kprobes: Introduce kprobes jump optimization
[linux-2.6.git] / kernel / kprobes.c
1 /*
2  *  Kernel Probes (KProbes)
3  *  kernel/kprobes.c
4  *
5  * This program is free software; you can redistribute it and/or modify
6  * it under the terms of the GNU General Public License as published by
7  * the Free Software Foundation; either version 2 of the License, or
8  * (at your option) any later version.
9  *
10  * This program is distributed in the hope that it will be useful,
11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13  * GNU General Public License for more details.
14  *
15  * You should have received a copy of the GNU General Public License
16  * along with this program; if not, write to the Free Software
17  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18  *
19  * Copyright (C) IBM Corporation, 2002, 2004
20  *
21  * 2002-Oct     Created by Vamsi Krishna S <vamsi_krishna@in.ibm.com> Kernel
22  *              Probes initial implementation (includes suggestions from
23  *              Rusty Russell).
24  * 2004-Aug     Updated by Prasanna S Panchamukhi <prasanna@in.ibm.com> with
25  *              hlists and exceptions notifier as suggested by Andi Kleen.
26  * 2004-July    Suparna Bhattacharya <suparna@in.ibm.com> added jumper probes
27  *              interface to access function arguments.
28  * 2004-Sep     Prasanna S Panchamukhi <prasanna@in.ibm.com> Changed Kprobes
29  *              exceptions notifier to be first on the priority list.
30  * 2005-May     Hien Nguyen <hien@us.ibm.com>, Jim Keniston
31  *              <jkenisto@us.ibm.com> and Prasanna S Panchamukhi
32  *              <prasanna@in.ibm.com> added function-return probes.
33  */
34 #include <linux/kprobes.h>
35 #include <linux/hash.h>
36 #include <linux/init.h>
37 #include <linux/slab.h>
38 #include <linux/stddef.h>
39 #include <linux/module.h>
40 #include <linux/moduleloader.h>
41 #include <linux/kallsyms.h>
42 #include <linux/freezer.h>
43 #include <linux/seq_file.h>
44 #include <linux/debugfs.h>
45 #include <linux/kdebug.h>
46 #include <linux/memory.h>
47 #include <linux/ftrace.h>
48 #include <linux/cpu.h>
49
50 #include <asm-generic/sections.h>
51 #include <asm/cacheflush.h>
52 #include <asm/errno.h>
53 #include <asm/uaccess.h>
54
55 #define KPROBE_HASH_BITS 6
56 #define KPROBE_TABLE_SIZE (1 << KPROBE_HASH_BITS)
57
58
59 /*
60  * Some oddball architectures like 64bit powerpc have function descriptors
61  * so this must be overridable.
62  */
63 #ifndef kprobe_lookup_name
64 #define kprobe_lookup_name(name, addr) \
65         addr = ((kprobe_opcode_t *)(kallsyms_lookup_name(name)))
66 #endif
67
68 static int kprobes_initialized;
69 static struct hlist_head kprobe_table[KPROBE_TABLE_SIZE];
70 static struct hlist_head kretprobe_inst_table[KPROBE_TABLE_SIZE];
71
72 /* NOTE: change this value only with kprobe_mutex held */
73 static bool kprobes_all_disarmed;
74
75 static DEFINE_MUTEX(kprobe_mutex);      /* Protects kprobe_table */
76 static DEFINE_PER_CPU(struct kprobe *, kprobe_instance) = NULL;
77 static struct {
78         spinlock_t lock ____cacheline_aligned_in_smp;
79 } kretprobe_table_locks[KPROBE_TABLE_SIZE];
80
81 static spinlock_t *kretprobe_table_lock_ptr(unsigned long hash)
82 {
83         return &(kretprobe_table_locks[hash].lock);
84 }
85
86 /*
87  * Normally, functions that we'd want to prohibit kprobes in, are marked
88  * __kprobes. But, there are cases where such functions already belong to
89  * a different section (__sched for preempt_schedule)
90  *
91  * For such cases, we now have a blacklist
92  */
93 static struct kprobe_blackpoint kprobe_blacklist[] = {
94         {"preempt_schedule",},
95         {"native_get_debugreg",},
96         {"irq_entries_start",},
97         {"common_interrupt",},
98         {"mcount",},    /* mcount can be called from everywhere */
99         {NULL}    /* Terminator */
100 };
101
102 #ifdef __ARCH_WANT_KPROBES_INSN_SLOT
103 /*
104  * kprobe->ainsn.insn points to the copy of the instruction to be
105  * single-stepped. x86_64, POWER4 and above have no-exec support and
106  * stepping on the instruction on a vmalloced/kmalloced/data page
107  * is a recipe for disaster
108  */
109 struct kprobe_insn_page {
110         struct list_head list;
111         kprobe_opcode_t *insns;         /* Page of instruction slots */
112         int nused;
113         int ngarbage;
114         char slot_used[];
115 };
116
117 #define KPROBE_INSN_PAGE_SIZE(slots)                    \
118         (offsetof(struct kprobe_insn_page, slot_used) + \
119          (sizeof(char) * (slots)))
120
121 struct kprobe_insn_cache {
122         struct list_head pages; /* list of kprobe_insn_page */
123         size_t insn_size;       /* size of instruction slot */
124         int nr_garbage;
125 };
126
127 static int slots_per_page(struct kprobe_insn_cache *c)
128 {
129         return PAGE_SIZE/(c->insn_size * sizeof(kprobe_opcode_t));
130 }
131
132 enum kprobe_slot_state {
133         SLOT_CLEAN = 0,
134         SLOT_DIRTY = 1,
135         SLOT_USED = 2,
136 };
137
138 static DEFINE_MUTEX(kprobe_insn_mutex); /* Protects kprobe_insn_slots */
139 static struct kprobe_insn_cache kprobe_insn_slots = {
140         .pages = LIST_HEAD_INIT(kprobe_insn_slots.pages),
141         .insn_size = MAX_INSN_SIZE,
142         .nr_garbage = 0,
143 };
144 static int __kprobes collect_garbage_slots(struct kprobe_insn_cache *c);
145
146 /**
147  * __get_insn_slot() - Find a slot on an executable page for an instruction.
148  * We allocate an executable page if there's no room on existing ones.
149  */
150 static kprobe_opcode_t __kprobes *__get_insn_slot(struct kprobe_insn_cache *c)
151 {
152         struct kprobe_insn_page *kip;
153
154  retry:
155         list_for_each_entry(kip, &c->pages, list) {
156                 if (kip->nused < slots_per_page(c)) {
157                         int i;
158                         for (i = 0; i < slots_per_page(c); i++) {
159                                 if (kip->slot_used[i] == SLOT_CLEAN) {
160                                         kip->slot_used[i] = SLOT_USED;
161                                         kip->nused++;
162                                         return kip->insns + (i * c->insn_size);
163                                 }
164                         }
165                         /* kip->nused is broken. Fix it. */
166                         kip->nused = slots_per_page(c);
167                         WARN_ON(1);
168                 }
169         }
170
171         /* If there are any garbage slots, collect it and try again. */
172         if (c->nr_garbage && collect_garbage_slots(c) == 0)
173                 goto retry;
174
175         /* All out of space.  Need to allocate a new page. */
176         kip = kmalloc(KPROBE_INSN_PAGE_SIZE(slots_per_page(c)), GFP_KERNEL);
177         if (!kip)
178                 return NULL;
179
180         /*
181          * Use module_alloc so this page is within +/- 2GB of where the
182          * kernel image and loaded module images reside. This is required
183          * so x86_64 can correctly handle the %rip-relative fixups.
184          */
185         kip->insns = module_alloc(PAGE_SIZE);
186         if (!kip->insns) {
187                 kfree(kip);
188                 return NULL;
189         }
190         INIT_LIST_HEAD(&kip->list);
191         memset(kip->slot_used, SLOT_CLEAN, slots_per_page(c));
192         kip->slot_used[0] = SLOT_USED;
193         kip->nused = 1;
194         kip->ngarbage = 0;
195         list_add(&kip->list, &c->pages);
196         return kip->insns;
197 }
198
199
200 kprobe_opcode_t __kprobes *get_insn_slot(void)
201 {
202         kprobe_opcode_t *ret = NULL;
203
204         mutex_lock(&kprobe_insn_mutex);
205         ret = __get_insn_slot(&kprobe_insn_slots);
206         mutex_unlock(&kprobe_insn_mutex);
207
208         return ret;
209 }
210
211 /* Return 1 if all garbages are collected, otherwise 0. */
212 static int __kprobes collect_one_slot(struct kprobe_insn_page *kip, int idx)
213 {
214         kip->slot_used[idx] = SLOT_CLEAN;
215         kip->nused--;
216         if (kip->nused == 0) {
217                 /*
218                  * Page is no longer in use.  Free it unless
219                  * it's the last one.  We keep the last one
220                  * so as not to have to set it up again the
221                  * next time somebody inserts a probe.
222                  */
223                 if (!list_is_singular(&kip->list)) {
224                         list_del(&kip->list);
225                         module_free(NULL, kip->insns);
226                         kfree(kip);
227                 }
228                 return 1;
229         }
230         return 0;
231 }
232
233 static int __kprobes collect_garbage_slots(struct kprobe_insn_cache *c)
234 {
235         struct kprobe_insn_page *kip, *next;
236
237         /* Ensure no-one is interrupted on the garbages */
238         synchronize_sched();
239
240         list_for_each_entry_safe(kip, next, &c->pages, list) {
241                 int i;
242                 if (kip->ngarbage == 0)
243                         continue;
244                 kip->ngarbage = 0;      /* we will collect all garbages */
245                 for (i = 0; i < slots_per_page(c); i++) {
246                         if (kip->slot_used[i] == SLOT_DIRTY &&
247                             collect_one_slot(kip, i))
248                                 break;
249                 }
250         }
251         c->nr_garbage = 0;
252         return 0;
253 }
254
255 static void __kprobes __free_insn_slot(struct kprobe_insn_cache *c,
256                                        kprobe_opcode_t *slot, int dirty)
257 {
258         struct kprobe_insn_page *kip;
259
260         list_for_each_entry(kip, &c->pages, list) {
261                 long idx = ((long)slot - (long)kip->insns) / c->insn_size;
262                 if (idx >= 0 && idx < slots_per_page(c)) {
263                         WARN_ON(kip->slot_used[idx] != SLOT_USED);
264                         if (dirty) {
265                                 kip->slot_used[idx] = SLOT_DIRTY;
266                                 kip->ngarbage++;
267                                 if (++c->nr_garbage > slots_per_page(c))
268                                         collect_garbage_slots(c);
269                         } else
270                                 collect_one_slot(kip, idx);
271                         return;
272                 }
273         }
274         /* Could not free this slot. */
275         WARN_ON(1);
276 }
277
278 void __kprobes free_insn_slot(kprobe_opcode_t * slot, int dirty)
279 {
280         mutex_lock(&kprobe_insn_mutex);
281         __free_insn_slot(&kprobe_insn_slots, slot, dirty);
282         mutex_unlock(&kprobe_insn_mutex);
283 }
284 #ifdef CONFIG_OPTPROBES
285 /* For optimized_kprobe buffer */
286 static DEFINE_MUTEX(kprobe_optinsn_mutex); /* Protects kprobe_optinsn_slots */
287 static struct kprobe_insn_cache kprobe_optinsn_slots = {
288         .pages = LIST_HEAD_INIT(kprobe_optinsn_slots.pages),
289         /* .insn_size is initialized later */
290         .nr_garbage = 0,
291 };
292 /* Get a slot for optimized_kprobe buffer */
293 kprobe_opcode_t __kprobes *get_optinsn_slot(void)
294 {
295         kprobe_opcode_t *ret = NULL;
296
297         mutex_lock(&kprobe_optinsn_mutex);
298         ret = __get_insn_slot(&kprobe_optinsn_slots);
299         mutex_unlock(&kprobe_optinsn_mutex);
300
301         return ret;
302 }
303
304 void __kprobes free_optinsn_slot(kprobe_opcode_t * slot, int dirty)
305 {
306         mutex_lock(&kprobe_optinsn_mutex);
307         __free_insn_slot(&kprobe_optinsn_slots, slot, dirty);
308         mutex_unlock(&kprobe_optinsn_mutex);
309 }
310 #endif
311 #endif
312
313 /* We have preemption disabled.. so it is safe to use __ versions */
314 static inline void set_kprobe_instance(struct kprobe *kp)
315 {
316         __get_cpu_var(kprobe_instance) = kp;
317 }
318
319 static inline void reset_kprobe_instance(void)
320 {
321         __get_cpu_var(kprobe_instance) = NULL;
322 }
323
324 /*
325  * This routine is called either:
326  *      - under the kprobe_mutex - during kprobe_[un]register()
327  *                              OR
328  *      - with preemption disabled - from arch/xxx/kernel/kprobes.c
329  */
330 struct kprobe __kprobes *get_kprobe(void *addr)
331 {
332         struct hlist_head *head;
333         struct hlist_node *node;
334         struct kprobe *p;
335
336         head = &kprobe_table[hash_ptr(addr, KPROBE_HASH_BITS)];
337         hlist_for_each_entry_rcu(p, node, head, hlist) {
338                 if (p->addr == addr)
339                         return p;
340         }
341
342         return NULL;
343 }
344
345 static int __kprobes aggr_pre_handler(struct kprobe *p, struct pt_regs *regs);
346
347 /* Return true if the kprobe is an aggregator */
348 static inline int kprobe_aggrprobe(struct kprobe *p)
349 {
350         return p->pre_handler == aggr_pre_handler;
351 }
352
353 /*
354  * Keep all fields in the kprobe consistent
355  */
356 static inline void copy_kprobe(struct kprobe *old_p, struct kprobe *p)
357 {
358         memcpy(&p->opcode, &old_p->opcode, sizeof(kprobe_opcode_t));
359         memcpy(&p->ainsn, &old_p->ainsn, sizeof(struct arch_specific_insn));
360 }
361
362 #ifdef CONFIG_OPTPROBES
363 /*
364  * Call all pre_handler on the list, but ignores its return value.
365  * This must be called from arch-dep optimized caller.
366  */
367 void __kprobes opt_pre_handler(struct kprobe *p, struct pt_regs *regs)
368 {
369         struct kprobe *kp;
370
371         list_for_each_entry_rcu(kp, &p->list, list) {
372                 if (kp->pre_handler && likely(!kprobe_disabled(kp))) {
373                         set_kprobe_instance(kp);
374                         kp->pre_handler(kp, regs);
375                 }
376                 reset_kprobe_instance();
377         }
378 }
379
380 /* Return true(!0) if the kprobe is ready for optimization. */
381 static inline int kprobe_optready(struct kprobe *p)
382 {
383         struct optimized_kprobe *op;
384
385         if (kprobe_aggrprobe(p)) {
386                 op = container_of(p, struct optimized_kprobe, kp);
387                 return arch_prepared_optinsn(&op->optinsn);
388         }
389
390         return 0;
391 }
392
393 /*
394  * Return an optimized kprobe whose optimizing code replaces
395  * instructions including addr (exclude breakpoint).
396  */
397 struct kprobe *__kprobes get_optimized_kprobe(unsigned long addr)
398 {
399         int i;
400         struct kprobe *p = NULL;
401         struct optimized_kprobe *op;
402
403         /* Don't check i == 0, since that is a breakpoint case. */
404         for (i = 1; !p && i < MAX_OPTIMIZED_LENGTH; i++)
405                 p = get_kprobe((void *)(addr - i));
406
407         if (p && kprobe_optready(p)) {
408                 op = container_of(p, struct optimized_kprobe, kp);
409                 if (arch_within_optimized_kprobe(op, addr))
410                         return p;
411         }
412
413         return NULL;
414 }
415
416 /* Optimization staging list, protected by kprobe_mutex */
417 static LIST_HEAD(optimizing_list);
418
419 static void kprobe_optimizer(struct work_struct *work);
420 static DECLARE_DELAYED_WORK(optimizing_work, kprobe_optimizer);
421 #define OPTIMIZE_DELAY 5
422
423 /* Kprobe jump optimizer */
424 static __kprobes void kprobe_optimizer(struct work_struct *work)
425 {
426         struct optimized_kprobe *op, *tmp;
427
428         /* Lock modules while optimizing kprobes */
429         mutex_lock(&module_mutex);
430         mutex_lock(&kprobe_mutex);
431         if (kprobes_all_disarmed)
432                 goto end;
433
434         /*
435          * Wait for quiesence period to ensure all running interrupts
436          * are done. Because optprobe may modify multiple instructions
437          * there is a chance that Nth instruction is interrupted. In that
438          * case, running interrupt can return to 2nd-Nth byte of jump
439          * instruction. This wait is for avoiding it.
440          */
441         synchronize_sched();
442
443         /*
444          * The optimization/unoptimization refers online_cpus via
445          * stop_machine() and cpu-hotplug modifies online_cpus.
446          * And same time, text_mutex will be held in cpu-hotplug and here.
447          * This combination can cause a deadlock (cpu-hotplug try to lock
448          * text_mutex but stop_machine can not be done because online_cpus
449          * has been changed)
450          * To avoid this deadlock, we need to call get_online_cpus()
451          * for preventing cpu-hotplug outside of text_mutex locking.
452          */
453         get_online_cpus();
454         mutex_lock(&text_mutex);
455         list_for_each_entry_safe(op, tmp, &optimizing_list, list) {
456                 WARN_ON(kprobe_disabled(&op->kp));
457                 if (arch_optimize_kprobe(op) < 0)
458                         op->kp.flags &= ~KPROBE_FLAG_OPTIMIZED;
459                 list_del_init(&op->list);
460         }
461         mutex_unlock(&text_mutex);
462         put_online_cpus();
463 end:
464         mutex_unlock(&kprobe_mutex);
465         mutex_unlock(&module_mutex);
466 }
467
468 /* Optimize kprobe if p is ready to be optimized */
469 static __kprobes void optimize_kprobe(struct kprobe *p)
470 {
471         struct optimized_kprobe *op;
472
473         /* Check if the kprobe is disabled or not ready for optimization. */
474         if (!kprobe_optready(p) ||
475             (kprobe_disabled(p) || kprobes_all_disarmed))
476                 return;
477
478         /* Both of break_handler and post_handler are not supported. */
479         if (p->break_handler || p->post_handler)
480                 return;
481
482         op = container_of(p, struct optimized_kprobe, kp);
483
484         /* Check there is no other kprobes at the optimized instructions */
485         if (arch_check_optimized_kprobe(op) < 0)
486                 return;
487
488         /* Check if it is already optimized. */
489         if (op->kp.flags & KPROBE_FLAG_OPTIMIZED)
490                 return;
491
492         op->kp.flags |= KPROBE_FLAG_OPTIMIZED;
493         list_add(&op->list, &optimizing_list);
494         if (!delayed_work_pending(&optimizing_work))
495                 schedule_delayed_work(&optimizing_work, OPTIMIZE_DELAY);
496 }
497
498 /* Unoptimize a kprobe if p is optimized */
499 static __kprobes void unoptimize_kprobe(struct kprobe *p)
500 {
501         struct optimized_kprobe *op;
502
503         if ((p->flags & KPROBE_FLAG_OPTIMIZED) && kprobe_aggrprobe(p)) {
504                 op = container_of(p, struct optimized_kprobe, kp);
505                 if (!list_empty(&op->list))
506                         /* Dequeue from the optimization queue */
507                         list_del_init(&op->list);
508                 else
509                         /* Replace jump with break */
510                         arch_unoptimize_kprobe(op);
511                 op->kp.flags &= ~KPROBE_FLAG_OPTIMIZED;
512         }
513 }
514
515 /* Remove optimized instructions */
516 static void __kprobes kill_optimized_kprobe(struct kprobe *p)
517 {
518         struct optimized_kprobe *op;
519
520         op = container_of(p, struct optimized_kprobe, kp);
521         if (!list_empty(&op->list)) {
522                 /* Dequeue from the optimization queue */
523                 list_del_init(&op->list);
524                 op->kp.flags &= ~KPROBE_FLAG_OPTIMIZED;
525         }
526         /* Don't unoptimize, because the target code will be freed. */
527         arch_remove_optimized_kprobe(op);
528 }
529
530 /* Try to prepare optimized instructions */
531 static __kprobes void prepare_optimized_kprobe(struct kprobe *p)
532 {
533         struct optimized_kprobe *op;
534
535         op = container_of(p, struct optimized_kprobe, kp);
536         arch_prepare_optimized_kprobe(op);
537 }
538
539 /* Free optimized instructions and optimized_kprobe */
540 static __kprobes void free_aggr_kprobe(struct kprobe *p)
541 {
542         struct optimized_kprobe *op;
543
544         op = container_of(p, struct optimized_kprobe, kp);
545         arch_remove_optimized_kprobe(op);
546         kfree(op);
547 }
548
549 /* Allocate new optimized_kprobe and try to prepare optimized instructions */
550 static __kprobes struct kprobe *alloc_aggr_kprobe(struct kprobe *p)
551 {
552         struct optimized_kprobe *op;
553
554         op = kzalloc(sizeof(struct optimized_kprobe), GFP_KERNEL);
555         if (!op)
556                 return NULL;
557
558         INIT_LIST_HEAD(&op->list);
559         op->kp.addr = p->addr;
560         arch_prepare_optimized_kprobe(op);
561
562         return &op->kp;
563 }
564
565 static void __kprobes init_aggr_kprobe(struct kprobe *ap, struct kprobe *p);
566
567 /*
568  * Prepare an optimized_kprobe and optimize it
569  * NOTE: p must be a normal registered kprobe
570  */
571 static __kprobes void try_to_optimize_kprobe(struct kprobe *p)
572 {
573         struct kprobe *ap;
574         struct optimized_kprobe *op;
575
576         ap = alloc_aggr_kprobe(p);
577         if (!ap)
578                 return;
579
580         op = container_of(ap, struct optimized_kprobe, kp);
581         if (!arch_prepared_optinsn(&op->optinsn)) {
582                 /* If failed to setup optimizing, fallback to kprobe */
583                 free_aggr_kprobe(ap);
584                 return;
585         }
586
587         init_aggr_kprobe(ap, p);
588         optimize_kprobe(ap);
589 }
590
591 static void __kprobes __arm_kprobe(struct kprobe *p)
592 {
593         struct kprobe *old_p;
594
595         /* Check collision with other optimized kprobes */
596         old_p = get_optimized_kprobe((unsigned long)p->addr);
597         if (unlikely(old_p))
598                 unoptimize_kprobe(old_p); /* Fallback to unoptimized kprobe */
599
600         arch_arm_kprobe(p);
601         optimize_kprobe(p);     /* Try to optimize (add kprobe to a list) */
602 }
603
604 static void __kprobes __disarm_kprobe(struct kprobe *p)
605 {
606         struct kprobe *old_p;
607
608         unoptimize_kprobe(p);   /* Try to unoptimize */
609         arch_disarm_kprobe(p);
610
611         /* If another kprobe was blocked, optimize it. */
612         old_p = get_optimized_kprobe((unsigned long)p->addr);
613         if (unlikely(old_p))
614                 optimize_kprobe(old_p);
615 }
616
617 #else /* !CONFIG_OPTPROBES */
618
619 #define optimize_kprobe(p)                      do {} while (0)
620 #define unoptimize_kprobe(p)                    do {} while (0)
621 #define kill_optimized_kprobe(p)                do {} while (0)
622 #define prepare_optimized_kprobe(p)             do {} while (0)
623 #define try_to_optimize_kprobe(p)               do {} while (0)
624 #define __arm_kprobe(p)                         arch_arm_kprobe(p)
625 #define __disarm_kprobe(p)                      arch_disarm_kprobe(p)
626
627 static __kprobes void free_aggr_kprobe(struct kprobe *p)
628 {
629         kfree(p);
630 }
631
632 static __kprobes struct kprobe *alloc_aggr_kprobe(struct kprobe *p)
633 {
634         return kzalloc(sizeof(struct kprobe), GFP_KERNEL);
635 }
636 #endif /* CONFIG_OPTPROBES */
637
638 /* Arm a kprobe with text_mutex */
639 static void __kprobes arm_kprobe(struct kprobe *kp)
640 {
641         /*
642          * Here, since __arm_kprobe() doesn't use stop_machine(),
643          * this doesn't cause deadlock on text_mutex. So, we don't
644          * need get_online_cpus().
645          */
646         mutex_lock(&text_mutex);
647         __arm_kprobe(kp);
648         mutex_unlock(&text_mutex);
649 }
650
651 /* Disarm a kprobe with text_mutex */
652 static void __kprobes disarm_kprobe(struct kprobe *kp)
653 {
654         get_online_cpus();      /* For avoiding text_mutex deadlock */
655         mutex_lock(&text_mutex);
656         __disarm_kprobe(kp);
657         mutex_unlock(&text_mutex);
658         put_online_cpus();
659 }
660
661 /*
662  * Aggregate handlers for multiple kprobes support - these handlers
663  * take care of invoking the individual kprobe handlers on p->list
664  */
665 static int __kprobes aggr_pre_handler(struct kprobe *p, struct pt_regs *regs)
666 {
667         struct kprobe *kp;
668
669         list_for_each_entry_rcu(kp, &p->list, list) {
670                 if (kp->pre_handler && likely(!kprobe_disabled(kp))) {
671                         set_kprobe_instance(kp);
672                         if (kp->pre_handler(kp, regs))
673                                 return 1;
674                 }
675                 reset_kprobe_instance();
676         }
677         return 0;
678 }
679
680 static void __kprobes aggr_post_handler(struct kprobe *p, struct pt_regs *regs,
681                                         unsigned long flags)
682 {
683         struct kprobe *kp;
684
685         list_for_each_entry_rcu(kp, &p->list, list) {
686                 if (kp->post_handler && likely(!kprobe_disabled(kp))) {
687                         set_kprobe_instance(kp);
688                         kp->post_handler(kp, regs, flags);
689                         reset_kprobe_instance();
690                 }
691         }
692 }
693
694 static int __kprobes aggr_fault_handler(struct kprobe *p, struct pt_regs *regs,
695                                         int trapnr)
696 {
697         struct kprobe *cur = __get_cpu_var(kprobe_instance);
698
699         /*
700          * if we faulted "during" the execution of a user specified
701          * probe handler, invoke just that probe's fault handler
702          */
703         if (cur && cur->fault_handler) {
704                 if (cur->fault_handler(cur, regs, trapnr))
705                         return 1;
706         }
707         return 0;
708 }
709
710 static int __kprobes aggr_break_handler(struct kprobe *p, struct pt_regs *regs)
711 {
712         struct kprobe *cur = __get_cpu_var(kprobe_instance);
713         int ret = 0;
714
715         if (cur && cur->break_handler) {
716                 if (cur->break_handler(cur, regs))
717                         ret = 1;
718         }
719         reset_kprobe_instance();
720         return ret;
721 }
722
723 /* Walks the list and increments nmissed count for multiprobe case */
724 void __kprobes kprobes_inc_nmissed_count(struct kprobe *p)
725 {
726         struct kprobe *kp;
727         if (!kprobe_aggrprobe(p)) {
728                 p->nmissed++;
729         } else {
730                 list_for_each_entry_rcu(kp, &p->list, list)
731                         kp->nmissed++;
732         }
733         return;
734 }
735
736 void __kprobes recycle_rp_inst(struct kretprobe_instance *ri,
737                                 struct hlist_head *head)
738 {
739         struct kretprobe *rp = ri->rp;
740
741         /* remove rp inst off the rprobe_inst_table */
742         hlist_del(&ri->hlist);
743         INIT_HLIST_NODE(&ri->hlist);
744         if (likely(rp)) {
745                 spin_lock(&rp->lock);
746                 hlist_add_head(&ri->hlist, &rp->free_instances);
747                 spin_unlock(&rp->lock);
748         } else
749                 /* Unregistering */
750                 hlist_add_head(&ri->hlist, head);
751 }
752
753 void __kprobes kretprobe_hash_lock(struct task_struct *tsk,
754                          struct hlist_head **head, unsigned long *flags)
755 {
756         unsigned long hash = hash_ptr(tsk, KPROBE_HASH_BITS);
757         spinlock_t *hlist_lock;
758
759         *head = &kretprobe_inst_table[hash];
760         hlist_lock = kretprobe_table_lock_ptr(hash);
761         spin_lock_irqsave(hlist_lock, *flags);
762 }
763
764 static void __kprobes kretprobe_table_lock(unsigned long hash,
765         unsigned long *flags)
766 {
767         spinlock_t *hlist_lock = kretprobe_table_lock_ptr(hash);
768         spin_lock_irqsave(hlist_lock, *flags);
769 }
770
771 void __kprobes kretprobe_hash_unlock(struct task_struct *tsk,
772         unsigned long *flags)
773 {
774         unsigned long hash = hash_ptr(tsk, KPROBE_HASH_BITS);
775         spinlock_t *hlist_lock;
776
777         hlist_lock = kretprobe_table_lock_ptr(hash);
778         spin_unlock_irqrestore(hlist_lock, *flags);
779 }
780
781 void __kprobes kretprobe_table_unlock(unsigned long hash, unsigned long *flags)
782 {
783         spinlock_t *hlist_lock = kretprobe_table_lock_ptr(hash);
784         spin_unlock_irqrestore(hlist_lock, *flags);
785 }
786
787 /*
788  * This function is called from finish_task_switch when task tk becomes dead,
789  * so that we can recycle any function-return probe instances associated
790  * with this task. These left over instances represent probed functions
791  * that have been called but will never return.
792  */
793 void __kprobes kprobe_flush_task(struct task_struct *tk)
794 {
795         struct kretprobe_instance *ri;
796         struct hlist_head *head, empty_rp;
797         struct hlist_node *node, *tmp;
798         unsigned long hash, flags = 0;
799
800         if (unlikely(!kprobes_initialized))
801                 /* Early boot.  kretprobe_table_locks not yet initialized. */
802                 return;
803
804         hash = hash_ptr(tk, KPROBE_HASH_BITS);
805         head = &kretprobe_inst_table[hash];
806         kretprobe_table_lock(hash, &flags);
807         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
808                 if (ri->task == tk)
809                         recycle_rp_inst(ri, &empty_rp);
810         }
811         kretprobe_table_unlock(hash, &flags);
812         INIT_HLIST_HEAD(&empty_rp);
813         hlist_for_each_entry_safe(ri, node, tmp, &empty_rp, hlist) {
814                 hlist_del(&ri->hlist);
815                 kfree(ri);
816         }
817 }
818
819 static inline void free_rp_inst(struct kretprobe *rp)
820 {
821         struct kretprobe_instance *ri;
822         struct hlist_node *pos, *next;
823
824         hlist_for_each_entry_safe(ri, pos, next, &rp->free_instances, hlist) {
825                 hlist_del(&ri->hlist);
826                 kfree(ri);
827         }
828 }
829
830 static void __kprobes cleanup_rp_inst(struct kretprobe *rp)
831 {
832         unsigned long flags, hash;
833         struct kretprobe_instance *ri;
834         struct hlist_node *pos, *next;
835         struct hlist_head *head;
836
837         /* No race here */
838         for (hash = 0; hash < KPROBE_TABLE_SIZE; hash++) {
839                 kretprobe_table_lock(hash, &flags);
840                 head = &kretprobe_inst_table[hash];
841                 hlist_for_each_entry_safe(ri, pos, next, head, hlist) {
842                         if (ri->rp == rp)
843                                 ri->rp = NULL;
844                 }
845                 kretprobe_table_unlock(hash, &flags);
846         }
847         free_rp_inst(rp);
848 }
849
850 /*
851 * Add the new probe to ap->list. Fail if this is the
852 * second jprobe at the address - two jprobes can't coexist
853 */
854 static int __kprobes add_new_kprobe(struct kprobe *ap, struct kprobe *p)
855 {
856         BUG_ON(kprobe_gone(ap) || kprobe_gone(p));
857
858         if (p->break_handler || p->post_handler)
859                 unoptimize_kprobe(ap);  /* Fall back to normal kprobe */
860
861         if (p->break_handler) {
862                 if (ap->break_handler)
863                         return -EEXIST;
864                 list_add_tail_rcu(&p->list, &ap->list);
865                 ap->break_handler = aggr_break_handler;
866         } else
867                 list_add_rcu(&p->list, &ap->list);
868         if (p->post_handler && !ap->post_handler)
869                 ap->post_handler = aggr_post_handler;
870
871         if (kprobe_disabled(ap) && !kprobe_disabled(p)) {
872                 ap->flags &= ~KPROBE_FLAG_DISABLED;
873                 if (!kprobes_all_disarmed)
874                         /* Arm the breakpoint again. */
875                         __arm_kprobe(ap);
876         }
877         return 0;
878 }
879
880 /*
881  * Fill in the required fields of the "manager kprobe". Replace the
882  * earlier kprobe in the hlist with the manager kprobe
883  */
884 static void __kprobes init_aggr_kprobe(struct kprobe *ap, struct kprobe *p)
885 {
886         /* Copy p's insn slot to ap */
887         copy_kprobe(p, ap);
888         flush_insn_slot(ap);
889         ap->addr = p->addr;
890         ap->flags = p->flags & ~KPROBE_FLAG_OPTIMIZED;
891         ap->pre_handler = aggr_pre_handler;
892         ap->fault_handler = aggr_fault_handler;
893         /* We don't care the kprobe which has gone. */
894         if (p->post_handler && !kprobe_gone(p))
895                 ap->post_handler = aggr_post_handler;
896         if (p->break_handler && !kprobe_gone(p))
897                 ap->break_handler = aggr_break_handler;
898
899         INIT_LIST_HEAD(&ap->list);
900         INIT_HLIST_NODE(&ap->hlist);
901
902         list_add_rcu(&p->list, &ap->list);
903         hlist_replace_rcu(&p->hlist, &ap->hlist);
904 }
905
906 /*
907  * This is the second or subsequent kprobe at the address - handle
908  * the intricacies
909  */
910 static int __kprobes register_aggr_kprobe(struct kprobe *old_p,
911                                           struct kprobe *p)
912 {
913         int ret = 0;
914         struct kprobe *ap = old_p;
915
916         if (!kprobe_aggrprobe(old_p)) {
917                 /* If old_p is not an aggr_kprobe, create new aggr_kprobe. */
918                 ap = alloc_aggr_kprobe(old_p);
919                 if (!ap)
920                         return -ENOMEM;
921                 init_aggr_kprobe(ap, old_p);
922         }
923
924         if (kprobe_gone(ap)) {
925                 /*
926                  * Attempting to insert new probe at the same location that
927                  * had a probe in the module vaddr area which already
928                  * freed. So, the instruction slot has already been
929                  * released. We need a new slot for the new probe.
930                  */
931                 ret = arch_prepare_kprobe(ap);
932                 if (ret)
933                         /*
934                          * Even if fail to allocate new slot, don't need to
935                          * free aggr_probe. It will be used next time, or
936                          * freed by unregister_kprobe.
937                          */
938                         return ret;
939
940                 /* Prepare optimized instructions if possible. */
941                 prepare_optimized_kprobe(ap);
942
943                 /*
944                  * Clear gone flag to prevent allocating new slot again, and
945                  * set disabled flag because it is not armed yet.
946                  */
947                 ap->flags = (ap->flags & ~KPROBE_FLAG_GONE)
948                             | KPROBE_FLAG_DISABLED;
949         }
950
951         /* Copy ap's insn slot to p */
952         copy_kprobe(ap, p);
953         return add_new_kprobe(ap, p);
954 }
955
956 /* Try to disable aggr_kprobe, and return 1 if succeeded.*/
957 static int __kprobes try_to_disable_aggr_kprobe(struct kprobe *p)
958 {
959         struct kprobe *kp;
960
961         list_for_each_entry_rcu(kp, &p->list, list) {
962                 if (!kprobe_disabled(kp))
963                         /*
964                          * There is an active probe on the list.
965                          * We can't disable aggr_kprobe.
966                          */
967                         return 0;
968         }
969         p->flags |= KPROBE_FLAG_DISABLED;
970         return 1;
971 }
972
973 static int __kprobes in_kprobes_functions(unsigned long addr)
974 {
975         struct kprobe_blackpoint *kb;
976
977         if (addr >= (unsigned long)__kprobes_text_start &&
978             addr < (unsigned long)__kprobes_text_end)
979                 return -EINVAL;
980         /*
981          * If there exists a kprobe_blacklist, verify and
982          * fail any probe registration in the prohibited area
983          */
984         for (kb = kprobe_blacklist; kb->name != NULL; kb++) {
985                 if (kb->start_addr) {
986                         if (addr >= kb->start_addr &&
987                             addr < (kb->start_addr + kb->range))
988                                 return -EINVAL;
989                 }
990         }
991         return 0;
992 }
993
994 /*
995  * If we have a symbol_name argument, look it up and add the offset field
996  * to it. This way, we can specify a relative address to a symbol.
997  */
998 static kprobe_opcode_t __kprobes *kprobe_addr(struct kprobe *p)
999 {
1000         kprobe_opcode_t *addr = p->addr;
1001         if (p->symbol_name) {
1002                 if (addr)
1003                         return NULL;
1004                 kprobe_lookup_name(p->symbol_name, addr);
1005         }
1006
1007         if (!addr)
1008                 return NULL;
1009         return (kprobe_opcode_t *)(((char *)addr) + p->offset);
1010 }
1011
1012 /* Check passed kprobe is valid and return kprobe in kprobe_table. */
1013 static struct kprobe * __kprobes __get_valid_kprobe(struct kprobe *p)
1014 {
1015         struct kprobe *old_p, *list_p;
1016
1017         old_p = get_kprobe(p->addr);
1018         if (unlikely(!old_p))
1019                 return NULL;
1020
1021         if (p != old_p) {
1022                 list_for_each_entry_rcu(list_p, &old_p->list, list)
1023                         if (list_p == p)
1024                         /* kprobe p is a valid probe */
1025                                 goto valid;
1026                 return NULL;
1027         }
1028 valid:
1029         return old_p;
1030 }
1031
1032 /* Return error if the kprobe is being re-registered */
1033 static inline int check_kprobe_rereg(struct kprobe *p)
1034 {
1035         int ret = 0;
1036         struct kprobe *old_p;
1037
1038         mutex_lock(&kprobe_mutex);
1039         old_p = __get_valid_kprobe(p);
1040         if (old_p)
1041                 ret = -EINVAL;
1042         mutex_unlock(&kprobe_mutex);
1043         return ret;
1044 }
1045
1046 int __kprobes register_kprobe(struct kprobe *p)
1047 {
1048         int ret = 0;
1049         struct kprobe *old_p;
1050         struct module *probed_mod;
1051         kprobe_opcode_t *addr;
1052
1053         addr = kprobe_addr(p);
1054         if (!addr)
1055                 return -EINVAL;
1056         p->addr = addr;
1057
1058         ret = check_kprobe_rereg(p);
1059         if (ret)
1060                 return ret;
1061
1062         preempt_disable();
1063         if (!kernel_text_address((unsigned long) p->addr) ||
1064             in_kprobes_functions((unsigned long) p->addr) ||
1065             ftrace_text_reserved(p->addr, p->addr)) {
1066                 preempt_enable();
1067                 return -EINVAL;
1068         }
1069
1070         /* User can pass only KPROBE_FLAG_DISABLED to register_kprobe */
1071         p->flags &= KPROBE_FLAG_DISABLED;
1072
1073         /*
1074          * Check if are we probing a module.
1075          */
1076         probed_mod = __module_text_address((unsigned long) p->addr);
1077         if (probed_mod) {
1078                 /*
1079                  * We must hold a refcount of the probed module while updating
1080                  * its code to prohibit unexpected unloading.
1081                  */
1082                 if (unlikely(!try_module_get(probed_mod))) {
1083                         preempt_enable();
1084                         return -EINVAL;
1085                 }
1086                 /*
1087                  * If the module freed .init.text, we couldn't insert
1088                  * kprobes in there.
1089                  */
1090                 if (within_module_init((unsigned long)p->addr, probed_mod) &&
1091                     probed_mod->state != MODULE_STATE_COMING) {
1092                         module_put(probed_mod);
1093                         preempt_enable();
1094                         return -EINVAL;
1095                 }
1096         }
1097         preempt_enable();
1098
1099         p->nmissed = 0;
1100         INIT_LIST_HEAD(&p->list);
1101         mutex_lock(&kprobe_mutex);
1102
1103         get_online_cpus();      /* For avoiding text_mutex deadlock. */
1104         mutex_lock(&text_mutex);
1105
1106         old_p = get_kprobe(p->addr);
1107         if (old_p) {
1108                 /* Since this may unoptimize old_p, locking text_mutex. */
1109                 ret = register_aggr_kprobe(old_p, p);
1110                 goto out;
1111         }
1112
1113         ret = arch_prepare_kprobe(p);
1114         if (ret)
1115                 goto out;
1116
1117         INIT_HLIST_NODE(&p->hlist);
1118         hlist_add_head_rcu(&p->hlist,
1119                        &kprobe_table[hash_ptr(p->addr, KPROBE_HASH_BITS)]);
1120
1121         if (!kprobes_all_disarmed && !kprobe_disabled(p))
1122                 __arm_kprobe(p);
1123
1124         /* Try to optimize kprobe */
1125         try_to_optimize_kprobe(p);
1126
1127 out:
1128         mutex_unlock(&text_mutex);
1129         put_online_cpus();
1130         mutex_unlock(&kprobe_mutex);
1131
1132         if (probed_mod)
1133                 module_put(probed_mod);
1134
1135         return ret;
1136 }
1137 EXPORT_SYMBOL_GPL(register_kprobe);
1138
1139 /*
1140  * Unregister a kprobe without a scheduler synchronization.
1141  */
1142 static int __kprobes __unregister_kprobe_top(struct kprobe *p)
1143 {
1144         struct kprobe *old_p, *list_p;
1145
1146         old_p = __get_valid_kprobe(p);
1147         if (old_p == NULL)
1148                 return -EINVAL;
1149
1150         if (old_p == p ||
1151             (kprobe_aggrprobe(old_p) &&
1152              list_is_singular(&old_p->list))) {
1153                 /*
1154                  * Only probe on the hash list. Disarm only if kprobes are
1155                  * enabled and not gone - otherwise, the breakpoint would
1156                  * already have been removed. We save on flushing icache.
1157                  */
1158                 if (!kprobes_all_disarmed && !kprobe_disabled(old_p))
1159                         disarm_kprobe(old_p);
1160                 hlist_del_rcu(&old_p->hlist);
1161         } else {
1162                 if (p->break_handler && !kprobe_gone(p))
1163                         old_p->break_handler = NULL;
1164                 if (p->post_handler && !kprobe_gone(p)) {
1165                         list_for_each_entry_rcu(list_p, &old_p->list, list) {
1166                                 if ((list_p != p) && (list_p->post_handler))
1167                                         goto noclean;
1168                         }
1169                         old_p->post_handler = NULL;
1170                 }
1171 noclean:
1172                 list_del_rcu(&p->list);
1173                 if (!kprobe_disabled(old_p)) {
1174                         try_to_disable_aggr_kprobe(old_p);
1175                         if (!kprobes_all_disarmed) {
1176                                 if (kprobe_disabled(old_p))
1177                                         disarm_kprobe(old_p);
1178                                 else
1179                                         /* Try to optimize this probe again */
1180                                         optimize_kprobe(old_p);
1181                         }
1182                 }
1183         }
1184         return 0;
1185 }
1186
1187 static void __kprobes __unregister_kprobe_bottom(struct kprobe *p)
1188 {
1189         struct kprobe *old_p;
1190
1191         if (list_empty(&p->list))
1192                 arch_remove_kprobe(p);
1193         else if (list_is_singular(&p->list)) {
1194                 /* "p" is the last child of an aggr_kprobe */
1195                 old_p = list_entry(p->list.next, struct kprobe, list);
1196                 list_del(&p->list);
1197                 arch_remove_kprobe(old_p);
1198                 free_aggr_kprobe(old_p);
1199         }
1200 }
1201
1202 int __kprobes register_kprobes(struct kprobe **kps, int num)
1203 {
1204         int i, ret = 0;
1205
1206         if (num <= 0)
1207                 return -EINVAL;
1208         for (i = 0; i < num; i++) {
1209                 ret = register_kprobe(kps[i]);
1210                 if (ret < 0) {
1211                         if (i > 0)
1212                                 unregister_kprobes(kps, i);
1213                         break;
1214                 }
1215         }
1216         return ret;
1217 }
1218 EXPORT_SYMBOL_GPL(register_kprobes);
1219
1220 void __kprobes unregister_kprobe(struct kprobe *p)
1221 {
1222         unregister_kprobes(&p, 1);
1223 }
1224 EXPORT_SYMBOL_GPL(unregister_kprobe);
1225
1226 void __kprobes unregister_kprobes(struct kprobe **kps, int num)
1227 {
1228         int i;
1229
1230         if (num <= 0)
1231                 return;
1232         mutex_lock(&kprobe_mutex);
1233         for (i = 0; i < num; i++)
1234                 if (__unregister_kprobe_top(kps[i]) < 0)
1235                         kps[i]->addr = NULL;
1236         mutex_unlock(&kprobe_mutex);
1237
1238         synchronize_sched();
1239         for (i = 0; i < num; i++)
1240                 if (kps[i]->addr)
1241                         __unregister_kprobe_bottom(kps[i]);
1242 }
1243 EXPORT_SYMBOL_GPL(unregister_kprobes);
1244
1245 static struct notifier_block kprobe_exceptions_nb = {
1246         .notifier_call = kprobe_exceptions_notify,
1247         .priority = 0x7fffffff /* we need to be notified first */
1248 };
1249
1250 unsigned long __weak arch_deref_entry_point(void *entry)
1251 {
1252         return (unsigned long)entry;
1253 }
1254
1255 int __kprobes register_jprobes(struct jprobe **jps, int num)
1256 {
1257         struct jprobe *jp;
1258         int ret = 0, i;
1259
1260         if (num <= 0)
1261                 return -EINVAL;
1262         for (i = 0; i < num; i++) {
1263                 unsigned long addr;
1264                 jp = jps[i];
1265                 addr = arch_deref_entry_point(jp->entry);
1266
1267                 if (!kernel_text_address(addr))
1268                         ret = -EINVAL;
1269                 else {
1270                         /* Todo: Verify probepoint is a function entry point */
1271                         jp->kp.pre_handler = setjmp_pre_handler;
1272                         jp->kp.break_handler = longjmp_break_handler;
1273                         ret = register_kprobe(&jp->kp);
1274                 }
1275                 if (ret < 0) {
1276                         if (i > 0)
1277                                 unregister_jprobes(jps, i);
1278                         break;
1279                 }
1280         }
1281         return ret;
1282 }
1283 EXPORT_SYMBOL_GPL(register_jprobes);
1284
1285 int __kprobes register_jprobe(struct jprobe *jp)
1286 {
1287         return register_jprobes(&jp, 1);
1288 }
1289 EXPORT_SYMBOL_GPL(register_jprobe);
1290
1291 void __kprobes unregister_jprobe(struct jprobe *jp)
1292 {
1293         unregister_jprobes(&jp, 1);
1294 }
1295 EXPORT_SYMBOL_GPL(unregister_jprobe);
1296
1297 void __kprobes unregister_jprobes(struct jprobe **jps, int num)
1298 {
1299         int i;
1300
1301         if (num <= 0)
1302                 return;
1303         mutex_lock(&kprobe_mutex);
1304         for (i = 0; i < num; i++)
1305                 if (__unregister_kprobe_top(&jps[i]->kp) < 0)
1306                         jps[i]->kp.addr = NULL;
1307         mutex_unlock(&kprobe_mutex);
1308
1309         synchronize_sched();
1310         for (i = 0; i < num; i++) {
1311                 if (jps[i]->kp.addr)
1312                         __unregister_kprobe_bottom(&jps[i]->kp);
1313         }
1314 }
1315 EXPORT_SYMBOL_GPL(unregister_jprobes);
1316
1317 #ifdef CONFIG_KRETPROBES
1318 /*
1319  * This kprobe pre_handler is registered with every kretprobe. When probe
1320  * hits it will set up the return probe.
1321  */
1322 static int __kprobes pre_handler_kretprobe(struct kprobe *p,
1323                                            struct pt_regs *regs)
1324 {
1325         struct kretprobe *rp = container_of(p, struct kretprobe, kp);
1326         unsigned long hash, flags = 0;
1327         struct kretprobe_instance *ri;
1328
1329         /*TODO: consider to only swap the RA after the last pre_handler fired */
1330         hash = hash_ptr(current, KPROBE_HASH_BITS);
1331         spin_lock_irqsave(&rp->lock, flags);
1332         if (!hlist_empty(&rp->free_instances)) {
1333                 ri = hlist_entry(rp->free_instances.first,
1334                                 struct kretprobe_instance, hlist);
1335                 hlist_del(&ri->hlist);
1336                 spin_unlock_irqrestore(&rp->lock, flags);
1337
1338                 ri->rp = rp;
1339                 ri->task = current;
1340
1341                 if (rp->entry_handler && rp->entry_handler(ri, regs))
1342                         return 0;
1343
1344                 arch_prepare_kretprobe(ri, regs);
1345
1346                 /* XXX(hch): why is there no hlist_move_head? */
1347                 INIT_HLIST_NODE(&ri->hlist);
1348                 kretprobe_table_lock(hash, &flags);
1349                 hlist_add_head(&ri->hlist, &kretprobe_inst_table[hash]);
1350                 kretprobe_table_unlock(hash, &flags);
1351         } else {
1352                 rp->nmissed++;
1353                 spin_unlock_irqrestore(&rp->lock, flags);
1354         }
1355         return 0;
1356 }
1357
1358 int __kprobes register_kretprobe(struct kretprobe *rp)
1359 {
1360         int ret = 0;
1361         struct kretprobe_instance *inst;
1362         int i;
1363         void *addr;
1364
1365         if (kretprobe_blacklist_size) {
1366                 addr = kprobe_addr(&rp->kp);
1367                 if (!addr)
1368                         return -EINVAL;
1369
1370                 for (i = 0; kretprobe_blacklist[i].name != NULL; i++) {
1371                         if (kretprobe_blacklist[i].addr == addr)
1372                                 return -EINVAL;
1373                 }
1374         }
1375
1376         rp->kp.pre_handler = pre_handler_kretprobe;
1377         rp->kp.post_handler = NULL;
1378         rp->kp.fault_handler = NULL;
1379         rp->kp.break_handler = NULL;
1380
1381         /* Pre-allocate memory for max kretprobe instances */
1382         if (rp->maxactive <= 0) {
1383 #ifdef CONFIG_PREEMPT
1384                 rp->maxactive = max_t(unsigned int, 10, 2*num_possible_cpus());
1385 #else
1386                 rp->maxactive = num_possible_cpus();
1387 #endif
1388         }
1389         spin_lock_init(&rp->lock);
1390         INIT_HLIST_HEAD(&rp->free_instances);
1391         for (i = 0; i < rp->maxactive; i++) {
1392                 inst = kmalloc(sizeof(struct kretprobe_instance) +
1393                                rp->data_size, GFP_KERNEL);
1394                 if (inst == NULL) {
1395                         free_rp_inst(rp);
1396                         return -ENOMEM;
1397                 }
1398                 INIT_HLIST_NODE(&inst->hlist);
1399                 hlist_add_head(&inst->hlist, &rp->free_instances);
1400         }
1401
1402         rp->nmissed = 0;
1403         /* Establish function entry probe point */
1404         ret = register_kprobe(&rp->kp);
1405         if (ret != 0)
1406                 free_rp_inst(rp);
1407         return ret;
1408 }
1409 EXPORT_SYMBOL_GPL(register_kretprobe);
1410
1411 int __kprobes register_kretprobes(struct kretprobe **rps, int num)
1412 {
1413         int ret = 0, i;
1414
1415         if (num <= 0)
1416                 return -EINVAL;
1417         for (i = 0; i < num; i++) {
1418                 ret = register_kretprobe(rps[i]);
1419                 if (ret < 0) {
1420                         if (i > 0)
1421                                 unregister_kretprobes(rps, i);
1422                         break;
1423                 }
1424         }
1425         return ret;
1426 }
1427 EXPORT_SYMBOL_GPL(register_kretprobes);
1428
1429 void __kprobes unregister_kretprobe(struct kretprobe *rp)
1430 {
1431         unregister_kretprobes(&rp, 1);
1432 }
1433 EXPORT_SYMBOL_GPL(unregister_kretprobe);
1434
1435 void __kprobes unregister_kretprobes(struct kretprobe **rps, int num)
1436 {
1437         int i;
1438
1439         if (num <= 0)
1440                 return;
1441         mutex_lock(&kprobe_mutex);
1442         for (i = 0; i < num; i++)
1443                 if (__unregister_kprobe_top(&rps[i]->kp) < 0)
1444                         rps[i]->kp.addr = NULL;
1445         mutex_unlock(&kprobe_mutex);
1446
1447         synchronize_sched();
1448         for (i = 0; i < num; i++) {
1449                 if (rps[i]->kp.addr) {
1450                         __unregister_kprobe_bottom(&rps[i]->kp);
1451                         cleanup_rp_inst(rps[i]);
1452                 }
1453         }
1454 }
1455 EXPORT_SYMBOL_GPL(unregister_kretprobes);
1456
1457 #else /* CONFIG_KRETPROBES */
1458 int __kprobes register_kretprobe(struct kretprobe *rp)
1459 {
1460         return -ENOSYS;
1461 }
1462 EXPORT_SYMBOL_GPL(register_kretprobe);
1463
1464 int __kprobes register_kretprobes(struct kretprobe **rps, int num)
1465 {
1466         return -ENOSYS;
1467 }
1468 EXPORT_SYMBOL_GPL(register_kretprobes);
1469
1470 void __kprobes unregister_kretprobe(struct kretprobe *rp)
1471 {
1472 }
1473 EXPORT_SYMBOL_GPL(unregister_kretprobe);
1474
1475 void __kprobes unregister_kretprobes(struct kretprobe **rps, int num)
1476 {
1477 }
1478 EXPORT_SYMBOL_GPL(unregister_kretprobes);
1479
1480 static int __kprobes pre_handler_kretprobe(struct kprobe *p,
1481                                            struct pt_regs *regs)
1482 {
1483         return 0;
1484 }
1485
1486 #endif /* CONFIG_KRETPROBES */
1487
1488 /* Set the kprobe gone and remove its instruction buffer. */
1489 static void __kprobes kill_kprobe(struct kprobe *p)
1490 {
1491         struct kprobe *kp;
1492
1493         p->flags |= KPROBE_FLAG_GONE;
1494         if (kprobe_aggrprobe(p)) {
1495                 /*
1496                  * If this is an aggr_kprobe, we have to list all the
1497                  * chained probes and mark them GONE.
1498                  */
1499                 list_for_each_entry_rcu(kp, &p->list, list)
1500                         kp->flags |= KPROBE_FLAG_GONE;
1501                 p->post_handler = NULL;
1502                 p->break_handler = NULL;
1503                 kill_optimized_kprobe(p);
1504         }
1505         /*
1506          * Here, we can remove insn_slot safely, because no thread calls
1507          * the original probed function (which will be freed soon) any more.
1508          */
1509         arch_remove_kprobe(p);
1510 }
1511
1512 void __kprobes dump_kprobe(struct kprobe *kp)
1513 {
1514         printk(KERN_WARNING "Dumping kprobe:\n");
1515         printk(KERN_WARNING "Name: %s\nAddress: %p\nOffset: %x\n",
1516                kp->symbol_name, kp->addr, kp->offset);
1517 }
1518
1519 /* Module notifier call back, checking kprobes on the module */
1520 static int __kprobes kprobes_module_callback(struct notifier_block *nb,
1521                                              unsigned long val, void *data)
1522 {
1523         struct module *mod = data;
1524         struct hlist_head *head;
1525         struct hlist_node *node;
1526         struct kprobe *p;
1527         unsigned int i;
1528         int checkcore = (val == MODULE_STATE_GOING);
1529
1530         if (val != MODULE_STATE_GOING && val != MODULE_STATE_LIVE)
1531                 return NOTIFY_DONE;
1532
1533         /*
1534          * When MODULE_STATE_GOING was notified, both of module .text and
1535          * .init.text sections would be freed. When MODULE_STATE_LIVE was
1536          * notified, only .init.text section would be freed. We need to
1537          * disable kprobes which have been inserted in the sections.
1538          */
1539         mutex_lock(&kprobe_mutex);
1540         for (i = 0; i < KPROBE_TABLE_SIZE; i++) {
1541                 head = &kprobe_table[i];
1542                 hlist_for_each_entry_rcu(p, node, head, hlist)
1543                         if (within_module_init((unsigned long)p->addr, mod) ||
1544                             (checkcore &&
1545                              within_module_core((unsigned long)p->addr, mod))) {
1546                                 /*
1547                                  * The vaddr this probe is installed will soon
1548                                  * be vfreed buy not synced to disk. Hence,
1549                                  * disarming the breakpoint isn't needed.
1550                                  */
1551                                 kill_kprobe(p);
1552                         }
1553         }
1554         mutex_unlock(&kprobe_mutex);
1555         return NOTIFY_DONE;
1556 }
1557
1558 static struct notifier_block kprobe_module_nb = {
1559         .notifier_call = kprobes_module_callback,
1560         .priority = 0
1561 };
1562
1563 static int __init init_kprobes(void)
1564 {
1565         int i, err = 0;
1566         unsigned long offset = 0, size = 0;
1567         char *modname, namebuf[128];
1568         const char *symbol_name;
1569         void *addr;
1570         struct kprobe_blackpoint *kb;
1571
1572         /* FIXME allocate the probe table, currently defined statically */
1573         /* initialize all list heads */
1574         for (i = 0; i < KPROBE_TABLE_SIZE; i++) {
1575                 INIT_HLIST_HEAD(&kprobe_table[i]);
1576                 INIT_HLIST_HEAD(&kretprobe_inst_table[i]);
1577                 spin_lock_init(&(kretprobe_table_locks[i].lock));
1578         }
1579
1580         /*
1581          * Lookup and populate the kprobe_blacklist.
1582          *
1583          * Unlike the kretprobe blacklist, we'll need to determine
1584          * the range of addresses that belong to the said functions,
1585          * since a kprobe need not necessarily be at the beginning
1586          * of a function.
1587          */
1588         for (kb = kprobe_blacklist; kb->name != NULL; kb++) {
1589                 kprobe_lookup_name(kb->name, addr);
1590                 if (!addr)
1591                         continue;
1592
1593                 kb->start_addr = (unsigned long)addr;
1594                 symbol_name = kallsyms_lookup(kb->start_addr,
1595                                 &size, &offset, &modname, namebuf);
1596                 if (!symbol_name)
1597                         kb->range = 0;
1598                 else
1599                         kb->range = size;
1600         }
1601
1602         if (kretprobe_blacklist_size) {
1603                 /* lookup the function address from its name */
1604                 for (i = 0; kretprobe_blacklist[i].name != NULL; i++) {
1605                         kprobe_lookup_name(kretprobe_blacklist[i].name,
1606                                            kretprobe_blacklist[i].addr);
1607                         if (!kretprobe_blacklist[i].addr)
1608                                 printk("kretprobe: lookup failed: %s\n",
1609                                        kretprobe_blacklist[i].name);
1610                 }
1611         }
1612
1613 #if defined(CONFIG_OPTPROBES) && defined(__ARCH_WANT_KPROBES_INSN_SLOT)
1614         /* Init kprobe_optinsn_slots */
1615         kprobe_optinsn_slots.insn_size = MAX_OPTINSN_SIZE;
1616 #endif
1617
1618         /* By default, kprobes are armed */
1619         kprobes_all_disarmed = false;
1620
1621         err = arch_init_kprobes();
1622         if (!err)
1623                 err = register_die_notifier(&kprobe_exceptions_nb);
1624         if (!err)
1625                 err = register_module_notifier(&kprobe_module_nb);
1626
1627         kprobes_initialized = (err == 0);
1628
1629         if (!err)
1630                 init_test_probes();
1631         return err;
1632 }
1633
1634 #ifdef CONFIG_DEBUG_FS
1635 static void __kprobes report_probe(struct seq_file *pi, struct kprobe *p,
1636                 const char *sym, int offset, char *modname, struct kprobe *pp)
1637 {
1638         char *kprobe_type;
1639
1640         if (p->pre_handler == pre_handler_kretprobe)
1641                 kprobe_type = "r";
1642         else if (p->pre_handler == setjmp_pre_handler)
1643                 kprobe_type = "j";
1644         else
1645                 kprobe_type = "k";
1646
1647         if (sym)
1648                 seq_printf(pi, "%p  %s  %s+0x%x  %s ",
1649                         p->addr, kprobe_type, sym, offset,
1650                         (modname ? modname : " "));
1651         else
1652                 seq_printf(pi, "%p  %s  %p ",
1653                         p->addr, kprobe_type, p->addr);
1654
1655         if (!pp)
1656                 pp = p;
1657         seq_printf(pi, "%s%s%s\n",
1658                 (kprobe_gone(p) ? "[GONE]" : ""),
1659                 ((kprobe_disabled(p) && !kprobe_gone(p)) ?  "[DISABLED]" : ""),
1660                 (kprobe_optimized(pp) ? "[OPTIMIZED]" : ""));
1661 }
1662
1663 static void __kprobes *kprobe_seq_start(struct seq_file *f, loff_t *pos)
1664 {
1665         return (*pos < KPROBE_TABLE_SIZE) ? pos : NULL;
1666 }
1667
1668 static void __kprobes *kprobe_seq_next(struct seq_file *f, void *v, loff_t *pos)
1669 {
1670         (*pos)++;
1671         if (*pos >= KPROBE_TABLE_SIZE)
1672                 return NULL;
1673         return pos;
1674 }
1675
1676 static void __kprobes kprobe_seq_stop(struct seq_file *f, void *v)
1677 {
1678         /* Nothing to do */
1679 }
1680
1681 static int __kprobes show_kprobe_addr(struct seq_file *pi, void *v)
1682 {
1683         struct hlist_head *head;
1684         struct hlist_node *node;
1685         struct kprobe *p, *kp;
1686         const char *sym = NULL;
1687         unsigned int i = *(loff_t *) v;
1688         unsigned long offset = 0;
1689         char *modname, namebuf[128];
1690
1691         head = &kprobe_table[i];
1692         preempt_disable();
1693         hlist_for_each_entry_rcu(p, node, head, hlist) {
1694                 sym = kallsyms_lookup((unsigned long)p->addr, NULL,
1695                                         &offset, &modname, namebuf);
1696                 if (kprobe_aggrprobe(p)) {
1697                         list_for_each_entry_rcu(kp, &p->list, list)
1698                                 report_probe(pi, kp, sym, offset, modname, p);
1699                 } else
1700                         report_probe(pi, p, sym, offset, modname, NULL);
1701         }
1702         preempt_enable();
1703         return 0;
1704 }
1705
1706 static const struct seq_operations kprobes_seq_ops = {
1707         .start = kprobe_seq_start,
1708         .next  = kprobe_seq_next,
1709         .stop  = kprobe_seq_stop,
1710         .show  = show_kprobe_addr
1711 };
1712
1713 static int __kprobes kprobes_open(struct inode *inode, struct file *filp)
1714 {
1715         return seq_open(filp, &kprobes_seq_ops);
1716 }
1717
1718 static const struct file_operations debugfs_kprobes_operations = {
1719         .open           = kprobes_open,
1720         .read           = seq_read,
1721         .llseek         = seq_lseek,
1722         .release        = seq_release,
1723 };
1724
1725 /* Disable one kprobe */
1726 int __kprobes disable_kprobe(struct kprobe *kp)
1727 {
1728         int ret = 0;
1729         struct kprobe *p;
1730
1731         mutex_lock(&kprobe_mutex);
1732
1733         /* Check whether specified probe is valid. */
1734         p = __get_valid_kprobe(kp);
1735         if (unlikely(p == NULL)) {
1736                 ret = -EINVAL;
1737                 goto out;
1738         }
1739
1740         /* If the probe is already disabled (or gone), just return */
1741         if (kprobe_disabled(kp))
1742                 goto out;
1743
1744         kp->flags |= KPROBE_FLAG_DISABLED;
1745         if (p != kp)
1746                 /* When kp != p, p is always enabled. */
1747                 try_to_disable_aggr_kprobe(p);
1748
1749         if (!kprobes_all_disarmed && kprobe_disabled(p))
1750                 disarm_kprobe(p);
1751 out:
1752         mutex_unlock(&kprobe_mutex);
1753         return ret;
1754 }
1755 EXPORT_SYMBOL_GPL(disable_kprobe);
1756
1757 /* Enable one kprobe */
1758 int __kprobes enable_kprobe(struct kprobe *kp)
1759 {
1760         int ret = 0;
1761         struct kprobe *p;
1762
1763         mutex_lock(&kprobe_mutex);
1764
1765         /* Check whether specified probe is valid. */
1766         p = __get_valid_kprobe(kp);
1767         if (unlikely(p == NULL)) {
1768                 ret = -EINVAL;
1769                 goto out;
1770         }
1771
1772         if (kprobe_gone(kp)) {
1773                 /* This kprobe has gone, we couldn't enable it. */
1774                 ret = -EINVAL;
1775                 goto out;
1776         }
1777
1778         if (p != kp)
1779                 kp->flags &= ~KPROBE_FLAG_DISABLED;
1780
1781         if (!kprobes_all_disarmed && kprobe_disabled(p)) {
1782                 p->flags &= ~KPROBE_FLAG_DISABLED;
1783                 arm_kprobe(p);
1784         }
1785 out:
1786         mutex_unlock(&kprobe_mutex);
1787         return ret;
1788 }
1789 EXPORT_SYMBOL_GPL(enable_kprobe);
1790
1791 static void __kprobes arm_all_kprobes(void)
1792 {
1793         struct hlist_head *head;
1794         struct hlist_node *node;
1795         struct kprobe *p;
1796         unsigned int i;
1797
1798         mutex_lock(&kprobe_mutex);
1799
1800         /* If kprobes are armed, just return */
1801         if (!kprobes_all_disarmed)
1802                 goto already_enabled;
1803
1804         /* Arming kprobes doesn't optimize kprobe itself */
1805         mutex_lock(&text_mutex);
1806         for (i = 0; i < KPROBE_TABLE_SIZE; i++) {
1807                 head = &kprobe_table[i];
1808                 hlist_for_each_entry_rcu(p, node, head, hlist)
1809                         if (!kprobe_disabled(p))
1810                                 __arm_kprobe(p);
1811         }
1812         mutex_unlock(&text_mutex);
1813
1814         kprobes_all_disarmed = false;
1815         printk(KERN_INFO "Kprobes globally enabled\n");
1816
1817 already_enabled:
1818         mutex_unlock(&kprobe_mutex);
1819         return;
1820 }
1821
1822 static void __kprobes disarm_all_kprobes(void)
1823 {
1824         struct hlist_head *head;
1825         struct hlist_node *node;
1826         struct kprobe *p;
1827         unsigned int i;
1828
1829         mutex_lock(&kprobe_mutex);
1830
1831         /* If kprobes are already disarmed, just return */
1832         if (kprobes_all_disarmed)
1833                 goto already_disabled;
1834
1835         kprobes_all_disarmed = true;
1836         printk(KERN_INFO "Kprobes globally disabled\n");
1837
1838         /*
1839          * Here we call get_online_cpus() for avoiding text_mutex deadlock,
1840          * because disarming may also unoptimize kprobes.
1841          */
1842         get_online_cpus();
1843         mutex_lock(&text_mutex);
1844         for (i = 0; i < KPROBE_TABLE_SIZE; i++) {
1845                 head = &kprobe_table[i];
1846                 hlist_for_each_entry_rcu(p, node, head, hlist) {
1847                         if (!arch_trampoline_kprobe(p) && !kprobe_disabled(p))
1848                                 __disarm_kprobe(p);
1849                 }
1850         }
1851
1852         mutex_unlock(&text_mutex);
1853         put_online_cpus();
1854         mutex_unlock(&kprobe_mutex);
1855         /* Allow all currently running kprobes to complete */
1856         synchronize_sched();
1857         return;
1858
1859 already_disabled:
1860         mutex_unlock(&kprobe_mutex);
1861         return;
1862 }
1863
1864 /*
1865  * XXX: The debugfs bool file interface doesn't allow for callbacks
1866  * when the bool state is switched. We can reuse that facility when
1867  * available
1868  */
1869 static ssize_t read_enabled_file_bool(struct file *file,
1870                char __user *user_buf, size_t count, loff_t *ppos)
1871 {
1872         char buf[3];
1873
1874         if (!kprobes_all_disarmed)
1875                 buf[0] = '1';
1876         else
1877                 buf[0] = '0';
1878         buf[1] = '\n';
1879         buf[2] = 0x00;
1880         return simple_read_from_buffer(user_buf, count, ppos, buf, 2);
1881 }
1882
1883 static ssize_t write_enabled_file_bool(struct file *file,
1884                const char __user *user_buf, size_t count, loff_t *ppos)
1885 {
1886         char buf[32];
1887         int buf_size;
1888
1889         buf_size = min(count, (sizeof(buf)-1));
1890         if (copy_from_user(buf, user_buf, buf_size))
1891                 return -EFAULT;
1892
1893         switch (buf[0]) {
1894         case 'y':
1895         case 'Y':
1896         case '1':
1897                 arm_all_kprobes();
1898                 break;
1899         case 'n':
1900         case 'N':
1901         case '0':
1902                 disarm_all_kprobes();
1903                 break;
1904         }
1905
1906         return count;
1907 }
1908
1909 static const struct file_operations fops_kp = {
1910         .read =         read_enabled_file_bool,
1911         .write =        write_enabled_file_bool,
1912 };
1913
1914 static int __kprobes debugfs_kprobe_init(void)
1915 {
1916         struct dentry *dir, *file;
1917         unsigned int value = 1;
1918
1919         dir = debugfs_create_dir("kprobes", NULL);
1920         if (!dir)
1921                 return -ENOMEM;
1922
1923         file = debugfs_create_file("list", 0444, dir, NULL,
1924                                 &debugfs_kprobes_operations);
1925         if (!file) {
1926                 debugfs_remove(dir);
1927                 return -ENOMEM;
1928         }
1929
1930         file = debugfs_create_file("enabled", 0600, dir,
1931                                         &value, &fops_kp);
1932         if (!file) {
1933                 debugfs_remove(dir);
1934                 return -ENOMEM;
1935         }
1936
1937         return 0;
1938 }
1939
1940 late_initcall(debugfs_kprobe_init);
1941 #endif /* CONFIG_DEBUG_FS */
1942
1943 module_init(init_kprobes);
1944
1945 /* defined in arch/.../kernel/kprobes.c */
1946 EXPORT_SYMBOL_GPL(jprobe_return);