arch/x86/kernel/kprobes/opt.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  *  Kernel Probes Jump Optimization (Optprobes)
   4  *
   5  * Copyright (C) IBM Corporation, 2002, 2004
   6  * Copyright (C) Hitachi Ltd., 2012
   7  */
   8 #include <linux/kprobes.h>
   9 #include <linux/perf_event.h>
  10 #include <linux/ptrace.h>
  11 #include <linux/string.h>
  12 #include <linux/slab.h>
  13 #include <linux/hardirq.h>
  14 #include <linux/preempt.h>
  15 #include <linux/extable.h>
  16 #include <linux/kdebug.h>
  17 #include <linux/kallsyms.h>
  18 #include <linux/ftrace.h>
  19 #include <linux/objtool.h>
  20 #include <linux/pgtable.h>
  21 #include <linux/static_call.h>
  22
  23 #include <asm/text-patching.h>
  24 #include <asm/cacheflush.h>
  25 #include <asm/desc.h>
  26 #include <linux/uaccess.h>
  27 #include <asm/alternative.h>
  28 #include <asm/insn.h>
  29 #include <asm/debugreg.h>
  30 #include <asm/set_memory.h>
  31 #include <asm/sections.h>
  32 #include <asm/nospec-branch.h>
  33
  34 #include "common.h"
  35
  36 unsigned long __recover_optprobed_insn(kprobe_opcode_t *buf, unsigned long addr)
  37 {
  38         struct optimized_kprobe *op;
  39         struct kprobe *kp;
  40         long offs;
  41         int i;
  42
  43         for (i = 0; i < JMP32_INSN_SIZE; i++) {
  44                 kp = get_kprobe((void *)addr - i);
  45                 /* This function only handles jump-optimized kprobe */
  46                 if (kp && kprobe_optimized(kp)) {
  47                         op = container_of(kp, struct optimized_kprobe, kp);
  48                         /* If op->list is not empty, op is under optimizing */
  49                         if (list_empty(&op->list))
  50                                 goto found;
  51                 }
  52         }
  53
  54         return addr;
  55 found:
  56         /*
  57          * If the kprobe can be optimized, original bytes which can be
  58          * overwritten by jump destination address. In this case, original
  59          * bytes must be recovered from op->optinsn.copied_insn buffer.
  60          */
  61         if (copy_from_kernel_nofault(buf, (void *)addr,
  62                 MAX_INSN_SIZE * sizeof(kprobe_opcode_t)))
  63                 return 0UL;
  64
  65         if (addr == (unsigned long)kp->addr) {
  66                 buf[0] = kp->opcode;
  67                 memcpy(buf + 1, op->optinsn.copied_insn, DISP32_SIZE);
  68         } else {
  69                 offs = addr - (unsigned long)kp->addr - 1;
  70                 memcpy(buf, op->optinsn.copied_insn + offs, DISP32_SIZE - offs);
  71         }
  72
  73         return (unsigned long)buf;
  74 }
  75
  76 static void synthesize_clac(kprobe_opcode_t *addr)
  77 {
  78         /*
  79          * Can't be static_cpu_has() due to how objtool treats this feature bit.
  80          * This isn't a fast path anyway.
  81          */
  82         if (!boot_cpu_has(X86_FEATURE_SMAP))
  83                 return;
  84
  85         /* Replace the NOP3 with CLAC */
  86         addr[0] = 0x0f;
  87         addr[1] = 0x01;
  88         addr[2] = 0xca;
  89 }
  90
  91 /* Insert a move instruction which sets a pointer to eax/rdi (1st arg). */
  92 static void synthesize_set_arg1(kprobe_opcode_t *addr, unsigned long val)
  93 {
  94 #ifdef CONFIG_X86_64
  95         *addr++ = 0x48;
  96         *addr++ = 0xbf;
  97 #else
  98         *addr++ = 0xb8;
  99 #endif
 100         *(unsigned long *)addr = val;
 101 }
 102
 103 asm (
 104                         ".pushsection .rodata\n"
 105                         "optprobe_template_func:\n"
 106                         ".global optprobe_template_entry\n"
 107                         "optprobe_template_entry:\n"
 108 #ifdef CONFIG_X86_64
 109                         /* We don't bother saving the ss register */
 110                         "       pushq %rsp\n"
 111                         "       pushfq\n"
 112                         ".global optprobe_template_clac\n"
 113                         "optprobe_template_clac:\n"
 114                         ASM_NOP3
 115                         SAVE_REGS_STRING
 116                         "       movq %rsp, %rsi\n"
 117                         ".global optprobe_template_val\n"
 118                         "optprobe_template_val:\n"
 119                         ASM_NOP5
 120                         ASM_NOP5
 121                         ".global optprobe_template_call\n"
 122                         "optprobe_template_call:\n"
 123                         ASM_NOP5
 124                         /* Move flags to rsp */
 125                         "       movq 18*8(%rsp), %rdx\n"
 126                         "       movq %rdx, 19*8(%rsp)\n"
 127                         RESTORE_REGS_STRING
 128                         /* Skip flags entry */
 129                         "       addq $8, %rsp\n"
 130                         "       popfq\n"
 131 #else /* CONFIG_X86_32 */
 132                         "       pushl %esp\n"
 133                         "       pushfl\n"
 134                         ".global optprobe_template_clac\n"
 135                         "optprobe_template_clac:\n"
 136                         ASM_NOP3
 137                         SAVE_REGS_STRING
 138                         "       movl %esp, %edx\n"
 139                         ".global optprobe_template_val\n"
 140                         "optprobe_template_val:\n"
 141                         ASM_NOP5
 142                         ".global optprobe_template_call\n"
 143                         "optprobe_template_call:\n"
 144                         ASM_NOP5
 145                         /* Move flags into esp */
 146                         "       movl 14*4(%esp), %edx\n"
 147                         "       movl %edx, 15*4(%esp)\n"
 148                         RESTORE_REGS_STRING
 149                         /* Skip flags entry */
 150                         "       addl $4, %esp\n"
 151                         "       popfl\n"
 152 #endif
 153                         ".global optprobe_template_end\n"
 154                         "optprobe_template_end:\n"
 155                         ".popsection\n");
 156
 157 void optprobe_template_func(void);
 158 STACK_FRAME_NON_STANDARD(optprobe_template_func);
 159
 160 #define TMPL_CLAC_IDX \
 161         ((long)optprobe_template_clac - (long)optprobe_template_entry)
 162 #define TMPL_MOVE_IDX \
 163         ((long)optprobe_template_val - (long)optprobe_template_entry)
 164 #define TMPL_CALL_IDX \
 165         ((long)optprobe_template_call - (long)optprobe_template_entry)
 166 #define TMPL_END_IDX \
 167         ((long)optprobe_template_end - (long)optprobe_template_entry)
 168
 169 /* Optimized kprobe call back function: called from optinsn */
 170 static void
 171 optimized_callback(struct optimized_kprobe *op, struct pt_regs *regs)
 172 {
 173         /* This is possible if op is under delayed unoptimizing */
 174         if (kprobe_disabled(&op->kp))
 175                 return;
 176
 177         preempt_disable();
 178         if (kprobe_running()) {
 179                 kprobes_inc_nmissed_count(&op->kp);
 180         } else {
 181                 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 182                 /* Save skipped registers */
 183                 regs->cs = __KERNEL_CS;
 184 #ifdef CONFIG_X86_32
 185                 regs->gs = 0;
 186 #endif
 187                 regs->ip = (unsigned long)op->kp.addr + INT3_INSN_SIZE;
 188                 regs->orig_ax = ~0UL;
 189
 190                 __this_cpu_write(current_kprobe, &op->kp);
 191                 kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 192                 opt_pre_handler(&op->kp, regs);
 193                 __this_cpu_write(current_kprobe, NULL);
 194         }
 195         preempt_enable();
 196 }
 197 NOKPROBE_SYMBOL(optimized_callback);
 198
 199 static int copy_optimized_instructions(u8 *dest, u8 *src, u8 *real)
 200 {
 201         struct insn insn;
 202         int len = 0, ret;
 203
 204         while (len < JMP32_INSN_SIZE) {
 205                 ret = __copy_instruction(dest + len, src + len, real + len, &insn);
 206                 if (!ret || !can_boost(&insn, src + len))
 207                         return -EINVAL;
 208                 len += ret;
 209         }
 210         /* Check whether the address range is reserved */
 211         if (ftrace_text_reserved(src, src + len - 1) ||
 212             alternatives_text_reserved(src, src + len - 1) ||
 213             jump_label_text_reserved(src, src + len - 1) ||
 214             static_call_text_reserved(src, src + len - 1))
 215                 return -EBUSY;
 216
 217         return len;
 218 }
 219
 220 /* Check whether insn is indirect jump */
 221 static int __insn_is_indirect_jump(struct insn *insn)
 222 {
 223         return ((insn->opcode.bytes[0] == 0xff &&
 224                 (X86_MODRM_REG(insn->modrm.value) & 6) == 4) || /* Jump */
 225                 insn->opcode.bytes[0] == 0xea); /* Segment based jump */
 226 }
 227
 228 /* Check whether insn jumps into specified address range */
 229 static int insn_jump_into_range(struct insn *insn, unsigned long start, int len)
 230 {
 231         unsigned long target = 0;
 232
 233         switch (insn->opcode.bytes[0]) {
 234         case 0xe0:      /* loopne */
 235         case 0xe1:      /* loope */
 236         case 0xe2:      /* loop */
 237         case 0xe3:      /* jcxz */
 238         case 0xe9:      /* near relative jump */
 239         case 0xeb:      /* short relative jump */
 240                 break;
 241         case 0x0f:
 242                 if ((insn->opcode.bytes[1] & 0xf0) == 0x80) /* jcc near */
 243                         break;
 244                 return 0;
 245         default:
 246                 if ((insn->opcode.bytes[0] & 0xf0) == 0x70) /* jcc short */
 247                         break;
 248                 return 0;
 249         }
 250         target = (unsigned long)insn->next_byte + insn->immediate.value;
 251
 252         return (start <= target && target <= start + len);
 253 }
 254
 255 static int insn_is_indirect_jump(struct insn *insn)
 256 {
 257         int ret = __insn_is_indirect_jump(insn);
 258
 259 #ifdef CONFIG_RETPOLINE
 260         /*
 261          * Jump to x86_indirect_thunk_* is treated as an indirect jump.
 262          * Note that even with CONFIG_RETPOLINE=y, the kernel compiled with
 263          * older gcc may use indirect jump. So we add this check instead of
 264          * replace indirect-jump check.
 265          */
 266         if (!ret)
 267                 ret = insn_jump_into_range(insn,
 268                                 (unsigned long)__indirect_thunk_start,
 269                                 (unsigned long)__indirect_thunk_end -
 270                                 (unsigned long)__indirect_thunk_start);
 271 #endif
 272         return ret;
 273 }
 274
 275 static bool is_padding_int3(unsigned long addr, unsigned long eaddr)
 276 {
 277         unsigned char ops;
 278
 279         for (; addr < eaddr; addr++) {
 280                 if (get_kernel_nofault(ops, (void *)addr) < 0 ||
 281                     ops != INT3_INSN_OPCODE)
 282                         return false;
 283         }
 284
 285         return true;
 286 }
 287
 288 /* Decode whole function to ensure any instructions don't jump into target */
 289 static int can_optimize(unsigned long paddr)
 290 {
 291         unsigned long addr, size = 0, offset = 0;
 292         struct insn insn;
 293         kprobe_opcode_t buf[MAX_INSN_SIZE];
 294
 295         /* Lookup symbol including addr */
 296         if (!kallsyms_lookup_size_offset(paddr, &size, &offset))
 297                 return 0;
 298
 299         /*
 300          * Do not optimize in the entry code due to the unstable
 301          * stack handling and registers setup.
 302          */
 303         if (((paddr >= (unsigned long)__entry_text_start) &&
 304              (paddr <  (unsigned long)__entry_text_end)))
 305                 return 0;
 306
 307         /* Check there is enough space for a relative jump. */
 308         if (size - offset < JMP32_INSN_SIZE)
 309                 return 0;
 310
 311         /* Decode instructions */
 312         addr = paddr - offset;
 313         while (addr < paddr - offset + size) { /* Decode until function end */
 314                 unsigned long recovered_insn;
 315                 if (search_exception_tables(addr))
 316                         /*
 317                          * Since some fixup code will jumps into this function,
 318                          * we can't optimize kprobe in this function.
 319                          */
 320                         return 0;
 321                 recovered_insn = recover_probed_instruction(buf, addr);
 322                 if (!recovered_insn)
 323                         return 0;
 324                 kernel_insn_init(&insn, (void *)recovered_insn, MAX_INSN_SIZE);
 325                 insn_get_length(&insn);
 326                 /*
 327                  * In the case of detecting unknown breakpoint, this could be
 328                  * a padding INT3 between functions. Let's check that all the
 329                  * rest of the bytes are also INT3.
 330                  */
 331                 if (insn.opcode.bytes[0] == INT3_INSN_OPCODE)
 332                         return is_padding_int3(addr, paddr - offset + size) ? 1 : 0;
 333
 334                 /* Recover address */
 335                 insn.kaddr = (void *)addr;
 336                 insn.next_byte = (void *)(addr + insn.length);
 337                 /* Check any instructions don't jump into target */
 338                 if (insn_is_indirect_jump(&insn) ||
 339                     insn_jump_into_range(&insn, paddr + INT3_INSN_SIZE,
 340                                          DISP32_SIZE))
 341                         return 0;
 342                 addr += insn.length;
 343         }
 344
 345         return 1;
 346 }
 347
 348 /* Check optimized_kprobe can actually be optimized. */
 349 int arch_check_optimized_kprobe(struct optimized_kprobe *op)
 350 {
 351         int i;
 352         struct kprobe *p;
 353
 354         for (i = 1; i < op->optinsn.size; i++) {
 355                 p = get_kprobe(op->kp.addr + i);
 356                 if (p && !kprobe_disabled(p))
 357                         return -EEXIST;
 358         }
 359
 360         return 0;
 361 }
 362
 363 /* Check the addr is within the optimized instructions. */
 364 int arch_within_optimized_kprobe(struct optimized_kprobe *op,
 365                                  unsigned long addr)
 366 {
 367         return ((unsigned long)op->kp.addr <= addr &&
 368                 (unsigned long)op->kp.addr + op->optinsn.size > addr);
 369 }
 370
 371 /* Free optimized instruction slot */
 372 static
 373 void __arch_remove_optimized_kprobe(struct optimized_kprobe *op, int dirty)
 374 {
 375         u8 *slot = op->optinsn.insn;
 376         if (slot) {
 377                 int len = TMPL_END_IDX + op->optinsn.size + JMP32_INSN_SIZE;
 378
 379                 /* Record the perf event before freeing the slot */
 380                 if (dirty)
 381                         perf_event_text_poke(slot, slot, len, NULL, 0);
 382
 383                 free_optinsn_slot(slot, dirty);
 384                 op->optinsn.insn = NULL;
 385                 op->optinsn.size = 0;
 386         }
 387 }
 388
 389 void arch_remove_optimized_kprobe(struct optimized_kprobe *op)
 390 {
 391         __arch_remove_optimized_kprobe(op, 1);
 392 }
 393
 394 /*
 395  * Copy replacing target instructions
 396  * Target instructions MUST be relocatable (checked inside)
 397  * This is called when new aggr(opt)probe is allocated or reused.
 398  */
 399 int arch_prepare_optimized_kprobe(struct optimized_kprobe *op,
 400                                   struct kprobe *__unused)
 401 {
 402         u8 *buf = NULL, *slot;
 403         int ret, len;
 404         long rel;
 405
 406         if (!can_optimize((unsigned long)op->kp.addr))
 407                 return -EILSEQ;
 408
 409         buf = kzalloc(MAX_OPTINSN_SIZE, GFP_KERNEL);
 410         if (!buf)
 411                 return -ENOMEM;
 412
 413         op->optinsn.insn = slot = get_optinsn_slot();
 414         if (!slot) {
 415                 ret = -ENOMEM;
 416                 goto out;
 417         }
 418
 419         /*
 420          * Verify if the address gap is in 2GB range, because this uses
 421          * a relative jump.
 422          */
 423         rel = (long)slot - (long)op->kp.addr + JMP32_INSN_SIZE;
 424         if (abs(rel) > 0x7fffffff) {
 425                 ret = -ERANGE;
 426                 goto err;
 427         }
 428
 429         /* Copy arch-dep-instance from template */
 430         memcpy(buf, optprobe_template_entry, TMPL_END_IDX);
 431
 432         /* Copy instructions into the out-of-line buffer */
 433         ret = copy_optimized_instructions(buf + TMPL_END_IDX, op->kp.addr,
 434                                           slot + TMPL_END_IDX);
 435         if (ret < 0)
 436                 goto err;
 437         op->optinsn.size = ret;
 438         len = TMPL_END_IDX + op->optinsn.size;
 439
 440         synthesize_clac(buf + TMPL_CLAC_IDX);
 441
 442         /* Set probe information */
 443         synthesize_set_arg1(buf + TMPL_MOVE_IDX, (unsigned long)op);
 444
 445         /* Set probe function call */
 446         synthesize_relcall(buf + TMPL_CALL_IDX,
 447                            slot + TMPL_CALL_IDX, optimized_callback);
 448
 449         /* Set returning jmp instruction at the tail of out-of-line buffer */
 450         synthesize_reljump(buf + len, slot + len,
 451                            (u8 *)op->kp.addr + op->optinsn.size);
 452         len += JMP32_INSN_SIZE;
 453
 454         /*
 455          * Note len = TMPL_END_IDX + op->optinsn.size + JMP32_INSN_SIZE is also
 456          * used in __arch_remove_optimized_kprobe().
 457          */
 458
 459         /* We have to use text_poke() for instruction buffer because it is RO */
 460         perf_event_text_poke(slot, NULL, 0, buf, len);
 461         text_poke(slot, buf, len);
 462
 463         ret = 0;
 464 out:
 465         kfree(buf);
 466         return ret;
 467
 468 err:
 469         __arch_remove_optimized_kprobe(op, 0);
 470         goto out;
 471 }
 472
 473 /*
 474  * Replace breakpoints (INT3) with relative jumps (JMP.d32).
 475  * Caller must call with locking kprobe_mutex and text_mutex.
 476  *
 477  * The caller will have installed a regular kprobe and after that issued
 478  * syncrhonize_rcu_tasks(), this ensures that the instruction(s) that live in
 479  * the 4 bytes after the INT3 are unused and can now be overwritten.
 480  */
 481 void arch_optimize_kprobes(struct list_head *oplist)
 482 {
 483         struct optimized_kprobe *op, *tmp;
 484         u8 insn_buff[JMP32_INSN_SIZE];
 485
 486         list_for_each_entry_safe(op, tmp, oplist, list) {
 487                 s32 rel = (s32)((long)op->optinsn.insn -
 488                         ((long)op->kp.addr + JMP32_INSN_SIZE));
 489
 490                 WARN_ON(kprobe_disabled(&op->kp));
 491
 492                 /* Backup instructions which will be replaced by jump address */
 493                 memcpy(op->optinsn.copied_insn, op->kp.addr + INT3_INSN_SIZE,
 494                        DISP32_SIZE);
 495
 496                 insn_buff[0] = JMP32_INSN_OPCODE;
 497                 *(s32 *)(&insn_buff[1]) = rel;
 498
 499                 text_poke_bp(op->kp.addr, insn_buff, JMP32_INSN_SIZE, NULL);
 500
 501                 list_del_init(&op->list);
 502         }
 503 }
 504
 505 /*
 506  * Replace a relative jump (JMP.d32) with a breakpoint (INT3).
 507  *
 508  * After that, we can restore the 4 bytes after the INT3 to undo what
 509  * arch_optimize_kprobes() scribbled. This is safe since those bytes will be
 510  * unused once the INT3 lands.
 511  */
 512 void arch_unoptimize_kprobe(struct optimized_kprobe *op)
 513 {
 514         u8 new[JMP32_INSN_SIZE] = { INT3_INSN_OPCODE, };
 515         u8 old[JMP32_INSN_SIZE];
 516         u8 *addr = op->kp.addr;
 517
 518         memcpy(old, op->kp.addr, JMP32_INSN_SIZE);
 519         memcpy(new + INT3_INSN_SIZE,
 520                op->optinsn.copied_insn,
 521                JMP32_INSN_SIZE - INT3_INSN_SIZE);
 522
 523         text_poke(addr, new, INT3_INSN_SIZE);
 524         text_poke_sync();
 525         text_poke(addr + INT3_INSN_SIZE,
 526                   new + INT3_INSN_SIZE,
 527                   JMP32_INSN_SIZE - INT3_INSN_SIZE);
 528         text_poke_sync();
 529
 530         perf_event_text_poke(op->kp.addr, old, JMP32_INSN_SIZE, new, JMP32_INSN_SIZE);
 531 }
 532
 533 /*
 534  * Recover original instructions and breakpoints from relative jumps.
 535  * Caller must call with locking kprobe_mutex.
 536  */
 537 extern void arch_unoptimize_kprobes(struct list_head *oplist,
 538                                     struct list_head *done_list)
 539 {
 540         struct optimized_kprobe *op, *tmp;
 541
 542         list_for_each_entry_safe(op, tmp, oplist, list) {
 543                 arch_unoptimize_kprobe(op);
 544                 list_move(&op->list, done_list);
 545         }
 546 }
 547
 548 int setup_detour_execution(struct kprobe *p, struct pt_regs *regs, int reenter)
 549 {
 550         struct optimized_kprobe *op;
 551
 552         if (p->flags & KPROBE_FLAG_OPTIMIZED) {
 553                 /* This kprobe is really able to run optimized path. */
 554                 op = container_of(p, struct optimized_kprobe, kp);
 555                 /* Detour through copied instructions */
 556                 regs->ip = (unsigned long)op->optinsn.insn + TMPL_END_IDX;
 557                 if (!reenter)
 558                         reset_current_kprobe();
 559                 return 1;
 560         }
 561         return 0;
 562 }
 563 NOKPROBE_SYMBOL(setup_detour_execution);