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
;
58 void (*call
)(const struct marker
*mdata
, /* Probe wrapper */
59 void *call_private
, const char *fmt
, ...);
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
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
,
102 const char *fmt
, ...)
108 * preempt_disable does two things : disabling preemption to make sure
109 * the teardown of the callbacks can be done correctly when they are in
110 * modules and they insure RCU read coherency.
113 ptype
= mdata
->ptype
;
114 if (likely(!ptype
)) {
115 marker_probe_func
*func
;
116 /* Must read the ptype before ptr. They are not data dependant,
117 * so we put an explicit smp_rmb() here. */
119 func
= mdata
->single
.func
;
120 /* Must read the ptr before private data. They are not data
121 * dependant, so we put an explicit smp_rmb() here. */
124 func(mdata
->single
.probe_private
, call_private
, fmt
, &args
);
127 struct marker_probe_closure
*multi
;
130 * multi points to an array, therefore accessing the array
131 * depends on reading multi. However, even in this case,
132 * we must insure that the pointer is read _before_ the array
133 * data. Same as rcu_dereference, but we need a full smp_rmb()
134 * in the fast path, so put the explicit barrier here.
136 smp_read_barrier_depends();
137 multi
= mdata
->multi
;
138 for (i
= 0; multi
[i
].func
; i
++) {
140 multi
[i
].func(multi
[i
].probe_private
, call_private
, fmt
,
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 * @fmt: format string
154 * @...: Variable argument list.
156 * Should be connected to markers "MARK_NOARGS".
158 void marker_probe_cb_noarg(const struct marker
*mdata
,
159 void *call_private
, const char *fmt
, ...)
161 va_list args
; /* not initialized */
165 ptype
= mdata
->ptype
;
166 if (likely(!ptype
)) {
167 marker_probe_func
*func
;
168 /* Must read the ptype before ptr. They are not data dependant,
169 * so we put an explicit smp_rmb() here. */
171 func
= mdata
->single
.func
;
172 /* Must read the ptr before private data. They are not data
173 * dependant, so we put an explicit smp_rmb() here. */
175 func(mdata
->single
.probe_private
, call_private
, fmt
, &args
);
177 struct marker_probe_closure
*multi
;
180 * multi points to an array, therefore accessing the array
181 * depends on reading multi. However, even in this case,
182 * we must insure that the pointer is read _before_ the array
183 * data. Same as rcu_dereference, but we need a full smp_rmb()
184 * in the fast path, so put the explicit barrier here.
186 smp_read_barrier_depends();
187 multi
= mdata
->multi
;
188 for (i
= 0; multi
[i
].func
; i
++)
189 multi
[i
].func(multi
[i
].probe_private
, call_private
, fmt
,
194 EXPORT_SYMBOL_GPL(marker_probe_cb_noarg
);
196 static void free_old_closure(struct rcu_head
*head
)
198 struct marker_entry
*entry
= container_of(head
,
199 struct marker_entry
, rcu
);
200 kfree(entry
->oldptr
);
201 /* Make sure we free the data before setting the pending flag to 0 */
203 entry
->rcu_pending
= 0;
206 static void debug_print_probes(struct marker_entry
*entry
)
214 printk(KERN_DEBUG
"Single probe : %p %p\n",
216 entry
->single
.probe_private
);
218 for (i
= 0; entry
->multi
[i
].func
; i
++)
219 printk(KERN_DEBUG
"Multi probe %d : %p %p\n", i
,
220 entry
->multi
[i
].func
,
221 entry
->multi
[i
].probe_private
);
225 static struct marker_probe_closure
*
226 marker_entry_add_probe(struct marker_entry
*entry
,
227 marker_probe_func
*probe
, void *probe_private
)
230 struct marker_probe_closure
*old
, *new;
234 debug_print_probes(entry
);
237 if (entry
->single
.func
== probe
&&
238 entry
->single
.probe_private
== probe_private
)
239 return ERR_PTR(-EBUSY
);
240 if (entry
->single
.func
== __mark_empty_function
) {
242 entry
->single
.func
= probe
;
243 entry
->single
.probe_private
= probe_private
;
246 debug_print_probes(entry
);
254 /* (N -> N+1), (N != 0, 1) probes */
255 for (nr_probes
= 0; old
[nr_probes
].func
; nr_probes
++)
256 if (old
[nr_probes
].func
== probe
257 && old
[nr_probes
].probe_private
259 return ERR_PTR(-EBUSY
);
261 /* + 2 : one for new probe, one for NULL func */
262 new = kzalloc((nr_probes
+ 2) * sizeof(struct marker_probe_closure
),
265 return ERR_PTR(-ENOMEM
);
267 new[0] = entry
->single
;
270 nr_probes
* sizeof(struct marker_probe_closure
));
271 new[nr_probes
].func
= probe
;
272 new[nr_probes
].probe_private
= probe_private
;
273 entry
->refcount
= nr_probes
+ 1;
276 debug_print_probes(entry
);
280 static struct marker_probe_closure
*
281 marker_entry_remove_probe(struct marker_entry
*entry
,
282 marker_probe_func
*probe
, void *probe_private
)
284 int nr_probes
= 0, nr_del
= 0, i
;
285 struct marker_probe_closure
*old
, *new;
289 debug_print_probes(entry
);
291 /* 0 -> N is an error */
292 WARN_ON(entry
->single
.func
== __mark_empty_function
);
294 WARN_ON(probe
&& entry
->single
.func
!= probe
);
295 WARN_ON(entry
->single
.probe_private
!= probe_private
);
296 entry
->single
.func
= __mark_empty_function
;
299 debug_print_probes(entry
);
302 /* (N -> M), (N > 1, M >= 0) probes */
303 for (nr_probes
= 0; old
[nr_probes
].func
; nr_probes
++) {
304 if ((!probe
|| old
[nr_probes
].func
== probe
)
305 && old
[nr_probes
].probe_private
311 if (nr_probes
- nr_del
== 0) {
312 /* N -> 0, (N > 1) */
313 entry
->single
.func
= __mark_empty_function
;
316 } else if (nr_probes
- nr_del
== 1) {
317 /* N -> 1, (N > 1) */
318 for (i
= 0; old
[i
].func
; i
++)
319 if ((probe
&& old
[i
].func
!= probe
) ||
320 old
[i
].probe_private
!= probe_private
)
321 entry
->single
= old
[i
];
326 /* N -> M, (N > 1, M > 1) */
328 new = kzalloc((nr_probes
- nr_del
+ 1)
329 * sizeof(struct marker_probe_closure
), GFP_KERNEL
);
331 return ERR_PTR(-ENOMEM
);
332 for (i
= 0; old
[i
].func
; i
++)
333 if ((probe
&& old
[i
].func
!= probe
) ||
334 old
[i
].probe_private
!= probe_private
)
336 entry
->refcount
= nr_probes
- nr_del
;
340 debug_print_probes(entry
);
345 * Get marker if the marker is present in the marker hash table.
346 * Must be called with markers_mutex held.
347 * Returns NULL if not present.
349 static struct marker_entry
*get_marker(const char *name
)
351 struct hlist_head
*head
;
352 struct hlist_node
*node
;
353 struct marker_entry
*e
;
354 u32 hash
= jhash(name
, strlen(name
), 0);
356 head
= &marker_table
[hash
& ((1 << MARKER_HASH_BITS
)-1)];
357 hlist_for_each_entry(e
, node
, head
, hlist
) {
358 if (!strcmp(name
, e
->name
))
365 * Add the marker to the marker hash table. Must be called with markers_mutex
368 static struct marker_entry
*add_marker(const char *name
, const char *format
)
370 struct hlist_head
*head
;
371 struct hlist_node
*node
;
372 struct marker_entry
*e
;
373 size_t name_len
= strlen(name
) + 1;
374 size_t format_len
= 0;
375 u32 hash
= jhash(name
, name_len
-1, 0);
378 format_len
= strlen(format
) + 1;
379 head
= &marker_table
[hash
& ((1 << MARKER_HASH_BITS
)-1)];
380 hlist_for_each_entry(e
, node
, head
, hlist
) {
381 if (!strcmp(name
, e
->name
)) {
383 "Marker %s busy\n", name
);
384 return ERR_PTR(-EBUSY
); /* Already there */
388 * Using kmalloc here to allocate a variable length element. Could
389 * cause some memory fragmentation if overused.
391 e
= kmalloc(sizeof(struct marker_entry
) + name_len
+ format_len
,
394 return ERR_PTR(-ENOMEM
);
395 memcpy(&e
->name
[0], name
, name_len
);
397 e
->format
= &e
->name
[name_len
];
398 memcpy(e
->format
, format
, format_len
);
399 if (strcmp(e
->format
, MARK_NOARGS
) == 0)
400 e
->call
= marker_probe_cb_noarg
;
402 e
->call
= marker_probe_cb
;
403 trace_mark(core_marker_format
, "name %s format %s",
407 e
->call
= marker_probe_cb
;
409 e
->single
.func
= __mark_empty_function
;
410 e
->single
.probe_private
= NULL
;
415 hlist_add_head(&e
->hlist
, head
);
420 * Remove the marker from the marker hash table. Must be called with mutex_lock
423 static int remove_marker(const char *name
)
425 struct hlist_head
*head
;
426 struct hlist_node
*node
;
427 struct marker_entry
*e
;
429 size_t len
= strlen(name
) + 1;
430 u32 hash
= jhash(name
, len
-1, 0);
432 head
= &marker_table
[hash
& ((1 << MARKER_HASH_BITS
)-1)];
433 hlist_for_each_entry(e
, node
, head
, hlist
) {
434 if (!strcmp(name
, e
->name
)) {
441 if (e
->single
.func
!= __mark_empty_function
)
443 hlist_del(&e
->hlist
);
444 /* Make sure the call_rcu has been executed */
452 * Set the mark_entry format to the format found in the element.
454 static int marker_set_format(struct marker_entry
**entry
, const char *format
)
456 struct marker_entry
*e
;
457 size_t name_len
= strlen((*entry
)->name
) + 1;
458 size_t format_len
= strlen(format
) + 1;
461 e
= kmalloc(sizeof(struct marker_entry
) + name_len
+ format_len
,
465 memcpy(&e
->name
[0], (*entry
)->name
, name_len
);
466 e
->format
= &e
->name
[name_len
];
467 memcpy(e
->format
, format
, format_len
);
468 if (strcmp(e
->format
, MARK_NOARGS
) == 0)
469 e
->call
= marker_probe_cb_noarg
;
471 e
->call
= marker_probe_cb
;
472 e
->single
= (*entry
)->single
;
473 e
->multi
= (*entry
)->multi
;
474 e
->ptype
= (*entry
)->ptype
;
475 e
->refcount
= (*entry
)->refcount
;
477 hlist_add_before(&e
->hlist
, &(*entry
)->hlist
);
478 hlist_del(&(*entry
)->hlist
);
479 /* Make sure the call_rcu has been executed */
480 if ((*entry
)->rcu_pending
)
484 trace_mark(core_marker_format
, "name %s format %s",
490 * Sets the probe callback corresponding to one marker.
492 static int set_marker(struct marker_entry
**entry
, struct marker
*elem
,
496 WARN_ON(strcmp((*entry
)->name
, elem
->name
) != 0);
498 if ((*entry
)->format
) {
499 if (strcmp((*entry
)->format
, elem
->format
) != 0) {
501 "Format mismatch for probe %s "
502 "(%s), marker (%s)\n",
509 ret
= marker_set_format(entry
, elem
->format
);
515 * probe_cb setup (statically known) is done here. It is
516 * asynchronous with the rest of execution, therefore we only
517 * pass from a "safe" callback (with argument) to an "unsafe"
518 * callback (does not set arguments).
520 elem
->call
= (*entry
)->call
;
523 * We only update the single probe private data when the ptr is
524 * set to a _non_ single probe! (0 -> 1 and N -> 1, N != 1)
526 WARN_ON(elem
->single
.func
!= __mark_empty_function
527 && elem
->single
.probe_private
528 != (*entry
)->single
.probe_private
&&
530 elem
->single
.probe_private
= (*entry
)->single
.probe_private
;
532 * Make sure the private data is valid when we update the
536 elem
->single
.func
= (*entry
)->single
.func
;
538 * We also make sure that the new probe callbacks array is consistent
539 * before setting a pointer to it.
541 rcu_assign_pointer(elem
->multi
, (*entry
)->multi
);
543 * Update the function or multi probe array pointer before setting the
547 elem
->ptype
= (*entry
)->ptype
;
548 elem
->state
= active
;
554 * Disable a marker and its probe callback.
555 * Note: only waiting an RCU period after setting elem->call to the empty
556 * function insures that the original callback is not used anymore. This insured
557 * by preempt_disable around the call site.
559 static void disable_marker(struct marker
*elem
)
561 /* leave "call" as is. It is known statically. */
563 elem
->single
.func
= __mark_empty_function
;
564 /* Update the function before setting the ptype */
566 elem
->ptype
= 0; /* single probe */
568 * Leave the private data and id there, because removal is racy and
569 * should be done only after an RCU period. These are never used until
570 * the next initialization anyway.
575 * marker_update_probe_range - Update a probe range
576 * @begin: beginning of the range
577 * @end: end of the range
579 * Updates the probe callback corresponding to a range of markers.
581 void marker_update_probe_range(struct marker
*begin
,
585 struct marker_entry
*mark_entry
;
587 mutex_lock(&markers_mutex
);
588 for (iter
= begin
; iter
< end
; iter
++) {
589 mark_entry
= get_marker(iter
->name
);
591 set_marker(&mark_entry
, iter
,
592 !!mark_entry
->refcount
);
594 * ignore error, continue
597 disable_marker(iter
);
600 mutex_unlock(&markers_mutex
);
604 * Update probes, removing the faulty probes.
606 * Internal callback only changed before the first probe is connected to it.
607 * Single probe private data can only be changed on 0 -> 1 and 2 -> 1
608 * transitions. All other transitions will leave the old private data valid.
609 * This makes the non-atomicity of the callback/private data updates valid.
611 * "special case" updates :
616 * Other updates all behave the same, just like the 2 -> 3 or 3 -> 2 updates.
617 * Site effect : marker_set_format may delete the marker entry (creating a
620 static void marker_update_probes(void)
622 /* Core kernel markers */
623 marker_update_probe_range(__start___markers
, __stop___markers
);
624 /* Markers in modules. */
625 module_update_markers();
629 * marker_probe_register - Connect a probe to a marker
631 * @format: format string
632 * @probe: probe handler
633 * @probe_private: probe private data
635 * private data must be a valid allocated memory address, or NULL.
636 * Returns 0 if ok, error value on error.
637 * The probe address must at least be aligned on the architecture pointer size.
639 int marker_probe_register(const char *name
, const char *format
,
640 marker_probe_func
*probe
, void *probe_private
)
642 struct marker_entry
*entry
;
644 struct marker_probe_closure
*old
;
646 mutex_lock(&markers_mutex
);
647 entry
= get_marker(name
);
649 entry
= add_marker(name
, format
);
651 ret
= PTR_ERR(entry
);
656 * If we detect that a call_rcu is pending for this marker,
657 * make sure it's executed now.
659 if (entry
->rcu_pending
)
661 old
= marker_entry_add_probe(entry
, probe
, probe_private
);
666 mutex_unlock(&markers_mutex
);
667 marker_update_probes(); /* may update entry */
668 mutex_lock(&markers_mutex
);
669 entry
= get_marker(name
);
672 entry
->rcu_pending
= 1;
673 /* write rcu_pending before calling the RCU callback */
675 #ifdef CONFIG_PREEMPT_RCU
676 synchronize_sched(); /* Until we have the call_rcu_sched() */
678 call_rcu(&entry
->rcu
, free_old_closure
);
680 mutex_unlock(&markers_mutex
);
683 EXPORT_SYMBOL_GPL(marker_probe_register
);
686 * marker_probe_unregister - Disconnect a probe from a marker
688 * @probe: probe function pointer
689 * @probe_private: probe private data
691 * Returns the private data given to marker_probe_register, or an ERR_PTR().
692 * We do not need to call a synchronize_sched to make sure the probes have
693 * finished running before doing a module unload, because the module unload
694 * itself uses stop_machine(), which insures that every preempt disabled section
697 int marker_probe_unregister(const char *name
,
698 marker_probe_func
*probe
, void *probe_private
)
700 struct marker_entry
*entry
;
701 struct marker_probe_closure
*old
;
704 mutex_lock(&markers_mutex
);
705 entry
= get_marker(name
);
708 if (entry
->rcu_pending
)
710 old
= marker_entry_remove_probe(entry
, probe
, probe_private
);
711 mutex_unlock(&markers_mutex
);
712 marker_update_probes(); /* may update entry */
713 mutex_lock(&markers_mutex
);
714 entry
= get_marker(name
);
718 entry
->rcu_pending
= 1;
719 /* write rcu_pending before calling the RCU callback */
721 #ifdef CONFIG_PREEMPT_RCU
722 synchronize_sched(); /* Until we have the call_rcu_sched() */
724 call_rcu(&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 #ifdef CONFIG_PREEMPT_RCU
803 synchronize_sched(); /* Until we have the call_rcu_sched() */
805 call_rcu(&entry
->rcu
, free_old_closure
);
806 remove_marker(entry
->name
); /* Ignore busy error message */
808 mutex_unlock(&markers_mutex
);
811 EXPORT_SYMBOL_GPL(marker_probe_unregister_private_data
);
814 * marker_get_private_data - Get a marker's probe private data
816 * @probe: probe to match
817 * @num: get the nth matching probe's private data
819 * Returns the nth private data pointer (starting from 0) matching, or an
821 * Returns the private data pointer, or an ERR_PTR.
822 * The private data pointer should _only_ be dereferenced if the caller is the
823 * owner of the data, or its content could vanish. This is mostly used to
824 * confirm that a caller is the owner of a registered probe.
826 void *marker_get_private_data(const char *name
, marker_probe_func
*probe
,
829 struct hlist_head
*head
;
830 struct hlist_node
*node
;
831 struct marker_entry
*e
;
832 size_t name_len
= strlen(name
) + 1;
833 u32 hash
= jhash(name
, name_len
-1, 0);
836 head
= &marker_table
[hash
& ((1 << MARKER_HASH_BITS
)-1)];
837 hlist_for_each_entry(e
, node
, head
, hlist
) {
838 if (!strcmp(name
, e
->name
)) {
840 if (num
== 0 && e
->single
.func
== probe
)
841 return e
->single
.probe_private
;
845 struct marker_probe_closure
*closure
;
848 for (i
= 0; closure
[i
].func
; i
++) {
849 if (closure
[i
].func
!= probe
)
852 return closure
[i
].probe_private
;
857 return ERR_PTR(-ENOENT
);
859 EXPORT_SYMBOL_GPL(marker_get_private_data
);