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