git.oblomov.eu Git - linux-2.6/blob - 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->eflags & (TF_MASK | IF_MASK));
 216         if (is_IF_modifier(p->opcode))
 217                 kcb->kprobe_saved_eflags &= ~IF_MASK;
 218 }
 219
 220 static void __kprobes prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
 221 {
 222         regs->eflags |= TF_MASK;
 223         regs->eflags &= ~IF_MASK;
 224         /*single step inline if the instruction is an int3*/
 225         if (p->opcode == BREAKPOINT_INSTRUCTION)
 226                 regs->eip = (unsigned long)p->addr;
 227         else
 228                 regs->eip = (unsigned long)p->ainsn.insn;
 229 }
 230
 231 /* Called with kretprobe_lock held */
 232 void __kprobes arch_prepare_kretprobe(struct kretprobe_instance *ri,
 233                                       struct pt_regs *regs)
 234 {
 235         unsigned long *sara = (unsigned long *)&regs->esp;
 236
 237         ri->ret_addr = (kprobe_opcode_t *) *sara;
 238
 239         /* Replace the return addr with trampoline addr */
 240         *sara = (unsigned long) &kretprobe_trampoline;
 241 }
 242
 243 /*
 244  * Interrupts are disabled on entry as trap3 is an interrupt gate and they
 245  * remain disabled thorough out this function.
 246  */
 247 static int __kprobes kprobe_handler(struct pt_regs *regs)
 248 {
 249         struct kprobe *p;
 250         int ret = 0;
 251         kprobe_opcode_t *addr;
 252         struct kprobe_ctlblk *kcb;
 253
 254         addr = (kprobe_opcode_t *)(regs->eip - sizeof(kprobe_opcode_t));
 255
 256         /*
 257          * We don't want to be preempted for the entire
 258          * duration of kprobe processing
 259          */
 260         preempt_disable();
 261         kcb = get_kprobe_ctlblk();
 262
 263         /* Check we're not actually recursing */
 264         if (kprobe_running()) {
 265                 p = get_kprobe(addr);
 266                 if (p) {
 267                         if (kcb->kprobe_status == KPROBE_HIT_SS &&
 268                                 *p->ainsn.insn == BREAKPOINT_INSTRUCTION) {
 269                                 regs->eflags &= ~TF_MASK;
 270                                 regs->eflags |= kcb->kprobe_saved_eflags;
 271                                 goto no_kprobe;
 272                         }
 273                         /* We have reentered the kprobe_handler(), since
 274                          * another probe was hit while within the handler.
 275                          * We here save the original kprobes variables and
 276                          * just single step on the instruction of the new probe
 277                          * without calling any user handlers.
 278                          */
 279                         save_previous_kprobe(kcb);
 280                         set_current_kprobe(p, regs, kcb);
 281                         kprobes_inc_nmissed_count(p);
 282                         prepare_singlestep(p, regs);
 283                         kcb->kprobe_status = KPROBE_REENTER;
 284                         return 1;
 285                 } else {
 286                         if (*addr != BREAKPOINT_INSTRUCTION) {
 287                         /* The breakpoint instruction was removed by
 288                          * another cpu right after we hit, no further
 289                          * handling of this interrupt is appropriate
 290                          */
 291                                 regs->eip -= sizeof(kprobe_opcode_t);
 292                                 ret = 1;
 293                                 goto no_kprobe;
 294                         }
 295                         p = __get_cpu_var(current_kprobe);
 296                         if (p->break_handler && p->break_handler(p, regs)) {
 297                                 goto ss_probe;
 298                         }
 299                 }
 300                 goto no_kprobe;
 301         }
 302
 303         p = get_kprobe(addr);
 304         if (!p) {
 305                 if (*addr != BREAKPOINT_INSTRUCTION) {
 306                         /*
 307                          * The breakpoint instruction was removed right
 308                          * after we hit it.  Another cpu has removed
 309                          * either a probepoint or a debugger breakpoint
 310                          * at this address.  In either case, no further
 311                          * handling of this interrupt is appropriate.
 312                          * Back up over the (now missing) int3 and run
 313                          * the original instruction.
 314                          */
 315                         regs->eip -= sizeof(kprobe_opcode_t);
 316                         ret = 1;
 317                 }
 318                 /* Not one of ours: let kernel handle it */
 319                 goto no_kprobe;
 320         }
 321
 322         set_current_kprobe(p, regs, kcb);
 323         kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 324
 325         if (p->pre_handler && p->pre_handler(p, regs))
 326                 /* handler has already set things up, so skip ss setup */
 327                 return 1;
 328
 329 ss_probe:
 330 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PM)
 331         if (p->ainsn.boostable == 1 && !p->post_handler){
 332                 /* Boost up -- we can execute copied instructions directly */
 333                 reset_current_kprobe();
 334                 regs->eip = (unsigned long)p->ainsn.insn;
 335                 preempt_enable_no_resched();
 336                 return 1;
 337         }
 338 #endif
 339         prepare_singlestep(p, regs);
 340         kcb->kprobe_status = KPROBE_HIT_SS;
 341         return 1;
 342
 343 no_kprobe:
 344         preempt_enable_no_resched();
 345         return ret;
 346 }
 347
 348 /*
 349  * For function-return probes, init_kprobes() establishes a probepoint
 350  * here. When a retprobed function returns, this probe is hit and
 351  * trampoline_probe_handler() runs, calling the kretprobe's handler.
 352  */
 353  void __kprobes kretprobe_trampoline_holder(void)
 354  {
 355         asm volatile ( ".global kretprobe_trampoline\n"
 356                         "kretprobe_trampoline: \n"
 357                         "       pushf\n"
 358                         /* skip cs, eip, orig_eax */
 359                         "       subl $12, %esp\n"
 360                         "       pushl %fs\n"
 361                         "       pushl %ds\n"
 362                         "       pushl %es\n"
 363                         "       pushl %eax\n"
 364                         "       pushl %ebp\n"
 365                         "       pushl %edi\n"
 366                         "       pushl %esi\n"
 367                         "       pushl %edx\n"
 368                         "       pushl %ecx\n"
 369                         "       pushl %ebx\n"
 370                         "       movl %esp, %eax\n"
 371                         "       call trampoline_handler\n"
 372                         /* move eflags to cs */
 373                         "       movl 52(%esp), %edx\n"
 374                         "       movl %edx, 48(%esp)\n"
 375                         /* save true return address on eflags */
 376                         "       movl %eax, 52(%esp)\n"
 377                         "       popl %ebx\n"
 378                         "       popl %ecx\n"
 379                         "       popl %edx\n"
 380                         "       popl %esi\n"
 381                         "       popl %edi\n"
 382                         "       popl %ebp\n"
 383                         "       popl %eax\n"
 384                         /* skip eip, orig_eax, es, ds, fs */
 385                         "       addl $20, %esp\n"
 386                         "       popf\n"
 387                         "       ret\n");
 388 }
 389
 390 /*
 391  * Called from kretprobe_trampoline
 392  */
 393 fastcall void *__kprobes trampoline_handler(struct pt_regs *regs)
 394 {
 395         struct kretprobe_instance *ri = NULL;
 396         struct hlist_head *head, empty_rp;
 397         struct hlist_node *node, *tmp;
 398         unsigned long flags, orig_ret_address = 0;
 399         unsigned long trampoline_address =(unsigned long)&kretprobe_trampoline;
 400
 401         INIT_HLIST_HEAD(&empty_rp);
 402         spin_lock_irqsave(&kretprobe_lock, flags);
 403         head = kretprobe_inst_table_head(current);
 404         /* fixup registers */
 405         regs->xcs = __KERNEL_CS | get_kernel_rpl();
 406         regs->eip = trampoline_address;
 407         regs->orig_eax = 0xffffffff;
 408
 409         /*
 410          * It is possible to have multiple instances associated with a given
 411          * task either because an multiple functions in the call path
 412          * have a return probe installed on them, and/or more then one return
 413          * return probe was registered for a target function.
 414          *
 415          * We can handle this because:
 416          *     - instances are always inserted at the head of the list
 417          *     - when multiple return probes are registered for the same
 418          *       function, the first instance's ret_addr will point to the
 419          *       real return address, and all the rest will point to
 420          *       kretprobe_trampoline
 421          */
 422         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
 423                 if (ri->task != current)
 424                         /* another task is sharing our hash bucket */
 425                         continue;
 426
 427                 if (ri->rp && ri->rp->handler){
 428                         __get_cpu_var(current_kprobe) = &ri->rp->kp;
 429                         get_kprobe_ctlblk()->kprobe_status = KPROBE_HIT_ACTIVE;
 430                         ri->rp->handler(ri, regs);
 431                         __get_cpu_var(current_kprobe) = NULL;
 432                 }
 433
 434                 orig_ret_address = (unsigned long)ri->ret_addr;
 435                 recycle_rp_inst(ri, &empty_rp);
 436
 437                 if (orig_ret_address != trampoline_address)
 438                         /*
 439                          * This is the real return address. Any other
 440                          * instances associated with this task are for
 441                          * other calls deeper on the call stack
 442                          */
 443                         break;
 444         }
 445
 446         kretprobe_assert(ri, orig_ret_address, trampoline_address);
 447         spin_unlock_irqrestore(&kretprobe_lock, flags);
 448
 449         hlist_for_each_entry_safe(ri, node, tmp, &empty_rp, hlist) {
 450                 hlist_del(&ri->hlist);
 451                 kfree(ri);
 452         }
 453         return (void*)orig_ret_address;
 454 }
 455
 456 /*
 457  * Called after single-stepping.  p->addr is the address of the
 458  * instruction whose first byte has been replaced by the "int 3"
 459  * instruction.  To avoid the SMP problems that can occur when we
 460  * temporarily put back the original opcode to single-step, we
 461  * single-stepped a copy of the instruction.  The address of this
 462  * copy is p->ainsn.insn.
 463  *
 464  * This function prepares to return from the post-single-step
 465  * interrupt.  We have to fix up the stack as follows:
 466  *
 467  * 0) Except in the case of absolute or indirect jump or call instructions,
 468  * the new eip is relative to the copied instruction.  We need to make
 469  * it relative to the original instruction.
 470  *
 471  * 1) If the single-stepped instruction was pushfl, then the TF and IF
 472  * flags are set in the just-pushed eflags, and may need to be cleared.
 473  *
 474  * 2) If the single-stepped instruction was a call, the return address
 475  * that is atop the stack is the address following the copied instruction.
 476  * We need to make it the address following the original instruction.
 477  *
 478  * This function also checks instruction size for preparing direct execution.
 479  */
 480 static void __kprobes resume_execution(struct kprobe *p,
 481                 struct pt_regs *regs, struct kprobe_ctlblk *kcb)
 482 {
 483         unsigned long *tos = (unsigned long *)&regs->esp;
 484         unsigned long copy_eip = (unsigned long)p->ainsn.insn;
 485         unsigned long orig_eip = (unsigned long)p->addr;
 486
 487         regs->eflags &= ~TF_MASK;
 488         switch (p->ainsn.insn[0]) {
 489         case 0x9c:              /* pushfl */
 490                 *tos &= ~(TF_MASK | IF_MASK);
 491                 *tos |= kcb->kprobe_old_eflags;
 492                 break;
 493         case 0xc2:              /* iret/ret/lret */
 494         case 0xc3:
 495         case 0xca:
 496         case 0xcb:
 497         case 0xcf:
 498         case 0xea:              /* jmp absolute -- eip is correct */
 499                 /* eip is already adjusted, no more changes required */
 500                 p->ainsn.boostable = 1;
 501                 goto no_change;
 502         case 0xe8:              /* call relative - Fix return addr */
 503                 *tos = orig_eip + (*tos - copy_eip);
 504                 break;
 505         case 0x9a:              /* call absolute -- same as call absolute, indirect */
 506                 *tos = orig_eip + (*tos - copy_eip);
 507                 goto no_change;
 508         case 0xff:
 509                 if ((p->ainsn.insn[1] & 0x30) == 0x10) {
 510                         /*
 511                          * call absolute, indirect
 512                          * Fix return addr; eip is correct.
 513                          * But this is not boostable
 514                          */
 515                         *tos = orig_eip + (*tos - copy_eip);
 516                         goto no_change;
 517                 } else if (((p->ainsn.insn[1] & 0x31) == 0x20) ||       /* jmp near, absolute indirect */
 518                            ((p->ainsn.insn[1] & 0x31) == 0x21)) {       /* jmp far, absolute indirect */
 519                         /* eip is correct. And this is boostable */
 520                         p->ainsn.boostable = 1;
 521                         goto no_change;
 522                 }
 523         default:
 524                 break;
 525         }
 526
 527         if (p->ainsn.boostable == 0) {
 528                 if ((regs->eip > copy_eip) &&
 529                     (regs->eip - copy_eip) + 5 < MAX_INSN_SIZE) {
 530                         /*
 531                          * These instructions can be executed directly if it
 532                          * jumps back to correct address.
 533                          */
 534                         set_jmp_op((void *)regs->eip,
 535                                    (void *)orig_eip + (regs->eip - copy_eip));
 536                         p->ainsn.boostable = 1;
 537                 } else {
 538                         p->ainsn.boostable = -1;
 539                 }
 540         }
 541
 542         regs->eip = orig_eip + (regs->eip - copy_eip);
 543
 544 no_change:
 545         return;
 546 }
 547
 548 /*
 549  * Interrupts are disabled on entry as trap1 is an interrupt gate and they
 550  * remain disabled thoroughout this function.
 551  */
 552 static int __kprobes post_kprobe_handler(struct pt_regs *regs)
 553 {
 554         struct kprobe *cur = kprobe_running();
 555         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 556
 557         if (!cur)
 558                 return 0;
 559
 560         if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
 561                 kcb->kprobe_status = KPROBE_HIT_SSDONE;
 562                 cur->post_handler(cur, regs, 0);
 563         }
 564
 565         resume_execution(cur, regs, kcb);
 566         regs->eflags |= kcb->kprobe_saved_eflags;
 567 #ifdef CONFIG_TRACE_IRQFLAGS_SUPPORT
 568         if (raw_irqs_disabled_flags(regs->eflags))
 569                 trace_hardirqs_off();
 570         else
 571                 trace_hardirqs_on();
 572 #endif
 573
 574         /*Restore back the original saved kprobes variables and continue. */
 575         if (kcb->kprobe_status == KPROBE_REENTER) {
 576                 restore_previous_kprobe(kcb);
 577                 goto out;
 578         }
 579         reset_current_kprobe();
 580 out:
 581         preempt_enable_no_resched();
 582
 583         /*
 584          * if somebody else is singlestepping across a probe point, eflags
 585          * will have TF set, in which case, continue the remaining processing
 586          * of do_debug, as if this is not a probe hit.
 587          */
 588         if (regs->eflags & TF_MASK)
 589                 return 0;
 590
 591         return 1;
 592 }
 593
 594 int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
 595 {
 596         struct kprobe *cur = kprobe_running();
 597         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 598
 599         switch(kcb->kprobe_status) {
 600         case KPROBE_HIT_SS:
 601         case KPROBE_REENTER:
 602                 /*
 603                  * We are here because the instruction being single
 604                  * stepped caused a page fault. We reset the current
 605                  * kprobe and the eip points back to the probe address
 606                  * and allow the page fault handler to continue as a
 607                  * normal page fault.
 608                  */
 609                 regs->eip = (unsigned long)cur->addr;
 610                 regs->eflags |= kcb->kprobe_old_eflags;
 611                 if (kcb->kprobe_status == KPROBE_REENTER)
 612                         restore_previous_kprobe(kcb);
 613                 else
 614                         reset_current_kprobe();
 615                 preempt_enable_no_resched();
 616                 break;
 617         case KPROBE_HIT_ACTIVE:
 618         case KPROBE_HIT_SSDONE:
 619                 /*
 620                  * We increment the nmissed count for accounting,
 621                  * we can also use npre/npostfault count for accouting
 622                  * these specific fault cases.
 623                  */
 624                 kprobes_inc_nmissed_count(cur);
 625
 626                 /*
 627                  * We come here because instructions in the pre/post
 628                  * handler caused the page_fault, this could happen
 629                  * if handler tries to access user space by
 630                  * copy_from_user(), get_user() etc. Let the
 631                  * user-specified handler try to fix it first.
 632                  */
 633                 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
 634                         return 1;
 635
 636                 /*
 637                  * In case the user-specified fault handler returned
 638                  * zero, try to fix up.
 639                  */
 640                 if (fixup_exception(regs))
 641                         return 1;
 642
 643                 /*
 644                  * fixup_exception() could not handle it,
 645                  * Let do_page_fault() fix it.
 646                  */
 647                 break;
 648         default:
 649                 break;
 650         }
 651         return 0;
 652 }
 653
 654 /*
 655  * Wrapper routine to for handling exceptions.
 656  */
 657 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
 658                                        unsigned long val, void *data)
 659 {
 660         struct die_args *args = (struct die_args *)data;
 661         int ret = NOTIFY_DONE;
 662
 663         if (args->regs && user_mode_vm(args->regs))
 664                 return ret;
 665
 666         switch (val) {
 667         case DIE_INT3:
 668                 if (kprobe_handler(args->regs))
 669                         ret = NOTIFY_STOP;
 670                 break;
 671         case DIE_DEBUG:
 672                 if (post_kprobe_handler(args->regs))
 673                         ret = NOTIFY_STOP;
 674                 break;
 675         case DIE_GPF:
 676                 /* kprobe_running() needs smp_processor_id() */
 677                 preempt_disable();
 678                 if (kprobe_running() &&
 679                     kprobe_fault_handler(args->regs, args->trapnr))
 680                         ret = NOTIFY_STOP;
 681                 preempt_enable();
 682                 break;
 683         default:
 684                 break;
 685         }
 686         return ret;
 687 }
 688
 689 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
 690 {
 691         struct jprobe *jp = container_of(p, struct jprobe, kp);
 692         unsigned long addr;
 693         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 694
 695         kcb->jprobe_saved_regs = *regs;
 696         kcb->jprobe_saved_esp = &regs->esp;
 697         addr = (unsigned long)(kcb->jprobe_saved_esp);
 698
 699         /*
 700          * TBD: As Linus pointed out, gcc assumes that the callee
 701          * owns the argument space and could overwrite it, e.g.
 702          * tailcall optimization. So, to be absolutely safe
 703          * we also save and restore enough stack bytes to cover
 704          * the argument area.
 705          */
 706         memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
 707                         MIN_STACK_SIZE(addr));
 708         regs->eflags &= ~IF_MASK;
 709         trace_hardirqs_off();
 710         regs->eip = (unsigned long)(jp->entry);
 711         return 1;
 712 }
 713
 714 void __kprobes jprobe_return(void)
 715 {
 716         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 717
 718         asm volatile ("       xchgl   %%ebx,%%esp     \n"
 719                       "       int3                      \n"
 720                       "       .globl jprobe_return_end  \n"
 721                       "       jprobe_return_end:        \n"
 722                       "       nop                       \n"::"b"
 723                       (kcb->jprobe_saved_esp):"memory");
 724 }
 725
 726 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
 727 {
 728         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 729         u8 *addr = (u8 *) (regs->eip - 1);
 730         unsigned long stack_addr = (unsigned long)(kcb->jprobe_saved_esp);
 731         struct jprobe *jp = container_of(p, struct jprobe, kp);
 732
 733         if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
 734                 if (&regs->esp != kcb->jprobe_saved_esp) {
 735                         struct pt_regs *saved_regs =
 736                             container_of(kcb->jprobe_saved_esp,
 737                                             struct pt_regs, esp);
 738                         printk("current esp %p does not match saved esp %p\n",
 739                                &regs->esp, kcb->jprobe_saved_esp);
 740                         printk("Saved registers for jprobe %p\n", jp);
 741                         show_registers(saved_regs);
 742                         printk("Current registers\n");
 743                         show_registers(regs);
 744                         BUG();
 745                 }
 746                 *regs = kcb->jprobe_saved_regs;
 747                 memcpy((kprobe_opcode_t *) stack_addr, kcb->jprobes_stack,
 748                        MIN_STACK_SIZE(stack_addr));
 749                 preempt_enable_no_resched();
 750                 return 1;
 751         }
 752         return 0;
 753 }
 754
 755 int __kprobes arch_trampoline_kprobe(struct kprobe *p)
 756 {
 757         return 0;
 758 }
 759
 760 int __init arch_init_kprobes(void)
 761 {
 762         return 0;
 763 }