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 #include <linux/moduleloader.h>
  42 #include <asm/cacheflush.h>
  43 #include <asm/pgtable.h>
  44 #include <asm/kdebug.h>
  45
  46 static DECLARE_MUTEX(kprobe_mutex);
  47
  48 /* kprobe_status settings */
  49 #define KPROBE_HIT_ACTIVE       0x00000001
  50 #define KPROBE_HIT_SS           0x00000002
  51
  52 static struct kprobe *current_kprobe;
  53 static unsigned long kprobe_status, kprobe_old_rflags, kprobe_saved_rflags;
  54 static struct pt_regs jprobe_saved_regs;
  55 static long *jprobe_saved_rsp;
  56 static kprobe_opcode_t *get_insn_slot(void);
  57 static void free_insn_slot(kprobe_opcode_t *slot);
  58 void jprobe_return_end(void);
  59
  60 /* copy of the kernel stack at the probe fire time */
  61 static kprobe_opcode_t jprobes_stack[MAX_STACK_SIZE];
  62
  63 /*
  64  * returns non-zero if opcode modifies the interrupt flag.
  65  */
  66 static inline int is_IF_modifier(kprobe_opcode_t *insn)
  67 {
  68         switch (*insn) {
  69         case 0xfa:              /* cli */
  70         case 0xfb:              /* sti */
  71         case 0xcf:              /* iret/iretd */
  72         case 0x9d:              /* popf/popfd */
  73                 return 1;
  74         }
  75
  76         if (*insn  >= 0x40 && *insn <= 0x4f && *++insn == 0xcf)
  77                 return 1;
  78         return 0;
  79 }
  80
  81 int arch_prepare_kprobe(struct kprobe *p)
  82 {
  83         /* insn: must be on special executable page on x86_64. */
  84         up(&kprobe_mutex);
  85         p->ainsn.insn = get_insn_slot();
  86         down(&kprobe_mutex);
  87         if (!p->ainsn.insn) {
  88                 return -ENOMEM;
  89         }
  90         return 0;
  91 }
  92
  93 /*
  94  * Determine if the instruction uses the %rip-relative addressing mode.
  95  * If it does, return the address of the 32-bit displacement word.
  96  * If not, return null.
  97  */
  98 static inline s32 *is_riprel(u8 *insn)
  99 {
 100 #define W(row,b0,b1,b2,b3,b4,b5,b6,b7,b8,b9,ba,bb,bc,bd,be,bf)                \
 101         (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
 102           (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
 103           (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
 104           (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
 105          << (row % 64))
 106         static const u64 onebyte_has_modrm[256 / 64] = {
 107                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
 108                 /*      -------------------------------         */
 109                 W(0x00, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 00 */
 110                 W(0x10, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 10 */
 111                 W(0x20, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 20 */
 112                 W(0x30, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0), /* 30 */
 113                 W(0x40, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 40 */
 114                 W(0x50, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 50 */
 115                 W(0x60, 0,0,1,1,0,0,0,0,0,1,0,1,0,0,0,0)| /* 60 */
 116                 W(0x70, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 70 */
 117                 W(0x80, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 80 */
 118                 W(0x90, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 90 */
 119                 W(0xa0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* a0 */
 120                 W(0xb0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* b0 */
 121                 W(0xc0, 1,1,0,0,1,1,1,1,0,0,0,0,0,0,0,0)| /* c0 */
 122                 W(0xd0, 1,1,1,1,0,0,0,0,1,1,1,1,1,1,1,1)| /* d0 */
 123                 W(0xe0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* e0 */
 124                 W(0xf0, 0,0,0,0,0,0,1,1,0,0,0,0,0,0,1,1)  /* f0 */
 125                 /*      -------------------------------         */
 126                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
 127         };
 128         static const u64 twobyte_has_modrm[256 / 64] = {
 129                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
 130                 /*      -------------------------------         */
 131                 W(0x00, 1,1,1,1,0,0,0,0,0,0,0,0,0,1,0,1)| /* 0f */
 132                 W(0x10, 1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0)| /* 1f */
 133                 W(0x20, 1,1,1,1,1,0,1,0,1,1,1,1,1,1,1,1)| /* 2f */
 134                 W(0x30, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 3f */
 135                 W(0x40, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 4f */
 136                 W(0x50, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 5f */
 137                 W(0x60, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 6f */
 138                 W(0x70, 1,1,1,1,1,1,1,0,0,0,0,0,1,1,1,1), /* 7f */
 139                 W(0x80, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 8f */
 140                 W(0x90, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 9f */
 141                 W(0xa0, 0,0,0,1,1,1,1,1,0,0,0,1,1,1,1,1)| /* af */
 142                 W(0xb0, 1,1,1,1,1,1,1,1,0,0,1,1,1,1,1,1), /* bf */
 143                 W(0xc0, 1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0)| /* cf */
 144                 W(0xd0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* df */
 145                 W(0xe0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* ef */
 146                 W(0xf0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,0)  /* ff */
 147                 /*      -------------------------------         */
 148                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
 149         };
 150 #undef  W
 151         int need_modrm;
 152
 153         /* Skip legacy instruction prefixes.  */
 154         while (1) {
 155                 switch (*insn) {
 156                 case 0x66:
 157                 case 0x67:
 158                 case 0x2e:
 159                 case 0x3e:
 160                 case 0x26:
 161                 case 0x64:
 162                 case 0x65:
 163                 case 0x36:
 164                 case 0xf0:
 165                 case 0xf3:
 166                 case 0xf2:
 167                         ++insn;
 168                         continue;
 169                 }
 170                 break;
 171         }
 172
 173         /* Skip REX instruction prefix.  */
 174         if ((*insn & 0xf0) == 0x40)
 175                 ++insn;
 176
 177         if (*insn == 0x0f) {    /* Two-byte opcode.  */
 178                 ++insn;
 179                 need_modrm = test_bit(*insn, twobyte_has_modrm);
 180         } else {                /* One-byte opcode.  */
 181                 need_modrm = test_bit(*insn, onebyte_has_modrm);
 182         }
 183
 184         if (need_modrm) {
 185                 u8 modrm = *++insn;
 186                 if ((modrm & 0xc7) == 0x05) { /* %rip+disp32 addressing mode */
 187                         /* Displacement follows ModRM byte.  */
 188                         return (s32 *) ++insn;
 189                 }
 190         }
 191
 192         /* No %rip-relative addressing mode here.  */
 193         return NULL;
 194 }
 195
 196 void arch_copy_kprobe(struct kprobe *p)
 197 {
 198         s32 *ripdisp;
 199         memcpy(p->ainsn.insn, p->addr, MAX_INSN_SIZE);
 200         ripdisp = is_riprel(p->ainsn.insn);
 201         if (ripdisp) {
 202                 /*
 203                  * The copied instruction uses the %rip-relative
 204                  * addressing mode.  Adjust the displacement for the
 205                  * difference between the original location of this
 206                  * instruction and the location of the copy that will
 207                  * actually be run.  The tricky bit here is making sure
 208                  * that the sign extension happens correctly in this
 209                  * calculation, since we need a signed 32-bit result to
 210                  * be sign-extended to 64 bits when it's added to the
 211                  * %rip value and yield the same 64-bit result that the
 212                  * sign-extension of the original signed 32-bit
 213                  * displacement would have given.
 214                  */
 215                 s64 disp = (u8 *) p->addr + *ripdisp - (u8 *) p->ainsn.insn;
 216                 BUG_ON((s64) (s32) disp != disp); /* Sanity check.  */
 217                 *ripdisp = disp;
 218         }
 219         p->opcode = *p->addr;
 220 }
 221
 222 void arch_arm_kprobe(struct kprobe *p)
 223 {
 224         *p->addr = BREAKPOINT_INSTRUCTION;
 225         flush_icache_range((unsigned long) p->addr,
 226                            (unsigned long) p->addr + sizeof(kprobe_opcode_t));
 227 }
 228
 229 void arch_disarm_kprobe(struct kprobe *p)
 230 {
 231         *p->addr = p->opcode;
 232         flush_icache_range((unsigned long) p->addr,
 233                            (unsigned long) p->addr + sizeof(kprobe_opcode_t));
 234 }
 235
 236 void arch_remove_kprobe(struct kprobe *p)
 237 {
 238         up(&kprobe_mutex);
 239         free_insn_slot(p->ainsn.insn);
 240         down(&kprobe_mutex);
 241 }
 242
 243 static void prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
 244 {
 245         regs->eflags |= TF_MASK;
 246         regs->eflags &= ~IF_MASK;
 247         /*single step inline if the instruction is an int3*/
 248         if (p->opcode == BREAKPOINT_INSTRUCTION)
 249                 regs->rip = (unsigned long)p->addr;
 250         else
 251                 regs->rip = (unsigned long)p->ainsn.insn;
 252 }
 253
 254 struct task_struct  *arch_get_kprobe_task(void *ptr)
 255 {
 256         return ((struct thread_info *) (((unsigned long) ptr) &
 257                                         (~(THREAD_SIZE -1))))->task;
 258 }
 259
 260 void arch_prepare_kretprobe(struct kretprobe *rp, struct pt_regs *regs)
 261 {
 262         unsigned long *sara = (unsigned long *)regs->rsp;
 263         struct kretprobe_instance *ri;
 264         static void *orig_ret_addr;
 265
 266         /*
 267          * Save the return address when the return probe hits
 268          * the first time, and use it to populate the (krprobe
 269          * instance)->ret_addr for subsequent return probes at
 270          * the same addrress since stack address would have
 271          * the kretprobe_trampoline by then.
 272          */
 273         if (((void*) *sara) != kretprobe_trampoline)
 274                 orig_ret_addr = (void*) *sara;
 275
 276         if ((ri = get_free_rp_inst(rp)) != NULL) {
 277                 ri->rp = rp;
 278                 ri->stack_addr = sara;
 279                 ri->ret_addr = orig_ret_addr;
 280                 add_rp_inst(ri);
 281                 /* Replace the return addr with trampoline addr */
 282                 *sara = (unsigned long) &kretprobe_trampoline;
 283         } else {
 284                 rp->nmissed++;
 285         }
 286 }
 287
 288 void arch_kprobe_flush_task(struct task_struct *tk)
 289 {
 290         struct kretprobe_instance *ri;
 291         while ((ri = get_rp_inst_tsk(tk)) != NULL) {
 292                 *((unsigned long *)(ri->stack_addr)) =
 293                                         (unsigned long) ri->ret_addr;
 294                 recycle_rp_inst(ri);
 295         }
 296 }
 297
 298 /*
 299  * Interrupts are disabled on entry as trap3 is an interrupt gate and they
 300  * remain disabled thorough out this function.
 301  */
 302 int kprobe_handler(struct pt_regs *regs)
 303 {
 304         struct kprobe *p;
 305         int ret = 0;
 306         kprobe_opcode_t *addr = (kprobe_opcode_t *)(regs->rip - sizeof(kprobe_opcode_t));
 307
 308         /* We're in an interrupt, but this is clear and BUG()-safe. */
 309         preempt_disable();
 310
 311         /* Check we're not actually recursing */
 312         if (kprobe_running()) {
 313                 /* We *are* holding lock here, so this is safe.
 314                    Disarm the probe we just hit, and ignore it. */
 315                 p = get_kprobe(addr);
 316                 if (p) {
 317                         if (kprobe_status == KPROBE_HIT_SS) {
 318                                 regs->eflags &= ~TF_MASK;
 319                                 regs->eflags |= kprobe_saved_rflags;
 320                                 unlock_kprobes();
 321                                 goto no_kprobe;
 322                         }
 323                         arch_disarm_kprobe(p);
 324                         regs->rip = (unsigned long)p->addr;
 325                         ret = 1;
 326                 } else {
 327                         p = current_kprobe;
 328                         if (p->break_handler && p->break_handler(p, regs)) {
 329                                 goto ss_probe;
 330                         }
 331                 }
 332                 /* If it's not ours, can't be delete race, (we hold lock). */
 333                 goto no_kprobe;
 334         }
 335
 336         lock_kprobes();
 337         p = get_kprobe(addr);
 338         if (!p) {
 339                 unlock_kprobes();
 340                 if (*addr != BREAKPOINT_INSTRUCTION) {
 341                         /*
 342                          * The breakpoint instruction was removed right
 343                          * after we hit it.  Another cpu has removed
 344                          * either a probepoint or a debugger breakpoint
 345                          * at this address.  In either case, no further
 346                          * handling of this interrupt is appropriate.
 347                          */
 348                         ret = 1;
 349                 }
 350                 /* Not one of ours: let kernel handle it */
 351                 goto no_kprobe;
 352         }
 353
 354         kprobe_status = KPROBE_HIT_ACTIVE;
 355         current_kprobe = p;
 356         kprobe_saved_rflags = kprobe_old_rflags
 357             = (regs->eflags & (TF_MASK | IF_MASK));
 358         if (is_IF_modifier(p->ainsn.insn))
 359                 kprobe_saved_rflags &= ~IF_MASK;
 360
 361         if (p->pre_handler && p->pre_handler(p, regs))
 362                 /* handler has already set things up, so skip ss setup */
 363                 return 1;
 364
 365 ss_probe:
 366         prepare_singlestep(p, regs);
 367         kprobe_status = KPROBE_HIT_SS;
 368         return 1;
 369
 370 no_kprobe:
 371         preempt_enable_no_resched();
 372         return ret;
 373 }
 374
 375 /*
 376  * For function-return probes, init_kprobes() establishes a probepoint
 377  * here. When a retprobed function returns, this probe is hit and
 378  * trampoline_probe_handler() runs, calling the kretprobe's handler.
 379  */
 380  void kretprobe_trampoline_holder(void)
 381  {
 382         asm volatile (  ".global kretprobe_trampoline\n"
 383                         "kretprobe_trampoline: \n"
 384                         "nop\n");
 385  }
 386
 387 /*
 388  * Called when we hit the probe point at kretprobe_trampoline
 389  */
 390 int trampoline_probe_handler(struct kprobe *p, struct pt_regs *regs)
 391 {
 392         struct task_struct *tsk;
 393         struct kretprobe_instance *ri;
 394         struct hlist_head *head;
 395         struct hlist_node *node;
 396         unsigned long *sara = (unsigned long *)regs->rsp - 1;
 397
 398         tsk = arch_get_kprobe_task(sara);
 399         head = kretprobe_inst_table_head(tsk);
 400
 401         hlist_for_each_entry(ri, node, head, hlist) {
 402                 if (ri->stack_addr == sara && ri->rp) {
 403                         if (ri->rp->handler)
 404                                 ri->rp->handler(ri, regs);
 405                 }
 406         }
 407         return 0;
 408 }
 409
 410 void trampoline_post_handler(struct kprobe *p, struct pt_regs *regs,
 411                                                 unsigned long flags)
 412 {
 413         struct kretprobe_instance *ri;
 414         /* RA already popped */
 415         unsigned long *sara = ((unsigned long *)regs->rsp) - 1;
 416
 417         while ((ri = get_rp_inst(sara))) {
 418                 regs->rip = (unsigned long)ri->ret_addr;
 419                 recycle_rp_inst(ri);
 420         }
 421         regs->eflags &= ~TF_MASK;
 422 }
 423
 424 /*
 425  * Called after single-stepping.  p->addr is the address of the
 426  * instruction whose first byte has been replaced by the "int 3"
 427  * instruction.  To avoid the SMP problems that can occur when we
 428  * temporarily put back the original opcode to single-step, we
 429  * single-stepped a copy of the instruction.  The address of this
 430  * copy is p->ainsn.insn.
 431  *
 432  * This function prepares to return from the post-single-step
 433  * interrupt.  We have to fix up the stack as follows:
 434  *
 435  * 0) Except in the case of absolute or indirect jump or call instructions,
 436  * the new rip is relative to the copied instruction.  We need to make
 437  * it relative to the original instruction.
 438  *
 439  * 1) If the single-stepped instruction was pushfl, then the TF and IF
 440  * flags are set in the just-pushed eflags, and may need to be cleared.
 441  *
 442  * 2) If the single-stepped instruction was a call, the return address
 443  * that is atop the stack is the address following the copied instruction.
 444  * We need to make it the address following the original instruction.
 445  */
 446 static void resume_execution(struct kprobe *p, struct pt_regs *regs)
 447 {
 448         unsigned long *tos = (unsigned long *)regs->rsp;
 449         unsigned long next_rip = 0;
 450         unsigned long copy_rip = (unsigned long)p->ainsn.insn;
 451         unsigned long orig_rip = (unsigned long)p->addr;
 452         kprobe_opcode_t *insn = p->ainsn.insn;
 453
 454         /*skip the REX prefix*/
 455         if (*insn >= 0x40 && *insn <= 0x4f)
 456                 insn++;
 457
 458         switch (*insn) {
 459         case 0x9c:              /* pushfl */
 460                 *tos &= ~(TF_MASK | IF_MASK);
 461                 *tos |= kprobe_old_rflags;
 462                 break;
 463         case 0xc3:              /* ret/lret */
 464         case 0xcb:
 465         case 0xc2:
 466         case 0xca:
 467                 regs->eflags &= ~TF_MASK;
 468                 /* rip is already adjusted, no more changes required*/
 469                 return;
 470         case 0xe8:              /* call relative - Fix return addr */
 471                 *tos = orig_rip + (*tos - copy_rip);
 472                 break;
 473         case 0xff:
 474                 if ((*insn & 0x30) == 0x10) {
 475                         /* call absolute, indirect */
 476                         /* Fix return addr; rip is correct. */
 477                         next_rip = regs->rip;
 478                         *tos = orig_rip + (*tos - copy_rip);
 479                 } else if (((*insn & 0x31) == 0x20) ||  /* jmp near, absolute indirect */
 480                            ((*insn & 0x31) == 0x21)) {  /* jmp far, absolute indirect */
 481                         /* rip is correct. */
 482                         next_rip = regs->rip;
 483                 }
 484                 break;
 485         case 0xea:              /* jmp absolute -- rip is correct */
 486                 next_rip = regs->rip;
 487                 break;
 488         default:
 489                 break;
 490         }
 491
 492         regs->eflags &= ~TF_MASK;
 493         if (next_rip) {
 494                 regs->rip = next_rip;
 495         } else {
 496                 regs->rip = orig_rip + (regs->rip - copy_rip);
 497         }
 498 }
 499
 500 /*
 501  * Interrupts are disabled on entry as trap1 is an interrupt gate and they
 502  * remain disabled thoroughout this function.  And we hold kprobe lock.
 503  */
 504 int post_kprobe_handler(struct pt_regs *regs)
 505 {
 506         if (!kprobe_running())
 507                 return 0;
 508
 509         if (current_kprobe->post_handler)
 510                 current_kprobe->post_handler(current_kprobe, regs, 0);
 511
 512         if (current_kprobe->post_handler != trampoline_post_handler)
 513                 resume_execution(current_kprobe, regs);
 514         regs->eflags |= kprobe_saved_rflags;
 515
 516         unlock_kprobes();
 517         preempt_enable_no_resched();
 518
 519         /*
 520          * if somebody else is singlestepping across a probe point, eflags
 521          * will have TF set, in which case, continue the remaining processing
 522          * of do_debug, as if this is not a probe hit.
 523          */
 524         if (regs->eflags & TF_MASK)
 525                 return 0;
 526
 527         return 1;
 528 }
 529
 530 /* Interrupts disabled, kprobe_lock held. */
 531 int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
 532 {
 533         if (current_kprobe->fault_handler
 534             && current_kprobe->fault_handler(current_kprobe, regs, trapnr))
 535                 return 1;
 536
 537         if (kprobe_status & KPROBE_HIT_SS) {
 538                 resume_execution(current_kprobe, regs);
 539                 regs->eflags |= kprobe_old_rflags;
 540
 541                 unlock_kprobes();
 542                 preempt_enable_no_resched();
 543         }
 544         return 0;
 545 }
 546
 547 /*
 548  * Wrapper routine for handling exceptions.
 549  */
 550 int kprobe_exceptions_notify(struct notifier_block *self, unsigned long val,
 551                              void *data)
 552 {
 553         struct die_args *args = (struct die_args *)data;
 554         switch (val) {
 555         case DIE_INT3:
 556                 if (kprobe_handler(args->regs))
 557                         return NOTIFY_STOP;
 558                 break;
 559         case DIE_DEBUG:
 560                 if (post_kprobe_handler(args->regs))
 561                         return NOTIFY_STOP;
 562                 break;
 563         case DIE_GPF:
 564                 if (kprobe_running() &&
 565                     kprobe_fault_handler(args->regs, args->trapnr))
 566                         return NOTIFY_STOP;
 567                 break;
 568         case DIE_PAGE_FAULT:
 569                 if (kprobe_running() &&
 570                     kprobe_fault_handler(args->regs, args->trapnr))
 571                         return NOTIFY_STOP;
 572                 break;
 573         default:
 574                 break;
 575         }
 576         return NOTIFY_DONE;
 577 }
 578
 579 int setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
 580 {
 581         struct jprobe *jp = container_of(p, struct jprobe, kp);
 582         unsigned long addr;
 583
 584         jprobe_saved_regs = *regs;
 585         jprobe_saved_rsp = (long *) regs->rsp;
 586         addr = (unsigned long)jprobe_saved_rsp;
 587         /*
 588          * As Linus pointed out, gcc assumes that the callee
 589          * owns the argument space and could overwrite it, e.g.
 590          * tailcall optimization. So, to be absolutely safe
 591          * we also save and restore enough stack bytes to cover
 592          * the argument area.
 593          */
 594         memcpy(jprobes_stack, (kprobe_opcode_t *) addr, MIN_STACK_SIZE(addr));
 595         regs->eflags &= ~IF_MASK;
 596         regs->rip = (unsigned long)(jp->entry);
 597         return 1;
 598 }
 599
 600 void jprobe_return(void)
 601 {
 602         preempt_enable_no_resched();
 603         asm volatile ("       xchg   %%rbx,%%rsp     \n"
 604                       "       int3                      \n"
 605                       "       .globl jprobe_return_end  \n"
 606                       "       jprobe_return_end:        \n"
 607                       "       nop                       \n"::"b"
 608                       (jprobe_saved_rsp):"memory");
 609 }
 610
 611 int longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
 612 {
 613         u8 *addr = (u8 *) (regs->rip - 1);
 614         unsigned long stack_addr = (unsigned long)jprobe_saved_rsp;
 615         struct jprobe *jp = container_of(p, struct jprobe, kp);
 616
 617         if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
 618                 if ((long *)regs->rsp != jprobe_saved_rsp) {
 619                         struct pt_regs *saved_regs =
 620                             container_of(jprobe_saved_rsp, struct pt_regs, rsp);
 621                         printk("current rsp %p does not match saved rsp %p\n",
 622                                (long *)regs->rsp, jprobe_saved_rsp);
 623                         printk("Saved registers for jprobe %p\n", jp);
 624                         show_registers(saved_regs);
 625                         printk("Current registers\n");
 626                         show_registers(regs);
 627                         BUG();
 628                 }
 629                 *regs = jprobe_saved_regs;
 630                 memcpy((kprobe_opcode_t *) stack_addr, jprobes_stack,
 631                        MIN_STACK_SIZE(stack_addr));
 632                 return 1;
 633         }
 634         return 0;
 635 }
 636
 637 /*
 638  * kprobe->ainsn.insn points to the copy of the instruction to be single-stepped.
 639  * By default on x86_64, pages we get from kmalloc or vmalloc are not
 640  * executable.  Single-stepping an instruction on such a page yields an
 641  * oops.  So instead of storing the instruction copies in their respective
 642  * kprobe objects, we allocate a page, map it executable, and store all the
 643  * instruction copies there.  (We can allocate additional pages if somebody
 644  * inserts a huge number of probes.)  Each page can hold up to INSNS_PER_PAGE
 645  * instruction slots, each of which is MAX_INSN_SIZE*sizeof(kprobe_opcode_t)
 646  * bytes.
 647  */
 648 #define INSNS_PER_PAGE (PAGE_SIZE/(MAX_INSN_SIZE*sizeof(kprobe_opcode_t)))
 649 struct kprobe_insn_page {
 650         struct hlist_node hlist;
 651         kprobe_opcode_t *insns;         /* page of instruction slots */
 652         char slot_used[INSNS_PER_PAGE];
 653         int nused;
 654 };
 655
 656 static struct hlist_head kprobe_insn_pages;
 657
 658 /**
 659  * get_insn_slot() - Find a slot on an executable page for an instruction.
 660  * We allocate an executable page if there's no room on existing ones.
 661  */
 662 static kprobe_opcode_t *get_insn_slot(void)
 663 {
 664         struct kprobe_insn_page *kip;
 665         struct hlist_node *pos;
 666
 667         hlist_for_each(pos, &kprobe_insn_pages) {
 668                 kip = hlist_entry(pos, struct kprobe_insn_page, hlist);
 669                 if (kip->nused < INSNS_PER_PAGE) {
 670                         int i;
 671                         for (i = 0; i < INSNS_PER_PAGE; i++) {
 672                                 if (!kip->slot_used[i]) {
 673                                         kip->slot_used[i] = 1;
 674                                         kip->nused++;
 675                                         return kip->insns + (i*MAX_INSN_SIZE);
 676                                 }
 677                         }
 678                         /* Surprise!  No unused slots.  Fix kip->nused. */
 679                         kip->nused = INSNS_PER_PAGE;
 680                 }
 681         }
 682
 683         /* All out of space.  Need to allocate a new page. Use slot 0.*/
 684         kip = kmalloc(sizeof(struct kprobe_insn_page), GFP_KERNEL);
 685         if (!kip) {
 686                 return NULL;
 687         }
 688
 689         /*
 690          * For the %rip-relative displacement fixups to be doable, we
 691          * need our instruction copy to be within +/- 2GB of any data it
 692          * might access via %rip.  That is, within 2GB of where the
 693          * kernel image and loaded module images reside.  So we allocate
 694          * a page in the module loading area.
 695          */
 696         kip->insns = module_alloc(PAGE_SIZE);
 697         if (!kip->insns) {
 698                 kfree(kip);
 699                 return NULL;
 700         }
 701         INIT_HLIST_NODE(&kip->hlist);
 702         hlist_add_head(&kip->hlist, &kprobe_insn_pages);
 703         memset(kip->slot_used, 0, INSNS_PER_PAGE);
 704         kip->slot_used[0] = 1;
 705         kip->nused = 1;
 706         return kip->insns;
 707 }
 708
 709 /**
 710  * free_insn_slot() - Free instruction slot obtained from get_insn_slot().
 711  */
 712 static void free_insn_slot(kprobe_opcode_t *slot)
 713 {
 714         struct kprobe_insn_page *kip;
 715         struct hlist_node *pos;
 716
 717         hlist_for_each(pos, &kprobe_insn_pages) {
 718                 kip = hlist_entry(pos, struct kprobe_insn_page, hlist);
 719                 if (kip->insns <= slot
 720                     && slot < kip->insns+(INSNS_PER_PAGE*MAX_INSN_SIZE)) {
 721                         int i = (slot - kip->insns) / MAX_INSN_SIZE;
 722                         kip->slot_used[i] = 0;
 723                         kip->nused--;
 724                         if (kip->nused == 0) {
 725                                 /*
 726                                  * Page is no longer in use.  Free it unless
 727                                  * it's the last one.  We keep the last one
 728                                  * so as not to have to set it up again the
 729                                  * next time somebody inserts a probe.
 730                                  */
 731                                 hlist_del(&kip->hlist);
 732                                 if (hlist_empty(&kprobe_insn_pages)) {
 733                                         INIT_HLIST_NODE(&kip->hlist);
 734                                         hlist_add_head(&kip->hlist,
 735                                                 &kprobe_insn_pages);
 736                                 } else {
 737                                         module_free(NULL, kip->insns);
 738                                         kfree(kip);
 739                                 }
 740                         }
 741                         return;
 742                 }
 743         }
 744 }