arch/x86/kernel/kprobes_32.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  * 2005-May     Hien Nguyen <hien@us.ibm.com>, Jim Keniston
  26  *              <jkenisto@us.ibm.com> and Prasanna S Panchamukhi
  27  *              <prasanna@in.ibm.com> added function-return probes.
  28  */
  29
  30 #include <linux/kprobes.h>
  31 #include <linux/ptrace.h>
  32 #include <linux/preempt.h>
  33 #include <linux/kdebug.h>
  34 #include <asm/cacheflush.h>
  35 #include <asm/desc.h>
  36 #include <asm/uaccess.h>
  37 #include <asm/alternative.h>
  38
  39 void jprobe_return_end(void);
  40
  41 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
  42 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
  43
  44 struct kretprobe_blackpoint kretprobe_blacklist[] = {
  45         {"__switch_to", }, /* This function switches only current task, but
  46                              doesn't switch kernel stack.*/
  47         {NULL, NULL}    /* Terminator */
  48 };
  49 const int kretprobe_blacklist_size = ARRAY_SIZE(kretprobe_blacklist);
  50
  51 /* insert a jmp code */
  52 static __always_inline void set_jmp_op(void *from, void *to)
  53 {
  54         struct __arch_jmp_op {
  55                 char op;
  56                 long raddr;
  57         } __attribute__((packed)) *jop;
  58         jop = (struct __arch_jmp_op *)from;
  59         jop->raddr = (long)(to) - ((long)(from) + 5);
  60         jop->op = RELATIVEJUMP_INSTRUCTION;
  61 }
  62
  63 /*
  64  * returns non-zero if opcodes can be boosted.
  65  */
  66 static __always_inline int can_boost(kprobe_opcode_t *opcodes)
  67 {
  68 #define W(row,b0,b1,b2,b3,b4,b5,b6,b7,b8,b9,ba,bb,bc,bd,be,bf)                \
  69         (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
  70           (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
  71           (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
  72           (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
  73          << (row % 32))
  74         /*
  75          * Undefined/reserved opcodes, conditional jump, Opcode Extension
  76          * Groups, and some special opcodes can not be boost.
  77          */
  78         static const unsigned long twobyte_is_boostable[256 / 32] = {
  79                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
  80                 /*      -------------------------------         */
  81                 W(0x00, 0,0,1,1,0,0,1,0,1,1,0,0,0,0,0,0)| /* 00 */
  82                 W(0x10, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 10 */
  83                 W(0x20, 1,1,1,1,0,0,0,0,0,0,0,0,0,0,0,0)| /* 20 */
  84                 W(0x30, 0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 30 */
  85                 W(0x40, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 40 */
  86                 W(0x50, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 50 */
  87                 W(0x60, 1,1,1,1,1,1,1,1,1,1,1,1,0,0,1,1)| /* 60 */
  88                 W(0x70, 0,0,0,0,1,1,1,1,0,0,0,0,0,0,1,1), /* 70 */
  89                 W(0x80, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 80 */
  90                 W(0x90, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1), /* 90 */
  91                 W(0xa0, 1,1,0,1,1,1,0,0,1,1,0,1,1,1,0,1)| /* a0 */
  92                 W(0xb0, 1,1,1,1,1,1,1,1,0,0,0,1,1,1,1,1), /* b0 */
  93                 W(0xc0, 1,1,0,0,0,0,0,0,1,1,1,1,1,1,1,1)| /* c0 */
  94                 W(0xd0, 0,1,1,1,0,1,0,0,1,1,0,1,1,1,0,1), /* d0 */
  95                 W(0xe0, 0,1,1,0,0,1,0,0,1,1,0,1,1,1,0,1)| /* e0 */
  96                 W(0xf0, 0,1,1,1,0,1,0,0,1,1,1,0,1,1,1,0)  /* f0 */
  97                 /*      -------------------------------         */
  98                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
  99         };
 100 #undef W
 101         kprobe_opcode_t opcode;
 102         kprobe_opcode_t *orig_opcodes = opcodes;
 103 retry:
 104         if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
 105                 return 0;
 106         opcode = *(opcodes++);
 107
 108         /* 2nd-byte opcode */
 109         if (opcode == 0x0f) {
 110                 if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
 111                         return 0;
 112                 return test_bit(*opcodes, twobyte_is_boostable);
 113         }
 114
 115         switch (opcode & 0xf0) {
 116         case 0x60:
 117                 if (0x63 < opcode && opcode < 0x67)
 118                         goto retry; /* prefixes */
 119                 /* can't boost Address-size override and bound */
 120                 return (opcode != 0x62 && opcode != 0x67);
 121         case 0x70:
 122                 return 0; /* can't boost conditional jump */
 123         case 0xc0:
 124                 /* can't boost software-interruptions */
 125                 return (0xc1 < opcode && opcode < 0xcc) || opcode == 0xcf;
 126         case 0xd0:
 127                 /* can boost AA* and XLAT */
 128                 return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
 129         case 0xe0:
 130                 /* can boost in/out and absolute jmps */
 131                 return ((opcode & 0x04) || opcode == 0xea);
 132         case 0xf0:
 133                 if ((opcode & 0x0c) == 0 && opcode != 0xf1)
 134                         goto retry; /* lock/rep(ne) prefix */
 135                 /* clear and set flags can be boost */
 136                 return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
 137         default:
 138                 if (opcode == 0x26 || opcode == 0x36 || opcode == 0x3e)
 139                         goto retry; /* prefixes */
 140                 /* can't boost CS override and call */
 141                 return (opcode != 0x2e && opcode != 0x9a);
 142         }
 143 }
 144
 145 /*
 146  * returns non-zero if opcode modifies the interrupt flag.
 147  */
 148 static int __kprobes is_IF_modifier(kprobe_opcode_t opcode)
 149 {
 150         switch (opcode) {
 151         case 0xfa:              /* cli */
 152         case 0xfb:              /* sti */
 153         case 0xcf:              /* iret/iretd */
 154         case 0x9d:              /* popf/popfd */
 155                 return 1;
 156         }
 157         return 0;
 158 }
 159
 160 int __kprobes arch_prepare_kprobe(struct kprobe *p)
 161 {
 162         /* insn: must be on special executable page on i386. */
 163         p->ainsn.insn = get_insn_slot();
 164         if (!p->ainsn.insn)
 165                 return -ENOMEM;
 166
 167         memcpy(p->ainsn.insn, p->addr, MAX_INSN_SIZE * sizeof(kprobe_opcode_t));
 168         p->opcode = *p->addr;
 169         if (can_boost(p->addr)) {
 170                 p->ainsn.boostable = 0;
 171         } else {
 172                 p->ainsn.boostable = -1;
 173         }
 174         return 0;
 175 }
 176
 177 void __kprobes arch_arm_kprobe(struct kprobe *p)
 178 {
 179         text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
 180 }
 181
 182 void __kprobes arch_disarm_kprobe(struct kprobe *p)
 183 {
 184         text_poke(p->addr, &p->opcode, 1);
 185 }
 186
 187 void __kprobes arch_remove_kprobe(struct kprobe *p)
 188 {
 189         mutex_lock(&kprobe_mutex);
 190         free_insn_slot(p->ainsn.insn, (p->ainsn.boostable == 1));
 191         mutex_unlock(&kprobe_mutex);
 192 }
 193
 194 static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
 195 {
 196         kcb->prev_kprobe.kp = kprobe_running();
 197         kcb->prev_kprobe.status = kcb->kprobe_status;
 198         kcb->prev_kprobe.old_eflags = kcb->kprobe_old_eflags;
 199         kcb->prev_kprobe.saved_eflags = kcb->kprobe_saved_eflags;
 200 }
 201
 202 static void __kprobes restore_previous_kprobe(struct kprobe_ctlblk *kcb)
 203 {
 204         __get_cpu_var(current_kprobe) = kcb->prev_kprobe.kp;
 205         kcb->kprobe_status = kcb->prev_kprobe.status;
 206         kcb->kprobe_old_eflags = kcb->prev_kprobe.old_eflags;
 207         kcb->kprobe_saved_eflags = kcb->prev_kprobe.saved_eflags;
 208 }
 209
 210 static void __kprobes set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
 211                                 struct kprobe_ctlblk *kcb)
 212 {
 213         __get_cpu_var(current_kprobe) = p;
 214         kcb->kprobe_saved_eflags = kcb->kprobe_old_eflags
 215                 = (regs->flags & (TF_MASK | IF_MASK));
 216         if (is_IF_modifier(p->opcode))
 217                 kcb->kprobe_saved_eflags &= ~IF_MASK;
 218 }
 219
 220 static __always_inline void clear_btf(void)
 221 {
 222         if (test_thread_flag(TIF_DEBUGCTLMSR))
 223                 wrmsr(MSR_IA32_DEBUGCTLMSR, 0, 0);
 224 }
 225
 226 static __always_inline void restore_btf(void)
 227 {
 228         if (test_thread_flag(TIF_DEBUGCTLMSR))
 229                 wrmsr(MSR_IA32_DEBUGCTLMSR, current->thread.debugctlmsr, 0);
 230 }
 231
 232 static void __kprobes prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
 233 {
 234         clear_btf();
 235         regs->flags |= TF_MASK;
 236         regs->flags &= ~IF_MASK;
 237         /*single step inline if the instruction is an int3*/
 238         if (p->opcode == BREAKPOINT_INSTRUCTION)
 239                 regs->ip = (unsigned long)p->addr;
 240         else
 241                 regs->ip = (unsigned long)p->ainsn.insn;
 242 }
 243
 244 /* Called with kretprobe_lock held */
 245 void __kprobes arch_prepare_kretprobe(struct kretprobe_instance *ri,
 246                                       struct pt_regs *regs)
 247 {
 248         unsigned long *sara = (unsigned long *)&regs->sp;
 249
 250         ri->ret_addr = (kprobe_opcode_t *) *sara;
 251
 252         /* Replace the return addr with trampoline addr */
 253         *sara = (unsigned long) &kretprobe_trampoline;
 254 }
 255
 256 /*
 257  * Interrupts are disabled on entry as trap3 is an interrupt gate and they
 258  * remain disabled thorough out this function.
 259  */
 260 static int __kprobes kprobe_handler(struct pt_regs *regs)
 261 {
 262         struct kprobe *p;
 263         int ret = 0;
 264         kprobe_opcode_t *addr;
 265         struct kprobe_ctlblk *kcb;
 266
 267         addr = (kprobe_opcode_t *)(regs->ip - sizeof(kprobe_opcode_t));
 268
 269         /*
 270          * We don't want to be preempted for the entire
 271          * duration of kprobe processing
 272          */
 273         preempt_disable();
 274         kcb = get_kprobe_ctlblk();
 275
 276         /* Check we're not actually recursing */
 277         if (kprobe_running()) {
 278                 p = get_kprobe(addr);
 279                 if (p) {
 280                         if (kcb->kprobe_status == KPROBE_HIT_SS &&
 281                                 *p->ainsn.insn == BREAKPOINT_INSTRUCTION) {
 282                                 regs->flags &= ~TF_MASK;
 283                                 regs->flags |= kcb->kprobe_saved_eflags;
 284                                 goto no_kprobe;
 285                         }
 286                         /* We have reentered the kprobe_handler(), since
 287                          * another probe was hit while within the handler.
 288                          * We here save the original kprobes variables and
 289                          * just single step on the instruction of the new probe
 290                          * without calling any user handlers.
 291                          */
 292                         save_previous_kprobe(kcb);
 293                         set_current_kprobe(p, regs, kcb);
 294                         kprobes_inc_nmissed_count(p);
 295                         prepare_singlestep(p, regs);
 296                         kcb->kprobe_status = KPROBE_REENTER;
 297                         return 1;
 298                 } else {
 299                         if (*addr != BREAKPOINT_INSTRUCTION) {
 300                         /* The breakpoint instruction was removed by
 301                          * another cpu right after we hit, no further
 302                          * handling of this interrupt is appropriate
 303                          */
 304                                 regs->ip -= sizeof(kprobe_opcode_t);
 305                                 ret = 1;
 306                                 goto no_kprobe;
 307                         }
 308                         p = __get_cpu_var(current_kprobe);
 309                         if (p->break_handler && p->break_handler(p, regs)) {
 310                                 goto ss_probe;
 311                         }
 312                 }
 313                 goto no_kprobe;
 314         }
 315
 316         p = get_kprobe(addr);
 317         if (!p) {
 318                 if (*addr != BREAKPOINT_INSTRUCTION) {
 319                         /*
 320                          * The breakpoint instruction was removed right
 321                          * after we hit it.  Another cpu has removed
 322                          * either a probepoint or a debugger breakpoint
 323                          * at this address.  In either case, no further
 324                          * handling of this interrupt is appropriate.
 325                          * Back up over the (now missing) int3 and run
 326                          * the original instruction.
 327                          */
 328                         regs->ip -= sizeof(kprobe_opcode_t);
 329                         ret = 1;
 330                 }
 331                 /* Not one of ours: let kernel handle it */
 332                 goto no_kprobe;
 333         }
 334
 335         set_current_kprobe(p, regs, kcb);
 336         kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 337
 338         if (p->pre_handler && p->pre_handler(p, regs))
 339                 /* handler has already set things up, so skip ss setup */
 340                 return 1;
 341
 342 ss_probe:
 343 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PM)
 344         if (p->ainsn.boostable == 1 && !p->post_handler){
 345                 /* Boost up -- we can execute copied instructions directly */
 346                 reset_current_kprobe();
 347                 regs->ip = (unsigned long)p->ainsn.insn;
 348                 preempt_enable_no_resched();
 349                 return 1;
 350         }
 351 #endif
 352         prepare_singlestep(p, regs);
 353         kcb->kprobe_status = KPROBE_HIT_SS;
 354         return 1;
 355
 356 no_kprobe:
 357         preempt_enable_no_resched();
 358         return ret;
 359 }
 360
 361 /*
 362  * For function-return probes, init_kprobes() establishes a probepoint
 363  * here. When a retprobed function returns, this probe is hit and
 364  * trampoline_probe_handler() runs, calling the kretprobe's handler.
 365  */
 366  void __kprobes kretprobe_trampoline_holder(void)
 367  {
 368         asm volatile ( ".global kretprobe_trampoline\n"
 369                         "kretprobe_trampoline: \n"
 370                         "       pushf\n"
 371                         /* skip cs, ip, orig_ax */
 372                         "       subl $12, %esp\n"
 373                         "       pushl %fs\n"
 374                         "       pushl %ds\n"
 375                         "       pushl %es\n"
 376                         "       pushl %eax\n"
 377                         "       pushl %ebp\n"
 378                         "       pushl %edi\n"
 379                         "       pushl %esi\n"
 380                         "       pushl %edx\n"
 381                         "       pushl %ecx\n"
 382                         "       pushl %ebx\n"
 383                         "       movl %esp, %eax\n"
 384                         "       call trampoline_handler\n"
 385                         /* move flags to cs */
 386                         "       movl 52(%esp), %edx\n"
 387                         "       movl %edx, 48(%esp)\n"
 388                         /* save true return address on flags */
 389                         "       movl %eax, 52(%esp)\n"
 390                         "       popl %ebx\n"
 391                         "       popl %ecx\n"
 392                         "       popl %edx\n"
 393                         "       popl %esi\n"
 394                         "       popl %edi\n"
 395                         "       popl %ebp\n"
 396                         "       popl %eax\n"
 397                         /* skip ip, orig_ax, es, ds, fs */
 398                         "       addl $20, %esp\n"
 399                         "       popf\n"
 400                         "       ret\n");
 401 }
 402
 403 /*
 404  * Called from kretprobe_trampoline
 405  */
 406 void *__kprobes trampoline_handler(struct pt_regs *regs)
 407 {
 408         struct kretprobe_instance *ri = NULL;
 409         struct hlist_head *head, empty_rp;
 410         struct hlist_node *node, *tmp;
 411         unsigned long flags, orig_ret_address = 0;
 412         unsigned long trampoline_address =(unsigned long)&kretprobe_trampoline;
 413
 414         INIT_HLIST_HEAD(&empty_rp);
 415         spin_lock_irqsave(&kretprobe_lock, flags);
 416         head = kretprobe_inst_table_head(current);
 417         /* fixup registers */
 418         regs->cs = __KERNEL_CS | get_kernel_rpl();
 419         regs->ip = trampoline_address;
 420         regs->orig_ax = 0xffffffff;
 421
 422         /*
 423          * It is possible to have multiple instances associated with a given
 424          * task either because an multiple functions in the call path
 425          * have a return probe installed on them, and/or more then one return
 426          * return probe was registered for a target function.
 427          *
 428          * We can handle this because:
 429          *     - instances are always inserted at the head of the list
 430          *     - when multiple return probes are registered for the same
 431          *       function, the first instance's ret_addr will point to the
 432          *       real return address, and all the rest will point to
 433          *       kretprobe_trampoline
 434          */
 435         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
 436                 if (ri->task != current)
 437                         /* another task is sharing our hash bucket */
 438                         continue;
 439
 440                 if (ri->rp && ri->rp->handler){
 441                         __get_cpu_var(current_kprobe) = &ri->rp->kp;
 442                         get_kprobe_ctlblk()->kprobe_status = KPROBE_HIT_ACTIVE;
 443                         ri->rp->handler(ri, regs);
 444                         __get_cpu_var(current_kprobe) = NULL;
 445                 }
 446
 447                 orig_ret_address = (unsigned long)ri->ret_addr;
 448                 recycle_rp_inst(ri, &empty_rp);
 449
 450                 if (orig_ret_address != trampoline_address)
 451                         /*
 452                          * This is the real return address. Any other
 453                          * instances associated with this task are for
 454                          * other calls deeper on the call stack
 455                          */
 456                         break;
 457         }
 458
 459         kretprobe_assert(ri, orig_ret_address, trampoline_address);
 460         spin_unlock_irqrestore(&kretprobe_lock, flags);
 461
 462         hlist_for_each_entry_safe(ri, node, tmp, &empty_rp, hlist) {
 463                 hlist_del(&ri->hlist);
 464                 kfree(ri);
 465         }
 466         return (void*)orig_ret_address;
 467 }
 468
 469 /*
 470  * Called after single-stepping.  p->addr is the address of the
 471  * instruction whose first byte has been replaced by the "int 3"
 472  * instruction.  To avoid the SMP problems that can occur when we
 473  * temporarily put back the original opcode to single-step, we
 474  * single-stepped a copy of the instruction.  The address of this
 475  * copy is p->ainsn.insn.
 476  *
 477  * This function prepares to return from the post-single-step
 478  * interrupt.  We have to fix up the stack as follows:
 479  *
 480  * 0) Except in the case of absolute or indirect jump or call instructions,
 481  * the new ip is relative to the copied instruction.  We need to make
 482  * it relative to the original instruction.
 483  *
 484  * 1) If the single-stepped instruction was pushfl, then the TF and IF
 485  * flags are set in the just-pushed flags, and may need to be cleared.
 486  *
 487  * 2) If the single-stepped instruction was a call, the return address
 488  * that is atop the stack is the address following the copied instruction.
 489  * We need to make it the address following the original instruction.
 490  *
 491  * This function also checks instruction size for preparing direct execution.
 492  */
 493 static void __kprobes resume_execution(struct kprobe *p,
 494                 struct pt_regs *regs, struct kprobe_ctlblk *kcb)
 495 {
 496         unsigned long *tos = (unsigned long *)&regs->sp;
 497         unsigned long copy_eip = (unsigned long)p->ainsn.insn;
 498         unsigned long orig_eip = (unsigned long)p->addr;
 499
 500         regs->flags &= ~TF_MASK;
 501         switch (p->ainsn.insn[0]) {
 502         case 0x9c:              /* pushfl */
 503                 *tos &= ~(TF_MASK | IF_MASK);
 504                 *tos |= kcb->kprobe_old_eflags;
 505                 break;
 506         case 0xc2:              /* iret/ret/lret */
 507         case 0xc3:
 508         case 0xca:
 509         case 0xcb:
 510         case 0xcf:
 511         case 0xea:              /* jmp absolute -- ip is correct */
 512                 /* ip is already adjusted, no more changes required */
 513                 p->ainsn.boostable = 1;
 514                 goto no_change;
 515         case 0xe8:              /* call relative - Fix return addr */
 516                 *tos = orig_eip + (*tos - copy_eip);
 517                 break;
 518         case 0x9a:              /* call absolute -- same as call absolute, indirect */
 519                 *tos = orig_eip + (*tos - copy_eip);
 520                 goto no_change;
 521         case 0xff:
 522                 if ((p->ainsn.insn[1] & 0x30) == 0x10) {
 523                         /*
 524                          * call absolute, indirect
 525                          * Fix return addr; ip is correct.
 526                          * But this is not boostable
 527                          */
 528                         *tos = orig_eip + (*tos - copy_eip);
 529                         goto no_change;
 530                 } else if (((p->ainsn.insn[1] & 0x31) == 0x20) ||       /* jmp near, absolute indirect */
 531                            ((p->ainsn.insn[1] & 0x31) == 0x21)) {       /* jmp far, absolute indirect */
 532                         /* ip is correct. And this is boostable */
 533                         p->ainsn.boostable = 1;
 534                         goto no_change;
 535                 }
 536         default:
 537                 break;
 538         }
 539
 540         if (p->ainsn.boostable == 0) {
 541                 if ((regs->ip > copy_eip) &&
 542                     (regs->ip - copy_eip) + 5 < MAX_INSN_SIZE) {
 543                         /*
 544                          * These instructions can be executed directly if it
 545                          * jumps back to correct address.
 546                          */
 547                         set_jmp_op((void *)regs->ip,
 548                                    (void *)orig_eip + (regs->ip - copy_eip));
 549                         p->ainsn.boostable = 1;
 550                 } else {
 551                         p->ainsn.boostable = -1;
 552                 }
 553         }
 554
 555         regs->ip = orig_eip + (regs->ip - copy_eip);
 556
 557 no_change:
 558         restore_btf();
 559
 560         return;
 561 }
 562
 563 /*
 564  * Interrupts are disabled on entry as trap1 is an interrupt gate and they
 565  * remain disabled thoroughout this function.
 566  */
 567 static int __kprobes post_kprobe_handler(struct pt_regs *regs)
 568 {
 569         struct kprobe *cur = kprobe_running();
 570         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 571
 572         if (!cur)
 573                 return 0;
 574
 575         if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
 576                 kcb->kprobe_status = KPROBE_HIT_SSDONE;
 577                 cur->post_handler(cur, regs, 0);
 578         }
 579
 580         resume_execution(cur, regs, kcb);
 581         regs->flags |= kcb->kprobe_saved_eflags;
 582         trace_hardirqs_fixup_flags(regs->flags);
 583
 584         /*Restore back the original saved kprobes variables and continue. */
 585         if (kcb->kprobe_status == KPROBE_REENTER) {
 586                 restore_previous_kprobe(kcb);
 587                 goto out;
 588         }
 589         reset_current_kprobe();
 590 out:
 591         preempt_enable_no_resched();
 592
 593         /*
 594          * if somebody else is singlestepping across a probe point, flags
 595          * will have TF set, in which case, continue the remaining processing
 596          * of do_debug, as if this is not a probe hit.
 597          */
 598         if (regs->flags & TF_MASK)
 599                 return 0;
 600
 601         return 1;
 602 }
 603
 604 int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
 605 {
 606         struct kprobe *cur = kprobe_running();
 607         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 608
 609         switch(kcb->kprobe_status) {
 610         case KPROBE_HIT_SS:
 611         case KPROBE_REENTER:
 612                 /*
 613                  * We are here because the instruction being single
 614                  * stepped caused a page fault. We reset the current
 615                  * kprobe and the ip points back to the probe address
 616                  * and allow the page fault handler to continue as a
 617                  * normal page fault.
 618                  */
 619                 regs->ip = (unsigned long)cur->addr;
 620                 regs->flags |= kcb->kprobe_old_eflags;
 621                 if (kcb->kprobe_status == KPROBE_REENTER)
 622                         restore_previous_kprobe(kcb);
 623                 else
 624                         reset_current_kprobe();
 625                 preempt_enable_no_resched();
 626                 break;
 627         case KPROBE_HIT_ACTIVE:
 628         case KPROBE_HIT_SSDONE:
 629                 /*
 630                  * We increment the nmissed count for accounting,
 631                  * we can also use npre/npostfault count for accouting
 632                  * these specific fault cases.
 633                  */
 634                 kprobes_inc_nmissed_count(cur);
 635
 636                 /*
 637                  * We come here because instructions in the pre/post
 638                  * handler caused the page_fault, this could happen
 639                  * if handler tries to access user space by
 640                  * copy_from_user(), get_user() etc. Let the
 641                  * user-specified handler try to fix it first.
 642                  */
 643                 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
 644                         return 1;
 645
 646                 /*
 647                  * In case the user-specified fault handler returned
 648                  * zero, try to fix up.
 649                  */
 650                 if (fixup_exception(regs))
 651                         return 1;
 652
 653                 /*
 654                  * fixup_exception() could not handle it,
 655                  * Let do_page_fault() fix it.
 656                  */
 657                 break;
 658         default:
 659                 break;
 660         }
 661         return 0;
 662 }
 663
 664 /*
 665  * Wrapper routine to for handling exceptions.
 666  */
 667 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
 668                                        unsigned long val, void *data)
 669 {
 670         struct die_args *args = (struct die_args *)data;
 671         int ret = NOTIFY_DONE;
 672
 673         if (args->regs && user_mode_vm(args->regs))
 674                 return ret;
 675
 676         switch (val) {
 677         case DIE_INT3:
 678                 if (kprobe_handler(args->regs))
 679                         ret = NOTIFY_STOP;
 680                 break;
 681         case DIE_DEBUG:
 682                 if (post_kprobe_handler(args->regs))
 683                         ret = NOTIFY_STOP;
 684                 break;
 685         case DIE_GPF:
 686                 /* kprobe_running() needs smp_processor_id() */
 687                 preempt_disable();
 688                 if (kprobe_running() &&
 689                     kprobe_fault_handler(args->regs, args->trapnr))
 690                         ret = NOTIFY_STOP;
 691                 preempt_enable();
 692                 break;
 693         default:
 694                 break;
 695         }
 696         return ret;
 697 }
 698
 699 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
 700 {
 701         struct jprobe *jp = container_of(p, struct jprobe, kp);
 702         unsigned long addr;
 703         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 704
 705         kcb->jprobe_saved_regs = *regs;
 706         kcb->jprobe_saved_esp = &regs->sp;
 707         addr = (unsigned long)(kcb->jprobe_saved_esp);
 708
 709         /*
 710          * TBD: As Linus pointed out, gcc assumes that the callee
 711          * owns the argument space and could overwrite it, e.g.
 712          * tailcall optimization. So, to be absolutely safe
 713          * we also save and restore enough stack bytes to cover
 714          * the argument area.
 715          */
 716         memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
 717                         MIN_STACK_SIZE(addr));
 718         regs->flags &= ~IF_MASK;
 719         trace_hardirqs_off();
 720         regs->ip = (unsigned long)(jp->entry);
 721         return 1;
 722 }
 723
 724 void __kprobes jprobe_return(void)
 725 {
 726         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 727
 728         asm volatile ("       xchgl   %%ebx,%%esp     \n"
 729                       "       int3                      \n"
 730                       "       .globl jprobe_return_end  \n"
 731                       "       jprobe_return_end:        \n"
 732                       "       nop                       \n"::"b"
 733                       (kcb->jprobe_saved_esp):"memory");
 734 }
 735
 736 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
 737 {
 738         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 739         u8 *addr = (u8 *) (regs->ip - 1);
 740         unsigned long stack_addr = (unsigned long)(kcb->jprobe_saved_esp);
 741         struct jprobe *jp = container_of(p, struct jprobe, kp);
 742
 743         if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
 744                 if (&regs->sp != kcb->jprobe_saved_esp) {
 745                         struct pt_regs *saved_regs = &kcb->jprobe_saved_regs;
 746                         printk("current sp %p does not match saved sp %p\n",
 747                                &regs->sp, kcb->jprobe_saved_esp);
 748                         printk("Saved registers for jprobe %p\n", jp);
 749                         show_registers(saved_regs);
 750                         printk("Current registers\n");
 751                         show_registers(regs);
 752                         BUG();
 753                 }
 754                 *regs = kcb->jprobe_saved_regs;
 755                 memcpy((kprobe_opcode_t *) stack_addr, kcb->jprobes_stack,
 756                        MIN_STACK_SIZE(stack_addr));
 757                 preempt_enable_no_resched();
 758                 return 1;
 759         }
 760         return 0;
 761 }
 762
 763 int __kprobes arch_trampoline_kprobe(struct kprobe *p)
 764 {
 765         return 0;
 766 }
 767
 768 int __init arch_init_kprobes(void)
 769 {
 770         return 0;
 771 }