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>
27 extern struct marker __start___markers
[];
28 extern struct marker __stop___markers
[];
30 /* Set to 1 to enable marker debug output */
31 const int marker_debug
;
34 * markers_mutex nests inside module_mutex. Markers mutex protects the builtin
35 * and module markers and the hash table.
37 static DEFINE_MUTEX(markers_mutex
);
40 * Marker hash table, containing the active markers.
41 * Protected by module_mutex.
43 #define MARKER_HASH_BITS 6
44 #define MARKER_TABLE_SIZE (1 << MARKER_HASH_BITS)
48 * It is used to make sure every handler has finished using its private data
49 * between two consecutive operation (add or remove) on a given marker. It is
50 * also used to delay the free of multiple probes array until a quiescent state
52 * marker entries modifications are protected by the markers_mutex.
55 struct hlist_node hlist
;
57 void (*call
)(const struct marker
*mdata
, /* Probe wrapper */
58 void *call_private
, const char *fmt
, ...);
59 struct marker_probe_closure single
;
60 struct marker_probe_closure
*multi
;
61 int refcount
; /* Number of times armed. 0 if disarmed. */
64 unsigned char rcu_pending
:1;
65 unsigned char ptype
:1;
66 char name
[0]; /* Contains name'\0'format'\0' */
69 static struct hlist_head marker_table
[MARKER_TABLE_SIZE
];
72 * __mark_empty_function - Empty probe callback
73 * @probe_private: probe private data
74 * @call_private: call site private data
76 * @...: variable argument list
78 * Empty callback provided as a probe to the markers. By providing this to a
79 * disabled marker, we make sure the execution flow is always valid even
80 * though the function pointer change and the marker enabling are two distinct
81 * operations that modifies the execution flow of preemptible code.
83 void __mark_empty_function(void *probe_private
, void *call_private
,
84 const char *fmt
, va_list *args
)
87 EXPORT_SYMBOL_GPL(__mark_empty_function
);
90 * marker_probe_cb Callback that prepares the variable argument list for probes.
91 * @mdata: pointer of type struct marker
92 * @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
,
101 const char *fmt
, ...)
107 * preempt_disable does two things : disabling preemption to make sure
108 * the teardown of the callbacks can be done correctly when they are in
109 * modules and they insure RCU read coherency.
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. */
123 func(mdata
->single
.probe_private
, call_private
, fmt
, &args
);
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
++) {
139 multi
[i
].func(multi
[i
].probe_private
, call_private
, fmt
,
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 * @fmt: format string
153 * @...: Variable argument list.
155 * Should be connected to markers "MARK_NOARGS".
157 void marker_probe_cb_noarg(const struct marker
*mdata
,
158 void *call_private
, const char *fmt
, ...)
160 va_list args
; /* not initialized */
164 ptype
= mdata
->ptype
;
165 if (likely(!ptype
)) {
166 marker_probe_func
*func
;
167 /* Must read the ptype before ptr. They are not data dependant,
168 * so we put an explicit smp_rmb() here. */
170 func
= mdata
->single
.func
;
171 /* Must read the ptr before private data. They are not data
172 * dependant, so we put an explicit smp_rmb() here. */
174 func(mdata
->single
.probe_private
, call_private
, fmt
, &args
);
176 struct marker_probe_closure
*multi
;
179 * multi points to an array, therefore accessing the array
180 * depends on reading multi. However, even in this case,
181 * we must insure that the pointer is read _before_ the array
182 * data. Same as rcu_dereference, but we need a full smp_rmb()
183 * in the fast path, so put the explicit barrier here.
185 smp_read_barrier_depends();
186 multi
= mdata
->multi
;
187 for (i
= 0; multi
[i
].func
; i
++)
188 multi
[i
].func(multi
[i
].probe_private
, call_private
, fmt
,
193 EXPORT_SYMBOL_GPL(marker_probe_cb_noarg
);
195 static void free_old_closure(struct rcu_head
*head
)
197 struct marker_entry
*entry
= container_of(head
,
198 struct marker_entry
, rcu
);
199 kfree(entry
->oldptr
);
200 /* Make sure we free the data before setting the pending flag to 0 */
202 entry
->rcu_pending
= 0;
205 static void debug_print_probes(struct marker_entry
*entry
)
213 printk(KERN_DEBUG
"Single probe : %p %p\n",
215 entry
->single
.probe_private
);
217 for (i
= 0; entry
->multi
[i
].func
; i
++)
218 printk(KERN_DEBUG
"Multi probe %d : %p %p\n", i
,
219 entry
->multi
[i
].func
,
220 entry
->multi
[i
].probe_private
);
224 static struct marker_probe_closure
*
225 marker_entry_add_probe(struct marker_entry
*entry
,
226 marker_probe_func
*probe
, void *probe_private
)
229 struct marker_probe_closure
*old
, *new;
233 debug_print_probes(entry
);
236 if (entry
->single
.func
== probe
&&
237 entry
->single
.probe_private
== probe_private
)
238 return ERR_PTR(-EBUSY
);
239 if (entry
->single
.func
== __mark_empty_function
) {
241 entry
->single
.func
= probe
;
242 entry
->single
.probe_private
= probe_private
;
245 debug_print_probes(entry
);
253 /* (N -> N+1), (N != 0, 1) probes */
254 for (nr_probes
= 0; old
[nr_probes
].func
; nr_probes
++)
255 if (old
[nr_probes
].func
== probe
256 && old
[nr_probes
].probe_private
258 return ERR_PTR(-EBUSY
);
260 /* + 2 : one for new probe, one for NULL func */
261 new = kzalloc((nr_probes
+ 2) * sizeof(struct marker_probe_closure
),
264 return ERR_PTR(-ENOMEM
);
266 new[0] = entry
->single
;
269 nr_probes
* sizeof(struct marker_probe_closure
));
270 new[nr_probes
].func
= probe
;
271 new[nr_probes
].probe_private
= probe_private
;
272 entry
->refcount
= nr_probes
+ 1;
275 debug_print_probes(entry
);
279 static struct marker_probe_closure
*
280 marker_entry_remove_probe(struct marker_entry
*entry
,
281 marker_probe_func
*probe
, void *probe_private
)
283 int nr_probes
= 0, nr_del
= 0, i
;
284 struct marker_probe_closure
*old
, *new;
288 debug_print_probes(entry
);
290 /* 0 -> N is an error */
291 WARN_ON(entry
->single
.func
== __mark_empty_function
);
293 WARN_ON(probe
&& entry
->single
.func
!= probe
);
294 WARN_ON(entry
->single
.probe_private
!= probe_private
);
295 entry
->single
.func
= __mark_empty_function
;
298 debug_print_probes(entry
);
301 /* (N -> M), (N > 1, M >= 0) probes */
302 for (nr_probes
= 0; old
[nr_probes
].func
; nr_probes
++) {
303 if ((!probe
|| old
[nr_probes
].func
== probe
)
304 && old
[nr_probes
].probe_private
310 if (nr_probes
- nr_del
== 0) {
311 /* N -> 0, (N > 1) */
312 entry
->single
.func
= __mark_empty_function
;
315 } else if (nr_probes
- nr_del
== 1) {
316 /* N -> 1, (N > 1) */
317 for (i
= 0; old
[i
].func
; i
++)
318 if ((probe
&& old
[i
].func
!= probe
) ||
319 old
[i
].probe_private
!= probe_private
)
320 entry
->single
= old
[i
];
325 /* N -> M, (N > 1, M > 1) */
327 new = kzalloc((nr_probes
- nr_del
+ 1)
328 * sizeof(struct marker_probe_closure
), GFP_KERNEL
);
330 return ERR_PTR(-ENOMEM
);
331 for (i
= 0; old
[i
].func
; i
++)
332 if ((probe
&& old
[i
].func
!= probe
) ||
333 old
[i
].probe_private
!= probe_private
)
335 entry
->refcount
= nr_probes
- nr_del
;
339 debug_print_probes(entry
);
344 * Get marker if the marker is present in the marker hash table.
345 * Must be called with markers_mutex held.
346 * Returns NULL if not present.
348 static struct marker_entry
*get_marker(const char *name
)
350 struct hlist_head
*head
;
351 struct hlist_node
*node
;
352 struct marker_entry
*e
;
353 u32 hash
= jhash(name
, strlen(name
), 0);
355 head
= &marker_table
[hash
& ((1 << MARKER_HASH_BITS
)-1)];
356 hlist_for_each_entry(e
, node
, head
, hlist
) {
357 if (!strcmp(name
, e
->name
))
364 * Add the marker to the marker hash table. Must be called with markers_mutex
367 static struct marker_entry
*add_marker(const char *name
, const char *format
)
369 struct hlist_head
*head
;
370 struct hlist_node
*node
;
371 struct marker_entry
*e
;
372 size_t name_len
= strlen(name
) + 1;
373 size_t format_len
= 0;
374 u32 hash
= jhash(name
, name_len
-1, 0);
377 format_len
= strlen(format
) + 1;
378 head
= &marker_table
[hash
& ((1 << MARKER_HASH_BITS
)-1)];
379 hlist_for_each_entry(e
, node
, head
, hlist
) {
380 if (!strcmp(name
, e
->name
)) {
382 "Marker %s busy\n", name
);
383 return ERR_PTR(-EBUSY
); /* Already there */
387 * Using kmalloc here to allocate a variable length element. Could
388 * cause some memory fragmentation if overused.
390 e
= kmalloc(sizeof(struct marker_entry
) + name_len
+ format_len
,
393 return ERR_PTR(-ENOMEM
);
394 memcpy(&e
->name
[0], name
, name_len
);
396 e
->format
= &e
->name
[name_len
];
397 memcpy(e
->format
, format
, format_len
);
398 if (strcmp(e
->format
, MARK_NOARGS
) == 0)
399 e
->call
= marker_probe_cb_noarg
;
401 e
->call
= marker_probe_cb
;
402 trace_mark(core_marker_format
, "name %s format %s",
406 e
->call
= marker_probe_cb
;
408 e
->single
.func
= __mark_empty_function
;
409 e
->single
.probe_private
= NULL
;
414 hlist_add_head(&e
->hlist
, head
);
419 * Remove the marker from the marker hash table. Must be called with mutex_lock
422 static int remove_marker(const char *name
)
424 struct hlist_head
*head
;
425 struct hlist_node
*node
;
426 struct marker_entry
*e
;
428 size_t len
= strlen(name
) + 1;
429 u32 hash
= jhash(name
, len
-1, 0);
431 head
= &marker_table
[hash
& ((1 << MARKER_HASH_BITS
)-1)];
432 hlist_for_each_entry(e
, node
, head
, hlist
) {
433 if (!strcmp(name
, e
->name
)) {
440 if (e
->single
.func
!= __mark_empty_function
)
442 hlist_del(&e
->hlist
);
443 /* Make sure the call_rcu has been executed */
451 * Set the mark_entry format to the format found in the element.
453 static int marker_set_format(struct marker_entry
**entry
, const char *format
)
455 struct marker_entry
*e
;
456 size_t name_len
= strlen((*entry
)->name
) + 1;
457 size_t format_len
= strlen(format
) + 1;
460 e
= kmalloc(sizeof(struct marker_entry
) + name_len
+ format_len
,
464 memcpy(&e
->name
[0], (*entry
)->name
, name_len
);
465 e
->format
= &e
->name
[name_len
];
466 memcpy(e
->format
, format
, format_len
);
467 if (strcmp(e
->format
, MARK_NOARGS
) == 0)
468 e
->call
= marker_probe_cb_noarg
;
470 e
->call
= marker_probe_cb
;
471 e
->single
= (*entry
)->single
;
472 e
->multi
= (*entry
)->multi
;
473 e
->ptype
= (*entry
)->ptype
;
474 e
->refcount
= (*entry
)->refcount
;
476 hlist_add_before(&e
->hlist
, &(*entry
)->hlist
);
477 hlist_del(&(*entry
)->hlist
);
478 /* Make sure the call_rcu has been executed */
479 if ((*entry
)->rcu_pending
)
483 trace_mark(core_marker_format
, "name %s format %s",
489 * Sets the probe callback corresponding to one marker.
491 static int set_marker(struct marker_entry
**entry
, struct marker
*elem
,
495 WARN_ON(strcmp((*entry
)->name
, elem
->name
) != 0);
497 if ((*entry
)->format
) {
498 if (strcmp((*entry
)->format
, elem
->format
) != 0) {
500 "Format mismatch for probe %s "
501 "(%s), marker (%s)\n",
508 ret
= marker_set_format(entry
, elem
->format
);
514 * probe_cb setup (statically known) is done here. It is
515 * asynchronous with the rest of execution, therefore we only
516 * pass from a "safe" callback (with argument) to an "unsafe"
517 * callback (does not set arguments).
519 elem
->call
= (*entry
)->call
;
522 * We only update the single probe private data when the ptr is
523 * set to a _non_ single probe! (0 -> 1 and N -> 1, N != 1)
525 WARN_ON(elem
->single
.func
!= __mark_empty_function
526 && elem
->single
.probe_private
527 != (*entry
)->single
.probe_private
&&
529 elem
->single
.probe_private
= (*entry
)->single
.probe_private
;
531 * Make sure the private data is valid when we update the
535 elem
->single
.func
= (*entry
)->single
.func
;
537 * We also make sure that the new probe callbacks array is consistent
538 * before setting a pointer to it.
540 rcu_assign_pointer(elem
->multi
, (*entry
)->multi
);
542 * Update the function or multi probe array pointer before setting the
546 elem
->ptype
= (*entry
)->ptype
;
547 elem
->state
= active
;
553 * Disable a marker and its probe callback.
554 * Note: only waiting an RCU period after setting elem->call to the empty
555 * function insures that the original callback is not used anymore. This insured
556 * by preempt_disable around the call site.
558 static void disable_marker(struct marker
*elem
)
560 /* leave "call" as is. It is known statically. */
562 elem
->single
.func
= __mark_empty_function
;
563 /* Update the function before setting the ptype */
565 elem
->ptype
= 0; /* single probe */
567 * Leave the private data and id there, because removal is racy and
568 * should be done only after an RCU period. These are never used until
569 * the next initialization anyway.
574 * marker_update_probe_range - Update a probe range
575 * @begin: beginning of the range
576 * @end: end of the range
578 * Updates the probe callback corresponding to a range of markers.
580 void marker_update_probe_range(struct marker
*begin
,
584 struct marker_entry
*mark_entry
;
586 mutex_lock(&markers_mutex
);
587 for (iter
= begin
; iter
< end
; iter
++) {
588 mark_entry
= get_marker(iter
->name
);
590 set_marker(&mark_entry
, iter
,
591 !!mark_entry
->refcount
);
593 * ignore error, continue
596 disable_marker(iter
);
599 mutex_unlock(&markers_mutex
);
603 * Update probes, removing the faulty probes.
605 * Internal callback only changed before the first probe is connected to it.
606 * Single probe private data can only be changed on 0 -> 1 and 2 -> 1
607 * transitions. All other transitions will leave the old private data valid.
608 * This makes the non-atomicity of the callback/private data updates valid.
610 * "special case" updates :
615 * Other updates all behave the same, just like the 2 -> 3 or 3 -> 2 updates.
616 * Site effect : marker_set_format may delete the marker entry (creating a
619 static void marker_update_probes(void)
621 /* Core kernel markers */
622 marker_update_probe_range(__start___markers
, __stop___markers
);
623 /* Markers in modules. */
624 module_update_markers();
628 * marker_probe_register - Connect a probe to a marker
630 * @format: format string
631 * @probe: probe handler
632 * @probe_private: probe private data
634 * private data must be a valid allocated memory address, or NULL.
635 * Returns 0 if ok, error value on error.
636 * The probe address must at least be aligned on the architecture pointer size.
638 int marker_probe_register(const char *name
, const char *format
,
639 marker_probe_func
*probe
, void *probe_private
)
641 struct marker_entry
*entry
;
643 struct marker_probe_closure
*old
;
645 mutex_lock(&markers_mutex
);
646 entry
= get_marker(name
);
648 entry
= add_marker(name
, format
);
650 ret
= PTR_ERR(entry
);
655 * If we detect that a call_rcu is pending for this marker,
656 * make sure it's executed now.
658 if (entry
->rcu_pending
)
660 old
= marker_entry_add_probe(entry
, probe
, probe_private
);
665 mutex_unlock(&markers_mutex
);
666 marker_update_probes(); /* may update entry */
667 mutex_lock(&markers_mutex
);
668 entry
= get_marker(name
);
671 entry
->rcu_pending
= 1;
672 /* write rcu_pending before calling the RCU callback */
674 #ifdef CONFIG_PREEMPT_RCU
675 synchronize_sched(); /* Until we have the call_rcu_sched() */
677 call_rcu(&entry
->rcu
, free_old_closure
);
679 mutex_unlock(&markers_mutex
);
682 EXPORT_SYMBOL_GPL(marker_probe_register
);
685 * marker_probe_unregister - Disconnect a probe from a marker
687 * @probe: probe function pointer
688 * @probe_private: probe private data
690 * Returns the private data given to marker_probe_register, or an ERR_PTR().
691 * We do not need to call a synchronize_sched to make sure the probes have
692 * finished running before doing a module unload, because the module unload
693 * itself uses stop_machine(), which insures that every preempt disabled section
696 int marker_probe_unregister(const char *name
,
697 marker_probe_func
*probe
, void *probe_private
)
699 struct marker_entry
*entry
;
700 struct marker_probe_closure
*old
;
703 mutex_lock(&markers_mutex
);
704 entry
= get_marker(name
);
707 if (entry
->rcu_pending
)
709 old
= marker_entry_remove_probe(entry
, probe
, probe_private
);
710 mutex_unlock(&markers_mutex
);
711 marker_update_probes(); /* may update entry */
712 mutex_lock(&markers_mutex
);
713 entry
= get_marker(name
);
717 entry
->rcu_pending
= 1;
718 /* write rcu_pending before calling the RCU callback */
720 #ifdef CONFIG_PREEMPT_RCU
721 synchronize_sched(); /* Until we have the call_rcu_sched() */
723 call_rcu(&entry
->rcu
, free_old_closure
);
724 remove_marker(name
); /* Ignore busy error message */
727 mutex_unlock(&markers_mutex
);
730 EXPORT_SYMBOL_GPL(marker_probe_unregister
);
732 static struct marker_entry
*
733 get_marker_from_private_data(marker_probe_func
*probe
, void *probe_private
)
735 struct marker_entry
*entry
;
737 struct hlist_head
*head
;
738 struct hlist_node
*node
;
740 for (i
= 0; i
< MARKER_TABLE_SIZE
; i
++) {
741 head
= &marker_table
[i
];
742 hlist_for_each_entry(entry
, node
, head
, hlist
) {
744 if (entry
->single
.func
== probe
745 && entry
->single
.probe_private
749 struct marker_probe_closure
*closure
;
750 closure
= entry
->multi
;
751 for (i
= 0; closure
[i
].func
; i
++) {
752 if (closure
[i
].func
== probe
&&
753 closure
[i
].probe_private
764 * marker_probe_unregister_private_data - Disconnect a probe from a marker
765 * @probe: probe function
766 * @probe_private: probe private data
768 * Unregister a probe by providing the registered private data.
769 * Only removes the first marker found in hash table.
770 * Return 0 on success or error value.
771 * We do not need to call a synchronize_sched to make sure the probes have
772 * finished running before doing a module unload, because the module unload
773 * itself uses stop_machine(), which insures that every preempt disabled section
776 int marker_probe_unregister_private_data(marker_probe_func
*probe
,
779 struct marker_entry
*entry
;
781 struct marker_probe_closure
*old
;
783 mutex_lock(&markers_mutex
);
784 entry
= get_marker_from_private_data(probe
, probe_private
);
789 if (entry
->rcu_pending
)
791 old
= marker_entry_remove_probe(entry
, NULL
, probe_private
);
792 mutex_unlock(&markers_mutex
);
793 marker_update_probes(); /* may update entry */
794 mutex_lock(&markers_mutex
);
795 entry
= get_marker_from_private_data(probe
, probe_private
);
798 entry
->rcu_pending
= 1;
799 /* write rcu_pending before calling the RCU callback */
801 #ifdef CONFIG_PREEMPT_RCU
802 synchronize_sched(); /* Until we have the call_rcu_sched() */
804 call_rcu(&entry
->rcu
, free_old_closure
);
805 remove_marker(entry
->name
); /* Ignore busy error message */
807 mutex_unlock(&markers_mutex
);
810 EXPORT_SYMBOL_GPL(marker_probe_unregister_private_data
);
813 * marker_get_private_data - Get a marker's probe private data
815 * @probe: probe to match
816 * @num: get the nth matching probe's private data
818 * Returns the nth private data pointer (starting from 0) matching, or an
820 * Returns the private data pointer, or an ERR_PTR.
821 * The private data pointer should _only_ be dereferenced if the caller is the
822 * owner of the data, or its content could vanish. This is mostly used to
823 * confirm that a caller is the owner of a registered probe.
825 void *marker_get_private_data(const char *name
, marker_probe_func
*probe
,
828 struct hlist_head
*head
;
829 struct hlist_node
*node
;
830 struct marker_entry
*e
;
831 size_t name_len
= strlen(name
) + 1;
832 u32 hash
= jhash(name
, name_len
-1, 0);
835 head
= &marker_table
[hash
& ((1 << MARKER_HASH_BITS
)-1)];
836 hlist_for_each_entry(e
, node
, head
, hlist
) {
837 if (!strcmp(name
, e
->name
)) {
839 if (num
== 0 && e
->single
.func
== probe
)
840 return e
->single
.probe_private
;
844 struct marker_probe_closure
*closure
;
847 for (i
= 0; closure
[i
].func
; i
++) {
848 if (closure
[i
].func
!= probe
)
851 return closure
[i
].probe_private
;
856 return ERR_PTR(-ENOENT
);
858 EXPORT_SYMBOL_GPL(marker_get_private_data
);