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/module.h>
  39 #include <linux/moduleloader.h>
  40 #include <linux/kallsyms.h>
  41 #include <asm-generic/sections.h>
  42 #include <asm/cacheflush.h>
  43 #include <asm/errno.h>
  44 #include <asm/kdebug.h>
  45
  46 #define KPROBE_HASH_BITS 6
  47 #define KPROBE_TABLE_SIZE (1 << KPROBE_HASH_BITS)
  48
  49
  50 /*
  51  * Some oddball architectures like 64bit powerpc have function descriptors
  52  * so this must be overridable.
  53  */
  54 #ifndef kprobe_lookup_name
  55 #define kprobe_lookup_name(name, addr) \
  56         addr = ((kprobe_opcode_t *)(kallsyms_lookup_name(name)))
  57 #endif
  58
  59 static struct hlist_head kprobe_table[KPROBE_TABLE_SIZE];
  60 static struct hlist_head kretprobe_inst_table[KPROBE_TABLE_SIZE];
  61 static atomic_t kprobe_count;
  62
  63 DEFINE_MUTEX(kprobe_mutex);             /* Protects kprobe_table */
  64 DEFINE_SPINLOCK(kretprobe_lock);        /* Protects kretprobe_inst_table */
  65 static DEFINE_PER_CPU(struct kprobe *, kprobe_instance) = NULL;
  66
  67 static struct notifier_block kprobe_page_fault_nb = {
  68         .notifier_call = kprobe_exceptions_notify,
  69         .priority = 0x7fffffff /* we need to notified first */
  70 };
  71
  72 #ifdef __ARCH_WANT_KPROBES_INSN_SLOT
  73 /*
  74  * kprobe->ainsn.insn points to the copy of the instruction to be
  75  * single-stepped. x86_64, POWER4 and above have no-exec support and
  76  * stepping on the instruction on a vmalloced/kmalloced/data page
  77  * is a recipe for disaster
  78  */
  79 #define INSNS_PER_PAGE  (PAGE_SIZE/(MAX_INSN_SIZE * sizeof(kprobe_opcode_t)))
  80
  81 struct kprobe_insn_page {
  82         struct hlist_node hlist;
  83         kprobe_opcode_t *insns;         /* Page of instruction slots */
  84         char slot_used[INSNS_PER_PAGE];
  85         int nused;
  86 };
  87
  88 static struct hlist_head kprobe_insn_pages;
  89
  90 /**
  91  * get_insn_slot() - Find a slot on an executable page for an instruction.
  92  * We allocate an executable page if there's no room on existing ones.
  93  */
  94 kprobe_opcode_t __kprobes *get_insn_slot(void)
  95 {
  96         struct kprobe_insn_page *kip;
  97         struct hlist_node *pos;
  98
  99         hlist_for_each(pos, &kprobe_insn_pages) {
 100                 kip = hlist_entry(pos, struct kprobe_insn_page, hlist);
 101                 if (kip->nused < INSNS_PER_PAGE) {
 102                         int i;
 103                         for (i = 0; i < INSNS_PER_PAGE; i++) {
 104                                 if (!kip->slot_used[i]) {
 105                                         kip->slot_used[i] = 1;
 106                                         kip->nused++;
 107                                         return kip->insns + (i * MAX_INSN_SIZE);
 108                                 }
 109                         }
 110                         /* Surprise!  No unused slots.  Fix kip->nused. */
 111                         kip->nused = INSNS_PER_PAGE;
 112                 }
 113         }
 114
 115         /* All out of space.  Need to allocate a new page. Use slot 0.*/
 116         kip = kmalloc(sizeof(struct kprobe_insn_page), GFP_KERNEL);
 117         if (!kip) {
 118                 return NULL;
 119         }
 120
 121         /*
 122          * Use module_alloc so this page is within +/- 2GB of where the
 123          * kernel image and loaded module images reside. This is required
 124          * so x86_64 can correctly handle the %rip-relative fixups.
 125          */
 126         kip->insns = module_alloc(PAGE_SIZE);
 127         if (!kip->insns) {
 128                 kfree(kip);
 129                 return NULL;
 130         }
 131         INIT_HLIST_NODE(&kip->hlist);
 132         hlist_add_head(&kip->hlist, &kprobe_insn_pages);
 133         memset(kip->slot_used, 0, INSNS_PER_PAGE);
 134         kip->slot_used[0] = 1;
 135         kip->nused = 1;
 136         return kip->insns;
 137 }
 138
 139 void __kprobes free_insn_slot(kprobe_opcode_t *slot)
 140 {
 141         struct kprobe_insn_page *kip;
 142         struct hlist_node *pos;
 143
 144         hlist_for_each(pos, &kprobe_insn_pages) {
 145                 kip = hlist_entry(pos, struct kprobe_insn_page, hlist);
 146                 if (kip->insns <= slot &&
 147                     slot < kip->insns + (INSNS_PER_PAGE * MAX_INSN_SIZE)) {
 148                         int i = (slot - kip->insns) / MAX_INSN_SIZE;
 149                         kip->slot_used[i] = 0;
 150                         kip->nused--;
 151                         if (kip->nused == 0) {
 152                                 /*
 153                                  * Page is no longer in use.  Free it unless
 154                                  * it's the last one.  We keep the last one
 155                                  * so as not to have to set it up again the
 156                                  * next time somebody inserts a probe.
 157                                  */
 158                                 hlist_del(&kip->hlist);
 159                                 if (hlist_empty(&kprobe_insn_pages)) {
 160                                         INIT_HLIST_NODE(&kip->hlist);
 161                                         hlist_add_head(&kip->hlist,
 162                                                 &kprobe_insn_pages);
 163                                 } else {
 164                                         module_free(NULL, kip->insns);
 165                                         kfree(kip);
 166                                 }
 167                         }
 168                         return;
 169                 }
 170         }
 171 }
 172 #endif
 173
 174 /* We have preemption disabled.. so it is safe to use __ versions */
 175 static inline void set_kprobe_instance(struct kprobe *kp)
 176 {
 177         __get_cpu_var(kprobe_instance) = kp;
 178 }
 179
 180 static inline void reset_kprobe_instance(void)
 181 {
 182         __get_cpu_var(kprobe_instance) = NULL;
 183 }
 184
 185 /*
 186  * This routine is called either:
 187  *      - under the kprobe_mutex - during kprobe_[un]register()
 188  *                              OR
 189  *      - with preemption disabled - from arch/xxx/kernel/kprobes.c
 190  */
 191 struct kprobe __kprobes *get_kprobe(void *addr)
 192 {
 193         struct hlist_head *head;
 194         struct hlist_node *node;
 195         struct kprobe *p;
 196
 197         head = &kprobe_table[hash_ptr(addr, KPROBE_HASH_BITS)];
 198         hlist_for_each_entry_rcu(p, node, head, hlist) {
 199                 if (p->addr == addr)
 200                         return p;
 201         }
 202         return NULL;
 203 }
 204
 205 /*
 206  * Aggregate handlers for multiple kprobes support - these handlers
 207  * take care of invoking the individual kprobe handlers on p->list
 208  */
 209 static int __kprobes aggr_pre_handler(struct kprobe *p, struct pt_regs *regs)
 210 {
 211         struct kprobe *kp;
 212
 213         list_for_each_entry_rcu(kp, &p->list, list) {
 214                 if (kp->pre_handler) {
 215                         set_kprobe_instance(kp);
 216                         if (kp->pre_handler(kp, regs))
 217                                 return 1;
 218                 }
 219                 reset_kprobe_instance();
 220         }
 221         return 0;
 222 }
 223
 224 static void __kprobes aggr_post_handler(struct kprobe *p, struct pt_regs *regs,
 225                                         unsigned long flags)
 226 {
 227         struct kprobe *kp;
 228
 229         list_for_each_entry_rcu(kp, &p->list, list) {
 230                 if (kp->post_handler) {
 231                         set_kprobe_instance(kp);
 232                         kp->post_handler(kp, regs, flags);
 233                         reset_kprobe_instance();
 234                 }
 235         }
 236         return;
 237 }
 238
 239 static int __kprobes aggr_fault_handler(struct kprobe *p, struct pt_regs *regs,
 240                                         int trapnr)
 241 {
 242         struct kprobe *cur = __get_cpu_var(kprobe_instance);
 243
 244         /*
 245          * if we faulted "during" the execution of a user specified
 246          * probe handler, invoke just that probe's fault handler
 247          */
 248         if (cur && cur->fault_handler) {
 249                 if (cur->fault_handler(cur, regs, trapnr))
 250                         return 1;
 251         }
 252         return 0;
 253 }
 254
 255 static int __kprobes aggr_break_handler(struct kprobe *p, struct pt_regs *regs)
 256 {
 257         struct kprobe *cur = __get_cpu_var(kprobe_instance);
 258         int ret = 0;
 259
 260         if (cur && cur->break_handler) {
 261                 if (cur->break_handler(cur, regs))
 262                         ret = 1;
 263         }
 264         reset_kprobe_instance();
 265         return ret;
 266 }
 267
 268 /* Walks the list and increments nmissed count for multiprobe case */
 269 void __kprobes kprobes_inc_nmissed_count(struct kprobe *p)
 270 {
 271         struct kprobe *kp;
 272         if (p->pre_handler != aggr_pre_handler) {
 273                 p->nmissed++;
 274         } else {
 275                 list_for_each_entry_rcu(kp, &p->list, list)
 276                         kp->nmissed++;
 277         }
 278         return;
 279 }
 280
 281 /* Called with kretprobe_lock held */
 282 struct kretprobe_instance __kprobes *get_free_rp_inst(struct kretprobe *rp)
 283 {
 284         struct hlist_node *node;
 285         struct kretprobe_instance *ri;
 286         hlist_for_each_entry(ri, node, &rp->free_instances, uflist)
 287                 return ri;
 288         return NULL;
 289 }
 290
 291 /* Called with kretprobe_lock held */
 292 static struct kretprobe_instance __kprobes *get_used_rp_inst(struct kretprobe
 293                                                               *rp)
 294 {
 295         struct hlist_node *node;
 296         struct kretprobe_instance *ri;
 297         hlist_for_each_entry(ri, node, &rp->used_instances, uflist)
 298                 return ri;
 299         return NULL;
 300 }
 301
 302 /* Called with kretprobe_lock held */
 303 void __kprobes add_rp_inst(struct kretprobe_instance *ri)
 304 {
 305         /*
 306          * Remove rp inst off the free list -
 307          * Add it back when probed function returns
 308          */
 309         hlist_del(&ri->uflist);
 310
 311         /* Add rp inst onto table */
 312         INIT_HLIST_NODE(&ri->hlist);
 313         hlist_add_head(&ri->hlist,
 314                         &kretprobe_inst_table[hash_ptr(ri->task, KPROBE_HASH_BITS)]);
 315
 316         /* Also add this rp inst to the used list. */
 317         INIT_HLIST_NODE(&ri->uflist);
 318         hlist_add_head(&ri->uflist, &ri->rp->used_instances);
 319 }
 320
 321 /* Called with kretprobe_lock held */
 322 void __kprobes recycle_rp_inst(struct kretprobe_instance *ri,
 323                                 struct hlist_head *head)
 324 {
 325         /* remove rp inst off the rprobe_inst_table */
 326         hlist_del(&ri->hlist);
 327         if (ri->rp) {
 328                 /* remove rp inst off the used list */
 329                 hlist_del(&ri->uflist);
 330                 /* put rp inst back onto the free list */
 331                 INIT_HLIST_NODE(&ri->uflist);
 332                 hlist_add_head(&ri->uflist, &ri->rp->free_instances);
 333         } else
 334                 /* Unregistering */
 335                 hlist_add_head(&ri->hlist, head);
 336 }
 337
 338 struct hlist_head __kprobes *kretprobe_inst_table_head(struct task_struct *tsk)
 339 {
 340         return &kretprobe_inst_table[hash_ptr(tsk, KPROBE_HASH_BITS)];
 341 }
 342
 343 /*
 344  * This function is called from finish_task_switch when task tk becomes dead,
 345  * so that we can recycle any function-return probe instances associated
 346  * with this task. These left over instances represent probed functions
 347  * that have been called but will never return.
 348  */
 349 void __kprobes kprobe_flush_task(struct task_struct *tk)
 350 {
 351         struct kretprobe_instance *ri;
 352         struct hlist_head *head, empty_rp;
 353         struct hlist_node *node, *tmp;
 354         unsigned long flags = 0;
 355
 356         INIT_HLIST_HEAD(&empty_rp);
 357         spin_lock_irqsave(&kretprobe_lock, flags);
 358         head = kretprobe_inst_table_head(tk);
 359         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
 360                 if (ri->task == tk)
 361                         recycle_rp_inst(ri, &empty_rp);
 362         }
 363         spin_unlock_irqrestore(&kretprobe_lock, flags);
 364
 365         hlist_for_each_entry_safe(ri, node, tmp, &empty_rp, hlist) {
 366                 hlist_del(&ri->hlist);
 367                 kfree(ri);
 368         }
 369 }
 370
 371 static inline void free_rp_inst(struct kretprobe *rp)
 372 {
 373         struct kretprobe_instance *ri;
 374         while ((ri = get_free_rp_inst(rp)) != NULL) {
 375                 hlist_del(&ri->uflist);
 376                 kfree(ri);
 377         }
 378 }
 379
 380 /*
 381  * Keep all fields in the kprobe consistent
 382  */
 383 static inline void copy_kprobe(struct kprobe *old_p, struct kprobe *p)
 384 {
 385         memcpy(&p->opcode, &old_p->opcode, sizeof(kprobe_opcode_t));
 386         memcpy(&p->ainsn, &old_p->ainsn, sizeof(struct arch_specific_insn));
 387 }
 388
 389 /*
 390 * Add the new probe to old_p->list. Fail if this is the
 391 * second jprobe at the address - two jprobes can't coexist
 392 */
 393 static int __kprobes add_new_kprobe(struct kprobe *old_p, struct kprobe *p)
 394 {
 395         if (p->break_handler) {
 396                 if (old_p->break_handler)
 397                         return -EEXIST;
 398                 list_add_tail_rcu(&p->list, &old_p->list);
 399                 old_p->break_handler = aggr_break_handler;
 400         } else
 401                 list_add_rcu(&p->list, &old_p->list);
 402         if (p->post_handler && !old_p->post_handler)
 403                 old_p->post_handler = aggr_post_handler;
 404         return 0;
 405 }
 406
 407 /*
 408  * Fill in the required fields of the "manager kprobe". Replace the
 409  * earlier kprobe in the hlist with the manager kprobe
 410  */
 411 static inline void add_aggr_kprobe(struct kprobe *ap, struct kprobe *p)
 412 {
 413         copy_kprobe(p, ap);
 414         flush_insn_slot(ap);
 415         ap->addr = p->addr;
 416         ap->pre_handler = aggr_pre_handler;
 417         ap->fault_handler = aggr_fault_handler;
 418         if (p->post_handler)
 419                 ap->post_handler = aggr_post_handler;
 420         if (p->break_handler)
 421                 ap->break_handler = aggr_break_handler;
 422
 423         INIT_LIST_HEAD(&ap->list);
 424         list_add_rcu(&p->list, &ap->list);
 425
 426         hlist_replace_rcu(&p->hlist, &ap->hlist);
 427 }
 428
 429 /*
 430  * This is the second or subsequent kprobe at the address - handle
 431  * the intricacies
 432  */
 433 static int __kprobes register_aggr_kprobe(struct kprobe *old_p,
 434                                           struct kprobe *p)
 435 {
 436         int ret = 0;
 437         struct kprobe *ap;
 438
 439         if (old_p->pre_handler == aggr_pre_handler) {
 440                 copy_kprobe(old_p, p);
 441                 ret = add_new_kprobe(old_p, p);
 442         } else {
 443                 ap = kzalloc(sizeof(struct kprobe), GFP_KERNEL);
 444                 if (!ap)
 445                         return -ENOMEM;
 446                 add_aggr_kprobe(ap, old_p);
 447                 copy_kprobe(ap, p);
 448                 ret = add_new_kprobe(ap, p);
 449         }
 450         return ret;
 451 }
 452
 453 static int __kprobes in_kprobes_functions(unsigned long addr)
 454 {
 455         if (addr >= (unsigned long)__kprobes_text_start
 456                 && addr < (unsigned long)__kprobes_text_end)
 457                 return -EINVAL;
 458         return 0;
 459 }
 460
 461 static int __kprobes __register_kprobe(struct kprobe *p,
 462         unsigned long called_from)
 463 {
 464         int ret = 0;
 465         struct kprobe *old_p;
 466         struct module *probed_mod;
 467
 468         /*
 469          * If we have a symbol_name argument look it up,
 470          * and add it to the address.  That way the addr
 471          * field can either be global or relative to a symbol.
 472          */
 473         if (p->symbol_name) {
 474                 if (p->addr)
 475                         return -EINVAL;
 476                 kprobe_lookup_name(p->symbol_name, p->addr);
 477         }
 478
 479         if (!p->addr)
 480                 return -EINVAL;
 481         p->addr = (kprobe_opcode_t *)(((char *)p->addr)+ p->offset);
 482
 483         if ((!kernel_text_address((unsigned long) p->addr)) ||
 484                 in_kprobes_functions((unsigned long) p->addr))
 485                 return -EINVAL;
 486
 487         p->mod_refcounted = 0;
 488         /* Check are we probing a module */
 489         if ((probed_mod = module_text_address((unsigned long) p->addr))) {
 490                 struct module *calling_mod = module_text_address(called_from);
 491                 /* We must allow modules to probe themself and
 492                  * in this case avoid incrementing the module refcount,
 493                  * so as to allow unloading of self probing modules.
 494                  */
 495                 if (calling_mod && (calling_mod != probed_mod)) {
 496                         if (unlikely(!try_module_get(probed_mod)))
 497                                 return -EINVAL;
 498                         p->mod_refcounted = 1;
 499                 } else
 500                         probed_mod = NULL;
 501         }
 502
 503         p->nmissed = 0;
 504         mutex_lock(&kprobe_mutex);
 505         old_p = get_kprobe(p->addr);
 506         if (old_p) {
 507                 ret = register_aggr_kprobe(old_p, p);
 508                 if (!ret)
 509                         atomic_inc(&kprobe_count);
 510                 goto out;
 511         }
 512
 513         if ((ret = arch_prepare_kprobe(p)) != 0)
 514                 goto out;
 515
 516         INIT_HLIST_NODE(&p->hlist);
 517         hlist_add_head_rcu(&p->hlist,
 518                        &kprobe_table[hash_ptr(p->addr, KPROBE_HASH_BITS)]);
 519
 520         if (atomic_add_return(1, &kprobe_count) == \
 521                                 (ARCH_INACTIVE_KPROBE_COUNT + 1))
 522                 register_page_fault_notifier(&kprobe_page_fault_nb);
 523
 524         arch_arm_kprobe(p);
 525
 526 out:
 527         mutex_unlock(&kprobe_mutex);
 528
 529         if (ret && probed_mod)
 530                 module_put(probed_mod);
 531         return ret;
 532 }
 533
 534 int __kprobes register_kprobe(struct kprobe *p)
 535 {
 536         return __register_kprobe(p,
 537                 (unsigned long)__builtin_return_address(0));
 538 }
 539
 540 void __kprobes unregister_kprobe(struct kprobe *p)
 541 {
 542         struct module *mod;
 543         struct kprobe *old_p, *list_p;
 544         int cleanup_p;
 545
 546         mutex_lock(&kprobe_mutex);
 547         old_p = get_kprobe(p->addr);
 548         if (unlikely(!old_p)) {
 549                 mutex_unlock(&kprobe_mutex);
 550                 return;
 551         }
 552         if (p != old_p) {
 553                 list_for_each_entry_rcu(list_p, &old_p->list, list)
 554                         if (list_p == p)
 555                         /* kprobe p is a valid probe */
 556                                 goto valid_p;
 557                 mutex_unlock(&kprobe_mutex);
 558                 return;
 559         }
 560 valid_p:
 561         if ((old_p == p) || ((old_p->pre_handler == aggr_pre_handler) &&
 562                 (p->list.next == &old_p->list) &&
 563                 (p->list.prev == &old_p->list))) {
 564                 /* Only probe on the hash list */
 565                 arch_disarm_kprobe(p);
 566                 hlist_del_rcu(&old_p->hlist);
 567                 cleanup_p = 1;
 568         } else {
 569                 list_del_rcu(&p->list);
 570                 cleanup_p = 0;
 571         }
 572
 573         mutex_unlock(&kprobe_mutex);
 574
 575         synchronize_sched();
 576         if (p->mod_refcounted &&
 577             (mod = module_text_address((unsigned long)p->addr)))
 578                 module_put(mod);
 579
 580         if (cleanup_p) {
 581                 if (p != old_p) {
 582                         list_del_rcu(&p->list);
 583                         kfree(old_p);
 584                 }
 585                 arch_remove_kprobe(p);
 586         } else {
 587                 mutex_lock(&kprobe_mutex);
 588                 if (p->break_handler)
 589                         old_p->break_handler = NULL;
 590                 if (p->post_handler){
 591                         list_for_each_entry_rcu(list_p, &old_p->list, list){
 592                                 if (list_p->post_handler){
 593                                         cleanup_p = 2;
 594                                         break;
 595                                 }
 596                         }
 597                         if (cleanup_p == 0)
 598                                 old_p->post_handler = NULL;
 599                 }
 600                 mutex_unlock(&kprobe_mutex);
 601         }
 602
 603         /* Call unregister_page_fault_notifier()
 604          * if no probes are active
 605          */
 606         mutex_lock(&kprobe_mutex);
 607         if (atomic_add_return(-1, &kprobe_count) == \
 608                                 ARCH_INACTIVE_KPROBE_COUNT)
 609                 unregister_page_fault_notifier(&kprobe_page_fault_nb);
 610         mutex_unlock(&kprobe_mutex);
 611         return;
 612 }
 613
 614 static struct notifier_block kprobe_exceptions_nb = {
 615         .notifier_call = kprobe_exceptions_notify,
 616         .priority = 0x7fffffff /* we need to be notified first */
 617 };
 618
 619
 620 int __kprobes register_jprobe(struct jprobe *jp)
 621 {
 622         /* Todo: Verify probepoint is a function entry point */
 623         jp->kp.pre_handler = setjmp_pre_handler;
 624         jp->kp.break_handler = longjmp_break_handler;
 625
 626         return __register_kprobe(&jp->kp,
 627                 (unsigned long)__builtin_return_address(0));
 628 }
 629
 630 void __kprobes unregister_jprobe(struct jprobe *jp)
 631 {
 632         unregister_kprobe(&jp->kp);
 633 }
 634
 635 #ifdef ARCH_SUPPORTS_KRETPROBES
 636
 637 /*
 638  * This kprobe pre_handler is registered with every kretprobe. When probe
 639  * hits it will set up the return probe.
 640  */
 641 static int __kprobes pre_handler_kretprobe(struct kprobe *p,
 642                                            struct pt_regs *regs)
 643 {
 644         struct kretprobe *rp = container_of(p, struct kretprobe, kp);
 645         unsigned long flags = 0;
 646
 647         /*TODO: consider to only swap the RA after the last pre_handler fired */
 648         spin_lock_irqsave(&kretprobe_lock, flags);
 649         arch_prepare_kretprobe(rp, regs);
 650         spin_unlock_irqrestore(&kretprobe_lock, flags);
 651         return 0;
 652 }
 653
 654 int __kprobes register_kretprobe(struct kretprobe *rp)
 655 {
 656         int ret = 0;
 657         struct kretprobe_instance *inst;
 658         int i;
 659
 660         rp->kp.pre_handler = pre_handler_kretprobe;
 661         rp->kp.post_handler = NULL;
 662         rp->kp.fault_handler = NULL;
 663         rp->kp.break_handler = NULL;
 664
 665         /* Pre-allocate memory for max kretprobe instances */
 666         if (rp->maxactive <= 0) {
 667 #ifdef CONFIG_PREEMPT
 668                 rp->maxactive = max(10, 2 * NR_CPUS);
 669 #else
 670                 rp->maxactive = NR_CPUS;
 671 #endif
 672         }
 673         INIT_HLIST_HEAD(&rp->used_instances);
 674         INIT_HLIST_HEAD(&rp->free_instances);
 675         for (i = 0; i < rp->maxactive; i++) {
 676                 inst = kmalloc(sizeof(struct kretprobe_instance), GFP_KERNEL);
 677                 if (inst == NULL) {
 678                         free_rp_inst(rp);
 679                         return -ENOMEM;
 680                 }
 681                 INIT_HLIST_NODE(&inst->uflist);
 682                 hlist_add_head(&inst->uflist, &rp->free_instances);
 683         }
 684
 685         rp->nmissed = 0;
 686         /* Establish function entry probe point */
 687         if ((ret = __register_kprobe(&rp->kp,
 688                 (unsigned long)__builtin_return_address(0))) != 0)
 689                 free_rp_inst(rp);
 690         return ret;
 691 }
 692
 693 #else /* ARCH_SUPPORTS_KRETPROBES */
 694
 695 int __kprobes register_kretprobe(struct kretprobe *rp)
 696 {
 697         return -ENOSYS;
 698 }
 699
 700 #endif /* ARCH_SUPPORTS_KRETPROBES */
 701
 702 void __kprobes unregister_kretprobe(struct kretprobe *rp)
 703 {
 704         unsigned long flags;
 705         struct kretprobe_instance *ri;
 706
 707         unregister_kprobe(&rp->kp);
 708         /* No race here */
 709         spin_lock_irqsave(&kretprobe_lock, flags);
 710         while ((ri = get_used_rp_inst(rp)) != NULL) {
 711                 ri->rp = NULL;
 712                 hlist_del(&ri->uflist);
 713         }
 714         spin_unlock_irqrestore(&kretprobe_lock, flags);
 715         free_rp_inst(rp);
 716 }
 717
 718 static int __init init_kprobes(void)
 719 {
 720         int i, err = 0;
 721
 722         /* FIXME allocate the probe table, currently defined statically */
 723         /* initialize all list heads */
 724         for (i = 0; i < KPROBE_TABLE_SIZE; i++) {
 725                 INIT_HLIST_HEAD(&kprobe_table[i]);
 726                 INIT_HLIST_HEAD(&kretprobe_inst_table[i]);
 727         }
 728         atomic_set(&kprobe_count, 0);
 729
 730         err = arch_init_kprobes();
 731         if (!err)
 732                 err = register_die_notifier(&kprobe_exceptions_nb);
 733
 734         return err;
 735 }
 736
 737 __initcall(init_kprobes);
 738
 739 EXPORT_SYMBOL_GPL(register_kprobe);
 740 EXPORT_SYMBOL_GPL(unregister_kprobe);
 741 EXPORT_SYMBOL_GPL(register_jprobe);
 742 EXPORT_SYMBOL_GPL(unregister_jprobe);
 743 EXPORT_SYMBOL_GPL(jprobe_return);
 744 EXPORT_SYMBOL_GPL(register_kretprobe);
 745 EXPORT_SYMBOL_GPL(unregister_kretprobe);
 746