git.oblomov.eu Git - linux-2.6/blob - arch/x86/kernel/kprobes_64.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  * 2004-Oct     Jim Keniston <kenistoj@us.ibm.com> and Prasanna S Panchamukhi
  26  *              <prasanna@in.ibm.com> adapted for x86_64
  27  * 2005-Mar     Roland McGrath <roland@redhat.com>
  28  *              Fixed to handle %rip-relative addressing mode correctly.
  29  * 2005-May     Rusty Lynch <rusty.lynch@intel.com>
  30  *              Added function return probes functionality
  31  */
  32
  33 #include <linux/kprobes.h>
  34 #include <linux/ptrace.h>
  35 #include <linux/string.h>
  36 #include <linux/slab.h>
  37 #include <linux/preempt.h>
  38 #include <linux/module.h>
  39 #include <linux/kdebug.h>
  40
  41 #include <asm/pgtable.h>
  42 #include <asm/uaccess.h>
  43 #include <asm/alternative.h>
  44
  45 void jprobe_return_end(void);
  46 static void __kprobes arch_copy_kprobe(struct kprobe *p);
  47
  48 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
  49 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
  50
  51 /*
  52  * returns non-zero if opcode modifies the interrupt flag.
  53  */
  54 static __always_inline int is_IF_modifier(kprobe_opcode_t *insn)
  55 {
  56         switch (*insn) {
  57         case 0xfa:              /* cli */
  58         case 0xfb:              /* sti */
  59         case 0xcf:              /* iret/iretd */
  60         case 0x9d:              /* popf/popfd */
  61                 return 1;
  62         }
  63
  64         if (*insn  >= 0x40 && *insn <= 0x4f && *++insn == 0xcf)
  65                 return 1;
  66         return 0;
  67 }
  68
  69 int __kprobes arch_prepare_kprobe(struct kprobe *p)
  70 {
  71         /* insn: must be on special executable page on x86_64. */
  72         p->ainsn.insn = get_insn_slot();
  73         if (!p->ainsn.insn) {
  74                 return -ENOMEM;
  75         }
  76         arch_copy_kprobe(p);
  77         return 0;
  78 }
  79
  80 /*
  81  * Determine if the instruction uses the %rip-relative addressing mode.
  82  * If it does, return the address of the 32-bit displacement word.
  83  * If not, return null.
  84  */
  85 static s32 __kprobes *is_riprel(u8 *insn)
  86 {
  87 #define W(row,b0,b1,b2,b3,b4,b5,b6,b7,b8,b9,ba,bb,bc,bd,be,bf)                \
  88         (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
  89           (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
  90           (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
  91           (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
  92          << (row % 64))
  93         static const u64 onebyte_has_modrm[256 / 64] = {
  94                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
  95                 /*      -------------------------------         */
  96                 W(0x00, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 00 */
  97                 W(0x10, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 10 */
  98                 W(0x20, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 20 */
  99                 W(0x30, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0), /* 30 */
 100                 W(0x40, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 40 */
 101                 W(0x50, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 50 */
 102                 W(0x60, 0,0,1,1,0,0,0,0,0,1,0,1,0,0,0,0)| /* 60 */
 103                 W(0x70, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 70 */
 104                 W(0x80, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 80 */
 105                 W(0x90, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 90 */
 106                 W(0xa0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* a0 */
 107                 W(0xb0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* b0 */
 108                 W(0xc0, 1,1,0,0,1,1,1,1,0,0,0,0,0,0,0,0)| /* c0 */
 109                 W(0xd0, 1,1,1,1,0,0,0,0,1,1,1,1,1,1,1,1)| /* d0 */
 110                 W(0xe0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* e0 */
 111                 W(0xf0, 0,0,0,0,0,0,1,1,0,0,0,0,0,0,1,1)  /* f0 */
 112                 /*      -------------------------------         */
 113                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
 114         };
 115         static const u64 twobyte_has_modrm[256 / 64] = {
 116                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
 117                 /*      -------------------------------         */
 118                 W(0x00, 1,1,1,1,0,0,0,0,0,0,0,0,0,1,0,1)| /* 0f */
 119                 W(0x10, 1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0)| /* 1f */
 120                 W(0x20, 1,1,1,1,1,0,1,0,1,1,1,1,1,1,1,1)| /* 2f */
 121                 W(0x30, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 3f */
 122                 W(0x40, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 4f */
 123                 W(0x50, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 5f */
 124                 W(0x60, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 6f */
 125                 W(0x70, 1,1,1,1,1,1,1,0,0,0,0,0,1,1,1,1), /* 7f */
 126                 W(0x80, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 8f */
 127                 W(0x90, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 9f */
 128                 W(0xa0, 0,0,0,1,1,1,1,1,0,0,0,1,1,1,1,1)| /* af */
 129                 W(0xb0, 1,1,1,1,1,1,1,1,0,0,1,1,1,1,1,1), /* bf */
 130                 W(0xc0, 1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0)| /* cf */
 131                 W(0xd0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* df */
 132                 W(0xe0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* ef */
 133                 W(0xf0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,0)  /* ff */
 134                 /*      -------------------------------         */
 135                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
 136         };
 137 #undef  W
 138         int need_modrm;
 139
 140         /* Skip legacy instruction prefixes.  */
 141         while (1) {
 142                 switch (*insn) {
 143                 case 0x66:
 144                 case 0x67:
 145                 case 0x2e:
 146                 case 0x3e:
 147                 case 0x26:
 148                 case 0x64:
 149                 case 0x65:
 150                 case 0x36:
 151                 case 0xf0:
 152                 case 0xf3:
 153                 case 0xf2:
 154                         ++insn;
 155                         continue;
 156                 }
 157                 break;
 158         }
 159
 160         /* Skip REX instruction prefix.  */
 161         if ((*insn & 0xf0) == 0x40)
 162                 ++insn;
 163
 164         if (*insn == 0x0f) {    /* Two-byte opcode.  */
 165                 ++insn;
 166                 need_modrm = test_bit(*insn, twobyte_has_modrm);
 167         } else {                /* One-byte opcode.  */
 168                 need_modrm = test_bit(*insn, onebyte_has_modrm);
 169         }
 170
 171         if (need_modrm) {
 172                 u8 modrm = *++insn;
 173                 if ((modrm & 0xc7) == 0x05) { /* %rip+disp32 addressing mode */
 174                         /* Displacement follows ModRM byte.  */
 175                         return (s32 *) ++insn;
 176                 }
 177         }
 178
 179         /* No %rip-relative addressing mode here.  */
 180         return NULL;
 181 }
 182
 183 static void __kprobes arch_copy_kprobe(struct kprobe *p)
 184 {
 185         s32 *ripdisp;
 186         memcpy(p->ainsn.insn, p->addr, MAX_INSN_SIZE);
 187         ripdisp = is_riprel(p->ainsn.insn);
 188         if (ripdisp) {
 189                 /*
 190                  * The copied instruction uses the %rip-relative
 191                  * addressing mode.  Adjust the displacement for the
 192                  * difference between the original location of this
 193                  * instruction and the location of the copy that will
 194                  * actually be run.  The tricky bit here is making sure
 195                  * that the sign extension happens correctly in this
 196                  * calculation, since we need a signed 32-bit result to
 197                  * be sign-extended to 64 bits when it's added to the
 198                  * %rip value and yield the same 64-bit result that the
 199                  * sign-extension of the original signed 32-bit
 200                  * displacement would have given.
 201                  */
 202                 s64 disp = (u8 *) p->addr + *ripdisp - (u8 *) p->ainsn.insn;
 203                 BUG_ON((s64) (s32) disp != disp); /* Sanity check.  */
 204                 *ripdisp = disp;
 205         }
 206         p->opcode = *p->addr;
 207 }
 208
 209 void __kprobes arch_arm_kprobe(struct kprobe *p)
 210 {
 211         text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
 212 }
 213
 214 void __kprobes arch_disarm_kprobe(struct kprobe *p)
 215 {
 216         text_poke(p->addr, &p->opcode, 1);
 217 }
 218
 219 void __kprobes arch_remove_kprobe(struct kprobe *p)
 220 {
 221         mutex_lock(&kprobe_mutex);
 222         free_insn_slot(p->ainsn.insn, 0);
 223         mutex_unlock(&kprobe_mutex);
 224 }
 225
 226 static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
 227 {
 228         kcb->prev_kprobe.kp = kprobe_running();
 229         kcb->prev_kprobe.status = kcb->kprobe_status;
 230         kcb->prev_kprobe.old_rflags = kcb->kprobe_old_rflags;
 231         kcb->prev_kprobe.saved_rflags = kcb->kprobe_saved_rflags;
 232 }
 233
 234 static void __kprobes restore_previous_kprobe(struct kprobe_ctlblk *kcb)
 235 {
 236         __get_cpu_var(current_kprobe) = kcb->prev_kprobe.kp;
 237         kcb->kprobe_status = kcb->prev_kprobe.status;
 238         kcb->kprobe_old_rflags = kcb->prev_kprobe.old_rflags;
 239         kcb->kprobe_saved_rflags = kcb->prev_kprobe.saved_rflags;
 240 }
 241
 242 static void __kprobes set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
 243                                 struct kprobe_ctlblk *kcb)
 244 {
 245         __get_cpu_var(current_kprobe) = p;
 246         kcb->kprobe_saved_rflags = kcb->kprobe_old_rflags
 247                 = (regs->eflags & (TF_MASK | IF_MASK));
 248         if (is_IF_modifier(p->ainsn.insn))
 249                 kcb->kprobe_saved_rflags &= ~IF_MASK;
 250 }
 251
 252 static void __kprobes prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
 253 {
 254         regs->eflags |= TF_MASK;
 255         regs->eflags &= ~IF_MASK;
 256         /*single step inline if the instruction is an int3*/
 257         if (p->opcode == BREAKPOINT_INSTRUCTION)
 258                 regs->rip = (unsigned long)p->addr;
 259         else
 260                 regs->rip = (unsigned long)p->ainsn.insn;
 261 }
 262
 263 /* Called with kretprobe_lock held */
 264 void __kprobes arch_prepare_kretprobe(struct kretprobe_instance *ri,
 265                                       struct pt_regs *regs)
 266 {
 267         unsigned long *sara = (unsigned long *)regs->rsp;
 268
 269         ri->ret_addr = (kprobe_opcode_t *) *sara;
 270         /* Replace the return addr with trampoline addr */
 271         *sara = (unsigned long) &kretprobe_trampoline;
 272 }
 273
 274 int __kprobes kprobe_handler(struct pt_regs *regs)
 275 {
 276         struct kprobe *p;
 277         int ret = 0;
 278         kprobe_opcode_t *addr = (kprobe_opcode_t *)(regs->rip - sizeof(kprobe_opcode_t));
 279         struct kprobe_ctlblk *kcb;
 280
 281         /*
 282          * We don't want to be preempted for the entire
 283          * duration of kprobe processing
 284          */
 285         preempt_disable();
 286         kcb = get_kprobe_ctlblk();
 287
 288         /* Check we're not actually recursing */
 289         if (kprobe_running()) {
 290                 p = get_kprobe(addr);
 291                 if (p) {
 292                         if (kcb->kprobe_status == KPROBE_HIT_SS &&
 293                                 *p->ainsn.insn == BREAKPOINT_INSTRUCTION) {
 294                                 regs->eflags &= ~TF_MASK;
 295                                 regs->eflags |= kcb->kprobe_saved_rflags;
 296                                 goto no_kprobe;
 297                         } else if (kcb->kprobe_status == KPROBE_HIT_SSDONE) {
 298                                 /* TODO: Provide re-entrancy from
 299                                  * post_kprobes_handler() and avoid exception
 300                                  * stack corruption while single-stepping on
 301                                  * the instruction of the new probe.
 302                                  */
 303                                 arch_disarm_kprobe(p);
 304                                 regs->rip = (unsigned long)p->addr;
 305                                 reset_current_kprobe();
 306                                 ret = 1;
 307                         } else {
 308                                 /* We have reentered the kprobe_handler(), since
 309                                  * another probe was hit while within the
 310                                  * handler. We here save the original kprobe
 311                                  * variables and just single step on instruction
 312                                  * of the new probe without calling any user
 313                                  * handlers.
 314                                  */
 315                                 save_previous_kprobe(kcb);
 316                                 set_current_kprobe(p, regs, kcb);
 317                                 kprobes_inc_nmissed_count(p);
 318                                 prepare_singlestep(p, regs);
 319                                 kcb->kprobe_status = KPROBE_REENTER;
 320                                 return 1;
 321                         }
 322                 } else {
 323                         if (*addr != BREAKPOINT_INSTRUCTION) {
 324                         /* The breakpoint instruction was removed by
 325                          * another cpu right after we hit, no further
 326                          * handling of this interrupt is appropriate
 327                          */
 328                                 regs->rip = (unsigned long)addr;
 329                                 ret = 1;
 330                                 goto no_kprobe;
 331                         }
 332                         p = __get_cpu_var(current_kprobe);
 333                         if (p->break_handler && p->break_handler(p, regs)) {
 334                                 goto ss_probe;
 335                         }
 336                 }
 337                 goto no_kprobe;
 338         }
 339
 340         p = get_kprobe(addr);
 341         if (!p) {
 342                 if (*addr != BREAKPOINT_INSTRUCTION) {
 343                         /*
 344                          * The breakpoint instruction was removed right
 345                          * after we hit it.  Another cpu has removed
 346                          * either a probepoint or a debugger breakpoint
 347                          * at this address.  In either case, no further
 348                          * handling of this interrupt is appropriate.
 349                          * Back up over the (now missing) int3 and run
 350                          * the original instruction.
 351                          */
 352                         regs->rip = (unsigned long)addr;
 353                         ret = 1;
 354                 }
 355                 /* Not one of ours: let kernel handle it */
 356                 goto no_kprobe;
 357         }
 358
 359         set_current_kprobe(p, regs, kcb);
 360         kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 361
 362         if (p->pre_handler && p->pre_handler(p, regs))
 363                 /* handler has already set things up, so skip ss setup */
 364                 return 1;
 365
 366 ss_probe:
 367         prepare_singlestep(p, regs);
 368         kcb->kprobe_status = KPROBE_HIT_SS;
 369         return 1;
 370
 371 no_kprobe:
 372         preempt_enable_no_resched();
 373         return ret;
 374 }
 375
 376 /*
 377  * For function-return probes, init_kprobes() establishes a probepoint
 378  * here. When a retprobed function returns, this probe is hit and
 379  * trampoline_probe_handler() runs, calling the kretprobe's handler.
 380  */
 381  void kretprobe_trampoline_holder(void)
 382  {
 383         asm volatile (  ".global kretprobe_trampoline\n"
 384                         "kretprobe_trampoline: \n"
 385                         "nop\n");
 386  }
 387
 388 /*
 389  * Called when we hit the probe point at kretprobe_trampoline
 390  */
 391 int __kprobes trampoline_probe_handler(struct kprobe *p, struct pt_regs *regs)
 392 {
 393         struct kretprobe_instance *ri = NULL;
 394         struct hlist_head *head, empty_rp;
 395         struct hlist_node *node, *tmp;
 396         unsigned long flags, orig_ret_address = 0;
 397         unsigned long trampoline_address =(unsigned long)&kretprobe_trampoline;
 398
 399         INIT_HLIST_HEAD(&empty_rp);
 400         spin_lock_irqsave(&kretprobe_lock, flags);
 401         head = kretprobe_inst_table_head(current);
 402
 403         /*
 404          * It is possible to have multiple instances associated with a given
 405          * task either because an multiple functions in the call path
 406          * have a return probe installed on them, and/or more then one return
 407          * return probe was registered for a target function.
 408          *
 409          * We can handle this because:
 410          *     - instances are always inserted at the head of the list
 411          *     - when multiple return probes are registered for the same
 412          *       function, the first instance's ret_addr will point to the
 413          *       real return address, and all the rest will point to
 414          *       kretprobe_trampoline
 415          */
 416         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
 417                 if (ri->task != current)
 418                         /* another task is sharing our hash bucket */
 419                         continue;
 420
 421                 if (ri->rp && ri->rp->handler)
 422                         ri->rp->handler(ri, regs);
 423
 424                 orig_ret_address = (unsigned long)ri->ret_addr;
 425                 recycle_rp_inst(ri, &empty_rp);
 426
 427                 if (orig_ret_address != trampoline_address)
 428                         /*
 429                          * This is the real return address. Any other
 430                          * instances associated with this task are for
 431                          * other calls deeper on the call stack
 432                          */
 433                         break;
 434         }
 435
 436         kretprobe_assert(ri, orig_ret_address, trampoline_address);
 437         regs->rip = orig_ret_address;
 438
 439         reset_current_kprobe();
 440         spin_unlock_irqrestore(&kretprobe_lock, flags);
 441         preempt_enable_no_resched();
 442
 443         hlist_for_each_entry_safe(ri, node, tmp, &empty_rp, hlist) {
 444                 hlist_del(&ri->hlist);
 445                 kfree(ri);
 446         }
 447         /*
 448          * By returning a non-zero value, we are telling
 449          * kprobe_handler() that we don't want the post_handler
 450          * to run (and have re-enabled preemption)
 451          */
 452         return 1;
 453 }
 454
 455 /*
 456  * Called after single-stepping.  p->addr is the address of the
 457  * instruction whose first byte has been replaced by the "int 3"
 458  * instruction.  To avoid the SMP problems that can occur when we
 459  * temporarily put back the original opcode to single-step, we
 460  * single-stepped a copy of the instruction.  The address of this
 461  * copy is p->ainsn.insn.
 462  *
 463  * This function prepares to return from the post-single-step
 464  * interrupt.  We have to fix up the stack as follows:
 465  *
 466  * 0) Except in the case of absolute or indirect jump or call instructions,
 467  * the new rip is relative to the copied instruction.  We need to make
 468  * it relative to the original instruction.
 469  *
 470  * 1) If the single-stepped instruction was pushfl, then the TF and IF
 471  * flags are set in the just-pushed eflags, and may need to be cleared.
 472  *
 473  * 2) If the single-stepped instruction was a call, the return address
 474  * that is atop the stack is the address following the copied instruction.
 475  * We need to make it the address following the original instruction.
 476  */
 477 static void __kprobes resume_execution(struct kprobe *p,
 478                 struct pt_regs *regs, struct kprobe_ctlblk *kcb)
 479 {
 480         unsigned long *tos = (unsigned long *)regs->rsp;
 481         unsigned long next_rip = 0;
 482         unsigned long copy_rip = (unsigned long)p->ainsn.insn;
 483         unsigned long orig_rip = (unsigned long)p->addr;
 484         kprobe_opcode_t *insn = p->ainsn.insn;
 485
 486         /*skip the REX prefix*/
 487         if (*insn >= 0x40 && *insn <= 0x4f)
 488                 insn++;
 489
 490         switch (*insn) {
 491         case 0x9c:              /* pushfl */
 492                 *tos &= ~(TF_MASK | IF_MASK);
 493                 *tos |= kcb->kprobe_old_rflags;
 494                 break;
 495         case 0xc3:              /* ret/lret */
 496         case 0xcb:
 497         case 0xc2:
 498         case 0xca:
 499                 regs->eflags &= ~TF_MASK;
 500                 /* rip is already adjusted, no more changes required*/
 501                 return;
 502         case 0xe8:              /* call relative - Fix return addr */
 503                 *tos = orig_rip + (*tos - copy_rip);
 504                 break;
 505         case 0xff:
 506                 if ((insn[1] & 0x30) == 0x10) {
 507                         /* call absolute, indirect */
 508                         /* Fix return addr; rip is correct. */
 509                         next_rip = regs->rip;
 510                         *tos = orig_rip + (*tos - copy_rip);
 511                 } else if (((insn[1] & 0x31) == 0x20) ||        /* jmp near, absolute indirect */
 512                            ((insn[1] & 0x31) == 0x21)) {        /* jmp far, absolute indirect */
 513                         /* rip is correct. */
 514                         next_rip = regs->rip;
 515                 }
 516                 break;
 517         case 0xea:              /* jmp absolute -- rip is correct */
 518                 next_rip = regs->rip;
 519                 break;
 520         default:
 521                 break;
 522         }
 523
 524         regs->eflags &= ~TF_MASK;
 525         if (next_rip) {
 526                 regs->rip = next_rip;
 527         } else {
 528                 regs->rip = orig_rip + (regs->rip - copy_rip);
 529         }
 530 }
 531
 532 int __kprobes post_kprobe_handler(struct pt_regs *regs)
 533 {
 534         struct kprobe *cur = kprobe_running();
 535         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 536
 537         if (!cur)
 538                 return 0;
 539
 540         if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
 541                 kcb->kprobe_status = KPROBE_HIT_SSDONE;
 542                 cur->post_handler(cur, regs, 0);
 543         }
 544
 545         resume_execution(cur, regs, kcb);
 546         regs->eflags |= kcb->kprobe_saved_rflags;
 547 #ifdef CONFIG_TRACE_IRQFLAGS_SUPPORT
 548         if (raw_irqs_disabled_flags(regs->eflags))
 549                 trace_hardirqs_off();
 550         else
 551                 trace_hardirqs_on();
 552 #endif
 553
 554         /* Restore the original saved kprobes variables and continue. */
 555         if (kcb->kprobe_status == KPROBE_REENTER) {
 556                 restore_previous_kprobe(kcb);
 557                 goto out;
 558         }
 559         reset_current_kprobe();
 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 int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
 575 {
 576         struct kprobe *cur = kprobe_running();
 577         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 578         const struct exception_table_entry *fixup;
 579
 580         switch(kcb->kprobe_status) {
 581         case KPROBE_HIT_SS:
 582         case KPROBE_REENTER:
 583                 /*
 584                  * We are here because the instruction being single
 585                  * stepped caused a page fault. We reset the current
 586                  * kprobe and the rip points back to the probe address
 587                  * and allow the page fault handler to continue as a
 588                  * normal page fault.
 589                  */
 590                 regs->rip = (unsigned long)cur->addr;
 591                 regs->eflags |= kcb->kprobe_old_rflags;
 592                 if (kcb->kprobe_status == KPROBE_REENTER)
 593                         restore_previous_kprobe(kcb);
 594                 else
 595                         reset_current_kprobe();
 596                 preempt_enable_no_resched();
 597                 break;
 598         case KPROBE_HIT_ACTIVE:
 599         case KPROBE_HIT_SSDONE:
 600                 /*
 601                  * We increment the nmissed count for accounting,
 602                  * we can also use npre/npostfault count for accouting
 603                  * these specific fault cases.
 604                  */
 605                 kprobes_inc_nmissed_count(cur);
 606
 607                 /*
 608                  * We come here because instructions in the pre/post
 609                  * handler caused the page_fault, this could happen
 610                  * if handler tries to access user space by
 611                  * copy_from_user(), get_user() etc. Let the
 612                  * user-specified handler try to fix it first.
 613                  */
 614                 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
 615                         return 1;
 616
 617                 /*
 618                  * In case the user-specified fault handler returned
 619                  * zero, try to fix up.
 620                  */
 621                 fixup = search_exception_tables(regs->rip);
 622                 if (fixup) {
 623                         regs->rip = fixup->fixup;
 624                         return 1;
 625                 }
 626
 627                 /*
 628                  * fixup() could not handle it,
 629                  * Let do_page_fault() fix it.
 630                  */
 631                 break;
 632         default:
 633                 break;
 634         }
 635         return 0;
 636 }
 637
 638 /*
 639  * Wrapper routine for handling exceptions.
 640  */
 641 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
 642                                        unsigned long val, void *data)
 643 {
 644         struct die_args *args = (struct die_args *)data;
 645         int ret = NOTIFY_DONE;
 646
 647         if (args->regs && user_mode(args->regs))
 648                 return ret;
 649
 650         switch (val) {
 651         case DIE_INT3:
 652                 if (kprobe_handler(args->regs))
 653                         ret = NOTIFY_STOP;
 654                 break;
 655         case DIE_DEBUG:
 656                 if (post_kprobe_handler(args->regs))
 657                         ret = NOTIFY_STOP;
 658                 break;
 659         case DIE_GPF:
 660         case DIE_PAGE_FAULT:
 661                 /* kprobe_running() needs smp_processor_id() */
 662                 preempt_disable();
 663                 if (kprobe_running() &&
 664                     kprobe_fault_handler(args->regs, args->trapnr))
 665                         ret = NOTIFY_STOP;
 666                 preempt_enable();
 667                 break;
 668         default:
 669                 break;
 670         }
 671         return ret;
 672 }
 673
 674 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
 675 {
 676         struct jprobe *jp = container_of(p, struct jprobe, kp);
 677         unsigned long addr;
 678         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 679
 680         kcb->jprobe_saved_regs = *regs;
 681         kcb->jprobe_saved_rsp = (long *) regs->rsp;
 682         addr = (unsigned long)(kcb->jprobe_saved_rsp);
 683         /*
 684          * As Linus pointed out, gcc assumes that the callee
 685          * owns the argument space and could overwrite it, e.g.
 686          * tailcall optimization. So, to be absolutely safe
 687          * we also save and restore enough stack bytes to cover
 688          * the argument area.
 689          */
 690         memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
 691                         MIN_STACK_SIZE(addr));
 692         regs->eflags &= ~IF_MASK;
 693         trace_hardirqs_off();
 694         regs->rip = (unsigned long)(jp->entry);
 695         return 1;
 696 }
 697
 698 void __kprobes jprobe_return(void)
 699 {
 700         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 701
 702         asm volatile ("       xchg   %%rbx,%%rsp     \n"
 703                       "       int3                      \n"
 704                       "       .globl jprobe_return_end  \n"
 705                       "       jprobe_return_end:        \n"
 706                       "       nop                       \n"::"b"
 707                       (kcb->jprobe_saved_rsp):"memory");
 708 }
 709
 710 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
 711 {
 712         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 713         u8 *addr = (u8 *) (regs->rip - 1);
 714         unsigned long stack_addr = (unsigned long)(kcb->jprobe_saved_rsp);
 715         struct jprobe *jp = container_of(p, struct jprobe, kp);
 716
 717         if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
 718                 if ((long *)regs->rsp != kcb->jprobe_saved_rsp) {
 719                         struct pt_regs *saved_regs =
 720                             container_of(kcb->jprobe_saved_rsp,
 721                                             struct pt_regs, rsp);
 722                         printk("current rsp %p does not match saved rsp %p\n",
 723                                (long *)regs->rsp, kcb->jprobe_saved_rsp);
 724                         printk("Saved registers for jprobe %p\n", jp);
 725                         show_registers(saved_regs);
 726                         printk("Current registers\n");
 727                         show_registers(regs);
 728                         BUG();
 729                 }
 730                 *regs = kcb->jprobe_saved_regs;
 731                 memcpy((kprobe_opcode_t *) stack_addr, kcb->jprobes_stack,
 732                        MIN_STACK_SIZE(stack_addr));
 733                 preempt_enable_no_resched();
 734                 return 1;
 735         }
 736         return 0;
 737 }
 738
 739 static struct kprobe trampoline_p = {
 740         .addr = (kprobe_opcode_t *) &kretprobe_trampoline,
 741         .pre_handler = trampoline_probe_handler
 742 };
 743
 744 int __init arch_init_kprobes(void)
 745 {
 746         return register_kprobe(&trampoline_p);
 747 }
 748
 749 int __kprobes arch_trampoline_kprobe(struct kprobe *p)
 750 {
 751         if (p->addr == (kprobe_opcode_t *)&kretprobe_trampoline)
 752                 return 1;
 753
 754         return 0;
 755 }