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          * This is non-const to keep gcc from statically optimizing it out, as
  79          * variable_test_bit makes gcc think only *(unsigned long*) is used.
  80          */
  81 static u32 twobyte_is_boostable[256 / 32] = {
  82         /*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f          */
  83         /*      ----------------------------------------------          */
  84         W(0x00, 0, 0, 1, 1, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0) | /* 00 */
  85         W(0x10, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 10 */
  86         W(0x20, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 20 */
  87         W(0x30, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 30 */
  88         W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
  89         W(0x50, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 50 */
  90         W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1) | /* 60 */
  91         W(0x70, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1) , /* 70 */
  92         W(0x80, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 80 */
  93         W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
  94         W(0xa0, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* a0 */
  95         W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1) , /* b0 */
  96         W(0xc0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1) | /* c0 */
  97         W(0xd0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) , /* d0 */
  98         W(0xe0, 0, 1, 1, 0, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* e0 */
  99         W(0xf0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 0)   /* f0 */
 100         /*      -----------------------------------------------         */
 101         /*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f          */
 102 };
 103 #undef W
 104
 105 struct kretprobe_blackpoint kretprobe_blacklist[] = {
 106         {"__switch_to", }, /* This function switches only current task, but
 107                               doesn't switch kernel stack.*/
 108         {NULL, NULL}    /* Terminator */
 109 };
 110 const int kretprobe_blacklist_size = ARRAY_SIZE(kretprobe_blacklist);
 111
 112 static void __kprobes __synthesize_relative_insn(void *from, void *to, u8 op)
 113 {
 114         struct __arch_relative_insn {
 115                 u8 op;
 116                 s32 raddr;
 117         } __attribute__((packed)) *insn;
 118
 119         insn = (struct __arch_relative_insn *)from;
 120         insn->raddr = (s32)((long)(to) - ((long)(from) + 5));
 121         insn->op = op;
 122 }
 123
 124 /* Insert a jump instruction at address 'from', which jumps to address 'to'.*/
 125 static void __kprobes synthesize_reljump(void *from, void *to)
 126 {
 127         __synthesize_relative_insn(from, to, RELATIVEJUMP_OPCODE);
 128 }
 129
 130 /*
 131  * Skip the prefixes of the instruction.
 132  */
 133 static kprobe_opcode_t *__kprobes skip_prefixes(kprobe_opcode_t *insn)
 134 {
 135         insn_attr_t attr;
 136
 137         attr = inat_get_opcode_attribute((insn_byte_t)*insn);
 138         while (inat_is_legacy_prefix(attr)) {
 139                 insn++;
 140                 attr = inat_get_opcode_attribute((insn_byte_t)*insn);
 141         }
 142 #ifdef CONFIG_X86_64
 143         if (inat_is_rex_prefix(attr))
 144                 insn++;
 145 #endif
 146         return insn;
 147 }
 148
 149 /*
 150  * Returns non-zero if opcode is boostable.
 151  * RIP relative instructions are adjusted at copying time in 64 bits mode
 152  */
 153 static int __kprobes can_boost(kprobe_opcode_t *opcodes)
 154 {
 155         kprobe_opcode_t opcode;
 156         kprobe_opcode_t *orig_opcodes = opcodes;
 157
 158         if (search_exception_tables((unsigned long)opcodes))
 159                 return 0;       /* Page fault may occur on this address. */
 160
 161 retry:
 162         if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
 163                 return 0;
 164         opcode = *(opcodes++);
 165
 166         /* 2nd-byte opcode */
 167         if (opcode == 0x0f) {
 168                 if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
 169                         return 0;
 170                 return test_bit(*opcodes,
 171                                 (unsigned long *)twobyte_is_boostable);
 172         }
 173
 174         switch (opcode & 0xf0) {
 175 #ifdef CONFIG_X86_64
 176         case 0x40:
 177                 goto retry; /* REX prefix is boostable */
 178 #endif
 179         case 0x60:
 180                 if (0x63 < opcode && opcode < 0x67)
 181                         goto retry; /* prefixes */
 182                 /* can't boost Address-size override and bound */
 183                 return (opcode != 0x62 && opcode != 0x67);
 184         case 0x70:
 185                 return 0; /* can't boost conditional jump */
 186         case 0xc0:
 187                 /* can't boost software-interruptions */
 188                 return (0xc1 < opcode && opcode < 0xcc) || opcode == 0xcf;
 189         case 0xd0:
 190                 /* can boost AA* and XLAT */
 191                 return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
 192         case 0xe0:
 193                 /* can boost in/out and absolute jmps */
 194                 return ((opcode & 0x04) || opcode == 0xea);
 195         case 0xf0:
 196                 if ((opcode & 0x0c) == 0 && opcode != 0xf1)
 197                         goto retry; /* lock/rep(ne) prefix */
 198                 /* clear and set flags are boostable */
 199                 return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
 200         default:
 201                 /* segment override prefixes are boostable */
 202                 if (opcode == 0x26 || opcode == 0x36 || opcode == 0x3e)
 203                         goto retry; /* prefixes */
 204                 /* CS override prefix and call are not boostable */
 205                 return (opcode != 0x2e && opcode != 0x9a);
 206         }
 207 }
 208
 209 /* Recover the probed instruction at addr for further analysis. */
 210 static int recover_probed_instruction(kprobe_opcode_t *buf, unsigned long addr)
 211 {
 212         struct kprobe *kp;
 213         kp = get_kprobe((void *)addr);
 214         if (!kp)
 215                 return -EINVAL;
 216
 217         /*
 218          *  Basically, kp->ainsn.insn has an original instruction.
 219          *  However, RIP-relative instruction can not do single-stepping
 220          *  at different place, __copy_instruction() tweaks the displacement of
 221          *  that instruction. In that case, we can't recover the instruction
 222          *  from the kp->ainsn.insn.
 223          *
 224          *  On the other hand, kp->opcode has a copy of the first byte of
 225          *  the probed instruction, which is overwritten by int3. And
 226          *  the instruction at kp->addr is not modified by kprobes except
 227          *  for the first byte, we can recover the original instruction
 228          *  from it and kp->opcode.
 229          */
 230         memcpy(buf, kp->addr, MAX_INSN_SIZE * sizeof(kprobe_opcode_t));
 231         buf[0] = kp->opcode;
 232         return 0;
 233 }
 234
 235 /* Check if paddr is at an instruction boundary */
 236 static int __kprobes can_probe(unsigned long paddr)
 237 {
 238         int ret;
 239         unsigned long addr, offset = 0;
 240         struct insn insn;
 241         kprobe_opcode_t buf[MAX_INSN_SIZE];
 242
 243         if (!kallsyms_lookup_size_offset(paddr, NULL, &offset))
 244                 return 0;
 245
 246         /* Decode instructions */
 247         addr = paddr - offset;
 248         while (addr < paddr) {
 249                 kernel_insn_init(&insn, (void *)addr);
 250                 insn_get_opcode(&insn);
 251
 252                 /*
 253                  * Check if the instruction has been modified by another
 254                  * kprobe, in which case we replace the breakpoint by the
 255                  * original instruction in our buffer.
 256                  */
 257                 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION) {
 258                         ret = recover_probed_instruction(buf, addr);
 259                         if (ret)
 260                                 /*
 261                                  * Another debugging subsystem might insert
 262                                  * this breakpoint. In that case, we can't
 263                                  * recover it.
 264                                  */
 265                                 return 0;
 266                         kernel_insn_init(&insn, buf);
 267                 }
 268                 insn_get_length(&insn);
 269                 addr += insn.length;
 270         }
 271
 272         return (addr == paddr);
 273 }
 274
 275 /*
 276  * Returns non-zero if opcode modifies the interrupt flag.
 277  */
 278 static int __kprobes is_IF_modifier(kprobe_opcode_t *insn)
 279 {
 280         /* Skip prefixes */
 281         insn = skip_prefixes(insn);
 282
 283         switch (*insn) {
 284         case 0xfa:              /* cli */
 285         case 0xfb:              /* sti */
 286         case 0xcf:              /* iret/iretd */
 287         case 0x9d:              /* popf/popfd */
 288                 return 1;
 289         }
 290
 291         return 0;
 292 }
 293
 294 /*
 295  * Copy an instruction and adjust the displacement if the instruction
 296  * uses the %rip-relative addressing mode.
 297  * If it does, Return the address of the 32-bit displacement word.
 298  * If not, return null.
 299  * Only applicable to 64-bit x86.
 300  */
 301 static int __kprobes __copy_instruction(u8 *dest, u8 *src, int recover)
 302 {
 303         struct insn insn;
 304         int ret;
 305         kprobe_opcode_t buf[MAX_INSN_SIZE];
 306
 307         kernel_insn_init(&insn, src);
 308         if (recover) {
 309                 insn_get_opcode(&insn);
 310                 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION) {
 311                         ret = recover_probed_instruction(buf,
 312                                                          (unsigned long)src);
 313                         if (ret)
 314                                 return 0;
 315                         kernel_insn_init(&insn, buf);
 316                 }
 317         }
 318         insn_get_length(&insn);
 319         memcpy(dest, insn.kaddr, insn.length);
 320
 321 #ifdef CONFIG_X86_64
 322         if (insn_rip_relative(&insn)) {
 323                 s64 newdisp;
 324                 u8 *disp;
 325                 kernel_insn_init(&insn, dest);
 326                 insn_get_displacement(&insn);
 327                 /*
 328                  * The copied instruction uses the %rip-relative addressing
 329                  * mode.  Adjust the displacement for the difference between
 330                  * the original location of this instruction and the location
 331                  * of the copy that will actually be run.  The tricky bit here
 332                  * is making sure that the sign extension happens correctly in
 333                  * this calculation, since we need a signed 32-bit result to
 334                  * be sign-extended to 64 bits when it's added to the %rip
 335                  * value and yield the same 64-bit result that the sign-
 336                  * extension of the original signed 32-bit displacement would
 337                  * have given.
 338                  */
 339                 newdisp = (u8 *) src + (s64) insn.displacement.value -
 340                           (u8 *) dest;
 341                 BUG_ON((s64) (s32) newdisp != newdisp); /* Sanity check.  */
 342                 disp = (u8 *) dest + insn_offset_displacement(&insn);
 343                 *(s32 *) disp = (s32) newdisp;
 344         }
 345 #endif
 346         return insn.length;
 347 }
 348
 349 static void __kprobes arch_copy_kprobe(struct kprobe *p)
 350 {
 351         /*
 352          * Copy an instruction without recovering int3, because it will be
 353          * put by another subsystem.
 354          */
 355         __copy_instruction(p->ainsn.insn, p->addr, 0);
 356
 357         if (can_boost(p->addr))
 358                 p->ainsn.boostable = 0;
 359         else
 360                 p->ainsn.boostable = -1;
 361
 362         p->opcode = *p->addr;
 363 }
 364
 365 int __kprobes arch_prepare_kprobe(struct kprobe *p)
 366 {
 367         if (alternatives_text_reserved(p->addr, p->addr))
 368                 return -EINVAL;
 369
 370         if (!can_probe((unsigned long)p->addr))
 371                 return -EILSEQ;
 372         /* insn: must be on special executable page on x86. */
 373         p->ainsn.insn = get_insn_slot();
 374         if (!p->ainsn.insn)
 375                 return -ENOMEM;
 376         arch_copy_kprobe(p);
 377         return 0;
 378 }
 379
 380 void __kprobes arch_arm_kprobe(struct kprobe *p)
 381 {
 382         text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
 383 }
 384
 385 void __kprobes arch_disarm_kprobe(struct kprobe *p)
 386 {
 387         text_poke(p->addr, &p->opcode, 1);
 388 }
 389
 390 void __kprobes arch_remove_kprobe(struct kprobe *p)
 391 {
 392         if (p->ainsn.insn) {
 393                 free_insn_slot(p->ainsn.insn, (p->ainsn.boostable == 1));
 394                 p->ainsn.insn = NULL;
 395         }
 396 }
 397
 398 static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
 399 {
 400         kcb->prev_kprobe.kp = kprobe_running();
 401         kcb->prev_kprobe.status = kcb->kprobe_status;
 402         kcb->prev_kprobe.old_flags = kcb->kprobe_old_flags;
 403         kcb->prev_kprobe.saved_flags = kcb->kprobe_saved_flags;
 404 }
 405
 406 static void __kprobes restore_previous_kprobe(struct kprobe_ctlblk *kcb)
 407 {
 408         __this_cpu_write(current_kprobe, kcb->prev_kprobe.kp);
 409         kcb->kprobe_status = kcb->prev_kprobe.status;
 410         kcb->kprobe_old_flags = kcb->prev_kprobe.old_flags;
 411         kcb->kprobe_saved_flags = kcb->prev_kprobe.saved_flags;
 412 }
 413
 414 static void __kprobes set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
 415                                 struct kprobe_ctlblk *kcb)
 416 {
 417         __this_cpu_write(current_kprobe, p);
 418         kcb->kprobe_saved_flags = kcb->kprobe_old_flags
 419                 = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
 420         if (is_IF_modifier(p->ainsn.insn))
 421                 kcb->kprobe_saved_flags &= ~X86_EFLAGS_IF;
 422 }
 423
 424 static void __kprobes clear_btf(void)
 425 {
 426         if (test_thread_flag(TIF_BLOCKSTEP)) {
 427                 unsigned long debugctl = get_debugctlmsr();
 428
 429                 debugctl &= ~DEBUGCTLMSR_BTF;
 430                 update_debugctlmsr(debugctl);
 431         }
 432 }
 433
 434 static void __kprobes restore_btf(void)
 435 {
 436         if (test_thread_flag(TIF_BLOCKSTEP)) {
 437                 unsigned long debugctl = get_debugctlmsr();
 438
 439                 debugctl |= DEBUGCTLMSR_BTF;
 440                 update_debugctlmsr(debugctl);
 441         }
 442 }
 443
 444 void __kprobes arch_prepare_kretprobe(struct kretprobe_instance *ri,
 445                                       struct pt_regs *regs)
 446 {
 447         unsigned long *sara = stack_addr(regs);
 448
 449         ri->ret_addr = (kprobe_opcode_t *) *sara;
 450
 451         /* Replace the return addr with trampoline addr */
 452         *sara = (unsigned long) &kretprobe_trampoline;
 453 }
 454
 455 #ifdef CONFIG_OPTPROBES
 456 static int  __kprobes setup_detour_execution(struct kprobe *p,
 457                                              struct pt_regs *regs,
 458                                              int reenter);
 459 #else
 460 #define setup_detour_execution(p, regs, reenter) (0)
 461 #endif
 462
 463 static void __kprobes setup_singlestep(struct kprobe *p, struct pt_regs *regs,
 464                                        struct kprobe_ctlblk *kcb, int reenter)
 465 {
 466         if (setup_detour_execution(p, regs, reenter))
 467                 return;
 468
 469 #if !defined(CONFIG_PREEMPT)
 470         if (p->ainsn.boostable == 1 && !p->post_handler) {
 471                 /* Boost up -- we can execute copied instructions directly */
 472                 if (!reenter)
 473                         reset_current_kprobe();
 474                 /*
 475                  * Reentering boosted probe doesn't reset current_kprobe,
 476                  * nor set current_kprobe, because it doesn't use single
 477                  * stepping.
 478                  */
 479                 regs->ip = (unsigned long)p->ainsn.insn;
 480                 preempt_enable_no_resched();
 481                 return;
 482         }
 483 #endif
 484         if (reenter) {
 485                 save_previous_kprobe(kcb);
 486                 set_current_kprobe(p, regs, kcb);
 487                 kcb->kprobe_status = KPROBE_REENTER;
 488         } else
 489                 kcb->kprobe_status = KPROBE_HIT_SS;
 490         /* Prepare real single stepping */
 491         clear_btf();
 492         regs->flags |= X86_EFLAGS_TF;
 493         regs->flags &= ~X86_EFLAGS_IF;
 494         /* single step inline if the instruction is an int3 */
 495         if (p->opcode == BREAKPOINT_INSTRUCTION)
 496                 regs->ip = (unsigned long)p->addr;
 497         else
 498                 regs->ip = (unsigned long)p->ainsn.insn;
 499 }
 500
 501 /*
 502  * We have reentered the kprobe_handler(), since another probe was hit while
 503  * within the handler. We save the original kprobes variables and just single
 504  * step on the instruction of the new probe without calling any user handlers.
 505  */
 506 static int __kprobes reenter_kprobe(struct kprobe *p, struct pt_regs *regs,
 507                                     struct kprobe_ctlblk *kcb)
 508 {
 509         switch (kcb->kprobe_status) {
 510         case KPROBE_HIT_SSDONE:
 511         case KPROBE_HIT_ACTIVE:
 512                 kprobes_inc_nmissed_count(p);
 513                 setup_singlestep(p, regs, kcb, 1);
 514                 break;
 515         case KPROBE_HIT_SS:
 516                 /* A probe has been hit in the codepath leading up to, or just
 517                  * after, single-stepping of a probed instruction. This entire
 518                  * codepath should strictly reside in .kprobes.text section.
 519                  * Raise a BUG or we'll continue in an endless reentering loop
 520                  * and eventually a stack overflow.
 521                  */
 522                 printk(KERN_WARNING "Unrecoverable kprobe detected at %p.\n",
 523                        p->addr);
 524                 dump_kprobe(p);
 525                 BUG();
 526         default:
 527                 /* impossible cases */
 528                 WARN_ON(1);
 529                 return 0;
 530         }
 531
 532         return 1;
 533 }
 534
 535 /*
 536  * Interrupts are disabled on entry as trap3 is an interrupt gate and they
 537  * remain disabled throughout this function.
 538  */
 539 static int __kprobes kprobe_handler(struct pt_regs *regs)
 540 {
 541         kprobe_opcode_t *addr;
 542         struct kprobe *p;
 543         struct kprobe_ctlblk *kcb;
 544
 545         addr = (kprobe_opcode_t *)(regs->ip - sizeof(kprobe_opcode_t));
 546         /*
 547          * We don't want to be preempted for the entire
 548          * duration of kprobe processing. We conditionally
 549          * re-enable preemption at the end of this function,
 550          * and also in reenter_kprobe() and setup_singlestep().
 551          */
 552         preempt_disable();
 553
 554         kcb = get_kprobe_ctlblk();
 555         p = get_kprobe(addr);
 556
 557         if (p) {
 558                 if (kprobe_running()) {
 559                         if (reenter_kprobe(p, regs, kcb))
 560                                 return 1;
 561                 } else {
 562                         set_current_kprobe(p, regs, kcb);
 563                         kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 564
 565                         /*
 566                          * If we have no pre-handler or it returned 0, we
 567                          * continue with normal processing.  If we have a
 568                          * pre-handler and it returned non-zero, it prepped
 569                          * for calling the break_handler below on re-entry
 570                          * for jprobe processing, so get out doing nothing
 571                          * more here.
 572                          */
 573                         if (!p->pre_handler || !p->pre_handler(p, regs))
 574                                 setup_singlestep(p, regs, kcb, 0);
 575                         return 1;
 576                 }
 577         } else if (*addr != BREAKPOINT_INSTRUCTION) {
 578                 /*
 579                  * The breakpoint instruction was removed right
 580                  * after we hit it.  Another cpu has removed
 581                  * either a probepoint or a debugger breakpoint
 582                  * at this address.  In either case, no further
 583                  * handling of this interrupt is appropriate.
 584                  * Back up over the (now missing) int3 and run
 585                  * the original instruction.
 586                  */
 587                 regs->ip = (unsigned long)addr;
 588                 preempt_enable_no_resched();
 589                 return 1;
 590         } else if (kprobe_running()) {
 591                 p = __this_cpu_read(current_kprobe);
 592                 if (p->break_handler && p->break_handler(p, regs)) {
 593                         setup_singlestep(p, regs, kcb, 0);
 594                         return 1;
 595                 }
 596         } /* else: not a kprobe fault; let the kernel handle it */
 597
 598         preempt_enable_no_resched();
 599         return 0;
 600 }
 601
 602 #ifdef CONFIG_X86_64
 603 #define SAVE_REGS_STRING                \
 604         /* Skip cs, ip, orig_ax. */     \
 605         "       subq $24, %rsp\n"       \
 606         "       pushq %rdi\n"           \
 607         "       pushq %rsi\n"           \
 608         "       pushq %rdx\n"           \
 609         "       pushq %rcx\n"           \
 610         "       pushq %rax\n"           \
 611         "       pushq %r8\n"            \
 612         "       pushq %r9\n"            \
 613         "       pushq %r10\n"           \
 614         "       pushq %r11\n"           \
 615         "       pushq %rbx\n"           \
 616         "       pushq %rbp\n"           \
 617         "       pushq %r12\n"           \
 618         "       pushq %r13\n"           \
 619         "       pushq %r14\n"           \
 620         "       pushq %r15\n"
 621 #define RESTORE_REGS_STRING             \
 622         "       popq %r15\n"            \
 623         "       popq %r14\n"            \
 624         "       popq %r13\n"            \
 625         "       popq %r12\n"            \
 626         "       popq %rbp\n"            \
 627         "       popq %rbx\n"            \
 628         "       popq %r11\n"            \
 629         "       popq %r10\n"            \
 630         "       popq %r9\n"             \
 631         "       popq %r8\n"             \
 632         "       popq %rax\n"            \
 633         "       popq %rcx\n"            \
 634         "       popq %rdx\n"            \
 635         "       popq %rsi\n"            \
 636         "       popq %rdi\n"            \
 637         /* Skip orig_ax, ip, cs */      \
 638         "       addq $24, %rsp\n"
 639 #else
 640 #define SAVE_REGS_STRING                \
 641         /* Skip cs, ip, orig_ax and gs. */      \
 642         "       subl $16, %esp\n"       \
 643         "       pushl %fs\n"            \
 644         "       pushl %es\n"            \
 645         "       pushl %ds\n"            \
 646         "       pushl %eax\n"           \
 647         "       pushl %ebp\n"           \
 648         "       pushl %edi\n"           \
 649         "       pushl %esi\n"           \
 650         "       pushl %edx\n"           \
 651         "       pushl %ecx\n"           \
 652         "       pushl %ebx\n"
 653 #define RESTORE_REGS_STRING             \
 654         "       popl %ebx\n"            \
 655         "       popl %ecx\n"            \
 656         "       popl %edx\n"            \
 657         "       popl %esi\n"            \
 658         "       popl %edi\n"            \
 659         "       popl %ebp\n"            \
 660         "       popl %eax\n"            \
 661         /* Skip ds, es, fs, gs, orig_ax, and ip. Note: don't pop cs here*/\
 662         "       addl $24, %esp\n"
 663 #endif
 664
 665 /*
 666  * When a retprobed function returns, this code saves registers and
 667  * calls trampoline_handler() runs, which calls the kretprobe's handler.
 668  */
 669 static void __used __kprobes kretprobe_trampoline_holder(void)
 670 {
 671         asm volatile (
 672                         ".global kretprobe_trampoline\n"
 673                         "kretprobe_trampoline: \n"
 674 #ifdef CONFIG_X86_64
 675                         /* We don't bother saving the ss register */
 676                         "       pushq %rsp\n"
 677                         "       pushfq\n"
 678                         SAVE_REGS_STRING
 679                         "       movq %rsp, %rdi\n"
 680                         "       call trampoline_handler\n"
 681                         /* Replace saved sp with true return address. */
 682                         "       movq %rax, 152(%rsp)\n"
 683                         RESTORE_REGS_STRING
 684                         "       popfq\n"
 685 #else
 686                         "       pushf\n"
 687                         SAVE_REGS_STRING
 688                         "       movl %esp, %eax\n"
 689                         "       call trampoline_handler\n"
 690                         /* Move flags to cs */
 691                         "       movl 56(%esp), %edx\n"
 692                         "       movl %edx, 52(%esp)\n"
 693                         /* Replace saved flags with true return address. */
 694                         "       movl %eax, 56(%esp)\n"
 695                         RESTORE_REGS_STRING
 696                         "       popf\n"
 697 #endif
 698                         "       ret\n");
 699 }
 700
 701 /*
 702  * Called from kretprobe_trampoline
 703  */
 704 static __used __kprobes void *trampoline_handler(struct pt_regs *regs)
 705 {
 706         struct kretprobe_instance *ri = NULL;
 707         struct hlist_head *head, empty_rp;
 708         struct hlist_node *node, *tmp;
 709         unsigned long flags, orig_ret_address = 0;
 710         unsigned long trampoline_address = (unsigned long)&kretprobe_trampoline;
 711         kprobe_opcode_t *correct_ret_addr = NULL;
 712
 713         INIT_HLIST_HEAD(&empty_rp);
 714         kretprobe_hash_lock(current, &head, &flags);
 715         /* fixup registers */
 716 #ifdef CONFIG_X86_64
 717         regs->cs = __KERNEL_CS;
 718 #else
 719         regs->cs = __KERNEL_CS | get_kernel_rpl();
 720         regs->gs = 0;
 721 #endif
 722         regs->ip = trampoline_address;
 723         regs->orig_ax = ~0UL;
 724
 725         /*
 726          * It is possible to have multiple instances associated with a given
 727          * task either because multiple functions in the call path have
 728          * return probes installed on them, and/or more than one
 729          * return probe was registered for a target function.
 730          *
 731          * We can handle this because:
 732          *     - instances are always pushed into the head of the list
 733          *     - when multiple return probes are registered for the same
 734          *       function, the (chronologically) first instance's ret_addr
 735          *       will be the real return address, and all the rest will
 736          *       point to kretprobe_trampoline.
 737          */
 738         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
 739                 if (ri->task != current)
 740                         /* another task is sharing our hash bucket */
 741                         continue;
 742
 743                 orig_ret_address = (unsigned long)ri->ret_addr;
 744
 745                 if (orig_ret_address != trampoline_address)
 746                         /*
 747                          * This is the real return address. Any other
 748                          * instances associated with this task are for
 749                          * other calls deeper on the call stack
 750                          */
 751                         break;
 752         }
 753
 754         kretprobe_assert(ri, orig_ret_address, trampoline_address);
 755
 756         correct_ret_addr = ri->ret_addr;
 757         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
 758                 if (ri->task != current)
 759                         /* another task is sharing our hash bucket */
 760                         continue;
 761
 762                 orig_ret_address = (unsigned long)ri->ret_addr;
 763                 if (ri->rp && ri->rp->handler) {
 764                         __this_cpu_write(current_kprobe, &ri->rp->kp);
 765                         get_kprobe_ctlblk()->kprobe_status = KPROBE_HIT_ACTIVE;
 766                         ri->ret_addr = correct_ret_addr;
 767                         ri->rp->handler(ri, regs);
 768                         __this_cpu_write(current_kprobe, NULL);
 769                 }
 770
 771                 recycle_rp_inst(ri, &empty_rp);
 772
 773                 if (orig_ret_address != trampoline_address)
 774                         /*
 775                          * This is the real return address. Any other
 776                          * instances associated with this task are for
 777                          * other calls deeper on the call stack
 778                          */
 779                         break;
 780         }
 781
 782         kretprobe_hash_unlock(current, &flags);
 783
 784         hlist_for_each_entry_safe(ri, node, tmp, &empty_rp, hlist) {
 785                 hlist_del(&ri->hlist);
 786                 kfree(ri);
 787         }
 788         return (void *)orig_ret_address;
 789 }
 790
 791 /*
 792  * Called after single-stepping.  p->addr is the address of the
 793  * instruction whose first byte has been replaced by the "int 3"
 794  * instruction.  To avoid the SMP problems that can occur when we
 795  * temporarily put back the original opcode to single-step, we
 796  * single-stepped a copy of the instruction.  The address of this
 797  * copy is p->ainsn.insn.
 798  *
 799  * This function prepares to return from the post-single-step
 800  * interrupt.  We have to fix up the stack as follows:
 801  *
 802  * 0) Except in the case of absolute or indirect jump or call instructions,
 803  * the new ip is relative to the copied instruction.  We need to make
 804  * it relative to the original instruction.
 805  *
 806  * 1) If the single-stepped instruction was pushfl, then the TF and IF
 807  * flags are set in the just-pushed flags, and may need to be cleared.
 808  *
 809  * 2) If the single-stepped instruction was a call, the return address
 810  * that is atop the stack is the address following the copied instruction.
 811  * We need to make it the address following the original instruction.
 812  *
 813  * If this is the first time we've single-stepped the instruction at
 814  * this probepoint, and the instruction is boostable, boost it: add a
 815  * jump instruction after the copied instruction, that jumps to the next
 816  * instruction after the probepoint.
 817  */
 818 static void __kprobes resume_execution(struct kprobe *p,
 819                 struct pt_regs *regs, struct kprobe_ctlblk *kcb)
 820 {
 821         unsigned long *tos = stack_addr(regs);
 822         unsigned long copy_ip = (unsigned long)p->ainsn.insn;
 823         unsigned long orig_ip = (unsigned long)p->addr;
 824         kprobe_opcode_t *insn = p->ainsn.insn;
 825
 826         /* Skip prefixes */
 827         insn = skip_prefixes(insn);
 828
 829         regs->flags &= ~X86_EFLAGS_TF;
 830         switch (*insn) {
 831         case 0x9c:      /* pushfl */
 832                 *tos &= ~(X86_EFLAGS_TF | X86_EFLAGS_IF);
 833                 *tos |= kcb->kprobe_old_flags;
 834                 break;
 835         case 0xc2:      /* iret/ret/lret */
 836         case 0xc3:
 837         case 0xca:
 838         case 0xcb:
 839         case 0xcf:
 840         case 0xea:      /* jmp absolute -- ip is correct */
 841                 /* ip is already adjusted, no more changes required */
 842                 p->ainsn.boostable = 1;
 843                 goto no_change;
 844         case 0xe8:      /* call relative - Fix return addr */
 845                 *tos = orig_ip + (*tos - copy_ip);
 846                 break;
 847 #ifdef CONFIG_X86_32
 848         case 0x9a:      /* call absolute -- same as call absolute, indirect */
 849                 *tos = orig_ip + (*tos - copy_ip);
 850                 goto no_change;
 851 #endif
 852         case 0xff:
 853                 if ((insn[1] & 0x30) == 0x10) {
 854                         /*
 855                          * call absolute, indirect
 856                          * Fix return addr; ip is correct.
 857                          * But this is not boostable
 858                          */
 859                         *tos = orig_ip + (*tos - copy_ip);
 860                         goto no_change;
 861                 } else if (((insn[1] & 0x31) == 0x20) ||
 862                            ((insn[1] & 0x31) == 0x21)) {
 863                         /*
 864                          * jmp near and far, absolute indirect
 865                          * ip is correct. And this is boostable
 866                          */
 867                         p->ainsn.boostable = 1;
 868                         goto no_change;
 869                 }
 870         default:
 871                 break;
 872         }
 873
 874         if (p->ainsn.boostable == 0) {
 875                 if ((regs->ip > copy_ip) &&
 876                     (regs->ip - copy_ip) + 5 < MAX_INSN_SIZE) {
 877                         /*
 878                          * These instructions can be executed directly if it
 879                          * jumps back to correct address.
 880                          */
 881                         synthesize_reljump((void *)regs->ip,
 882                                 (void *)orig_ip + (regs->ip - copy_ip));
 883                         p->ainsn.boostable = 1;
 884                 } else {
 885                         p->ainsn.boostable = -1;
 886                 }
 887         }
 888
 889         regs->ip += orig_ip - copy_ip;
 890
 891 no_change:
 892         restore_btf();
 893 }
 894
 895 /*
 896  * Interrupts are disabled on entry as trap1 is an interrupt gate and they
 897  * remain disabled throughout this function.
 898  */
 899 static int __kprobes post_kprobe_handler(struct pt_regs *regs)
 900 {
 901         struct kprobe *cur = kprobe_running();
 902         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 903
 904         if (!cur)
 905                 return 0;
 906
 907         resume_execution(cur, regs, kcb);
 908         regs->flags |= kcb->kprobe_saved_flags;
 909
 910         if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
 911                 kcb->kprobe_status = KPROBE_HIT_SSDONE;
 912                 cur->post_handler(cur, regs, 0);
 913         }
 914
 915         /* Restore back the original saved kprobes variables and continue. */
 916         if (kcb->kprobe_status == KPROBE_REENTER) {
 917                 restore_previous_kprobe(kcb);
 918                 goto out;
 919         }
 920         reset_current_kprobe();
 921 out:
 922         preempt_enable_no_resched();
 923
 924         /*
 925          * if somebody else is singlestepping across a probe point, flags
 926          * will have TF set, in which case, continue the remaining processing
 927          * of do_debug, as if this is not a probe hit.
 928          */
 929         if (regs->flags & X86_EFLAGS_TF)
 930                 return 0;
 931
 932         return 1;
 933 }
 934
 935 int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
 936 {
 937         struct kprobe *cur = kprobe_running();
 938         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 939
 940         switch (kcb->kprobe_status) {
 941         case KPROBE_HIT_SS:
 942         case KPROBE_REENTER:
 943                 /*
 944                  * We are here because the instruction being single
 945                  * stepped caused a page fault. We reset the current
 946                  * kprobe and the ip points back to the probe address
 947                  * and allow the page fault handler to continue as a
 948                  * normal page fault.
 949                  */
 950                 regs->ip = (unsigned long)cur->addr;
 951                 regs->flags |= kcb->kprobe_old_flags;
 952                 if (kcb->kprobe_status == KPROBE_REENTER)
 953                         restore_previous_kprobe(kcb);
 954                 else
 955                         reset_current_kprobe();
 956                 preempt_enable_no_resched();
 957                 break;
 958         case KPROBE_HIT_ACTIVE:
 959         case KPROBE_HIT_SSDONE:
 960                 /*
 961                  * We increment the nmissed count for accounting,
 962                  * we can also use npre/npostfault count for accounting
 963                  * these specific fault cases.
 964                  */
 965                 kprobes_inc_nmissed_count(cur);
 966
 967                 /*
 968                  * We come here because instructions in the pre/post
 969                  * handler caused the page_fault, this could happen
 970                  * if handler tries to access user space by
 971                  * copy_from_user(), get_user() etc. Let the
 972                  * user-specified handler try to fix it first.
 973                  */
 974                 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
 975                         return 1;
 976
 977                 /*
 978                  * In case the user-specified fault handler returned
 979                  * zero, try to fix up.
 980                  */
 981                 if (fixup_exception(regs))
 982                         return 1;
 983
 984                 /*
 985                  * fixup routine could not handle it,
 986                  * Let do_page_fault() fix it.
 987                  */
 988                 break;
 989         default:
 990                 break;
 991         }
 992         return 0;
 993 }
 994
 995 /*
 996  * Wrapper routine for handling exceptions.
 997  */
 998 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
 999                                        unsigned long val, void *data)
1000 {
1001         struct die_args *args = data;
1002         int ret = NOTIFY_DONE;
1003
1004         if (args->regs && user_mode_vm(args->regs))
1005                 return ret;
1006
1007         switch (val) {
1008         case DIE_INT3:
1009                 if (kprobe_handler(args->regs))
1010                         ret = NOTIFY_STOP;
1011                 break;
1012         case DIE_DEBUG:
1013                 if (post_kprobe_handler(args->regs)) {
1014                         /*
1015                          * Reset the BS bit in dr6 (pointed by args->err) to
1016                          * denote completion of processing
1017                          */
1018                         (*(unsigned long *)ERR_PTR(args->err)) &= ~DR_STEP;
1019                         ret = NOTIFY_STOP;
1020                 }
1021                 break;
1022         case DIE_GPF:
1023                 /*
1024                  * To be potentially processing a kprobe fault and to
1025                  * trust the result from kprobe_running(), we have
1026                  * be non-preemptible.
1027                  */
1028                 if (!preemptible() && kprobe_running() &&
1029                     kprobe_fault_handler(args->regs, args->trapnr))
1030                         ret = NOTIFY_STOP;
1031                 break;
1032         default:
1033                 break;
1034         }
1035         return ret;
1036 }
1037
1038 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
1039 {
1040         struct jprobe *jp = container_of(p, struct jprobe, kp);
1041         unsigned long addr;
1042         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1043
1044         kcb->jprobe_saved_regs = *regs;
1045         kcb->jprobe_saved_sp = stack_addr(regs);
1046         addr = (unsigned long)(kcb->jprobe_saved_sp);
1047
1048         /*
1049          * As Linus pointed out, gcc assumes that the callee
1050          * owns the argument space and could overwrite it, e.g.
1051          * tailcall optimization. So, to be absolutely safe
1052          * we also save and restore enough stack bytes to cover
1053          * the argument area.
1054          */
1055         memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
1056                MIN_STACK_SIZE(addr));
1057         regs->flags &= ~X86_EFLAGS_IF;
1058         trace_hardirqs_off();
1059         regs->ip = (unsigned long)(jp->entry);
1060         return 1;
1061 }
1062
1063 void __kprobes jprobe_return(void)
1064 {
1065         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1066
1067         asm volatile (
1068 #ifdef CONFIG_X86_64
1069                         "       xchg   %%rbx,%%rsp      \n"
1070 #else
1071                         "       xchgl   %%ebx,%%esp     \n"
1072 #endif
1073                         "       int3                    \n"
1074                         "       .globl jprobe_return_end\n"
1075                         "       jprobe_return_end:      \n"
1076                         "       nop                     \n"::"b"
1077                         (kcb->jprobe_saved_sp):"memory");
1078 }
1079
1080 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
1081 {
1082         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1083         u8 *addr = (u8 *) (regs->ip - 1);
1084         struct jprobe *jp = container_of(p, struct jprobe, kp);
1085
1086         if ((addr > (u8 *) jprobe_return) &&
1087             (addr < (u8 *) jprobe_return_end)) {
1088                 if (stack_addr(regs) != kcb->jprobe_saved_sp) {
1089                         struct pt_regs *saved_regs = &kcb->jprobe_saved_regs;
1090                         printk(KERN_ERR
1091                                "current sp %p does not match saved sp %p\n",
1092                                stack_addr(regs), kcb->jprobe_saved_sp);
1093                         printk(KERN_ERR "Saved registers for jprobe %p\n", jp);
1094                         show_registers(saved_regs);
1095                         printk(KERN_ERR "Current registers\n");
1096                         show_registers(regs);
1097                         BUG();
1098                 }
1099                 *regs = kcb->jprobe_saved_regs;
1100                 memcpy((kprobe_opcode_t *)(kcb->jprobe_saved_sp),
1101                        kcb->jprobes_stack,
1102                        MIN_STACK_SIZE(kcb->jprobe_saved_sp));
1103                 preempt_enable_no_resched();
1104                 return 1;
1105         }
1106         return 0;
1107 }
1108
1109
1110 #ifdef CONFIG_OPTPROBES
1111
1112 /* Insert a call instruction at address 'from', which calls address 'to'.*/
1113 static void __kprobes synthesize_relcall(void *from, void *to)
1114 {
1115         __synthesize_relative_insn(from, to, RELATIVECALL_OPCODE);
1116 }
1117
1118 /* Insert a move instruction which sets a pointer to eax/rdi (1st arg). */
1119 static void __kprobes synthesize_set_arg1(kprobe_opcode_t *addr,
1120                                           unsigned long val)
1121 {
1122 #ifdef CONFIG_X86_64
1123         *addr++ = 0x48;
1124         *addr++ = 0xbf;
1125 #else
1126         *addr++ = 0xb8;
1127 #endif
1128         *(unsigned long *)addr = val;
1129 }
1130
1131 static void __used __kprobes kprobes_optinsn_template_holder(void)
1132 {
1133         asm volatile (
1134                         ".global optprobe_template_entry\n"
1135                         "optprobe_template_entry: \n"
1136 #ifdef CONFIG_X86_64
1137                         /* We don't bother saving the ss register */
1138                         "       pushq %rsp\n"
1139                         "       pushfq\n"
1140                         SAVE_REGS_STRING
1141                         "       movq %rsp, %rsi\n"
1142                         ".global optprobe_template_val\n"
1143                         "optprobe_template_val: \n"
1144                         ASM_NOP5
1145                         ASM_NOP5
1146                         ".global optprobe_template_call\n"
1147                         "optprobe_template_call: \n"
1148                         ASM_NOP5
1149                         /* Move flags to rsp */
1150                         "       movq 144(%rsp), %rdx\n"
1151                         "       movq %rdx, 152(%rsp)\n"
1152                         RESTORE_REGS_STRING
1153                         /* Skip flags entry */
1154                         "       addq $8, %rsp\n"
1155                         "       popfq\n"
1156 #else /* CONFIG_X86_32 */
1157                         "       pushf\n"
1158                         SAVE_REGS_STRING
1159                         "       movl %esp, %edx\n"
1160                         ".global optprobe_template_val\n"
1161                         "optprobe_template_val: \n"
1162                         ASM_NOP5
1163                         ".global optprobe_template_call\n"
1164                         "optprobe_template_call: \n"
1165                         ASM_NOP5
1166                         RESTORE_REGS_STRING
1167                         "       addl $4, %esp\n"        /* skip cs */
1168                         "       popf\n"
1169 #endif
1170                         ".global optprobe_template_end\n"
1171                         "optprobe_template_end: \n");
1172 }
1173
1174 #define TMPL_MOVE_IDX \
1175         ((long)&optprobe_template_val - (long)&optprobe_template_entry)
1176 #define TMPL_CALL_IDX \
1177         ((long)&optprobe_template_call - (long)&optprobe_template_entry)
1178 #define TMPL_END_IDX \
1179         ((long)&optprobe_template_end - (long)&optprobe_template_entry)
1180
1181 #define INT3_SIZE sizeof(kprobe_opcode_t)
1182
1183 /* Optimized kprobe call back function: called from optinsn */
1184 static void __kprobes optimized_callback(struct optimized_kprobe *op,
1185                                          struct pt_regs *regs)
1186 {
1187         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1188         unsigned long flags;
1189
1190         /* This is possible if op is under delayed unoptimizing */
1191         if (kprobe_disabled(&op->kp))
1192                 return;
1193
1194         local_irq_save(flags);
1195         if (kprobe_running()) {
1196                 kprobes_inc_nmissed_count(&op->kp);
1197         } else {
1198                 /* Save skipped registers */
1199 #ifdef CONFIG_X86_64
1200                 regs->cs = __KERNEL_CS;
1201 #else
1202                 regs->cs = __KERNEL_CS | get_kernel_rpl();
1203                 regs->gs = 0;
1204 #endif
1205                 regs->ip = (unsigned long)op->kp.addr + INT3_SIZE;
1206                 regs->orig_ax = ~0UL;
1207
1208                 __this_cpu_write(current_kprobe, &op->kp);
1209                 kcb->kprobe_status = KPROBE_HIT_ACTIVE;
1210                 opt_pre_handler(&op->kp, regs);
1211                 __this_cpu_write(current_kprobe, NULL);
1212         }
1213         local_irq_restore(flags);
1214 }
1215
1216 static int __kprobes copy_optimized_instructions(u8 *dest, u8 *src)
1217 {
1218         int len = 0, ret;
1219
1220         while (len < RELATIVEJUMP_SIZE) {
1221                 ret = __copy_instruction(dest + len, src + len, 1);
1222                 if (!ret || !can_boost(dest + len))
1223                         return -EINVAL;
1224                 len += ret;
1225         }
1226         /* Check whether the address range is reserved */
1227         if (ftrace_text_reserved(src, src + len - 1) ||
1228             alternatives_text_reserved(src, src + len - 1) ||
1229             jump_label_text_reserved(src, src + len - 1))
1230                 return -EBUSY;
1231
1232         return len;
1233 }
1234
1235 /* Check whether insn is indirect jump */
1236 static int __kprobes insn_is_indirect_jump(struct insn *insn)
1237 {
1238         return ((insn->opcode.bytes[0] == 0xff &&
1239                 (X86_MODRM_REG(insn->modrm.value) & 6) == 4) || /* Jump */
1240                 insn->opcode.bytes[0] == 0xea); /* Segment based jump */
1241 }
1242
1243 /* Check whether insn jumps into specified address range */
1244 static int insn_jump_into_range(struct insn *insn, unsigned long start, int len)
1245 {
1246         unsigned long target = 0;
1247
1248         switch (insn->opcode.bytes[0]) {
1249         case 0xe0:      /* loopne */
1250         case 0xe1:      /* loope */
1251         case 0xe2:      /* loop */
1252         case 0xe3:      /* jcxz */
1253         case 0xe9:      /* near relative jump */
1254         case 0xeb:      /* short relative jump */
1255                 break;
1256         case 0x0f:
1257                 if ((insn->opcode.bytes[1] & 0xf0) == 0x80) /* jcc near */
1258                         break;
1259                 return 0;
1260         default:
1261                 if ((insn->opcode.bytes[0] & 0xf0) == 0x70) /* jcc short */
1262                         break;
1263                 return 0;
1264         }
1265         target = (unsigned long)insn->next_byte + insn->immediate.value;
1266
1267         return (start <= target && target <= start + len);
1268 }
1269
1270 /* Decode whole function to ensure any instructions don't jump into target */
1271 static int __kprobes can_optimize(unsigned long paddr)
1272 {
1273         int ret;
1274         unsigned long addr, size = 0, offset = 0;
1275         struct insn insn;
1276         kprobe_opcode_t buf[MAX_INSN_SIZE];
1277
1278         /* Lookup symbol including addr */
1279         if (!kallsyms_lookup_size_offset(paddr, &size, &offset))
1280                 return 0;
1281
1282         /*
1283          * Do not optimize in the entry code due to the unstable
1284          * stack handling.
1285          */
1286         if ((paddr >= (unsigned long )__entry_text_start) &&
1287             (paddr <  (unsigned long )__entry_text_end))
1288                 return 0;
1289
1290         /* Check there is enough space for a relative jump. */
1291         if (size - offset < RELATIVEJUMP_SIZE)
1292                 return 0;
1293
1294         /* Decode instructions */
1295         addr = paddr - offset;
1296         while (addr < paddr - offset + size) { /* Decode until function end */
1297                 if (search_exception_tables(addr))
1298                         /*
1299                          * Since some fixup code will jumps into this function,
1300                          * we can't optimize kprobe in this function.
1301                          */
1302                         return 0;
1303                 kernel_insn_init(&insn, (void *)addr);
1304                 insn_get_opcode(&insn);
1305                 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION) {
1306                         ret = recover_probed_instruction(buf, addr);
1307                         if (ret)
1308                                 return 0;
1309                         kernel_insn_init(&insn, buf);
1310                 }
1311                 insn_get_length(&insn);
1312                 /* Recover address */
1313                 insn.kaddr = (void *)addr;
1314                 insn.next_byte = (void *)(addr + insn.length);
1315                 /* Check any instructions don't jump into target */
1316                 if (insn_is_indirect_jump(&insn) ||
1317                     insn_jump_into_range(&insn, paddr + INT3_SIZE,
1318                                          RELATIVE_ADDR_SIZE))
1319                         return 0;
1320                 addr += insn.length;
1321         }
1322
1323         return 1;
1324 }
1325
1326 /* Check optimized_kprobe can actually be optimized. */
1327 int __kprobes arch_check_optimized_kprobe(struct optimized_kprobe *op)
1328 {
1329         int i;
1330         struct kprobe *p;
1331
1332         for (i = 1; i < op->optinsn.size; i++) {
1333                 p = get_kprobe(op->kp.addr + i);
1334                 if (p && !kprobe_disabled(p))
1335                         return -EEXIST;
1336         }
1337
1338         return 0;
1339 }
1340
1341 /* Check the addr is within the optimized instructions. */
1342 int __kprobes arch_within_optimized_kprobe(struct optimized_kprobe *op,
1343                                            unsigned long addr)
1344 {
1345         return ((unsigned long)op->kp.addr <= addr &&
1346                 (unsigned long)op->kp.addr + op->optinsn.size > addr);
1347 }
1348
1349 /* Free optimized instruction slot */
1350 static __kprobes
1351 void __arch_remove_optimized_kprobe(struct optimized_kprobe *op, int dirty)
1352 {
1353         if (op->optinsn.insn) {
1354                 free_optinsn_slot(op->optinsn.insn, dirty);
1355                 op->optinsn.insn = NULL;
1356                 op->optinsn.size = 0;
1357         }
1358 }
1359
1360 void __kprobes arch_remove_optimized_kprobe(struct optimized_kprobe *op)
1361 {
1362         __arch_remove_optimized_kprobe(op, 1);
1363 }
1364
1365 /*
1366  * Copy replacing target instructions
1367  * Target instructions MUST be relocatable (checked inside)
1368  */
1369 int __kprobes arch_prepare_optimized_kprobe(struct optimized_kprobe *op)
1370 {
1371         u8 *buf;
1372         int ret;
1373         long rel;
1374
1375         if (!can_optimize((unsigned long)op->kp.addr))
1376                 return -EILSEQ;
1377
1378         op->optinsn.insn = get_optinsn_slot();
1379         if (!op->optinsn.insn)
1380                 return -ENOMEM;
1381
1382         /*
1383          * Verify if the address gap is in 2GB range, because this uses
1384          * a relative jump.
1385          */
1386         rel = (long)op->optinsn.insn - (long)op->kp.addr + RELATIVEJUMP_SIZE;
1387         if (abs(rel) > 0x7fffffff)
1388                 return -ERANGE;
1389
1390         buf = (u8 *)op->optinsn.insn;
1391
1392         /* Copy instructions into the out-of-line buffer */
1393         ret = copy_optimized_instructions(buf + TMPL_END_IDX, op->kp.addr);
1394         if (ret < 0) {
1395                 __arch_remove_optimized_kprobe(op, 0);
1396                 return ret;
1397         }
1398         op->optinsn.size = ret;
1399
1400         /* Copy arch-dep-instance from template */
1401         memcpy(buf, &optprobe_template_entry, TMPL_END_IDX);
1402
1403         /* Set probe information */
1404         synthesize_set_arg1(buf + TMPL_MOVE_IDX, (unsigned long)op);
1405
1406         /* Set probe function call */
1407         synthesize_relcall(buf + TMPL_CALL_IDX, optimized_callback);
1408
1409         /* Set returning jmp instruction at the tail of out-of-line buffer */
1410         synthesize_reljump(buf + TMPL_END_IDX + op->optinsn.size,
1411                            (u8 *)op->kp.addr + op->optinsn.size);
1412
1413         flush_icache_range((unsigned long) buf,
1414                            (unsigned long) buf + TMPL_END_IDX +
1415                            op->optinsn.size + RELATIVEJUMP_SIZE);
1416         return 0;
1417 }
1418
1419 #define MAX_OPTIMIZE_PROBES 256
1420 static struct text_poke_param *jump_poke_params;
1421 static struct jump_poke_buffer {
1422         u8 buf[RELATIVEJUMP_SIZE];
1423 } *jump_poke_bufs;
1424
1425 static void __kprobes setup_optimize_kprobe(struct text_poke_param *tprm,
1426                                             u8 *insn_buf,
1427                                             struct optimized_kprobe *op)
1428 {
1429         s32 rel = (s32)((long)op->optinsn.insn -
1430                         ((long)op->kp.addr + RELATIVEJUMP_SIZE));
1431
1432         /* Backup instructions which will be replaced by jump address */
1433         memcpy(op->optinsn.copied_insn, op->kp.addr + INT3_SIZE,
1434                RELATIVE_ADDR_SIZE);
1435
1436         insn_buf[0] = RELATIVEJUMP_OPCODE;
1437         *(s32 *)(&insn_buf[1]) = rel;
1438
1439         tprm->addr = op->kp.addr;
1440         tprm->opcode = insn_buf;
1441         tprm->len = RELATIVEJUMP_SIZE;
1442 }
1443
1444 /*
1445  * Replace breakpoints (int3) with relative jumps.
1446  * Caller must call with locking kprobe_mutex and text_mutex.
1447  */
1448 void __kprobes arch_optimize_kprobes(struct list_head *oplist)
1449 {
1450         struct optimized_kprobe *op, *tmp;
1451         int c = 0;
1452
1453         list_for_each_entry_safe(op, tmp, oplist, list) {
1454                 WARN_ON(kprobe_disabled(&op->kp));
1455                 /* Setup param */
1456                 setup_optimize_kprobe(&jump_poke_params[c],
1457                                       jump_poke_bufs[c].buf, op);
1458                 list_del_init(&op->list);
1459                 if (++c >= MAX_OPTIMIZE_PROBES)
1460                         break;
1461         }
1462
1463         /*
1464          * text_poke_smp doesn't support NMI/MCE code modifying.
1465          * However, since kprobes itself also doesn't support NMI/MCE
1466          * code probing, it's not a problem.
1467          */
1468         text_poke_smp_batch(jump_poke_params, c);
1469 }
1470
1471 static void __kprobes setup_unoptimize_kprobe(struct text_poke_param *tprm,
1472                                               u8 *insn_buf,
1473                                               struct optimized_kprobe *op)
1474 {
1475         /* Set int3 to first byte for kprobes */
1476         insn_buf[0] = BREAKPOINT_INSTRUCTION;
1477         memcpy(insn_buf + 1, op->optinsn.copied_insn, RELATIVE_ADDR_SIZE);
1478
1479         tprm->addr = op->kp.addr;
1480         tprm->opcode = insn_buf;
1481         tprm->len = RELATIVEJUMP_SIZE;
1482 }
1483
1484 /*
1485  * Recover original instructions and breakpoints from relative jumps.
1486  * Caller must call with locking kprobe_mutex.
1487  */
1488 extern void arch_unoptimize_kprobes(struct list_head *oplist,
1489                                     struct list_head *done_list)
1490 {
1491         struct optimized_kprobe *op, *tmp;
1492         int c = 0;
1493
1494         list_for_each_entry_safe(op, tmp, oplist, list) {
1495                 /* Setup param */
1496                 setup_unoptimize_kprobe(&jump_poke_params[c],
1497                                         jump_poke_bufs[c].buf, op);
1498                 list_move(&op->list, done_list);
1499                 if (++c >= MAX_OPTIMIZE_PROBES)
1500                         break;
1501         }
1502
1503         /*
1504          * text_poke_smp doesn't support NMI/MCE code modifying.
1505          * However, since kprobes itself also doesn't support NMI/MCE
1506          * code probing, it's not a problem.
1507          */
1508         text_poke_smp_batch(jump_poke_params, c);
1509 }
1510
1511 /* Replace a relative jump with a breakpoint (int3).  */
1512 void __kprobes arch_unoptimize_kprobe(struct optimized_kprobe *op)
1513 {
1514         u8 buf[RELATIVEJUMP_SIZE];
1515
1516         /* Set int3 to first byte for kprobes */
1517         buf[0] = BREAKPOINT_INSTRUCTION;
1518         memcpy(buf + 1, op->optinsn.copied_insn, RELATIVE_ADDR_SIZE);
1519         text_poke_smp(op->kp.addr, buf, RELATIVEJUMP_SIZE);
1520 }
1521
1522 static int  __kprobes setup_detour_execution(struct kprobe *p,
1523                                              struct pt_regs *regs,
1524                                              int reenter)
1525 {
1526         struct optimized_kprobe *op;
1527
1528         if (p->flags & KPROBE_FLAG_OPTIMIZED) {
1529                 /* This kprobe is really able to run optimized path. */
1530                 op = container_of(p, struct optimized_kprobe, kp);
1531                 /* Detour through copied instructions */
1532                 regs->ip = (unsigned long)op->optinsn.insn + TMPL_END_IDX;
1533                 if (!reenter)
1534                         reset_current_kprobe();
1535                 preempt_enable_no_resched();
1536                 return 1;
1537         }
1538         return 0;
1539 }
1540
1541 static int __kprobes init_poke_params(void)
1542 {
1543         /* Allocate code buffer and parameter array */
1544         jump_poke_bufs = kmalloc(sizeof(struct jump_poke_buffer) *
1545                                  MAX_OPTIMIZE_PROBES, GFP_KERNEL);
1546         if (!jump_poke_bufs)
1547                 return -ENOMEM;
1548
1549         jump_poke_params = kmalloc(sizeof(struct text_poke_param) *
1550                                    MAX_OPTIMIZE_PROBES, GFP_KERNEL);
1551         if (!jump_poke_params) {
1552                 kfree(jump_poke_bufs);
1553                 jump_poke_bufs = NULL;
1554                 return -ENOMEM;
1555         }
1556
1557         return 0;
1558 }
1559 #else   /* !CONFIG_OPTPROBES */
1560 static int __kprobes init_poke_params(void)
1561 {
1562         return 0;
1563 }
1564 #endif
1565
1566 int __init arch_init_kprobes(void)
1567 {
1568         return init_poke_params();
1569 }
1570
1571 int __kprobes arch_trampoline_kprobe(struct kprobe *p)
1572 {
1573         return 0;
1574 }