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