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