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 * Read mdata->ptype before mdata->multi.
132 multi
= mdata
->multi
;
134 * multi points to an array, therefore accessing the array
135 * depends on reading multi. However, even in this case,
136 * we must insure that the pointer is read _before_ the array
137 * data. Same as rcu_dereference, but we need a full smp_rmb()
138 * in the fast path, so put the explicit barrier here.
140 smp_read_barrier_depends();
141 for (i
= 0; multi
[i
].func
; i
++) {
142 va_start(args
, call_private
);
143 multi
[i
].func(multi
[i
].probe_private
, call_private
,
144 mdata
->format
, &args
);
150 EXPORT_SYMBOL_GPL(marker_probe_cb
);
153 * marker_probe_cb Callback that does not prepare the variable argument list.
154 * @mdata: pointer of type struct marker
155 * @call_private: caller site private data
156 * @...: Variable argument list.
158 * Should be connected to markers "MARK_NOARGS".
160 void marker_probe_cb_noarg(const struct marker
*mdata
, void *call_private
, ...)
162 va_list args
; /* not initialized */
166 ptype
= mdata
->ptype
;
167 if (likely(!ptype
)) {
168 marker_probe_func
*func
;
169 /* Must read the ptype before ptr. They are not data dependant,
170 * so we put an explicit smp_rmb() here. */
172 func
= mdata
->single
.func
;
173 /* Must read the ptr before private data. They are not data
174 * dependant, so we put an explicit smp_rmb() here. */
176 func(mdata
->single
.probe_private
, call_private
, mdata
->format
,
179 struct marker_probe_closure
*multi
;
182 * Read mdata->ptype before mdata->multi.
185 multi
= mdata
->multi
;
187 * multi points to an array, therefore accessing the array
188 * depends on reading multi. However, even in this case,
189 * we must insure that the pointer is read _before_ the array
190 * data. Same as rcu_dereference, but we need a full smp_rmb()
191 * in the fast path, so put the explicit barrier here.
193 smp_read_barrier_depends();
194 for (i
= 0; multi
[i
].func
; i
++)
195 multi
[i
].func(multi
[i
].probe_private
, call_private
,
196 mdata
->format
, &args
);
200 EXPORT_SYMBOL_GPL(marker_probe_cb_noarg
);
202 static void free_old_closure(struct rcu_head
*head
)
204 struct marker_entry
*entry
= container_of(head
,
205 struct marker_entry
, rcu
);
206 kfree(entry
->oldptr
);
207 /* Make sure we free the data before setting the pending flag to 0 */
209 entry
->rcu_pending
= 0;
212 static void debug_print_probes(struct marker_entry
*entry
)
220 printk(KERN_DEBUG
"Single probe : %p %p\n",
222 entry
->single
.probe_private
);
224 for (i
= 0; entry
->multi
[i
].func
; i
++)
225 printk(KERN_DEBUG
"Multi probe %d : %p %p\n", i
,
226 entry
->multi
[i
].func
,
227 entry
->multi
[i
].probe_private
);
231 static struct marker_probe_closure
*
232 marker_entry_add_probe(struct marker_entry
*entry
,
233 marker_probe_func
*probe
, void *probe_private
)
236 struct marker_probe_closure
*old
, *new;
240 debug_print_probes(entry
);
243 if (entry
->single
.func
== probe
&&
244 entry
->single
.probe_private
== probe_private
)
245 return ERR_PTR(-EBUSY
);
246 if (entry
->single
.func
== __mark_empty_function
) {
248 entry
->single
.func
= probe
;
249 entry
->single
.probe_private
= probe_private
;
252 debug_print_probes(entry
);
260 /* (N -> N+1), (N != 0, 1) probes */
261 for (nr_probes
= 0; old
[nr_probes
].func
; nr_probes
++)
262 if (old
[nr_probes
].func
== probe
263 && old
[nr_probes
].probe_private
265 return ERR_PTR(-EBUSY
);
267 /* + 2 : one for new probe, one for NULL func */
268 new = kzalloc((nr_probes
+ 2) * sizeof(struct marker_probe_closure
),
271 return ERR_PTR(-ENOMEM
);
273 new[0] = entry
->single
;
276 nr_probes
* sizeof(struct marker_probe_closure
));
277 new[nr_probes
].func
= probe
;
278 new[nr_probes
].probe_private
= probe_private
;
279 entry
->refcount
= nr_probes
+ 1;
282 debug_print_probes(entry
);
286 static struct marker_probe_closure
*
287 marker_entry_remove_probe(struct marker_entry
*entry
,
288 marker_probe_func
*probe
, void *probe_private
)
290 int nr_probes
= 0, nr_del
= 0, i
;
291 struct marker_probe_closure
*old
, *new;
295 debug_print_probes(entry
);
297 /* 0 -> N is an error */
298 WARN_ON(entry
->single
.func
== __mark_empty_function
);
300 WARN_ON(probe
&& entry
->single
.func
!= probe
);
301 WARN_ON(entry
->single
.probe_private
!= probe_private
);
302 entry
->single
.func
= __mark_empty_function
;
305 debug_print_probes(entry
);
308 /* (N -> M), (N > 1, M >= 0) probes */
309 for (nr_probes
= 0; old
[nr_probes
].func
; nr_probes
++) {
310 if ((!probe
|| old
[nr_probes
].func
== probe
)
311 && old
[nr_probes
].probe_private
317 if (nr_probes
- nr_del
== 0) {
318 /* N -> 0, (N > 1) */
319 entry
->single
.func
= __mark_empty_function
;
322 } else if (nr_probes
- nr_del
== 1) {
323 /* N -> 1, (N > 1) */
324 for (i
= 0; old
[i
].func
; i
++)
325 if ((probe
&& old
[i
].func
!= probe
) ||
326 old
[i
].probe_private
!= probe_private
)
327 entry
->single
= old
[i
];
332 /* N -> M, (N > 1, M > 1) */
334 new = kzalloc((nr_probes
- nr_del
+ 1)
335 * sizeof(struct marker_probe_closure
), GFP_KERNEL
);
337 return ERR_PTR(-ENOMEM
);
338 for (i
= 0; old
[i
].func
; i
++)
339 if ((probe
&& old
[i
].func
!= probe
) ||
340 old
[i
].probe_private
!= probe_private
)
342 entry
->refcount
= nr_probes
- nr_del
;
346 debug_print_probes(entry
);
351 * Get marker if the marker is present in the marker hash table.
352 * Must be called with markers_mutex held.
353 * Returns NULL if not present.
355 static struct marker_entry
*get_marker(const char *name
)
357 struct hlist_head
*head
;
358 struct hlist_node
*node
;
359 struct marker_entry
*e
;
360 u32 hash
= jhash(name
, strlen(name
), 0);
362 head
= &marker_table
[hash
& ((1 << MARKER_HASH_BITS
)-1)];
363 hlist_for_each_entry(e
, node
, head
, hlist
) {
364 if (!strcmp(name
, e
->name
))
371 * Add the marker to the marker hash table. Must be called with markers_mutex
374 static struct marker_entry
*add_marker(const char *name
, const char *format
)
376 struct hlist_head
*head
;
377 struct hlist_node
*node
;
378 struct marker_entry
*e
;
379 size_t name_len
= strlen(name
) + 1;
380 size_t format_len
= 0;
381 u32 hash
= jhash(name
, name_len
-1, 0);
384 format_len
= strlen(format
) + 1;
385 head
= &marker_table
[hash
& ((1 << MARKER_HASH_BITS
)-1)];
386 hlist_for_each_entry(e
, node
, head
, hlist
) {
387 if (!strcmp(name
, e
->name
)) {
389 "Marker %s busy\n", name
);
390 return ERR_PTR(-EBUSY
); /* Already there */
394 * Using kmalloc here to allocate a variable length element. Could
395 * cause some memory fragmentation if overused.
397 e
= kmalloc(sizeof(struct marker_entry
) + name_len
+ format_len
,
400 return ERR_PTR(-ENOMEM
);
401 memcpy(&e
->name
[0], name
, name_len
);
403 e
->format
= &e
->name
[name_len
];
404 memcpy(e
->format
, format
, format_len
);
405 if (strcmp(e
->format
, MARK_NOARGS
) == 0)
406 e
->call
= marker_probe_cb_noarg
;
408 e
->call
= marker_probe_cb
;
409 trace_mark(core_marker_format
, "name %s format %s",
413 e
->call
= marker_probe_cb
;
415 e
->single
.func
= __mark_empty_function
;
416 e
->single
.probe_private
= NULL
;
421 hlist_add_head(&e
->hlist
, head
);
426 * Remove the marker from the marker hash table. Must be called with mutex_lock
429 static int remove_marker(const char *name
)
431 struct hlist_head
*head
;
432 struct hlist_node
*node
;
433 struct marker_entry
*e
;
435 size_t len
= strlen(name
) + 1;
436 u32 hash
= jhash(name
, len
-1, 0);
438 head
= &marker_table
[hash
& ((1 << MARKER_HASH_BITS
)-1)];
439 hlist_for_each_entry(e
, node
, head
, hlist
) {
440 if (!strcmp(name
, e
->name
)) {
447 if (e
->single
.func
!= __mark_empty_function
)
449 hlist_del(&e
->hlist
);
450 /* Make sure the call_rcu has been executed */
458 * Set the mark_entry format to the format found in the element.
460 static int marker_set_format(struct marker_entry
**entry
, const char *format
)
462 struct marker_entry
*e
;
463 size_t name_len
= strlen((*entry
)->name
) + 1;
464 size_t format_len
= strlen(format
) + 1;
467 e
= kmalloc(sizeof(struct marker_entry
) + name_len
+ format_len
,
471 memcpy(&e
->name
[0], (*entry
)->name
, name_len
);
472 e
->format
= &e
->name
[name_len
];
473 memcpy(e
->format
, format
, format_len
);
474 if (strcmp(e
->format
, MARK_NOARGS
) == 0)
475 e
->call
= marker_probe_cb_noarg
;
477 e
->call
= marker_probe_cb
;
478 e
->single
= (*entry
)->single
;
479 e
->multi
= (*entry
)->multi
;
480 e
->ptype
= (*entry
)->ptype
;
481 e
->refcount
= (*entry
)->refcount
;
483 hlist_add_before(&e
->hlist
, &(*entry
)->hlist
);
484 hlist_del(&(*entry
)->hlist
);
485 /* Make sure the call_rcu has been executed */
486 if ((*entry
)->rcu_pending
)
490 trace_mark(core_marker_format
, "name %s format %s",
496 * Sets the probe callback corresponding to one marker.
498 static int set_marker(struct marker_entry
**entry
, struct marker
*elem
,
502 WARN_ON(strcmp((*entry
)->name
, elem
->name
) != 0);
504 if ((*entry
)->format
) {
505 if (strcmp((*entry
)->format
, elem
->format
) != 0) {
507 "Format mismatch for probe %s "
508 "(%s), marker (%s)\n",
515 ret
= marker_set_format(entry
, elem
->format
);
521 * probe_cb setup (statically known) is done here. It is
522 * asynchronous with the rest of execution, therefore we only
523 * pass from a "safe" callback (with argument) to an "unsafe"
524 * callback (does not set arguments).
526 elem
->call
= (*entry
)->call
;
529 * We only update the single probe private data when the ptr is
530 * set to a _non_ single probe! (0 -> 1 and N -> 1, N != 1)
532 WARN_ON(elem
->single
.func
!= __mark_empty_function
533 && elem
->single
.probe_private
534 != (*entry
)->single
.probe_private
&&
536 elem
->single
.probe_private
= (*entry
)->single
.probe_private
;
538 * Make sure the private data is valid when we update the
542 elem
->single
.func
= (*entry
)->single
.func
;
544 * We also make sure that the new probe callbacks array is consistent
545 * before setting a pointer to it.
547 rcu_assign_pointer(elem
->multi
, (*entry
)->multi
);
549 * Update the function or multi probe array pointer before setting the
553 elem
->ptype
= (*entry
)->ptype
;
554 elem
->state
= active
;
560 * Disable a marker and its probe callback.
561 * Note: only waiting an RCU period after setting elem->call to the empty
562 * function insures that the original callback is not used anymore. This insured
563 * by preempt_disable around the call site.
565 static void disable_marker(struct marker
*elem
)
567 /* leave "call" as is. It is known statically. */
569 elem
->single
.func
= __mark_empty_function
;
570 /* Update the function before setting the ptype */
572 elem
->ptype
= 0; /* single probe */
574 * Leave the private data and id there, because removal is racy and
575 * should be done only after an RCU period. These are never used until
576 * the next initialization anyway.
581 * marker_update_probe_range - Update a probe range
582 * @begin: beginning of the range
583 * @end: end of the range
585 * Updates the probe callback corresponding to a range of markers.
587 void marker_update_probe_range(struct marker
*begin
,
591 struct marker_entry
*mark_entry
;
593 mutex_lock(&markers_mutex
);
594 for (iter
= begin
; iter
< end
; iter
++) {
595 mark_entry
= get_marker(iter
->name
);
597 set_marker(&mark_entry
, iter
,
598 !!mark_entry
->refcount
);
600 * ignore error, continue
603 disable_marker(iter
);
606 mutex_unlock(&markers_mutex
);
610 * Update probes, removing the faulty probes.
612 * Internal callback only changed before the first probe is connected to it.
613 * Single probe private data can only be changed on 0 -> 1 and 2 -> 1
614 * transitions. All other transitions will leave the old private data valid.
615 * This makes the non-atomicity of the callback/private data updates valid.
617 * "special case" updates :
622 * Other updates all behave the same, just like the 2 -> 3 or 3 -> 2 updates.
623 * Site effect : marker_set_format may delete the marker entry (creating a
626 static void marker_update_probes(void)
628 /* Core kernel markers */
629 marker_update_probe_range(__start___markers
, __stop___markers
);
630 /* Markers in modules. */
631 module_update_markers();
635 * marker_probe_register - Connect a probe to a marker
637 * @format: format string
638 * @probe: probe handler
639 * @probe_private: probe private data
641 * private data must be a valid allocated memory address, or NULL.
642 * Returns 0 if ok, error value on error.
643 * The probe address must at least be aligned on the architecture pointer size.
645 int marker_probe_register(const char *name
, const char *format
,
646 marker_probe_func
*probe
, void *probe_private
)
648 struct marker_entry
*entry
;
650 struct marker_probe_closure
*old
;
652 mutex_lock(&markers_mutex
);
653 entry
= get_marker(name
);
655 entry
= add_marker(name
, format
);
657 ret
= PTR_ERR(entry
);
662 * If we detect that a call_rcu is pending for this marker,
663 * make sure it's executed now.
665 if (entry
->rcu_pending
)
667 old
= marker_entry_add_probe(entry
, probe
, probe_private
);
672 mutex_unlock(&markers_mutex
);
673 marker_update_probes(); /* may update entry */
674 mutex_lock(&markers_mutex
);
675 entry
= get_marker(name
);
678 entry
->rcu_pending
= 1;
679 /* write rcu_pending before calling the RCU callback */
681 call_rcu_sched(&entry
->rcu
, free_old_closure
);
683 mutex_unlock(&markers_mutex
);
686 EXPORT_SYMBOL_GPL(marker_probe_register
);
689 * marker_probe_unregister - Disconnect a probe from a marker
691 * @probe: probe function pointer
692 * @probe_private: probe private data
694 * Returns the private data given to marker_probe_register, or an ERR_PTR().
695 * We do not need to call a synchronize_sched to make sure the probes have
696 * finished running before doing a module unload, because the module unload
697 * itself uses stop_machine(), which insures that every preempt disabled section
700 int marker_probe_unregister(const char *name
,
701 marker_probe_func
*probe
, void *probe_private
)
703 struct marker_entry
*entry
;
704 struct marker_probe_closure
*old
;
707 mutex_lock(&markers_mutex
);
708 entry
= get_marker(name
);
711 if (entry
->rcu_pending
)
713 old
= marker_entry_remove_probe(entry
, probe
, probe_private
);
714 mutex_unlock(&markers_mutex
);
715 marker_update_probes(); /* may update entry */
716 mutex_lock(&markers_mutex
);
717 entry
= get_marker(name
);
721 entry
->rcu_pending
= 1;
722 /* write rcu_pending before calling the RCU callback */
724 call_rcu_sched(&entry
->rcu
, free_old_closure
);
725 remove_marker(name
); /* Ignore busy error message */
728 mutex_unlock(&markers_mutex
);
731 EXPORT_SYMBOL_GPL(marker_probe_unregister
);
733 static struct marker_entry
*
734 get_marker_from_private_data(marker_probe_func
*probe
, void *probe_private
)
736 struct marker_entry
*entry
;
738 struct hlist_head
*head
;
739 struct hlist_node
*node
;
741 for (i
= 0; i
< MARKER_TABLE_SIZE
; i
++) {
742 head
= &marker_table
[i
];
743 hlist_for_each_entry(entry
, node
, head
, hlist
) {
745 if (entry
->single
.func
== probe
746 && entry
->single
.probe_private
750 struct marker_probe_closure
*closure
;
751 closure
= entry
->multi
;
752 for (i
= 0; closure
[i
].func
; i
++) {
753 if (closure
[i
].func
== probe
&&
754 closure
[i
].probe_private
765 * marker_probe_unregister_private_data - Disconnect a probe from a marker
766 * @probe: probe function
767 * @probe_private: probe private data
769 * Unregister a probe by providing the registered private data.
770 * Only removes the first marker found in hash table.
771 * Return 0 on success or error value.
772 * We do not need to call a synchronize_sched to make sure the probes have
773 * finished running before doing a module unload, because the module unload
774 * itself uses stop_machine(), which insures that every preempt disabled section
777 int marker_probe_unregister_private_data(marker_probe_func
*probe
,
780 struct marker_entry
*entry
;
782 struct marker_probe_closure
*old
;
784 mutex_lock(&markers_mutex
);
785 entry
= get_marker_from_private_data(probe
, probe_private
);
790 if (entry
->rcu_pending
)
792 old
= marker_entry_remove_probe(entry
, NULL
, probe_private
);
793 mutex_unlock(&markers_mutex
);
794 marker_update_probes(); /* may update entry */
795 mutex_lock(&markers_mutex
);
796 entry
= get_marker_from_private_data(probe
, probe_private
);
799 entry
->rcu_pending
= 1;
800 /* write rcu_pending before calling the RCU callback */
802 call_rcu_sched(&entry
->rcu
, free_old_closure
);
803 remove_marker(entry
->name
); /* Ignore busy error message */
805 mutex_unlock(&markers_mutex
);
808 EXPORT_SYMBOL_GPL(marker_probe_unregister_private_data
);
811 * marker_get_private_data - Get a marker's probe private data
813 * @probe: probe to match
814 * @num: get the nth matching probe's private data
816 * Returns the nth private data pointer (starting from 0) matching, or an
818 * Returns the private data pointer, or an ERR_PTR.
819 * The private data pointer should _only_ be dereferenced if the caller is the
820 * owner of the data, or its content could vanish. This is mostly used to
821 * confirm that a caller is the owner of a registered probe.
823 void *marker_get_private_data(const char *name
, marker_probe_func
*probe
,
826 struct hlist_head
*head
;
827 struct hlist_node
*node
;
828 struct marker_entry
*e
;
829 size_t name_len
= strlen(name
) + 1;
830 u32 hash
= jhash(name
, name_len
-1, 0);
833 head
= &marker_table
[hash
& ((1 << MARKER_HASH_BITS
)-1)];
834 hlist_for_each_entry(e
, node
, head
, hlist
) {
835 if (!strcmp(name
, e
->name
)) {
837 if (num
== 0 && e
->single
.func
== probe
)
838 return e
->single
.probe_private
;
842 struct marker_probe_closure
*closure
;
845 for (i
= 0; closure
[i
].func
; i
++) {
846 if (closure
[i
].func
!= probe
)
849 return closure
[i
].probe_private
;
854 return ERR_PTR(-ENOENT
);
856 EXPORT_SYMBOL_GPL(marker_get_private_data
);