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 #include <asm/nospec-branch.h>
  62
  63 void jprobe_return_end(void);
  64
  65 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
  66 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
  67
  68 #define stack_addr(regs) ((unsigned long *)kernel_stack_pointer(regs))
  69
  70 #define W(row, b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, ba, bb, bc, bd, be, bf)\
  71         (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
  72           (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
  73           (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
  74           (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
  75          << (row % 32))
  76         /*
  77          * Undefined/reserved opcodes, conditional jump, Opcode Extension
  78          * Groups, and some special opcodes can not boost.
  79          * This is non-const and volatile to keep gcc from statically
  80          * optimizing it out, as variable_test_bit makes gcc think only
  81          * *(unsigned long*) is used.
  82          */
  83 static volatile u32 twobyte_is_boostable[256 / 32] = {
  84         /*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f          */
  85         /*      ----------------------------------------------          */
  86         W(0x00, 0, 0, 1, 1, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0) | /* 00 */
  87         W(0x10, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 10 */
  88         W(0x20, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 20 */
  89         W(0x30, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 30 */
  90         W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
  91         W(0x50, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 50 */
  92         W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1) | /* 60 */
  93         W(0x70, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1) , /* 70 */
  94         W(0x80, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 80 */
  95         W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
  96         W(0xa0, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* a0 */
  97         W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1) , /* b0 */
  98         W(0xc0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1) | /* c0 */
  99         W(0xd0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) , /* d0 */
 100         W(0xe0, 0, 1, 1, 0, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* e0 */
 101         W(0xf0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 0)   /* f0 */
 102         /*      -----------------------------------------------         */
 103         /*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f          */
 104 };
 105 #undef W
 106
 107 struct kretprobe_blackpoint kretprobe_blacklist[] = {
 108         {"__switch_to", }, /* This function switches only current task, but
 109                               doesn't switch kernel stack.*/
 110         {NULL, NULL}    /* Terminator */
 111 };
 112 const int kretprobe_blacklist_size = ARRAY_SIZE(kretprobe_blacklist);
 113
 114 static void __kprobes __synthesize_relative_insn(void *from, void *to, u8 op)
 115 {
 116         struct __arch_relative_insn {
 117                 u8 op;
 118                 s32 raddr;
 119         } __attribute__((packed)) *insn;
 120
 121         insn = (struct __arch_relative_insn *)from;
 122         insn->raddr = (s32)((long)(to) - ((long)(from) + 5));
 123         insn->op = op;
 124 }
 125
 126 /* Insert a jump instruction at address 'from', which jumps to address 'to'.*/
 127 static void __kprobes synthesize_reljump(void *from, void *to)
 128 {
 129         __synthesize_relative_insn(from, to, RELATIVEJUMP_OPCODE);
 130 }
 131
 132 /*
 133  * Skip the prefixes of the instruction.
 134  */
 135 static kprobe_opcode_t *__kprobes skip_prefixes(kprobe_opcode_t *insn)
 136 {
 137         insn_attr_t attr;
 138
 139         attr = inat_get_opcode_attribute((insn_byte_t)*insn);
 140         while (inat_is_legacy_prefix(attr)) {
 141                 insn++;
 142                 attr = inat_get_opcode_attribute((insn_byte_t)*insn);
 143         }
 144 #ifdef CONFIG_X86_64
 145         if (inat_is_rex_prefix(attr))
 146                 insn++;
 147 #endif
 148         return insn;
 149 }
 150
 151 /*
 152  * Returns non-zero if opcode is boostable.
 153  * RIP relative instructions are adjusted at copying time in 64 bits mode
 154  */
 155 static int __kprobes can_boost(kprobe_opcode_t *opcodes)
 156 {
 157         kprobe_opcode_t opcode;
 158         kprobe_opcode_t *orig_opcodes = opcodes;
 159
 160         if (search_exception_tables((unsigned long)opcodes))
 161                 return 0;       /* Page fault may occur on this address. */
 162
 163 retry:
 164         if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
 165                 return 0;
 166         opcode = *(opcodes++);
 167
 168         /* 2nd-byte opcode */
 169         if (opcode == 0x0f) {
 170                 if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
 171                         return 0;
 172                 return test_bit(*opcodes,
 173                                 (unsigned long *)twobyte_is_boostable);
 174         }
 175
 176         switch (opcode & 0xf0) {
 177 #ifdef CONFIG_X86_64
 178         case 0x40:
 179                 goto retry; /* REX prefix is boostable */
 180 #endif
 181         case 0x60:
 182                 if (0x63 < opcode && opcode < 0x67)
 183                         goto retry; /* prefixes */
 184                 /* can't boost Address-size override and bound */
 185                 return (opcode != 0x62 && opcode != 0x67);
 186         case 0x70:
 187                 return 0; /* can't boost conditional jump */
 188         case 0xc0:
 189                 /* can't boost software-interruptions */
 190                 return (0xc1 < opcode && opcode < 0xcc) || opcode == 0xcf;
 191         case 0xd0:
 192                 /* can boost AA* and XLAT */
 193                 return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
 194         case 0xe0:
 195                 /* can boost in/out and absolute jmps */
 196                 return ((opcode & 0x04) || opcode == 0xea);
 197         case 0xf0:
 198                 if ((opcode & 0x0c) == 0 && opcode != 0xf1)
 199                         goto retry; /* lock/rep(ne) prefix */
 200                 /* clear and set flags are boostable */
 201                 return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
 202         default:
 203                 /* segment override prefixes are boostable */
 204                 if (opcode == 0x26 || opcode == 0x36 || opcode == 0x3e)
 205                         goto retry; /* prefixes */
 206                 /* CS override prefix and call are not boostable */
 207                 return (opcode != 0x2e && opcode != 0x9a);
 208         }
 209 }
 210
 211 /* Recover the probed instruction at addr for further analysis. */
 212 static int recover_probed_instruction(kprobe_opcode_t *buf, unsigned long addr)
 213 {
 214         struct kprobe *kp;
 215         kp = get_kprobe((void *)addr);
 216         if (!kp)
 217                 return -EINVAL;
 218
 219         /*
 220          *  Basically, kp->ainsn.insn has an original instruction.
 221          *  However, RIP-relative instruction can not do single-stepping
 222          *  at different place, __copy_instruction() tweaks the displacement of
 223          *  that instruction. In that case, we can't recover the instruction
 224          *  from the kp->ainsn.insn.
 225          *
 226          *  On the other hand, kp->opcode has a copy of the first byte of
 227          *  the probed instruction, which is overwritten by int3. And
 228          *  the instruction at kp->addr is not modified by kprobes except
 229          *  for the first byte, we can recover the original instruction
 230          *  from it and kp->opcode.
 231          */
 232         memcpy(buf, kp->addr, MAX_INSN_SIZE * sizeof(kprobe_opcode_t));
 233         buf[0] = kp->opcode;
 234         return 0;
 235 }
 236
 237 /* Check if paddr is at an instruction boundary */
 238 static int __kprobes can_probe(unsigned long paddr)
 239 {
 240         int ret;
 241         unsigned long addr, offset = 0;
 242         struct insn insn;
 243         kprobe_opcode_t buf[MAX_INSN_SIZE];
 244
 245         if (!kallsyms_lookup_size_offset(paddr, NULL, &offset))
 246                 return 0;
 247
 248         /* Decode instructions */
 249         addr = paddr - offset;
 250         while (addr < paddr) {
 251                 kernel_insn_init(&insn, (void *)addr);
 252                 insn_get_opcode(&insn);
 253
 254                 /*
 255                  * Check if the instruction has been modified by another
 256                  * kprobe, in which case we replace the breakpoint by the
 257                  * original instruction in our buffer.
 258                  */
 259                 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION) {
 260                         ret = recover_probed_instruction(buf, addr);
 261                         if (ret)
 262                                 /*
 263                                  * Another debugging subsystem might insert
 264                                  * this breakpoint. In that case, we can't
 265                                  * recover it.
 266                                  */
 267                                 return 0;
 268                         kernel_insn_init(&insn, buf);
 269                 }
 270                 insn_get_length(&insn);
 271                 addr += insn.length;
 272         }
 273
 274         return (addr == paddr);
 275 }
 276
 277 /*
 278  * Returns non-zero if opcode modifies the interrupt flag.
 279  */
 280 static int __kprobes is_IF_modifier(kprobe_opcode_t *insn)
 281 {
 282         /* Skip prefixes */
 283         insn = skip_prefixes(insn);
 284
 285         switch (*insn) {
 286         case 0xfa:              /* cli */
 287         case 0xfb:              /* sti */
 288         case 0xcf:              /* iret/iretd */
 289         case 0x9d:              /* popf/popfd */
 290                 return 1;
 291         }
 292
 293         return 0;
 294 }
 295
 296 /*
 297  * Copy an instruction and adjust the displacement if the instruction
 298  * uses the %rip-relative addressing mode.
 299  * If it does, Return the address of the 32-bit displacement word.
 300  * If not, return null.
 301  * Only applicable to 64-bit x86.
 302  */
 303 static int __kprobes __copy_instruction(u8 *dest, u8 *src, int recover)
 304 {
 305         struct insn insn;
 306         int ret;
 307         kprobe_opcode_t buf[MAX_INSN_SIZE];
 308
 309         kernel_insn_init(&insn, src);
 310         if (recover) {
 311                 insn_get_opcode(&insn);
 312                 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION) {
 313                         ret = recover_probed_instruction(buf,
 314                                                          (unsigned long)src);
 315                         if (ret)
 316                                 return 0;
 317                         kernel_insn_init(&insn, buf);
 318                 }
 319         }
 320         insn_get_length(&insn);
 321         memcpy(dest, insn.kaddr, insn.length);
 322
 323 #ifdef CONFIG_X86_64
 324         if (insn_rip_relative(&insn)) {
 325                 s64 newdisp;
 326                 u8 *disp;
 327                 kernel_insn_init(&insn, dest);
 328                 insn_get_displacement(&insn);
 329                 /*
 330                  * The copied instruction uses the %rip-relative addressing
 331                  * mode.  Adjust the displacement for the difference between
 332                  * the original location of this instruction and the location
 333                  * of the copy that will actually be run.  The tricky bit here
 334                  * is making sure that the sign extension happens correctly in
 335                  * this calculation, since we need a signed 32-bit result to
 336                  * be sign-extended to 64 bits when it's added to the %rip
 337                  * value and yield the same 64-bit result that the sign-
 338                  * extension of the original signed 32-bit displacement would
 339                  * have given.
 340                  */
 341                 newdisp = (u8 *) src + (s64) insn.displacement.value -
 342                           (u8 *) dest;
 343                 BUG_ON((s64) (s32) newdisp != newdisp); /* Sanity check.  */
 344                 disp = (u8 *) dest + insn_offset_displacement(&insn);
 345                 *(s32 *) disp = (s32) newdisp;
 346         }
 347 #endif
 348         return insn.length;
 349 }
 350
 351 static void __kprobes arch_copy_kprobe(struct kprobe *p)
 352 {
 353         /*
 354          * Copy an instruction without recovering int3, because it will be
 355          * put by another subsystem.
 356          */
 357         __copy_instruction(p->ainsn.insn, p->addr, 0);
 358
 359         if (can_boost(p->addr))
 360                 p->ainsn.boostable = 0;
 361         else
 362                 p->ainsn.boostable = -1;
 363
 364         p->opcode = *p->addr;
 365 }
 366
 367 int __kprobes arch_prepare_kprobe(struct kprobe *p)
 368 {
 369         if (alternatives_text_reserved(p->addr, p->addr))
 370                 return -EINVAL;
 371
 372         if (!can_probe((unsigned long)p->addr))
 373                 return -EILSEQ;
 374         /* insn: must be on special executable page on x86. */
 375         p->ainsn.insn = get_insn_slot();
 376         if (!p->ainsn.insn)
 377                 return -ENOMEM;
 378         arch_copy_kprobe(p);
 379         return 0;
 380 }
 381
 382 void __kprobes arch_arm_kprobe(struct kprobe *p)
 383 {
 384         text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
 385 }
 386
 387 void __kprobes arch_disarm_kprobe(struct kprobe *p)
 388 {
 389         text_poke(p->addr, &p->opcode, 1);
 390 }
 391
 392 void __kprobes arch_remove_kprobe(struct kprobe *p)
 393 {
 394         if (p->ainsn.insn) {
 395                 free_insn_slot(p->ainsn.insn, (p->ainsn.boostable == 1));
 396                 p->ainsn.insn = NULL;
 397         }
 398 }
 399
 400 static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
 401 {
 402         kcb->prev_kprobe.kp = kprobe_running();
 403         kcb->prev_kprobe.status = kcb->kprobe_status;
 404         kcb->prev_kprobe.old_flags = kcb->kprobe_old_flags;
 405         kcb->prev_kprobe.saved_flags = kcb->kprobe_saved_flags;
 406 }
 407
 408 static void __kprobes restore_previous_kprobe(struct kprobe_ctlblk *kcb)
 409 {
 410         __this_cpu_write(current_kprobe, kcb->prev_kprobe.kp);
 411         kcb->kprobe_status = kcb->prev_kprobe.status;
 412         kcb->kprobe_old_flags = kcb->prev_kprobe.old_flags;
 413         kcb->kprobe_saved_flags = kcb->prev_kprobe.saved_flags;
 414 }
 415
 416 static void __kprobes set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
 417                                 struct kprobe_ctlblk *kcb)
 418 {
 419         __this_cpu_write(current_kprobe, p);
 420         kcb->kprobe_saved_flags = kcb->kprobe_old_flags
 421                 = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
 422         if (is_IF_modifier(p->ainsn.insn))
 423                 kcb->kprobe_saved_flags &= ~X86_EFLAGS_IF;
 424 }
 425
 426 static void __kprobes clear_btf(void)
 427 {
 428         if (test_thread_flag(TIF_BLOCKSTEP)) {
 429                 unsigned long debugctl = get_debugctlmsr();
 430
 431                 debugctl &= ~DEBUGCTLMSR_BTF;
 432                 update_debugctlmsr(debugctl);
 433         }
 434 }
 435
 436 static void __kprobes restore_btf(void)
 437 {
 438         if (test_thread_flag(TIF_BLOCKSTEP)) {
 439                 unsigned long debugctl = get_debugctlmsr();
 440
 441                 debugctl |= DEBUGCTLMSR_BTF;
 442                 update_debugctlmsr(debugctl);
 443         }
 444 }
 445
 446 void __kprobes arch_prepare_kretprobe(struct kretprobe_instance *ri,
 447                                       struct pt_regs *regs)
 448 {
 449         unsigned long *sara = stack_addr(regs);
 450
 451         ri->ret_addr = (kprobe_opcode_t *) *sara;
 452
 453         /* Replace the return addr with trampoline addr */
 454         *sara = (unsigned long) &kretprobe_trampoline;
 455 }
 456
 457 #ifdef CONFIG_OPTPROBES
 458 static int  __kprobes setup_detour_execution(struct kprobe *p,
 459                                              struct pt_regs *regs,
 460                                              int reenter);
 461 #else
 462 #define setup_detour_execution(p, regs, reenter) (0)
 463 #endif
 464
 465 static void __kprobes setup_singlestep(struct kprobe *p, struct pt_regs *regs,
 466                                        struct kprobe_ctlblk *kcb, int reenter)
 467 {
 468         if (setup_detour_execution(p, regs, reenter))
 469                 return;
 470
 471 #if !defined(CONFIG_PREEMPT)
 472         if (p->ainsn.boostable == 1 && !p->post_handler) {
 473                 /* Boost up -- we can execute copied instructions directly */
 474                 if (!reenter)
 475                         reset_current_kprobe();
 476                 /*
 477                  * Reentering boosted probe doesn't reset current_kprobe,
 478                  * nor set current_kprobe, because it doesn't use single
 479                  * stepping.
 480                  */
 481                 regs->ip = (unsigned long)p->ainsn.insn;
 482                 preempt_enable_no_resched();
 483                 return;
 484         }
 485 #endif
 486         if (reenter) {
 487                 save_previous_kprobe(kcb);
 488                 set_current_kprobe(p, regs, kcb);
 489                 kcb->kprobe_status = KPROBE_REENTER;
 490         } else
 491                 kcb->kprobe_status = KPROBE_HIT_SS;
 492         /* Prepare real single stepping */
 493         clear_btf();
 494         regs->flags |= X86_EFLAGS_TF;
 495         regs->flags &= ~X86_EFLAGS_IF;
 496         /* single step inline if the instruction is an int3 */
 497         if (p->opcode == BREAKPOINT_INSTRUCTION)
 498                 regs->ip = (unsigned long)p->addr;
 499         else
 500                 regs->ip = (unsigned long)p->ainsn.insn;
 501 }
 502
 503 /*
 504  * We have reentered the kprobe_handler(), since another probe was hit while
 505  * within the handler. We save the original kprobes variables and just single
 506  * step on the instruction of the new probe without calling any user handlers.
 507  */
 508 static int __kprobes reenter_kprobe(struct kprobe *p, struct pt_regs *regs,
 509                                     struct kprobe_ctlblk *kcb)
 510 {
 511         switch (kcb->kprobe_status) {
 512         case KPROBE_HIT_SSDONE:
 513         case KPROBE_HIT_ACTIVE:
 514                 kprobes_inc_nmissed_count(p);
 515                 setup_singlestep(p, regs, kcb, 1);
 516                 break;
 517         case KPROBE_HIT_SS:
 518                 /* A probe has been hit in the codepath leading up to, or just
 519                  * after, single-stepping of a probed instruction. This entire
 520                  * codepath should strictly reside in .kprobes.text section.
 521                  * Raise a BUG or we'll continue in an endless reentering loop
 522                  * and eventually a stack overflow.
 523                  */
 524                 printk(KERN_WARNING "Unrecoverable kprobe detected at %p.\n",
 525                        p->addr);
 526                 dump_kprobe(p);
 527                 BUG();
 528         default:
 529                 /* impossible cases */
 530                 WARN_ON(1);
 531                 return 0;
 532         }
 533
 534         return 1;
 535 }
 536
 537 /*
 538  * Interrupts are disabled on entry as trap3 is an interrupt gate and they
 539  * remain disabled throughout this function.
 540  */
 541 static int __kprobes kprobe_handler(struct pt_regs *regs)
 542 {
 543         kprobe_opcode_t *addr;
 544         struct kprobe *p;
 545         struct kprobe_ctlblk *kcb;
 546
 547         addr = (kprobe_opcode_t *)(regs->ip - sizeof(kprobe_opcode_t));
 548         /*
 549          * We don't want to be preempted for the entire
 550          * duration of kprobe processing. We conditionally
 551          * re-enable preemption at the end of this function,
 552          * and also in reenter_kprobe() and setup_singlestep().
 553          */
 554         preempt_disable();
 555
 556         kcb = get_kprobe_ctlblk();
 557         p = get_kprobe(addr);
 558
 559         if (p) {
 560                 if (kprobe_running()) {
 561                         if (reenter_kprobe(p, regs, kcb))
 562                                 return 1;
 563                 } else {
 564                         set_current_kprobe(p, regs, kcb);
 565                         kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 566
 567                         /*
 568                          * If we have no pre-handler or it returned 0, we
 569                          * continue with normal processing.  If we have a
 570                          * pre-handler and it returned non-zero, it prepped
 571                          * for calling the break_handler below on re-entry
 572                          * for jprobe processing, so get out doing nothing
 573                          * more here.
 574                          */
 575                         if (!p->pre_handler || !p->pre_handler(p, regs))
 576                                 setup_singlestep(p, regs, kcb, 0);
 577                         return 1;
 578                 }
 579         } else if (*addr != BREAKPOINT_INSTRUCTION) {
 580                 /*
 581                  * The breakpoint instruction was removed right
 582                  * after we hit it.  Another cpu has removed
 583                  * either a probepoint or a debugger breakpoint
 584                  * at this address.  In either case, no further
 585                  * handling of this interrupt is appropriate.
 586                  * Back up over the (now missing) int3 and run
 587                  * the original instruction.
 588                  */
 589                 regs->ip = (unsigned long)addr;
 590                 preempt_enable_no_resched();
 591                 return 1;
 592         } else if (kprobe_running()) {
 593                 p = __this_cpu_read(current_kprobe);
 594                 if (p->break_handler && p->break_handler(p, regs)) {
 595                         setup_singlestep(p, regs, kcb, 0);
 596                         return 1;
 597                 }
 598         } /* else: not a kprobe fault; let the kernel handle it */
 599
 600         preempt_enable_no_resched();
 601         return 0;
 602 }
 603
 604 #ifdef CONFIG_X86_64
 605 #define SAVE_REGS_STRING                \
 606         /* Skip cs, ip, orig_ax. */     \
 607         "       subq $24, %rsp\n"       \
 608         "       pushq %rdi\n"           \
 609         "       pushq %rsi\n"           \
 610         "       pushq %rdx\n"           \
 611         "       pushq %rcx\n"           \
 612         "       pushq %rax\n"           \
 613         "       pushq %r8\n"            \
 614         "       pushq %r9\n"            \
 615         "       pushq %r10\n"           \
 616         "       pushq %r11\n"           \
 617         "       pushq %rbx\n"           \
 618         "       pushq %rbp\n"           \
 619         "       pushq %r12\n"           \
 620         "       pushq %r13\n"           \
 621         "       pushq %r14\n"           \
 622         "       pushq %r15\n"
 623 #define RESTORE_REGS_STRING             \
 624         "       popq %r15\n"            \
 625         "       popq %r14\n"            \
 626         "       popq %r13\n"            \
 627         "       popq %r12\n"            \
 628         "       popq %rbp\n"            \
 629         "       popq %rbx\n"            \
 630         "       popq %r11\n"            \
 631         "       popq %r10\n"            \
 632         "       popq %r9\n"             \
 633         "       popq %r8\n"             \
 634         "       popq %rax\n"            \
 635         "       popq %rcx\n"            \
 636         "       popq %rdx\n"            \
 637         "       popq %rsi\n"            \
 638         "       popq %rdi\n"            \
 639         /* Skip orig_ax, ip, cs */      \
 640         "       addq $24, %rsp\n"
 641 #else
 642 #define SAVE_REGS_STRING                \
 643         /* Skip cs, ip, orig_ax and gs. */      \
 644         "       subl $16, %esp\n"       \
 645         "       pushl %fs\n"            \
 646         "       pushl %es\n"            \
 647         "       pushl %ds\n"            \
 648         "       pushl %eax\n"           \
 649         "       pushl %ebp\n"           \
 650         "       pushl %edi\n"           \
 651         "       pushl %esi\n"           \
 652         "       pushl %edx\n"           \
 653         "       pushl %ecx\n"           \
 654         "       pushl %ebx\n"
 655 #define RESTORE_REGS_STRING             \
 656         "       popl %ebx\n"            \
 657         "       popl %ecx\n"            \
 658         "       popl %edx\n"            \
 659         "       popl %esi\n"            \
 660         "       popl %edi\n"            \
 661         "       popl %ebp\n"            \
 662         "       popl %eax\n"            \
 663         /* Skip ds, es, fs, gs, orig_ax, and ip. Note: don't pop cs here*/\
 664         "       addl $24, %esp\n"
 665 #endif
 666
 667 /*
 668  * When a retprobed function returns, this code saves registers and
 669  * calls trampoline_handler() runs, which calls the kretprobe's handler.
 670  */
 671 static void __used __kprobes kretprobe_trampoline_holder(void)
 672 {
 673         asm volatile (
 674                         ".global kretprobe_trampoline\n"
 675                         "kretprobe_trampoline: \n"
 676 #ifdef CONFIG_X86_64
 677                         /* We don't bother saving the ss register */
 678                         "       pushq %rsp\n"
 679                         "       pushfq\n"
 680                         SAVE_REGS_STRING
 681                         "       movq %rsp, %rdi\n"
 682                         "       call trampoline_handler\n"
 683                         /* Replace saved sp with true return address. */
 684                         "       movq %rax, 152(%rsp)\n"
 685                         RESTORE_REGS_STRING
 686                         "       popfq\n"
 687 #else
 688                         "       pushf\n"
 689                         SAVE_REGS_STRING
 690                         "       movl %esp, %eax\n"
 691                         "       call trampoline_handler\n"
 692                         /* Move flags to cs */
 693                         "       movl 56(%esp), %edx\n"
 694                         "       movl %edx, 52(%esp)\n"
 695                         /* Replace saved flags with true return address. */
 696                         "       movl %eax, 56(%esp)\n"
 697                         RESTORE_REGS_STRING
 698                         "       popf\n"
 699 #endif
 700                         "       ret\n");
 701 }
 702
 703 /*
 704  * Called from kretprobe_trampoline
 705  */
 706 static __used __kprobes void *trampoline_handler(struct pt_regs *regs)
 707 {
 708         struct kretprobe_instance *ri = NULL;
 709         struct hlist_head *head, empty_rp;
 710         struct hlist_node *node, *tmp;
 711         unsigned long flags, orig_ret_address = 0;
 712         unsigned long trampoline_address = (unsigned long)&kretprobe_trampoline;
 713         kprobe_opcode_t *correct_ret_addr = NULL;
 714
 715         INIT_HLIST_HEAD(&empty_rp);
 716         kretprobe_hash_lock(current, &head, &flags);
 717         /* fixup registers */
 718 #ifdef CONFIG_X86_64
 719         regs->cs = __KERNEL_CS;
 720 #else
 721         regs->cs = __KERNEL_CS | get_kernel_rpl();
 722         regs->gs = 0;
 723 #endif
 724         regs->ip = trampoline_address;
 725         regs->orig_ax = ~0UL;
 726
 727         /*
 728          * It is possible to have multiple instances associated with a given
 729          * task either because multiple functions in the call path have
 730          * return probes installed on them, and/or more than one
 731          * return probe was registered for a target function.
 732          *
 733          * We can handle this because:
 734          *     - instances are always pushed into the head of the list
 735          *     - when multiple return probes are registered for the same
 736          *       function, the (chronologically) first instance's ret_addr
 737          *       will be the real return address, and all the rest will
 738          *       point to kretprobe_trampoline.
 739          */
 740         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
 741                 if (ri->task != current)
 742                         /* another task is sharing our hash bucket */
 743                         continue;
 744
 745                 orig_ret_address = (unsigned long)ri->ret_addr;
 746
 747                 if (orig_ret_address != trampoline_address)
 748                         /*
 749                          * This is the real return address. Any other
 750                          * instances associated with this task are for
 751                          * other calls deeper on the call stack
 752                          */
 753                         break;
 754         }
 755
 756         kretprobe_assert(ri, orig_ret_address, trampoline_address);
 757
 758         correct_ret_addr = ri->ret_addr;
 759         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
 760                 if (ri->task != current)
 761                         /* another task is sharing our hash bucket */
 762                         continue;
 763
 764                 orig_ret_address = (unsigned long)ri->ret_addr;
 765                 if (ri->rp && ri->rp->handler) {
 766                         __this_cpu_write(current_kprobe, &ri->rp->kp);
 767                         get_kprobe_ctlblk()->kprobe_status = KPROBE_HIT_ACTIVE;
 768                         ri->ret_addr = correct_ret_addr;
 769                         ri->rp->handler(ri, regs);
 770                         __this_cpu_write(current_kprobe, NULL);
 771                 }
 772
 773                 recycle_rp_inst(ri, &empty_rp);
 774
 775                 if (orig_ret_address != trampoline_address)
 776                         /*
 777                          * This is the real return address. Any other
 778                          * instances associated with this task are for
 779                          * other calls deeper on the call stack
 780                          */
 781                         break;
 782         }
 783
 784         kretprobe_hash_unlock(current, &flags);
 785
 786         hlist_for_each_entry_safe(ri, node, tmp, &empty_rp, hlist) {
 787                 hlist_del(&ri->hlist);
 788                 kfree(ri);
 789         }
 790         return (void *)orig_ret_address;
 791 }
 792
 793 /*
 794  * Called after single-stepping.  p->addr is the address of the
 795  * instruction whose first byte has been replaced by the "int 3"
 796  * instruction.  To avoid the SMP problems that can occur when we
 797  * temporarily put back the original opcode to single-step, we
 798  * single-stepped a copy of the instruction.  The address of this
 799  * copy is p->ainsn.insn.
 800  *
 801  * This function prepares to return from the post-single-step
 802  * interrupt.  We have to fix up the stack as follows:
 803  *
 804  * 0) Except in the case of absolute or indirect jump or call instructions,
 805  * the new ip is relative to the copied instruction.  We need to make
 806  * it relative to the original instruction.
 807  *
 808  * 1) If the single-stepped instruction was pushfl, then the TF and IF
 809  * flags are set in the just-pushed flags, and may need to be cleared.
 810  *
 811  * 2) If the single-stepped instruction was a call, the return address
 812  * that is atop the stack is the address following the copied instruction.
 813  * We need to make it the address following the original instruction.
 814  *
 815  * If this is the first time we've single-stepped the instruction at
 816  * this probepoint, and the instruction is boostable, boost it: add a
 817  * jump instruction after the copied instruction, that jumps to the next
 818  * instruction after the probepoint.
 819  */
 820 static void __kprobes resume_execution(struct kprobe *p,
 821                 struct pt_regs *regs, struct kprobe_ctlblk *kcb)
 822 {
 823         unsigned long *tos = stack_addr(regs);
 824         unsigned long copy_ip = (unsigned long)p->ainsn.insn;
 825         unsigned long orig_ip = (unsigned long)p->addr;
 826         kprobe_opcode_t *insn = p->ainsn.insn;
 827
 828         /* Skip prefixes */
 829         insn = skip_prefixes(insn);
 830
 831         regs->flags &= ~X86_EFLAGS_TF;
 832         switch (*insn) {
 833         case 0x9c:      /* pushfl */
 834                 *tos &= ~(X86_EFLAGS_TF | X86_EFLAGS_IF);
 835                 *tos |= kcb->kprobe_old_flags;
 836                 break;
 837         case 0xc2:      /* iret/ret/lret */
 838         case 0xc3:
 839         case 0xca:
 840         case 0xcb:
 841         case 0xcf:
 842         case 0xea:      /* jmp absolute -- ip is correct */
 843                 /* ip is already adjusted, no more changes required */
 844                 p->ainsn.boostable = 1;
 845                 goto no_change;
 846         case 0xe8:      /* call relative - Fix return addr */
 847                 *tos = orig_ip + (*tos - copy_ip);
 848                 break;
 849 #ifdef CONFIG_X86_32
 850         case 0x9a:      /* call absolute -- same as call absolute, indirect */
 851                 *tos = orig_ip + (*tos - copy_ip);
 852                 goto no_change;
 853 #endif
 854         case 0xff:
 855                 if ((insn[1] & 0x30) == 0x10) {
 856                         /*
 857                          * call absolute, indirect
 858                          * Fix return addr; ip is correct.
 859                          * But this is not boostable
 860                          */
 861                         *tos = orig_ip + (*tos - copy_ip);
 862                         goto no_change;
 863                 } else if (((insn[1] & 0x31) == 0x20) ||
 864                            ((insn[1] & 0x31) == 0x21)) {
 865                         /*
 866                          * jmp near and far, absolute indirect
 867                          * ip is correct. And this is boostable
 868                          */
 869                         p->ainsn.boostable = 1;
 870                         goto no_change;
 871                 }
 872         default:
 873                 break;
 874         }
 875
 876         if (p->ainsn.boostable == 0) {
 877                 if ((regs->ip > copy_ip) &&
 878                     (regs->ip - copy_ip) + 5 < MAX_INSN_SIZE) {
 879                         /*
 880                          * These instructions can be executed directly if it
 881                          * jumps back to correct address.
 882                          */
 883                         synthesize_reljump((void *)regs->ip,
 884                                 (void *)orig_ip + (regs->ip - copy_ip));
 885                         p->ainsn.boostable = 1;
 886                 } else {
 887                         p->ainsn.boostable = -1;
 888                 }
 889         }
 890
 891         regs->ip += orig_ip - copy_ip;
 892
 893 no_change:
 894         restore_btf();
 895 }
 896
 897 /*
 898  * Interrupts are disabled on entry as trap1 is an interrupt gate and they
 899  * remain disabled throughout this function.
 900  */
 901 static int __kprobes post_kprobe_handler(struct pt_regs *regs)
 902 {
 903         struct kprobe *cur = kprobe_running();
 904         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 905
 906         if (!cur)
 907                 return 0;
 908
 909         resume_execution(cur, regs, kcb);
 910         regs->flags |= kcb->kprobe_saved_flags;
 911
 912         if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
 913                 kcb->kprobe_status = KPROBE_HIT_SSDONE;
 914                 cur->post_handler(cur, regs, 0);
 915         }
 916
 917         /* Restore back the original saved kprobes variables and continue. */
 918         if (kcb->kprobe_status == KPROBE_REENTER) {
 919                 restore_previous_kprobe(kcb);
 920                 goto out;
 921         }
 922         reset_current_kprobe();
 923 out:
 924         preempt_enable_no_resched();
 925
 926         /*
 927          * if somebody else is singlestepping across a probe point, flags
 928          * will have TF set, in which case, continue the remaining processing
 929          * of do_debug, as if this is not a probe hit.
 930          */
 931         if (regs->flags & X86_EFLAGS_TF)
 932                 return 0;
 933
 934         return 1;
 935 }
 936
 937 int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
 938 {
 939         struct kprobe *cur = kprobe_running();
 940         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 941
 942         switch (kcb->kprobe_status) {
 943         case KPROBE_HIT_SS:
 944         case KPROBE_REENTER:
 945                 /*
 946                  * We are here because the instruction being single
 947                  * stepped caused a page fault. We reset the current
 948                  * kprobe and the ip points back to the probe address
 949                  * and allow the page fault handler to continue as a
 950                  * normal page fault.
 951                  */
 952                 regs->ip = (unsigned long)cur->addr;
 953                 /*
 954                  * Trap flag (TF) has been set here because this fault
 955                  * happened where the single stepping will be done.
 956                  * So clear it by resetting the current kprobe:
 957                  */
 958                 regs->flags &= ~X86_EFLAGS_TF;
 959
 960                 /*
 961                  * If the TF flag was set before the kprobe hit,
 962                  * don't touch it:
 963                  */
 964                 regs->flags |= kcb->kprobe_old_flags;
 965
 966                 if (kcb->kprobe_status == KPROBE_REENTER)
 967                         restore_previous_kprobe(kcb);
 968                 else
 969                         reset_current_kprobe();
 970                 preempt_enable_no_resched();
 971                 break;
 972         case KPROBE_HIT_ACTIVE:
 973         case KPROBE_HIT_SSDONE:
 974                 /*
 975                  * We increment the nmissed count for accounting,
 976                  * we can also use npre/npostfault count for accounting
 977                  * these specific fault cases.
 978                  */
 979                 kprobes_inc_nmissed_count(cur);
 980
 981                 /*
 982                  * We come here because instructions in the pre/post
 983                  * handler caused the page_fault, this could happen
 984                  * if handler tries to access user space by
 985                  * copy_from_user(), get_user() etc. Let the
 986                  * user-specified handler try to fix it first.
 987                  */
 988                 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
 989                         return 1;
 990
 991                 /*
 992                  * In case the user-specified fault handler returned
 993                  * zero, try to fix up.
 994                  */
 995                 if (fixup_exception(regs))
 996                         return 1;
 997
 998                 /*
 999                  * fixup routine could not handle it,
1000                  * Let do_page_fault() fix it.
1001                  */
1002                 break;
1003         default:
1004                 break;
1005         }
1006         return 0;
1007 }
1008
1009 /*
1010  * Wrapper routine for handling exceptions.
1011  */
1012 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
1013                                        unsigned long val, void *data)
1014 {
1015         struct die_args *args = data;
1016         int ret = NOTIFY_DONE;
1017
1018         if (args->regs && user_mode_vm(args->regs))
1019                 return ret;
1020
1021         switch (val) {
1022         case DIE_INT3:
1023                 if (kprobe_handler(args->regs))
1024                         ret = NOTIFY_STOP;
1025                 break;
1026         case DIE_DEBUG:
1027                 if (post_kprobe_handler(args->regs)) {
1028                         /*
1029                          * Reset the BS bit in dr6 (pointed by args->err) to
1030                          * denote completion of processing
1031                          */
1032                         (*(unsigned long *)ERR_PTR(args->err)) &= ~DR_STEP;
1033                         ret = NOTIFY_STOP;
1034                 }
1035                 break;
1036         case DIE_GPF:
1037                 /*
1038                  * To be potentially processing a kprobe fault and to
1039                  * trust the result from kprobe_running(), we have
1040                  * be non-preemptible.
1041                  */
1042                 if (!preemptible() && kprobe_running() &&
1043                     kprobe_fault_handler(args->regs, args->trapnr))
1044                         ret = NOTIFY_STOP;
1045                 break;
1046         default:
1047                 break;
1048         }
1049         return ret;
1050 }
1051
1052 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
1053 {
1054         struct jprobe *jp = container_of(p, struct jprobe, kp);
1055         unsigned long addr;
1056         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1057
1058         kcb->jprobe_saved_regs = *regs;
1059         kcb->jprobe_saved_sp = stack_addr(regs);
1060         addr = (unsigned long)(kcb->jprobe_saved_sp);
1061
1062         /*
1063          * As Linus pointed out, gcc assumes that the callee
1064          * owns the argument space and could overwrite it, e.g.
1065          * tailcall optimization. So, to be absolutely safe
1066          * we also save and restore enough stack bytes to cover
1067          * the argument area.
1068          */
1069         memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
1070                MIN_STACK_SIZE(addr));
1071         regs->flags &= ~X86_EFLAGS_IF;
1072         trace_hardirqs_off();
1073         regs->ip = (unsigned long)(jp->entry);
1074
1075         /*
1076          * jprobes use jprobe_return() which skips the normal return
1077          * path of the function, and this messes up the accounting of the
1078          * function graph tracer to get messed up.
1079          *
1080          * Pause function graph tracing while performing the jprobe function.
1081          */
1082         pause_graph_tracing();
1083         return 1;
1084 }
1085
1086 void __kprobes jprobe_return(void)
1087 {
1088         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1089
1090         asm volatile (
1091 #ifdef CONFIG_X86_64
1092                         "       xchg   %%rbx,%%rsp      \n"
1093 #else
1094                         "       xchgl   %%ebx,%%esp     \n"
1095 #endif
1096                         "       int3                    \n"
1097                         "       .globl jprobe_return_end\n"
1098                         "       jprobe_return_end:      \n"
1099                         "       nop                     \n"::"b"
1100                         (kcb->jprobe_saved_sp):"memory");
1101 }
1102
1103 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
1104 {
1105         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1106         u8 *addr = (u8 *) (regs->ip - 1);
1107         struct jprobe *jp = container_of(p, struct jprobe, kp);
1108         void *saved_sp = kcb->jprobe_saved_sp;
1109
1110         if ((addr > (u8 *) jprobe_return) &&
1111             (addr < (u8 *) jprobe_return_end)) {
1112                 if (stack_addr(regs) != saved_sp) {
1113                         struct pt_regs *saved_regs = &kcb->jprobe_saved_regs;
1114                         printk(KERN_ERR
1115                                "current sp %p does not match saved sp %p\n",
1116                                stack_addr(regs), saved_sp);
1117                         printk(KERN_ERR "Saved registers for jprobe %p\n", jp);
1118                         show_registers(saved_regs);
1119                         printk(KERN_ERR "Current registers\n");
1120                         show_registers(regs);
1121                         BUG();
1122                 }
1123                 /* It's OK to start function graph tracing again */
1124                 unpause_graph_tracing();
1125                 *regs = kcb->jprobe_saved_regs;
1126                 memcpy(saved_sp, kcb->jprobes_stack, MIN_STACK_SIZE(saved_sp));
1127                 preempt_enable_no_resched();
1128                 return 1;
1129         }
1130         return 0;
1131 }
1132
1133
1134 #ifdef CONFIG_OPTPROBES
1135
1136 /* Insert a call instruction at address 'from', which calls address 'to'.*/
1137 static void __kprobes synthesize_relcall(void *from, void *to)
1138 {
1139         __synthesize_relative_insn(from, to, RELATIVECALL_OPCODE);
1140 }
1141
1142 /* Insert a move instruction which sets a pointer to eax/rdi (1st arg). */
1143 static void __kprobes synthesize_set_arg1(kprobe_opcode_t *addr,
1144                                           unsigned long val)
1145 {
1146 #ifdef CONFIG_X86_64
1147         *addr++ = 0x48;
1148         *addr++ = 0xbf;
1149 #else
1150         *addr++ = 0xb8;
1151 #endif
1152         *(unsigned long *)addr = val;
1153 }
1154
1155 static void __used __kprobes kprobes_optinsn_template_holder(void)
1156 {
1157         asm volatile (
1158                         ".global optprobe_template_entry\n"
1159                         "optprobe_template_entry: \n"
1160 #ifdef CONFIG_X86_64
1161                         /* We don't bother saving the ss register */
1162                         "       pushq %rsp\n"
1163                         "       pushfq\n"
1164                         SAVE_REGS_STRING
1165                         "       movq %rsp, %rsi\n"
1166                         ".global optprobe_template_val\n"
1167                         "optprobe_template_val: \n"
1168                         ASM_NOP5
1169                         ASM_NOP5
1170                         ".global optprobe_template_call\n"
1171                         "optprobe_template_call: \n"
1172                         ASM_NOP5
1173                         /* Move flags to rsp */
1174                         "       movq 144(%rsp), %rdx\n"
1175                         "       movq %rdx, 152(%rsp)\n"
1176                         RESTORE_REGS_STRING
1177                         /* Skip flags entry */
1178                         "       addq $8, %rsp\n"
1179                         "       popfq\n"
1180 #else /* CONFIG_X86_32 */
1181                         "       pushf\n"
1182                         SAVE_REGS_STRING
1183                         "       movl %esp, %edx\n"
1184                         ".global optprobe_template_val\n"
1185                         "optprobe_template_val: \n"
1186                         ASM_NOP5
1187                         ".global optprobe_template_call\n"
1188                         "optprobe_template_call: \n"
1189                         ASM_NOP5
1190                         RESTORE_REGS_STRING
1191                         "       addl $4, %esp\n"        /* skip cs */
1192                         "       popf\n"
1193 #endif
1194                         ".global optprobe_template_end\n"
1195                         "optprobe_template_end: \n");
1196 }
1197
1198 #define TMPL_MOVE_IDX \
1199         ((long)&optprobe_template_val - (long)&optprobe_template_entry)
1200 #define TMPL_CALL_IDX \
1201         ((long)&optprobe_template_call - (long)&optprobe_template_entry)
1202 #define TMPL_END_IDX \
1203         ((long)&optprobe_template_end - (long)&optprobe_template_entry)
1204
1205 #define INT3_SIZE sizeof(kprobe_opcode_t)
1206
1207 /* Optimized kprobe call back function: called from optinsn */
1208 static void __kprobes optimized_callback(struct optimized_kprobe *op,
1209                                          struct pt_regs *regs)
1210 {
1211         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1212         unsigned long flags;
1213
1214         /* This is possible if op is under delayed unoptimizing */
1215         if (kprobe_disabled(&op->kp))
1216                 return;
1217
1218         local_irq_save(flags);
1219         if (kprobe_running()) {
1220                 kprobes_inc_nmissed_count(&op->kp);
1221         } else {
1222                 /* Save skipped registers */
1223 #ifdef CONFIG_X86_64
1224                 regs->cs = __KERNEL_CS;
1225 #else
1226                 regs->cs = __KERNEL_CS | get_kernel_rpl();
1227                 regs->gs = 0;
1228 #endif
1229                 regs->ip = (unsigned long)op->kp.addr + INT3_SIZE;
1230                 regs->orig_ax = ~0UL;
1231
1232                 __this_cpu_write(current_kprobe, &op->kp);
1233                 kcb->kprobe_status = KPROBE_HIT_ACTIVE;
1234                 opt_pre_handler(&op->kp, regs);
1235                 __this_cpu_write(current_kprobe, NULL);
1236         }
1237         local_irq_restore(flags);
1238 }
1239
1240 static int __kprobes copy_optimized_instructions(u8 *dest, u8 *src)
1241 {
1242         int len = 0, ret;
1243
1244         while (len < RELATIVEJUMP_SIZE) {
1245                 ret = __copy_instruction(dest + len, src + len, 1);
1246                 if (!ret || !can_boost(dest + len))
1247                         return -EINVAL;
1248                 len += ret;
1249         }
1250         /* Check whether the address range is reserved */
1251         if (ftrace_text_reserved(src, src + len - 1) ||
1252             alternatives_text_reserved(src, src + len - 1) ||
1253             jump_label_text_reserved(src, src + len - 1))
1254                 return -EBUSY;
1255
1256         return len;
1257 }
1258
1259 /* Check whether insn is indirect jump */
1260 static int __kprobes __insn_is_indirect_jump(struct insn *insn)
1261 {
1262         return ((insn->opcode.bytes[0] == 0xff &&
1263                 (X86_MODRM_REG(insn->modrm.value) & 6) == 4) || /* Jump */
1264                 insn->opcode.bytes[0] == 0xea); /* Segment based jump */
1265 }
1266
1267 /* Check whether insn jumps into specified address range */
1268 static int insn_jump_into_range(struct insn *insn, unsigned long start, int len)
1269 {
1270         unsigned long target = 0;
1271
1272         switch (insn->opcode.bytes[0]) {
1273         case 0xe0:      /* loopne */
1274         case 0xe1:      /* loope */
1275         case 0xe2:      /* loop */
1276         case 0xe3:      /* jcxz */
1277         case 0xe9:      /* near relative jump */
1278         case 0xeb:      /* short relative jump */
1279                 break;
1280         case 0x0f:
1281                 if ((insn->opcode.bytes[1] & 0xf0) == 0x80) /* jcc near */
1282                         break;
1283                 return 0;
1284         default:
1285                 if ((insn->opcode.bytes[0] & 0xf0) == 0x70) /* jcc short */
1286                         break;
1287                 return 0;
1288         }
1289         target = (unsigned long)insn->next_byte + insn->immediate.value;
1290
1291         return (start <= target && target <= start + len);
1292 }
1293
1294 static int __kprobes insn_is_indirect_jump(struct insn *insn)
1295 {
1296         int ret = __insn_is_indirect_jump(insn);
1297
1298 #ifdef CONFIG_RETPOLINE
1299         /*
1300          * Jump to x86_indirect_thunk_* is treated as an indirect jump.
1301          * Note that even with CONFIG_RETPOLINE=y, the kernel compiled with
1302          * older gcc may use indirect jump. So we add this check instead of
1303          * replace indirect-jump check.
1304          */
1305         if (!ret)
1306                 ret = insn_jump_into_range(insn,
1307                                 (unsigned long)__indirect_thunk_start,
1308                                 (unsigned long)__indirect_thunk_end -
1309                                 (unsigned long)__indirect_thunk_start);
1310 #endif
1311         return ret;
1312 }
1313
1314 /* Decode whole function to ensure any instructions don't jump into target */
1315 static int __kprobes can_optimize(unsigned long paddr)
1316 {
1317         int ret;
1318         unsigned long addr, size = 0, offset = 0;
1319         struct insn insn;
1320         kprobe_opcode_t buf[MAX_INSN_SIZE];
1321
1322         /* Lookup symbol including addr */
1323         if (!kallsyms_lookup_size_offset(paddr, &size, &offset))
1324                 return 0;
1325
1326         /*
1327          * Do not optimize in the entry code due to the unstable
1328          * stack handling.
1329          */
1330         if ((paddr >= (unsigned long )__entry_text_start) &&
1331             (paddr <  (unsigned long )__entry_text_end))
1332                 return 0;
1333
1334         /* Check there is enough space for a relative jump. */
1335         if (size - offset < RELATIVEJUMP_SIZE)
1336                 return 0;
1337
1338         /* Decode instructions */
1339         addr = paddr - offset;
1340         while (addr < paddr - offset + size) { /* Decode until function end */
1341                 if (search_exception_tables(addr))
1342                         /*
1343                          * Since some fixup code will jumps into this function,
1344                          * we can't optimize kprobe in this function.
1345                          */
1346                         return 0;
1347                 kernel_insn_init(&insn, (void *)addr);
1348                 insn_get_opcode(&insn);
1349                 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION) {
1350                         ret = recover_probed_instruction(buf, addr);
1351                         if (ret)
1352                                 return 0;
1353                         kernel_insn_init(&insn, buf);
1354                 }
1355                 insn_get_length(&insn);
1356                 /* Recover address */
1357                 insn.kaddr = (void *)addr;
1358                 insn.next_byte = (void *)(addr + insn.length);
1359                 /* Check any instructions don't jump into target */
1360                 if (insn_is_indirect_jump(&insn) ||
1361                     insn_jump_into_range(&insn, paddr + INT3_SIZE,
1362                                          RELATIVE_ADDR_SIZE))
1363                         return 0;
1364                 addr += insn.length;
1365         }
1366
1367         return 1;
1368 }
1369
1370 /* Check optimized_kprobe can actually be optimized. */
1371 int __kprobes arch_check_optimized_kprobe(struct optimized_kprobe *op)
1372 {
1373         int i;
1374         struct kprobe *p;
1375
1376         for (i = 1; i < op->optinsn.size; i++) {
1377                 p = get_kprobe(op->kp.addr + i);
1378                 if (p && !kprobe_disabled(p))
1379                         return -EEXIST;
1380         }
1381
1382         return 0;
1383 }
1384
1385 /* Check the addr is within the optimized instructions. */
1386 int __kprobes arch_within_optimized_kprobe(struct optimized_kprobe *op,
1387                                            unsigned long addr)
1388 {
1389         return ((unsigned long)op->kp.addr <= addr &&
1390                 (unsigned long)op->kp.addr + op->optinsn.size > addr);
1391 }
1392
1393 /* Free optimized instruction slot */
1394 static __kprobes
1395 void __arch_remove_optimized_kprobe(struct optimized_kprobe *op, int dirty)
1396 {
1397         if (op->optinsn.insn) {
1398                 free_optinsn_slot(op->optinsn.insn, dirty);
1399                 op->optinsn.insn = NULL;
1400                 op->optinsn.size = 0;
1401         }
1402 }
1403
1404 void __kprobes arch_remove_optimized_kprobe(struct optimized_kprobe *op)
1405 {
1406         __arch_remove_optimized_kprobe(op, 1);
1407 }
1408
1409 /*
1410  * Copy replacing target instructions
1411  * Target instructions MUST be relocatable (checked inside)
1412  */
1413 int __kprobes arch_prepare_optimized_kprobe(struct optimized_kprobe *op)
1414 {
1415         u8 *buf;
1416         int ret;
1417         long rel;
1418
1419         if (!can_optimize((unsigned long)op->kp.addr))
1420                 return -EILSEQ;
1421
1422         op->optinsn.insn = get_optinsn_slot();
1423         if (!op->optinsn.insn)
1424                 return -ENOMEM;
1425
1426         /*
1427          * Verify if the address gap is in 2GB range, because this uses
1428          * a relative jump.
1429          */
1430         rel = (long)op->optinsn.insn - (long)op->kp.addr + RELATIVEJUMP_SIZE;
1431         if (abs(rel) > 0x7fffffff)
1432                 return -ERANGE;
1433
1434         buf = (u8 *)op->optinsn.insn;
1435
1436         /* Copy instructions into the out-of-line buffer */
1437         ret = copy_optimized_instructions(buf + TMPL_END_IDX, op->kp.addr);
1438         if (ret < 0) {
1439                 __arch_remove_optimized_kprobe(op, 0);
1440                 return ret;
1441         }
1442         op->optinsn.size = ret;
1443
1444         /* Copy arch-dep-instance from template */
1445         memcpy(buf, &optprobe_template_entry, TMPL_END_IDX);
1446
1447         /* Set probe information */
1448         synthesize_set_arg1(buf + TMPL_MOVE_IDX, (unsigned long)op);
1449
1450         /* Set probe function call */
1451         synthesize_relcall(buf + TMPL_CALL_IDX, optimized_callback);
1452
1453         /* Set returning jmp instruction at the tail of out-of-line buffer */
1454         synthesize_reljump(buf + TMPL_END_IDX + op->optinsn.size,
1455                            (u8 *)op->kp.addr + op->optinsn.size);
1456
1457         flush_icache_range((unsigned long) buf,
1458                            (unsigned long) buf + TMPL_END_IDX +
1459                            op->optinsn.size + RELATIVEJUMP_SIZE);
1460         return 0;
1461 }
1462
1463 #define MAX_OPTIMIZE_PROBES 256
1464 static struct text_poke_param *jump_poke_params;
1465 static struct jump_poke_buffer {
1466         u8 buf[RELATIVEJUMP_SIZE];
1467 } *jump_poke_bufs;
1468
1469 static void __kprobes setup_optimize_kprobe(struct text_poke_param *tprm,
1470                                             u8 *insn_buf,
1471                                             struct optimized_kprobe *op)
1472 {
1473         s32 rel = (s32)((long)op->optinsn.insn -
1474                         ((long)op->kp.addr + RELATIVEJUMP_SIZE));
1475
1476         /* Backup instructions which will be replaced by jump address */
1477         memcpy(op->optinsn.copied_insn, op->kp.addr + INT3_SIZE,
1478                RELATIVE_ADDR_SIZE);
1479
1480         insn_buf[0] = RELATIVEJUMP_OPCODE;
1481         *(s32 *)(&insn_buf[1]) = rel;
1482
1483         tprm->addr = op->kp.addr;
1484         tprm->opcode = insn_buf;
1485         tprm->len = RELATIVEJUMP_SIZE;
1486 }
1487
1488 /*
1489  * Replace breakpoints (int3) with relative jumps.
1490  * Caller must call with locking kprobe_mutex and text_mutex.
1491  */
1492 void __kprobes arch_optimize_kprobes(struct list_head *oplist)
1493 {
1494         struct optimized_kprobe *op, *tmp;
1495         int c = 0;
1496
1497         list_for_each_entry_safe(op, tmp, oplist, list) {
1498                 WARN_ON(kprobe_disabled(&op->kp));
1499                 /* Setup param */
1500                 setup_optimize_kprobe(&jump_poke_params[c],
1501                                       jump_poke_bufs[c].buf, op);
1502                 list_del_init(&op->list);
1503                 if (++c >= MAX_OPTIMIZE_PROBES)
1504                         break;
1505         }
1506
1507         /*
1508          * text_poke_smp doesn't support NMI/MCE code modifying.
1509          * However, since kprobes itself also doesn't support NMI/MCE
1510          * code probing, it's not a problem.
1511          */
1512         text_poke_smp_batch(jump_poke_params, c);
1513 }
1514
1515 static void __kprobes setup_unoptimize_kprobe(struct text_poke_param *tprm,
1516                                               u8 *insn_buf,
1517                                               struct optimized_kprobe *op)
1518 {
1519         /* Set int3 to first byte for kprobes */
1520         insn_buf[0] = BREAKPOINT_INSTRUCTION;
1521         memcpy(insn_buf + 1, op->optinsn.copied_insn, RELATIVE_ADDR_SIZE);
1522
1523         tprm->addr = op->kp.addr;
1524         tprm->opcode = insn_buf;
1525         tprm->len = RELATIVEJUMP_SIZE;
1526 }
1527
1528 /*
1529  * Recover original instructions and breakpoints from relative jumps.
1530  * Caller must call with locking kprobe_mutex.
1531  */
1532 extern void arch_unoptimize_kprobes(struct list_head *oplist,
1533                                     struct list_head *done_list)
1534 {
1535         struct optimized_kprobe *op, *tmp;
1536         int c = 0;
1537
1538         list_for_each_entry_safe(op, tmp, oplist, list) {
1539                 /* Setup param */
1540                 setup_unoptimize_kprobe(&jump_poke_params[c],
1541                                         jump_poke_bufs[c].buf, op);
1542                 list_move(&op->list, done_list);
1543                 if (++c >= MAX_OPTIMIZE_PROBES)
1544                         break;
1545         }
1546
1547         /*
1548          * text_poke_smp doesn't support NMI/MCE code modifying.
1549          * However, since kprobes itself also doesn't support NMI/MCE
1550          * code probing, it's not a problem.
1551          */
1552         text_poke_smp_batch(jump_poke_params, c);
1553 }
1554
1555 /* Replace a relative jump with a breakpoint (int3).  */
1556 void __kprobes arch_unoptimize_kprobe(struct optimized_kprobe *op)
1557 {
1558         u8 buf[RELATIVEJUMP_SIZE];
1559
1560         /* Set int3 to first byte for kprobes */
1561         buf[0] = BREAKPOINT_INSTRUCTION;
1562         memcpy(buf + 1, op->optinsn.copied_insn, RELATIVE_ADDR_SIZE);
1563         text_poke_smp(op->kp.addr, buf, RELATIVEJUMP_SIZE);
1564 }
1565
1566 static int  __kprobes setup_detour_execution(struct kprobe *p,
1567                                              struct pt_regs *regs,
1568                                              int reenter)
1569 {
1570         struct optimized_kprobe *op;
1571
1572         if (p->flags & KPROBE_FLAG_OPTIMIZED) {
1573                 /* This kprobe is really able to run optimized path. */
1574                 op = container_of(p, struct optimized_kprobe, kp);
1575                 /* Detour through copied instructions */
1576                 regs->ip = (unsigned long)op->optinsn.insn + TMPL_END_IDX;
1577                 if (!reenter)
1578                         reset_current_kprobe();
1579                 preempt_enable_no_resched();
1580                 return 1;
1581         }
1582         return 0;
1583 }
1584
1585 static int __kprobes init_poke_params(void)
1586 {
1587         /* Allocate code buffer and parameter array */
1588         jump_poke_bufs = kmalloc(sizeof(struct jump_poke_buffer) *
1589                                  MAX_OPTIMIZE_PROBES, GFP_KERNEL);
1590         if (!jump_poke_bufs)
1591                 return -ENOMEM;
1592
1593         jump_poke_params = kmalloc(sizeof(struct text_poke_param) *
1594                                    MAX_OPTIMIZE_PROBES, GFP_KERNEL);
1595         if (!jump_poke_params) {
1596                 kfree(jump_poke_bufs);
1597                 jump_poke_bufs = NULL;
1598                 return -ENOMEM;
1599         }
1600
1601         return 0;
1602 }
1603 #else   /* !CONFIG_OPTPROBES */
1604 static int __kprobes init_poke_params(void)
1605 {
1606         return 0;
1607 }
1608 #endif
1609
1610 int __init arch_init_kprobes(void)
1611 {
1612         return init_poke_params();
1613 }
1614
1615 int __kprobes arch_trampoline_kprobe(struct kprobe *p)
1616 {
1617         return 0;
1618 }