352 u32 hash = jhash(name, strlen(name), 0);
354 head = &marker_table[hash & ((1 << MARKER_HASH_BITS)-1)];
355 hlist_for_each_entry(e, node, head, hlist) {
356 if (!strcmp(name, e->name))
363 * Add the marker to the marker hash table. Must be called with markers_mutex
366 static struct marker_entry *add_marker(const char *name, const char *format)
368 struct hlist_head *head;
369 struct hlist_node *node;
370 struct marker_entry *e;
371 size_t name_len = strlen(name) + 1;
372 size_t format_len = 0;
373 u32 hash = jhash(name, name_len-1, 0);
376 format_len = strlen(format) + 1;
377 head = &marker_table[hash & ((1 << MARKER_HASH_BITS)-1)];
378 hlist_for_each_entry(e, node, head, hlist) {
379 if (!strcmp(name, e->name)) {
381 "Marker %s busy\n", name);
382 return ERR_PTR(-EBUSY); /* Already there */
386 * Using kmalloc here to allocate a variable length element. Could
387 * cause some memory fragmentation if overused.
389 e = kmalloc(sizeof(struct marker_entry) + name_len + format_len,
392 return ERR_PTR(-ENOMEM);
393 memcpy(&e->name[0], name, name_len);
395 e->format = &e->name[name_len];
396 memcpy(e->format, format, format_len);
397 if (strcmp(e->format, MARK_NOARGS) == 0)
398 e->call = marker_probe_cb_noarg;
400 e->call = marker_probe_cb;
401 trace_mark(core_marker_format, "name %s format %s",
405 e->call = marker_probe_cb;
407 e->single.func = __mark_empty_function;
408 e->single.probe_private = NULL;
413 hlist_add_head(&e->hlist, head);
418 * Remove the marker from the marker hash table. Must be called with mutex_lock
421 static int remove_marker(const char *name)
423 struct hlist_head *head;
424 struct hlist_node *node;
425 struct marker_entry *e;
427 size_t len = strlen(name) + 1;
428 u32 hash = jhash(name, len-1, 0);
430 head = &marker_table[hash & ((1 << MARKER_HASH_BITS)-1)];
431 hlist_for_each_entry(e, node, head, hlist) {
432 if (!strcmp(name, e->name)) {
439 if (e->single.func != __mark_empty_function)
441 hlist_del(&e->hlist);
442 /* Make sure the call_rcu has been executed */
450 * Set the mark_entry format to the format found in the element.
452 static int marker_set_format(struct marker_entry **entry, const char *format)
454 struct marker_entry *e;
455 size_t name_len = strlen((*entry)->name) + 1;
456 size_t format_len = strlen(format) + 1;
459 e = kmalloc(sizeof(struct marker_entry) + name_len + format_len,
463 memcpy(&e->name[0], (*entry)->name, name_len);
464 e->format = &e->name[name_len];
465 memcpy(e->format, format, format_len);
466 if (strcmp(e->format, MARK_NOARGS) == 0)
467 e->call = marker_probe_cb_noarg;
469 e->call = marker_probe_cb;
470 e->single = (*entry)->single;
471 e->multi = (*entry)->multi;
472 e->ptype = (*entry)->ptype;
473 e->refcount = (*entry)->refcount;
475 hlist_add_before(&e->hlist, &(*entry)->hlist);
476 hlist_del(&(*entry)->hlist);
477 /* Make sure the call_rcu has been executed */
478 if ((*entry)->rcu_pending)
482 trace_mark(core_marker_format, "name %s format %s",
488 * Sets the probe callback corresponding to one marker.
490 static int set_marker(struct marker_entry **entry, struct marker *elem,
494 WARN_ON(strcmp((*entry)->name, elem->name) != 0);
496 if ((*entry)->format) {
497 if (strcmp((*entry)->format, elem->format) != 0) {
499 "Format mismatch for probe %s "
500 "(%s), marker (%s)\n",
507 ret = marker_set_format(entry, elem->format);
513 * probe_cb setup (statically known) is done here. It is
514 * asynchronous with the rest of execution, therefore we only
515 * pass from a "safe" callback (with argument) to an "unsafe"
516 * callback (does not set arguments).
518 elem->call = (*entry)->call;
521 * We only update the single probe private data when the ptr is
522 * set to a _non_ single probe! (0 -> 1 and N -> 1, N != 1)
524 WARN_ON(elem->single.func != __mark_empty_function
525 && elem->single.probe_private
526 != (*entry)->single.probe_private &&
528 elem->single.probe_private = (*entry)->single.probe_private;
530 * Make sure the private data is valid when we update the
534 elem->single.func = (*entry)->single.func;
536 * We also make sure that the new probe callbacks array is consistent
537 * before setting a pointer to it.
539 rcu_assign_pointer(elem->multi, (*entry)->multi);
541 * Update the function or multi probe array pointer before setting the
545 elem->ptype = (*entry)->ptype;
546 elem->state = active;
552 * Disable a marker and its probe callback.
553 * Note: only waiting an RCU period after setting elem->call to the empty
554 * function insures that the original callback is not used anymore. This insured
555 * by preempt_disable around the call site.
557 static void disable_marker(struct marker *elem)
559 /* leave "call" as is. It is known statically. */
561 elem->single.func = __mark_empty_function;
562 /* Update the function before setting the ptype */
564 elem->ptype = 0; /* single probe */
566 * Leave the private data and id there, because removal is racy and
567 * should be done only after an RCU period. These are never used until
568 * the next initialization anyway.
573 * marker_update_probe_range - Update a probe range
574 * @begin: beginning of the range
575 * @end: end of the range
577 * Updates the probe callback corresponding to a range of markers.
579 void marker_update_probe_range(struct marker *begin,
583 struct marker_entry *mark_entry;
585 mutex_lock(&markers_mutex);
586 for (iter = begin; iter < end; iter++) {
587 mark_entry = get_marker(iter->name);
589 set_marker(&mark_entry, iter,
590 !!mark_entry->refcount);
592 * ignore error, continue
595 disable_marker(iter);
598 mutex_unlock(&markers_mutex);
602 * Update probes, removing the faulty probes.
604 * Internal callback only changed before the first probe is connected to it.
605 * Single probe private data can only be changed on 0 -> 1 and 2 -> 1
606 * transitions. All other transitions will leave the old private data valid.
607 * This makes the non-atomicity of the callback/private data updates valid.
609 * "special case" updates :
614 * Other updates all behave the same, just like the 2 -> 3 or 3 -> 2 updates.
615 * Site effect : marker_set_format may delete the marker entry (creating a
618 static void marker_update_probes(void)
620 /* Core kernel markers */
621 marker_update_probe_range(__start___markers, __stop___markers);
622 /* Markers in modules. */
623 module_update_markers();
627 * marker_probe_register - Connect a probe to a marker
629 * @format: format string
630 * @probe: probe handler
631 * @probe_private: probe private data
633 * private data must be a valid allocated memory address, or NULL.
634 * Returns 0 if ok, error value on error.
635 * The probe address must at least be aligned on the architecture pointer size.
637 int marker_probe_register(const char *name, const char *format,
638 marker_probe_func *probe, void *probe_private)
640 struct marker_entry *entry;
642 struct marker_probe_closure *old;
644 mutex_lock(&markers_mutex);
645 entry = get_marker(name);
647 entry = add_marker(name, format);
649 ret = PTR_ERR(entry);
654 * If we detect that a call_rcu is pending for this marker,
655 * make sure it's executed now.
657 if (entry->rcu_pending)
659 old = marker_entry_add_probe(entry, probe, probe_private);
664 mutex_unlock(&markers_mutex);
665 marker_update_probes(); /* may update entry */
666 mutex_lock(&markers_mutex);
667 entry = get_marker(name);
670 entry->rcu_pending = 1;
671 /* write rcu_pending before calling the RCU callback */
673 #ifdef CONFIG_PREEMPT_RCU
674 synchronize_sched(); /* Until we have the call_rcu_sched() */
676 call_rcu(&entry->rcu, free_old_closure);
678 mutex_unlock(&markers_mutex);
681 EXPORT_SYMBOL_GPL(marker_probe_register);
684 * marker_probe_unregister - Disconnect a probe from a marker
686 * @probe: probe function pointer
687 * @probe_private: probe private data
689 * Returns the private data given to marker_probe_register, or an ERR_PTR().
690 * We do not need to call a synchronize_sched to make sure the probes have
691 * finished running before doing a module unload, because the module unload
692 * itself uses stop_machine(), which insures that every preempt disabled section
695 int marker_probe_unregister(const char *name,
696 marker_probe_func *probe, void *probe_private)
698 struct marker_entry *entry;
699 struct marker_probe_closure *old;
702 mutex_lock(&markers_mutex);
703 entry = get_marker(name);
706 if (entry->rcu_pending)
708 old = marker_entry_remove_probe(entry, probe, probe_private);
709 mutex_unlock(&markers_mutex);
710 marker_update_probes(); /* may update entry */
711 mutex_lock(&markers_mutex);
712 entry = get_marker(name);
716 entry->rcu_pending = 1;
717 /* write rcu_pending before calling the RCU callback */
719 #ifdef CONFIG_PREEMPT_RCU
720 synchronize_sched(); /* Until we have the call_rcu_sched() */
722 call_rcu(&entry->rcu, free_old_closure);
723 remove_marker(name); /* Ignore busy error message */
726 mutex_unlock(&markers_mutex);
729 EXPORT_SYMBOL_GPL(marker_probe_unregister);
731 static struct marker_entry *
732 get_marker_from_private_data(marker_probe_func *probe, void *probe_private)
734 struct marker_entry *entry;
736 struct hlist_head *head;
737 struct hlist_node *node;
739 for (i = 0; i < MARKER_TABLE_SIZE; i++) {
740 head = &marker_table[i];
741 hlist_for_each_entry(entry, node, head, hlist) {
743 if (entry->single.func == probe
744 && entry->single.probe_private
748 struct marker_probe_closure *closure;
749 closure = entry->multi;
750 for (i = 0; closure[i].func; i++) {
751 if (closure[i].func == probe &&
752 closure[i].probe_private
763 * marker_probe_unregister_private_data - Disconnect a probe from a marker
764 * @probe: probe function
765 * @probe_private: probe private data
767 * Unregister a probe by providing the registered private data.
768 * Only removes the first marker found in hash table.
769 * Return 0 on success or error value.
770 * We do not need to call a synchronize_sched to make sure the probes have
771 * finished running before doing a module unload, because the module unload
772 * itself uses stop_machine(), which insures that every preempt disabled section
775 int marker_probe_unregister_private_data(marker_probe_func *probe,
778 struct marker_entry *entry;
780 struct marker_probe_closure *old;
782 mutex_lock(&markers_mutex);
783 entry = get_marker_from_private_data(probe, probe_private);
788 if (entry->rcu_pending)
790 old = marker_entry_remove_probe(entry, NULL, probe_private);
791 mutex_unlock(&markers_mutex);
792 marker_update_probes(); /* may update entry */
793 mutex_lock(&markers_mutex);
794 entry = get_marker_from_private_data(probe, probe_private);
797 entry->rcu_pending = 1;
798 /* write rcu_pending before calling the RCU callback */
800 #ifdef CONFIG_PREEMPT_RCU
801 synchronize_sched(); /* Until we have the call_rcu_sched() */
803 call_rcu(&entry->rcu, free_old_closure);
804 remove_marker(entry->name); /* Ignore busy error message */
806 mutex_unlock(&markers_mutex);
809 EXPORT_SYMBOL_GPL(marker_probe_unregister_private_data);
812 * marker_get_private_data - Get a marker's probe private data
814 * @probe: probe to match
815 * @num: get the nth matching probe's private data
817 * Returns the nth private data pointer (starting from 0) matching, or an
819 * Returns the private data pointer, or an ERR_PTR.
820 * The private data pointer should _only_ be dereferenced if the caller is the
821 * owner of the data, or its content could vanish. This is mostly used to
822 * confirm that a caller is the owner of a registered probe.
824 void *marker_get_private_data(const char *name, marker_probe_func *probe,
827 struct hlist_head *head;
828 struct hlist_node *node;
829 struct marker_entry *e;
830 size_t name_len = strlen(name) + 1;
831 u32 hash = jhash(name, name_len-1, 0);
834 head = &marker_table[hash & ((1 << MARKER_HASH_BITS)-1)];
835 hlist_for_each_entry(e, node, head, hlist) {
836 if (!strcmp(name, e->name)) {
838 if (num == 0 && e->single.func == probe)
839 return e->single.probe_private;
843 struct marker_probe_closure *closure;
846 for (i = 0; closure[i].func; i++) {
847 if (closure[i].func != probe)
850 return closure[i].probe_private;
855 return ERR_PTR(-ENOENT);
857 EXPORT_SYMBOL_GPL(marker_get_private_data);