arch/x86_64/kernel/kprobes.c

   1 /*
   2  *  Kernel Probes (KProbes)
   3  *  arch/x86_64/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  * 2004-Oct     Jim Keniston <kenistoj@us.ibm.com> and Prasanna S Panchamukhi
  27  *              <prasanna@in.ibm.com> adapted for x86_64
  28  * 2005-Mar     Roland McGrath <roland@redhat.com>
  29  *              Fixed to handle %rip-relative addressing mode correctly.
  30  * 2005-May     Rusty Lynch <rusty.lynch@intel.com>
  31  *              Added function return probes functionality
  32  */
  33
  34 #include <linux/config.h>
  35 #include <linux/kprobes.h>
  36 #include <linux/ptrace.h>
  37 #include <linux/spinlock.h>
  38 #include <linux/string.h>
  39 #include <linux/slab.h>
  40 #include <linux/preempt.h>
  41
  42 #include <asm/cacheflush.h>
  43 #include <asm/pgtable.h>
  44 #include <asm/kdebug.h>
  45
  46 static DECLARE_MUTEX(kprobe_mutex);
  47
  48 static struct kprobe *current_kprobe;
  49 static unsigned long kprobe_status, kprobe_old_rflags, kprobe_saved_rflags;
  50 static struct kprobe *kprobe_prev;
  51 static unsigned long kprobe_status_prev, kprobe_old_rflags_prev, kprobe_saved_rflags_prev;
  52 static struct pt_regs jprobe_saved_regs;
  53 static long *jprobe_saved_rsp;
  54 void jprobe_return_end(void);
  55
  56 /* copy of the kernel stack at the probe fire time */
  57 static kprobe_opcode_t jprobes_stack[MAX_STACK_SIZE];
  58
  59 /*
  60  * returns non-zero if opcode modifies the interrupt flag.
  61  */
  62 static inline int is_IF_modifier(kprobe_opcode_t *insn)
  63 {
  64         switch (*insn) {
  65         case 0xfa:              /* cli */
  66         case 0xfb:              /* sti */
  67         case 0xcf:              /* iret/iretd */
  68         case 0x9d:              /* popf/popfd */
  69                 return 1;
  70         }
  71
  72         if (*insn  >= 0x40 && *insn <= 0x4f && *++insn == 0xcf)
  73                 return 1;
  74         return 0;
  75 }
  76
  77 int __kprobes arch_prepare_kprobe(struct kprobe *p)
  78 {
  79         /* insn: must be on special executable page on x86_64. */
  80         up(&kprobe_mutex);
  81         p->ainsn.insn = get_insn_slot();
  82         down(&kprobe_mutex);
  83         if (!p->ainsn.insn) {
  84                 return -ENOMEM;
  85         }
  86         return 0;
  87 }
  88
  89 /*
  90  * Determine if the instruction uses the %rip-relative addressing mode.
  91  * If it does, return the address of the 32-bit displacement word.
  92  * If not, return null.
  93  */
  94 static inline s32 *is_riprel(u8 *insn)
  95 {
  96 #define W(row,b0,b1,b2,b3,b4,b5,b6,b7,b8,b9,ba,bb,bc,bd,be,bf)                \
  97         (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
  98           (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
  99           (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
 100           (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
 101          << (row % 64))
 102         static const u64 onebyte_has_modrm[256 / 64] = {
 103                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
 104                 /*      -------------------------------         */
 105                 W(0x00, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 00 */
 106                 W(0x10, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 10 */
 107                 W(0x20, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 20 */
 108                 W(0x30, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0), /* 30 */
 109                 W(0x40, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 40 */
 110                 W(0x50, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 50 */
 111                 W(0x60, 0,0,1,1,0,0,0,0,0,1,0,1,0,0,0,0)| /* 60 */
 112                 W(0x70, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 70 */
 113                 W(0x80, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 80 */
 114                 W(0x90, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 90 */
 115                 W(0xa0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* a0 */
 116                 W(0xb0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* b0 */
 117                 W(0xc0, 1,1,0,0,1,1,1,1,0,0,0,0,0,0,0,0)| /* c0 */
 118                 W(0xd0, 1,1,1,1,0,0,0,0,1,1,1,1,1,1,1,1)| /* d0 */
 119                 W(0xe0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* e0 */
 120                 W(0xf0, 0,0,0,0,0,0,1,1,0,0,0,0,0,0,1,1)  /* f0 */
 121                 /*      -------------------------------         */
 122                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
 123         };
 124         static const u64 twobyte_has_modrm[256 / 64] = {
 125                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
 126                 /*      -------------------------------         */
 127                 W(0x00, 1,1,1,1,0,0,0,0,0,0,0,0,0,1,0,1)| /* 0f */
 128                 W(0x10, 1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0)| /* 1f */
 129                 W(0x20, 1,1,1,1,1,0,1,0,1,1,1,1,1,1,1,1)| /* 2f */
 130                 W(0x30, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 3f */
 131                 W(0x40, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 4f */
 132                 W(0x50, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 5f */
 133                 W(0x60, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 6f */
 134                 W(0x70, 1,1,1,1,1,1,1,0,0,0,0,0,1,1,1,1), /* 7f */
 135                 W(0x80, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 8f */
 136                 W(0x90, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 9f */
 137                 W(0xa0, 0,0,0,1,1,1,1,1,0,0,0,1,1,1,1,1)| /* af */
 138                 W(0xb0, 1,1,1,1,1,1,1,1,0,0,1,1,1,1,1,1), /* bf */
 139                 W(0xc0, 1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0)| /* cf */
 140                 W(0xd0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* df */
 141                 W(0xe0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* ef */
 142                 W(0xf0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,0)  /* ff */
 143                 /*      -------------------------------         */
 144                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
 145         };
 146 #undef  W
 147         int need_modrm;
 148
 149         /* Skip legacy instruction prefixes.  */
 150         while (1) {
 151                 switch (*insn) {
 152                 case 0x66:
 153                 case 0x67:
 154                 case 0x2e:
 155                 case 0x3e:
 156                 case 0x26:
 157                 case 0x64:
 158                 case 0x65:
 159                 case 0x36:
 160                 case 0xf0:
 161                 case 0xf3:
 162                 case 0xf2:
 163                         ++insn;
 164                         continue;
 165                 }
 166                 break;
 167         }
 168
 169         /* Skip REX instruction prefix.  */
 170         if ((*insn & 0xf0) == 0x40)
 171                 ++insn;
 172
 173         if (*insn == 0x0f) {    /* Two-byte opcode.  */
 174                 ++insn;
 175                 need_modrm = test_bit(*insn, twobyte_has_modrm);
 176         } else {                /* One-byte opcode.  */
 177                 need_modrm = test_bit(*insn, onebyte_has_modrm);
 178         }
 179
 180         if (need_modrm) {
 181                 u8 modrm = *++insn;
 182                 if ((modrm & 0xc7) == 0x05) { /* %rip+disp32 addressing mode */
 183                         /* Displacement follows ModRM byte.  */
 184                         return (s32 *) ++insn;
 185                 }
 186         }
 187
 188         /* No %rip-relative addressing mode here.  */
 189         return NULL;
 190 }
 191
 192 void __kprobes arch_copy_kprobe(struct kprobe *p)
 193 {
 194         s32 *ripdisp;
 195         memcpy(p->ainsn.insn, p->addr, MAX_INSN_SIZE);
 196         ripdisp = is_riprel(p->ainsn.insn);
 197         if (ripdisp) {
 198                 /*
 199                  * The copied instruction uses the %rip-relative
 200                  * addressing mode.  Adjust the displacement for the
 201                  * difference between the original location of this
 202                  * instruction and the location of the copy that will
 203                  * actually be run.  The tricky bit here is making sure
 204                  * that the sign extension happens correctly in this
 205                  * calculation, since we need a signed 32-bit result to
 206                  * be sign-extended to 64 bits when it's added to the
 207                  * %rip value and yield the same 64-bit result that the
 208                  * sign-extension of the original signed 32-bit
 209                  * displacement would have given.
 210                  */
 211                 s64 disp = (u8 *) p->addr + *ripdisp - (u8 *) p->ainsn.insn;
 212                 BUG_ON((s64) (s32) disp != disp); /* Sanity check.  */
 213                 *ripdisp = disp;
 214         }
 215         p->opcode = *p->addr;
 216 }
 217
 218 void __kprobes arch_arm_kprobe(struct kprobe *p)
 219 {
 220         *p->addr = BREAKPOINT_INSTRUCTION;
 221         flush_icache_range((unsigned long) p->addr,
 222                            (unsigned long) p->addr + sizeof(kprobe_opcode_t));
 223 }
 224
 225 void __kprobes arch_disarm_kprobe(struct kprobe *p)
 226 {
 227         *p->addr = p->opcode;
 228         flush_icache_range((unsigned long) p->addr,
 229                            (unsigned long) p->addr + sizeof(kprobe_opcode_t));
 230 }
 231
 232 void __kprobes arch_remove_kprobe(struct kprobe *p)
 233 {
 234         up(&kprobe_mutex);
 235         free_insn_slot(p->ainsn.insn);
 236         down(&kprobe_mutex);
 237 }
 238
 239 static inline void save_previous_kprobe(void)
 240 {
 241         kprobe_prev = current_kprobe;
 242         kprobe_status_prev = kprobe_status;
 243         kprobe_old_rflags_prev = kprobe_old_rflags;
 244         kprobe_saved_rflags_prev = kprobe_saved_rflags;
 245 }
 246
 247 static inline void restore_previous_kprobe(void)
 248 {
 249         current_kprobe = kprobe_prev;
 250         kprobe_status = kprobe_status_prev;
 251         kprobe_old_rflags = kprobe_old_rflags_prev;
 252         kprobe_saved_rflags = kprobe_saved_rflags_prev;
 253 }
 254
 255 static inline void set_current_kprobe(struct kprobe *p, struct pt_regs *regs)
 256 {
 257         current_kprobe = p;
 258         kprobe_saved_rflags = kprobe_old_rflags
 259                 = (regs->eflags & (TF_MASK | IF_MASK));
 260         if (is_IF_modifier(p->ainsn.insn))
 261                 kprobe_saved_rflags &= ~IF_MASK;
 262 }
 263
 264 static void __kprobes prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
 265 {
 266         regs->eflags |= TF_MASK;
 267         regs->eflags &= ~IF_MASK;
 268         /*single step inline if the instruction is an int3*/
 269         if (p->opcode == BREAKPOINT_INSTRUCTION)
 270                 regs->rip = (unsigned long)p->addr;
 271         else
 272                 regs->rip = (unsigned long)p->ainsn.insn;
 273 }
 274
 275 void __kprobes arch_prepare_kretprobe(struct kretprobe *rp,
 276                                       struct pt_regs *regs)
 277 {
 278         unsigned long *sara = (unsigned long *)regs->rsp;
 279         struct kretprobe_instance *ri;
 280
 281         if ((ri = get_free_rp_inst(rp)) != NULL) {
 282                 ri->rp = rp;
 283                 ri->task = current;
 284                 ri->ret_addr = (kprobe_opcode_t *) *sara;
 285
 286                 /* Replace the return addr with trampoline addr */
 287                 *sara = (unsigned long) &kretprobe_trampoline;
 288
 289                 add_rp_inst(ri);
 290         } else {
 291                 rp->nmissed++;
 292         }
 293 }
 294
 295 /*
 296  * Interrupts are disabled on entry as trap3 is an interrupt gate and they
 297  * remain disabled thorough out this function.
 298  */
 299 int __kprobes kprobe_handler(struct pt_regs *regs)
 300 {
 301         struct kprobe *p;
 302         int ret = 0;
 303         kprobe_opcode_t *addr = (kprobe_opcode_t *)(regs->rip - sizeof(kprobe_opcode_t));
 304
 305         /* We're in an interrupt, but this is clear and BUG()-safe. */
 306         preempt_disable();
 307
 308         /* Check we're not actually recursing */
 309         if (kprobe_running()) {
 310                 /* We *are* holding lock here, so this is safe.
 311                    Disarm the probe we just hit, and ignore it. */
 312                 p = get_kprobe(addr);
 313                 if (p) {
 314                         if (kprobe_status == KPROBE_HIT_SS) {
 315                                 regs->eflags &= ~TF_MASK;
 316                                 regs->eflags |= kprobe_saved_rflags;
 317                                 unlock_kprobes();
 318                                 goto no_kprobe;
 319                         } else if (kprobe_status == KPROBE_HIT_SSDONE) {
 320                                 /* TODO: Provide re-entrancy from
 321                                  * post_kprobes_handler() and avoid exception
 322                                  * stack corruption while single-stepping on
 323                                  * the instruction of the new probe.
 324                                  */
 325                                 arch_disarm_kprobe(p);
 326                                 regs->rip = (unsigned long)p->addr;
 327                                 ret = 1;
 328                         } else {
 329                                 /* We have reentered the kprobe_handler(), since
 330                                  * another probe was hit while within the
 331                                  * handler. We here save the original kprobe
 332                                  * variables and just single step on instruction
 333                                  * of the new probe without calling any user
 334                                  * handlers.
 335                                  */
 336                                 save_previous_kprobe();
 337                                 set_current_kprobe(p, regs);
 338                                 p->nmissed++;
 339                                 prepare_singlestep(p, regs);
 340                                 kprobe_status = KPROBE_REENTER;
 341                                 return 1;
 342                         }
 343                 } else {
 344                         p = current_kprobe;
 345                         if (p->break_handler && p->break_handler(p, regs)) {
 346                                 goto ss_probe;
 347                         }
 348                 }
 349                 /* If it's not ours, can't be delete race, (we hold lock). */
 350                 goto no_kprobe;
 351         }
 352
 353         lock_kprobes();
 354         p = get_kprobe(addr);
 355         if (!p) {
 356                 unlock_kprobes();
 357                 if (*addr != BREAKPOINT_INSTRUCTION) {
 358                         /*
 359                          * The breakpoint instruction was removed right
 360                          * after we hit it.  Another cpu has removed
 361                          * either a probepoint or a debugger breakpoint
 362                          * at this address.  In either case, no further
 363                          * handling of this interrupt is appropriate.
 364                          */
 365                         ret = 1;
 366                 }
 367                 /* Not one of ours: let kernel handle it */
 368                 goto no_kprobe;
 369         }
 370
 371         kprobe_status = KPROBE_HIT_ACTIVE;
 372         set_current_kprobe(p, regs);
 373
 374         if (p->pre_handler && p->pre_handler(p, regs))
 375                 /* handler has already set things up, so skip ss setup */
 376                 return 1;
 377
 378 ss_probe:
 379         prepare_singlestep(p, regs);
 380         kprobe_status = KPROBE_HIT_SS;
 381         return 1;
 382
 383 no_kprobe:
 384         preempt_enable_no_resched();
 385         return ret;
 386 }
 387
 388 /*
 389  * For function-return probes, init_kprobes() establishes a probepoint
 390  * here. When a retprobed function returns, this probe is hit and
 391  * trampoline_probe_handler() runs, calling the kretprobe's handler.
 392  */
 393  void kretprobe_trampoline_holder(void)
 394  {
 395         asm volatile (  ".global kretprobe_trampoline\n"
 396                         "kretprobe_trampoline: \n"
 397                         "nop\n");
 398  }
 399
 400 /*
 401  * Called when we hit the probe point at kretprobe_trampoline
 402  */
 403 int __kprobes trampoline_probe_handler(struct kprobe *p, struct pt_regs *regs)
 404 {
 405         struct kretprobe_instance *ri = NULL;
 406         struct hlist_head *head;
 407         struct hlist_node *node, *tmp;
 408         unsigned long orig_ret_address = 0;
 409         unsigned long trampoline_address =(unsigned long)&kretprobe_trampoline;
 410
 411         head = kretprobe_inst_table_head(current);
 412
 413         /*
 414          * It is possible to have multiple instances associated with a given
 415          * task either because an multiple functions in the call path
 416          * have a return probe installed on them, and/or more then one return
 417          * return probe was registered for a target function.
 418          *
 419          * We can handle this because:
 420          *     - instances are always inserted at the head of the list
 421          *     - when multiple return probes are registered for the same
 422          *       function, the first instance's ret_addr will point to the
 423          *       real return address, and all the rest will point to
 424          *       kretprobe_trampoline
 425          */
 426         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
 427                 if (ri->task != current)
 428                         /* another task is sharing our hash bucket */
 429                         continue;
 430
 431                 if (ri->rp && ri->rp->handler)
 432                         ri->rp->handler(ri, regs);
 433
 434                 orig_ret_address = (unsigned long)ri->ret_addr;
 435                 recycle_rp_inst(ri);
 436
 437                 if (orig_ret_address != trampoline_address)
 438                         /*
 439                          * This is the real return address. Any other
 440                          * instances associated with this task are for
 441                          * other calls deeper on the call stack
 442                          */
 443                         break;
 444         }
 445
 446         BUG_ON(!orig_ret_address || (orig_ret_address == trampoline_address));
 447         regs->rip = orig_ret_address;
 448
 449         unlock_kprobes();
 450         preempt_enable_no_resched();
 451
 452         /*
 453          * By returning a non-zero value, we are telling
 454          * kprobe_handler() that we have handled unlocking
 455          * and re-enabling preemption.
 456          */
 457         return 1;
 458 }
 459
 460 /*
 461  * Called after single-stepping.  p->addr is the address of the
 462  * instruction whose first byte has been replaced by the "int 3"
 463  * instruction.  To avoid the SMP problems that can occur when we
 464  * temporarily put back the original opcode to single-step, we
 465  * single-stepped a copy of the instruction.  The address of this
 466  * copy is p->ainsn.insn.
 467  *
 468  * This function prepares to return from the post-single-step
 469  * interrupt.  We have to fix up the stack as follows:
 470  *
 471  * 0) Except in the case of absolute or indirect jump or call instructions,
 472  * the new rip is relative to the copied instruction.  We need to make
 473  * it relative to the original instruction.
 474  *
 475  * 1) If the single-stepped instruction was pushfl, then the TF and IF
 476  * flags are set in the just-pushed eflags, and may need to be cleared.
 477  *
 478  * 2) If the single-stepped instruction was a call, the return address
 479  * that is atop the stack is the address following the copied instruction.
 480  * We need to make it the address following the original instruction.
 481  */
 482 static void __kprobes resume_execution(struct kprobe *p, struct pt_regs *regs)
 483 {
 484         unsigned long *tos = (unsigned long *)regs->rsp;
 485         unsigned long next_rip = 0;
 486         unsigned long copy_rip = (unsigned long)p->ainsn.insn;
 487         unsigned long orig_rip = (unsigned long)p->addr;
 488         kprobe_opcode_t *insn = p->ainsn.insn;
 489
 490         /*skip the REX prefix*/
 491         if (*insn >= 0x40 && *insn <= 0x4f)
 492                 insn++;
 493
 494         switch (*insn) {
 495         case 0x9c:              /* pushfl */
 496                 *tos &= ~(TF_MASK | IF_MASK);
 497                 *tos |= kprobe_old_rflags;
 498                 break;
 499         case 0xc3:              /* ret/lret */
 500         case 0xcb:
 501         case 0xc2:
 502         case 0xca:
 503                 regs->eflags &= ~TF_MASK;
 504                 /* rip is already adjusted, no more changes required*/
 505                 return;
 506         case 0xe8:              /* call relative - Fix return addr */
 507                 *tos = orig_rip + (*tos - copy_rip);
 508                 break;
 509         case 0xff:
 510                 if ((*insn & 0x30) == 0x10) {
 511                         /* call absolute, indirect */
 512                         /* Fix return addr; rip is correct. */
 513                         next_rip = regs->rip;
 514                         *tos = orig_rip + (*tos - copy_rip);
 515                 } else if (((*insn & 0x31) == 0x20) ||  /* jmp near, absolute indirect */
 516                            ((*insn & 0x31) == 0x21)) {  /* jmp far, absolute indirect */
 517                         /* rip is correct. */
 518                         next_rip = regs->rip;
 519                 }
 520                 break;
 521         case 0xea:              /* jmp absolute -- rip is correct */
 522                 next_rip = regs->rip;
 523                 break;
 524         default:
 525                 break;
 526         }
 527
 528         regs->eflags &= ~TF_MASK;
 529         if (next_rip) {
 530                 regs->rip = next_rip;
 531         } else {
 532                 regs->rip = orig_rip + (regs->rip - copy_rip);
 533         }
 534 }
 535
 536 /*
 537  * Interrupts are disabled on entry as trap1 is an interrupt gate and they
 538  * remain disabled thoroughout this function.  And we hold kprobe lock.
 539  */
 540 int __kprobes post_kprobe_handler(struct pt_regs *regs)
 541 {
 542         if (!kprobe_running())
 543                 return 0;
 544
 545         if ((kprobe_status != KPROBE_REENTER) && current_kprobe->post_handler) {
 546                 kprobe_status = KPROBE_HIT_SSDONE;
 547                 current_kprobe->post_handler(current_kprobe, regs, 0);
 548         }
 549
 550         resume_execution(current_kprobe, regs);
 551         regs->eflags |= kprobe_saved_rflags;
 552
 553         /* Restore the original saved kprobes variables and continue. */
 554         if (kprobe_status == KPROBE_REENTER) {
 555                 restore_previous_kprobe();
 556                 goto out;
 557         } else {
 558                 unlock_kprobes();
 559         }
 560 out:
 561         preempt_enable_no_resched();
 562
 563         /*
 564          * if somebody else is singlestepping across a probe point, eflags
 565          * will have TF set, in which case, continue the remaining processing
 566          * of do_debug, as if this is not a probe hit.
 567          */
 568         if (regs->eflags & TF_MASK)
 569                 return 0;
 570
 571         return 1;
 572 }
 573
 574 /* Interrupts disabled, kprobe_lock held. */
 575 int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
 576 {
 577         if (current_kprobe->fault_handler
 578             && current_kprobe->fault_handler(current_kprobe, regs, trapnr))
 579                 return 1;
 580
 581         if (kprobe_status & KPROBE_HIT_SS) {
 582                 resume_execution(current_kprobe, regs);
 583                 regs->eflags |= kprobe_old_rflags;
 584
 585                 unlock_kprobes();
 586                 preempt_enable_no_resched();
 587         }
 588         return 0;
 589 }
 590
 591 /*
 592  * Wrapper routine for handling exceptions.
 593  */
 594 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
 595                                        unsigned long val, void *data)
 596 {
 597         struct die_args *args = (struct die_args *)data;
 598         switch (val) {
 599         case DIE_INT3:
 600                 if (kprobe_handler(args->regs))
 601                         return NOTIFY_STOP;
 602                 break;
 603         case DIE_DEBUG:
 604                 if (post_kprobe_handler(args->regs))
 605                         return NOTIFY_STOP;
 606                 break;
 607         case DIE_GPF:
 608                 if (kprobe_running() &&
 609                     kprobe_fault_handler(args->regs, args->trapnr))
 610                         return NOTIFY_STOP;
 611                 break;
 612         case DIE_PAGE_FAULT:
 613                 if (kprobe_running() &&
 614                     kprobe_fault_handler(args->regs, args->trapnr))
 615                         return NOTIFY_STOP;
 616                 break;
 617         default:
 618                 break;
 619         }
 620         return NOTIFY_DONE;
 621 }
 622
 623 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
 624 {
 625         struct jprobe *jp = container_of(p, struct jprobe, kp);
 626         unsigned long addr;
 627
 628         jprobe_saved_regs = *regs;
 629         jprobe_saved_rsp = (long *) regs->rsp;
 630         addr = (unsigned long)jprobe_saved_rsp;
 631         /*
 632          * As Linus pointed out, gcc assumes that the callee
 633          * owns the argument space and could overwrite it, e.g.
 634          * tailcall optimization. So, to be absolutely safe
 635          * we also save and restore enough stack bytes to cover
 636          * the argument area.
 637          */
 638         memcpy(jprobes_stack, (kprobe_opcode_t *) addr, MIN_STACK_SIZE(addr));
 639         regs->eflags &= ~IF_MASK;
 640         regs->rip = (unsigned long)(jp->entry);
 641         return 1;
 642 }
 643
 644 void __kprobes jprobe_return(void)
 645 {
 646         preempt_enable_no_resched();
 647         asm volatile ("       xchg   %%rbx,%%rsp     \n"
 648                       "       int3                      \n"
 649                       "       .globl jprobe_return_end  \n"
 650                       "       jprobe_return_end:        \n"
 651                       "       nop                       \n"::"b"
 652                       (jprobe_saved_rsp):"memory");
 653 }
 654
 655 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
 656 {
 657         u8 *addr = (u8 *) (regs->rip - 1);
 658         unsigned long stack_addr = (unsigned long)jprobe_saved_rsp;
 659         struct jprobe *jp = container_of(p, struct jprobe, kp);
 660
 661         if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
 662                 if ((long *)regs->rsp != jprobe_saved_rsp) {
 663                         struct pt_regs *saved_regs =
 664                             container_of(jprobe_saved_rsp, struct pt_regs, rsp);
 665                         printk("current rsp %p does not match saved rsp %p\n",
 666                                (long *)regs->rsp, jprobe_saved_rsp);
 667                         printk("Saved registers for jprobe %p\n", jp);
 668                         show_registers(saved_regs);
 669                         printk("Current registers\n");
 670                         show_registers(regs);
 671                         BUG();
 672                 }
 673                 *regs = jprobe_saved_regs;
 674                 memcpy((kprobe_opcode_t *) stack_addr, jprobes_stack,
 675                        MIN_STACK_SIZE(stack_addr));
 676                 return 1;
 677         }
 678         return 0;
 679 }
 680
 681 static struct kprobe trampoline_p = {
 682         .addr = (kprobe_opcode_t *) &kretprobe_trampoline,
 683         .pre_handler = trampoline_probe_handler
 684 };
 685
 686 int __init arch_init_kprobes(void)
 687 {
 688         return register_kprobe(&trampoline_p);
 689 }