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 and volatile to keep gcc from statically
  79          * optimizing it out, as variable_test_bit makes gcc think only
  80          * *(unsigned long*) is used.
  81          */
  82 static volatile u32 twobyte_is_boostable[256 / 32] = {
  83         /*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f          */
  84         /*      ----------------------------------------------          */
  85         W(0x00, 0, 0, 1, 1, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0) | /* 00 */
  86         W(0x10, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 10 */
  87         W(0x20, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 20 */
  88         W(0x30, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 30 */
  89         W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
  90         W(0x50, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 50 */
  91         W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1) | /* 60 */
  92         W(0x70, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1) , /* 70 */
  93         W(0x80, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 80 */
  94         W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
  95         W(0xa0, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* a0 */
  96         W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1) , /* b0 */
  97         W(0xc0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1) | /* c0 */
  98         W(0xd0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) , /* d0 */
  99         W(0xe0, 0, 1, 1, 0, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* e0 */
 100         W(0xf0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 0)   /* f0 */
 101         /*      -----------------------------------------------         */
 102         /*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f          */
 103 };
 104 #undef W
 105
 106 struct kretprobe_blackpoint kretprobe_blacklist[] = {
 107         {"__switch_to", }, /* This function switches only current task, but
 108                               doesn't switch kernel stack.*/
 109         {NULL, NULL}    /* Terminator */
 110 };
 111 const int kretprobe_blacklist_size = ARRAY_SIZE(kretprobe_blacklist);
 112
 113 static void __kprobes __synthesize_relative_insn(void *from, void *to, u8 op)
 114 {
 115         struct __arch_relative_insn {
 116                 u8 op;
 117                 s32 raddr;
 118         } __attribute__((packed)) *insn;
 119
 120         insn = (struct __arch_relative_insn *)from;
 121         insn->raddr = (s32)((long)(to) - ((long)(from) + 5));
 122         insn->op = op;
 123 }
 124
 125 /* Insert a jump instruction at address 'from', which jumps to address 'to'.*/
 126 static void __kprobes synthesize_reljump(void *from, void *to)
 127 {
 128         __synthesize_relative_insn(from, to, RELATIVEJUMP_OPCODE);
 129 }
 130
 131 /*
 132  * Skip the prefixes of the instruction.
 133  */
 134 static kprobe_opcode_t *__kprobes skip_prefixes(kprobe_opcode_t *insn)
 135 {
 136         insn_attr_t attr;
 137
 138         attr = inat_get_opcode_attribute((insn_byte_t)*insn);
 139         while (inat_is_legacy_prefix(attr)) {
 140                 insn++;
 141                 attr = inat_get_opcode_attribute((insn_byte_t)*insn);
 142         }
 143 #ifdef CONFIG_X86_64
 144         if (inat_is_rex_prefix(attr))
 145                 insn++;
 146 #endif
 147         return insn;
 148 }
 149
 150 /*
 151  * Returns non-zero if opcode is boostable.
 152  * RIP relative instructions are adjusted at copying time in 64 bits mode
 153  */
 154 static int __kprobes can_boost(kprobe_opcode_t *opcodes)
 155 {
 156         kprobe_opcode_t opcode;
 157         kprobe_opcode_t *orig_opcodes = opcodes;
 158
 159         if (search_exception_tables((unsigned long)opcodes))
 160                 return 0;       /* Page fault may occur on this address. */
 161
 162 retry:
 163         if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
 164                 return 0;
 165         opcode = *(opcodes++);
 166
 167         /* 2nd-byte opcode */
 168         if (opcode == 0x0f) {
 169                 if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
 170                         return 0;
 171                 return test_bit(*opcodes,
 172                                 (unsigned long *)twobyte_is_boostable);
 173         }
 174
 175         switch (opcode & 0xf0) {
 176 #ifdef CONFIG_X86_64
 177         case 0x40:
 178                 goto retry; /* REX prefix is boostable */
 179 #endif
 180         case 0x60:
 181                 if (0x63 < opcode && opcode < 0x67)
 182                         goto retry; /* prefixes */
 183                 /* can't boost Address-size override and bound */
 184                 return (opcode != 0x62 && opcode != 0x67);
 185         case 0x70:
 186                 return 0; /* can't boost conditional jump */
 187         case 0xc0:
 188                 /* can't boost software-interruptions */
 189                 return (0xc1 < opcode && opcode < 0xcc) || opcode == 0xcf;
 190         case 0xd0:
 191                 /* can boost AA* and XLAT */
 192                 return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
 193         case 0xe0:
 194                 /* can boost in/out and absolute jmps */
 195                 return ((opcode & 0x04) || opcode == 0xea);
 196         case 0xf0:
 197                 if ((opcode & 0x0c) == 0 && opcode != 0xf1)
 198                         goto retry; /* lock/rep(ne) prefix */
 199                 /* clear and set flags are boostable */
 200                 return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
 201         default:
 202                 /* segment override prefixes are boostable */
 203                 if (opcode == 0x26 || opcode == 0x36 || opcode == 0x3e)
 204                         goto retry; /* prefixes */
 205                 /* CS override prefix and call are not boostable */
 206                 return (opcode != 0x2e && opcode != 0x9a);
 207         }
 208 }
 209
 210 /* Recover the probed instruction at addr for further analysis. */
 211 static int recover_probed_instruction(kprobe_opcode_t *buf, unsigned long addr)
 212 {
 213         struct kprobe *kp;
 214         kp = get_kprobe((void *)addr);
 215         if (!kp)
 216                 return -EINVAL;
 217
 218         /*
 219          *  Basically, kp->ainsn.insn has an original instruction.
 220          *  However, RIP-relative instruction can not do single-stepping
 221          *  at different place, __copy_instruction() tweaks the displacement of
 222          *  that instruction. In that case, we can't recover the instruction
 223          *  from the kp->ainsn.insn.
 224          *
 225          *  On the other hand, kp->opcode has a copy of the first byte of
 226          *  the probed instruction, which is overwritten by int3. And
 227          *  the instruction at kp->addr is not modified by kprobes except
 228          *  for the first byte, we can recover the original instruction
 229          *  from it and kp->opcode.
 230          */
 231         memcpy(buf, kp->addr, MAX_INSN_SIZE * sizeof(kprobe_opcode_t));
 232         buf[0] = kp->opcode;
 233         return 0;
 234 }
 235
 236 /* Check if paddr is at an instruction boundary */
 237 static int __kprobes can_probe(unsigned long paddr)
 238 {
 239         int ret;
 240         unsigned long addr, offset = 0;
 241         struct insn insn;
 242         kprobe_opcode_t buf[MAX_INSN_SIZE];
 243
 244         if (!kallsyms_lookup_size_offset(paddr, NULL, &offset))
 245                 return 0;
 246
 247         /* Decode instructions */
 248         addr = paddr - offset;
 249         while (addr < paddr) {
 250                 kernel_insn_init(&insn, (void *)addr);
 251                 insn_get_opcode(&insn);
 252
 253                 /*
 254                  * Check if the instruction has been modified by another
 255                  * kprobe, in which case we replace the breakpoint by the
 256                  * original instruction in our buffer.
 257                  */
 258                 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION) {
 259                         ret = recover_probed_instruction(buf, addr);
 260                         if (ret)
 261                                 /*
 262                                  * Another debugging subsystem might insert
 263                                  * this breakpoint. In that case, we can't
 264                                  * recover it.
 265                                  */
 266                                 return 0;
 267                         kernel_insn_init(&insn, buf);
 268                 }
 269                 insn_get_length(&insn);
 270                 addr += insn.length;
 271         }
 272
 273         return (addr == paddr);
 274 }
 275
 276 /*
 277  * Returns non-zero if opcode modifies the interrupt flag.
 278  */
 279 static int __kprobes is_IF_modifier(kprobe_opcode_t *insn)
 280 {
 281         /* Skip prefixes */
 282         insn = skip_prefixes(insn);
 283
 284         switch (*insn) {
 285         case 0xfa:              /* cli */
 286         case 0xfb:              /* sti */
 287         case 0xcf:              /* iret/iretd */
 288         case 0x9d:              /* popf/popfd */
 289                 return 1;
 290         }
 291
 292         return 0;
 293 }
 294
 295 /*
 296  * Copy an instruction and adjust the displacement if the instruction
 297  * uses the %rip-relative addressing mode.
 298  * If it does, Return the address of the 32-bit displacement word.
 299  * If not, return null.
 300  * Only applicable to 64-bit x86.
 301  */
 302 static int __kprobes __copy_instruction(u8 *dest, u8 *src, int recover)
 303 {
 304         struct insn insn;
 305         int ret;
 306         kprobe_opcode_t buf[MAX_INSN_SIZE];
 307
 308         kernel_insn_init(&insn, src);
 309         if (recover) {
 310                 insn_get_opcode(&insn);
 311                 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION) {
 312                         ret = recover_probed_instruction(buf,
 313                                                          (unsigned long)src);
 314                         if (ret)
 315                                 return 0;
 316                         kernel_insn_init(&insn, buf);
 317                 }
 318         }
 319         insn_get_length(&insn);
 320         memcpy(dest, insn.kaddr, insn.length);
 321
 322 #ifdef CONFIG_X86_64
 323         if (insn_rip_relative(&insn)) {
 324                 s64 newdisp;
 325                 u8 *disp;
 326                 kernel_insn_init(&insn, dest);
 327                 insn_get_displacement(&insn);
 328                 /*
 329                  * The copied instruction uses the %rip-relative addressing
 330                  * mode.  Adjust the displacement for the difference between
 331                  * the original location of this instruction and the location
 332                  * of the copy that will actually be run.  The tricky bit here
 333                  * is making sure that the sign extension happens correctly in
 334                  * this calculation, since we need a signed 32-bit result to
 335                  * be sign-extended to 64 bits when it's added to the %rip
 336                  * value and yield the same 64-bit result that the sign-
 337                  * extension of the original signed 32-bit displacement would
 338                  * have given.
 339                  */
 340                 newdisp = (u8 *) src + (s64) insn.displacement.value -
 341                           (u8 *) dest;
 342                 BUG_ON((s64) (s32) newdisp != newdisp); /* Sanity check.  */
 343                 disp = (u8 *) dest + insn_offset_displacement(&insn);
 344                 *(s32 *) disp = (s32) newdisp;
 345         }
 346 #endif
 347         return insn.length;
 348 }
 349
 350 static void __kprobes arch_copy_kprobe(struct kprobe *p)
 351 {
 352         /*
 353          * Copy an instruction without recovering int3, because it will be
 354          * put by another subsystem.
 355          */
 356         __copy_instruction(p->ainsn.insn, p->addr, 0);
 357
 358         if (can_boost(p->addr))
 359                 p->ainsn.boostable = 0;
 360         else
 361                 p->ainsn.boostable = -1;
 362
 363         p->opcode = *p->addr;
 364 }
 365
 366 int __kprobes arch_prepare_kprobe(struct kprobe *p)
 367 {
 368         if (alternatives_text_reserved(p->addr, p->addr))
 369                 return -EINVAL;
 370
 371         if (!can_probe((unsigned long)p->addr))
 372                 return -EILSEQ;
 373         /* insn: must be on special executable page on x86. */
 374         p->ainsn.insn = get_insn_slot();
 375         if (!p->ainsn.insn)
 376                 return -ENOMEM;
 377         arch_copy_kprobe(p);
 378         return 0;
 379 }
 380
 381 void __kprobes arch_arm_kprobe(struct kprobe *p)
 382 {
 383         text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
 384 }
 385
 386 void __kprobes arch_disarm_kprobe(struct kprobe *p)
 387 {
 388         text_poke(p->addr, &p->opcode, 1);
 389 }
 390
 391 void __kprobes arch_remove_kprobe(struct kprobe *p)
 392 {
 393         if (p->ainsn.insn) {
 394                 free_insn_slot(p->ainsn.insn, (p->ainsn.boostable == 1));
 395                 p->ainsn.insn = NULL;
 396         }
 397 }
 398
 399 static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
 400 {
 401         kcb->prev_kprobe.kp = kprobe_running();
 402         kcb->prev_kprobe.status = kcb->kprobe_status;
 403         kcb->prev_kprobe.old_flags = kcb->kprobe_old_flags;
 404         kcb->prev_kprobe.saved_flags = kcb->kprobe_saved_flags;
 405 }
 406
 407 static void __kprobes restore_previous_kprobe(struct kprobe_ctlblk *kcb)
 408 {
 409         __this_cpu_write(current_kprobe, kcb->prev_kprobe.kp);
 410         kcb->kprobe_status = kcb->prev_kprobe.status;
 411         kcb->kprobe_old_flags = kcb->prev_kprobe.old_flags;
 412         kcb->kprobe_saved_flags = kcb->prev_kprobe.saved_flags;
 413 }
 414
 415 static void __kprobes set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
 416                                 struct kprobe_ctlblk *kcb)
 417 {
 418         __this_cpu_write(current_kprobe, p);
 419         kcb->kprobe_saved_flags = kcb->kprobe_old_flags
 420                 = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
 421         if (is_IF_modifier(p->ainsn.insn))
 422                 kcb->kprobe_saved_flags &= ~X86_EFLAGS_IF;
 423 }
 424
 425 static void __kprobes clear_btf(void)
 426 {
 427         if (test_thread_flag(TIF_BLOCKSTEP)) {
 428                 unsigned long debugctl = get_debugctlmsr();
 429
 430                 debugctl &= ~DEBUGCTLMSR_BTF;
 431                 update_debugctlmsr(debugctl);
 432         }
 433 }
 434
 435 static void __kprobes restore_btf(void)
 436 {
 437         if (test_thread_flag(TIF_BLOCKSTEP)) {
 438                 unsigned long debugctl = get_debugctlmsr();
 439
 440                 debugctl |= DEBUGCTLMSR_BTF;
 441                 update_debugctlmsr(debugctl);
 442         }
 443 }
 444
 445 void __kprobes arch_prepare_kretprobe(struct kretprobe_instance *ri,
 446                                       struct pt_regs *regs)
 447 {
 448         unsigned long *sara = stack_addr(regs);
 449
 450         ri->ret_addr = (kprobe_opcode_t *) *sara;
 451
 452         /* Replace the return addr with trampoline addr */
 453         *sara = (unsigned long) &kretprobe_trampoline;
 454 }
 455
 456 #ifdef CONFIG_OPTPROBES
 457 static int  __kprobes setup_detour_execution(struct kprobe *p,
 458                                              struct pt_regs *regs,
 459                                              int reenter);
 460 #else
 461 #define setup_detour_execution(p, regs, reenter) (0)
 462 #endif
 463
 464 static void __kprobes setup_singlestep(struct kprobe *p, struct pt_regs *regs,
 465                                        struct kprobe_ctlblk *kcb, int reenter)
 466 {
 467         if (setup_detour_execution(p, regs, reenter))
 468                 return;
 469
 470 #if !defined(CONFIG_PREEMPT)
 471         if (p->ainsn.boostable == 1 && !p->post_handler) {
 472                 /* Boost up -- we can execute copied instructions directly */
 473                 if (!reenter)
 474                         reset_current_kprobe();
 475                 /*
 476                  * Reentering boosted probe doesn't reset current_kprobe,
 477                  * nor set current_kprobe, because it doesn't use single
 478                  * stepping.
 479                  */
 480                 regs->ip = (unsigned long)p->ainsn.insn;
 481                 preempt_enable_no_resched();
 482                 return;
 483         }
 484 #endif
 485         if (reenter) {
 486                 save_previous_kprobe(kcb);
 487                 set_current_kprobe(p, regs, kcb);
 488                 kcb->kprobe_status = KPROBE_REENTER;
 489         } else
 490                 kcb->kprobe_status = KPROBE_HIT_SS;
 491         /* Prepare real single stepping */
 492         clear_btf();
 493         regs->flags |= X86_EFLAGS_TF;
 494         regs->flags &= ~X86_EFLAGS_IF;
 495         /* single step inline if the instruction is an int3 */
 496         if (p->opcode == BREAKPOINT_INSTRUCTION)
 497                 regs->ip = (unsigned long)p->addr;
 498         else
 499                 regs->ip = (unsigned long)p->ainsn.insn;
 500 }
 501
 502 /*
 503  * We have reentered the kprobe_handler(), since another probe was hit while
 504  * within the handler. We save the original kprobes variables and just single
 505  * step on the instruction of the new probe without calling any user handlers.
 506  */
 507 static int __kprobes reenter_kprobe(struct kprobe *p, struct pt_regs *regs,
 508                                     struct kprobe_ctlblk *kcb)
 509 {
 510         switch (kcb->kprobe_status) {
 511         case KPROBE_HIT_SSDONE:
 512         case KPROBE_HIT_ACTIVE:
 513                 kprobes_inc_nmissed_count(p);
 514                 setup_singlestep(p, regs, kcb, 1);
 515                 break;
 516         case KPROBE_HIT_SS:
 517                 /* A probe has been hit in the codepath leading up to, or just
 518                  * after, single-stepping of a probed instruction. This entire
 519                  * codepath should strictly reside in .kprobes.text section.
 520                  * Raise a BUG or we'll continue in an endless reentering loop
 521                  * and eventually a stack overflow.
 522                  */
 523                 printk(KERN_WARNING "Unrecoverable kprobe detected at %p.\n",
 524                        p->addr);
 525                 dump_kprobe(p);
 526                 BUG();
 527         default:
 528                 /* impossible cases */
 529                 WARN_ON(1);
 530                 return 0;
 531         }
 532
 533         return 1;
 534 }
 535
 536 /*
 537  * Interrupts are disabled on entry as trap3 is an interrupt gate and they
 538  * remain disabled throughout this function.
 539  */
 540 static int __kprobes kprobe_handler(struct pt_regs *regs)
 541 {
 542         kprobe_opcode_t *addr;
 543         struct kprobe *p;
 544         struct kprobe_ctlblk *kcb;
 545
 546         addr = (kprobe_opcode_t *)(regs->ip - sizeof(kprobe_opcode_t));
 547         /*
 548          * We don't want to be preempted for the entire
 549          * duration of kprobe processing. We conditionally
 550          * re-enable preemption at the end of this function,
 551          * and also in reenter_kprobe() and setup_singlestep().
 552          */
 553         preempt_disable();
 554
 555         kcb = get_kprobe_ctlblk();
 556         p = get_kprobe(addr);
 557
 558         if (p) {
 559                 if (kprobe_running()) {
 560                         if (reenter_kprobe(p, regs, kcb))
 561                                 return 1;
 562                 } else {
 563                         set_current_kprobe(p, regs, kcb);
 564                         kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 565
 566                         /*
 567                          * If we have no pre-handler or it returned 0, we
 568                          * continue with normal processing.  If we have a
 569                          * pre-handler and it returned non-zero, it prepped
 570                          * for calling the break_handler below on re-entry
 571                          * for jprobe processing, so get out doing nothing
 572                          * more here.
 573                          */
 574                         if (!p->pre_handler || !p->pre_handler(p, regs))
 575                                 setup_singlestep(p, regs, kcb, 0);
 576                         return 1;
 577                 }
 578         } else if (*addr != BREAKPOINT_INSTRUCTION) {
 579                 /*
 580                  * The breakpoint instruction was removed right
 581                  * after we hit it.  Another cpu has removed
 582                  * either a probepoint or a debugger breakpoint
 583                  * at this address.  In either case, no further
 584                  * handling of this interrupt is appropriate.
 585                  * Back up over the (now missing) int3 and run
 586                  * the original instruction.
 587                  */
 588                 regs->ip = (unsigned long)addr;
 589                 preempt_enable_no_resched();
 590                 return 1;
 591         } else if (kprobe_running()) {
 592                 p = __this_cpu_read(current_kprobe);
 593                 if (p->break_handler && p->break_handler(p, regs)) {
 594                         setup_singlestep(p, regs, kcb, 0);
 595                         return 1;
 596                 }
 597         } /* else: not a kprobe fault; let the kernel handle it */
 598
 599         preempt_enable_no_resched();
 600         return 0;
 601 }
 602
 603 #ifdef CONFIG_X86_64
 604 #define SAVE_REGS_STRING                \
 605         /* Skip cs, ip, orig_ax. */     \
 606         "       subq $24, %rsp\n"       \
 607         "       pushq %rdi\n"           \
 608         "       pushq %rsi\n"           \
 609         "       pushq %rdx\n"           \
 610         "       pushq %rcx\n"           \
 611         "       pushq %rax\n"           \
 612         "       pushq %r8\n"            \
 613         "       pushq %r9\n"            \
 614         "       pushq %r10\n"           \
 615         "       pushq %r11\n"           \
 616         "       pushq %rbx\n"           \
 617         "       pushq %rbp\n"           \
 618         "       pushq %r12\n"           \
 619         "       pushq %r13\n"           \
 620         "       pushq %r14\n"           \
 621         "       pushq %r15\n"
 622 #define RESTORE_REGS_STRING             \
 623         "       popq %r15\n"            \
 624         "       popq %r14\n"            \
 625         "       popq %r13\n"            \
 626         "       popq %r12\n"            \
 627         "       popq %rbp\n"            \
 628         "       popq %rbx\n"            \
 629         "       popq %r11\n"            \
 630         "       popq %r10\n"            \
 631         "       popq %r9\n"             \
 632         "       popq %r8\n"             \
 633         "       popq %rax\n"            \
 634         "       popq %rcx\n"            \
 635         "       popq %rdx\n"            \
 636         "       popq %rsi\n"            \
 637         "       popq %rdi\n"            \
 638         /* Skip orig_ax, ip, cs */      \
 639         "       addq $24, %rsp\n"
 640 #else
 641 #define SAVE_REGS_STRING                \
 642         /* Skip cs, ip, orig_ax and gs. */      \
 643         "       subl $16, %esp\n"       \
 644         "       pushl %fs\n"            \
 645         "       pushl %es\n"            \
 646         "       pushl %ds\n"            \
 647         "       pushl %eax\n"           \
 648         "       pushl %ebp\n"           \
 649         "       pushl %edi\n"           \
 650         "       pushl %esi\n"           \
 651         "       pushl %edx\n"           \
 652         "       pushl %ecx\n"           \
 653         "       pushl %ebx\n"
 654 #define RESTORE_REGS_STRING             \
 655         "       popl %ebx\n"            \
 656         "       popl %ecx\n"            \
 657         "       popl %edx\n"            \
 658         "       popl %esi\n"            \
 659         "       popl %edi\n"            \
 660         "       popl %ebp\n"            \
 661         "       popl %eax\n"            \
 662         /* Skip ds, es, fs, gs, orig_ax, and ip. Note: don't pop cs here*/\
 663         "       addl $24, %esp\n"
 664 #endif
 665
 666 /*
 667  * When a retprobed function returns, this code saves registers and
 668  * calls trampoline_handler() runs, which calls the kretprobe's handler.
 669  */
 670 static void __used __kprobes kretprobe_trampoline_holder(void)
 671 {
 672         asm volatile (
 673                         ".global kretprobe_trampoline\n"
 674                         "kretprobe_trampoline: \n"
 675 #ifdef CONFIG_X86_64
 676                         /* We don't bother saving the ss register */
 677                         "       pushq %rsp\n"
 678                         "       pushfq\n"
 679                         SAVE_REGS_STRING
 680                         "       movq %rsp, %rdi\n"
 681                         "       call trampoline_handler\n"
 682                         /* Replace saved sp with true return address. */
 683                         "       movq %rax, 152(%rsp)\n"
 684                         RESTORE_REGS_STRING
 685                         "       popfq\n"
 686 #else
 687                         "       pushf\n"
 688                         SAVE_REGS_STRING
 689                         "       movl %esp, %eax\n"
 690                         "       call trampoline_handler\n"
 691                         /* Move flags to cs */
 692                         "       movl 56(%esp), %edx\n"
 693                         "       movl %edx, 52(%esp)\n"
 694                         /* Replace saved flags with true return address. */
 695                         "       movl %eax, 56(%esp)\n"
 696                         RESTORE_REGS_STRING
 697                         "       popf\n"
 698 #endif
 699                         "       ret\n");
 700 }
 701
 702 /*
 703  * Called from kretprobe_trampoline
 704  */
 705 static __used __kprobes void *trampoline_handler(struct pt_regs *regs)
 706 {
 707         struct kretprobe_instance *ri = NULL;
 708         struct hlist_head *head, empty_rp;
 709         struct hlist_node *node, *tmp;
 710         unsigned long flags, orig_ret_address = 0;
 711         unsigned long trampoline_address = (unsigned long)&kretprobe_trampoline;
 712         kprobe_opcode_t *correct_ret_addr = NULL;
 713
 714         INIT_HLIST_HEAD(&empty_rp);
 715         kretprobe_hash_lock(current, &head, &flags);
 716         /* fixup registers */
 717 #ifdef CONFIG_X86_64
 718         regs->cs = __KERNEL_CS;
 719 #else
 720         regs->cs = __KERNEL_CS | get_kernel_rpl();
 721         regs->gs = 0;
 722 #endif
 723         regs->ip = trampoline_address;
 724         regs->orig_ax = ~0UL;
 725
 726         /*
 727          * It is possible to have multiple instances associated with a given
 728          * task either because multiple functions in the call path have
 729          * return probes installed on them, and/or more than one
 730          * return probe was registered for a target function.
 731          *
 732          * We can handle this because:
 733          *     - instances are always pushed into the head of the list
 734          *     - when multiple return probes are registered for the same
 735          *       function, the (chronologically) first instance's ret_addr
 736          *       will be the real return address, and all the rest will
 737          *       point to kretprobe_trampoline.
 738          */
 739         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
 740                 if (ri->task != current)
 741                         /* another task is sharing our hash bucket */
 742                         continue;
 743
 744                 orig_ret_address = (unsigned long)ri->ret_addr;
 745
 746                 if (orig_ret_address != trampoline_address)
 747                         /*
 748                          * This is the real return address. Any other
 749                          * instances associated with this task are for
 750                          * other calls deeper on the call stack
 751                          */
 752                         break;
 753         }
 754
 755         kretprobe_assert(ri, orig_ret_address, trampoline_address);
 756
 757         correct_ret_addr = ri->ret_addr;
 758         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
 759                 if (ri->task != current)
 760                         /* another task is sharing our hash bucket */
 761                         continue;
 762
 763                 orig_ret_address = (unsigned long)ri->ret_addr;
 764                 if (ri->rp && ri->rp->handler) {
 765                         __this_cpu_write(current_kprobe, &ri->rp->kp);
 766                         get_kprobe_ctlblk()->kprobe_status = KPROBE_HIT_ACTIVE;
 767                         ri->ret_addr = correct_ret_addr;
 768                         ri->rp->handler(ri, regs);
 769                         __this_cpu_write(current_kprobe, NULL);
 770                 }
 771
 772                 recycle_rp_inst(ri, &empty_rp);
 773
 774                 if (orig_ret_address != trampoline_address)
 775                         /*
 776                          * This is the real return address. Any other
 777                          * instances associated with this task are for
 778                          * other calls deeper on the call stack
 779                          */
 780                         break;
 781         }
 782
 783         kretprobe_hash_unlock(current, &flags);
 784
 785         hlist_for_each_entry_safe(ri, node, tmp, &empty_rp, hlist) {
 786                 hlist_del(&ri->hlist);
 787                 kfree(ri);
 788         }
 789         return (void *)orig_ret_address;
 790 }
 791
 792 /*
 793  * Called after single-stepping.  p->addr is the address of the
 794  * instruction whose first byte has been replaced by the "int 3"
 795  * instruction.  To avoid the SMP problems that can occur when we
 796  * temporarily put back the original opcode to single-step, we
 797  * single-stepped a copy of the instruction.  The address of this
 798  * copy is p->ainsn.insn.
 799  *
 800  * This function prepares to return from the post-single-step
 801  * interrupt.  We have to fix up the stack as follows:
 802  *
 803  * 0) Except in the case of absolute or indirect jump or call instructions,
 804  * the new ip is relative to the copied instruction.  We need to make
 805  * it relative to the original instruction.
 806  *
 807  * 1) If the single-stepped instruction was pushfl, then the TF and IF
 808  * flags are set in the just-pushed flags, and may need to be cleared.
 809  *
 810  * 2) If the single-stepped instruction was a call, the return address
 811  * that is atop the stack is the address following the copied instruction.
 812  * We need to make it the address following the original instruction.
 813  *
 814  * If this is the first time we've single-stepped the instruction at
 815  * this probepoint, and the instruction is boostable, boost it: add a
 816  * jump instruction after the copied instruction, that jumps to the next
 817  * instruction after the probepoint.
 818  */
 819 static void __kprobes resume_execution(struct kprobe *p,
 820                 struct pt_regs *regs, struct kprobe_ctlblk *kcb)
 821 {
 822         unsigned long *tos = stack_addr(regs);
 823         unsigned long copy_ip = (unsigned long)p->ainsn.insn;
 824         unsigned long orig_ip = (unsigned long)p->addr;
 825         kprobe_opcode_t *insn = p->ainsn.insn;
 826
 827         /* Skip prefixes */
 828         insn = skip_prefixes(insn);
 829
 830         regs->flags &= ~X86_EFLAGS_TF;
 831         switch (*insn) {
 832         case 0x9c:      /* pushfl */
 833                 *tos &= ~(X86_EFLAGS_TF | X86_EFLAGS_IF);
 834                 *tos |= kcb->kprobe_old_flags;
 835                 break;
 836         case 0xc2:      /* iret/ret/lret */
 837         case 0xc3:
 838         case 0xca:
 839         case 0xcb:
 840         case 0xcf:
 841         case 0xea:      /* jmp absolute -- ip is correct */
 842                 /* ip is already adjusted, no more changes required */
 843                 p->ainsn.boostable = 1;
 844                 goto no_change;
 845         case 0xe8:      /* call relative - Fix return addr */
 846                 *tos = orig_ip + (*tos - copy_ip);
 847                 break;
 848 #ifdef CONFIG_X86_32
 849         case 0x9a:      /* call absolute -- same as call absolute, indirect */
 850                 *tos = orig_ip + (*tos - copy_ip);
 851                 goto no_change;
 852 #endif
 853         case 0xff:
 854                 if ((insn[1] & 0x30) == 0x10) {
 855                         /*
 856                          * call absolute, indirect
 857                          * Fix return addr; ip is correct.
 858                          * But this is not boostable
 859                          */
 860                         *tos = orig_ip + (*tos - copy_ip);
 861                         goto no_change;
 862                 } else if (((insn[1] & 0x31) == 0x20) ||
 863                            ((insn[1] & 0x31) == 0x21)) {
 864                         /*
 865                          * jmp near and far, absolute indirect
 866                          * ip is correct. And this is boostable
 867                          */
 868                         p->ainsn.boostable = 1;
 869                         goto no_change;
 870                 }
 871         default:
 872                 break;
 873         }
 874
 875         if (p->ainsn.boostable == 0) {
 876                 if ((regs->ip > copy_ip) &&
 877                     (regs->ip - copy_ip) + 5 < MAX_INSN_SIZE) {
 878                         /*
 879                          * These instructions can be executed directly if it
 880                          * jumps back to correct address.
 881                          */
 882                         synthesize_reljump((void *)regs->ip,
 883                                 (void *)orig_ip + (regs->ip - copy_ip));
 884                         p->ainsn.boostable = 1;
 885                 } else {
 886                         p->ainsn.boostable = -1;
 887                 }
 888         }
 889
 890         regs->ip += orig_ip - copy_ip;
 891
 892 no_change:
 893         restore_btf();
 894 }
 895
 896 /*
 897  * Interrupts are disabled on entry as trap1 is an interrupt gate and they
 898  * remain disabled throughout this function.
 899  */
 900 static int __kprobes post_kprobe_handler(struct pt_regs *regs)
 901 {
 902         struct kprobe *cur = kprobe_running();
 903         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 904
 905         if (!cur)
 906                 return 0;
 907
 908         resume_execution(cur, regs, kcb);
 909         regs->flags |= kcb->kprobe_saved_flags;
 910
 911         if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
 912                 kcb->kprobe_status = KPROBE_HIT_SSDONE;
 913                 cur->post_handler(cur, regs, 0);
 914         }
 915
 916         /* Restore back the original saved kprobes variables and continue. */
 917         if (kcb->kprobe_status == KPROBE_REENTER) {
 918                 restore_previous_kprobe(kcb);
 919                 goto out;
 920         }
 921         reset_current_kprobe();
 922 out:
 923         preempt_enable_no_resched();
 924
 925         /*
 926          * if somebody else is singlestepping across a probe point, flags
 927          * will have TF set, in which case, continue the remaining processing
 928          * of do_debug, as if this is not a probe hit.
 929          */
 930         if (regs->flags & X86_EFLAGS_TF)
 931                 return 0;
 932
 933         return 1;
 934 }
 935
 936 int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
 937 {
 938         struct kprobe *cur = kprobe_running();
 939         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 940
 941         switch (kcb->kprobe_status) {
 942         case KPROBE_HIT_SS:
 943         case KPROBE_REENTER:
 944                 /*
 945                  * We are here because the instruction being single
 946                  * stepped caused a page fault. We reset the current
 947                  * kprobe and the ip points back to the probe address
 948                  * and allow the page fault handler to continue as a
 949                  * normal page fault.
 950                  */
 951                 regs->ip = (unsigned long)cur->addr;
 952                 regs->flags |= kcb->kprobe_old_flags;
 953                 if (kcb->kprobe_status == KPROBE_REENTER)
 954                         restore_previous_kprobe(kcb);
 955                 else
 956                         reset_current_kprobe();
 957                 preempt_enable_no_resched();
 958                 break;
 959         case KPROBE_HIT_ACTIVE:
 960         case KPROBE_HIT_SSDONE:
 961                 /*
 962                  * We increment the nmissed count for accounting,
 963                  * we can also use npre/npostfault count for accounting
 964                  * these specific fault cases.
 965                  */
 966                 kprobes_inc_nmissed_count(cur);
 967
 968                 /*
 969                  * We come here because instructions in the pre/post
 970                  * handler caused the page_fault, this could happen
 971                  * if handler tries to access user space by
 972                  * copy_from_user(), get_user() etc. Let the
 973                  * user-specified handler try to fix it first.
 974                  */
 975                 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
 976                         return 1;
 977
 978                 /*
 979                  * In case the user-specified fault handler returned
 980                  * zero, try to fix up.
 981                  */
 982                 if (fixup_exception(regs))
 983                         return 1;
 984
 985                 /*
 986                  * fixup routine could not handle it,
 987                  * Let do_page_fault() fix it.
 988                  */
 989                 break;
 990         default:
 991                 break;
 992         }
 993         return 0;
 994 }
 995
 996 /*
 997  * Wrapper routine for handling exceptions.
 998  */
 999 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
1000                                        unsigned long val, void *data)
1001 {
1002         struct die_args *args = data;
1003         int ret = NOTIFY_DONE;
1004
1005         if (args->regs && user_mode_vm(args->regs))
1006                 return ret;
1007
1008         switch (val) {
1009         case DIE_INT3:
1010                 if (kprobe_handler(args->regs))
1011                         ret = NOTIFY_STOP;
1012                 break;
1013         case DIE_DEBUG:
1014                 if (post_kprobe_handler(args->regs)) {
1015                         /*
1016                          * Reset the BS bit in dr6 (pointed by args->err) to
1017                          * denote completion of processing
1018                          */
1019                         (*(unsigned long *)ERR_PTR(args->err)) &= ~DR_STEP;
1020                         ret = NOTIFY_STOP;
1021                 }
1022                 break;
1023         case DIE_GPF:
1024                 /*
1025                  * To be potentially processing a kprobe fault and to
1026                  * trust the result from kprobe_running(), we have
1027                  * be non-preemptible.
1028                  */
1029                 if (!preemptible() && kprobe_running() &&
1030                     kprobe_fault_handler(args->regs, args->trapnr))
1031                         ret = NOTIFY_STOP;
1032                 break;
1033         default:
1034                 break;
1035         }
1036         return ret;
1037 }
1038
1039 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
1040 {
1041         struct jprobe *jp = container_of(p, struct jprobe, kp);
1042         unsigned long addr;
1043         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1044
1045         kcb->jprobe_saved_regs = *regs;
1046         kcb->jprobe_saved_sp = stack_addr(regs);
1047         addr = (unsigned long)(kcb->jprobe_saved_sp);
1048
1049         /*
1050          * As Linus pointed out, gcc assumes that the callee
1051          * owns the argument space and could overwrite it, e.g.
1052          * tailcall optimization. So, to be absolutely safe
1053          * we also save and restore enough stack bytes to cover
1054          * the argument area.
1055          */
1056         memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
1057                MIN_STACK_SIZE(addr));
1058         regs->flags &= ~X86_EFLAGS_IF;
1059         trace_hardirqs_off();
1060         regs->ip = (unsigned long)(jp->entry);
1061         return 1;
1062 }
1063
1064 void __kprobes jprobe_return(void)
1065 {
1066         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1067
1068         asm volatile (
1069 #ifdef CONFIG_X86_64
1070                         "       xchg   %%rbx,%%rsp      \n"
1071 #else
1072                         "       xchgl   %%ebx,%%esp     \n"
1073 #endif
1074                         "       int3                    \n"
1075                         "       .globl jprobe_return_end\n"
1076                         "       jprobe_return_end:      \n"
1077                         "       nop                     \n"::"b"
1078                         (kcb->jprobe_saved_sp):"memory");
1079 }
1080
1081 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
1082 {
1083         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1084         u8 *addr = (u8 *) (regs->ip - 1);
1085         struct jprobe *jp = container_of(p, struct jprobe, kp);
1086
1087         if ((addr > (u8 *) jprobe_return) &&
1088             (addr < (u8 *) jprobe_return_end)) {
1089                 if (stack_addr(regs) != kcb->jprobe_saved_sp) {
1090                         struct pt_regs *saved_regs = &kcb->jprobe_saved_regs;
1091                         printk(KERN_ERR
1092                                "current sp %p does not match saved sp %p\n",
1093                                stack_addr(regs), kcb->jprobe_saved_sp);
1094                         printk(KERN_ERR "Saved registers for jprobe %p\n", jp);
1095                         show_registers(saved_regs);
1096                         printk(KERN_ERR "Current registers\n");
1097                         show_registers(regs);
1098                         BUG();
1099                 }
1100                 *regs = kcb->jprobe_saved_regs;
1101                 memcpy((kprobe_opcode_t *)(kcb->jprobe_saved_sp),
1102                        kcb->jprobes_stack,
1103                        MIN_STACK_SIZE(kcb->jprobe_saved_sp));
1104                 preempt_enable_no_resched();
1105                 return 1;
1106         }
1107         return 0;
1108 }
1109
1110
1111 #ifdef CONFIG_OPTPROBES
1112
1113 /* Insert a call instruction at address 'from', which calls address 'to'.*/
1114 static void __kprobes synthesize_relcall(void *from, void *to)
1115 {
1116         __synthesize_relative_insn(from, to, RELATIVECALL_OPCODE);
1117 }
1118
1119 /* Insert a move instruction which sets a pointer to eax/rdi (1st arg). */
1120 static void __kprobes synthesize_set_arg1(kprobe_opcode_t *addr,
1121                                           unsigned long val)
1122 {
1123 #ifdef CONFIG_X86_64
1124         *addr++ = 0x48;
1125         *addr++ = 0xbf;
1126 #else
1127         *addr++ = 0xb8;
1128 #endif
1129         *(unsigned long *)addr = val;
1130 }
1131
1132 static void __used __kprobes kprobes_optinsn_template_holder(void)
1133 {
1134         asm volatile (
1135                         ".global optprobe_template_entry\n"
1136                         "optprobe_template_entry: \n"
1137 #ifdef CONFIG_X86_64
1138                         /* We don't bother saving the ss register */
1139                         "       pushq %rsp\n"
1140                         "       pushfq\n"
1141                         SAVE_REGS_STRING
1142                         "       movq %rsp, %rsi\n"
1143                         ".global optprobe_template_val\n"
1144                         "optprobe_template_val: \n"
1145                         ASM_NOP5
1146                         ASM_NOP5
1147                         ".global optprobe_template_call\n"
1148                         "optprobe_template_call: \n"
1149                         ASM_NOP5
1150                         /* Move flags to rsp */
1151                         "       movq 144(%rsp), %rdx\n"
1152                         "       movq %rdx, 152(%rsp)\n"
1153                         RESTORE_REGS_STRING
1154                         /* Skip flags entry */
1155                         "       addq $8, %rsp\n"
1156                         "       popfq\n"
1157 #else /* CONFIG_X86_32 */
1158                         "       pushf\n"
1159                         SAVE_REGS_STRING
1160                         "       movl %esp, %edx\n"
1161                         ".global optprobe_template_val\n"
1162                         "optprobe_template_val: \n"
1163                         ASM_NOP5
1164                         ".global optprobe_template_call\n"
1165                         "optprobe_template_call: \n"
1166                         ASM_NOP5
1167                         RESTORE_REGS_STRING
1168                         "       addl $4, %esp\n"        /* skip cs */
1169                         "       popf\n"
1170 #endif
1171                         ".global optprobe_template_end\n"
1172                         "optprobe_template_end: \n");
1173 }
1174
1175 #define TMPL_MOVE_IDX \
1176         ((long)&optprobe_template_val - (long)&optprobe_template_entry)
1177 #define TMPL_CALL_IDX \
1178         ((long)&optprobe_template_call - (long)&optprobe_template_entry)
1179 #define TMPL_END_IDX \
1180         ((long)&optprobe_template_end - (long)&optprobe_template_entry)
1181
1182 #define INT3_SIZE sizeof(kprobe_opcode_t)
1183
1184 /* Optimized kprobe call back function: called from optinsn */
1185 static void __kprobes optimized_callback(struct optimized_kprobe *op,
1186                                          struct pt_regs *regs)
1187 {
1188         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1189         unsigned long flags;
1190
1191         /* This is possible if op is under delayed unoptimizing */
1192         if (kprobe_disabled(&op->kp))
1193                 return;
1194
1195         local_irq_save(flags);
1196         if (kprobe_running()) {
1197                 kprobes_inc_nmissed_count(&op->kp);
1198         } else {
1199                 /* Save skipped registers */
1200 #ifdef CONFIG_X86_64
1201                 regs->cs = __KERNEL_CS;
1202 #else
1203                 regs->cs = __KERNEL_CS | get_kernel_rpl();
1204                 regs->gs = 0;
1205 #endif
1206                 regs->ip = (unsigned long)op->kp.addr + INT3_SIZE;
1207                 regs->orig_ax = ~0UL;
1208
1209                 __this_cpu_write(current_kprobe, &op->kp);
1210                 kcb->kprobe_status = KPROBE_HIT_ACTIVE;
1211                 opt_pre_handler(&op->kp, regs);
1212                 __this_cpu_write(current_kprobe, NULL);
1213         }
1214         local_irq_restore(flags);
1215 }
1216
1217 static int __kprobes copy_optimized_instructions(u8 *dest, u8 *src)
1218 {
1219         int len = 0, ret;
1220
1221         while (len < RELATIVEJUMP_SIZE) {
1222                 ret = __copy_instruction(dest + len, src + len, 1);
1223                 if (!ret || !can_boost(dest + len))
1224                         return -EINVAL;
1225                 len += ret;
1226         }
1227         /* Check whether the address range is reserved */
1228         if (ftrace_text_reserved(src, src + len - 1) ||
1229             alternatives_text_reserved(src, src + len - 1) ||
1230             jump_label_text_reserved(src, src + len - 1))
1231                 return -EBUSY;
1232
1233         return len;
1234 }
1235
1236 /* Check whether insn is indirect jump */
1237 static int __kprobes insn_is_indirect_jump(struct insn *insn)
1238 {
1239         return ((insn->opcode.bytes[0] == 0xff &&
1240                 (X86_MODRM_REG(insn->modrm.value) & 6) == 4) || /* Jump */
1241                 insn->opcode.bytes[0] == 0xea); /* Segment based jump */
1242 }
1243
1244 /* Check whether insn jumps into specified address range */
1245 static int insn_jump_into_range(struct insn *insn, unsigned long start, int len)
1246 {
1247         unsigned long target = 0;
1248
1249         switch (insn->opcode.bytes[0]) {
1250         case 0xe0:      /* loopne */
1251         case 0xe1:      /* loope */
1252         case 0xe2:      /* loop */
1253         case 0xe3:      /* jcxz */
1254         case 0xe9:      /* near relative jump */
1255         case 0xeb:      /* short relative jump */
1256                 break;
1257         case 0x0f:
1258                 if ((insn->opcode.bytes[1] & 0xf0) == 0x80) /* jcc near */
1259                         break;
1260                 return 0;
1261         default:
1262                 if ((insn->opcode.bytes[0] & 0xf0) == 0x70) /* jcc short */
1263                         break;
1264                 return 0;
1265         }
1266         target = (unsigned long)insn->next_byte + insn->immediate.value;
1267
1268         return (start <= target && target <= start + len);
1269 }
1270
1271 /* Decode whole function to ensure any instructions don't jump into target */
1272 static int __kprobes can_optimize(unsigned long paddr)
1273 {
1274         int ret;
1275         unsigned long addr, size = 0, offset = 0;
1276         struct insn insn;
1277         kprobe_opcode_t buf[MAX_INSN_SIZE];
1278
1279         /* Lookup symbol including addr */
1280         if (!kallsyms_lookup_size_offset(paddr, &size, &offset))
1281                 return 0;
1282
1283         /*
1284          * Do not optimize in the entry code due to the unstable
1285          * stack handling.
1286          */
1287         if ((paddr >= (unsigned long )__entry_text_start) &&
1288             (paddr <  (unsigned long )__entry_text_end))
1289                 return 0;
1290
1291         /* Check there is enough space for a relative jump. */
1292         if (size - offset < RELATIVEJUMP_SIZE)
1293                 return 0;
1294
1295         /* Decode instructions */
1296         addr = paddr - offset;
1297         while (addr < paddr - offset + size) { /* Decode until function end */
1298                 if (search_exception_tables(addr))
1299                         /*
1300                          * Since some fixup code will jumps into this function,
1301                          * we can't optimize kprobe in this function.
1302                          */
1303                         return 0;
1304                 kernel_insn_init(&insn, (void *)addr);
1305                 insn_get_opcode(&insn);
1306                 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION) {
1307                         ret = recover_probed_instruction(buf, addr);
1308                         if (ret)
1309                                 return 0;
1310                         kernel_insn_init(&insn, buf);
1311                 }
1312                 insn_get_length(&insn);
1313                 /* Recover address */
1314                 insn.kaddr = (void *)addr;
1315                 insn.next_byte = (void *)(addr + insn.length);
1316                 /* Check any instructions don't jump into target */
1317                 if (insn_is_indirect_jump(&insn) ||
1318                     insn_jump_into_range(&insn, paddr + INT3_SIZE,
1319                                          RELATIVE_ADDR_SIZE))
1320                         return 0;
1321                 addr += insn.length;
1322         }
1323
1324         return 1;
1325 }
1326
1327 /* Check optimized_kprobe can actually be optimized. */
1328 int __kprobes arch_check_optimized_kprobe(struct optimized_kprobe *op)
1329 {
1330         int i;
1331         struct kprobe *p;
1332
1333         for (i = 1; i < op->optinsn.size; i++) {
1334                 p = get_kprobe(op->kp.addr + i);
1335                 if (p && !kprobe_disabled(p))
1336                         return -EEXIST;
1337         }
1338
1339         return 0;
1340 }
1341
1342 /* Check the addr is within the optimized instructions. */
1343 int __kprobes arch_within_optimized_kprobe(struct optimized_kprobe *op,
1344                                            unsigned long addr)
1345 {
1346         return ((unsigned long)op->kp.addr <= addr &&
1347                 (unsigned long)op->kp.addr + op->optinsn.size > addr);
1348 }
1349
1350 /* Free optimized instruction slot */
1351 static __kprobes
1352 void __arch_remove_optimized_kprobe(struct optimized_kprobe *op, int dirty)
1353 {
1354         if (op->optinsn.insn) {
1355                 free_optinsn_slot(op->optinsn.insn, dirty);
1356                 op->optinsn.insn = NULL;
1357                 op->optinsn.size = 0;
1358         }
1359 }
1360
1361 void __kprobes arch_remove_optimized_kprobe(struct optimized_kprobe *op)
1362 {
1363         __arch_remove_optimized_kprobe(op, 1);
1364 }
1365
1366 /*
1367  * Copy replacing target instructions
1368  * Target instructions MUST be relocatable (checked inside)
1369  */
1370 int __kprobes arch_prepare_optimized_kprobe(struct optimized_kprobe *op)
1371 {
1372         u8 *buf;
1373         int ret;
1374         long rel;
1375
1376         if (!can_optimize((unsigned long)op->kp.addr))
1377                 return -EILSEQ;
1378
1379         op->optinsn.insn = get_optinsn_slot();
1380         if (!op->optinsn.insn)
1381                 return -ENOMEM;
1382
1383         /*
1384          * Verify if the address gap is in 2GB range, because this uses
1385          * a relative jump.
1386          */
1387         rel = (long)op->optinsn.insn - (long)op->kp.addr + RELATIVEJUMP_SIZE;
1388         if (abs(rel) > 0x7fffffff)
1389                 return -ERANGE;
1390
1391         buf = (u8 *)op->optinsn.insn;
1392
1393         /* Copy instructions into the out-of-line buffer */
1394         ret = copy_optimized_instructions(buf + TMPL_END_IDX, op->kp.addr);
1395         if (ret < 0) {
1396                 __arch_remove_optimized_kprobe(op, 0);
1397                 return ret;
1398         }
1399         op->optinsn.size = ret;
1400
1401         /* Copy arch-dep-instance from template */
1402         memcpy(buf, &optprobe_template_entry, TMPL_END_IDX);
1403
1404         /* Set probe information */
1405         synthesize_set_arg1(buf + TMPL_MOVE_IDX, (unsigned long)op);
1406
1407         /* Set probe function call */
1408         synthesize_relcall(buf + TMPL_CALL_IDX, optimized_callback);
1409
1410         /* Set returning jmp instruction at the tail of out-of-line buffer */
1411         synthesize_reljump(buf + TMPL_END_IDX + op->optinsn.size,
1412                            (u8 *)op->kp.addr + op->optinsn.size);
1413
1414         flush_icache_range((unsigned long) buf,
1415                            (unsigned long) buf + TMPL_END_IDX +
1416                            op->optinsn.size + RELATIVEJUMP_SIZE);
1417         return 0;
1418 }
1419
1420 #define MAX_OPTIMIZE_PROBES 256
1421 static struct text_poke_param *jump_poke_params;
1422 static struct jump_poke_buffer {
1423         u8 buf[RELATIVEJUMP_SIZE];
1424 } *jump_poke_bufs;
1425
1426 static void __kprobes setup_optimize_kprobe(struct text_poke_param *tprm,
1427                                             u8 *insn_buf,
1428                                             struct optimized_kprobe *op)
1429 {
1430         s32 rel = (s32)((long)op->optinsn.insn -
1431                         ((long)op->kp.addr + RELATIVEJUMP_SIZE));
1432
1433         /* Backup instructions which will be replaced by jump address */
1434         memcpy(op->optinsn.copied_insn, op->kp.addr + INT3_SIZE,
1435                RELATIVE_ADDR_SIZE);
1436
1437         insn_buf[0] = RELATIVEJUMP_OPCODE;
1438         *(s32 *)(&insn_buf[1]) = rel;
1439
1440         tprm->addr = op->kp.addr;
1441         tprm->opcode = insn_buf;
1442         tprm->len = RELATIVEJUMP_SIZE;
1443 }
1444
1445 /*
1446  * Replace breakpoints (int3) with relative jumps.
1447  * Caller must call with locking kprobe_mutex and text_mutex.
1448  */
1449 void __kprobes arch_optimize_kprobes(struct list_head *oplist)
1450 {
1451         struct optimized_kprobe *op, *tmp;
1452         int c = 0;
1453
1454         list_for_each_entry_safe(op, tmp, oplist, list) {
1455                 WARN_ON(kprobe_disabled(&op->kp));
1456                 /* Setup param */
1457                 setup_optimize_kprobe(&jump_poke_params[c],
1458                                       jump_poke_bufs[c].buf, op);
1459                 list_del_init(&op->list);
1460                 if (++c >= MAX_OPTIMIZE_PROBES)
1461                         break;
1462         }
1463
1464         /*
1465          * text_poke_smp doesn't support NMI/MCE code modifying.
1466          * However, since kprobes itself also doesn't support NMI/MCE
1467          * code probing, it's not a problem.
1468          */
1469         text_poke_smp_batch(jump_poke_params, c);
1470 }
1471
1472 static void __kprobes setup_unoptimize_kprobe(struct text_poke_param *tprm,
1473                                               u8 *insn_buf,
1474                                               struct optimized_kprobe *op)
1475 {
1476         /* Set int3 to first byte for kprobes */
1477         insn_buf[0] = BREAKPOINT_INSTRUCTION;
1478         memcpy(insn_buf + 1, op->optinsn.copied_insn, RELATIVE_ADDR_SIZE);
1479
1480         tprm->addr = op->kp.addr;
1481         tprm->opcode = insn_buf;
1482         tprm->len = RELATIVEJUMP_SIZE;
1483 }
1484
1485 /*
1486  * Recover original instructions and breakpoints from relative jumps.
1487  * Caller must call with locking kprobe_mutex.
1488  */
1489 extern void arch_unoptimize_kprobes(struct list_head *oplist,
1490                                     struct list_head *done_list)
1491 {
1492         struct optimized_kprobe *op, *tmp;
1493         int c = 0;
1494
1495         list_for_each_entry_safe(op, tmp, oplist, list) {
1496                 /* Setup param */
1497                 setup_unoptimize_kprobe(&jump_poke_params[c],
1498                                         jump_poke_bufs[c].buf, op);
1499                 list_move(&op->list, done_list);
1500                 if (++c >= MAX_OPTIMIZE_PROBES)
1501                         break;
1502         }
1503
1504         /*
1505          * text_poke_smp doesn't support NMI/MCE code modifying.
1506          * However, since kprobes itself also doesn't support NMI/MCE
1507          * code probing, it's not a problem.
1508          */
1509         text_poke_smp_batch(jump_poke_params, c);
1510 }
1511
1512 /* Replace a relative jump with a breakpoint (int3).  */
1513 void __kprobes arch_unoptimize_kprobe(struct optimized_kprobe *op)
1514 {
1515         u8 buf[RELATIVEJUMP_SIZE];
1516
1517         /* Set int3 to first byte for kprobes */
1518         buf[0] = BREAKPOINT_INSTRUCTION;
1519         memcpy(buf + 1, op->optinsn.copied_insn, RELATIVE_ADDR_SIZE);
1520         text_poke_smp(op->kp.addr, buf, RELATIVEJUMP_SIZE);
1521 }
1522
1523 static int  __kprobes setup_detour_execution(struct kprobe *p,
1524                                              struct pt_regs *regs,
1525                                              int reenter)
1526 {
1527         struct optimized_kprobe *op;
1528
1529         if (p->flags & KPROBE_FLAG_OPTIMIZED) {
1530                 /* This kprobe is really able to run optimized path. */
1531                 op = container_of(p, struct optimized_kprobe, kp);
1532                 /* Detour through copied instructions */
1533                 regs->ip = (unsigned long)op->optinsn.insn + TMPL_END_IDX;
1534                 if (!reenter)
1535                         reset_current_kprobe();
1536                 preempt_enable_no_resched();
1537                 return 1;
1538         }
1539         return 0;
1540 }
1541
1542 static int __kprobes init_poke_params(void)
1543 {
1544         /* Allocate code buffer and parameter array */
1545         jump_poke_bufs = kmalloc(sizeof(struct jump_poke_buffer) *
1546                                  MAX_OPTIMIZE_PROBES, GFP_KERNEL);
1547         if (!jump_poke_bufs)
1548                 return -ENOMEM;
1549
1550         jump_poke_params = kmalloc(sizeof(struct text_poke_param) *
1551                                    MAX_OPTIMIZE_PROBES, GFP_KERNEL);
1552         if (!jump_poke_params) {
1553                 kfree(jump_poke_bufs);
1554                 jump_poke_bufs = NULL;
1555                 return -ENOMEM;
1556         }
1557
1558         return 0;
1559 }
1560 #else   /* !CONFIG_OPTPROBES */
1561 static int __kprobes init_poke_params(void)
1562 {
1563         return 0;
1564 }
1565 #endif
1566
1567 int __init arch_init_kprobes(void)
1568 {
1569         return init_poke_params();
1570 }
1571
1572 int __kprobes arch_trampoline_kprobe(struct kprobe *p)
1573 {
1574         return 0;
1575 }