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 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 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 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 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 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 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 arch_prepare_kretprobe(struct kretprobe *rp, struct pt_regs *regs)
 276 {
 277         unsigned long *sara = (unsigned long *)regs->rsp;
 278         struct kretprobe_instance *ri;
 279
 280         if ((ri = get_free_rp_inst(rp)) != NULL) {
 281                 ri->rp = rp;
 282                 ri->task = current;
 283                 ri->ret_addr = (kprobe_opcode_t *) *sara;
 284
 285                 /* Replace the return addr with trampoline addr */
 286                 *sara = (unsigned long) &kretprobe_trampoline;
 287
 288                 add_rp_inst(ri);
 289         } else {
 290                 rp->nmissed++;
 291         }
 292 }
 293
 294 /*
 295  * Interrupts are disabled on entry as trap3 is an interrupt gate and they
 296  * remain disabled thorough out this function.
 297  */
 298 int kprobe_handler(struct pt_regs *regs)
 299 {
 300         struct kprobe *p;
 301         int ret = 0;
 302         kprobe_opcode_t *addr = (kprobe_opcode_t *)(regs->rip - sizeof(kprobe_opcode_t));
 303
 304         /* We're in an interrupt, but this is clear and BUG()-safe. */
 305         preempt_disable();
 306
 307         /* Check we're not actually recursing */
 308         if (kprobe_running()) {
 309                 /* We *are* holding lock here, so this is safe.
 310                    Disarm the probe we just hit, and ignore it. */
 311                 p = get_kprobe(addr);
 312                 if (p) {
 313                         if (kprobe_status == KPROBE_HIT_SS) {
 314                                 regs->eflags &= ~TF_MASK;
 315                                 regs->eflags |= kprobe_saved_rflags;
 316                                 unlock_kprobes();
 317                                 goto no_kprobe;
 318                         } else if (kprobe_status == KPROBE_HIT_SSDONE) {
 319                                 /* TODO: Provide re-entrancy from
 320                                  * post_kprobes_handler() and avoid exception
 321                                  * stack corruption while single-stepping on
 322                                  * the instruction of the new probe.
 323                                  */
 324                                 arch_disarm_kprobe(p);
 325                                 regs->rip = (unsigned long)p->addr;
 326                                 ret = 1;
 327                         } else {
 328                                 /* We have reentered the kprobe_handler(), since
 329                                  * another probe was hit while within the
 330                                  * handler. We here save the original kprobe
 331                                  * variables and just single step on instruction
 332                                  * of the new probe without calling any user
 333                                  * handlers.
 334                                  */
 335                                 save_previous_kprobe();
 336                                 set_current_kprobe(p, regs);
 337                                 p->nmissed++;
 338                                 prepare_singlestep(p, regs);
 339                                 kprobe_status = KPROBE_REENTER;
 340                                 return 1;
 341                         }
 342                 } else {
 343                         p = current_kprobe;
 344                         if (p->break_handler && p->break_handler(p, regs)) {
 345                                 goto ss_probe;
 346                         }
 347                 }
 348                 /* If it's not ours, can't be delete race, (we hold lock). */
 349                 goto no_kprobe;
 350         }
 351
 352         lock_kprobes();
 353         p = get_kprobe(addr);
 354         if (!p) {
 355                 unlock_kprobes();
 356                 if (*addr != BREAKPOINT_INSTRUCTION) {
 357                         /*
 358                          * The breakpoint instruction was removed right
 359                          * after we hit it.  Another cpu has removed
 360                          * either a probepoint or a debugger breakpoint
 361                          * at this address.  In either case, no further
 362                          * handling of this interrupt is appropriate.
 363                          */
 364                         ret = 1;
 365                 }
 366                 /* Not one of ours: let kernel handle it */
 367                 goto no_kprobe;
 368         }
 369
 370         kprobe_status = KPROBE_HIT_ACTIVE;
 371         set_current_kprobe(p, regs);
 372
 373         if (p->pre_handler && p->pre_handler(p, regs))
 374                 /* handler has already set things up, so skip ss setup */
 375                 return 1;
 376
 377 ss_probe:
 378         prepare_singlestep(p, regs);
 379         kprobe_status = KPROBE_HIT_SS;
 380         return 1;
 381
 382 no_kprobe:
 383         preempt_enable_no_resched();
 384         return ret;
 385 }
 386
 387 /*
 388  * For function-return probes, init_kprobes() establishes a probepoint
 389  * here. When a retprobed function returns, this probe is hit and
 390  * trampoline_probe_handler() runs, calling the kretprobe's handler.
 391  */
 392  void kretprobe_trampoline_holder(void)
 393  {
 394         asm volatile (  ".global kretprobe_trampoline\n"
 395                         "kretprobe_trampoline: \n"
 396                         "nop\n");
 397  }
 398
 399 /*
 400  * Called when we hit the probe point at kretprobe_trampoline
 401  */
 402 int trampoline_probe_handler(struct kprobe *p, struct pt_regs *regs)
 403 {
 404         struct kretprobe_instance *ri = NULL;
 405         struct hlist_head *head;
 406         struct hlist_node *node, *tmp;
 407         unsigned long orig_ret_address = 0;
 408         unsigned long trampoline_address =(unsigned long)&kretprobe_trampoline;
 409
 410         head = kretprobe_inst_table_head(current);
 411
 412         /*
 413          * It is possible to have multiple instances associated with a given
 414          * task either because an multiple functions in the call path
 415          * have a return probe installed on them, and/or more then one return
 416          * return probe was registered for a target function.
 417          *
 418          * We can handle this because:
 419          *     - instances are always inserted at the head of the list
 420          *     - when multiple return probes are registered for the same
 421          *       function, the first instance's ret_addr will point to the
 422          *       real return address, and all the rest will point to
 423          *       kretprobe_trampoline
 424          */
 425         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
 426                 if (ri->task != current)
 427                         /* another task is sharing our hash bucket */
 428                         continue;
 429
 430                 if (ri->rp && ri->rp->handler)
 431                         ri->rp->handler(ri, regs);
 432
 433                 orig_ret_address = (unsigned long)ri->ret_addr;
 434                 recycle_rp_inst(ri);
 435
 436                 if (orig_ret_address != trampoline_address)
 437                         /*
 438                          * This is the real return address. Any other
 439                          * instances associated with this task are for
 440                          * other calls deeper on the call stack
 441                          */
 442                         break;
 443         }
 444
 445         BUG_ON(!orig_ret_address || (orig_ret_address == trampoline_address));
 446         regs->rip = orig_ret_address;
 447
 448         unlock_kprobes();
 449         preempt_enable_no_resched();
 450
 451         /*
 452          * By returning a non-zero value, we are telling
 453          * kprobe_handler() that we have handled unlocking
 454          * and re-enabling preemption.
 455          */
 456         return 1;
 457 }
 458
 459 /*
 460  * Called after single-stepping.  p->addr is the address of the
 461  * instruction whose first byte has been replaced by the "int 3"
 462  * instruction.  To avoid the SMP problems that can occur when we
 463  * temporarily put back the original opcode to single-step, we
 464  * single-stepped a copy of the instruction.  The address of this
 465  * copy is p->ainsn.insn.
 466  *
 467  * This function prepares to return from the post-single-step
 468  * interrupt.  We have to fix up the stack as follows:
 469  *
 470  * 0) Except in the case of absolute or indirect jump or call instructions,
 471  * the new rip is relative to the copied instruction.  We need to make
 472  * it relative to the original instruction.
 473  *
 474  * 1) If the single-stepped instruction was pushfl, then the TF and IF
 475  * flags are set in the just-pushed eflags, and may need to be cleared.
 476  *
 477  * 2) If the single-stepped instruction was a call, the return address
 478  * that is atop the stack is the address following the copied instruction.
 479  * We need to make it the address following the original instruction.
 480  */
 481 static void resume_execution(struct kprobe *p, struct pt_regs *regs)
 482 {
 483         unsigned long *tos = (unsigned long *)regs->rsp;
 484         unsigned long next_rip = 0;
 485         unsigned long copy_rip = (unsigned long)p->ainsn.insn;
 486         unsigned long orig_rip = (unsigned long)p->addr;
 487         kprobe_opcode_t *insn = p->ainsn.insn;
 488
 489         /*skip the REX prefix*/
 490         if (*insn >= 0x40 && *insn <= 0x4f)
 491                 insn++;
 492
 493         switch (*insn) {
 494         case 0x9c:              /* pushfl */
 495                 *tos &= ~(TF_MASK | IF_MASK);
 496                 *tos |= kprobe_old_rflags;
 497                 break;
 498         case 0xc3:              /* ret/lret */
 499         case 0xcb:
 500         case 0xc2:
 501         case 0xca:
 502                 regs->eflags &= ~TF_MASK;
 503                 /* rip is already adjusted, no more changes required*/
 504                 return;
 505         case 0xe8:              /* call relative - Fix return addr */
 506                 *tos = orig_rip + (*tos - copy_rip);
 507                 break;
 508         case 0xff:
 509                 if ((*insn & 0x30) == 0x10) {
 510                         /* call absolute, indirect */
 511                         /* Fix return addr; rip is correct. */
 512                         next_rip = regs->rip;
 513                         *tos = orig_rip + (*tos - copy_rip);
 514                 } else if (((*insn & 0x31) == 0x20) ||  /* jmp near, absolute indirect */
 515                            ((*insn & 0x31) == 0x21)) {  /* jmp far, absolute indirect */
 516                         /* rip is correct. */
 517                         next_rip = regs->rip;
 518                 }
 519                 break;
 520         case 0xea:              /* jmp absolute -- rip is correct */
 521                 next_rip = regs->rip;
 522                 break;
 523         default:
 524                 break;
 525         }
 526
 527         regs->eflags &= ~TF_MASK;
 528         if (next_rip) {
 529                 regs->rip = next_rip;
 530         } else {
 531                 regs->rip = orig_rip + (regs->rip - copy_rip);
 532         }
 533 }
 534
 535 /*
 536  * Interrupts are disabled on entry as trap1 is an interrupt gate and they
 537  * remain disabled thoroughout this function.  And we hold kprobe lock.
 538  */
 539 int post_kprobe_handler(struct pt_regs *regs)
 540 {
 541         if (!kprobe_running())
 542                 return 0;
 543
 544         if ((kprobe_status != KPROBE_REENTER) && current_kprobe->post_handler) {
 545                 kprobe_status = KPROBE_HIT_SSDONE;
 546                 current_kprobe->post_handler(current_kprobe, regs, 0);
 547         }
 548
 549         resume_execution(current_kprobe, regs);
 550         regs->eflags |= kprobe_saved_rflags;
 551
 552         /* Restore the original saved kprobes variables and continue. */
 553         if (kprobe_status == KPROBE_REENTER) {
 554                 restore_previous_kprobe();
 555                 goto out;
 556         } else {
 557                 unlock_kprobes();
 558         }
 559 out:
 560         preempt_enable_no_resched();
 561
 562         /*
 563          * if somebody else is singlestepping across a probe point, eflags
 564          * will have TF set, in which case, continue the remaining processing
 565          * of do_debug, as if this is not a probe hit.
 566          */
 567         if (regs->eflags & TF_MASK)
 568                 return 0;
 569
 570         return 1;
 571 }
 572
 573 /* Interrupts disabled, kprobe_lock held. */
 574 int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
 575 {
 576         if (current_kprobe->fault_handler
 577             && current_kprobe->fault_handler(current_kprobe, regs, trapnr))
 578                 return 1;
 579
 580         if (kprobe_status & KPROBE_HIT_SS) {
 581                 resume_execution(current_kprobe, regs);
 582                 regs->eflags |= kprobe_old_rflags;
 583
 584                 unlock_kprobes();
 585                 preempt_enable_no_resched();
 586         }
 587         return 0;
 588 }
 589
 590 /*
 591  * Wrapper routine for handling exceptions.
 592  */
 593 int kprobe_exceptions_notify(struct notifier_block *self, unsigned long val,
 594                              void *data)
 595 {
 596         struct die_args *args = (struct die_args *)data;
 597         switch (val) {
 598         case DIE_INT3:
 599                 if (kprobe_handler(args->regs))
 600                         return NOTIFY_STOP;
 601                 break;
 602         case DIE_DEBUG:
 603                 if (post_kprobe_handler(args->regs))
 604                         return NOTIFY_STOP;
 605                 break;
 606         case DIE_GPF:
 607                 if (kprobe_running() &&
 608                     kprobe_fault_handler(args->regs, args->trapnr))
 609                         return NOTIFY_STOP;
 610                 break;
 611         case DIE_PAGE_FAULT:
 612                 if (kprobe_running() &&
 613                     kprobe_fault_handler(args->regs, args->trapnr))
 614                         return NOTIFY_STOP;
 615                 break;
 616         default:
 617                 break;
 618         }
 619         return NOTIFY_DONE;
 620 }
 621
 622 int setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
 623 {
 624         struct jprobe *jp = container_of(p, struct jprobe, kp);
 625         unsigned long addr;
 626
 627         jprobe_saved_regs = *regs;
 628         jprobe_saved_rsp = (long *) regs->rsp;
 629         addr = (unsigned long)jprobe_saved_rsp;
 630         /*
 631          * As Linus pointed out, gcc assumes that the callee
 632          * owns the argument space and could overwrite it, e.g.
 633          * tailcall optimization. So, to be absolutely safe
 634          * we also save and restore enough stack bytes to cover
 635          * the argument area.
 636          */
 637         memcpy(jprobes_stack, (kprobe_opcode_t *) addr, MIN_STACK_SIZE(addr));
 638         regs->eflags &= ~IF_MASK;
 639         regs->rip = (unsigned long)(jp->entry);
 640         return 1;
 641 }
 642
 643 void jprobe_return(void)
 644 {
 645         preempt_enable_no_resched();
 646         asm volatile ("       xchg   %%rbx,%%rsp     \n"
 647                       "       int3                      \n"
 648                       "       .globl jprobe_return_end  \n"
 649                       "       jprobe_return_end:        \n"
 650                       "       nop                       \n"::"b"
 651                       (jprobe_saved_rsp):"memory");
 652 }
 653
 654 int longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
 655 {
 656         u8 *addr = (u8 *) (regs->rip - 1);
 657         unsigned long stack_addr = (unsigned long)jprobe_saved_rsp;
 658         struct jprobe *jp = container_of(p, struct jprobe, kp);
 659
 660         if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
 661                 if ((long *)regs->rsp != jprobe_saved_rsp) {
 662                         struct pt_regs *saved_regs =
 663                             container_of(jprobe_saved_rsp, struct pt_regs, rsp);
 664                         printk("current rsp %p does not match saved rsp %p\n",
 665                                (long *)regs->rsp, jprobe_saved_rsp);
 666                         printk("Saved registers for jprobe %p\n", jp);
 667                         show_registers(saved_regs);
 668                         printk("Current registers\n");
 669                         show_registers(regs);
 670                         BUG();
 671                 }
 672                 *regs = jprobe_saved_regs;
 673                 memcpy((kprobe_opcode_t *) stack_addr, jprobes_stack,
 674                        MIN_STACK_SIZE(stack_addr));
 675                 return 1;
 676         }
 677         return 0;
 678 }
 679
 680 static struct kprobe trampoline_p = {
 681         .addr = (kprobe_opcode_t *) &kretprobe_trampoline,
 682         .pre_handler = trampoline_probe_handler
 683 };
 684
 685 int __init arch_init_kprobes(void)
 686 {
 687         return register_kprobe(&trampoline_p);
 688 }