git.oblomov.eu Git - linux-2.6/blob - arch/i386/kernel/kprobes.c

   1 /*
   2  *  Kernel Probes (KProbes)
   3  *  arch/i386/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 contributions from
  23  *              Rusty Russell).
  24  * 2004-July    Suparna Bhattacharya <suparna@in.ibm.com> added jumper probes
  25  *              interface to access function arguments.
  26  * 2005-May     Hien Nguyen <hien@us.ibm.com>, Jim Keniston
  27  *              <jkenisto@us.ibm.com> and Prasanna S Panchamukhi
  28  *              <prasanna@in.ibm.com> added function-return probes.
  29  */
  30
  31 #include <linux/config.h>
  32 #include <linux/kprobes.h>
  33 #include <linux/ptrace.h>
  34 #include <linux/preempt.h>
  35 #include <asm/cacheflush.h>
  36 #include <asm/kdebug.h>
  37 #include <asm/desc.h>
  38 #include <asm/uaccess.h>
  39
  40 void jprobe_return_end(void);
  41
  42 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
  43 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
  44
  45 /* insert a jmp code */
  46 static __always_inline void set_jmp_op(void *from, void *to)
  47 {
  48         struct __arch_jmp_op {
  49                 char op;
  50                 long raddr;
  51         } __attribute__((packed)) *jop;
  52         jop = (struct __arch_jmp_op *)from;
  53         jop->raddr = (long)(to) - ((long)(from) + 5);
  54         jop->op = RELATIVEJUMP_INSTRUCTION;
  55 }
  56
  57 /*
  58  * returns non-zero if opcodes can be boosted.
  59  */
  60 static __always_inline int can_boost(kprobe_opcode_t opcode)
  61 {
  62         switch (opcode & 0xf0 ) {
  63         case 0x70:
  64                 return 0; /* can't boost conditional jump */
  65         case 0x90:
  66                 /* can't boost call and pushf */
  67                 return opcode != 0x9a && opcode != 0x9c;
  68         case 0xc0:
  69                 /* can't boost undefined opcodes and soft-interruptions */
  70                 return (0xc1 < opcode && opcode < 0xc6) ||
  71                         (0xc7 < opcode && opcode < 0xcc) || opcode == 0xcf;
  72         case 0xd0:
  73                 /* can boost AA* and XLAT */
  74                 return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
  75         case 0xe0:
  76                 /* can boost in/out and (may be) jmps */
  77                 return (0xe3 < opcode && opcode != 0xe8);
  78         case 0xf0:
  79                 /* clear and set flags can be boost */
  80                 return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
  81         default:
  82                 /* currently, can't boost 2 bytes opcodes */
  83                 return opcode != 0x0f;
  84         }
  85 }
  86
  87
  88 /*
  89  * returns non-zero if opcode modifies the interrupt flag.
  90  */
  91 static int __kprobes is_IF_modifier(kprobe_opcode_t opcode)
  92 {
  93         switch (opcode) {
  94         case 0xfa:              /* cli */
  95         case 0xfb:              /* sti */
  96         case 0xcf:              /* iret/iretd */
  97         case 0x9d:              /* popf/popfd */
  98                 return 1;
  99         }
 100         return 0;
 101 }
 102
 103 int __kprobes arch_prepare_kprobe(struct kprobe *p)
 104 {
 105         /* insn: must be on special executable page on i386. */
 106         p->ainsn.insn = get_insn_slot();
 107         if (!p->ainsn.insn)
 108                 return -ENOMEM;
 109
 110         memcpy(p->ainsn.insn, p->addr, MAX_INSN_SIZE * sizeof(kprobe_opcode_t));
 111         p->opcode = *p->addr;
 112         if (can_boost(p->opcode)) {
 113                 p->ainsn.boostable = 0;
 114         } else {
 115                 p->ainsn.boostable = -1;
 116         }
 117         return 0;
 118 }
 119
 120 void __kprobes arch_arm_kprobe(struct kprobe *p)
 121 {
 122         *p->addr = BREAKPOINT_INSTRUCTION;
 123         flush_icache_range((unsigned long) p->addr,
 124                            (unsigned long) p->addr + sizeof(kprobe_opcode_t));
 125 }
 126
 127 void __kprobes arch_disarm_kprobe(struct kprobe *p)
 128 {
 129         *p->addr = p->opcode;
 130         flush_icache_range((unsigned long) p->addr,
 131                            (unsigned long) p->addr + sizeof(kprobe_opcode_t));
 132 }
 133
 134 void __kprobes arch_remove_kprobe(struct kprobe *p)
 135 {
 136         mutex_lock(&kprobe_mutex);
 137         free_insn_slot(p->ainsn.insn);
 138         mutex_unlock(&kprobe_mutex);
 139 }
 140
 141 static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
 142 {
 143         kcb->prev_kprobe.kp = kprobe_running();
 144         kcb->prev_kprobe.status = kcb->kprobe_status;
 145         kcb->prev_kprobe.old_eflags = kcb->kprobe_old_eflags;
 146         kcb->prev_kprobe.saved_eflags = kcb->kprobe_saved_eflags;
 147 }
 148
 149 static void __kprobes restore_previous_kprobe(struct kprobe_ctlblk *kcb)
 150 {
 151         __get_cpu_var(current_kprobe) = kcb->prev_kprobe.kp;
 152         kcb->kprobe_status = kcb->prev_kprobe.status;
 153         kcb->kprobe_old_eflags = kcb->prev_kprobe.old_eflags;
 154         kcb->kprobe_saved_eflags = kcb->prev_kprobe.saved_eflags;
 155 }
 156
 157 static void __kprobes set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
 158                                 struct kprobe_ctlblk *kcb)
 159 {
 160         __get_cpu_var(current_kprobe) = p;
 161         kcb->kprobe_saved_eflags = kcb->kprobe_old_eflags
 162                 = (regs->eflags & (TF_MASK | IF_MASK));
 163         if (is_IF_modifier(p->opcode))
 164                 kcb->kprobe_saved_eflags &= ~IF_MASK;
 165 }
 166
 167 static void __kprobes prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
 168 {
 169         regs->eflags |= TF_MASK;
 170         regs->eflags &= ~IF_MASK;
 171         /*single step inline if the instruction is an int3*/
 172         if (p->opcode == BREAKPOINT_INSTRUCTION)
 173                 regs->eip = (unsigned long)p->addr;
 174         else
 175                 regs->eip = (unsigned long)p->ainsn.insn;
 176 }
 177
 178 /* Called with kretprobe_lock held */
 179 void __kprobes arch_prepare_kretprobe(struct kretprobe *rp,
 180                                       struct pt_regs *regs)
 181 {
 182         unsigned long *sara = (unsigned long *)&regs->esp;
 183         struct kretprobe_instance *ri;
 184
 185         if ((ri = get_free_rp_inst(rp)) != NULL) {
 186                 ri->rp = rp;
 187                 ri->task = current;
 188                 ri->ret_addr = (kprobe_opcode_t *) *sara;
 189
 190                 /* Replace the return addr with trampoline addr */
 191                 *sara = (unsigned long) &kretprobe_trampoline;
 192
 193                 add_rp_inst(ri);
 194         } else {
 195                 rp->nmissed++;
 196         }
 197 }
 198
 199 /*
 200  * Interrupts are disabled on entry as trap3 is an interrupt gate and they
 201  * remain disabled thorough out this function.
 202  */
 203 static int __kprobes kprobe_handler(struct pt_regs *regs)
 204 {
 205         struct kprobe *p;
 206         int ret = 0;
 207         kprobe_opcode_t *addr;
 208         struct kprobe_ctlblk *kcb;
 209 #ifdef CONFIG_PREEMPT
 210         unsigned pre_preempt_count = preempt_count();
 211 #endif /* CONFIG_PREEMPT */
 212
 213         addr = (kprobe_opcode_t *)(regs->eip - sizeof(kprobe_opcode_t));
 214
 215         /*
 216          * We don't want to be preempted for the entire
 217          * duration of kprobe processing
 218          */
 219         preempt_disable();
 220         kcb = get_kprobe_ctlblk();
 221
 222         /* Check we're not actually recursing */
 223         if (kprobe_running()) {
 224                 p = get_kprobe(addr);
 225                 if (p) {
 226                         if (kcb->kprobe_status == KPROBE_HIT_SS &&
 227                                 *p->ainsn.insn == BREAKPOINT_INSTRUCTION) {
 228                                 regs->eflags &= ~TF_MASK;
 229                                 regs->eflags |= kcb->kprobe_saved_eflags;
 230                                 goto no_kprobe;
 231                         }
 232                         /* We have reentered the kprobe_handler(), since
 233                          * another probe was hit while within the handler.
 234                          * We here save the original kprobes variables and
 235                          * just single step on the instruction of the new probe
 236                          * without calling any user handlers.
 237                          */
 238                         save_previous_kprobe(kcb);
 239                         set_current_kprobe(p, regs, kcb);
 240                         kprobes_inc_nmissed_count(p);
 241                         prepare_singlestep(p, regs);
 242                         kcb->kprobe_status = KPROBE_REENTER;
 243                         return 1;
 244                 } else {
 245                         if (*addr != BREAKPOINT_INSTRUCTION) {
 246                         /* The breakpoint instruction was removed by
 247                          * another cpu right after we hit, no further
 248                          * handling of this interrupt is appropriate
 249                          */
 250                                 regs->eip -= sizeof(kprobe_opcode_t);
 251                                 ret = 1;
 252                                 goto no_kprobe;
 253                         }
 254                         p = __get_cpu_var(current_kprobe);
 255                         if (p->break_handler && p->break_handler(p, regs)) {
 256                                 goto ss_probe;
 257                         }
 258                 }
 259                 goto no_kprobe;
 260         }
 261
 262         p = get_kprobe(addr);
 263         if (!p) {
 264                 if (*addr != BREAKPOINT_INSTRUCTION) {
 265                         /*
 266                          * The breakpoint instruction was removed right
 267                          * after we hit it.  Another cpu has removed
 268                          * either a probepoint or a debugger breakpoint
 269                          * at this address.  In either case, no further
 270                          * handling of this interrupt is appropriate.
 271                          * Back up over the (now missing) int3 and run
 272                          * the original instruction.
 273                          */
 274                         regs->eip -= sizeof(kprobe_opcode_t);
 275                         ret = 1;
 276                 }
 277                 /* Not one of ours: let kernel handle it */
 278                 goto no_kprobe;
 279         }
 280
 281         set_current_kprobe(p, regs, kcb);
 282         kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 283
 284         if (p->pre_handler && p->pre_handler(p, regs))
 285                 /* handler has already set things up, so skip ss setup */
 286                 return 1;
 287
 288         if (p->ainsn.boostable == 1 &&
 289 #ifdef CONFIG_PREEMPT
 290             !(pre_preempt_count) && /*
 291                                        * This enables booster when the direct
 292                                        * execution path aren't preempted.
 293                                        */
 294 #endif /* CONFIG_PREEMPT */
 295             !p->post_handler && !p->break_handler ) {
 296                 /* Boost up -- we can execute copied instructions directly */
 297                 reset_current_kprobe();
 298                 regs->eip = (unsigned long)p->ainsn.insn;
 299                 preempt_enable_no_resched();
 300                 return 1;
 301         }
 302
 303 ss_probe:
 304         prepare_singlestep(p, regs);
 305         kcb->kprobe_status = KPROBE_HIT_SS;
 306         return 1;
 307
 308 no_kprobe:
 309         preempt_enable_no_resched();
 310         return ret;
 311 }
 312
 313 /*
 314  * For function-return probes, init_kprobes() establishes a probepoint
 315  * here. When a retprobed function returns, this probe is hit and
 316  * trampoline_probe_handler() runs, calling the kretprobe's handler.
 317  */
 318  void __kprobes kretprobe_trampoline_holder(void)
 319  {
 320         asm volatile ( ".global kretprobe_trampoline\n"
 321                         "kretprobe_trampoline: \n"
 322                         "       pushf\n"
 323                         /* skip cs, eip, orig_eax, es, ds */
 324                         "       subl $20, %esp\n"
 325                         "       pushl %eax\n"
 326                         "       pushl %ebp\n"
 327                         "       pushl %edi\n"
 328                         "       pushl %esi\n"
 329                         "       pushl %edx\n"
 330                         "       pushl %ecx\n"
 331                         "       pushl %ebx\n"
 332                         "       movl %esp, %eax\n"
 333                         "       call trampoline_handler\n"
 334                         /* move eflags to cs */
 335                         "       movl 48(%esp), %edx\n"
 336                         "       movl %edx, 44(%esp)\n"
 337                         /* save true return address on eflags */
 338                         "       movl %eax, 48(%esp)\n"
 339                         "       popl %ebx\n"
 340                         "       popl %ecx\n"
 341                         "       popl %edx\n"
 342                         "       popl %esi\n"
 343                         "       popl %edi\n"
 344                         "       popl %ebp\n"
 345                         "       popl %eax\n"
 346                         /* skip eip, orig_eax, es, ds */
 347                         "       addl $16, %esp\n"
 348                         "       popf\n"
 349                         "       ret\n");
 350 }
 351
 352 /*
 353  * Called from kretprobe_trampoline
 354  */
 355 fastcall void *__kprobes trampoline_handler(struct pt_regs *regs)
 356 {
 357         struct kretprobe_instance *ri = NULL;
 358         struct hlist_head *head;
 359         struct hlist_node *node, *tmp;
 360         unsigned long flags, orig_ret_address = 0;
 361         unsigned long trampoline_address =(unsigned long)&kretprobe_trampoline;
 362
 363         spin_lock_irqsave(&kretprobe_lock, flags);
 364         head = kretprobe_inst_table_head(current);
 365
 366         /*
 367          * It is possible to have multiple instances associated with a given
 368          * task either because an multiple functions in the call path
 369          * have a return probe installed on them, and/or more then one return
 370          * return probe was registered for a target function.
 371          *
 372          * We can handle this because:
 373          *     - instances are always inserted at the head of the list
 374          *     - when multiple return probes are registered for the same
 375          *       function, the first instance's ret_addr will point to the
 376          *       real return address, and all the rest will point to
 377          *       kretprobe_trampoline
 378          */
 379         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
 380                 if (ri->task != current)
 381                         /* another task is sharing our hash bucket */
 382                         continue;
 383
 384                 if (ri->rp && ri->rp->handler){
 385                         __get_cpu_var(current_kprobe) = &ri->rp->kp;
 386                         ri->rp->handler(ri, regs);
 387                         __get_cpu_var(current_kprobe) = NULL;
 388                 }
 389
 390                 orig_ret_address = (unsigned long)ri->ret_addr;
 391                 recycle_rp_inst(ri);
 392
 393                 if (orig_ret_address != trampoline_address)
 394                         /*
 395                          * This is the real return address. Any other
 396                          * instances associated with this task are for
 397                          * other calls deeper on the call stack
 398                          */
 399                         break;
 400         }
 401
 402         BUG_ON(!orig_ret_address || (orig_ret_address == trampoline_address));
 403
 404         spin_unlock_irqrestore(&kretprobe_lock, flags);
 405
 406         return (void*)orig_ret_address;
 407 }
 408
 409 /*
 410  * Called after single-stepping.  p->addr is the address of the
 411  * instruction whose first byte has been replaced by the "int 3"
 412  * instruction.  To avoid the SMP problems that can occur when we
 413  * temporarily put back the original opcode to single-step, we
 414  * single-stepped a copy of the instruction.  The address of this
 415  * copy is p->ainsn.insn.
 416  *
 417  * This function prepares to return from the post-single-step
 418  * interrupt.  We have to fix up the stack as follows:
 419  *
 420  * 0) Except in the case of absolute or indirect jump or call instructions,
 421  * the new eip is relative to the copied instruction.  We need to make
 422  * it relative to the original instruction.
 423  *
 424  * 1) If the single-stepped instruction was pushfl, then the TF and IF
 425  * flags are set in the just-pushed eflags, and may need to be cleared.
 426  *
 427  * 2) If the single-stepped instruction was a call, the return address
 428  * that is atop the stack is the address following the copied instruction.
 429  * We need to make it the address following the original instruction.
 430  *
 431  * This function also checks instruction size for preparing direct execution.
 432  */
 433 static void __kprobes resume_execution(struct kprobe *p,
 434                 struct pt_regs *regs, struct kprobe_ctlblk *kcb)
 435 {
 436         unsigned long *tos = (unsigned long *)&regs->esp;
 437         unsigned long copy_eip = (unsigned long)p->ainsn.insn;
 438         unsigned long orig_eip = (unsigned long)p->addr;
 439
 440         regs->eflags &= ~TF_MASK;
 441         switch (p->ainsn.insn[0]) {
 442         case 0x9c:              /* pushfl */
 443                 *tos &= ~(TF_MASK | IF_MASK);
 444                 *tos |= kcb->kprobe_old_eflags;
 445                 break;
 446         case 0xc2:              /* iret/ret/lret */
 447         case 0xc3:
 448         case 0xca:
 449         case 0xcb:
 450         case 0xcf:
 451         case 0xea:              /* jmp absolute -- eip is correct */
 452                 /* eip is already adjusted, no more changes required */
 453                 p->ainsn.boostable = 1;
 454                 goto no_change;
 455         case 0xe8:              /* call relative - Fix return addr */
 456                 *tos = orig_eip + (*tos - copy_eip);
 457                 break;
 458         case 0x9a:              /* call absolute -- same as call absolute, indirect */
 459                 *tos = orig_eip + (*tos - copy_eip);
 460                 goto no_change;
 461         case 0xff:
 462                 if ((p->ainsn.insn[1] & 0x30) == 0x10) {
 463                         /*
 464                          * call absolute, indirect
 465                          * Fix return addr; eip is correct.
 466                          * But this is not boostable
 467                          */
 468                         *tos = orig_eip + (*tos - copy_eip);
 469                         goto no_change;
 470                 } else if (((p->ainsn.insn[1] & 0x31) == 0x20) ||       /* jmp near, absolute indirect */
 471                            ((p->ainsn.insn[1] & 0x31) == 0x21)) {       /* jmp far, absolute indirect */
 472                         /* eip is correct. And this is boostable */
 473                         p->ainsn.boostable = 1;
 474                         goto no_change;
 475                 }
 476         default:
 477                 break;
 478         }
 479
 480         if (p->ainsn.boostable == 0) {
 481                 if ((regs->eip > copy_eip) &&
 482                     (regs->eip - copy_eip) + 5 < MAX_INSN_SIZE) {
 483                         /*
 484                          * These instructions can be executed directly if it
 485                          * jumps back to correct address.
 486                          */
 487                         set_jmp_op((void *)regs->eip,
 488                                    (void *)orig_eip + (regs->eip - copy_eip));
 489                         p->ainsn.boostable = 1;
 490                 } else {
 491                         p->ainsn.boostable = -1;
 492                 }
 493         }
 494
 495         regs->eip = orig_eip + (regs->eip - copy_eip);
 496
 497 no_change:
 498         return;
 499 }
 500
 501 /*
 502  * Interrupts are disabled on entry as trap1 is an interrupt gate and they
 503  * remain disabled thoroughout this function.
 504  */
 505 static int __kprobes post_kprobe_handler(struct pt_regs *regs)
 506 {
 507         struct kprobe *cur = kprobe_running();
 508         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 509
 510         if (!cur)
 511                 return 0;
 512
 513         if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
 514                 kcb->kprobe_status = KPROBE_HIT_SSDONE;
 515                 cur->post_handler(cur, regs, 0);
 516         }
 517
 518         resume_execution(cur, regs, kcb);
 519         regs->eflags |= kcb->kprobe_saved_eflags;
 520
 521         /*Restore back the original saved kprobes variables and continue. */
 522         if (kcb->kprobe_status == KPROBE_REENTER) {
 523                 restore_previous_kprobe(kcb);
 524                 goto out;
 525         }
 526         reset_current_kprobe();
 527 out:
 528         preempt_enable_no_resched();
 529
 530         /*
 531          * if somebody else is singlestepping across a probe point, eflags
 532          * will have TF set, in which case, continue the remaining processing
 533          * of do_debug, as if this is not a probe hit.
 534          */
 535         if (regs->eflags & TF_MASK)
 536                 return 0;
 537
 538         return 1;
 539 }
 540
 541 static int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
 542 {
 543         struct kprobe *cur = kprobe_running();
 544         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 545
 546         switch(kcb->kprobe_status) {
 547         case KPROBE_HIT_SS:
 548         case KPROBE_REENTER:
 549                 /*
 550                  * We are here because the instruction being single
 551                  * stepped caused a page fault. We reset the current
 552                  * kprobe and the eip points back to the probe address
 553                  * and allow the page fault handler to continue as a
 554                  * normal page fault.
 555                  */
 556                 regs->eip = (unsigned long)cur->addr;
 557                 regs->eflags |= kcb->kprobe_old_eflags;
 558                 if (kcb->kprobe_status == KPROBE_REENTER)
 559                         restore_previous_kprobe(kcb);
 560                 else
 561                         reset_current_kprobe();
 562                 preempt_enable_no_resched();
 563                 break;
 564         case KPROBE_HIT_ACTIVE:
 565         case KPROBE_HIT_SSDONE:
 566                 /*
 567                  * We increment the nmissed count for accounting,
 568                  * we can also use npre/npostfault count for accouting
 569                  * these specific fault cases.
 570                  */
 571                 kprobes_inc_nmissed_count(cur);
 572
 573                 /*
 574                  * We come here because instructions in the pre/post
 575                  * handler caused the page_fault, this could happen
 576                  * if handler tries to access user space by
 577                  * copy_from_user(), get_user() etc. Let the
 578                  * user-specified handler try to fix it first.
 579                  */
 580                 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
 581                         return 1;
 582
 583                 /*
 584                  * In case the user-specified fault handler returned
 585                  * zero, try to fix up.
 586                  */
 587                 if (fixup_exception(regs))
 588                         return 1;
 589
 590                 /*
 591                  * fixup_exception() could not handle it,
 592                  * Let do_page_fault() fix it.
 593                  */
 594                 break;
 595         default:
 596                 break;
 597         }
 598         return 0;
 599 }
 600
 601 /*
 602  * Wrapper routine to for handling exceptions.
 603  */
 604 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
 605                                        unsigned long val, void *data)
 606 {
 607         struct die_args *args = (struct die_args *)data;
 608         int ret = NOTIFY_DONE;
 609
 610         if (args->regs && user_mode_vm(args->regs))
 611                 return ret;
 612
 613         switch (val) {
 614         case DIE_INT3:
 615                 if (kprobe_handler(args->regs))
 616                         ret = NOTIFY_STOP;
 617                 break;
 618         case DIE_DEBUG:
 619                 if (post_kprobe_handler(args->regs))
 620                         ret = NOTIFY_STOP;
 621                 break;
 622         case DIE_GPF:
 623         case DIE_PAGE_FAULT:
 624                 /* kprobe_running() needs smp_processor_id() */
 625                 preempt_disable();
 626                 if (kprobe_running() &&
 627                     kprobe_fault_handler(args->regs, args->trapnr))
 628                         ret = NOTIFY_STOP;
 629                 preempt_enable();
 630                 break;
 631         default:
 632                 break;
 633         }
 634         return ret;
 635 }
 636
 637 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
 638 {
 639         struct jprobe *jp = container_of(p, struct jprobe, kp);
 640         unsigned long addr;
 641         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 642
 643         kcb->jprobe_saved_regs = *regs;
 644         kcb->jprobe_saved_esp = &regs->esp;
 645         addr = (unsigned long)(kcb->jprobe_saved_esp);
 646
 647         /*
 648          * TBD: As Linus pointed out, gcc assumes that the callee
 649          * owns the argument space and could overwrite it, e.g.
 650          * tailcall optimization. So, to be absolutely safe
 651          * we also save and restore enough stack bytes to cover
 652          * the argument area.
 653          */
 654         memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
 655                         MIN_STACK_SIZE(addr));
 656         regs->eflags &= ~IF_MASK;
 657         regs->eip = (unsigned long)(jp->entry);
 658         return 1;
 659 }
 660
 661 void __kprobes jprobe_return(void)
 662 {
 663         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 664
 665         asm volatile ("       xchgl   %%ebx,%%esp     \n"
 666                       "       int3                      \n"
 667                       "       .globl jprobe_return_end  \n"
 668                       "       jprobe_return_end:        \n"
 669                       "       nop                       \n"::"b"
 670                       (kcb->jprobe_saved_esp):"memory");
 671 }
 672
 673 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
 674 {
 675         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 676         u8 *addr = (u8 *) (regs->eip - 1);
 677         unsigned long stack_addr = (unsigned long)(kcb->jprobe_saved_esp);
 678         struct jprobe *jp = container_of(p, struct jprobe, kp);
 679
 680         if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
 681                 if (&regs->esp != kcb->jprobe_saved_esp) {
 682                         struct pt_regs *saved_regs =
 683                             container_of(kcb->jprobe_saved_esp,
 684                                             struct pt_regs, esp);
 685                         printk("current esp %p does not match saved esp %p\n",
 686                                &regs->esp, kcb->jprobe_saved_esp);
 687                         printk("Saved registers for jprobe %p\n", jp);
 688                         show_registers(saved_regs);
 689                         printk("Current registers\n");
 690                         show_registers(regs);
 691                         BUG();
 692                 }
 693                 *regs = kcb->jprobe_saved_regs;
 694                 memcpy((kprobe_opcode_t *) stack_addr, kcb->jprobes_stack,
 695                        MIN_STACK_SIZE(stack_addr));
 696                 preempt_enable_no_resched();
 697                 return 1;
 698         }
 699         return 0;
 700 }
 701
 702 int __init arch_init_kprobes(void)
 703 {
 704         return 0;
 705 }