git.oblomov.eu Git - linux-2.6/blob - arch/x86/kernel/kprobes_32.c

   1 /*
   2  *  Kernel Probes (KProbes)
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License as published by
   6  * the Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write to the Free Software
  16  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  17  *
  18  * Copyright (C) IBM Corporation, 2002, 2004
  19  *
  20  * 2002-Oct     Created by Vamsi Krishna S <vamsi_krishna@in.ibm.com> Kernel
  21  *              Probes initial implementation ( includes contributions from
  22  *              Rusty Russell).
  23  * 2004-July    Suparna Bhattacharya <suparna@in.ibm.com> added jumper probes
  24  *              interface to access function arguments.
  25  * 2005-May     Hien Nguyen <hien@us.ibm.com>, Jim Keniston
  26  *              <jkenisto@us.ibm.com> and Prasanna S Panchamukhi
  27  *              <prasanna@in.ibm.com> added function-return probes.
  28  */
  29
  30 #include <linux/kprobes.h>
  31 #include <linux/ptrace.h>
  32 #include <linux/preempt.h>
  33 #include <linux/kdebug.h>
  34 #include <asm/cacheflush.h>
  35 #include <asm/desc.h>
  36 #include <asm/uaccess.h>
  37 #include <asm/alternative.h>
  38
  39 void jprobe_return_end(void);
  40
  41 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
  42 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
  43
  44 /* insert a jmp code */
  45 static __always_inline void set_jmp_op(void *from, void *to)
  46 {
  47         struct __arch_jmp_op {
  48                 char op;
  49                 long raddr;
  50         } __attribute__((packed)) *jop;
  51         jop = (struct __arch_jmp_op *)from;
  52         jop->raddr = (long)(to) - ((long)(from) + 5);
  53         jop->op = RELATIVEJUMP_INSTRUCTION;
  54 }
  55
  56 /*
  57  * returns non-zero if opcodes can be boosted.
  58  */
  59 static __always_inline int can_boost(kprobe_opcode_t *opcodes)
  60 {
  61 #define W(row,b0,b1,b2,b3,b4,b5,b6,b7,b8,b9,ba,bb,bc,bd,be,bf)                \
  62         (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
  63           (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
  64           (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
  65           (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
  66          << (row % 32))
  67         /*
  68          * Undefined/reserved opcodes, conditional jump, Opcode Extension
  69          * Groups, and some special opcodes can not be boost.
  70          */
  71         static const unsigned long twobyte_is_boostable[256 / 32] = {
  72                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
  73                 /*      -------------------------------         */
  74                 W(0x00, 0,0,1,1,0,0,1,0,1,1,0,0,0,0,0,0)| /* 00 */
  75                 W(0x10, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 10 */
  76                 W(0x20, 1,1,1,1,0,0,0,0,0,0,0,0,0,0,0,0)| /* 20 */
  77                 W(0x30, 0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 30 */
  78                 W(0x40, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 40 */
  79                 W(0x50, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 50 */
  80                 W(0x60, 1,1,1,1,1,1,1,1,1,1,1,1,0,0,1,1)| /* 60 */
  81                 W(0x70, 0,0,0,0,1,1,1,1,0,0,0,0,0,0,1,1), /* 70 */
  82                 W(0x80, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 80 */
  83                 W(0x90, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1), /* 90 */
  84                 W(0xa0, 1,1,0,1,1,1,0,0,1,1,0,1,1,1,0,1)| /* a0 */
  85                 W(0xb0, 1,1,1,1,1,1,1,1,0,0,0,1,1,1,1,1), /* b0 */
  86                 W(0xc0, 1,1,0,0,0,0,0,0,1,1,1,1,1,1,1,1)| /* c0 */
  87                 W(0xd0, 0,1,1,1,0,1,0,0,1,1,0,1,1,1,0,1), /* d0 */
  88                 W(0xe0, 0,1,1,0,0,1,0,0,1,1,0,1,1,1,0,1)| /* e0 */
  89                 W(0xf0, 0,1,1,1,0,1,0,0,1,1,1,0,1,1,1,0)  /* f0 */
  90                 /*      -------------------------------         */
  91                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
  92         };
  93 #undef W
  94         kprobe_opcode_t opcode;
  95         kprobe_opcode_t *orig_opcodes = opcodes;
  96 retry:
  97         if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
  98                 return 0;
  99         opcode = *(opcodes++);
 100
 101         /* 2nd-byte opcode */
 102         if (opcode == 0x0f) {
 103                 if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
 104                         return 0;
 105                 return test_bit(*opcodes, twobyte_is_boostable);
 106         }
 107
 108         switch (opcode & 0xf0) {
 109         case 0x60:
 110                 if (0x63 < opcode && opcode < 0x67)
 111                         goto retry; /* prefixes */
 112                 /* can't boost Address-size override and bound */
 113                 return (opcode != 0x62 && opcode != 0x67);
 114         case 0x70:
 115                 return 0; /* can't boost conditional jump */
 116         case 0xc0:
 117                 /* can't boost software-interruptions */
 118                 return (0xc1 < opcode && opcode < 0xcc) || opcode == 0xcf;
 119         case 0xd0:
 120                 /* can boost AA* and XLAT */
 121                 return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
 122         case 0xe0:
 123                 /* can boost in/out and absolute jmps */
 124                 return ((opcode & 0x04) || opcode == 0xea);
 125         case 0xf0:
 126                 if ((opcode & 0x0c) == 0 && opcode != 0xf1)
 127                         goto retry; /* lock/rep(ne) prefix */
 128                 /* clear and set flags can be boost */
 129                 return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
 130         default:
 131                 if (opcode == 0x26 || opcode == 0x36 || opcode == 0x3e)
 132                         goto retry; /* prefixes */
 133                 /* can't boost CS override and call */
 134                 return (opcode != 0x2e && opcode != 0x9a);
 135         }
 136 }
 137
 138 /*
 139  * returns non-zero if opcode modifies the interrupt flag.
 140  */
 141 static int __kprobes is_IF_modifier(kprobe_opcode_t opcode)
 142 {
 143         switch (opcode) {
 144         case 0xfa:              /* cli */
 145         case 0xfb:              /* sti */
 146         case 0xcf:              /* iret/iretd */
 147         case 0x9d:              /* popf/popfd */
 148                 return 1;
 149         }
 150         return 0;
 151 }
 152
 153 int __kprobes arch_prepare_kprobe(struct kprobe *p)
 154 {
 155         /* insn: must be on special executable page on i386. */
 156         p->ainsn.insn = get_insn_slot();
 157         if (!p->ainsn.insn)
 158                 return -ENOMEM;
 159
 160         memcpy(p->ainsn.insn, p->addr, MAX_INSN_SIZE * sizeof(kprobe_opcode_t));
 161         p->opcode = *p->addr;
 162         if (can_boost(p->addr)) {
 163                 p->ainsn.boostable = 0;
 164         } else {
 165                 p->ainsn.boostable = -1;
 166         }
 167         return 0;
 168 }
 169
 170 void __kprobes arch_arm_kprobe(struct kprobe *p)
 171 {
 172         text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
 173 }
 174
 175 void __kprobes arch_disarm_kprobe(struct kprobe *p)
 176 {
 177         text_poke(p->addr, &p->opcode, 1);
 178 }
 179
 180 void __kprobes arch_remove_kprobe(struct kprobe *p)
 181 {
 182         mutex_lock(&kprobe_mutex);
 183         free_insn_slot(p->ainsn.insn, (p->ainsn.boostable == 1));
 184         mutex_unlock(&kprobe_mutex);
 185 }
 186
 187 static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
 188 {
 189         kcb->prev_kprobe.kp = kprobe_running();
 190         kcb->prev_kprobe.status = kcb->kprobe_status;
 191         kcb->prev_kprobe.old_eflags = kcb->kprobe_old_eflags;
 192         kcb->prev_kprobe.saved_eflags = kcb->kprobe_saved_eflags;
 193 }
 194
 195 static void __kprobes restore_previous_kprobe(struct kprobe_ctlblk *kcb)
 196 {
 197         __get_cpu_var(current_kprobe) = kcb->prev_kprobe.kp;
 198         kcb->kprobe_status = kcb->prev_kprobe.status;
 199         kcb->kprobe_old_eflags = kcb->prev_kprobe.old_eflags;
 200         kcb->kprobe_saved_eflags = kcb->prev_kprobe.saved_eflags;
 201 }
 202
 203 static void __kprobes set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
 204                                 struct kprobe_ctlblk *kcb)
 205 {
 206         __get_cpu_var(current_kprobe) = p;
 207         kcb->kprobe_saved_eflags = kcb->kprobe_old_eflags
 208                 = (regs->eflags & (TF_MASK | IF_MASK));
 209         if (is_IF_modifier(p->opcode))
 210                 kcb->kprobe_saved_eflags &= ~IF_MASK;
 211 }
 212
 213 static void __kprobes prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
 214 {
 215         regs->eflags |= TF_MASK;
 216         regs->eflags &= ~IF_MASK;
 217         /*single step inline if the instruction is an int3*/
 218         if (p->opcode == BREAKPOINT_INSTRUCTION)
 219                 regs->eip = (unsigned long)p->addr;
 220         else
 221                 regs->eip = (unsigned long)p->ainsn.insn;
 222 }
 223
 224 /* Called with kretprobe_lock held */
 225 void __kprobes arch_prepare_kretprobe(struct kretprobe_instance *ri,
 226                                       struct pt_regs *regs)
 227 {
 228         unsigned long *sara = (unsigned long *)&regs->esp;
 229
 230         ri->ret_addr = (kprobe_opcode_t *) *sara;
 231
 232         /* Replace the return addr with trampoline addr */
 233         *sara = (unsigned long) &kretprobe_trampoline;
 234 }
 235
 236 /*
 237  * Interrupts are disabled on entry as trap3 is an interrupt gate and they
 238  * remain disabled thorough out this function.
 239  */
 240 static int __kprobes kprobe_handler(struct pt_regs *regs)
 241 {
 242         struct kprobe *p;
 243         int ret = 0;
 244         kprobe_opcode_t *addr;
 245         struct kprobe_ctlblk *kcb;
 246
 247         addr = (kprobe_opcode_t *)(regs->eip - sizeof(kprobe_opcode_t));
 248
 249         /*
 250          * We don't want to be preempted for the entire
 251          * duration of kprobe processing
 252          */
 253         preempt_disable();
 254         kcb = get_kprobe_ctlblk();
 255
 256         /* Check we're not actually recursing */
 257         if (kprobe_running()) {
 258                 p = get_kprobe(addr);
 259                 if (p) {
 260                         if (kcb->kprobe_status == KPROBE_HIT_SS &&
 261                                 *p->ainsn.insn == BREAKPOINT_INSTRUCTION) {
 262                                 regs->eflags &= ~TF_MASK;
 263                                 regs->eflags |= kcb->kprobe_saved_eflags;
 264                                 goto no_kprobe;
 265                         }
 266                         /* We have reentered the kprobe_handler(), since
 267                          * another probe was hit while within the handler.
 268                          * We here save the original kprobes variables and
 269                          * just single step on the instruction of the new probe
 270                          * without calling any user handlers.
 271                          */
 272                         save_previous_kprobe(kcb);
 273                         set_current_kprobe(p, regs, kcb);
 274                         kprobes_inc_nmissed_count(p);
 275                         prepare_singlestep(p, regs);
 276                         kcb->kprobe_status = KPROBE_REENTER;
 277                         return 1;
 278                 } else {
 279                         if (*addr != BREAKPOINT_INSTRUCTION) {
 280                         /* The breakpoint instruction was removed by
 281                          * another cpu right after we hit, no further
 282                          * handling of this interrupt is appropriate
 283                          */
 284                                 regs->eip -= sizeof(kprobe_opcode_t);
 285                                 ret = 1;
 286                                 goto no_kprobe;
 287                         }
 288                         p = __get_cpu_var(current_kprobe);
 289                         if (p->break_handler && p->break_handler(p, regs)) {
 290                                 goto ss_probe;
 291                         }
 292                 }
 293                 goto no_kprobe;
 294         }
 295
 296         p = get_kprobe(addr);
 297         if (!p) {
 298                 if (*addr != BREAKPOINT_INSTRUCTION) {
 299                         /*
 300                          * The breakpoint instruction was removed right
 301                          * after we hit it.  Another cpu has removed
 302                          * either a probepoint or a debugger breakpoint
 303                          * at this address.  In either case, no further
 304                          * handling of this interrupt is appropriate.
 305                          * Back up over the (now missing) int3 and run
 306                          * the original instruction.
 307                          */
 308                         regs->eip -= sizeof(kprobe_opcode_t);
 309                         ret = 1;
 310                 }
 311                 /* Not one of ours: let kernel handle it */
 312                 goto no_kprobe;
 313         }
 314
 315         set_current_kprobe(p, regs, kcb);
 316         kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 317
 318         if (p->pre_handler && p->pre_handler(p, regs))
 319                 /* handler has already set things up, so skip ss setup */
 320                 return 1;
 321
 322 ss_probe:
 323 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PM)
 324         if (p->ainsn.boostable == 1 && !p->post_handler){
 325                 /* Boost up -- we can execute copied instructions directly */
 326                 reset_current_kprobe();
 327                 regs->eip = (unsigned long)p->ainsn.insn;
 328                 preempt_enable_no_resched();
 329                 return 1;
 330         }
 331 #endif
 332         prepare_singlestep(p, regs);
 333         kcb->kprobe_status = KPROBE_HIT_SS;
 334         return 1;
 335
 336 no_kprobe:
 337         preempt_enable_no_resched();
 338         return ret;
 339 }
 340
 341 /*
 342  * For function-return probes, init_kprobes() establishes a probepoint
 343  * here. When a retprobed function returns, this probe is hit and
 344  * trampoline_probe_handler() runs, calling the kretprobe's handler.
 345  */
 346  void __kprobes kretprobe_trampoline_holder(void)
 347  {
 348         asm volatile ( ".global kretprobe_trampoline\n"
 349                         "kretprobe_trampoline: \n"
 350                         "       pushf\n"
 351                         /* skip cs, eip, orig_eax */
 352                         "       subl $12, %esp\n"
 353                         "       pushl %fs\n"
 354                         "       pushl %ds\n"
 355                         "       pushl %es\n"
 356                         "       pushl %eax\n"
 357                         "       pushl %ebp\n"
 358                         "       pushl %edi\n"
 359                         "       pushl %esi\n"
 360                         "       pushl %edx\n"
 361                         "       pushl %ecx\n"
 362                         "       pushl %ebx\n"
 363                         "       movl %esp, %eax\n"
 364                         "       call trampoline_handler\n"
 365                         /* move eflags to cs */
 366                         "       movl 52(%esp), %edx\n"
 367                         "       movl %edx, 48(%esp)\n"
 368                         /* save true return address on eflags */
 369                         "       movl %eax, 52(%esp)\n"
 370                         "       popl %ebx\n"
 371                         "       popl %ecx\n"
 372                         "       popl %edx\n"
 373                         "       popl %esi\n"
 374                         "       popl %edi\n"
 375                         "       popl %ebp\n"
 376                         "       popl %eax\n"
 377                         /* skip eip, orig_eax, es, ds, fs */
 378                         "       addl $20, %esp\n"
 379                         "       popf\n"
 380                         "       ret\n");
 381 }
 382
 383 /*
 384  * Called from kretprobe_trampoline
 385  */
 386 fastcall void *__kprobes trampoline_handler(struct pt_regs *regs)
 387 {
 388         struct kretprobe_instance *ri = NULL;
 389         struct hlist_head *head, empty_rp;
 390         struct hlist_node *node, *tmp;
 391         unsigned long flags, orig_ret_address = 0;
 392         unsigned long trampoline_address =(unsigned long)&kretprobe_trampoline;
 393
 394         INIT_HLIST_HEAD(&empty_rp);
 395         spin_lock_irqsave(&kretprobe_lock, flags);
 396         head = kretprobe_inst_table_head(current);
 397         /* fixup registers */
 398         regs->xcs = __KERNEL_CS | get_kernel_rpl();
 399         regs->eip = trampoline_address;
 400         regs->orig_eax = 0xffffffff;
 401
 402         /*
 403          * It is possible to have multiple instances associated with a given
 404          * task either because an multiple functions in the call path
 405          * have a return probe installed on them, and/or more then one return
 406          * return probe was registered for a target function.
 407          *
 408          * We can handle this because:
 409          *     - instances are always inserted at the head of the list
 410          *     - when multiple return probes are registered for the same
 411          *       function, the first instance's ret_addr will point to the
 412          *       real return address, and all the rest will point to
 413          *       kretprobe_trampoline
 414          */
 415         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
 416                 if (ri->task != current)
 417                         /* another task is sharing our hash bucket */
 418                         continue;
 419
 420                 if (ri->rp && ri->rp->handler){
 421                         __get_cpu_var(current_kprobe) = &ri->rp->kp;
 422                         get_kprobe_ctlblk()->kprobe_status = KPROBE_HIT_ACTIVE;
 423                         ri->rp->handler(ri, regs);
 424                         __get_cpu_var(current_kprobe) = NULL;
 425                 }
 426
 427                 orig_ret_address = (unsigned long)ri->ret_addr;
 428                 recycle_rp_inst(ri, &empty_rp);
 429
 430                 if (orig_ret_address != trampoline_address)
 431                         /*
 432                          * This is the real return address. Any other
 433                          * instances associated with this task are for
 434                          * other calls deeper on the call stack
 435                          */
 436                         break;
 437         }
 438
 439         kretprobe_assert(ri, orig_ret_address, trampoline_address);
 440         spin_unlock_irqrestore(&kretprobe_lock, flags);
 441
 442         hlist_for_each_entry_safe(ri, node, tmp, &empty_rp, hlist) {
 443                 hlist_del(&ri->hlist);
 444                 kfree(ri);
 445         }
 446         return (void*)orig_ret_address;
 447 }
 448
 449 /*
 450  * Called after single-stepping.  p->addr is the address of the
 451  * instruction whose first byte has been replaced by the "int 3"
 452  * instruction.  To avoid the SMP problems that can occur when we
 453  * temporarily put back the original opcode to single-step, we
 454  * single-stepped a copy of the instruction.  The address of this
 455  * copy is p->ainsn.insn.
 456  *
 457  * This function prepares to return from the post-single-step
 458  * interrupt.  We have to fix up the stack as follows:
 459  *
 460  * 0) Except in the case of absolute or indirect jump or call instructions,
 461  * the new eip is relative to the copied instruction.  We need to make
 462  * it relative to the original instruction.
 463  *
 464  * 1) If the single-stepped instruction was pushfl, then the TF and IF
 465  * flags are set in the just-pushed eflags, and may need to be cleared.
 466  *
 467  * 2) If the single-stepped instruction was a call, the return address
 468  * that is atop the stack is the address following the copied instruction.
 469  * We need to make it the address following the original instruction.
 470  *
 471  * This function also checks instruction size for preparing direct execution.
 472  */
 473 static void __kprobes resume_execution(struct kprobe *p,
 474                 struct pt_regs *regs, struct kprobe_ctlblk *kcb)
 475 {
 476         unsigned long *tos = (unsigned long *)&regs->esp;
 477         unsigned long copy_eip = (unsigned long)p->ainsn.insn;
 478         unsigned long orig_eip = (unsigned long)p->addr;
 479
 480         regs->eflags &= ~TF_MASK;
 481         switch (p->ainsn.insn[0]) {
 482         case 0x9c:              /* pushfl */
 483                 *tos &= ~(TF_MASK | IF_MASK);
 484                 *tos |= kcb->kprobe_old_eflags;
 485                 break;
 486         case 0xc2:              /* iret/ret/lret */
 487         case 0xc3:
 488         case 0xca:
 489         case 0xcb:
 490         case 0xcf:
 491         case 0xea:              /* jmp absolute -- eip is correct */
 492                 /* eip is already adjusted, no more changes required */
 493                 p->ainsn.boostable = 1;
 494                 goto no_change;
 495         case 0xe8:              /* call relative - Fix return addr */
 496                 *tos = orig_eip + (*tos - copy_eip);
 497                 break;
 498         case 0x9a:              /* call absolute -- same as call absolute, indirect */
 499                 *tos = orig_eip + (*tos - copy_eip);
 500                 goto no_change;
 501         case 0xff:
 502                 if ((p->ainsn.insn[1] & 0x30) == 0x10) {
 503                         /*
 504                          * call absolute, indirect
 505                          * Fix return addr; eip is correct.
 506                          * But this is not boostable
 507                          */
 508                         *tos = orig_eip + (*tos - copy_eip);
 509                         goto no_change;
 510                 } else if (((p->ainsn.insn[1] & 0x31) == 0x20) ||       /* jmp near, absolute indirect */
 511                            ((p->ainsn.insn[1] & 0x31) == 0x21)) {       /* jmp far, absolute indirect */
 512                         /* eip is correct. And this is boostable */
 513                         p->ainsn.boostable = 1;
 514                         goto no_change;
 515                 }
 516         default:
 517                 break;
 518         }
 519
 520         if (p->ainsn.boostable == 0) {
 521                 if ((regs->eip > copy_eip) &&
 522                     (regs->eip - copy_eip) + 5 < MAX_INSN_SIZE) {
 523                         /*
 524                          * These instructions can be executed directly if it
 525                          * jumps back to correct address.
 526                          */
 527                         set_jmp_op((void *)regs->eip,
 528                                    (void *)orig_eip + (regs->eip - copy_eip));
 529                         p->ainsn.boostable = 1;
 530                 } else {
 531                         p->ainsn.boostable = -1;
 532                 }
 533         }
 534
 535         regs->eip = orig_eip + (regs->eip - copy_eip);
 536
 537 no_change:
 538         return;
 539 }
 540
 541 /*
 542  * Interrupts are disabled on entry as trap1 is an interrupt gate and they
 543  * remain disabled thoroughout this function.
 544  */
 545 static int __kprobes post_kprobe_handler(struct pt_regs *regs)
 546 {
 547         struct kprobe *cur = kprobe_running();
 548         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 549
 550         if (!cur)
 551                 return 0;
 552
 553         if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
 554                 kcb->kprobe_status = KPROBE_HIT_SSDONE;
 555                 cur->post_handler(cur, regs, 0);
 556         }
 557
 558         resume_execution(cur, regs, kcb);
 559         regs->eflags |= kcb->kprobe_saved_eflags;
 560 #ifdef CONFIG_TRACE_IRQFLAGS_SUPPORT
 561         if (raw_irqs_disabled_flags(regs->eflags))
 562                 trace_hardirqs_off();
 563         else
 564                 trace_hardirqs_on();
 565 #endif
 566
 567         /*Restore back the original saved kprobes variables and continue. */
 568         if (kcb->kprobe_status == KPROBE_REENTER) {
 569                 restore_previous_kprobe(kcb);
 570                 goto out;
 571         }
 572         reset_current_kprobe();
 573 out:
 574         preempt_enable_no_resched();
 575
 576         /*
 577          * if somebody else is singlestepping across a probe point, eflags
 578          * will have TF set, in which case, continue the remaining processing
 579          * of do_debug, as if this is not a probe hit.
 580          */
 581         if (regs->eflags & TF_MASK)
 582                 return 0;
 583
 584         return 1;
 585 }
 586
 587 int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
 588 {
 589         struct kprobe *cur = kprobe_running();
 590         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 591
 592         switch(kcb->kprobe_status) {
 593         case KPROBE_HIT_SS:
 594         case KPROBE_REENTER:
 595                 /*
 596                  * We are here because the instruction being single
 597                  * stepped caused a page fault. We reset the current
 598                  * kprobe and the eip points back to the probe address
 599                  * and allow the page fault handler to continue as a
 600                  * normal page fault.
 601                  */
 602                 regs->eip = (unsigned long)cur->addr;
 603                 regs->eflags |= kcb->kprobe_old_eflags;
 604                 if (kcb->kprobe_status == KPROBE_REENTER)
 605                         restore_previous_kprobe(kcb);
 606                 else
 607                         reset_current_kprobe();
 608                 preempt_enable_no_resched();
 609                 break;
 610         case KPROBE_HIT_ACTIVE:
 611         case KPROBE_HIT_SSDONE:
 612                 /*
 613                  * We increment the nmissed count for accounting,
 614                  * we can also use npre/npostfault count for accouting
 615                  * these specific fault cases.
 616                  */
 617                 kprobes_inc_nmissed_count(cur);
 618
 619                 /*
 620                  * We come here because instructions in the pre/post
 621                  * handler caused the page_fault, this could happen
 622                  * if handler tries to access user space by
 623                  * copy_from_user(), get_user() etc. Let the
 624                  * user-specified handler try to fix it first.
 625                  */
 626                 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
 627                         return 1;
 628
 629                 /*
 630                  * In case the user-specified fault handler returned
 631                  * zero, try to fix up.
 632                  */
 633                 if (fixup_exception(regs))
 634                         return 1;
 635
 636                 /*
 637                  * fixup_exception() could not handle it,
 638                  * Let do_page_fault() fix it.
 639                  */
 640                 break;
 641         default:
 642                 break;
 643         }
 644         return 0;
 645 }
 646
 647 /*
 648  * Wrapper routine to for handling exceptions.
 649  */
 650 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
 651                                        unsigned long val, void *data)
 652 {
 653         struct die_args *args = (struct die_args *)data;
 654         int ret = NOTIFY_DONE;
 655
 656         if (args->regs && user_mode_vm(args->regs))
 657                 return ret;
 658
 659         switch (val) {
 660         case DIE_INT3:
 661                 if (kprobe_handler(args->regs))
 662                         ret = NOTIFY_STOP;
 663                 break;
 664         case DIE_DEBUG:
 665                 if (post_kprobe_handler(args->regs))
 666                         ret = NOTIFY_STOP;
 667                 break;
 668         case DIE_GPF:
 669                 /* kprobe_running() needs smp_processor_id() */
 670                 preempt_disable();
 671                 if (kprobe_running() &&
 672                     kprobe_fault_handler(args->regs, args->trapnr))
 673                         ret = NOTIFY_STOP;
 674                 preempt_enable();
 675                 break;
 676         default:
 677                 break;
 678         }
 679         return ret;
 680 }
 681
 682 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
 683 {
 684         struct jprobe *jp = container_of(p, struct jprobe, kp);
 685         unsigned long addr;
 686         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 687
 688         kcb->jprobe_saved_regs = *regs;
 689         kcb->jprobe_saved_esp = &regs->esp;
 690         addr = (unsigned long)(kcb->jprobe_saved_esp);
 691
 692         /*
 693          * TBD: As Linus pointed out, gcc assumes that the callee
 694          * owns the argument space and could overwrite it, e.g.
 695          * tailcall optimization. So, to be absolutely safe
 696          * we also save and restore enough stack bytes to cover
 697          * the argument area.
 698          */
 699         memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
 700                         MIN_STACK_SIZE(addr));
 701         regs->eflags &= ~IF_MASK;
 702         trace_hardirqs_off();
 703         regs->eip = (unsigned long)(jp->entry);
 704         return 1;
 705 }
 706
 707 void __kprobes jprobe_return(void)
 708 {
 709         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 710
 711         asm volatile ("       xchgl   %%ebx,%%esp     \n"
 712                       "       int3                      \n"
 713                       "       .globl jprobe_return_end  \n"
 714                       "       jprobe_return_end:        \n"
 715                       "       nop                       \n"::"b"
 716                       (kcb->jprobe_saved_esp):"memory");
 717 }
 718
 719 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
 720 {
 721         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 722         u8 *addr = (u8 *) (regs->eip - 1);
 723         unsigned long stack_addr = (unsigned long)(kcb->jprobe_saved_esp);
 724         struct jprobe *jp = container_of(p, struct jprobe, kp);
 725
 726         if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
 727                 if (&regs->esp != kcb->jprobe_saved_esp) {
 728                         struct pt_regs *saved_regs =
 729                             container_of(kcb->jprobe_saved_esp,
 730                                             struct pt_regs, esp);
 731                         printk("current esp %p does not match saved esp %p\n",
 732                                &regs->esp, kcb->jprobe_saved_esp);
 733                         printk("Saved registers for jprobe %p\n", jp);
 734                         show_registers(saved_regs);
 735                         printk("Current registers\n");
 736                         show_registers(regs);
 737                         BUG();
 738                 }
 739                 *regs = kcb->jprobe_saved_regs;
 740                 memcpy((kprobe_opcode_t *) stack_addr, kcb->jprobes_stack,
 741                        MIN_STACK_SIZE(stack_addr));
 742                 preempt_enable_no_resched();
 743                 return 1;
 744         }
 745         return 0;
 746 }
 747
 748 int __kprobes arch_trampoline_kprobe(struct kprobe *p)
 749 {
 750         return 0;
 751 }
 752
 753 int __init arch_init_kprobes(void)
 754 {
 755         return 0;
 756 }