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                         "       pushq $" __stringify(__KERNEL_DS) "\n"
 110                         /* Save the 'sp - 8', this will be fixed later. */
 111                         "       pushq %rsp\n"
 112                         "       pushfq\n"
 113                         ".global optprobe_template_clac\n"
 114                         "optprobe_template_clac:\n"
 115                         ASM_NOP3
 116                         SAVE_REGS_STRING
 117                         "       movq %rsp, %rsi\n"
 118                         ".global optprobe_template_val\n"
 119                         "optprobe_template_val:\n"
 120                         ASM_NOP5
 121                         ASM_NOP5
 122                         ".global optprobe_template_call\n"
 123                         "optprobe_template_call:\n"
 124                         ASM_NOP5
 125                         /* Copy 'regs->flags' into 'regs->ss'. */
 126                         "       movq 18*8(%rsp), %rdx\n"
 127                         "       movq %rdx, 20*8(%rsp)\n"
 128                         RESTORE_REGS_STRING
 129                         /* Skip 'regs->flags' and 'regs->sp'. */
 130                         "       addq $16, %rsp\n"
 131                         /* And pop flags register from 'regs->ss'. */
 132                         "       popfq\n"
 133 #else /* CONFIG_X86_32 */
 134                         "       pushl %ss\n"
 135                         /* Save the 'sp - 4', this will be fixed later. */
 136                         "       pushl %esp\n"
 137                         "       pushfl\n"
 138                         ".global optprobe_template_clac\n"
 139                         "optprobe_template_clac:\n"
 140                         ASM_NOP3
 141                         SAVE_REGS_STRING
 142                         "       movl %esp, %edx\n"
 143                         ".global optprobe_template_val\n"
 144                         "optprobe_template_val:\n"
 145                         ASM_NOP5
 146                         ".global optprobe_template_call\n"
 147                         "optprobe_template_call:\n"
 148                         ASM_NOP5
 149                         /* Copy 'regs->flags' into 'regs->ss'. */
 150                         "       movl 14*4(%esp), %edx\n"
 151                         "       movl %edx, 16*4(%esp)\n"
 152                         RESTORE_REGS_STRING
 153                         /* Skip 'regs->flags' and 'regs->sp'. */
 154                         "       addl $8, %esp\n"
 155                         /* And pop flags register from 'regs->ss'. */
 156                         "       popfl\n"
 157 #endif
 158                         ".global optprobe_template_end\n"
 159                         "optprobe_template_end:\n"
 160                         ".popsection\n");
 161
 162 void optprobe_template_func(void);
 163 STACK_FRAME_NON_STANDARD(optprobe_template_func);
 164
 165 #define TMPL_CLAC_IDX \
 166         ((long)optprobe_template_clac - (long)optprobe_template_entry)
 167 #define TMPL_MOVE_IDX \
 168         ((long)optprobe_template_val - (long)optprobe_template_entry)
 169 #define TMPL_CALL_IDX \
 170         ((long)optprobe_template_call - (long)optprobe_template_entry)
 171 #define TMPL_END_IDX \
 172         ((long)optprobe_template_end - (long)optprobe_template_entry)
 173
 174 /* Optimized kprobe call back function: called from optinsn */
 175 static void
 176 optimized_callback(struct optimized_kprobe *op, struct pt_regs *regs)
 177 {
 178         /* This is possible if op is under delayed unoptimizing */
 179         if (kprobe_disabled(&op->kp))
 180                 return;
 181
 182         preempt_disable();
 183         if (kprobe_running()) {
 184                 kprobes_inc_nmissed_count(&op->kp);
 185         } else {
 186                 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 187                 /* Adjust stack pointer */
 188                 regs->sp += sizeof(long);
 189                 /* Save skipped registers */
 190                 regs->cs = __KERNEL_CS;
 191 #ifdef CONFIG_X86_32
 192                 regs->gs = 0;
 193 #endif
 194                 regs->ip = (unsigned long)op->kp.addr + INT3_INSN_SIZE;
 195                 regs->orig_ax = ~0UL;
 196
 197                 __this_cpu_write(current_kprobe, &op->kp);
 198                 kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 199                 opt_pre_handler(&op->kp, regs);
 200                 __this_cpu_write(current_kprobe, NULL);
 201         }
 202         preempt_enable();
 203 }
 204 NOKPROBE_SYMBOL(optimized_callback);
 205
 206 static int copy_optimized_instructions(u8 *dest, u8 *src, u8 *real)
 207 {
 208         struct insn insn;
 209         int len = 0, ret;
 210
 211         while (len < JMP32_INSN_SIZE) {
 212                 ret = __copy_instruction(dest + len, src + len, real + len, &insn);
 213                 if (!ret || !can_boost(&insn, src + len))
 214                         return -EINVAL;
 215                 len += ret;
 216         }
 217         /* Check whether the address range is reserved */
 218         if (ftrace_text_reserved(src, src + len - 1) ||
 219             alternatives_text_reserved(src, src + len - 1) ||
 220             jump_label_text_reserved(src, src + len - 1) ||
 221             static_call_text_reserved(src, src + len - 1))
 222                 return -EBUSY;
 223
 224         return len;
 225 }
 226
 227 /* Check whether insn is indirect jump */
 228 static int __insn_is_indirect_jump(struct insn *insn)
 229 {
 230         return ((insn->opcode.bytes[0] == 0xff &&
 231                 (X86_MODRM_REG(insn->modrm.value) & 6) == 4) || /* Jump */
 232                 insn->opcode.bytes[0] == 0xea); /* Segment based jump */
 233 }
 234
 235 /* Check whether insn jumps into specified address range */
 236 static int insn_jump_into_range(struct insn *insn, unsigned long start, int len)
 237 {
 238         unsigned long target = 0;
 239
 240         switch (insn->opcode.bytes[0]) {
 241         case 0xe0:      /* loopne */
 242         case 0xe1:      /* loope */
 243         case 0xe2:      /* loop */
 244         case 0xe3:      /* jcxz */
 245         case 0xe9:      /* near relative jump */
 246         case 0xeb:      /* short relative jump */
 247                 break;
 248         case 0x0f:
 249                 if ((insn->opcode.bytes[1] & 0xf0) == 0x80) /* jcc near */
 250                         break;
 251                 return 0;
 252         default:
 253                 if ((insn->opcode.bytes[0] & 0xf0) == 0x70) /* jcc short */
 254                         break;
 255                 return 0;
 256         }
 257         target = (unsigned long)insn->next_byte + insn->immediate.value;
 258
 259         return (start <= target && target <= start + len);
 260 }
 261
 262 static int insn_is_indirect_jump(struct insn *insn)
 263 {
 264         int ret = __insn_is_indirect_jump(insn);
 265
 266 #ifdef CONFIG_RETPOLINE
 267         /*
 268          * Jump to x86_indirect_thunk_* is treated as an indirect jump.
 269          * Note that even with CONFIG_RETPOLINE=y, the kernel compiled with
 270          * older gcc may use indirect jump. So we add this check instead of
 271          * replace indirect-jump check.
 272          */
 273         if (!ret)
 274                 ret = insn_jump_into_range(insn,
 275                                 (unsigned long)__indirect_thunk_start,
 276                                 (unsigned long)__indirect_thunk_end -
 277                                 (unsigned long)__indirect_thunk_start);
 278 #endif
 279         return ret;
 280 }
 281
 282 static bool is_padding_int3(unsigned long addr, unsigned long eaddr)
 283 {
 284         unsigned char ops;
 285
 286         for (; addr < eaddr; addr++) {
 287                 if (get_kernel_nofault(ops, (void *)addr) < 0 ||
 288                     ops != INT3_INSN_OPCODE)
 289                         return false;
 290         }
 291
 292         return true;
 293 }
 294
 295 /* Decode whole function to ensure any instructions don't jump into target */
 296 static int can_optimize(unsigned long paddr)
 297 {
 298         unsigned long addr, size = 0, offset = 0;
 299         struct insn insn;
 300         kprobe_opcode_t buf[MAX_INSN_SIZE];
 301
 302         /* Lookup symbol including addr */
 303         if (!kallsyms_lookup_size_offset(paddr, &size, &offset))
 304                 return 0;
 305
 306         /*
 307          * Do not optimize in the entry code due to the unstable
 308          * stack handling and registers setup.
 309          */
 310         if (((paddr >= (unsigned long)__entry_text_start) &&
 311              (paddr <  (unsigned long)__entry_text_end)))
 312                 return 0;
 313
 314         /* Check there is enough space for a relative jump. */
 315         if (size - offset < JMP32_INSN_SIZE)
 316                 return 0;
 317
 318         /* Decode instructions */
 319         addr = paddr - offset;
 320         while (addr < paddr - offset + size) { /* Decode until function end */
 321                 unsigned long recovered_insn;
 322                 int ret;
 323
 324                 if (search_exception_tables(addr))
 325                         /*
 326                          * Since some fixup code will jumps into this function,
 327                          * we can't optimize kprobe in this function.
 328                          */
 329                         return 0;
 330                 recovered_insn = recover_probed_instruction(buf, addr);
 331                 if (!recovered_insn)
 332                         return 0;
 333
 334                 ret = insn_decode_kernel(&insn, (void *)recovered_insn);
 335                 if (ret < 0)
 336                         return 0;
 337
 338                 /*
 339                  * In the case of detecting unknown breakpoint, this could be
 340                  * a padding INT3 between functions. Let's check that all the
 341                  * rest of the bytes are also INT3.
 342                  */
 343                 if (insn.opcode.bytes[0] == INT3_INSN_OPCODE)
 344                         return is_padding_int3(addr, paddr - offset + size) ? 1 : 0;
 345
 346                 /* Recover address */
 347                 insn.kaddr = (void *)addr;
 348                 insn.next_byte = (void *)(addr + insn.length);
 349                 /* Check any instructions don't jump into target */
 350                 if (insn_is_indirect_jump(&insn) ||
 351                     insn_jump_into_range(&insn, paddr + INT3_INSN_SIZE,
 352                                          DISP32_SIZE))
 353                         return 0;
 354                 addr += insn.length;
 355         }
 356
 357         return 1;
 358 }
 359
 360 /* Check optimized_kprobe can actually be optimized. */
 361 int arch_check_optimized_kprobe(struct optimized_kprobe *op)
 362 {
 363         int i;
 364         struct kprobe *p;
 365
 366         for (i = 1; i < op->optinsn.size; i++) {
 367                 p = get_kprobe(op->kp.addr + i);
 368                 if (p && !kprobe_disabled(p))
 369                         return -EEXIST;
 370         }
 371
 372         return 0;
 373 }
 374
 375 /* Check the addr is within the optimized instructions. */
 376 int arch_within_optimized_kprobe(struct optimized_kprobe *op,
 377                                  kprobe_opcode_t *addr)
 378 {
 379         return (op->kp.addr <= addr &&
 380                 op->kp.addr + op->optinsn.size > addr);
 381 }
 382
 383 /* Free optimized instruction slot */
 384 static
 385 void __arch_remove_optimized_kprobe(struct optimized_kprobe *op, int dirty)
 386 {
 387         u8 *slot = op->optinsn.insn;
 388         if (slot) {
 389                 int len = TMPL_END_IDX + op->optinsn.size + JMP32_INSN_SIZE;
 390
 391                 /* Record the perf event before freeing the slot */
 392                 if (dirty)
 393                         perf_event_text_poke(slot, slot, len, NULL, 0);
 394
 395                 free_optinsn_slot(slot, dirty);
 396                 op->optinsn.insn = NULL;
 397                 op->optinsn.size = 0;
 398         }
 399 }
 400
 401 void arch_remove_optimized_kprobe(struct optimized_kprobe *op)
 402 {
 403         __arch_remove_optimized_kprobe(op, 1);
 404 }
 405
 406 /*
 407  * Copy replacing target instructions
 408  * Target instructions MUST be relocatable (checked inside)
 409  * This is called when new aggr(opt)probe is allocated or reused.
 410  */
 411 int arch_prepare_optimized_kprobe(struct optimized_kprobe *op,
 412                                   struct kprobe *__unused)
 413 {
 414         u8 *buf = NULL, *slot;
 415         int ret, len;
 416         long rel;
 417
 418         if (!can_optimize((unsigned long)op->kp.addr))
 419                 return -EILSEQ;
 420
 421         buf = kzalloc(MAX_OPTINSN_SIZE, GFP_KERNEL);
 422         if (!buf)
 423                 return -ENOMEM;
 424
 425         op->optinsn.insn = slot = get_optinsn_slot();
 426         if (!slot) {
 427                 ret = -ENOMEM;
 428                 goto out;
 429         }
 430
 431         /*
 432          * Verify if the address gap is in 2GB range, because this uses
 433          * a relative jump.
 434          */
 435         rel = (long)slot - (long)op->kp.addr + JMP32_INSN_SIZE;
 436         if (abs(rel) > 0x7fffffff) {
 437                 ret = -ERANGE;
 438                 goto err;
 439         }
 440
 441         /* Copy arch-dep-instance from template */
 442         memcpy(buf, optprobe_template_entry, TMPL_END_IDX);
 443
 444         /* Copy instructions into the out-of-line buffer */
 445         ret = copy_optimized_instructions(buf + TMPL_END_IDX, op->kp.addr,
 446                                           slot + TMPL_END_IDX);
 447         if (ret < 0)
 448                 goto err;
 449         op->optinsn.size = ret;
 450         len = TMPL_END_IDX + op->optinsn.size;
 451
 452         synthesize_clac(buf + TMPL_CLAC_IDX);
 453
 454         /* Set probe information */
 455         synthesize_set_arg1(buf + TMPL_MOVE_IDX, (unsigned long)op);
 456
 457         /* Set probe function call */
 458         synthesize_relcall(buf + TMPL_CALL_IDX,
 459                            slot + TMPL_CALL_IDX, optimized_callback);
 460
 461         /* Set returning jmp instruction at the tail of out-of-line buffer */
 462         synthesize_reljump(buf + len, slot + len,
 463                            (u8 *)op->kp.addr + op->optinsn.size);
 464         len += JMP32_INSN_SIZE;
 465
 466         /*
 467          * Note len = TMPL_END_IDX + op->optinsn.size + JMP32_INSN_SIZE is also
 468          * used in __arch_remove_optimized_kprobe().
 469          */
 470
 471         /* We have to use text_poke() for instruction buffer because it is RO */
 472         perf_event_text_poke(slot, NULL, 0, buf, len);
 473         text_poke(slot, buf, len);
 474
 475         ret = 0;
 476 out:
 477         kfree(buf);
 478         return ret;
 479
 480 err:
 481         __arch_remove_optimized_kprobe(op, 0);
 482         goto out;
 483 }
 484
 485 /*
 486  * Replace breakpoints (INT3) with relative jumps (JMP.d32).
 487  * Caller must call with locking kprobe_mutex and text_mutex.
 488  *
 489  * The caller will have installed a regular kprobe and after that issued
 490  * syncrhonize_rcu_tasks(), this ensures that the instruction(s) that live in
 491  * the 4 bytes after the INT3 are unused and can now be overwritten.
 492  */
 493 void arch_optimize_kprobes(struct list_head *oplist)
 494 {
 495         struct optimized_kprobe *op, *tmp;
 496         u8 insn_buff[JMP32_INSN_SIZE];
 497
 498         list_for_each_entry_safe(op, tmp, oplist, list) {
 499                 s32 rel = (s32)((long)op->optinsn.insn -
 500                         ((long)op->kp.addr + JMP32_INSN_SIZE));
 501
 502                 WARN_ON(kprobe_disabled(&op->kp));
 503
 504                 /* Backup instructions which will be replaced by jump address */
 505                 memcpy(op->optinsn.copied_insn, op->kp.addr + INT3_INSN_SIZE,
 506                        DISP32_SIZE);
 507
 508                 insn_buff[0] = JMP32_INSN_OPCODE;
 509                 *(s32 *)(&insn_buff[1]) = rel;
 510
 511                 text_poke_bp(op->kp.addr, insn_buff, JMP32_INSN_SIZE, NULL);
 512
 513                 list_del_init(&op->list);
 514         }
 515 }
 516
 517 /*
 518  * Replace a relative jump (JMP.d32) with a breakpoint (INT3).
 519  *
 520  * After that, we can restore the 4 bytes after the INT3 to undo what
 521  * arch_optimize_kprobes() scribbled. This is safe since those bytes will be
 522  * unused once the INT3 lands.
 523  */
 524 void arch_unoptimize_kprobe(struct optimized_kprobe *op)
 525 {
 526         u8 new[JMP32_INSN_SIZE] = { INT3_INSN_OPCODE, };
 527         u8 old[JMP32_INSN_SIZE];
 528         u8 *addr = op->kp.addr;
 529
 530         memcpy(old, op->kp.addr, JMP32_INSN_SIZE);
 531         memcpy(new + INT3_INSN_SIZE,
 532                op->optinsn.copied_insn,
 533                JMP32_INSN_SIZE - INT3_INSN_SIZE);
 534
 535         text_poke(addr, new, INT3_INSN_SIZE);
 536         text_poke_sync();
 537         text_poke(addr + INT3_INSN_SIZE,
 538                   new + INT3_INSN_SIZE,
 539                   JMP32_INSN_SIZE - INT3_INSN_SIZE);
 540         text_poke_sync();
 541
 542         perf_event_text_poke(op->kp.addr, old, JMP32_INSN_SIZE, new, JMP32_INSN_SIZE);
 543 }
 544
 545 /*
 546  * Recover original instructions and breakpoints from relative jumps.
 547  * Caller must call with locking kprobe_mutex.
 548  */
 549 extern void arch_unoptimize_kprobes(struct list_head *oplist,
 550                                     struct list_head *done_list)
 551 {
 552         struct optimized_kprobe *op, *tmp;
 553
 554         list_for_each_entry_safe(op, tmp, oplist, list) {
 555                 arch_unoptimize_kprobe(op);
 556                 list_move(&op->list, done_list);
 557         }
 558 }
 559
 560 int setup_detour_execution(struct kprobe *p, struct pt_regs *regs, int reenter)
 561 {
 562         struct optimized_kprobe *op;
 563
 564         if (p->flags & KPROBE_FLAG_OPTIMIZED) {
 565                 /* This kprobe is really able to run optimized path. */
 566                 op = container_of(p, struct optimized_kprobe, kp);
 567                 /* Detour through copied instructions */
 568                 regs->ip = (unsigned long)op->optinsn.insn + TMPL_END_IDX;
 569                 if (!reenter)
 570                         reset_current_kprobe();
 571                 return 1;
 572         }
 573         return 0;
 574 }
 575 NOKPROBE_SYMBOL(setup_detour_execution);