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 void jprobe_return_end(void);
  48
  49 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
  50 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
  51
  52 /*
  53  * returns non-zero if opcode modifies the interrupt flag.
  54  */
  55 static inline int is_IF_modifier(kprobe_opcode_t *insn)
  56 {
  57         switch (*insn) {
  58         case 0xfa:              /* cli */
  59         case 0xfb:              /* sti */
  60         case 0xcf:              /* iret/iretd */
  61         case 0x9d:              /* popf/popfd */
  62                 return 1;
  63         }
  64
  65         if (*insn  >= 0x40 && *insn <= 0x4f && *++insn == 0xcf)
  66                 return 1;
  67         return 0;
  68 }
  69
  70 int __kprobes arch_prepare_kprobe(struct kprobe *p)
  71 {
  72         /* insn: must be on special executable page on x86_64. */
  73         down(&kprobe_mutex);
  74         p->ainsn.insn = get_insn_slot();
  75         up(&kprobe_mutex);
  76         if (!p->ainsn.insn) {
  77                 return -ENOMEM;
  78         }
  79         return 0;
  80 }
  81
  82 /*
  83  * Determine if the instruction uses the %rip-relative addressing mode.
  84  * If it does, return the address of the 32-bit displacement word.
  85  * If not, return null.
  86  */
  87 static inline s32 *is_riprel(u8 *insn)
  88 {
  89 #define W(row,b0,b1,b2,b3,b4,b5,b6,b7,b8,b9,ba,bb,bc,bd,be,bf)                \
  90         (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
  91           (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
  92           (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
  93           (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
  94          << (row % 64))
  95         static const u64 onebyte_has_modrm[256 / 64] = {
  96                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
  97                 /*      -------------------------------         */
  98                 W(0x00, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 00 */
  99                 W(0x10, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 10 */
 100                 W(0x20, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 20 */
 101                 W(0x30, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0), /* 30 */
 102                 W(0x40, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 40 */
 103                 W(0x50, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 50 */
 104                 W(0x60, 0,0,1,1,0,0,0,0,0,1,0,1,0,0,0,0)| /* 60 */
 105                 W(0x70, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 70 */
 106                 W(0x80, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 80 */
 107                 W(0x90, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 90 */
 108                 W(0xa0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* a0 */
 109                 W(0xb0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* b0 */
 110                 W(0xc0, 1,1,0,0,1,1,1,1,0,0,0,0,0,0,0,0)| /* c0 */
 111                 W(0xd0, 1,1,1,1,0,0,0,0,1,1,1,1,1,1,1,1)| /* d0 */
 112                 W(0xe0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* e0 */
 113                 W(0xf0, 0,0,0,0,0,0,1,1,0,0,0,0,0,0,1,1)  /* f0 */
 114                 /*      -------------------------------         */
 115                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
 116         };
 117         static const u64 twobyte_has_modrm[256 / 64] = {
 118                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
 119                 /*      -------------------------------         */
 120                 W(0x00, 1,1,1,1,0,0,0,0,0,0,0,0,0,1,0,1)| /* 0f */
 121                 W(0x10, 1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0)| /* 1f */
 122                 W(0x20, 1,1,1,1,1,0,1,0,1,1,1,1,1,1,1,1)| /* 2f */
 123                 W(0x30, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 3f */
 124                 W(0x40, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 4f */
 125                 W(0x50, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 5f */
 126                 W(0x60, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 6f */
 127                 W(0x70, 1,1,1,1,1,1,1,0,0,0,0,0,1,1,1,1), /* 7f */
 128                 W(0x80, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 8f */
 129                 W(0x90, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 9f */
 130                 W(0xa0, 0,0,0,1,1,1,1,1,0,0,0,1,1,1,1,1)| /* af */
 131                 W(0xb0, 1,1,1,1,1,1,1,1,0,0,1,1,1,1,1,1), /* bf */
 132                 W(0xc0, 1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0)| /* cf */
 133                 W(0xd0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* df */
 134                 W(0xe0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* ef */
 135                 W(0xf0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,0)  /* ff */
 136                 /*      -------------------------------         */
 137                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
 138         };
 139 #undef  W
 140         int need_modrm;
 141
 142         /* Skip legacy instruction prefixes.  */
 143         while (1) {
 144                 switch (*insn) {
 145                 case 0x66:
 146                 case 0x67:
 147                 case 0x2e:
 148                 case 0x3e:
 149                 case 0x26:
 150                 case 0x64:
 151                 case 0x65:
 152                 case 0x36:
 153                 case 0xf0:
 154                 case 0xf3:
 155                 case 0xf2:
 156                         ++insn;
 157                         continue;
 158                 }
 159                 break;
 160         }
 161
 162         /* Skip REX instruction prefix.  */
 163         if ((*insn & 0xf0) == 0x40)
 164                 ++insn;
 165
 166         if (*insn == 0x0f) {    /* Two-byte opcode.  */
 167                 ++insn;
 168                 need_modrm = test_bit(*insn, twobyte_has_modrm);
 169         } else {                /* One-byte opcode.  */
 170                 need_modrm = test_bit(*insn, onebyte_has_modrm);
 171         }
 172
 173         if (need_modrm) {
 174                 u8 modrm = *++insn;
 175                 if ((modrm & 0xc7) == 0x05) { /* %rip+disp32 addressing mode */
 176                         /* Displacement follows ModRM byte.  */
 177                         return (s32 *) ++insn;
 178                 }
 179         }
 180
 181         /* No %rip-relative addressing mode here.  */
 182         return NULL;
 183 }
 184
 185 void __kprobes arch_copy_kprobe(struct kprobe *p)
 186 {
 187         s32 *ripdisp;
 188         memcpy(p->ainsn.insn, p->addr, MAX_INSN_SIZE);
 189         ripdisp = is_riprel(p->ainsn.insn);
 190         if (ripdisp) {
 191                 /*
 192                  * The copied instruction uses the %rip-relative
 193                  * addressing mode.  Adjust the displacement for the
 194                  * difference between the original location of this
 195                  * instruction and the location of the copy that will
 196                  * actually be run.  The tricky bit here is making sure
 197                  * that the sign extension happens correctly in this
 198                  * calculation, since we need a signed 32-bit result to
 199                  * be sign-extended to 64 bits when it's added to the
 200                  * %rip value and yield the same 64-bit result that the
 201                  * sign-extension of the original signed 32-bit
 202                  * displacement would have given.
 203                  */
 204                 s64 disp = (u8 *) p->addr + *ripdisp - (u8 *) p->ainsn.insn;
 205                 BUG_ON((s64) (s32) disp != disp); /* Sanity check.  */
 206                 *ripdisp = disp;
 207         }
 208         p->opcode = *p->addr;
 209 }
 210
 211 void __kprobes arch_arm_kprobe(struct kprobe *p)
 212 {
 213         *p->addr = BREAKPOINT_INSTRUCTION;
 214         flush_icache_range((unsigned long) p->addr,
 215                            (unsigned long) p->addr + sizeof(kprobe_opcode_t));
 216 }
 217
 218 void __kprobes arch_disarm_kprobe(struct kprobe *p)
 219 {
 220         *p->addr = p->opcode;
 221         flush_icache_range((unsigned long) p->addr,
 222                            (unsigned long) p->addr + sizeof(kprobe_opcode_t));
 223 }
 224
 225 void __kprobes arch_remove_kprobe(struct kprobe *p)
 226 {
 227         down(&kprobe_mutex);
 228         free_insn_slot(p->ainsn.insn);
 229         up(&kprobe_mutex);
 230 }
 231
 232 static inline void save_previous_kprobe(struct kprobe_ctlblk *kcb)
 233 {
 234         kcb->prev_kprobe.kp = kprobe_running();
 235         kcb->prev_kprobe.status = kcb->kprobe_status;
 236         kcb->prev_kprobe.old_rflags = kcb->kprobe_old_rflags;
 237         kcb->prev_kprobe.saved_rflags = kcb->kprobe_saved_rflags;
 238 }
 239
 240 static inline void restore_previous_kprobe(struct kprobe_ctlblk *kcb)
 241 {
 242         __get_cpu_var(current_kprobe) = kcb->prev_kprobe.kp;
 243         kcb->kprobe_status = kcb->prev_kprobe.status;
 244         kcb->kprobe_old_rflags = kcb->prev_kprobe.old_rflags;
 245         kcb->kprobe_saved_rflags = kcb->prev_kprobe.saved_rflags;
 246 }
 247
 248 static inline void set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
 249                                 struct kprobe_ctlblk *kcb)
 250 {
 251         __get_cpu_var(current_kprobe) = p;
 252         kcb->kprobe_saved_rflags = kcb->kprobe_old_rflags
 253                 = (regs->eflags & (TF_MASK | IF_MASK));
 254         if (is_IF_modifier(p->ainsn.insn))
 255                 kcb->kprobe_saved_rflags &= ~IF_MASK;
 256 }
 257
 258 static void __kprobes prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
 259 {
 260         regs->eflags |= TF_MASK;
 261         regs->eflags &= ~IF_MASK;
 262         /*single step inline if the instruction is an int3*/
 263         if (p->opcode == BREAKPOINT_INSTRUCTION)
 264                 regs->rip = (unsigned long)p->addr;
 265         else
 266                 regs->rip = (unsigned long)p->ainsn.insn;
 267 }
 268
 269 void __kprobes arch_prepare_kretprobe(struct kretprobe *rp,
 270                                       struct pt_regs *regs)
 271 {
 272         unsigned long *sara = (unsigned long *)regs->rsp;
 273         struct kretprobe_instance *ri;
 274
 275         if ((ri = get_free_rp_inst(rp)) != NULL) {
 276                 ri->rp = rp;
 277                 ri->task = current;
 278                 ri->ret_addr = (kprobe_opcode_t *) *sara;
 279
 280                 /* Replace the return addr with trampoline addr */
 281                 *sara = (unsigned long) &kretprobe_trampoline;
 282
 283                 add_rp_inst(ri);
 284         } else {
 285                 rp->nmissed++;
 286         }
 287 }
 288
 289 /*
 290  * Interrupts are disabled on entry as trap3 is an interrupt gate and they
 291  * remain disabled thorough out this function.
 292  */
 293 int __kprobes kprobe_handler(struct pt_regs *regs)
 294 {
 295         struct kprobe *p;
 296         int ret = 0;
 297         kprobe_opcode_t *addr = (kprobe_opcode_t *)(regs->rip - sizeof(kprobe_opcode_t));
 298         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 299
 300         /* Check we're not actually recursing */
 301         if (kprobe_running()) {
 302                 /* We *are* holding lock here, so this is safe.
 303                    Disarm the probe we just hit, and ignore it. */
 304                 p = get_kprobe(addr);
 305                 if (p) {
 306                         if (kcb->kprobe_status == KPROBE_HIT_SS &&
 307                                 *p->ainsn.insn == BREAKPOINT_INSTRUCTION) {
 308                                 regs->eflags &= ~TF_MASK;
 309                                 regs->eflags |= kcb->kprobe_saved_rflags;
 310                                 unlock_kprobes();
 311                                 goto no_kprobe;
 312                         } else if (kcb->kprobe_status == KPROBE_HIT_SSDONE) {
 313                                 /* TODO: Provide re-entrancy from
 314                                  * post_kprobes_handler() and avoid exception
 315                                  * stack corruption while single-stepping on
 316                                  * the instruction of the new probe.
 317                                  */
 318                                 arch_disarm_kprobe(p);
 319                                 regs->rip = (unsigned long)p->addr;
 320                                 reset_current_kprobe();
 321                                 ret = 1;
 322                         } else {
 323                                 /* We have reentered the kprobe_handler(), since
 324                                  * another probe was hit while within the
 325                                  * handler. We here save the original kprobe
 326                                  * variables and just single step on instruction
 327                                  * of the new probe without calling any user
 328                                  * handlers.
 329                                  */
 330                                 save_previous_kprobe(kcb);
 331                                 set_current_kprobe(p, regs, kcb);
 332                                 p->nmissed++;
 333                                 prepare_singlestep(p, regs);
 334                                 kcb->kprobe_status = KPROBE_REENTER;
 335                                 return 1;
 336                         }
 337                 } else {
 338                         p = __get_cpu_var(current_kprobe);
 339                         if (p->break_handler && p->break_handler(p, regs)) {
 340                                 goto ss_probe;
 341                         }
 342                 }
 343                 /* If it's not ours, can't be delete race, (we hold lock). */
 344                 goto no_kprobe;
 345         }
 346
 347         lock_kprobes();
 348         p = get_kprobe(addr);
 349         if (!p) {
 350                 unlock_kprobes();
 351                 if (*addr != BREAKPOINT_INSTRUCTION) {
 352                         /*
 353                          * The breakpoint instruction was removed right
 354                          * after we hit it.  Another cpu has removed
 355                          * either a probepoint or a debugger breakpoint
 356                          * at this address.  In either case, no further
 357                          * handling of this interrupt is appropriate.
 358                          * Back up over the (now missing) int3 and run
 359                          * the original instruction.
 360                          */
 361                         regs->rip = (unsigned long)addr;
 362                         ret = 1;
 363                 }
 364                 /* Not one of ours: let kernel handle it */
 365                 goto no_kprobe;
 366         }
 367
 368         /*
 369          * This preempt_disable() matches the preempt_enable_no_resched()
 370          * in post_kprobe_handler()
 371          */
 372         preempt_disable();
 373         set_current_kprobe(p, regs, kcb);
 374         kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 375
 376         if (p->pre_handler && p->pre_handler(p, regs))
 377                 /* handler has already set things up, so skip ss setup */
 378                 return 1;
 379
 380 ss_probe:
 381         prepare_singlestep(p, regs);
 382         kcb->kprobe_status = KPROBE_HIT_SS;
 383         return 1;
 384
 385 no_kprobe:
 386         return ret;
 387 }
 388
 389 /*
 390  * For function-return probes, init_kprobes() establishes a probepoint
 391  * here. When a retprobed function returns, this probe is hit and
 392  * trampoline_probe_handler() runs, calling the kretprobe's handler.
 393  */
 394  void kretprobe_trampoline_holder(void)
 395  {
 396         asm volatile (  ".global kretprobe_trampoline\n"
 397                         "kretprobe_trampoline: \n"
 398                         "nop\n");
 399  }
 400
 401 /*
 402  * Called when we hit the probe point at kretprobe_trampoline
 403  */
 404 int __kprobes trampoline_probe_handler(struct kprobe *p, struct pt_regs *regs)
 405 {
 406         struct kretprobe_instance *ri = NULL;
 407         struct hlist_head *head;
 408         struct hlist_node *node, *tmp;
 409         unsigned long orig_ret_address = 0;
 410         unsigned long trampoline_address =(unsigned long)&kretprobe_trampoline;
 411
 412         head = kretprobe_inst_table_head(current);
 413
 414         /*
 415          * It is possible to have multiple instances associated with a given
 416          * task either because an multiple functions in the call path
 417          * have a return probe installed on them, and/or more then one return
 418          * return probe was registered for a target function.
 419          *
 420          * We can handle this because:
 421          *     - instances are always inserted at the head of the list
 422          *     - when multiple return probes are registered for the same
 423          *       function, the first instance's ret_addr will point to the
 424          *       real return address, and all the rest will point to
 425          *       kretprobe_trampoline
 426          */
 427         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
 428                 if (ri->task != current)
 429                         /* another task is sharing our hash bucket */
 430                         continue;
 431
 432                 if (ri->rp && ri->rp->handler)
 433                         ri->rp->handler(ri, regs);
 434
 435                 orig_ret_address = (unsigned long)ri->ret_addr;
 436                 recycle_rp_inst(ri);
 437
 438                 if (orig_ret_address != trampoline_address)
 439                         /*
 440                          * This is the real return address. Any other
 441                          * instances associated with this task are for
 442                          * other calls deeper on the call stack
 443                          */
 444                         break;
 445         }
 446
 447         BUG_ON(!orig_ret_address || (orig_ret_address == trampoline_address));
 448         regs->rip = orig_ret_address;
 449
 450         reset_current_kprobe();
 451         unlock_kprobes();
 452         preempt_enable_no_resched();
 453
 454         /*
 455          * By returning a non-zero value, we are telling
 456          * kprobe_handler() that we have handled unlocking
 457          * and re-enabling preemption
 458          */
 459         return 1;
 460 }
 461
 462 /*
 463  * Called after single-stepping.  p->addr is the address of the
 464  * instruction whose first byte has been replaced by the "int 3"
 465  * instruction.  To avoid the SMP problems that can occur when we
 466  * temporarily put back the original opcode to single-step, we
 467  * single-stepped a copy of the instruction.  The address of this
 468  * copy is p->ainsn.insn.
 469  *
 470  * This function prepares to return from the post-single-step
 471  * interrupt.  We have to fix up the stack as follows:
 472  *
 473  * 0) Except in the case of absolute or indirect jump or call instructions,
 474  * the new rip is relative to the copied instruction.  We need to make
 475  * it relative to the original instruction.
 476  *
 477  * 1) If the single-stepped instruction was pushfl, then the TF and IF
 478  * flags are set in the just-pushed eflags, and may need to be cleared.
 479  *
 480  * 2) If the single-stepped instruction was a call, the return address
 481  * that is atop the stack is the address following the copied instruction.
 482  * We need to make it the address following the original instruction.
 483  */
 484 static void __kprobes resume_execution(struct kprobe *p,
 485                 struct pt_regs *regs, struct kprobe_ctlblk *kcb)
 486 {
 487         unsigned long *tos = (unsigned long *)regs->rsp;
 488         unsigned long next_rip = 0;
 489         unsigned long copy_rip = (unsigned long)p->ainsn.insn;
 490         unsigned long orig_rip = (unsigned long)p->addr;
 491         kprobe_opcode_t *insn = p->ainsn.insn;
 492
 493         /*skip the REX prefix*/
 494         if (*insn >= 0x40 && *insn <= 0x4f)
 495                 insn++;
 496
 497         switch (*insn) {
 498         case 0x9c:              /* pushfl */
 499                 *tos &= ~(TF_MASK | IF_MASK);
 500                 *tos |= kcb->kprobe_old_rflags;
 501                 break;
 502         case 0xc3:              /* ret/lret */
 503         case 0xcb:
 504         case 0xc2:
 505         case 0xca:
 506                 regs->eflags &= ~TF_MASK;
 507                 /* rip is already adjusted, no more changes required*/
 508                 return;
 509         case 0xe8:              /* call relative - Fix return addr */
 510                 *tos = orig_rip + (*tos - copy_rip);
 511                 break;
 512         case 0xff:
 513                 if ((*insn & 0x30) == 0x10) {
 514                         /* call absolute, indirect */
 515                         /* Fix return addr; rip is correct. */
 516                         next_rip = regs->rip;
 517                         *tos = orig_rip + (*tos - copy_rip);
 518                 } else if (((*insn & 0x31) == 0x20) ||  /* jmp near, absolute indirect */
 519                            ((*insn & 0x31) == 0x21)) {  /* jmp far, absolute indirect */
 520                         /* rip is correct. */
 521                         next_rip = regs->rip;
 522                 }
 523                 break;
 524         case 0xea:              /* jmp absolute -- rip is correct */
 525                 next_rip = regs->rip;
 526                 break;
 527         default:
 528                 break;
 529         }
 530
 531         regs->eflags &= ~TF_MASK;
 532         if (next_rip) {
 533                 regs->rip = next_rip;
 534         } else {
 535                 regs->rip = orig_rip + (regs->rip - copy_rip);
 536         }
 537 }
 538
 539 /*
 540  * Interrupts are disabled on entry as trap1 is an interrupt gate and they
 541  * remain disabled thoroughout this function.  And we hold kprobe lock.
 542  */
 543 int __kprobes post_kprobe_handler(struct pt_regs *regs)
 544 {
 545         struct kprobe *cur = kprobe_running();
 546         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 547
 548         if (!cur)
 549                 return 0;
 550
 551         if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
 552                 kcb->kprobe_status = KPROBE_HIT_SSDONE;
 553                 cur->post_handler(cur, regs, 0);
 554         }
 555
 556         resume_execution(cur, regs, kcb);
 557         regs->eflags |= kcb->kprobe_saved_rflags;
 558
 559         /* Restore the original saved kprobes variables and continue. */
 560         if (kcb->kprobe_status == KPROBE_REENTER) {
 561                 restore_previous_kprobe(kcb);
 562                 goto out;
 563         } else {
 564                 unlock_kprobes();
 565         }
 566         reset_current_kprobe();
 567 out:
 568         preempt_enable_no_resched();
 569
 570         /*
 571          * if somebody else is singlestepping across a probe point, eflags
 572          * will have TF set, in which case, continue the remaining processing
 573          * of do_debug, as if this is not a probe hit.
 574          */
 575         if (regs->eflags & TF_MASK)
 576                 return 0;
 577
 578         return 1;
 579 }
 580
 581 /* Interrupts disabled, kprobe_lock held. */
 582 int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
 583 {
 584         struct kprobe *cur = kprobe_running();
 585         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 586
 587         if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
 588                 return 1;
 589
 590         if (kcb->kprobe_status & KPROBE_HIT_SS) {
 591                 resume_execution(cur, regs, kcb);
 592                 regs->eflags |= kcb->kprobe_old_rflags;
 593
 594                 reset_current_kprobe();
 595                 unlock_kprobes();
 596                 preempt_enable_no_resched();
 597         }
 598         return 0;
 599 }
 600
 601 /*
 602  * Wrapper routine for handling exceptions.
 603  */
 604 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
 605                                        unsigned long val, void *data)
 606 {
 607         struct die_args *args = (struct die_args *)data;
 608         int ret = NOTIFY_DONE;
 609
 610         preempt_disable();
 611         switch (val) {
 612         case DIE_INT3:
 613                 if (kprobe_handler(args->regs))
 614                         ret = NOTIFY_STOP;
 615                 break;
 616         case DIE_DEBUG:
 617                 if (post_kprobe_handler(args->regs))
 618                         ret = NOTIFY_STOP;
 619                 break;
 620         case DIE_GPF:
 621         case DIE_PAGE_FAULT:
 622                 if (kprobe_running() &&
 623                     kprobe_fault_handler(args->regs, args->trapnr))
 624                         ret = NOTIFY_STOP;
 625                 break;
 626         default:
 627                 break;
 628         }
 629         preempt_enable();
 630         return ret;
 631 }
 632
 633 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
 634 {
 635         struct jprobe *jp = container_of(p, struct jprobe, kp);
 636         unsigned long addr;
 637         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 638
 639         kcb->jprobe_saved_regs = *regs;
 640         kcb->jprobe_saved_rsp = (long *) regs->rsp;
 641         addr = (unsigned long)(kcb->jprobe_saved_rsp);
 642         /*
 643          * As Linus pointed out, gcc assumes that the callee
 644          * owns the argument space and could overwrite it, e.g.
 645          * tailcall optimization. So, to be absolutely safe
 646          * we also save and restore enough stack bytes to cover
 647          * the argument area.
 648          */
 649         memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
 650                         MIN_STACK_SIZE(addr));
 651         regs->eflags &= ~IF_MASK;
 652         regs->rip = (unsigned long)(jp->entry);
 653         return 1;
 654 }
 655
 656 void __kprobes jprobe_return(void)
 657 {
 658         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 659
 660         asm volatile ("       xchg   %%rbx,%%rsp     \n"
 661                       "       int3                      \n"
 662                       "       .globl jprobe_return_end  \n"
 663                       "       jprobe_return_end:        \n"
 664                       "       nop                       \n"::"b"
 665                       (kcb->jprobe_saved_rsp):"memory");
 666 }
 667
 668 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
 669 {
 670         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 671         u8 *addr = (u8 *) (regs->rip - 1);
 672         unsigned long stack_addr = (unsigned long)(kcb->jprobe_saved_rsp);
 673         struct jprobe *jp = container_of(p, struct jprobe, kp);
 674
 675         if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
 676                 if ((long *)regs->rsp != kcb->jprobe_saved_rsp) {
 677                         struct pt_regs *saved_regs =
 678                             container_of(kcb->jprobe_saved_rsp,
 679                                             struct pt_regs, rsp);
 680                         printk("current rsp %p does not match saved rsp %p\n",
 681                                (long *)regs->rsp, kcb->jprobe_saved_rsp);
 682                         printk("Saved registers for jprobe %p\n", jp);
 683                         show_registers(saved_regs);
 684                         printk("Current registers\n");
 685                         show_registers(regs);
 686                         BUG();
 687                 }
 688                 *regs = kcb->jprobe_saved_regs;
 689                 memcpy((kprobe_opcode_t *) stack_addr, kcb->jprobes_stack,
 690                        MIN_STACK_SIZE(stack_addr));
 691                 return 1;
 692         }
 693         return 0;
 694 }
 695
 696 static struct kprobe trampoline_p = {
 697         .addr = (kprobe_opcode_t *) &kretprobe_trampoline,
 698         .pre_handler = trampoline_probe_handler
 699 };
 700
 701 int __init arch_init_kprobes(void)
 702 {
 703         return register_kprobe(&trampoline_p);
 704 }