arch/x86/kernel/kprobes.c

   1 /*
   2  *  Kernel Probes (KProbes)
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License as published by
   6  * the Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write to the Free Software
  16  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  17  *
  18  * Copyright (C) IBM Corporation, 2002, 2004
  19  *
  20  * 2002-Oct     Created by Vamsi Krishna S <vamsi_krishna@in.ibm.com> Kernel
  21  *              Probes initial implementation ( includes contributions from
  22  *              Rusty Russell).
  23  * 2004-July    Suparna Bhattacharya <suparna@in.ibm.com> added jumper probes
  24  *              interface to access function arguments.
  25  * 2004-Oct     Jim Keniston <jkenisto@us.ibm.com> and Prasanna S Panchamukhi
  26  *              <prasanna@in.ibm.com> adapted for x86_64 from i386.
  27  * 2005-Mar     Roland McGrath <roland@redhat.com>
  28  *              Fixed to handle %rip-relative addressing mode correctly.
  29  * 2005-May     Hien Nguyen <hien@us.ibm.com>, Jim Keniston
  30  *              <jkenisto@us.ibm.com> and Prasanna S Panchamukhi
  31  *              <prasanna@in.ibm.com> added function-return probes.
  32  * 2005-May     Rusty Lynch <rusty.lynch@intel.com>
  33  *              Added function return probes functionality
  34  * 2006-Feb     Masami Hiramatsu <hiramatu@sdl.hitachi.co.jp> added
  35  *              kprobe-booster and kretprobe-booster for i386.
  36  * 2007-Dec     Masami Hiramatsu <mhiramat@redhat.com> added kprobe-booster
  37  *              and kretprobe-booster for x86-64
  38  * 2007-Dec     Masami Hiramatsu <mhiramat@redhat.com>, Arjan van de Ven
  39  *              <arjan@infradead.org> and Jim Keniston <jkenisto@us.ibm.com>
  40  *              unified x86 kprobes code.
  41  */
  42
  43 #include <linux/kprobes.h>
  44 #include <linux/ptrace.h>
  45 #include <linux/string.h>
  46 #include <linux/slab.h>
  47 #include <linux/hardirq.h>
  48 #include <linux/preempt.h>
  49 #include <linux/module.h>
  50 #include <linux/kdebug.h>
  51 #include <linux/kallsyms.h>
  52 #include <linux/ftrace.h>
  53
  54 #include <asm/cacheflush.h>
  55 #include <asm/desc.h>
  56 #include <asm/pgtable.h>
  57 #include <asm/uaccess.h>
  58 #include <asm/alternative.h>
  59 #include <asm/insn.h>
  60 #include <asm/debugreg.h>
  61
  62 void jprobe_return_end(void);
  63
  64 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
  65 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
  66
  67 #define stack_addr(regs) ((unsigned long *)kernel_stack_pointer(regs))
  68
  69 #define W(row, b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, ba, bb, bc, bd, be, bf)\
  70         (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
  71           (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
  72           (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
  73           (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
  74          << (row % 32))
  75         /*
  76          * Undefined/reserved opcodes, conditional jump, Opcode Extension
  77          * Groups, and some special opcodes can not boost.
  78          */
  79 static const u32 twobyte_is_boostable[256 / 32] = {
  80         /*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f          */
  81         /*      ----------------------------------------------          */
  82         W(0x00, 0, 0, 1, 1, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0) | /* 00 */
  83         W(0x10, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 10 */
  84         W(0x20, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 20 */
  85         W(0x30, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 30 */
  86         W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
  87         W(0x50, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 50 */
  88         W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1) | /* 60 */
  89         W(0x70, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1) , /* 70 */
  90         W(0x80, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 80 */
  91         W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
  92         W(0xa0, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* a0 */
  93         W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1) , /* b0 */
  94         W(0xc0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1) | /* c0 */
  95         W(0xd0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) , /* d0 */
  96         W(0xe0, 0, 1, 1, 0, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* e0 */
  97         W(0xf0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 0)   /* f0 */
  98         /*      -----------------------------------------------         */
  99         /*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f          */
 100 };
 101 #undef W
 102
 103 struct kretprobe_blackpoint kretprobe_blacklist[] = {
 104         {"__switch_to", }, /* This function switches only current task, but
 105                               doesn't switch kernel stack.*/
 106         {NULL, NULL}    /* Terminator */
 107 };
 108 const int kretprobe_blacklist_size = ARRAY_SIZE(kretprobe_blacklist);
 109
 110 static void __kprobes __synthesize_relative_insn(void *from, void *to, u8 op)
 111 {
 112         struct __arch_relative_insn {
 113                 u8 op;
 114                 s32 raddr;
 115         } __attribute__((packed)) *insn;
 116
 117         insn = (struct __arch_relative_insn *)from;
 118         insn->raddr = (s32)((long)(to) - ((long)(from) + 5));
 119         insn->op = op;
 120 }
 121
 122 /* Insert a jump instruction at address 'from', which jumps to address 'to'.*/
 123 static void __kprobes synthesize_reljump(void *from, void *to)
 124 {
 125         __synthesize_relative_insn(from, to, RELATIVEJUMP_OPCODE);
 126 }
 127
 128 /*
 129  * Skip the prefixes of the instruction.
 130  */
 131 static kprobe_opcode_t *__kprobes skip_prefixes(kprobe_opcode_t *insn)
 132 {
 133         insn_attr_t attr;
 134
 135         attr = inat_get_opcode_attribute((insn_byte_t)*insn);
 136         while (inat_is_legacy_prefix(attr)) {
 137                 insn++;
 138                 attr = inat_get_opcode_attribute((insn_byte_t)*insn);
 139         }
 140 #ifdef CONFIG_X86_64
 141         if (inat_is_rex_prefix(attr))
 142                 insn++;
 143 #endif
 144         return insn;
 145 }
 146
 147 /*
 148  * Returns non-zero if opcode is boostable.
 149  * RIP relative instructions are adjusted at copying time in 64 bits mode
 150  */
 151 static int __kprobes can_boost(kprobe_opcode_t *opcodes)
 152 {
 153         kprobe_opcode_t opcode;
 154         kprobe_opcode_t *orig_opcodes = opcodes;
 155
 156         if (search_exception_tables((unsigned long)opcodes))
 157                 return 0;       /* Page fault may occur on this address. */
 158
 159 retry:
 160         if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
 161                 return 0;
 162         opcode = *(opcodes++);
 163
 164         /* 2nd-byte opcode */
 165         if (opcode == 0x0f) {
 166                 if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
 167                         return 0;
 168                 return test_bit(*opcodes,
 169                                 (unsigned long *)twobyte_is_boostable);
 170         }
 171
 172         switch (opcode & 0xf0) {
 173 #ifdef CONFIG_X86_64
 174         case 0x40:
 175                 goto retry; /* REX prefix is boostable */
 176 #endif
 177         case 0x60:
 178                 if (0x63 < opcode && opcode < 0x67)
 179                         goto retry; /* prefixes */
 180                 /* can't boost Address-size override and bound */
 181                 return (opcode != 0x62 && opcode != 0x67);
 182         case 0x70:
 183                 return 0; /* can't boost conditional jump */
 184         case 0xc0:
 185                 /* can't boost software-interruptions */
 186                 return (0xc1 < opcode && opcode < 0xcc) || opcode == 0xcf;
 187         case 0xd0:
 188                 /* can boost AA* and XLAT */
 189                 return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
 190         case 0xe0:
 191                 /* can boost in/out and absolute jmps */
 192                 return ((opcode & 0x04) || opcode == 0xea);
 193         case 0xf0:
 194                 if ((opcode & 0x0c) == 0 && opcode != 0xf1)
 195                         goto retry; /* lock/rep(ne) prefix */
 196                 /* clear and set flags are boostable */
 197                 return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
 198         default:
 199                 /* segment override prefixes are boostable */
 200                 if (opcode == 0x26 || opcode == 0x36 || opcode == 0x3e)
 201                         goto retry; /* prefixes */
 202                 /* CS override prefix and call are not boostable */
 203                 return (opcode != 0x2e && opcode != 0x9a);
 204         }
 205 }
 206
 207 /* Recover the probed instruction at addr for further analysis. */
 208 static int recover_probed_instruction(kprobe_opcode_t *buf, unsigned long addr)
 209 {
 210         struct kprobe *kp;
 211         kp = get_kprobe((void *)addr);
 212         if (!kp)
 213                 return -EINVAL;
 214
 215         /*
 216          *  Basically, kp->ainsn.insn has an original instruction.
 217          *  However, RIP-relative instruction can not do single-stepping
 218          *  at different place, __copy_instruction() tweaks the displacement of
 219          *  that instruction. In that case, we can't recover the instruction
 220          *  from the kp->ainsn.insn.
 221          *
 222          *  On the other hand, kp->opcode has a copy of the first byte of
 223          *  the probed instruction, which is overwritten by int3. And
 224          *  the instruction at kp->addr is not modified by kprobes except
 225          *  for the first byte, we can recover the original instruction
 226          *  from it and kp->opcode.
 227          */
 228         memcpy(buf, kp->addr, MAX_INSN_SIZE * sizeof(kprobe_opcode_t));
 229         buf[0] = kp->opcode;
 230         return 0;
 231 }
 232
 233 /* Check if paddr is at an instruction boundary */
 234 static int __kprobes can_probe(unsigned long paddr)
 235 {
 236         int ret;
 237         unsigned long addr, offset = 0;
 238         struct insn insn;
 239         kprobe_opcode_t buf[MAX_INSN_SIZE];
 240
 241         if (!kallsyms_lookup_size_offset(paddr, NULL, &offset))
 242                 return 0;
 243
 244         /* Decode instructions */
 245         addr = paddr - offset;
 246         while (addr < paddr) {
 247                 kernel_insn_init(&insn, (void *)addr);
 248                 insn_get_opcode(&insn);
 249
 250                 /*
 251                  * Check if the instruction has been modified by another
 252                  * kprobe, in which case we replace the breakpoint by the
 253                  * original instruction in our buffer.
 254                  */
 255                 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION) {
 256                         ret = recover_probed_instruction(buf, addr);
 257                         if (ret)
 258                                 /*
 259                                  * Another debugging subsystem might insert
 260                                  * this breakpoint. In that case, we can't
 261                                  * recover it.
 262                                  */
 263                                 return 0;
 264                         kernel_insn_init(&insn, buf);
 265                 }
 266                 insn_get_length(&insn);
 267                 addr += insn.length;
 268         }
 269
 270         return (addr == paddr);
 271 }
 272
 273 /*
 274  * Returns non-zero if opcode modifies the interrupt flag.
 275  */
 276 static int __kprobes is_IF_modifier(kprobe_opcode_t *insn)
 277 {
 278         /* Skip prefixes */
 279         insn = skip_prefixes(insn);
 280
 281         switch (*insn) {
 282         case 0xfa:              /* cli */
 283         case 0xfb:              /* sti */
 284         case 0xcf:              /* iret/iretd */
 285         case 0x9d:              /* popf/popfd */
 286                 return 1;
 287         }
 288
 289         return 0;
 290 }
 291
 292 /*
 293  * Copy an instruction and adjust the displacement if the instruction
 294  * uses the %rip-relative addressing mode.
 295  * If it does, Return the address of the 32-bit displacement word.
 296  * If not, return null.
 297  * Only applicable to 64-bit x86.
 298  */
 299 static int __kprobes __copy_instruction(u8 *dest, u8 *src, int recover)
 300 {
 301         struct insn insn;
 302         int ret;
 303         kprobe_opcode_t buf[MAX_INSN_SIZE];
 304
 305         kernel_insn_init(&insn, src);
 306         if (recover) {
 307                 insn_get_opcode(&insn);
 308                 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION) {
 309                         ret = recover_probed_instruction(buf,
 310                                                          (unsigned long)src);
 311                         if (ret)
 312                                 return 0;
 313                         kernel_insn_init(&insn, buf);
 314                 }
 315         }
 316         insn_get_length(&insn);
 317         memcpy(dest, insn.kaddr, insn.length);
 318
 319 #ifdef CONFIG_X86_64
 320         if (insn_rip_relative(&insn)) {
 321                 s64 newdisp;
 322                 u8 *disp;
 323                 kernel_insn_init(&insn, dest);
 324                 insn_get_displacement(&insn);
 325                 /*
 326                  * The copied instruction uses the %rip-relative addressing
 327                  * mode.  Adjust the displacement for the difference between
 328                  * the original location of this instruction and the location
 329                  * of the copy that will actually be run.  The tricky bit here
 330                  * is making sure that the sign extension happens correctly in
 331                  * this calculation, since we need a signed 32-bit result to
 332                  * be sign-extended to 64 bits when it's added to the %rip
 333                  * value and yield the same 64-bit result that the sign-
 334                  * extension of the original signed 32-bit displacement would
 335                  * have given.
 336                  */
 337                 newdisp = (u8 *) src + (s64) insn.displacement.value -
 338                           (u8 *) dest;
 339                 BUG_ON((s64) (s32) newdisp != newdisp); /* Sanity check.  */
 340                 disp = (u8 *) dest + insn_offset_displacement(&insn);
 341                 *(s32 *) disp = (s32) newdisp;
 342         }
 343 #endif
 344         return insn.length;
 345 }
 346
 347 static void __kprobes arch_copy_kprobe(struct kprobe *p)
 348 {
 349         /*
 350          * Copy an instruction without recovering int3, because it will be
 351          * put by another subsystem.
 352          */
 353         __copy_instruction(p->ainsn.insn, p->addr, 0);
 354
 355         if (can_boost(p->addr))
 356                 p->ainsn.boostable = 0;
 357         else
 358                 p->ainsn.boostable = -1;
 359
 360         p->opcode = *p->addr;
 361 }
 362
 363 int __kprobes arch_prepare_kprobe(struct kprobe *p)
 364 {
 365         if (alternatives_text_reserved(p->addr, p->addr))
 366                 return -EINVAL;
 367
 368         if (!can_probe((unsigned long)p->addr))
 369                 return -EILSEQ;
 370         /* insn: must be on special executable page on x86. */
 371         p->ainsn.insn = get_insn_slot();
 372         if (!p->ainsn.insn)
 373                 return -ENOMEM;
 374         arch_copy_kprobe(p);
 375         return 0;
 376 }
 377
 378 void __kprobes arch_arm_kprobe(struct kprobe *p)
 379 {
 380         text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
 381 }
 382
 383 void __kprobes arch_disarm_kprobe(struct kprobe *p)
 384 {
 385         text_poke(p->addr, &p->opcode, 1);
 386 }
 387
 388 void __kprobes arch_remove_kprobe(struct kprobe *p)
 389 {
 390         if (p->ainsn.insn) {
 391                 free_insn_slot(p->ainsn.insn, (p->ainsn.boostable == 1));
 392                 p->ainsn.insn = NULL;
 393         }
 394 }
 395
 396 static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
 397 {
 398         kcb->prev_kprobe.kp = kprobe_running();
 399         kcb->prev_kprobe.status = kcb->kprobe_status;
 400         kcb->prev_kprobe.old_flags = kcb->kprobe_old_flags;
 401         kcb->prev_kprobe.saved_flags = kcb->kprobe_saved_flags;
 402 }
 403
 404 static void __kprobes restore_previous_kprobe(struct kprobe_ctlblk *kcb)
 405 {
 406         __this_cpu_write(current_kprobe, kcb->prev_kprobe.kp);
 407         kcb->kprobe_status = kcb->prev_kprobe.status;
 408         kcb->kprobe_old_flags = kcb->prev_kprobe.old_flags;
 409         kcb->kprobe_saved_flags = kcb->prev_kprobe.saved_flags;
 410 }
 411
 412 static void __kprobes set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
 413                                 struct kprobe_ctlblk *kcb)
 414 {
 415         __this_cpu_write(current_kprobe, p);
 416         kcb->kprobe_saved_flags = kcb->kprobe_old_flags
 417                 = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
 418         if (is_IF_modifier(p->ainsn.insn))
 419                 kcb->kprobe_saved_flags &= ~X86_EFLAGS_IF;
 420 }
 421
 422 static void __kprobes clear_btf(void)
 423 {
 424         if (test_thread_flag(TIF_BLOCKSTEP)) {
 425                 unsigned long debugctl = get_debugctlmsr();
 426
 427                 debugctl &= ~DEBUGCTLMSR_BTF;
 428                 update_debugctlmsr(debugctl);
 429         }
 430 }
 431
 432 static void __kprobes restore_btf(void)
 433 {
 434         if (test_thread_flag(TIF_BLOCKSTEP)) {
 435                 unsigned long debugctl = get_debugctlmsr();
 436
 437                 debugctl |= DEBUGCTLMSR_BTF;
 438                 update_debugctlmsr(debugctl);
 439         }
 440 }
 441
 442 void __kprobes arch_prepare_kretprobe(struct kretprobe_instance *ri,
 443                                       struct pt_regs *regs)
 444 {
 445         unsigned long *sara = stack_addr(regs);
 446
 447         ri->ret_addr = (kprobe_opcode_t *) *sara;
 448
 449         /* Replace the return addr with trampoline addr */
 450         *sara = (unsigned long) &kretprobe_trampoline;
 451 }
 452
 453 #ifdef CONFIG_OPTPROBES
 454 static int  __kprobes setup_detour_execution(struct kprobe *p,
 455                                              struct pt_regs *regs,
 456                                              int reenter);
 457 #else
 458 #define setup_detour_execution(p, regs, reenter) (0)
 459 #endif
 460
 461 static void __kprobes setup_singlestep(struct kprobe *p, struct pt_regs *regs,
 462                                        struct kprobe_ctlblk *kcb, int reenter)
 463 {
 464         if (setup_detour_execution(p, regs, reenter))
 465                 return;
 466
 467 #if !defined(CONFIG_PREEMPT)
 468         if (p->ainsn.boostable == 1 && !p->post_handler) {
 469                 /* Boost up -- we can execute copied instructions directly */
 470                 if (!reenter)
 471                         reset_current_kprobe();
 472                 /*
 473                  * Reentering boosted probe doesn't reset current_kprobe,
 474                  * nor set current_kprobe, because it doesn't use single
 475                  * stepping.
 476                  */
 477                 regs->ip = (unsigned long)p->ainsn.insn;
 478                 preempt_enable_no_resched();
 479                 return;
 480         }
 481 #endif
 482         if (reenter) {
 483                 save_previous_kprobe(kcb);
 484                 set_current_kprobe(p, regs, kcb);
 485                 kcb->kprobe_status = KPROBE_REENTER;
 486         } else
 487                 kcb->kprobe_status = KPROBE_HIT_SS;
 488         /* Prepare real single stepping */
 489         clear_btf();
 490         regs->flags |= X86_EFLAGS_TF;
 491         regs->flags &= ~X86_EFLAGS_IF;
 492         /* single step inline if the instruction is an int3 */
 493         if (p->opcode == BREAKPOINT_INSTRUCTION)
 494                 regs->ip = (unsigned long)p->addr;
 495         else
 496                 regs->ip = (unsigned long)p->ainsn.insn;
 497 }
 498
 499 /*
 500  * We have reentered the kprobe_handler(), since another probe was hit while
 501  * within the handler. We save the original kprobes variables and just single
 502  * step on the instruction of the new probe without calling any user handlers.
 503  */
 504 static int __kprobes reenter_kprobe(struct kprobe *p, struct pt_regs *regs,
 505                                     struct kprobe_ctlblk *kcb)
 506 {
 507         switch (kcb->kprobe_status) {
 508         case KPROBE_HIT_SSDONE:
 509         case KPROBE_HIT_ACTIVE:
 510                 kprobes_inc_nmissed_count(p);
 511                 setup_singlestep(p, regs, kcb, 1);
 512                 break;
 513         case KPROBE_HIT_SS:
 514                 /* A probe has been hit in the codepath leading up to, or just
 515                  * after, single-stepping of a probed instruction. This entire
 516                  * codepath should strictly reside in .kprobes.text section.
 517                  * Raise a BUG or we'll continue in an endless reentering loop
 518                  * and eventually a stack overflow.
 519                  */
 520                 printk(KERN_WARNING "Unrecoverable kprobe detected at %p.\n",
 521                        p->addr);
 522                 dump_kprobe(p);
 523                 BUG();
 524         default:
 525                 /* impossible cases */
 526                 WARN_ON(1);
 527                 return 0;
 528         }
 529
 530         return 1;
 531 }
 532
 533 /*
 534  * Interrupts are disabled on entry as trap3 is an interrupt gate and they
 535  * remain disabled throughout this function.
 536  */
 537 static int __kprobes kprobe_handler(struct pt_regs *regs)
 538 {
 539         kprobe_opcode_t *addr;
 540         struct kprobe *p;
 541         struct kprobe_ctlblk *kcb;
 542
 543         addr = (kprobe_opcode_t *)(regs->ip - sizeof(kprobe_opcode_t));
 544         /*
 545          * We don't want to be preempted for the entire
 546          * duration of kprobe processing. We conditionally
 547          * re-enable preemption at the end of this function,
 548          * and also in reenter_kprobe() and setup_singlestep().
 549          */
 550         preempt_disable();
 551
 552         kcb = get_kprobe_ctlblk();
 553         p = get_kprobe(addr);
 554
 555         if (p) {
 556                 if (kprobe_running()) {
 557                         if (reenter_kprobe(p, regs, kcb))
 558                                 return 1;
 559                 } else {
 560                         set_current_kprobe(p, regs, kcb);
 561                         kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 562
 563                         /*
 564                          * If we have no pre-handler or it returned 0, we
 565                          * continue with normal processing.  If we have a
 566                          * pre-handler and it returned non-zero, it prepped
 567                          * for calling the break_handler below on re-entry
 568                          * for jprobe processing, so get out doing nothing
 569                          * more here.
 570                          */
 571                         if (!p->pre_handler || !p->pre_handler(p, regs))
 572                                 setup_singlestep(p, regs, kcb, 0);
 573                         return 1;
 574                 }
 575         } else if (*addr != BREAKPOINT_INSTRUCTION) {
 576                 /*
 577                  * The breakpoint instruction was removed right
 578                  * after we hit it.  Another cpu has removed
 579                  * either a probepoint or a debugger breakpoint
 580                  * at this address.  In either case, no further
 581                  * handling of this interrupt is appropriate.
 582                  * Back up over the (now missing) int3 and run
 583                  * the original instruction.
 584                  */
 585                 regs->ip = (unsigned long)addr;
 586                 preempt_enable_no_resched();
 587                 return 1;
 588         } else if (kprobe_running()) {
 589                 p = __this_cpu_read(current_kprobe);
 590                 if (p->break_handler && p->break_handler(p, regs)) {
 591                         setup_singlestep(p, regs, kcb, 0);
 592                         return 1;
 593                 }
 594         } /* else: not a kprobe fault; let the kernel handle it */
 595
 596         preempt_enable_no_resched();
 597         return 0;
 598 }
 599
 600 #ifdef CONFIG_X86_64
 601 #define SAVE_REGS_STRING                \
 602         /* Skip cs, ip, orig_ax. */     \
 603         "       subq $24, %rsp\n"       \
 604         "       pushq %rdi\n"           \
 605         "       pushq %rsi\n"           \
 606         "       pushq %rdx\n"           \
 607         "       pushq %rcx\n"           \
 608         "       pushq %rax\n"           \
 609         "       pushq %r8\n"            \
 610         "       pushq %r9\n"            \
 611         "       pushq %r10\n"           \
 612         "       pushq %r11\n"           \
 613         "       pushq %rbx\n"           \
 614         "       pushq %rbp\n"           \
 615         "       pushq %r12\n"           \
 616         "       pushq %r13\n"           \
 617         "       pushq %r14\n"           \
 618         "       pushq %r15\n"
 619 #define RESTORE_REGS_STRING             \
 620         "       popq %r15\n"            \
 621         "       popq %r14\n"            \
 622         "       popq %r13\n"            \
 623         "       popq %r12\n"            \
 624         "       popq %rbp\n"            \
 625         "       popq %rbx\n"            \
 626         "       popq %r11\n"            \
 627         "       popq %r10\n"            \
 628         "       popq %r9\n"             \
 629         "       popq %r8\n"             \
 630         "       popq %rax\n"            \
 631         "       popq %rcx\n"            \
 632         "       popq %rdx\n"            \
 633         "       popq %rsi\n"            \
 634         "       popq %rdi\n"            \
 635         /* Skip orig_ax, ip, cs */      \
 636         "       addq $24, %rsp\n"
 637 #else
 638 #define SAVE_REGS_STRING                \
 639         /* Skip cs, ip, orig_ax and gs. */      \
 640         "       subl $16, %esp\n"       \
 641         "       pushl %fs\n"            \
 642         "       pushl %es\n"            \
 643         "       pushl %ds\n"            \
 644         "       pushl %eax\n"           \
 645         "       pushl %ebp\n"           \
 646         "       pushl %edi\n"           \
 647         "       pushl %esi\n"           \
 648         "       pushl %edx\n"           \
 649         "       pushl %ecx\n"           \
 650         "       pushl %ebx\n"
 651 #define RESTORE_REGS_STRING             \
 652         "       popl %ebx\n"            \
 653         "       popl %ecx\n"            \
 654         "       popl %edx\n"            \
 655         "       popl %esi\n"            \
 656         "       popl %edi\n"            \
 657         "       popl %ebp\n"            \
 658         "       popl %eax\n"            \
 659         /* Skip ds, es, fs, gs, orig_ax, and ip. Note: don't pop cs here*/\
 660         "       addl $24, %esp\n"
 661 #endif
 662
 663 /*
 664  * When a retprobed function returns, this code saves registers and
 665  * calls trampoline_handler() runs, which calls the kretprobe's handler.
 666  */
 667 static void __used __kprobes kretprobe_trampoline_holder(void)
 668 {
 669         asm volatile (
 670                         ".global kretprobe_trampoline\n"
 671                         "kretprobe_trampoline: \n"
 672 #ifdef CONFIG_X86_64
 673                         /* We don't bother saving the ss register */
 674                         "       pushq %rsp\n"
 675                         "       pushfq\n"
 676                         SAVE_REGS_STRING
 677                         "       movq %rsp, %rdi\n"
 678                         "       call trampoline_handler\n"
 679                         /* Replace saved sp with true return address. */
 680                         "       movq %rax, 152(%rsp)\n"
 681                         RESTORE_REGS_STRING
 682                         "       popfq\n"
 683 #else
 684                         "       pushf\n"
 685                         SAVE_REGS_STRING
 686                         "       movl %esp, %eax\n"
 687                         "       call trampoline_handler\n"
 688                         /* Move flags to cs */
 689                         "       movl 56(%esp), %edx\n"
 690                         "       movl %edx, 52(%esp)\n"
 691                         /* Replace saved flags with true return address. */
 692                         "       movl %eax, 56(%esp)\n"
 693                         RESTORE_REGS_STRING
 694                         "       popf\n"
 695 #endif
 696                         "       ret\n");
 697 }
 698
 699 /*
 700  * Called from kretprobe_trampoline
 701  */
 702 static __used __kprobes void *trampoline_handler(struct pt_regs *regs)
 703 {
 704         struct kretprobe_instance *ri = NULL;
 705         struct hlist_head *head, empty_rp;
 706         struct hlist_node *node, *tmp;
 707         unsigned long flags, orig_ret_address = 0;
 708         unsigned long trampoline_address = (unsigned long)&kretprobe_trampoline;
 709         kprobe_opcode_t *correct_ret_addr = NULL;
 710
 711         INIT_HLIST_HEAD(&empty_rp);
 712         kretprobe_hash_lock(current, &head, &flags);
 713         /* fixup registers */
 714 #ifdef CONFIG_X86_64
 715         regs->cs = __KERNEL_CS;
 716 #else
 717         regs->cs = __KERNEL_CS | get_kernel_rpl();
 718         regs->gs = 0;
 719 #endif
 720         regs->ip = trampoline_address;
 721         regs->orig_ax = ~0UL;
 722
 723         /*
 724          * It is possible to have multiple instances associated with a given
 725          * task either because multiple functions in the call path have
 726          * return probes installed on them, and/or more than one
 727          * return probe was registered for a target function.
 728          *
 729          * We can handle this because:
 730          *     - instances are always pushed into the head of the list
 731          *     - when multiple return probes are registered for the same
 732          *       function, the (chronologically) first instance's ret_addr
 733          *       will be the real return address, and all the rest will
 734          *       point to kretprobe_trampoline.
 735          */
 736         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
 737                 if (ri->task != current)
 738                         /* another task is sharing our hash bucket */
 739                         continue;
 740
 741                 orig_ret_address = (unsigned long)ri->ret_addr;
 742
 743                 if (orig_ret_address != trampoline_address)
 744                         /*
 745                          * This is the real return address. Any other
 746                          * instances associated with this task are for
 747                          * other calls deeper on the call stack
 748                          */
 749                         break;
 750         }
 751
 752         kretprobe_assert(ri, orig_ret_address, trampoline_address);
 753
 754         correct_ret_addr = ri->ret_addr;
 755         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
 756                 if (ri->task != current)
 757                         /* another task is sharing our hash bucket */
 758                         continue;
 759
 760                 orig_ret_address = (unsigned long)ri->ret_addr;
 761                 if (ri->rp && ri->rp->handler) {
 762                         __this_cpu_write(current_kprobe, &ri->rp->kp);
 763                         get_kprobe_ctlblk()->kprobe_status = KPROBE_HIT_ACTIVE;
 764                         ri->ret_addr = correct_ret_addr;
 765                         ri->rp->handler(ri, regs);
 766                         __this_cpu_write(current_kprobe, NULL);
 767                 }
 768
 769                 recycle_rp_inst(ri, &empty_rp);
 770
 771                 if (orig_ret_address != trampoline_address)
 772                         /*
 773                          * This is the real return address. Any other
 774                          * instances associated with this task are for
 775                          * other calls deeper on the call stack
 776                          */
 777                         break;
 778         }
 779
 780         kretprobe_hash_unlock(current, &flags);
 781
 782         hlist_for_each_entry_safe(ri, node, tmp, &empty_rp, hlist) {
 783                 hlist_del(&ri->hlist);
 784                 kfree(ri);
 785         }
 786         return (void *)orig_ret_address;
 787 }
 788
 789 /*
 790  * Called after single-stepping.  p->addr is the address of the
 791  * instruction whose first byte has been replaced by the "int 3"
 792  * instruction.  To avoid the SMP problems that can occur when we
 793  * temporarily put back the original opcode to single-step, we
 794  * single-stepped a copy of the instruction.  The address of this
 795  * copy is p->ainsn.insn.
 796  *
 797  * This function prepares to return from the post-single-step
 798  * interrupt.  We have to fix up the stack as follows:
 799  *
 800  * 0) Except in the case of absolute or indirect jump or call instructions,
 801  * the new ip is relative to the copied instruction.  We need to make
 802  * it relative to the original instruction.
 803  *
 804  * 1) If the single-stepped instruction was pushfl, then the TF and IF
 805  * flags are set in the just-pushed flags, and may need to be cleared.
 806  *
 807  * 2) If the single-stepped instruction was a call, the return address
 808  * that is atop the stack is the address following the copied instruction.
 809  * We need to make it the address following the original instruction.
 810  *
 811  * If this is the first time we've single-stepped the instruction at
 812  * this probepoint, and the instruction is boostable, boost it: add a
 813  * jump instruction after the copied instruction, that jumps to the next
 814  * instruction after the probepoint.
 815  */
 816 static void __kprobes resume_execution(struct kprobe *p,
 817                 struct pt_regs *regs, struct kprobe_ctlblk *kcb)
 818 {
 819         unsigned long *tos = stack_addr(regs);
 820         unsigned long copy_ip = (unsigned long)p->ainsn.insn;
 821         unsigned long orig_ip = (unsigned long)p->addr;
 822         kprobe_opcode_t *insn = p->ainsn.insn;
 823
 824         /* Skip prefixes */
 825         insn = skip_prefixes(insn);
 826
 827         regs->flags &= ~X86_EFLAGS_TF;
 828         switch (*insn) {
 829         case 0x9c:      /* pushfl */
 830                 *tos &= ~(X86_EFLAGS_TF | X86_EFLAGS_IF);
 831                 *tos |= kcb->kprobe_old_flags;
 832                 break;
 833         case 0xc2:      /* iret/ret/lret */
 834         case 0xc3:
 835         case 0xca:
 836         case 0xcb:
 837         case 0xcf:
 838         case 0xea:      /* jmp absolute -- ip is correct */
 839                 /* ip is already adjusted, no more changes required */
 840                 p->ainsn.boostable = 1;
 841                 goto no_change;
 842         case 0xe8:      /* call relative - Fix return addr */
 843                 *tos = orig_ip + (*tos - copy_ip);
 844                 break;
 845 #ifdef CONFIG_X86_32
 846         case 0x9a:      /* call absolute -- same as call absolute, indirect */
 847                 *tos = orig_ip + (*tos - copy_ip);
 848                 goto no_change;
 849 #endif
 850         case 0xff:
 851                 if ((insn[1] & 0x30) == 0x10) {
 852                         /*
 853                          * call absolute, indirect
 854                          * Fix return addr; ip is correct.
 855                          * But this is not boostable
 856                          */
 857                         *tos = orig_ip + (*tos - copy_ip);
 858                         goto no_change;
 859                 } else if (((insn[1] & 0x31) == 0x20) ||
 860                            ((insn[1] & 0x31) == 0x21)) {
 861                         /*
 862                          * jmp near and far, absolute indirect
 863                          * ip is correct. And this is boostable
 864                          */
 865                         p->ainsn.boostable = 1;
 866                         goto no_change;
 867                 }
 868         default:
 869                 break;
 870         }
 871
 872         if (p->ainsn.boostable == 0) {
 873                 if ((regs->ip > copy_ip) &&
 874                     (regs->ip - copy_ip) + 5 < MAX_INSN_SIZE) {
 875                         /*
 876                          * These instructions can be executed directly if it
 877                          * jumps back to correct address.
 878                          */
 879                         synthesize_reljump((void *)regs->ip,
 880                                 (void *)orig_ip + (regs->ip - copy_ip));
 881                         p->ainsn.boostable = 1;
 882                 } else {
 883                         p->ainsn.boostable = -1;
 884                 }
 885         }
 886
 887         regs->ip += orig_ip - copy_ip;
 888
 889 no_change:
 890         restore_btf();
 891 }
 892
 893 /*
 894  * Interrupts are disabled on entry as trap1 is an interrupt gate and they
 895  * remain disabled throughout this function.
 896  */
 897 static int __kprobes post_kprobe_handler(struct pt_regs *regs)
 898 {
 899         struct kprobe *cur = kprobe_running();
 900         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 901
 902         if (!cur)
 903                 return 0;
 904
 905         resume_execution(cur, regs, kcb);
 906         regs->flags |= kcb->kprobe_saved_flags;
 907
 908         if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
 909                 kcb->kprobe_status = KPROBE_HIT_SSDONE;
 910                 cur->post_handler(cur, regs, 0);
 911         }
 912
 913         /* Restore back the original saved kprobes variables and continue. */
 914         if (kcb->kprobe_status == KPROBE_REENTER) {
 915                 restore_previous_kprobe(kcb);
 916                 goto out;
 917         }
 918         reset_current_kprobe();
 919 out:
 920         preempt_enable_no_resched();
 921
 922         /*
 923          * if somebody else is singlestepping across a probe point, flags
 924          * will have TF set, in which case, continue the remaining processing
 925          * of do_debug, as if this is not a probe hit.
 926          */
 927         if (regs->flags & X86_EFLAGS_TF)
 928                 return 0;
 929
 930         return 1;
 931 }
 932
 933 int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
 934 {
 935         struct kprobe *cur = kprobe_running();
 936         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 937
 938         switch (kcb->kprobe_status) {
 939         case KPROBE_HIT_SS:
 940         case KPROBE_REENTER:
 941                 /*
 942                  * We are here because the instruction being single
 943                  * stepped caused a page fault. We reset the current
 944                  * kprobe and the ip points back to the probe address
 945                  * and allow the page fault handler to continue as a
 946                  * normal page fault.
 947                  */
 948                 regs->ip = (unsigned long)cur->addr;
 949                 regs->flags |= kcb->kprobe_old_flags;
 950                 if (kcb->kprobe_status == KPROBE_REENTER)
 951                         restore_previous_kprobe(kcb);
 952                 else
 953                         reset_current_kprobe();
 954                 preempt_enable_no_resched();
 955                 break;
 956         case KPROBE_HIT_ACTIVE:
 957         case KPROBE_HIT_SSDONE:
 958                 /*
 959                  * We increment the nmissed count for accounting,
 960                  * we can also use npre/npostfault count for accounting
 961                  * these specific fault cases.
 962                  */
 963                 kprobes_inc_nmissed_count(cur);
 964
 965                 /*
 966                  * We come here because instructions in the pre/post
 967                  * handler caused the page_fault, this could happen
 968                  * if handler tries to access user space by
 969                  * copy_from_user(), get_user() etc. Let the
 970                  * user-specified handler try to fix it first.
 971                  */
 972                 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
 973                         return 1;
 974
 975                 /*
 976                  * In case the user-specified fault handler returned
 977                  * zero, try to fix up.
 978                  */
 979                 if (fixup_exception(regs))
 980                         return 1;
 981
 982                 /*
 983                  * fixup routine could not handle it,
 984                  * Let do_page_fault() fix it.
 985                  */
 986                 break;
 987         default:
 988                 break;
 989         }
 990         return 0;
 991 }
 992
 993 /*
 994  * Wrapper routine for handling exceptions.
 995  */
 996 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
 997                                        unsigned long val, void *data)
 998 {
 999         struct die_args *args = data;
1000         int ret = NOTIFY_DONE;
1001
1002         if (args->regs && user_mode_vm(args->regs))
1003                 return ret;
1004
1005         switch (val) {
1006         case DIE_INT3:
1007                 if (kprobe_handler(args->regs))
1008                         ret = NOTIFY_STOP;
1009                 break;
1010         case DIE_DEBUG:
1011                 if (post_kprobe_handler(args->regs)) {
1012                         /*
1013                          * Reset the BS bit in dr6 (pointed by args->err) to
1014                          * denote completion of processing
1015                          */
1016                         (*(unsigned long *)ERR_PTR(args->err)) &= ~DR_STEP;
1017                         ret = NOTIFY_STOP;
1018                 }
1019                 break;
1020         case DIE_GPF:
1021                 /*
1022                  * To be potentially processing a kprobe fault and to
1023                  * trust the result from kprobe_running(), we have
1024                  * be non-preemptible.
1025                  */
1026                 if (!preemptible() && kprobe_running() &&
1027                     kprobe_fault_handler(args->regs, args->trapnr))
1028                         ret = NOTIFY_STOP;
1029                 break;
1030         default:
1031                 break;
1032         }
1033         return ret;
1034 }
1035
1036 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
1037 {
1038         struct jprobe *jp = container_of(p, struct jprobe, kp);
1039         unsigned long addr;
1040         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1041
1042         kcb->jprobe_saved_regs = *regs;
1043         kcb->jprobe_saved_sp = stack_addr(regs);
1044         addr = (unsigned long)(kcb->jprobe_saved_sp);
1045
1046         /*
1047          * As Linus pointed out, gcc assumes that the callee
1048          * owns the argument space and could overwrite it, e.g.
1049          * tailcall optimization. So, to be absolutely safe
1050          * we also save and restore enough stack bytes to cover
1051          * the argument area.
1052          */
1053         memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
1054                MIN_STACK_SIZE(addr));
1055         regs->flags &= ~X86_EFLAGS_IF;
1056         trace_hardirqs_off();
1057         regs->ip = (unsigned long)(jp->entry);
1058         return 1;
1059 }
1060
1061 void __kprobes jprobe_return(void)
1062 {
1063         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1064
1065         asm volatile (
1066 #ifdef CONFIG_X86_64
1067                         "       xchg   %%rbx,%%rsp      \n"
1068 #else
1069                         "       xchgl   %%ebx,%%esp     \n"
1070 #endif
1071                         "       int3                    \n"
1072                         "       .globl jprobe_return_end\n"
1073                         "       jprobe_return_end:      \n"
1074                         "       nop                     \n"::"b"
1075                         (kcb->jprobe_saved_sp):"memory");
1076 }
1077
1078 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
1079 {
1080         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1081         u8 *addr = (u8 *) (regs->ip - 1);
1082         struct jprobe *jp = container_of(p, struct jprobe, kp);
1083
1084         if ((addr > (u8 *) jprobe_return) &&
1085             (addr < (u8 *) jprobe_return_end)) {
1086                 if (stack_addr(regs) != kcb->jprobe_saved_sp) {
1087                         struct pt_regs *saved_regs = &kcb->jprobe_saved_regs;
1088                         printk(KERN_ERR
1089                                "current sp %p does not match saved sp %p\n",
1090                                stack_addr(regs), kcb->jprobe_saved_sp);
1091                         printk(KERN_ERR "Saved registers for jprobe %p\n", jp);
1092                         show_registers(saved_regs);
1093                         printk(KERN_ERR "Current registers\n");
1094                         show_registers(regs);
1095                         BUG();
1096                 }
1097                 *regs = kcb->jprobe_saved_regs;
1098                 memcpy((kprobe_opcode_t *)(kcb->jprobe_saved_sp),
1099                        kcb->jprobes_stack,
1100                        MIN_STACK_SIZE(kcb->jprobe_saved_sp));
1101                 preempt_enable_no_resched();
1102                 return 1;
1103         }
1104         return 0;
1105 }
1106
1107
1108 #ifdef CONFIG_OPTPROBES
1109
1110 /* Insert a call instruction at address 'from', which calls address 'to'.*/
1111 static void __kprobes synthesize_relcall(void *from, void *to)
1112 {
1113         __synthesize_relative_insn(from, to, RELATIVECALL_OPCODE);
1114 }
1115
1116 /* Insert a move instruction which sets a pointer to eax/rdi (1st arg). */
1117 static void __kprobes synthesize_set_arg1(kprobe_opcode_t *addr,
1118                                           unsigned long val)
1119 {
1120 #ifdef CONFIG_X86_64
1121         *addr++ = 0x48;
1122         *addr++ = 0xbf;
1123 #else
1124         *addr++ = 0xb8;
1125 #endif
1126         *(unsigned long *)addr = val;
1127 }
1128
1129 static void __used __kprobes kprobes_optinsn_template_holder(void)
1130 {
1131         asm volatile (
1132                         ".global optprobe_template_entry\n"
1133                         "optprobe_template_entry: \n"
1134 #ifdef CONFIG_X86_64
1135                         /* We don't bother saving the ss register */
1136                         "       pushq %rsp\n"
1137                         "       pushfq\n"
1138                         SAVE_REGS_STRING
1139                         "       movq %rsp, %rsi\n"
1140                         ".global optprobe_template_val\n"
1141                         "optprobe_template_val: \n"
1142                         ASM_NOP5
1143                         ASM_NOP5
1144                         ".global optprobe_template_call\n"
1145                         "optprobe_template_call: \n"
1146                         ASM_NOP5
1147                         /* Move flags to rsp */
1148                         "       movq 144(%rsp), %rdx\n"
1149                         "       movq %rdx, 152(%rsp)\n"
1150                         RESTORE_REGS_STRING
1151                         /* Skip flags entry */
1152                         "       addq $8, %rsp\n"
1153                         "       popfq\n"
1154 #else /* CONFIG_X86_32 */
1155                         "       pushf\n"
1156                         SAVE_REGS_STRING
1157                         "       movl %esp, %edx\n"
1158                         ".global optprobe_template_val\n"
1159                         "optprobe_template_val: \n"
1160                         ASM_NOP5
1161                         ".global optprobe_template_call\n"
1162                         "optprobe_template_call: \n"
1163                         ASM_NOP5
1164                         RESTORE_REGS_STRING
1165                         "       addl $4, %esp\n"        /* skip cs */
1166                         "       popf\n"
1167 #endif
1168                         ".global optprobe_template_end\n"
1169                         "optprobe_template_end: \n");
1170 }
1171
1172 #define TMPL_MOVE_IDX \
1173         ((long)&optprobe_template_val - (long)&optprobe_template_entry)
1174 #define TMPL_CALL_IDX \
1175         ((long)&optprobe_template_call - (long)&optprobe_template_entry)
1176 #define TMPL_END_IDX \
1177         ((long)&optprobe_template_end - (long)&optprobe_template_entry)
1178
1179 #define INT3_SIZE sizeof(kprobe_opcode_t)
1180
1181 /* Optimized kprobe call back function: called from optinsn */
1182 static void __kprobes optimized_callback(struct optimized_kprobe *op,
1183                                          struct pt_regs *regs)
1184 {
1185         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1186         unsigned long flags;
1187
1188         /* This is possible if op is under delayed unoptimizing */
1189         if (kprobe_disabled(&op->kp))
1190                 return;
1191
1192         local_irq_save(flags);
1193         if (kprobe_running()) {
1194                 kprobes_inc_nmissed_count(&op->kp);
1195         } else {
1196                 /* Save skipped registers */
1197 #ifdef CONFIG_X86_64
1198                 regs->cs = __KERNEL_CS;
1199 #else
1200                 regs->cs = __KERNEL_CS | get_kernel_rpl();
1201                 regs->gs = 0;
1202 #endif
1203                 regs->ip = (unsigned long)op->kp.addr + INT3_SIZE;
1204                 regs->orig_ax = ~0UL;
1205
1206                 __this_cpu_write(current_kprobe, &op->kp);
1207                 kcb->kprobe_status = KPROBE_HIT_ACTIVE;
1208                 opt_pre_handler(&op->kp, regs);
1209                 __this_cpu_write(current_kprobe, NULL);
1210         }
1211         local_irq_restore(flags);
1212 }
1213
1214 static int __kprobes copy_optimized_instructions(u8 *dest, u8 *src)
1215 {
1216         int len = 0, ret;
1217
1218         while (len < RELATIVEJUMP_SIZE) {
1219                 ret = __copy_instruction(dest + len, src + len, 1);
1220                 if (!ret || !can_boost(dest + len))
1221                         return -EINVAL;
1222                 len += ret;
1223         }
1224         /* Check whether the address range is reserved */
1225         if (ftrace_text_reserved(src, src + len - 1) ||
1226             alternatives_text_reserved(src, src + len - 1) ||
1227             jump_label_text_reserved(src, src + len - 1))
1228                 return -EBUSY;
1229
1230         return len;
1231 }
1232
1233 /* Check whether insn is indirect jump */
1234 static int __kprobes insn_is_indirect_jump(struct insn *insn)
1235 {
1236         return ((insn->opcode.bytes[0] == 0xff &&
1237                 (X86_MODRM_REG(insn->modrm.value) & 6) == 4) || /* Jump */
1238                 insn->opcode.bytes[0] == 0xea); /* Segment based jump */
1239 }
1240
1241 /* Check whether insn jumps into specified address range */
1242 static int insn_jump_into_range(struct insn *insn, unsigned long start, int len)
1243 {
1244         unsigned long target = 0;
1245
1246         switch (insn->opcode.bytes[0]) {
1247         case 0xe0:      /* loopne */
1248         case 0xe1:      /* loope */
1249         case 0xe2:      /* loop */
1250         case 0xe3:      /* jcxz */
1251         case 0xe9:      /* near relative jump */
1252         case 0xeb:      /* short relative jump */
1253                 break;
1254         case 0x0f:
1255                 if ((insn->opcode.bytes[1] & 0xf0) == 0x80) /* jcc near */
1256                         break;
1257                 return 0;
1258         default:
1259                 if ((insn->opcode.bytes[0] & 0xf0) == 0x70) /* jcc short */
1260                         break;
1261                 return 0;
1262         }
1263         target = (unsigned long)insn->next_byte + insn->immediate.value;
1264
1265         return (start <= target && target <= start + len);
1266 }
1267
1268 /* Decode whole function to ensure any instructions don't jump into target */
1269 static int __kprobes can_optimize(unsigned long paddr)
1270 {
1271         int ret;
1272         unsigned long addr, size = 0, offset = 0;
1273         struct insn insn;
1274         kprobe_opcode_t buf[MAX_INSN_SIZE];
1275
1276         /* Lookup symbol including addr */
1277         if (!kallsyms_lookup_size_offset(paddr, &size, &offset))
1278                 return 0;
1279
1280         /*
1281          * Do not optimize in the entry code due to the unstable
1282          * stack handling.
1283          */
1284         if ((paddr >= (unsigned long )__entry_text_start) &&
1285             (paddr <  (unsigned long )__entry_text_end))
1286                 return 0;
1287
1288         /* Check there is enough space for a relative jump. */
1289         if (size - offset < RELATIVEJUMP_SIZE)
1290                 return 0;
1291
1292         /* Decode instructions */
1293         addr = paddr - offset;
1294         while (addr < paddr - offset + size) { /* Decode until function end */
1295                 if (search_exception_tables(addr))
1296                         /*
1297                          * Since some fixup code will jumps into this function,
1298                          * we can't optimize kprobe in this function.
1299                          */
1300                         return 0;
1301                 kernel_insn_init(&insn, (void *)addr);
1302                 insn_get_opcode(&insn);
1303                 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION) {
1304                         ret = recover_probed_instruction(buf, addr);
1305                         if (ret)
1306                                 return 0;
1307                         kernel_insn_init(&insn, buf);
1308                 }
1309                 insn_get_length(&insn);
1310                 /* Recover address */
1311                 insn.kaddr = (void *)addr;
1312                 insn.next_byte = (void *)(addr + insn.length);
1313                 /* Check any instructions don't jump into target */
1314                 if (insn_is_indirect_jump(&insn) ||
1315                     insn_jump_into_range(&insn, paddr + INT3_SIZE,
1316                                          RELATIVE_ADDR_SIZE))
1317                         return 0;
1318                 addr += insn.length;
1319         }
1320
1321         return 1;
1322 }
1323
1324 /* Check optimized_kprobe can actually be optimized. */
1325 int __kprobes arch_check_optimized_kprobe(struct optimized_kprobe *op)
1326 {
1327         int i;
1328         struct kprobe *p;
1329
1330         for (i = 1; i < op->optinsn.size; i++) {
1331                 p = get_kprobe(op->kp.addr + i);
1332                 if (p && !kprobe_disabled(p))
1333                         return -EEXIST;
1334         }
1335
1336         return 0;
1337 }
1338
1339 /* Check the addr is within the optimized instructions. */
1340 int __kprobes arch_within_optimized_kprobe(struct optimized_kprobe *op,
1341                                            unsigned long addr)
1342 {
1343         return ((unsigned long)op->kp.addr <= addr &&
1344                 (unsigned long)op->kp.addr + op->optinsn.size > addr);
1345 }
1346
1347 /* Free optimized instruction slot */
1348 static __kprobes
1349 void __arch_remove_optimized_kprobe(struct optimized_kprobe *op, int dirty)
1350 {
1351         if (op->optinsn.insn) {
1352                 free_optinsn_slot(op->optinsn.insn, dirty);
1353                 op->optinsn.insn = NULL;
1354                 op->optinsn.size = 0;
1355         }
1356 }
1357
1358 void __kprobes arch_remove_optimized_kprobe(struct optimized_kprobe *op)
1359 {
1360         __arch_remove_optimized_kprobe(op, 1);
1361 }
1362
1363 /*
1364  * Copy replacing target instructions
1365  * Target instructions MUST be relocatable (checked inside)
1366  */
1367 int __kprobes arch_prepare_optimized_kprobe(struct optimized_kprobe *op)
1368 {
1369         u8 *buf;
1370         int ret;
1371         long rel;
1372
1373         if (!can_optimize((unsigned long)op->kp.addr))
1374                 return -EILSEQ;
1375
1376         op->optinsn.insn = get_optinsn_slot();
1377         if (!op->optinsn.insn)
1378                 return -ENOMEM;
1379
1380         /*
1381          * Verify if the address gap is in 2GB range, because this uses
1382          * a relative jump.
1383          */
1384         rel = (long)op->optinsn.insn - (long)op->kp.addr + RELATIVEJUMP_SIZE;
1385         if (abs(rel) > 0x7fffffff)
1386                 return -ERANGE;
1387
1388         buf = (u8 *)op->optinsn.insn;
1389
1390         /* Copy instructions into the out-of-line buffer */
1391         ret = copy_optimized_instructions(buf + TMPL_END_IDX, op->kp.addr);
1392         if (ret < 0) {
1393                 __arch_remove_optimized_kprobe(op, 0);
1394                 return ret;
1395         }
1396         op->optinsn.size = ret;
1397
1398         /* Copy arch-dep-instance from template */
1399         memcpy(buf, &optprobe_template_entry, TMPL_END_IDX);
1400
1401         /* Set probe information */
1402         synthesize_set_arg1(buf + TMPL_MOVE_IDX, (unsigned long)op);
1403
1404         /* Set probe function call */
1405         synthesize_relcall(buf + TMPL_CALL_IDX, optimized_callback);
1406
1407         /* Set returning jmp instruction at the tail of out-of-line buffer */
1408         synthesize_reljump(buf + TMPL_END_IDX + op->optinsn.size,
1409                            (u8 *)op->kp.addr + op->optinsn.size);
1410
1411         flush_icache_range((unsigned long) buf,
1412                            (unsigned long) buf + TMPL_END_IDX +
1413                            op->optinsn.size + RELATIVEJUMP_SIZE);
1414         return 0;
1415 }
1416
1417 #define MAX_OPTIMIZE_PROBES 256
1418 static struct text_poke_param *jump_poke_params;
1419 static struct jump_poke_buffer {
1420         u8 buf[RELATIVEJUMP_SIZE];
1421 } *jump_poke_bufs;
1422
1423 static void __kprobes setup_optimize_kprobe(struct text_poke_param *tprm,
1424                                             u8 *insn_buf,
1425                                             struct optimized_kprobe *op)
1426 {
1427         s32 rel = (s32)((long)op->optinsn.insn -
1428                         ((long)op->kp.addr + RELATIVEJUMP_SIZE));
1429
1430         /* Backup instructions which will be replaced by jump address */
1431         memcpy(op->optinsn.copied_insn, op->kp.addr + INT3_SIZE,
1432                RELATIVE_ADDR_SIZE);
1433
1434         insn_buf[0] = RELATIVEJUMP_OPCODE;
1435         *(s32 *)(&insn_buf[1]) = rel;
1436
1437         tprm->addr = op->kp.addr;
1438         tprm->opcode = insn_buf;
1439         tprm->len = RELATIVEJUMP_SIZE;
1440 }
1441
1442 /*
1443  * Replace breakpoints (int3) with relative jumps.
1444  * Caller must call with locking kprobe_mutex and text_mutex.
1445  */
1446 void __kprobes arch_optimize_kprobes(struct list_head *oplist)
1447 {
1448         struct optimized_kprobe *op, *tmp;
1449         int c = 0;
1450
1451         list_for_each_entry_safe(op, tmp, oplist, list) {
1452                 WARN_ON(kprobe_disabled(&op->kp));
1453                 /* Setup param */
1454                 setup_optimize_kprobe(&jump_poke_params[c],
1455                                       jump_poke_bufs[c].buf, op);
1456                 list_del_init(&op->list);
1457                 if (++c >= MAX_OPTIMIZE_PROBES)
1458                         break;
1459         }
1460
1461         /*
1462          * text_poke_smp doesn't support NMI/MCE code modifying.
1463          * However, since kprobes itself also doesn't support NMI/MCE
1464          * code probing, it's not a problem.
1465          */
1466         text_poke_smp_batch(jump_poke_params, c);
1467 }
1468
1469 static void __kprobes setup_unoptimize_kprobe(struct text_poke_param *tprm,
1470                                               u8 *insn_buf,
1471                                               struct optimized_kprobe *op)
1472 {
1473         /* Set int3 to first byte for kprobes */
1474         insn_buf[0] = BREAKPOINT_INSTRUCTION;
1475         memcpy(insn_buf + 1, op->optinsn.copied_insn, RELATIVE_ADDR_SIZE);
1476
1477         tprm->addr = op->kp.addr;
1478         tprm->opcode = insn_buf;
1479         tprm->len = RELATIVEJUMP_SIZE;
1480 }
1481
1482 /*
1483  * Recover original instructions and breakpoints from relative jumps.
1484  * Caller must call with locking kprobe_mutex.
1485  */
1486 extern void arch_unoptimize_kprobes(struct list_head *oplist,
1487                                     struct list_head *done_list)
1488 {
1489         struct optimized_kprobe *op, *tmp;
1490         int c = 0;
1491
1492         list_for_each_entry_safe(op, tmp, oplist, list) {
1493                 /* Setup param */
1494                 setup_unoptimize_kprobe(&jump_poke_params[c],
1495                                         jump_poke_bufs[c].buf, op);
1496                 list_move(&op->list, done_list);
1497                 if (++c >= MAX_OPTIMIZE_PROBES)
1498                         break;
1499         }
1500
1501         /*
1502          * text_poke_smp doesn't support NMI/MCE code modifying.
1503          * However, since kprobes itself also doesn't support NMI/MCE
1504          * code probing, it's not a problem.
1505          */
1506         text_poke_smp_batch(jump_poke_params, c);
1507 }
1508
1509 /* Replace a relative jump with a breakpoint (int3).  */
1510 void __kprobes arch_unoptimize_kprobe(struct optimized_kprobe *op)
1511 {
1512         u8 buf[RELATIVEJUMP_SIZE];
1513
1514         /* Set int3 to first byte for kprobes */
1515         buf[0] = BREAKPOINT_INSTRUCTION;
1516         memcpy(buf + 1, op->optinsn.copied_insn, RELATIVE_ADDR_SIZE);
1517         text_poke_smp(op->kp.addr, buf, RELATIVEJUMP_SIZE);
1518 }
1519
1520 static int  __kprobes setup_detour_execution(struct kprobe *p,
1521                                              struct pt_regs *regs,
1522                                              int reenter)
1523 {
1524         struct optimized_kprobe *op;
1525
1526         if (p->flags & KPROBE_FLAG_OPTIMIZED) {
1527                 /* This kprobe is really able to run optimized path. */
1528                 op = container_of(p, struct optimized_kprobe, kp);
1529                 /* Detour through copied instructions */
1530                 regs->ip = (unsigned long)op->optinsn.insn + TMPL_END_IDX;
1531                 if (!reenter)
1532                         reset_current_kprobe();
1533                 preempt_enable_no_resched();
1534                 return 1;
1535         }
1536         return 0;
1537 }
1538
1539 static int __kprobes init_poke_params(void)
1540 {
1541         /* Allocate code buffer and parameter array */
1542         jump_poke_bufs = kmalloc(sizeof(struct jump_poke_buffer) *
1543                                  MAX_OPTIMIZE_PROBES, GFP_KERNEL);
1544         if (!jump_poke_bufs)
1545                 return -ENOMEM;
1546
1547         jump_poke_params = kmalloc(sizeof(struct text_poke_param) *
1548                                    MAX_OPTIMIZE_PROBES, GFP_KERNEL);
1549         if (!jump_poke_params) {
1550                 kfree(jump_poke_bufs);
1551                 jump_poke_bufs = NULL;
1552                 return -ENOMEM;
1553         }
1554
1555         return 0;
1556 }
1557 #else   /* !CONFIG_OPTPROBES */
1558 static int __kprobes init_poke_params(void)
1559 {
1560         return 0;
1561 }
1562 #endif
1563
1564 int __init arch_init_kprobes(void)
1565 {
1566         return init_poke_params();
1567 }
1568
1569 int __kprobes arch_trampoline_kprobe(struct kprobe *p)
1570 {
1571         return 0;
1572 }