arch/x86/kernel/kprobes/opt.c

   1 /*
   2  *  Kernel Probes Jump Optimization (Optprobes)
   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  * Copyright (C) Hitachi Ltd., 2012
  20  */
  21 #include <linux/kprobes.h>
  22 #include <linux/ptrace.h>
  23 #include <linux/string.h>
  24 #include <linux/slab.h>
  25 #include <linux/hardirq.h>
  26 #include <linux/preempt.h>
  27 #include <linux/extable.h>
  28 #include <linux/kdebug.h>
  29 #include <linux/kallsyms.h>
  30 #include <linux/ftrace.h>
  31 #include <linux/frame.h>
  32
  33 #include <asm/text-patching.h>
  34 #include <asm/cacheflush.h>
  35 #include <asm/desc.h>
  36 #include <asm/pgtable.h>
  37 #include <linux/uaccess.h>
  38 #include <asm/alternative.h>
  39 #include <asm/insn.h>
  40 #include <asm/debugreg.h>
  41 #include <asm/set_memory.h>
  42 #include <asm/sections.h>
  43 #include <asm/nospec-branch.h>
  44
  45 #include "common.h"
  46
  47 unsigned long __recover_optprobed_insn(kprobe_opcode_t *buf, unsigned long addr)
  48 {
  49         struct optimized_kprobe *op;
  50         struct kprobe *kp;
  51         long offs;
  52         int i;
  53
  54         for (i = 0; i < RELATIVEJUMP_SIZE; i++) {
  55                 kp = get_kprobe((void *)addr - i);
  56                 /* This function only handles jump-optimized kprobe */
  57                 if (kp && kprobe_optimized(kp)) {
  58                         op = container_of(kp, struct optimized_kprobe, kp);
  59                         /* If op is optimized or under unoptimizing */
  60                         if (list_empty(&op->list) || optprobe_queued_unopt(op))
  61                                 goto found;
  62                 }
  63         }
  64
  65         return addr;
  66 found:
  67         /*
  68          * If the kprobe can be optimized, original bytes which can be
  69          * overwritten by jump destination address. In this case, original
  70          * bytes must be recovered from op->optinsn.copied_insn buffer.
  71          */
  72         if (probe_kernel_read(buf, (void *)addr,
  73                 MAX_INSN_SIZE * sizeof(kprobe_opcode_t)))
  74                 return 0UL;
  75
  76         if (addr == (unsigned long)kp->addr) {
  77                 buf[0] = kp->opcode;
  78                 memcpy(buf + 1, op->optinsn.copied_insn, RELATIVE_ADDR_SIZE);
  79         } else {
  80                 offs = addr - (unsigned long)kp->addr - 1;
  81                 memcpy(buf, op->optinsn.copied_insn + offs, RELATIVE_ADDR_SIZE - offs);
  82         }
  83
  84         return (unsigned long)buf;
  85 }
  86
  87 /* Insert a move instruction which sets a pointer to eax/rdi (1st arg). */
  88 static void synthesize_set_arg1(kprobe_opcode_t *addr, unsigned long val)
  89 {
  90 #ifdef CONFIG_X86_64
  91         *addr++ = 0x48;
  92         *addr++ = 0xbf;
  93 #else
  94         *addr++ = 0xb8;
  95 #endif
  96         *(unsigned long *)addr = val;
  97 }
  98
  99 asm (
 100                         "optprobe_template_func:\n"
 101                         ".global optprobe_template_entry\n"
 102                         "optprobe_template_entry:\n"
 103 #ifdef CONFIG_X86_64
 104                         /* We don't bother saving the ss register */
 105                         "       pushq %rsp\n"
 106                         "       pushfq\n"
 107                         SAVE_REGS_STRING
 108                         "       movq %rsp, %rsi\n"
 109                         ".global optprobe_template_val\n"
 110                         "optprobe_template_val:\n"
 111                         ASM_NOP5
 112                         ASM_NOP5
 113                         ".global optprobe_template_call\n"
 114                         "optprobe_template_call:\n"
 115                         ASM_NOP5
 116                         /* Move flags to rsp */
 117                         "       movq 144(%rsp), %rdx\n"
 118                         "       movq %rdx, 152(%rsp)\n"
 119                         RESTORE_REGS_STRING
 120                         /* Skip flags entry */
 121                         "       addq $8, %rsp\n"
 122                         "       popfq\n"
 123 #else /* CONFIG_X86_32 */
 124                         "       pushf\n"
 125                         SAVE_REGS_STRING
 126                         "       movl %esp, %edx\n"
 127                         ".global optprobe_template_val\n"
 128                         "optprobe_template_val:\n"
 129                         ASM_NOP5
 130                         ".global optprobe_template_call\n"
 131                         "optprobe_template_call:\n"
 132                         ASM_NOP5
 133                         RESTORE_REGS_STRING
 134                         "       addl $4, %esp\n"        /* skip cs */
 135                         "       popf\n"
 136 #endif
 137                         ".global optprobe_template_end\n"
 138                         "optprobe_template_end:\n"
 139                         ".type optprobe_template_func, @function\n"
 140                         ".size optprobe_template_func, .-optprobe_template_func\n");
 141
 142 void optprobe_template_func(void);
 143 STACK_FRAME_NON_STANDARD(optprobe_template_func);
 144 NOKPROBE_SYMBOL(optprobe_template_func);
 145 NOKPROBE_SYMBOL(optprobe_template_entry);
 146 NOKPROBE_SYMBOL(optprobe_template_val);
 147 NOKPROBE_SYMBOL(optprobe_template_call);
 148 NOKPROBE_SYMBOL(optprobe_template_end);
 149
 150 #define TMPL_MOVE_IDX \
 151         ((long)optprobe_template_val - (long)optprobe_template_entry)
 152 #define TMPL_CALL_IDX \
 153         ((long)optprobe_template_call - (long)optprobe_template_entry)
 154 #define TMPL_END_IDX \
 155         ((long)optprobe_template_end - (long)optprobe_template_entry)
 156
 157 #define INT3_SIZE sizeof(kprobe_opcode_t)
 158
 159 /* Optimized kprobe call back function: called from optinsn */
 160 static void
 161 optimized_callback(struct optimized_kprobe *op, struct pt_regs *regs)
 162 {
 163         /* This is possible if op is under delayed unoptimizing */
 164         if (kprobe_disabled(&op->kp))
 165                 return;
 166
 167         preempt_disable();
 168         if (kprobe_running()) {
 169                 kprobes_inc_nmissed_count(&op->kp);
 170         } else {
 171                 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 172                 /* Save skipped registers */
 173 #ifdef CONFIG_X86_64
 174                 regs->cs = __KERNEL_CS;
 175 #else
 176                 regs->cs = __KERNEL_CS | get_kernel_rpl();
 177                 regs->gs = 0;
 178 #endif
 179                 regs->ip = (unsigned long)op->kp.addr + INT3_SIZE;
 180                 regs->orig_ax = ~0UL;
 181
 182                 __this_cpu_write(current_kprobe, &op->kp);
 183                 kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 184                 opt_pre_handler(&op->kp, regs);
 185                 __this_cpu_write(current_kprobe, NULL);
 186         }
 187         preempt_enable();
 188 }
 189 NOKPROBE_SYMBOL(optimized_callback);
 190
 191 static int copy_optimized_instructions(u8 *dest, u8 *src, u8 *real)
 192 {
 193         struct insn insn;
 194         int len = 0, ret;
 195
 196         while (len < RELATIVEJUMP_SIZE) {
 197                 ret = __copy_instruction(dest + len, src + len, real + len, &insn);
 198                 if (!ret || !can_boost(&insn, src + len))
 199                         return -EINVAL;
 200                 len += ret;
 201         }
 202         /* Check whether the address range is reserved */
 203         if (ftrace_text_reserved(src, src + len - 1) ||
 204             alternatives_text_reserved(src, src + len - 1) ||
 205             jump_label_text_reserved(src, src + len - 1))
 206                 return -EBUSY;
 207
 208         return len;
 209 }
 210
 211 /* Check whether insn is indirect jump */
 212 static int __insn_is_indirect_jump(struct insn *insn)
 213 {
 214         return ((insn->opcode.bytes[0] == 0xff &&
 215                 (X86_MODRM_REG(insn->modrm.value) & 6) == 4) || /* Jump */
 216                 insn->opcode.bytes[0] == 0xea); /* Segment based jump */
 217 }
 218
 219 /* Check whether insn jumps into specified address range */
 220 static int insn_jump_into_range(struct insn *insn, unsigned long start, int len)
 221 {
 222         unsigned long target = 0;
 223
 224         switch (insn->opcode.bytes[0]) {
 225         case 0xe0:      /* loopne */
 226         case 0xe1:      /* loope */
 227         case 0xe2:      /* loop */
 228         case 0xe3:      /* jcxz */
 229         case 0xe9:      /* near relative jump */
 230         case 0xeb:      /* short relative jump */
 231                 break;
 232         case 0x0f:
 233                 if ((insn->opcode.bytes[1] & 0xf0) == 0x80) /* jcc near */
 234                         break;
 235                 return 0;
 236         default:
 237                 if ((insn->opcode.bytes[0] & 0xf0) == 0x70) /* jcc short */
 238                         break;
 239                 return 0;
 240         }
 241         target = (unsigned long)insn->next_byte + insn->immediate.value;
 242
 243         return (start <= target && target <= start + len);
 244 }
 245
 246 static int insn_is_indirect_jump(struct insn *insn)
 247 {
 248         int ret = __insn_is_indirect_jump(insn);
 249
 250 #ifdef CONFIG_RETPOLINE
 251         /*
 252          * Jump to x86_indirect_thunk_* is treated as an indirect jump.
 253          * Note that even with CONFIG_RETPOLINE=y, the kernel compiled with
 254          * older gcc may use indirect jump. So we add this check instead of
 255          * replace indirect-jump check.
 256          */
 257         if (!ret)
 258                 ret = insn_jump_into_range(insn,
 259                                 (unsigned long)__indirect_thunk_start,
 260                                 (unsigned long)__indirect_thunk_end -
 261                                 (unsigned long)__indirect_thunk_start);
 262 #endif
 263         return ret;
 264 }
 265
 266 /* Decode whole function to ensure any instructions don't jump into target */
 267 static int can_optimize(unsigned long paddr)
 268 {
 269         unsigned long addr, size = 0, offset = 0;
 270         struct insn insn;
 271         kprobe_opcode_t buf[MAX_INSN_SIZE];
 272
 273         /* Lookup symbol including addr */
 274         if (!kallsyms_lookup_size_offset(paddr, &size, &offset))
 275                 return 0;
 276
 277         /*
 278          * Do not optimize in the entry code due to the unstable
 279          * stack handling and registers setup.
 280          */
 281         if (((paddr >= (unsigned long)__entry_text_start) &&
 282              (paddr <  (unsigned long)__entry_text_end)) ||
 283             ((paddr >= (unsigned long)__irqentry_text_start) &&
 284              (paddr <  (unsigned long)__irqentry_text_end)))
 285                 return 0;
 286
 287         /* Check there is enough space for a relative jump. */
 288         if (size - offset < RELATIVEJUMP_SIZE)
 289                 return 0;
 290
 291         /* Decode instructions */
 292         addr = paddr - offset;
 293         while (addr < paddr - offset + size) { /* Decode until function end */
 294                 unsigned long recovered_insn;
 295                 if (search_exception_tables(addr))
 296                         /*
 297                          * Since some fixup code will jumps into this function,
 298                          * we can't optimize kprobe in this function.
 299                          */
 300                         return 0;
 301                 recovered_insn = recover_probed_instruction(buf, addr);
 302                 if (!recovered_insn)
 303                         return 0;
 304                 kernel_insn_init(&insn, (void *)recovered_insn, MAX_INSN_SIZE);
 305                 insn_get_length(&insn);
 306                 /* Another subsystem puts a breakpoint */
 307                 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
 308                         return 0;
 309                 /* Recover address */
 310                 insn.kaddr = (void *)addr;
 311                 insn.next_byte = (void *)(addr + insn.length);
 312                 /* Check any instructions don't jump into target */
 313                 if (insn_is_indirect_jump(&insn) ||
 314                     insn_jump_into_range(&insn, paddr + INT3_SIZE,
 315                                          RELATIVE_ADDR_SIZE))
 316                         return 0;
 317                 addr += insn.length;
 318         }
 319
 320         return 1;
 321 }
 322
 323 /* Check optimized_kprobe can actually be optimized. */
 324 int arch_check_optimized_kprobe(struct optimized_kprobe *op)
 325 {
 326         int i;
 327         struct kprobe *p;
 328
 329         for (i = 1; i < op->optinsn.size; i++) {
 330                 p = get_kprobe(op->kp.addr + i);
 331                 if (p && !kprobe_disarmed(p))
 332                         return -EEXIST;
 333         }
 334
 335         return 0;
 336 }
 337
 338 /* Check the addr is within the optimized instructions. */
 339 int arch_within_optimized_kprobe(struct optimized_kprobe *op,
 340                                  unsigned long addr)
 341 {
 342         return ((unsigned long)op->kp.addr <= addr &&
 343                 (unsigned long)op->kp.addr + op->optinsn.size > addr);
 344 }
 345
 346 /* Free optimized instruction slot */
 347 static
 348 void __arch_remove_optimized_kprobe(struct optimized_kprobe *op, int dirty)
 349 {
 350         if (op->optinsn.insn) {
 351                 free_optinsn_slot(op->optinsn.insn, dirty);
 352                 op->optinsn.insn = NULL;
 353                 op->optinsn.size = 0;
 354         }
 355 }
 356
 357 void arch_remove_optimized_kprobe(struct optimized_kprobe *op)
 358 {
 359         __arch_remove_optimized_kprobe(op, 1);
 360 }
 361
 362 /*
 363  * Copy replacing target instructions
 364  * Target instructions MUST be relocatable (checked inside)
 365  * This is called when new aggr(opt)probe is allocated or reused.
 366  */
 367 int arch_prepare_optimized_kprobe(struct optimized_kprobe *op,
 368                                   struct kprobe *__unused)
 369 {
 370         u8 *buf = NULL, *slot;
 371         int ret, len;
 372         long rel;
 373
 374         if (!can_optimize((unsigned long)op->kp.addr))
 375                 return -EILSEQ;
 376
 377         buf = kzalloc(MAX_OPTINSN_SIZE, GFP_KERNEL);
 378         if (!buf)
 379                 return -ENOMEM;
 380
 381         op->optinsn.insn = slot = get_optinsn_slot();
 382         if (!slot) {
 383                 ret = -ENOMEM;
 384                 goto out;
 385         }
 386
 387         /*
 388          * Verify if the address gap is in 2GB range, because this uses
 389          * a relative jump.
 390          */
 391         rel = (long)slot - (long)op->kp.addr + RELATIVEJUMP_SIZE;
 392         if (abs(rel) > 0x7fffffff) {
 393                 ret = -ERANGE;
 394                 goto err;
 395         }
 396
 397         /* Copy arch-dep-instance from template */
 398         memcpy(buf, optprobe_template_entry, TMPL_END_IDX);
 399
 400         /* Copy instructions into the out-of-line buffer */
 401         ret = copy_optimized_instructions(buf + TMPL_END_IDX, op->kp.addr,
 402                                           slot + TMPL_END_IDX);
 403         if (ret < 0)
 404                 goto err;
 405         op->optinsn.size = ret;
 406         len = TMPL_END_IDX + op->optinsn.size;
 407
 408         /* Set probe information */
 409         synthesize_set_arg1(buf + TMPL_MOVE_IDX, (unsigned long)op);
 410
 411         /* Set probe function call */
 412         synthesize_relcall(buf + TMPL_CALL_IDX,
 413                            slot + TMPL_CALL_IDX, optimized_callback);
 414
 415         /* Set returning jmp instruction at the tail of out-of-line buffer */
 416         synthesize_reljump(buf + len, slot + len,
 417                            (u8 *)op->kp.addr + op->optinsn.size);
 418         len += RELATIVEJUMP_SIZE;
 419
 420         /* We have to use text_poke for instuction buffer because it is RO */
 421         text_poke(slot, buf, len);
 422         ret = 0;
 423 out:
 424         kfree(buf);
 425         return ret;
 426
 427 err:
 428         __arch_remove_optimized_kprobe(op, 0);
 429         goto out;
 430 }
 431
 432 /*
 433  * Replace breakpoints (int3) with relative jumps.
 434  * Caller must call with locking kprobe_mutex and text_mutex.
 435  */
 436 void arch_optimize_kprobes(struct list_head *oplist)
 437 {
 438         struct optimized_kprobe *op, *tmp;
 439         u8 insn_buf[RELATIVEJUMP_SIZE];
 440
 441         list_for_each_entry_safe(op, tmp, oplist, list) {
 442                 s32 rel = (s32)((long)op->optinsn.insn -
 443                         ((long)op->kp.addr + RELATIVEJUMP_SIZE));
 444
 445                 WARN_ON(kprobe_disabled(&op->kp));
 446
 447                 /* Backup instructions which will be replaced by jump address */
 448                 memcpy(op->optinsn.copied_insn, op->kp.addr + INT3_SIZE,
 449                        RELATIVE_ADDR_SIZE);
 450
 451                 insn_buf[0] = RELATIVEJUMP_OPCODE;
 452                 *(s32 *)(&insn_buf[1]) = rel;
 453
 454                 text_poke_bp(op->kp.addr, insn_buf, RELATIVEJUMP_SIZE,
 455                              op->optinsn.insn);
 456
 457                 list_del_init(&op->list);
 458         }
 459 }
 460
 461 /* Replace a relative jump with a breakpoint (int3).  */
 462 void arch_unoptimize_kprobe(struct optimized_kprobe *op)
 463 {
 464         u8 insn_buf[RELATIVEJUMP_SIZE];
 465
 466         /* Set int3 to first byte for kprobes */
 467         insn_buf[0] = BREAKPOINT_INSTRUCTION;
 468         memcpy(insn_buf + 1, op->optinsn.copied_insn, RELATIVE_ADDR_SIZE);
 469         text_poke_bp(op->kp.addr, insn_buf, RELATIVEJUMP_SIZE,
 470                      op->optinsn.insn);
 471 }
 472
 473 /*
 474  * Recover original instructions and breakpoints from relative jumps.
 475  * Caller must call with locking kprobe_mutex.
 476  */
 477 extern void arch_unoptimize_kprobes(struct list_head *oplist,
 478                                     struct list_head *done_list)
 479 {
 480         struct optimized_kprobe *op, *tmp;
 481
 482         list_for_each_entry_safe(op, tmp, oplist, list) {
 483                 arch_unoptimize_kprobe(op);
 484                 list_move(&op->list, done_list);
 485         }
 486 }
 487
 488 int setup_detour_execution(struct kprobe *p, struct pt_regs *regs, int reenter)
 489 {
 490         struct optimized_kprobe *op;
 491
 492         if (p->flags & KPROBE_FLAG_OPTIMIZED) {
 493                 /* This kprobe is really able to run optimized path. */
 494                 op = container_of(p, struct optimized_kprobe, kp);
 495                 /* Detour through copied instructions */
 496                 regs->ip = (unsigned long)op->optinsn.insn + TMPL_END_IDX;
 497                 if (!reenter)
 498                         reset_current_kprobe();
 499                 return 1;
 500         }
 501         return 0;
 502 }
 503 NOKPROBE_SYMBOL(setup_detour_execution);