1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com 2 * 3 * This program is free software; you can redistribute it and/or 4 * modify it under the terms of version 2 of the GNU General Public 5 * License as published by the Free Software Foundation. 6 * 7 * This program is distributed in the hope that it will be useful, but 8 * WITHOUT ANY WARRANTY; without even the implied warranty of 9 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 10 * General Public License for more details. 11 */ 12 #include <linux/kernel.h> 13 #include <linux/types.h> 14 #include <linux/slab.h> 15 #include <linux/bpf.h> 16 #include <linux/filter.h> 17 #include <net/netlink.h> 18 #include <linux/file.h> 19 #include <linux/vmalloc.h> 20 21 /* bpf_check() is a static code analyzer that walks eBPF program 22 * instruction by instruction and updates register/stack state. 23 * All paths of conditional branches are analyzed until 'bpf_exit' insn. 24 * 25 * The first pass is depth-first-search to check that the program is a DAG. 26 * It rejects the following programs: 27 * - larger than BPF_MAXINSNS insns 28 * - if loop is present (detected via back-edge) 29 * - unreachable insns exist (shouldn't be a forest. program = one function) 30 * - out of bounds or malformed jumps 31 * The second pass is all possible path descent from the 1st insn. 32 * Since it's analyzing all pathes through the program, the length of the 33 * analysis is limited to 32k insn, which may be hit even if total number of 34 * insn is less then 4K, but there are too many branches that change stack/regs. 35 * Number of 'branches to be analyzed' is limited to 1k 36 * 37 * On entry to each instruction, each register has a type, and the instruction 38 * changes the types of the registers depending on instruction semantics. 39 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is 40 * copied to R1. 41 * 42 * All registers are 64-bit. 43 * R0 - return register 44 * R1-R5 argument passing registers 45 * R6-R9 callee saved registers 46 * R10 - frame pointer read-only 47 * 48 * At the start of BPF program the register R1 contains a pointer to bpf_context 49 * and has type PTR_TO_CTX. 50 * 51 * Verifier tracks arithmetic operations on pointers in case: 52 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10), 53 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20), 54 * 1st insn copies R10 (which has FRAME_PTR) type into R1 55 * and 2nd arithmetic instruction is pattern matched to recognize 56 * that it wants to construct a pointer to some element within stack. 57 * So after 2nd insn, the register R1 has type PTR_TO_STACK 58 * (and -20 constant is saved for further stack bounds checking). 59 * Meaning that this reg is a pointer to stack plus known immediate constant. 60 * 61 * Most of the time the registers have UNKNOWN_VALUE type, which 62 * means the register has some value, but it's not a valid pointer. 63 * (like pointer plus pointer becomes UNKNOWN_VALUE type) 64 * 65 * When verifier sees load or store instructions the type of base register 66 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, FRAME_PTR. These are three pointer 67 * types recognized by check_mem_access() function. 68 * 69 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value' 70 * and the range of [ptr, ptr + map's value_size) is accessible. 71 * 72 * registers used to pass values to function calls are checked against 73 * function argument constraints. 74 * 75 * ARG_PTR_TO_MAP_KEY is one of such argument constraints. 76 * It means that the register type passed to this function must be 77 * PTR_TO_STACK and it will be used inside the function as 78 * 'pointer to map element key' 79 * 80 * For example the argument constraints for bpf_map_lookup_elem(): 81 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL, 82 * .arg1_type = ARG_CONST_MAP_PTR, 83 * .arg2_type = ARG_PTR_TO_MAP_KEY, 84 * 85 * ret_type says that this function returns 'pointer to map elem value or null' 86 * function expects 1st argument to be a const pointer to 'struct bpf_map' and 87 * 2nd argument should be a pointer to stack, which will be used inside 88 * the helper function as a pointer to map element key. 89 * 90 * On the kernel side the helper function looks like: 91 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 92 * { 93 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1; 94 * void *key = (void *) (unsigned long) r2; 95 * void *value; 96 * 97 * here kernel can access 'key' and 'map' pointers safely, knowing that 98 * [key, key + map->key_size) bytes are valid and were initialized on 99 * the stack of eBPF program. 100 * } 101 * 102 * Corresponding eBPF program may look like: 103 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR 104 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK 105 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP 106 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem), 107 * here verifier looks at prototype of map_lookup_elem() and sees: 108 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok, 109 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes 110 * 111 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far, 112 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits 113 * and were initialized prior to this call. 114 * If it's ok, then verifier allows this BPF_CALL insn and looks at 115 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets 116 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function 117 * returns ether pointer to map value or NULL. 118 * 119 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off' 120 * insn, the register holding that pointer in the true branch changes state to 121 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false 122 * branch. See check_cond_jmp_op(). 123 * 124 * After the call R0 is set to return type of the function and registers R1-R5 125 * are set to NOT_INIT to indicate that they are no longer readable. 126 */ 127 128 /* types of values stored in eBPF registers */ 129 enum bpf_reg_type { 130 NOT_INIT = 0, /* nothing was written into register */ 131 UNKNOWN_VALUE, /* reg doesn't contain a valid pointer */ 132 PTR_TO_CTX, /* reg points to bpf_context */ 133 CONST_PTR_TO_MAP, /* reg points to struct bpf_map */ 134 PTR_TO_MAP_VALUE, /* reg points to map element value */ 135 PTR_TO_MAP_VALUE_OR_NULL,/* points to map elem value or NULL */ 136 FRAME_PTR, /* reg == frame_pointer */ 137 PTR_TO_STACK, /* reg == frame_pointer + imm */ 138 CONST_IMM, /* constant integer value */ 139 }; 140 141 struct reg_state { 142 enum bpf_reg_type type; 143 union { 144 /* valid when type == CONST_IMM | PTR_TO_STACK */ 145 int imm; 146 147 /* valid when type == CONST_PTR_TO_MAP | PTR_TO_MAP_VALUE | 148 * PTR_TO_MAP_VALUE_OR_NULL 149 */ 150 struct bpf_map *map_ptr; 151 }; 152 }; 153 154 enum bpf_stack_slot_type { 155 STACK_INVALID, /* nothing was stored in this stack slot */ 156 STACK_SPILL, /* register spilled into stack */ 157 STACK_MISC /* BPF program wrote some data into this slot */ 158 }; 159 160 #define BPF_REG_SIZE 8 /* size of eBPF register in bytes */ 161 162 /* state of the program: 163 * type of all registers and stack info 164 */ 165 struct verifier_state { 166 struct reg_state regs[MAX_BPF_REG]; 167 u8 stack_slot_type[MAX_BPF_STACK]; 168 struct reg_state spilled_regs[MAX_BPF_STACK / BPF_REG_SIZE]; 169 }; 170 171 /* linked list of verifier states used to prune search */ 172 struct verifier_state_list { 173 struct verifier_state state; 174 struct verifier_state_list *next; 175 }; 176 177 /* verifier_state + insn_idx are pushed to stack when branch is encountered */ 178 struct verifier_stack_elem { 179 /* verifer state is 'st' 180 * before processing instruction 'insn_idx' 181 * and after processing instruction 'prev_insn_idx' 182 */ 183 struct verifier_state st; 184 int insn_idx; 185 int prev_insn_idx; 186 struct verifier_stack_elem *next; 187 }; 188 189 #define MAX_USED_MAPS 64 /* max number of maps accessed by one eBPF program */ 190 191 /* single container for all structs 192 * one verifier_env per bpf_check() call 193 */ 194 struct verifier_env { 195 struct bpf_prog *prog; /* eBPF program being verified */ 196 struct verifier_stack_elem *head; /* stack of verifier states to be processed */ 197 int stack_size; /* number of states to be processed */ 198 struct verifier_state cur_state; /* current verifier state */ 199 struct verifier_state_list **explored_states; /* search pruning optimization */ 200 struct bpf_map *used_maps[MAX_USED_MAPS]; /* array of map's used by eBPF program */ 201 u32 used_map_cnt; /* number of used maps */ 202 bool allow_ptr_leaks; 203 }; 204 205 /* verbose verifier prints what it's seeing 206 * bpf_check() is called under lock, so no race to access these global vars 207 */ 208 static u32 log_level, log_size, log_len; 209 static char *log_buf; 210 211 static DEFINE_MUTEX(bpf_verifier_lock); 212 213 /* log_level controls verbosity level of eBPF verifier. 214 * verbose() is used to dump the verification trace to the log, so the user 215 * can figure out what's wrong with the program 216 */ 217 static __printf(1, 2) void verbose(const char *fmt, ...) 218 { 219 va_list args; 220 221 if (log_level == 0 || log_len >= log_size - 1) 222 return; 223 224 va_start(args, fmt); 225 log_len += vscnprintf(log_buf + log_len, log_size - log_len, fmt, args); 226 va_end(args); 227 } 228 229 /* string representation of 'enum bpf_reg_type' */ 230 static const char * const reg_type_str[] = { 231 [NOT_INIT] = "?", 232 [UNKNOWN_VALUE] = "inv", 233 [PTR_TO_CTX] = "ctx", 234 [CONST_PTR_TO_MAP] = "map_ptr", 235 [PTR_TO_MAP_VALUE] = "map_value", 236 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null", 237 [FRAME_PTR] = "fp", 238 [PTR_TO_STACK] = "fp", 239 [CONST_IMM] = "imm", 240 }; 241 242 static void print_verifier_state(struct verifier_env *env) 243 { 244 enum bpf_reg_type t; 245 int i; 246 247 for (i = 0; i < MAX_BPF_REG; i++) { 248 t = env->cur_state.regs[i].type; 249 if (t == NOT_INIT) 250 continue; 251 verbose(" R%d=%s", i, reg_type_str[t]); 252 if (t == CONST_IMM || t == PTR_TO_STACK) 253 verbose("%d", env->cur_state.regs[i].imm); 254 else if (t == CONST_PTR_TO_MAP || t == PTR_TO_MAP_VALUE || 255 t == PTR_TO_MAP_VALUE_OR_NULL) 256 verbose("(ks=%d,vs=%d)", 257 env->cur_state.regs[i].map_ptr->key_size, 258 env->cur_state.regs[i].map_ptr->value_size); 259 } 260 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) { 261 if (env->cur_state.stack_slot_type[i] == STACK_SPILL) 262 verbose(" fp%d=%s", -MAX_BPF_STACK + i, 263 reg_type_str[env->cur_state.spilled_regs[i / BPF_REG_SIZE].type]); 264 } 265 verbose("\n"); 266 } 267 268 static const char *const bpf_class_string[] = { 269 [BPF_LD] = "ld", 270 [BPF_LDX] = "ldx", 271 [BPF_ST] = "st", 272 [BPF_STX] = "stx", 273 [BPF_ALU] = "alu", 274 [BPF_JMP] = "jmp", 275 [BPF_RET] = "BUG", 276 [BPF_ALU64] = "alu64", 277 }; 278 279 static const char *const bpf_alu_string[16] = { 280 [BPF_ADD >> 4] = "+=", 281 [BPF_SUB >> 4] = "-=", 282 [BPF_MUL >> 4] = "*=", 283 [BPF_DIV >> 4] = "/=", 284 [BPF_OR >> 4] = "|=", 285 [BPF_AND >> 4] = "&=", 286 [BPF_LSH >> 4] = "<<=", 287 [BPF_RSH >> 4] = ">>=", 288 [BPF_NEG >> 4] = "neg", 289 [BPF_MOD >> 4] = "%=", 290 [BPF_XOR >> 4] = "^=", 291 [BPF_MOV >> 4] = "=", 292 [BPF_ARSH >> 4] = "s>>=", 293 [BPF_END >> 4] = "endian", 294 }; 295 296 static const char *const bpf_ldst_string[] = { 297 [BPF_W >> 3] = "u32", 298 [BPF_H >> 3] = "u16", 299 [BPF_B >> 3] = "u8", 300 [BPF_DW >> 3] = "u64", 301 }; 302 303 static const char *const bpf_jmp_string[16] = { 304 [BPF_JA >> 4] = "jmp", 305 [BPF_JEQ >> 4] = "==", 306 [BPF_JGT >> 4] = ">", 307 [BPF_JGE >> 4] = ">=", 308 [BPF_JSET >> 4] = "&", 309 [BPF_JNE >> 4] = "!=", 310 [BPF_JSGT >> 4] = "s>", 311 [BPF_JSGE >> 4] = "s>=", 312 [BPF_CALL >> 4] = "call", 313 [BPF_EXIT >> 4] = "exit", 314 }; 315 316 static void print_bpf_insn(struct bpf_insn *insn) 317 { 318 u8 class = BPF_CLASS(insn->code); 319 320 if (class == BPF_ALU || class == BPF_ALU64) { 321 if (BPF_SRC(insn->code) == BPF_X) 322 verbose("(%02x) %sr%d %s %sr%d\n", 323 insn->code, class == BPF_ALU ? "(u32) " : "", 324 insn->dst_reg, 325 bpf_alu_string[BPF_OP(insn->code) >> 4], 326 class == BPF_ALU ? "(u32) " : "", 327 insn->src_reg); 328 else 329 verbose("(%02x) %sr%d %s %s%d\n", 330 insn->code, class == BPF_ALU ? "(u32) " : "", 331 insn->dst_reg, 332 bpf_alu_string[BPF_OP(insn->code) >> 4], 333 class == BPF_ALU ? "(u32) " : "", 334 insn->imm); 335 } else if (class == BPF_STX) { 336 if (BPF_MODE(insn->code) == BPF_MEM) 337 verbose("(%02x) *(%s *)(r%d %+d) = r%d\n", 338 insn->code, 339 bpf_ldst_string[BPF_SIZE(insn->code) >> 3], 340 insn->dst_reg, 341 insn->off, insn->src_reg); 342 else if (BPF_MODE(insn->code) == BPF_XADD) 343 verbose("(%02x) lock *(%s *)(r%d %+d) += r%d\n", 344 insn->code, 345 bpf_ldst_string[BPF_SIZE(insn->code) >> 3], 346 insn->dst_reg, insn->off, 347 insn->src_reg); 348 else 349 verbose("BUG_%02x\n", insn->code); 350 } else if (class == BPF_ST) { 351 if (BPF_MODE(insn->code) != BPF_MEM) { 352 verbose("BUG_st_%02x\n", insn->code); 353 return; 354 } 355 verbose("(%02x) *(%s *)(r%d %+d) = %d\n", 356 insn->code, 357 bpf_ldst_string[BPF_SIZE(insn->code) >> 3], 358 insn->dst_reg, 359 insn->off, insn->imm); 360 } else if (class == BPF_LDX) { 361 if (BPF_MODE(insn->code) != BPF_MEM) { 362 verbose("BUG_ldx_%02x\n", insn->code); 363 return; 364 } 365 verbose("(%02x) r%d = *(%s *)(r%d %+d)\n", 366 insn->code, insn->dst_reg, 367 bpf_ldst_string[BPF_SIZE(insn->code) >> 3], 368 insn->src_reg, insn->off); 369 } else if (class == BPF_LD) { 370 if (BPF_MODE(insn->code) == BPF_ABS) { 371 verbose("(%02x) r0 = *(%s *)skb[%d]\n", 372 insn->code, 373 bpf_ldst_string[BPF_SIZE(insn->code) >> 3], 374 insn->imm); 375 } else if (BPF_MODE(insn->code) == BPF_IND) { 376 verbose("(%02x) r0 = *(%s *)skb[r%d + %d]\n", 377 insn->code, 378 bpf_ldst_string[BPF_SIZE(insn->code) >> 3], 379 insn->src_reg, insn->imm); 380 } else if (BPF_MODE(insn->code) == BPF_IMM) { 381 verbose("(%02x) r%d = 0x%x\n", 382 insn->code, insn->dst_reg, insn->imm); 383 } else { 384 verbose("BUG_ld_%02x\n", insn->code); 385 return; 386 } 387 } else if (class == BPF_JMP) { 388 u8 opcode = BPF_OP(insn->code); 389 390 if (opcode == BPF_CALL) { 391 verbose("(%02x) call %d\n", insn->code, insn->imm); 392 } else if (insn->code == (BPF_JMP | BPF_JA)) { 393 verbose("(%02x) goto pc%+d\n", 394 insn->code, insn->off); 395 } else if (insn->code == (BPF_JMP | BPF_EXIT)) { 396 verbose("(%02x) exit\n", insn->code); 397 } else if (BPF_SRC(insn->code) == BPF_X) { 398 verbose("(%02x) if r%d %s r%d goto pc%+d\n", 399 insn->code, insn->dst_reg, 400 bpf_jmp_string[BPF_OP(insn->code) >> 4], 401 insn->src_reg, insn->off); 402 } else { 403 verbose("(%02x) if r%d %s 0x%x goto pc%+d\n", 404 insn->code, insn->dst_reg, 405 bpf_jmp_string[BPF_OP(insn->code) >> 4], 406 insn->imm, insn->off); 407 } 408 } else { 409 verbose("(%02x) %s\n", insn->code, bpf_class_string[class]); 410 } 411 } 412 413 static int pop_stack(struct verifier_env *env, int *prev_insn_idx) 414 { 415 struct verifier_stack_elem *elem; 416 int insn_idx; 417 418 if (env->head == NULL) 419 return -1; 420 421 memcpy(&env->cur_state, &env->head->st, sizeof(env->cur_state)); 422 insn_idx = env->head->insn_idx; 423 if (prev_insn_idx) 424 *prev_insn_idx = env->head->prev_insn_idx; 425 elem = env->head->next; 426 kfree(env->head); 427 env->head = elem; 428 env->stack_size--; 429 return insn_idx; 430 } 431 432 static struct verifier_state *push_stack(struct verifier_env *env, int insn_idx, 433 int prev_insn_idx) 434 { 435 struct verifier_stack_elem *elem; 436 437 elem = kmalloc(sizeof(struct verifier_stack_elem), GFP_KERNEL); 438 if (!elem) 439 goto err; 440 441 memcpy(&elem->st, &env->cur_state, sizeof(env->cur_state)); 442 elem->insn_idx = insn_idx; 443 elem->prev_insn_idx = prev_insn_idx; 444 elem->next = env->head; 445 env->head = elem; 446 env->stack_size++; 447 if (env->stack_size > 1024) { 448 verbose("BPF program is too complex\n"); 449 goto err; 450 } 451 return &elem->st; 452 err: 453 /* pop all elements and return */ 454 while (pop_stack(env, NULL) >= 0); 455 return NULL; 456 } 457 458 #define CALLER_SAVED_REGS 6 459 static const int caller_saved[CALLER_SAVED_REGS] = { 460 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5 461 }; 462 463 static void init_reg_state(struct reg_state *regs) 464 { 465 int i; 466 467 for (i = 0; i < MAX_BPF_REG; i++) { 468 regs[i].type = NOT_INIT; 469 regs[i].imm = 0; 470 regs[i].map_ptr = NULL; 471 } 472 473 /* frame pointer */ 474 regs[BPF_REG_FP].type = FRAME_PTR; 475 476 /* 1st arg to a function */ 477 regs[BPF_REG_1].type = PTR_TO_CTX; 478 } 479 480 static void mark_reg_unknown_value(struct reg_state *regs, u32 regno) 481 { 482 BUG_ON(regno >= MAX_BPF_REG); 483 regs[regno].type = UNKNOWN_VALUE; 484 regs[regno].imm = 0; 485 regs[regno].map_ptr = NULL; 486 } 487 488 enum reg_arg_type { 489 SRC_OP, /* register is used as source operand */ 490 DST_OP, /* register is used as destination operand */ 491 DST_OP_NO_MARK /* same as above, check only, don't mark */ 492 }; 493 494 static int check_reg_arg(struct reg_state *regs, u32 regno, 495 enum reg_arg_type t) 496 { 497 if (regno >= MAX_BPF_REG) { 498 verbose("R%d is invalid\n", regno); 499 return -EINVAL; 500 } 501 502 if (t == SRC_OP) { 503 /* check whether register used as source operand can be read */ 504 if (regs[regno].type == NOT_INIT) { 505 verbose("R%d !read_ok\n", regno); 506 return -EACCES; 507 } 508 } else { 509 /* check whether register used as dest operand can be written to */ 510 if (regno == BPF_REG_FP) { 511 verbose("frame pointer is read only\n"); 512 return -EACCES; 513 } 514 if (t == DST_OP) 515 mark_reg_unknown_value(regs, regno); 516 } 517 return 0; 518 } 519 520 static int bpf_size_to_bytes(int bpf_size) 521 { 522 if (bpf_size == BPF_W) 523 return 4; 524 else if (bpf_size == BPF_H) 525 return 2; 526 else if (bpf_size == BPF_B) 527 return 1; 528 else if (bpf_size == BPF_DW) 529 return 8; 530 else 531 return -EINVAL; 532 } 533 534 static bool is_spillable_regtype(enum bpf_reg_type type) 535 { 536 switch (type) { 537 case PTR_TO_MAP_VALUE: 538 case PTR_TO_MAP_VALUE_OR_NULL: 539 case PTR_TO_STACK: 540 case PTR_TO_CTX: 541 case FRAME_PTR: 542 case CONST_PTR_TO_MAP: 543 return true; 544 default: 545 return false; 546 } 547 } 548 549 /* check_stack_read/write functions track spill/fill of registers, 550 * stack boundary and alignment are checked in check_mem_access() 551 */ 552 static int check_stack_write(struct verifier_state *state, int off, int size, 553 int value_regno) 554 { 555 int i; 556 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0, 557 * so it's aligned access and [off, off + size) are within stack limits 558 */ 559 560 if (value_regno >= 0 && 561 is_spillable_regtype(state->regs[value_regno].type)) { 562 563 /* register containing pointer is being spilled into stack */ 564 if (size != BPF_REG_SIZE) { 565 verbose("invalid size of register spill\n"); 566 return -EACCES; 567 } 568 569 /* save register state */ 570 state->spilled_regs[(MAX_BPF_STACK + off) / BPF_REG_SIZE] = 571 state->regs[value_regno]; 572 573 for (i = 0; i < BPF_REG_SIZE; i++) 574 state->stack_slot_type[MAX_BPF_STACK + off + i] = STACK_SPILL; 575 } else { 576 /* regular write of data into stack */ 577 state->spilled_regs[(MAX_BPF_STACK + off) / BPF_REG_SIZE] = 578 (struct reg_state) {}; 579 580 for (i = 0; i < size; i++) 581 state->stack_slot_type[MAX_BPF_STACK + off + i] = STACK_MISC; 582 } 583 return 0; 584 } 585 586 static int check_stack_read(struct verifier_state *state, int off, int size, 587 int value_regno) 588 { 589 u8 *slot_type; 590 int i; 591 592 slot_type = &state->stack_slot_type[MAX_BPF_STACK + off]; 593 594 if (slot_type[0] == STACK_SPILL) { 595 if (size != BPF_REG_SIZE) { 596 verbose("invalid size of register spill\n"); 597 return -EACCES; 598 } 599 for (i = 1; i < BPF_REG_SIZE; i++) { 600 if (slot_type[i] != STACK_SPILL) { 601 verbose("corrupted spill memory\n"); 602 return -EACCES; 603 } 604 } 605 606 if (value_regno >= 0) 607 /* restore register state from stack */ 608 state->regs[value_regno] = 609 state->spilled_regs[(MAX_BPF_STACK + off) / BPF_REG_SIZE]; 610 return 0; 611 } else { 612 for (i = 0; i < size; i++) { 613 if (slot_type[i] != STACK_MISC) { 614 verbose("invalid read from stack off %d+%d size %d\n", 615 off, i, size); 616 return -EACCES; 617 } 618 } 619 if (value_regno >= 0) 620 /* have read misc data from the stack */ 621 mark_reg_unknown_value(state->regs, value_regno); 622 return 0; 623 } 624 } 625 626 /* check read/write into map element returned by bpf_map_lookup_elem() */ 627 static int check_map_access(struct verifier_env *env, u32 regno, int off, 628 int size) 629 { 630 struct bpf_map *map = env->cur_state.regs[regno].map_ptr; 631 632 if (off < 0 || off + size > map->value_size) { 633 verbose("invalid access to map value, value_size=%d off=%d size=%d\n", 634 map->value_size, off, size); 635 return -EACCES; 636 } 637 return 0; 638 } 639 640 /* check access to 'struct bpf_context' fields */ 641 static int check_ctx_access(struct verifier_env *env, int off, int size, 642 enum bpf_access_type t) 643 { 644 if (env->prog->aux->ops->is_valid_access && 645 env->prog->aux->ops->is_valid_access(off, size, t)) 646 return 0; 647 648 verbose("invalid bpf_context access off=%d size=%d\n", off, size); 649 return -EACCES; 650 } 651 652 static bool is_pointer_value(struct verifier_env *env, int regno) 653 { 654 if (env->allow_ptr_leaks) 655 return false; 656 657 switch (env->cur_state.regs[regno].type) { 658 case UNKNOWN_VALUE: 659 case CONST_IMM: 660 return false; 661 default: 662 return true; 663 } 664 } 665 666 /* check whether memory at (regno + off) is accessible for t = (read | write) 667 * if t==write, value_regno is a register which value is stored into memory 668 * if t==read, value_regno is a register which will receive the value from memory 669 * if t==write && value_regno==-1, some unknown value is stored into memory 670 * if t==read && value_regno==-1, don't care what we read from memory 671 */ 672 static int check_mem_access(struct verifier_env *env, u32 regno, int off, 673 int bpf_size, enum bpf_access_type t, 674 int value_regno) 675 { 676 struct verifier_state *state = &env->cur_state; 677 int size, err = 0; 678 679 if (state->regs[regno].type == PTR_TO_STACK) 680 off += state->regs[regno].imm; 681 682 size = bpf_size_to_bytes(bpf_size); 683 if (size < 0) 684 return size; 685 686 if (off % size != 0) { 687 verbose("misaligned access off %d size %d\n", off, size); 688 return -EACCES; 689 } 690 691 if (state->regs[regno].type == PTR_TO_MAP_VALUE) { 692 if (t == BPF_WRITE && value_regno >= 0 && 693 is_pointer_value(env, value_regno)) { 694 verbose("R%d leaks addr into map\n", value_regno); 695 return -EACCES; 696 } 697 err = check_map_access(env, regno, off, size); 698 if (!err && t == BPF_READ && value_regno >= 0) 699 mark_reg_unknown_value(state->regs, value_regno); 700 701 } else if (state->regs[regno].type == PTR_TO_CTX) { 702 if (t == BPF_WRITE && value_regno >= 0 && 703 is_pointer_value(env, value_regno)) { 704 verbose("R%d leaks addr into ctx\n", value_regno); 705 return -EACCES; 706 } 707 err = check_ctx_access(env, off, size, t); 708 if (!err && t == BPF_READ && value_regno >= 0) 709 mark_reg_unknown_value(state->regs, value_regno); 710 711 } else if (state->regs[regno].type == FRAME_PTR || 712 state->regs[regno].type == PTR_TO_STACK) { 713 if (off >= 0 || off < -MAX_BPF_STACK) { 714 verbose("invalid stack off=%d size=%d\n", off, size); 715 return -EACCES; 716 } 717 if (t == BPF_WRITE) { 718 if (!env->allow_ptr_leaks && 719 state->stack_slot_type[MAX_BPF_STACK + off] == STACK_SPILL && 720 size != BPF_REG_SIZE) { 721 verbose("attempt to corrupt spilled pointer on stack\n"); 722 return -EACCES; 723 } 724 err = check_stack_write(state, off, size, value_regno); 725 } else { 726 err = check_stack_read(state, off, size, value_regno); 727 } 728 } else { 729 verbose("R%d invalid mem access '%s'\n", 730 regno, reg_type_str[state->regs[regno].type]); 731 return -EACCES; 732 } 733 return err; 734 } 735 736 static int check_xadd(struct verifier_env *env, struct bpf_insn *insn) 737 { 738 struct reg_state *regs = env->cur_state.regs; 739 int err; 740 741 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) || 742 insn->imm != 0) { 743 verbose("BPF_XADD uses reserved fields\n"); 744 return -EINVAL; 745 } 746 747 /* check src1 operand */ 748 err = check_reg_arg(regs, insn->src_reg, SRC_OP); 749 if (err) 750 return err; 751 752 /* check src2 operand */ 753 err = check_reg_arg(regs, insn->dst_reg, SRC_OP); 754 if (err) 755 return err; 756 757 /* check whether atomic_add can read the memory */ 758 err = check_mem_access(env, insn->dst_reg, insn->off, 759 BPF_SIZE(insn->code), BPF_READ, -1); 760 if (err) 761 return err; 762 763 /* check whether atomic_add can write into the same memory */ 764 return check_mem_access(env, insn->dst_reg, insn->off, 765 BPF_SIZE(insn->code), BPF_WRITE, -1); 766 } 767 768 /* when register 'regno' is passed into function that will read 'access_size' 769 * bytes from that pointer, make sure that it's within stack boundary 770 * and all elements of stack are initialized 771 */ 772 static int check_stack_boundary(struct verifier_env *env, int regno, 773 int access_size, bool zero_size_allowed) 774 { 775 struct verifier_state *state = &env->cur_state; 776 struct reg_state *regs = state->regs; 777 int off, i; 778 779 if (regs[regno].type != PTR_TO_STACK) { 780 if (zero_size_allowed && access_size == 0 && 781 regs[regno].type == CONST_IMM && 782 regs[regno].imm == 0) 783 return 0; 784 785 verbose("R%d type=%s expected=%s\n", regno, 786 reg_type_str[regs[regno].type], 787 reg_type_str[PTR_TO_STACK]); 788 return -EACCES; 789 } 790 791 off = regs[regno].imm; 792 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 || 793 access_size <= 0) { 794 verbose("invalid stack type R%d off=%d access_size=%d\n", 795 regno, off, access_size); 796 return -EACCES; 797 } 798 799 for (i = 0; i < access_size; i++) { 800 if (state->stack_slot_type[MAX_BPF_STACK + off + i] != STACK_MISC) { 801 verbose("invalid indirect read from stack off %d+%d size %d\n", 802 off, i, access_size); 803 return -EACCES; 804 } 805 } 806 return 0; 807 } 808 809 static int check_func_arg(struct verifier_env *env, u32 regno, 810 enum bpf_arg_type arg_type, struct bpf_map **mapp) 811 { 812 struct reg_state *reg = env->cur_state.regs + regno; 813 enum bpf_reg_type expected_type; 814 int err = 0; 815 816 if (arg_type == ARG_DONTCARE) 817 return 0; 818 819 if (reg->type == NOT_INIT) { 820 verbose("R%d !read_ok\n", regno); 821 return -EACCES; 822 } 823 824 if (arg_type == ARG_ANYTHING) { 825 if (is_pointer_value(env, regno)) { 826 verbose("R%d leaks addr into helper function\n", regno); 827 return -EACCES; 828 } 829 return 0; 830 } 831 832 if (arg_type == ARG_PTR_TO_MAP_KEY || 833 arg_type == ARG_PTR_TO_MAP_VALUE) { 834 expected_type = PTR_TO_STACK; 835 } else if (arg_type == ARG_CONST_STACK_SIZE || 836 arg_type == ARG_CONST_STACK_SIZE_OR_ZERO) { 837 expected_type = CONST_IMM; 838 } else if (arg_type == ARG_CONST_MAP_PTR) { 839 expected_type = CONST_PTR_TO_MAP; 840 } else if (arg_type == ARG_PTR_TO_CTX) { 841 expected_type = PTR_TO_CTX; 842 } else if (arg_type == ARG_PTR_TO_STACK) { 843 expected_type = PTR_TO_STACK; 844 /* One exception here. In case function allows for NULL to be 845 * passed in as argument, it's a CONST_IMM type. Final test 846 * happens during stack boundary checking. 847 */ 848 if (reg->type == CONST_IMM && reg->imm == 0) 849 expected_type = CONST_IMM; 850 } else { 851 verbose("unsupported arg_type %d\n", arg_type); 852 return -EFAULT; 853 } 854 855 if (reg->type != expected_type) { 856 verbose("R%d type=%s expected=%s\n", regno, 857 reg_type_str[reg->type], reg_type_str[expected_type]); 858 return -EACCES; 859 } 860 861 if (arg_type == ARG_CONST_MAP_PTR) { 862 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */ 863 *mapp = reg->map_ptr; 864 865 } else if (arg_type == ARG_PTR_TO_MAP_KEY) { 866 /* bpf_map_xxx(..., map_ptr, ..., key) call: 867 * check that [key, key + map->key_size) are within 868 * stack limits and initialized 869 */ 870 if (!*mapp) { 871 /* in function declaration map_ptr must come before 872 * map_key, so that it's verified and known before 873 * we have to check map_key here. Otherwise it means 874 * that kernel subsystem misconfigured verifier 875 */ 876 verbose("invalid map_ptr to access map->key\n"); 877 return -EACCES; 878 } 879 err = check_stack_boundary(env, regno, (*mapp)->key_size, 880 false); 881 } else if (arg_type == ARG_PTR_TO_MAP_VALUE) { 882 /* bpf_map_xxx(..., map_ptr, ..., value) call: 883 * check [value, value + map->value_size) validity 884 */ 885 if (!*mapp) { 886 /* kernel subsystem misconfigured verifier */ 887 verbose("invalid map_ptr to access map->value\n"); 888 return -EACCES; 889 } 890 err = check_stack_boundary(env, regno, (*mapp)->value_size, 891 false); 892 } else if (arg_type == ARG_CONST_STACK_SIZE || 893 arg_type == ARG_CONST_STACK_SIZE_OR_ZERO) { 894 bool zero_size_allowed = (arg_type == ARG_CONST_STACK_SIZE_OR_ZERO); 895 896 /* bpf_xxx(..., buf, len) call will access 'len' bytes 897 * from stack pointer 'buf'. Check it 898 * note: regno == len, regno - 1 == buf 899 */ 900 if (regno == 0) { 901 /* kernel subsystem misconfigured verifier */ 902 verbose("ARG_CONST_STACK_SIZE cannot be first argument\n"); 903 return -EACCES; 904 } 905 err = check_stack_boundary(env, regno - 1, reg->imm, 906 zero_size_allowed); 907 } 908 909 return err; 910 } 911 912 static int check_map_func_compatibility(struct bpf_map *map, int func_id) 913 { 914 if (!map) 915 return 0; 916 917 /* We need a two way check, first is from map perspective ... */ 918 switch (map->map_type) { 919 case BPF_MAP_TYPE_PROG_ARRAY: 920 if (func_id != BPF_FUNC_tail_call) 921 goto error; 922 break; 923 case BPF_MAP_TYPE_PERF_EVENT_ARRAY: 924 if (func_id != BPF_FUNC_perf_event_read && 925 func_id != BPF_FUNC_perf_event_output) 926 goto error; 927 break; 928 case BPF_MAP_TYPE_STACK_TRACE: 929 if (func_id != BPF_FUNC_get_stackid) 930 goto error; 931 break; 932 default: 933 break; 934 } 935 936 /* ... and second from the function itself. */ 937 switch (func_id) { 938 case BPF_FUNC_tail_call: 939 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY) 940 goto error; 941 break; 942 case BPF_FUNC_perf_event_read: 943 case BPF_FUNC_perf_event_output: 944 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY) 945 goto error; 946 break; 947 case BPF_FUNC_get_stackid: 948 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE) 949 goto error; 950 break; 951 default: 952 break; 953 } 954 955 return 0; 956 error: 957 verbose("cannot pass map_type %d into func %d\n", 958 map->map_type, func_id); 959 return -EINVAL; 960 } 961 962 static int check_call(struct verifier_env *env, int func_id) 963 { 964 struct verifier_state *state = &env->cur_state; 965 const struct bpf_func_proto *fn = NULL; 966 struct reg_state *regs = state->regs; 967 struct bpf_map *map = NULL; 968 struct reg_state *reg; 969 int i, err; 970 971 /* find function prototype */ 972 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) { 973 verbose("invalid func %d\n", func_id); 974 return -EINVAL; 975 } 976 977 if (env->prog->aux->ops->get_func_proto) 978 fn = env->prog->aux->ops->get_func_proto(func_id); 979 980 if (!fn) { 981 verbose("unknown func %d\n", func_id); 982 return -EINVAL; 983 } 984 985 /* eBPF programs must be GPL compatible to use GPL-ed functions */ 986 if (!env->prog->gpl_compatible && fn->gpl_only) { 987 verbose("cannot call GPL only function from proprietary program\n"); 988 return -EINVAL; 989 } 990 991 /* check args */ 992 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &map); 993 if (err) 994 return err; 995 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &map); 996 if (err) 997 return err; 998 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &map); 999 if (err) 1000 return err; 1001 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &map); 1002 if (err) 1003 return err; 1004 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &map); 1005 if (err) 1006 return err; 1007 1008 /* reset caller saved regs */ 1009 for (i = 0; i < CALLER_SAVED_REGS; i++) { 1010 reg = regs + caller_saved[i]; 1011 reg->type = NOT_INIT; 1012 reg->imm = 0; 1013 } 1014 1015 /* update return register */ 1016 if (fn->ret_type == RET_INTEGER) { 1017 regs[BPF_REG_0].type = UNKNOWN_VALUE; 1018 } else if (fn->ret_type == RET_VOID) { 1019 regs[BPF_REG_0].type = NOT_INIT; 1020 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL) { 1021 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL; 1022 /* remember map_ptr, so that check_map_access() 1023 * can check 'value_size' boundary of memory access 1024 * to map element returned from bpf_map_lookup_elem() 1025 */ 1026 if (map == NULL) { 1027 verbose("kernel subsystem misconfigured verifier\n"); 1028 return -EINVAL; 1029 } 1030 regs[BPF_REG_0].map_ptr = map; 1031 } else { 1032 verbose("unknown return type %d of func %d\n", 1033 fn->ret_type, func_id); 1034 return -EINVAL; 1035 } 1036 1037 err = check_map_func_compatibility(map, func_id); 1038 if (err) 1039 return err; 1040 1041 return 0; 1042 } 1043 1044 /* check validity of 32-bit and 64-bit arithmetic operations */ 1045 static int check_alu_op(struct verifier_env *env, struct bpf_insn *insn) 1046 { 1047 struct reg_state *regs = env->cur_state.regs; 1048 u8 opcode = BPF_OP(insn->code); 1049 int err; 1050 1051 if (opcode == BPF_END || opcode == BPF_NEG) { 1052 if (opcode == BPF_NEG) { 1053 if (BPF_SRC(insn->code) != 0 || 1054 insn->src_reg != BPF_REG_0 || 1055 insn->off != 0 || insn->imm != 0) { 1056 verbose("BPF_NEG uses reserved fields\n"); 1057 return -EINVAL; 1058 } 1059 } else { 1060 if (insn->src_reg != BPF_REG_0 || insn->off != 0 || 1061 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64)) { 1062 verbose("BPF_END uses reserved fields\n"); 1063 return -EINVAL; 1064 } 1065 } 1066 1067 /* check src operand */ 1068 err = check_reg_arg(regs, insn->dst_reg, SRC_OP); 1069 if (err) 1070 return err; 1071 1072 if (is_pointer_value(env, insn->dst_reg)) { 1073 verbose("R%d pointer arithmetic prohibited\n", 1074 insn->dst_reg); 1075 return -EACCES; 1076 } 1077 1078 /* check dest operand */ 1079 err = check_reg_arg(regs, insn->dst_reg, DST_OP); 1080 if (err) 1081 return err; 1082 1083 } else if (opcode == BPF_MOV) { 1084 1085 if (BPF_SRC(insn->code) == BPF_X) { 1086 if (insn->imm != 0 || insn->off != 0) { 1087 verbose("BPF_MOV uses reserved fields\n"); 1088 return -EINVAL; 1089 } 1090 1091 /* check src operand */ 1092 err = check_reg_arg(regs, insn->src_reg, SRC_OP); 1093 if (err) 1094 return err; 1095 } else { 1096 if (insn->src_reg != BPF_REG_0 || insn->off != 0) { 1097 verbose("BPF_MOV uses reserved fields\n"); 1098 return -EINVAL; 1099 } 1100 } 1101 1102 /* check dest operand */ 1103 err = check_reg_arg(regs, insn->dst_reg, DST_OP); 1104 if (err) 1105 return err; 1106 1107 if (BPF_SRC(insn->code) == BPF_X) { 1108 if (BPF_CLASS(insn->code) == BPF_ALU64) { 1109 /* case: R1 = R2 1110 * copy register state to dest reg 1111 */ 1112 regs[insn->dst_reg] = regs[insn->src_reg]; 1113 } else { 1114 if (is_pointer_value(env, insn->src_reg)) { 1115 verbose("R%d partial copy of pointer\n", 1116 insn->src_reg); 1117 return -EACCES; 1118 } 1119 regs[insn->dst_reg].type = UNKNOWN_VALUE; 1120 regs[insn->dst_reg].map_ptr = NULL; 1121 } 1122 } else { 1123 /* case: R = imm 1124 * remember the value we stored into this reg 1125 */ 1126 regs[insn->dst_reg].type = CONST_IMM; 1127 regs[insn->dst_reg].imm = insn->imm; 1128 } 1129 1130 } else if (opcode > BPF_END) { 1131 verbose("invalid BPF_ALU opcode %x\n", opcode); 1132 return -EINVAL; 1133 1134 } else { /* all other ALU ops: and, sub, xor, add, ... */ 1135 1136 bool stack_relative = false; 1137 1138 if (BPF_SRC(insn->code) == BPF_X) { 1139 if (insn->imm != 0 || insn->off != 0) { 1140 verbose("BPF_ALU uses reserved fields\n"); 1141 return -EINVAL; 1142 } 1143 /* check src1 operand */ 1144 err = check_reg_arg(regs, insn->src_reg, SRC_OP); 1145 if (err) 1146 return err; 1147 } else { 1148 if (insn->src_reg != BPF_REG_0 || insn->off != 0) { 1149 verbose("BPF_ALU uses reserved fields\n"); 1150 return -EINVAL; 1151 } 1152 } 1153 1154 /* check src2 operand */ 1155 err = check_reg_arg(regs, insn->dst_reg, SRC_OP); 1156 if (err) 1157 return err; 1158 1159 if ((opcode == BPF_MOD || opcode == BPF_DIV) && 1160 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) { 1161 verbose("div by zero\n"); 1162 return -EINVAL; 1163 } 1164 1165 if ((opcode == BPF_LSH || opcode == BPF_RSH || 1166 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) { 1167 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32; 1168 1169 if (insn->imm < 0 || insn->imm >= size) { 1170 verbose("invalid shift %d\n", insn->imm); 1171 return -EINVAL; 1172 } 1173 } 1174 1175 /* pattern match 'bpf_add Rx, imm' instruction */ 1176 if (opcode == BPF_ADD && BPF_CLASS(insn->code) == BPF_ALU64 && 1177 regs[insn->dst_reg].type == FRAME_PTR && 1178 BPF_SRC(insn->code) == BPF_K) { 1179 stack_relative = true; 1180 } else if (is_pointer_value(env, insn->dst_reg)) { 1181 verbose("R%d pointer arithmetic prohibited\n", 1182 insn->dst_reg); 1183 return -EACCES; 1184 } else if (BPF_SRC(insn->code) == BPF_X && 1185 is_pointer_value(env, insn->src_reg)) { 1186 verbose("R%d pointer arithmetic prohibited\n", 1187 insn->src_reg); 1188 return -EACCES; 1189 } 1190 1191 /* check dest operand */ 1192 err = check_reg_arg(regs, insn->dst_reg, DST_OP); 1193 if (err) 1194 return err; 1195 1196 if (stack_relative) { 1197 regs[insn->dst_reg].type = PTR_TO_STACK; 1198 regs[insn->dst_reg].imm = insn->imm; 1199 } 1200 } 1201 1202 return 0; 1203 } 1204 1205 static int check_cond_jmp_op(struct verifier_env *env, 1206 struct bpf_insn *insn, int *insn_idx) 1207 { 1208 struct reg_state *regs = env->cur_state.regs; 1209 struct verifier_state *other_branch; 1210 u8 opcode = BPF_OP(insn->code); 1211 int err; 1212 1213 if (opcode > BPF_EXIT) { 1214 verbose("invalid BPF_JMP opcode %x\n", opcode); 1215 return -EINVAL; 1216 } 1217 1218 if (BPF_SRC(insn->code) == BPF_X) { 1219 if (insn->imm != 0) { 1220 verbose("BPF_JMP uses reserved fields\n"); 1221 return -EINVAL; 1222 } 1223 1224 /* check src1 operand */ 1225 err = check_reg_arg(regs, insn->src_reg, SRC_OP); 1226 if (err) 1227 return err; 1228 1229 if (is_pointer_value(env, insn->src_reg)) { 1230 verbose("R%d pointer comparison prohibited\n", 1231 insn->src_reg); 1232 return -EACCES; 1233 } 1234 } else { 1235 if (insn->src_reg != BPF_REG_0) { 1236 verbose("BPF_JMP uses reserved fields\n"); 1237 return -EINVAL; 1238 } 1239 } 1240 1241 /* check src2 operand */ 1242 err = check_reg_arg(regs, insn->dst_reg, SRC_OP); 1243 if (err) 1244 return err; 1245 1246 /* detect if R == 0 where R was initialized to zero earlier */ 1247 if (BPF_SRC(insn->code) == BPF_K && 1248 (opcode == BPF_JEQ || opcode == BPF_JNE) && 1249 regs[insn->dst_reg].type == CONST_IMM && 1250 regs[insn->dst_reg].imm == insn->imm) { 1251 if (opcode == BPF_JEQ) { 1252 /* if (imm == imm) goto pc+off; 1253 * only follow the goto, ignore fall-through 1254 */ 1255 *insn_idx += insn->off; 1256 return 0; 1257 } else { 1258 /* if (imm != imm) goto pc+off; 1259 * only follow fall-through branch, since 1260 * that's where the program will go 1261 */ 1262 return 0; 1263 } 1264 } 1265 1266 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx); 1267 if (!other_branch) 1268 return -EFAULT; 1269 1270 /* detect if R == 0 where R is returned value from bpf_map_lookup_elem() */ 1271 if (BPF_SRC(insn->code) == BPF_K && 1272 insn->imm == 0 && (opcode == BPF_JEQ || 1273 opcode == BPF_JNE) && 1274 regs[insn->dst_reg].type == PTR_TO_MAP_VALUE_OR_NULL) { 1275 if (opcode == BPF_JEQ) { 1276 /* next fallthrough insn can access memory via 1277 * this register 1278 */ 1279 regs[insn->dst_reg].type = PTR_TO_MAP_VALUE; 1280 /* branch targer cannot access it, since reg == 0 */ 1281 other_branch->regs[insn->dst_reg].type = CONST_IMM; 1282 other_branch->regs[insn->dst_reg].imm = 0; 1283 } else { 1284 other_branch->regs[insn->dst_reg].type = PTR_TO_MAP_VALUE; 1285 regs[insn->dst_reg].type = CONST_IMM; 1286 regs[insn->dst_reg].imm = 0; 1287 } 1288 } else if (is_pointer_value(env, insn->dst_reg)) { 1289 verbose("R%d pointer comparison prohibited\n", insn->dst_reg); 1290 return -EACCES; 1291 } else if (BPF_SRC(insn->code) == BPF_K && 1292 (opcode == BPF_JEQ || opcode == BPF_JNE)) { 1293 1294 if (opcode == BPF_JEQ) { 1295 /* detect if (R == imm) goto 1296 * and in the target state recognize that R = imm 1297 */ 1298 other_branch->regs[insn->dst_reg].type = CONST_IMM; 1299 other_branch->regs[insn->dst_reg].imm = insn->imm; 1300 } else { 1301 /* detect if (R != imm) goto 1302 * and in the fall-through state recognize that R = imm 1303 */ 1304 regs[insn->dst_reg].type = CONST_IMM; 1305 regs[insn->dst_reg].imm = insn->imm; 1306 } 1307 } 1308 if (log_level) 1309 print_verifier_state(env); 1310 return 0; 1311 } 1312 1313 /* return the map pointer stored inside BPF_LD_IMM64 instruction */ 1314 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn) 1315 { 1316 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32; 1317 1318 return (struct bpf_map *) (unsigned long) imm64; 1319 } 1320 1321 /* verify BPF_LD_IMM64 instruction */ 1322 static int check_ld_imm(struct verifier_env *env, struct bpf_insn *insn) 1323 { 1324 struct reg_state *regs = env->cur_state.regs; 1325 int err; 1326 1327 if (BPF_SIZE(insn->code) != BPF_DW) { 1328 verbose("invalid BPF_LD_IMM insn\n"); 1329 return -EINVAL; 1330 } 1331 if (insn->off != 0) { 1332 verbose("BPF_LD_IMM64 uses reserved fields\n"); 1333 return -EINVAL; 1334 } 1335 1336 err = check_reg_arg(regs, insn->dst_reg, DST_OP); 1337 if (err) 1338 return err; 1339 1340 if (insn->src_reg == 0) 1341 /* generic move 64-bit immediate into a register */ 1342 return 0; 1343 1344 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */ 1345 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD); 1346 1347 regs[insn->dst_reg].type = CONST_PTR_TO_MAP; 1348 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn); 1349 return 0; 1350 } 1351 1352 static bool may_access_skb(enum bpf_prog_type type) 1353 { 1354 switch (type) { 1355 case BPF_PROG_TYPE_SOCKET_FILTER: 1356 case BPF_PROG_TYPE_SCHED_CLS: 1357 case BPF_PROG_TYPE_SCHED_ACT: 1358 return true; 1359 default: 1360 return false; 1361 } 1362 } 1363 1364 /* verify safety of LD_ABS|LD_IND instructions: 1365 * - they can only appear in the programs where ctx == skb 1366 * - since they are wrappers of function calls, they scratch R1-R5 registers, 1367 * preserve R6-R9, and store return value into R0 1368 * 1369 * Implicit input: 1370 * ctx == skb == R6 == CTX 1371 * 1372 * Explicit input: 1373 * SRC == any register 1374 * IMM == 32-bit immediate 1375 * 1376 * Output: 1377 * R0 - 8/16/32-bit skb data converted to cpu endianness 1378 */ 1379 static int check_ld_abs(struct verifier_env *env, struct bpf_insn *insn) 1380 { 1381 struct reg_state *regs = env->cur_state.regs; 1382 u8 mode = BPF_MODE(insn->code); 1383 struct reg_state *reg; 1384 int i, err; 1385 1386 if (!may_access_skb(env->prog->type)) { 1387 verbose("BPF_LD_ABS|IND instructions not allowed for this program type\n"); 1388 return -EINVAL; 1389 } 1390 1391 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 || 1392 BPF_SIZE(insn->code) == BPF_DW || 1393 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) { 1394 verbose("BPF_LD_ABS uses reserved fields\n"); 1395 return -EINVAL; 1396 } 1397 1398 /* check whether implicit source operand (register R6) is readable */ 1399 err = check_reg_arg(regs, BPF_REG_6, SRC_OP); 1400 if (err) 1401 return err; 1402 1403 if (regs[BPF_REG_6].type != PTR_TO_CTX) { 1404 verbose("at the time of BPF_LD_ABS|IND R6 != pointer to skb\n"); 1405 return -EINVAL; 1406 } 1407 1408 if (mode == BPF_IND) { 1409 /* check explicit source operand */ 1410 err = check_reg_arg(regs, insn->src_reg, SRC_OP); 1411 if (err) 1412 return err; 1413 } 1414 1415 /* reset caller saved regs to unreadable */ 1416 for (i = 0; i < CALLER_SAVED_REGS; i++) { 1417 reg = regs + caller_saved[i]; 1418 reg->type = NOT_INIT; 1419 reg->imm = 0; 1420 } 1421 1422 /* mark destination R0 register as readable, since it contains 1423 * the value fetched from the packet 1424 */ 1425 regs[BPF_REG_0].type = UNKNOWN_VALUE; 1426 return 0; 1427 } 1428 1429 /* non-recursive DFS pseudo code 1430 * 1 procedure DFS-iterative(G,v): 1431 * 2 label v as discovered 1432 * 3 let S be a stack 1433 * 4 S.push(v) 1434 * 5 while S is not empty 1435 * 6 t <- S.pop() 1436 * 7 if t is what we're looking for: 1437 * 8 return t 1438 * 9 for all edges e in G.adjacentEdges(t) do 1439 * 10 if edge e is already labelled 1440 * 11 continue with the next edge 1441 * 12 w <- G.adjacentVertex(t,e) 1442 * 13 if vertex w is not discovered and not explored 1443 * 14 label e as tree-edge 1444 * 15 label w as discovered 1445 * 16 S.push(w) 1446 * 17 continue at 5 1447 * 18 else if vertex w is discovered 1448 * 19 label e as back-edge 1449 * 20 else 1450 * 21 // vertex w is explored 1451 * 22 label e as forward- or cross-edge 1452 * 23 label t as explored 1453 * 24 S.pop() 1454 * 1455 * convention: 1456 * 0x10 - discovered 1457 * 0x11 - discovered and fall-through edge labelled 1458 * 0x12 - discovered and fall-through and branch edges labelled 1459 * 0x20 - explored 1460 */ 1461 1462 enum { 1463 DISCOVERED = 0x10, 1464 EXPLORED = 0x20, 1465 FALLTHROUGH = 1, 1466 BRANCH = 2, 1467 }; 1468 1469 #define STATE_LIST_MARK ((struct verifier_state_list *) -1L) 1470 1471 static int *insn_stack; /* stack of insns to process */ 1472 static int cur_stack; /* current stack index */ 1473 static int *insn_state; 1474 1475 /* t, w, e - match pseudo-code above: 1476 * t - index of current instruction 1477 * w - next instruction 1478 * e - edge 1479 */ 1480 static int push_insn(int t, int w, int e, struct verifier_env *env) 1481 { 1482 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH)) 1483 return 0; 1484 1485 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH)) 1486 return 0; 1487 1488 if (w < 0 || w >= env->prog->len) { 1489 verbose("jump out of range from insn %d to %d\n", t, w); 1490 return -EINVAL; 1491 } 1492 1493 if (e == BRANCH) 1494 /* mark branch target for state pruning */ 1495 env->explored_states[w] = STATE_LIST_MARK; 1496 1497 if (insn_state[w] == 0) { 1498 /* tree-edge */ 1499 insn_state[t] = DISCOVERED | e; 1500 insn_state[w] = DISCOVERED; 1501 if (cur_stack >= env->prog->len) 1502 return -E2BIG; 1503 insn_stack[cur_stack++] = w; 1504 return 1; 1505 } else if ((insn_state[w] & 0xF0) == DISCOVERED) { 1506 verbose("back-edge from insn %d to %d\n", t, w); 1507 return -EINVAL; 1508 } else if (insn_state[w] == EXPLORED) { 1509 /* forward- or cross-edge */ 1510 insn_state[t] = DISCOVERED | e; 1511 } else { 1512 verbose("insn state internal bug\n"); 1513 return -EFAULT; 1514 } 1515 return 0; 1516 } 1517 1518 /* non-recursive depth-first-search to detect loops in BPF program 1519 * loop == back-edge in directed graph 1520 */ 1521 static int check_cfg(struct verifier_env *env) 1522 { 1523 struct bpf_insn *insns = env->prog->insnsi; 1524 int insn_cnt = env->prog->len; 1525 int ret = 0; 1526 int i, t; 1527 1528 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL); 1529 if (!insn_state) 1530 return -ENOMEM; 1531 1532 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL); 1533 if (!insn_stack) { 1534 kfree(insn_state); 1535 return -ENOMEM; 1536 } 1537 1538 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */ 1539 insn_stack[0] = 0; /* 0 is the first instruction */ 1540 cur_stack = 1; 1541 1542 peek_stack: 1543 if (cur_stack == 0) 1544 goto check_state; 1545 t = insn_stack[cur_stack - 1]; 1546 1547 if (BPF_CLASS(insns[t].code) == BPF_JMP) { 1548 u8 opcode = BPF_OP(insns[t].code); 1549 1550 if (opcode == BPF_EXIT) { 1551 goto mark_explored; 1552 } else if (opcode == BPF_CALL) { 1553 ret = push_insn(t, t + 1, FALLTHROUGH, env); 1554 if (ret == 1) 1555 goto peek_stack; 1556 else if (ret < 0) 1557 goto err_free; 1558 } else if (opcode == BPF_JA) { 1559 if (BPF_SRC(insns[t].code) != BPF_K) { 1560 ret = -EINVAL; 1561 goto err_free; 1562 } 1563 /* unconditional jump with single edge */ 1564 ret = push_insn(t, t + insns[t].off + 1, 1565 FALLTHROUGH, env); 1566 if (ret == 1) 1567 goto peek_stack; 1568 else if (ret < 0) 1569 goto err_free; 1570 /* tell verifier to check for equivalent states 1571 * after every call and jump 1572 */ 1573 if (t + 1 < insn_cnt) 1574 env->explored_states[t + 1] = STATE_LIST_MARK; 1575 } else { 1576 /* conditional jump with two edges */ 1577 ret = push_insn(t, t + 1, FALLTHROUGH, env); 1578 if (ret == 1) 1579 goto peek_stack; 1580 else if (ret < 0) 1581 goto err_free; 1582 1583 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env); 1584 if (ret == 1) 1585 goto peek_stack; 1586 else if (ret < 0) 1587 goto err_free; 1588 } 1589 } else { 1590 /* all other non-branch instructions with single 1591 * fall-through edge 1592 */ 1593 ret = push_insn(t, t + 1, FALLTHROUGH, env); 1594 if (ret == 1) 1595 goto peek_stack; 1596 else if (ret < 0) 1597 goto err_free; 1598 } 1599 1600 mark_explored: 1601 insn_state[t] = EXPLORED; 1602 if (cur_stack-- <= 0) { 1603 verbose("pop stack internal bug\n"); 1604 ret = -EFAULT; 1605 goto err_free; 1606 } 1607 goto peek_stack; 1608 1609 check_state: 1610 for (i = 0; i < insn_cnt; i++) { 1611 if (insn_state[i] != EXPLORED) { 1612 verbose("unreachable insn %d\n", i); 1613 ret = -EINVAL; 1614 goto err_free; 1615 } 1616 } 1617 ret = 0; /* cfg looks good */ 1618 1619 err_free: 1620 kfree(insn_state); 1621 kfree(insn_stack); 1622 return ret; 1623 } 1624 1625 /* compare two verifier states 1626 * 1627 * all states stored in state_list are known to be valid, since 1628 * verifier reached 'bpf_exit' instruction through them 1629 * 1630 * this function is called when verifier exploring different branches of 1631 * execution popped from the state stack. If it sees an old state that has 1632 * more strict register state and more strict stack state then this execution 1633 * branch doesn't need to be explored further, since verifier already 1634 * concluded that more strict state leads to valid finish. 1635 * 1636 * Therefore two states are equivalent if register state is more conservative 1637 * and explored stack state is more conservative than the current one. 1638 * Example: 1639 * explored current 1640 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC) 1641 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC) 1642 * 1643 * In other words if current stack state (one being explored) has more 1644 * valid slots than old one that already passed validation, it means 1645 * the verifier can stop exploring and conclude that current state is valid too 1646 * 1647 * Similarly with registers. If explored state has register type as invalid 1648 * whereas register type in current state is meaningful, it means that 1649 * the current state will reach 'bpf_exit' instruction safely 1650 */ 1651 static bool states_equal(struct verifier_state *old, struct verifier_state *cur) 1652 { 1653 int i; 1654 1655 for (i = 0; i < MAX_BPF_REG; i++) { 1656 if (memcmp(&old->regs[i], &cur->regs[i], 1657 sizeof(old->regs[0])) != 0) { 1658 if (old->regs[i].type == NOT_INIT || 1659 (old->regs[i].type == UNKNOWN_VALUE && 1660 cur->regs[i].type != NOT_INIT)) 1661 continue; 1662 return false; 1663 } 1664 } 1665 1666 for (i = 0; i < MAX_BPF_STACK; i++) { 1667 if (old->stack_slot_type[i] == STACK_INVALID) 1668 continue; 1669 if (old->stack_slot_type[i] != cur->stack_slot_type[i]) 1670 /* Ex: old explored (safe) state has STACK_SPILL in 1671 * this stack slot, but current has has STACK_MISC -> 1672 * this verifier states are not equivalent, 1673 * return false to continue verification of this path 1674 */ 1675 return false; 1676 if (i % BPF_REG_SIZE) 1677 continue; 1678 if (memcmp(&old->spilled_regs[i / BPF_REG_SIZE], 1679 &cur->spilled_regs[i / BPF_REG_SIZE], 1680 sizeof(old->spilled_regs[0]))) 1681 /* when explored and current stack slot types are 1682 * the same, check that stored pointers types 1683 * are the same as well. 1684 * Ex: explored safe path could have stored 1685 * (struct reg_state) {.type = PTR_TO_STACK, .imm = -8} 1686 * but current path has stored: 1687 * (struct reg_state) {.type = PTR_TO_STACK, .imm = -16} 1688 * such verifier states are not equivalent. 1689 * return false to continue verification of this path 1690 */ 1691 return false; 1692 else 1693 continue; 1694 } 1695 return true; 1696 } 1697 1698 static int is_state_visited(struct verifier_env *env, int insn_idx) 1699 { 1700 struct verifier_state_list *new_sl; 1701 struct verifier_state_list *sl; 1702 1703 sl = env->explored_states[insn_idx]; 1704 if (!sl) 1705 /* this 'insn_idx' instruction wasn't marked, so we will not 1706 * be doing state search here 1707 */ 1708 return 0; 1709 1710 while (sl != STATE_LIST_MARK) { 1711 if (states_equal(&sl->state, &env->cur_state)) 1712 /* reached equivalent register/stack state, 1713 * prune the search 1714 */ 1715 return 1; 1716 sl = sl->next; 1717 } 1718 1719 /* there were no equivalent states, remember current one. 1720 * technically the current state is not proven to be safe yet, 1721 * but it will either reach bpf_exit (which means it's safe) or 1722 * it will be rejected. Since there are no loops, we won't be 1723 * seeing this 'insn_idx' instruction again on the way to bpf_exit 1724 */ 1725 new_sl = kmalloc(sizeof(struct verifier_state_list), GFP_USER); 1726 if (!new_sl) 1727 return -ENOMEM; 1728 1729 /* add new state to the head of linked list */ 1730 memcpy(&new_sl->state, &env->cur_state, sizeof(env->cur_state)); 1731 new_sl->next = env->explored_states[insn_idx]; 1732 env->explored_states[insn_idx] = new_sl; 1733 return 0; 1734 } 1735 1736 static int do_check(struct verifier_env *env) 1737 { 1738 struct verifier_state *state = &env->cur_state; 1739 struct bpf_insn *insns = env->prog->insnsi; 1740 struct reg_state *regs = state->regs; 1741 int insn_cnt = env->prog->len; 1742 int insn_idx, prev_insn_idx = 0; 1743 int insn_processed = 0; 1744 bool do_print_state = false; 1745 1746 init_reg_state(regs); 1747 insn_idx = 0; 1748 for (;;) { 1749 struct bpf_insn *insn; 1750 u8 class; 1751 int err; 1752 1753 if (insn_idx >= insn_cnt) { 1754 verbose("invalid insn idx %d insn_cnt %d\n", 1755 insn_idx, insn_cnt); 1756 return -EFAULT; 1757 } 1758 1759 insn = &insns[insn_idx]; 1760 class = BPF_CLASS(insn->code); 1761 1762 if (++insn_processed > 32768) { 1763 verbose("BPF program is too large. Proccessed %d insn\n", 1764 insn_processed); 1765 return -E2BIG; 1766 } 1767 1768 err = is_state_visited(env, insn_idx); 1769 if (err < 0) 1770 return err; 1771 if (err == 1) { 1772 /* found equivalent state, can prune the search */ 1773 if (log_level) { 1774 if (do_print_state) 1775 verbose("\nfrom %d to %d: safe\n", 1776 prev_insn_idx, insn_idx); 1777 else 1778 verbose("%d: safe\n", insn_idx); 1779 } 1780 goto process_bpf_exit; 1781 } 1782 1783 if (log_level && do_print_state) { 1784 verbose("\nfrom %d to %d:", prev_insn_idx, insn_idx); 1785 print_verifier_state(env); 1786 do_print_state = false; 1787 } 1788 1789 if (log_level) { 1790 verbose("%d: ", insn_idx); 1791 print_bpf_insn(insn); 1792 } 1793 1794 if (class == BPF_ALU || class == BPF_ALU64) { 1795 err = check_alu_op(env, insn); 1796 if (err) 1797 return err; 1798 1799 } else if (class == BPF_LDX) { 1800 enum bpf_reg_type src_reg_type; 1801 1802 /* check for reserved fields is already done */ 1803 1804 /* check src operand */ 1805 err = check_reg_arg(regs, insn->src_reg, SRC_OP); 1806 if (err) 1807 return err; 1808 1809 err = check_reg_arg(regs, insn->dst_reg, DST_OP_NO_MARK); 1810 if (err) 1811 return err; 1812 1813 src_reg_type = regs[insn->src_reg].type; 1814 1815 /* check that memory (src_reg + off) is readable, 1816 * the state of dst_reg will be updated by this func 1817 */ 1818 err = check_mem_access(env, insn->src_reg, insn->off, 1819 BPF_SIZE(insn->code), BPF_READ, 1820 insn->dst_reg); 1821 if (err) 1822 return err; 1823 1824 if (BPF_SIZE(insn->code) != BPF_W) { 1825 insn_idx++; 1826 continue; 1827 } 1828 1829 if (insn->imm == 0) { 1830 /* saw a valid insn 1831 * dst_reg = *(u32 *)(src_reg + off) 1832 * use reserved 'imm' field to mark this insn 1833 */ 1834 insn->imm = src_reg_type; 1835 1836 } else if (src_reg_type != insn->imm && 1837 (src_reg_type == PTR_TO_CTX || 1838 insn->imm == PTR_TO_CTX)) { 1839 /* ABuser program is trying to use the same insn 1840 * dst_reg = *(u32*) (src_reg + off) 1841 * with different pointer types: 1842 * src_reg == ctx in one branch and 1843 * src_reg == stack|map in some other branch. 1844 * Reject it. 1845 */ 1846 verbose("same insn cannot be used with different pointers\n"); 1847 return -EINVAL; 1848 } 1849 1850 } else if (class == BPF_STX) { 1851 enum bpf_reg_type dst_reg_type; 1852 1853 if (BPF_MODE(insn->code) == BPF_XADD) { 1854 err = check_xadd(env, insn); 1855 if (err) 1856 return err; 1857 insn_idx++; 1858 continue; 1859 } 1860 1861 /* check src1 operand */ 1862 err = check_reg_arg(regs, insn->src_reg, SRC_OP); 1863 if (err) 1864 return err; 1865 /* check src2 operand */ 1866 err = check_reg_arg(regs, insn->dst_reg, SRC_OP); 1867 if (err) 1868 return err; 1869 1870 dst_reg_type = regs[insn->dst_reg].type; 1871 1872 /* check that memory (dst_reg + off) is writeable */ 1873 err = check_mem_access(env, insn->dst_reg, insn->off, 1874 BPF_SIZE(insn->code), BPF_WRITE, 1875 insn->src_reg); 1876 if (err) 1877 return err; 1878 1879 if (insn->imm == 0) { 1880 insn->imm = dst_reg_type; 1881 } else if (dst_reg_type != insn->imm && 1882 (dst_reg_type == PTR_TO_CTX || 1883 insn->imm == PTR_TO_CTX)) { 1884 verbose("same insn cannot be used with different pointers\n"); 1885 return -EINVAL; 1886 } 1887 1888 } else if (class == BPF_ST) { 1889 if (BPF_MODE(insn->code) != BPF_MEM || 1890 insn->src_reg != BPF_REG_0) { 1891 verbose("BPF_ST uses reserved fields\n"); 1892 return -EINVAL; 1893 } 1894 /* check src operand */ 1895 err = check_reg_arg(regs, insn->dst_reg, SRC_OP); 1896 if (err) 1897 return err; 1898 1899 /* check that memory (dst_reg + off) is writeable */ 1900 err = check_mem_access(env, insn->dst_reg, insn->off, 1901 BPF_SIZE(insn->code), BPF_WRITE, 1902 -1); 1903 if (err) 1904 return err; 1905 1906 } else if (class == BPF_JMP) { 1907 u8 opcode = BPF_OP(insn->code); 1908 1909 if (opcode == BPF_CALL) { 1910 if (BPF_SRC(insn->code) != BPF_K || 1911 insn->off != 0 || 1912 insn->src_reg != BPF_REG_0 || 1913 insn->dst_reg != BPF_REG_0) { 1914 verbose("BPF_CALL uses reserved fields\n"); 1915 return -EINVAL; 1916 } 1917 1918 err = check_call(env, insn->imm); 1919 if (err) 1920 return err; 1921 1922 } else if (opcode == BPF_JA) { 1923 if (BPF_SRC(insn->code) != BPF_K || 1924 insn->imm != 0 || 1925 insn->src_reg != BPF_REG_0 || 1926 insn->dst_reg != BPF_REG_0) { 1927 verbose("BPF_JA uses reserved fields\n"); 1928 return -EINVAL; 1929 } 1930 1931 insn_idx += insn->off + 1; 1932 continue; 1933 1934 } else if (opcode == BPF_EXIT) { 1935 if (BPF_SRC(insn->code) != BPF_K || 1936 insn->imm != 0 || 1937 insn->src_reg != BPF_REG_0 || 1938 insn->dst_reg != BPF_REG_0) { 1939 verbose("BPF_EXIT uses reserved fields\n"); 1940 return -EINVAL; 1941 } 1942 1943 /* eBPF calling convetion is such that R0 is used 1944 * to return the value from eBPF program. 1945 * Make sure that it's readable at this time 1946 * of bpf_exit, which means that program wrote 1947 * something into it earlier 1948 */ 1949 err = check_reg_arg(regs, BPF_REG_0, SRC_OP); 1950 if (err) 1951 return err; 1952 1953 if (is_pointer_value(env, BPF_REG_0)) { 1954 verbose("R0 leaks addr as return value\n"); 1955 return -EACCES; 1956 } 1957 1958 process_bpf_exit: 1959 insn_idx = pop_stack(env, &prev_insn_idx); 1960 if (insn_idx < 0) { 1961 break; 1962 } else { 1963 do_print_state = true; 1964 continue; 1965 } 1966 } else { 1967 err = check_cond_jmp_op(env, insn, &insn_idx); 1968 if (err) 1969 return err; 1970 } 1971 } else if (class == BPF_LD) { 1972 u8 mode = BPF_MODE(insn->code); 1973 1974 if (mode == BPF_ABS || mode == BPF_IND) { 1975 err = check_ld_abs(env, insn); 1976 if (err) 1977 return err; 1978 1979 } else if (mode == BPF_IMM) { 1980 err = check_ld_imm(env, insn); 1981 if (err) 1982 return err; 1983 1984 insn_idx++; 1985 } else { 1986 verbose("invalid BPF_LD mode\n"); 1987 return -EINVAL; 1988 } 1989 } else { 1990 verbose("unknown insn class %d\n", class); 1991 return -EINVAL; 1992 } 1993 1994 insn_idx++; 1995 } 1996 1997 return 0; 1998 } 1999 2000 /* look for pseudo eBPF instructions that access map FDs and 2001 * replace them with actual map pointers 2002 */ 2003 static int replace_map_fd_with_map_ptr(struct verifier_env *env) 2004 { 2005 struct bpf_insn *insn = env->prog->insnsi; 2006 int insn_cnt = env->prog->len; 2007 int i, j; 2008 2009 for (i = 0; i < insn_cnt; i++, insn++) { 2010 if (BPF_CLASS(insn->code) == BPF_LDX && 2011 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) { 2012 verbose("BPF_LDX uses reserved fields\n"); 2013 return -EINVAL; 2014 } 2015 2016 if (BPF_CLASS(insn->code) == BPF_STX && 2017 ((BPF_MODE(insn->code) != BPF_MEM && 2018 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) { 2019 verbose("BPF_STX uses reserved fields\n"); 2020 return -EINVAL; 2021 } 2022 2023 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) { 2024 struct bpf_map *map; 2025 struct fd f; 2026 2027 if (i == insn_cnt - 1 || insn[1].code != 0 || 2028 insn[1].dst_reg != 0 || insn[1].src_reg != 0 || 2029 insn[1].off != 0) { 2030 verbose("invalid bpf_ld_imm64 insn\n"); 2031 return -EINVAL; 2032 } 2033 2034 if (insn->src_reg == 0) 2035 /* valid generic load 64-bit imm */ 2036 goto next_insn; 2037 2038 if (insn->src_reg != BPF_PSEUDO_MAP_FD) { 2039 verbose("unrecognized bpf_ld_imm64 insn\n"); 2040 return -EINVAL; 2041 } 2042 2043 f = fdget(insn->imm); 2044 map = __bpf_map_get(f); 2045 if (IS_ERR(map)) { 2046 verbose("fd %d is not pointing to valid bpf_map\n", 2047 insn->imm); 2048 return PTR_ERR(map); 2049 } 2050 2051 /* store map pointer inside BPF_LD_IMM64 instruction */ 2052 insn[0].imm = (u32) (unsigned long) map; 2053 insn[1].imm = ((u64) (unsigned long) map) >> 32; 2054 2055 /* check whether we recorded this map already */ 2056 for (j = 0; j < env->used_map_cnt; j++) 2057 if (env->used_maps[j] == map) { 2058 fdput(f); 2059 goto next_insn; 2060 } 2061 2062 if (env->used_map_cnt >= MAX_USED_MAPS) { 2063 fdput(f); 2064 return -E2BIG; 2065 } 2066 2067 /* hold the map. If the program is rejected by verifier, 2068 * the map will be released by release_maps() or it 2069 * will be used by the valid program until it's unloaded 2070 * and all maps are released in free_bpf_prog_info() 2071 */ 2072 map = bpf_map_inc(map, false); 2073 if (IS_ERR(map)) { 2074 fdput(f); 2075 return PTR_ERR(map); 2076 } 2077 env->used_maps[env->used_map_cnt++] = map; 2078 2079 fdput(f); 2080 next_insn: 2081 insn++; 2082 i++; 2083 } 2084 } 2085 2086 /* now all pseudo BPF_LD_IMM64 instructions load valid 2087 * 'struct bpf_map *' into a register instead of user map_fd. 2088 * These pointers will be used later by verifier to validate map access. 2089 */ 2090 return 0; 2091 } 2092 2093 /* drop refcnt of maps used by the rejected program */ 2094 static void release_maps(struct verifier_env *env) 2095 { 2096 int i; 2097 2098 for (i = 0; i < env->used_map_cnt; i++) 2099 bpf_map_put(env->used_maps[i]); 2100 } 2101 2102 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */ 2103 static void convert_pseudo_ld_imm64(struct verifier_env *env) 2104 { 2105 struct bpf_insn *insn = env->prog->insnsi; 2106 int insn_cnt = env->prog->len; 2107 int i; 2108 2109 for (i = 0; i < insn_cnt; i++, insn++) 2110 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW)) 2111 insn->src_reg = 0; 2112 } 2113 2114 static void adjust_branches(struct bpf_prog *prog, int pos, int delta) 2115 { 2116 struct bpf_insn *insn = prog->insnsi; 2117 int insn_cnt = prog->len; 2118 int i; 2119 2120 for (i = 0; i < insn_cnt; i++, insn++) { 2121 if (BPF_CLASS(insn->code) != BPF_JMP || 2122 BPF_OP(insn->code) == BPF_CALL || 2123 BPF_OP(insn->code) == BPF_EXIT) 2124 continue; 2125 2126 /* adjust offset of jmps if necessary */ 2127 if (i < pos && i + insn->off + 1 > pos) 2128 insn->off += delta; 2129 else if (i > pos + delta && i + insn->off + 1 <= pos + delta) 2130 insn->off -= delta; 2131 } 2132 } 2133 2134 /* convert load instructions that access fields of 'struct __sk_buff' 2135 * into sequence of instructions that access fields of 'struct sk_buff' 2136 */ 2137 static int convert_ctx_accesses(struct verifier_env *env) 2138 { 2139 struct bpf_insn *insn = env->prog->insnsi; 2140 int insn_cnt = env->prog->len; 2141 struct bpf_insn insn_buf[16]; 2142 struct bpf_prog *new_prog; 2143 u32 cnt; 2144 int i; 2145 enum bpf_access_type type; 2146 2147 if (!env->prog->aux->ops->convert_ctx_access) 2148 return 0; 2149 2150 for (i = 0; i < insn_cnt; i++, insn++) { 2151 if (insn->code == (BPF_LDX | BPF_MEM | BPF_W)) 2152 type = BPF_READ; 2153 else if (insn->code == (BPF_STX | BPF_MEM | BPF_W)) 2154 type = BPF_WRITE; 2155 else 2156 continue; 2157 2158 if (insn->imm != PTR_TO_CTX) { 2159 /* clear internal mark */ 2160 insn->imm = 0; 2161 continue; 2162 } 2163 2164 cnt = env->prog->aux->ops-> 2165 convert_ctx_access(type, insn->dst_reg, insn->src_reg, 2166 insn->off, insn_buf, env->prog); 2167 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) { 2168 verbose("bpf verifier is misconfigured\n"); 2169 return -EINVAL; 2170 } 2171 2172 if (cnt == 1) { 2173 memcpy(insn, insn_buf, sizeof(*insn)); 2174 continue; 2175 } 2176 2177 /* several new insns need to be inserted. Make room for them */ 2178 insn_cnt += cnt - 1; 2179 new_prog = bpf_prog_realloc(env->prog, 2180 bpf_prog_size(insn_cnt), 2181 GFP_USER); 2182 if (!new_prog) 2183 return -ENOMEM; 2184 2185 new_prog->len = insn_cnt; 2186 2187 memmove(new_prog->insnsi + i + cnt, new_prog->insns + i + 1, 2188 sizeof(*insn) * (insn_cnt - i - cnt)); 2189 2190 /* copy substitute insns in place of load instruction */ 2191 memcpy(new_prog->insnsi + i, insn_buf, sizeof(*insn) * cnt); 2192 2193 /* adjust branches in the whole program */ 2194 adjust_branches(new_prog, i, cnt - 1); 2195 2196 /* keep walking new program and skip insns we just inserted */ 2197 env->prog = new_prog; 2198 insn = new_prog->insnsi + i + cnt - 1; 2199 i += cnt - 1; 2200 } 2201 2202 return 0; 2203 } 2204 2205 static void free_states(struct verifier_env *env) 2206 { 2207 struct verifier_state_list *sl, *sln; 2208 int i; 2209 2210 if (!env->explored_states) 2211 return; 2212 2213 for (i = 0; i < env->prog->len; i++) { 2214 sl = env->explored_states[i]; 2215 2216 if (sl) 2217 while (sl != STATE_LIST_MARK) { 2218 sln = sl->next; 2219 kfree(sl); 2220 sl = sln; 2221 } 2222 } 2223 2224 kfree(env->explored_states); 2225 } 2226 2227 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr) 2228 { 2229 char __user *log_ubuf = NULL; 2230 struct verifier_env *env; 2231 int ret = -EINVAL; 2232 2233 if ((*prog)->len <= 0 || (*prog)->len > BPF_MAXINSNS) 2234 return -E2BIG; 2235 2236 /* 'struct verifier_env' can be global, but since it's not small, 2237 * allocate/free it every time bpf_check() is called 2238 */ 2239 env = kzalloc(sizeof(struct verifier_env), GFP_KERNEL); 2240 if (!env) 2241 return -ENOMEM; 2242 2243 env->prog = *prog; 2244 2245 /* grab the mutex to protect few globals used by verifier */ 2246 mutex_lock(&bpf_verifier_lock); 2247 2248 if (attr->log_level || attr->log_buf || attr->log_size) { 2249 /* user requested verbose verifier output 2250 * and supplied buffer to store the verification trace 2251 */ 2252 log_level = attr->log_level; 2253 log_ubuf = (char __user *) (unsigned long) attr->log_buf; 2254 log_size = attr->log_size; 2255 log_len = 0; 2256 2257 ret = -EINVAL; 2258 /* log_* values have to be sane */ 2259 if (log_size < 128 || log_size > UINT_MAX >> 8 || 2260 log_level == 0 || log_ubuf == NULL) 2261 goto free_env; 2262 2263 ret = -ENOMEM; 2264 log_buf = vmalloc(log_size); 2265 if (!log_buf) 2266 goto free_env; 2267 } else { 2268 log_level = 0; 2269 } 2270 2271 ret = replace_map_fd_with_map_ptr(env); 2272 if (ret < 0) 2273 goto skip_full_check; 2274 2275 env->explored_states = kcalloc(env->prog->len, 2276 sizeof(struct verifier_state_list *), 2277 GFP_USER); 2278 ret = -ENOMEM; 2279 if (!env->explored_states) 2280 goto skip_full_check; 2281 2282 ret = check_cfg(env); 2283 if (ret < 0) 2284 goto skip_full_check; 2285 2286 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN); 2287 2288 ret = do_check(env); 2289 2290 skip_full_check: 2291 while (pop_stack(env, NULL) >= 0); 2292 free_states(env); 2293 2294 if (ret == 0) 2295 /* program is valid, convert *(u32*)(ctx + off) accesses */ 2296 ret = convert_ctx_accesses(env); 2297 2298 if (log_level && log_len >= log_size - 1) { 2299 BUG_ON(log_len >= log_size); 2300 /* verifier log exceeded user supplied buffer */ 2301 ret = -ENOSPC; 2302 /* fall through to return what was recorded */ 2303 } 2304 2305 /* copy verifier log back to user space including trailing zero */ 2306 if (log_level && copy_to_user(log_ubuf, log_buf, log_len + 1) != 0) { 2307 ret = -EFAULT; 2308 goto free_log_buf; 2309 } 2310 2311 if (ret == 0 && env->used_map_cnt) { 2312 /* if program passed verifier, update used_maps in bpf_prog_info */ 2313 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt, 2314 sizeof(env->used_maps[0]), 2315 GFP_KERNEL); 2316 2317 if (!env->prog->aux->used_maps) { 2318 ret = -ENOMEM; 2319 goto free_log_buf; 2320 } 2321 2322 memcpy(env->prog->aux->used_maps, env->used_maps, 2323 sizeof(env->used_maps[0]) * env->used_map_cnt); 2324 env->prog->aux->used_map_cnt = env->used_map_cnt; 2325 2326 /* program is valid. Convert pseudo bpf_ld_imm64 into generic 2327 * bpf_ld_imm64 instructions 2328 */ 2329 convert_pseudo_ld_imm64(env); 2330 } 2331 2332 free_log_buf: 2333 if (log_level) 2334 vfree(log_buf); 2335 free_env: 2336 if (!env->prog->aux->used_maps) 2337 /* if we didn't copy map pointers into bpf_prog_info, release 2338 * them now. Otherwise free_bpf_prog_info() will release them. 2339 */ 2340 release_maps(env); 2341 *prog = env->prog; 2342 kfree(env); 2343 mutex_unlock(&bpf_verifier_lock); 2344 return ret; 2345 } 2346