| 1 | /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com |
| 2 | * Copyright (c) 2016 Facebook |
| 3 | * |
| 4 | * This program is free software; you can redistribute it and/or |
| 5 | * modify it under the terms of version 2 of the GNU General Public |
| 6 | * License as published by the Free Software Foundation. |
| 7 | * |
| 8 | * This program is distributed in the hope that it will be useful, but |
| 9 | * WITHOUT ANY WARRANTY; without even the implied warranty of |
| 10 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 11 | * General Public License for more details. |
| 12 | */ |
| 13 | #include <linux/kernel.h> |
| 14 | #include <linux/types.h> |
| 15 | #include <linux/slab.h> |
| 16 | #include <linux/bpf.h> |
| 17 | #include <linux/bpf_verifier.h> |
| 18 | #include <linux/filter.h> |
| 19 | #include <net/netlink.h> |
| 20 | #include <linux/file.h> |
| 21 | #include <linux/vmalloc.h> |
| 22 | #include <linux/stringify.h> |
| 23 | #include <linux/bsearch.h> |
| 24 | #include <linux/sort.h> |
| 25 | #include <linux/perf_event.h> |
| 26 | |
| 27 | #include "disasm.h" |
| 28 | |
| 29 | static const struct bpf_verifier_ops * const bpf_verifier_ops[] = { |
| 30 | #define BPF_PROG_TYPE(_id, _name) \ |
| 31 | [_id] = & _name ## _verifier_ops, |
| 32 | #define BPF_MAP_TYPE(_id, _ops) |
| 33 | #include <linux/bpf_types.h> |
| 34 | #undef BPF_PROG_TYPE |
| 35 | #undef BPF_MAP_TYPE |
| 36 | }; |
| 37 | |
| 38 | /* bpf_check() is a static code analyzer that walks eBPF program |
| 39 | * instruction by instruction and updates register/stack state. |
| 40 | * All paths of conditional branches are analyzed until 'bpf_exit' insn. |
| 41 | * |
| 42 | * The first pass is depth-first-search to check that the program is a DAG. |
| 43 | * It rejects the following programs: |
| 44 | * - larger than BPF_MAXINSNS insns |
| 45 | * - if loop is present (detected via back-edge) |
| 46 | * - unreachable insns exist (shouldn't be a forest. program = one function) |
| 47 | * - out of bounds or malformed jumps |
| 48 | * The second pass is all possible path descent from the 1st insn. |
| 49 | * Since it's analyzing all pathes through the program, the length of the |
| 50 | * analysis is limited to 64k insn, which may be hit even if total number of |
| 51 | * insn is less then 4K, but there are too many branches that change stack/regs. |
| 52 | * Number of 'branches to be analyzed' is limited to 1k |
| 53 | * |
| 54 | * On entry to each instruction, each register has a type, and the instruction |
| 55 | * changes the types of the registers depending on instruction semantics. |
| 56 | * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is |
| 57 | * copied to R1. |
| 58 | * |
| 59 | * All registers are 64-bit. |
| 60 | * R0 - return register |
| 61 | * R1-R5 argument passing registers |
| 62 | * R6-R9 callee saved registers |
| 63 | * R10 - frame pointer read-only |
| 64 | * |
| 65 | * At the start of BPF program the register R1 contains a pointer to bpf_context |
| 66 | * and has type PTR_TO_CTX. |
| 67 | * |
| 68 | * Verifier tracks arithmetic operations on pointers in case: |
| 69 | * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10), |
| 70 | * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20), |
| 71 | * 1st insn copies R10 (which has FRAME_PTR) type into R1 |
| 72 | * and 2nd arithmetic instruction is pattern matched to recognize |
| 73 | * that it wants to construct a pointer to some element within stack. |
| 74 | * So after 2nd insn, the register R1 has type PTR_TO_STACK |
| 75 | * (and -20 constant is saved for further stack bounds checking). |
| 76 | * Meaning that this reg is a pointer to stack plus known immediate constant. |
| 77 | * |
| 78 | * Most of the time the registers have SCALAR_VALUE type, which |
| 79 | * means the register has some value, but it's not a valid pointer. |
| 80 | * (like pointer plus pointer becomes SCALAR_VALUE type) |
| 81 | * |
| 82 | * When verifier sees load or store instructions the type of base register |
| 83 | * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK. These are three pointer |
| 84 | * types recognized by check_mem_access() function. |
| 85 | * |
| 86 | * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value' |
| 87 | * and the range of [ptr, ptr + map's value_size) is accessible. |
| 88 | * |
| 89 | * registers used to pass values to function calls are checked against |
| 90 | * function argument constraints. |
| 91 | * |
| 92 | * ARG_PTR_TO_MAP_KEY is one of such argument constraints. |
| 93 | * It means that the register type passed to this function must be |
| 94 | * PTR_TO_STACK and it will be used inside the function as |
| 95 | * 'pointer to map element key' |
| 96 | * |
| 97 | * For example the argument constraints for bpf_map_lookup_elem(): |
| 98 | * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL, |
| 99 | * .arg1_type = ARG_CONST_MAP_PTR, |
| 100 | * .arg2_type = ARG_PTR_TO_MAP_KEY, |
| 101 | * |
| 102 | * ret_type says that this function returns 'pointer to map elem value or null' |
| 103 | * function expects 1st argument to be a const pointer to 'struct bpf_map' and |
| 104 | * 2nd argument should be a pointer to stack, which will be used inside |
| 105 | * the helper function as a pointer to map element key. |
| 106 | * |
| 107 | * On the kernel side the helper function looks like: |
| 108 | * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) |
| 109 | * { |
| 110 | * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1; |
| 111 | * void *key = (void *) (unsigned long) r2; |
| 112 | * void *value; |
| 113 | * |
| 114 | * here kernel can access 'key' and 'map' pointers safely, knowing that |
| 115 | * [key, key + map->key_size) bytes are valid and were initialized on |
| 116 | * the stack of eBPF program. |
| 117 | * } |
| 118 | * |
| 119 | * Corresponding eBPF program may look like: |
| 120 | * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR |
| 121 | * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK |
| 122 | * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP |
| 123 | * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem), |
| 124 | * here verifier looks at prototype of map_lookup_elem() and sees: |
| 125 | * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok, |
| 126 | * Now verifier knows that this map has key of R1->map_ptr->key_size bytes |
| 127 | * |
| 128 | * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far, |
| 129 | * Now verifier checks that [R2, R2 + map's key_size) are within stack limits |
| 130 | * and were initialized prior to this call. |
| 131 | * If it's ok, then verifier allows this BPF_CALL insn and looks at |
| 132 | * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets |
| 133 | * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function |
| 134 | * returns ether pointer to map value or NULL. |
| 135 | * |
| 136 | * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off' |
| 137 | * insn, the register holding that pointer in the true branch changes state to |
| 138 | * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false |
| 139 | * branch. See check_cond_jmp_op(). |
| 140 | * |
| 141 | * After the call R0 is set to return type of the function and registers R1-R5 |
| 142 | * are set to NOT_INIT to indicate that they are no longer readable. |
| 143 | */ |
| 144 | |
| 145 | /* verifier_state + insn_idx are pushed to stack when branch is encountered */ |
| 146 | struct bpf_verifier_stack_elem { |
| 147 | /* verifer state is 'st' |
| 148 | * before processing instruction 'insn_idx' |
| 149 | * and after processing instruction 'prev_insn_idx' |
| 150 | */ |
| 151 | struct bpf_verifier_state st; |
| 152 | int insn_idx; |
| 153 | int prev_insn_idx; |
| 154 | struct bpf_verifier_stack_elem *next; |
| 155 | }; |
| 156 | |
| 157 | #define BPF_COMPLEXITY_LIMIT_INSNS 131072 |
| 158 | #define BPF_COMPLEXITY_LIMIT_STACK 1024 |
| 159 | |
| 160 | #define BPF_MAP_PTR_POISON ((void *)0xeB9F + POISON_POINTER_DELTA) |
| 161 | |
| 162 | struct bpf_call_arg_meta { |
| 163 | struct bpf_map *map_ptr; |
| 164 | bool raw_mode; |
| 165 | bool pkt_access; |
| 166 | int regno; |
| 167 | int access_size; |
| 168 | s64 msize_smax_value; |
| 169 | u64 msize_umax_value; |
| 170 | }; |
| 171 | |
| 172 | static DEFINE_MUTEX(bpf_verifier_lock); |
| 173 | |
| 174 | void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt, |
| 175 | va_list args) |
| 176 | { |
| 177 | unsigned int n; |
| 178 | |
| 179 | n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args); |
| 180 | |
| 181 | WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1, |
| 182 | "verifier log line truncated - local buffer too short\n"); |
| 183 | |
| 184 | n = min(log->len_total - log->len_used - 1, n); |
| 185 | log->kbuf[n] = '\0'; |
| 186 | |
| 187 | if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1)) |
| 188 | log->len_used += n; |
| 189 | else |
| 190 | log->ubuf = NULL; |
| 191 | } |
| 192 | |
| 193 | /* log_level controls verbosity level of eBPF verifier. |
| 194 | * bpf_verifier_log_write() is used to dump the verification trace to the log, |
| 195 | * so the user can figure out what's wrong with the program |
| 196 | */ |
| 197 | __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env, |
| 198 | const char *fmt, ...) |
| 199 | { |
| 200 | va_list args; |
| 201 | |
| 202 | if (!bpf_verifier_log_needed(&env->log)) |
| 203 | return; |
| 204 | |
| 205 | va_start(args, fmt); |
| 206 | bpf_verifier_vlog(&env->log, fmt, args); |
| 207 | va_end(args); |
| 208 | } |
| 209 | EXPORT_SYMBOL_GPL(bpf_verifier_log_write); |
| 210 | |
| 211 | __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...) |
| 212 | { |
| 213 | struct bpf_verifier_env *env = private_data; |
| 214 | va_list args; |
| 215 | |
| 216 | if (!bpf_verifier_log_needed(&env->log)) |
| 217 | return; |
| 218 | |
| 219 | va_start(args, fmt); |
| 220 | bpf_verifier_vlog(&env->log, fmt, args); |
| 221 | va_end(args); |
| 222 | } |
| 223 | |
| 224 | static bool type_is_pkt_pointer(enum bpf_reg_type type) |
| 225 | { |
| 226 | return type == PTR_TO_PACKET || |
| 227 | type == PTR_TO_PACKET_META; |
| 228 | } |
| 229 | |
| 230 | /* string representation of 'enum bpf_reg_type' */ |
| 231 | static const char * const reg_type_str[] = { |
| 232 | [NOT_INIT] = "?", |
| 233 | [SCALAR_VALUE] = "inv", |
| 234 | [PTR_TO_CTX] = "ctx", |
| 235 | [CONST_PTR_TO_MAP] = "map_ptr", |
| 236 | [PTR_TO_MAP_VALUE] = "map_value", |
| 237 | [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null", |
| 238 | [PTR_TO_STACK] = "fp", |
| 239 | [PTR_TO_PACKET] = "pkt", |
| 240 | [PTR_TO_PACKET_META] = "pkt_meta", |
| 241 | [PTR_TO_PACKET_END] = "pkt_end", |
| 242 | }; |
| 243 | |
| 244 | static void print_liveness(struct bpf_verifier_env *env, |
| 245 | enum bpf_reg_liveness live) |
| 246 | { |
| 247 | if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN)) |
| 248 | verbose(env, "_"); |
| 249 | if (live & REG_LIVE_READ) |
| 250 | verbose(env, "r"); |
| 251 | if (live & REG_LIVE_WRITTEN) |
| 252 | verbose(env, "w"); |
| 253 | } |
| 254 | |
| 255 | static struct bpf_func_state *func(struct bpf_verifier_env *env, |
| 256 | const struct bpf_reg_state *reg) |
| 257 | { |
| 258 | struct bpf_verifier_state *cur = env->cur_state; |
| 259 | |
| 260 | return cur->frame[reg->frameno]; |
| 261 | } |
| 262 | |
| 263 | static void print_verifier_state(struct bpf_verifier_env *env, |
| 264 | const struct bpf_func_state *state) |
| 265 | { |
| 266 | const struct bpf_reg_state *reg; |
| 267 | enum bpf_reg_type t; |
| 268 | int i; |
| 269 | |
| 270 | if (state->frameno) |
| 271 | verbose(env, " frame%d:", state->frameno); |
| 272 | for (i = 0; i < MAX_BPF_REG; i++) { |
| 273 | reg = &state->regs[i]; |
| 274 | t = reg->type; |
| 275 | if (t == NOT_INIT) |
| 276 | continue; |
| 277 | verbose(env, " R%d", i); |
| 278 | print_liveness(env, reg->live); |
| 279 | verbose(env, "=%s", reg_type_str[t]); |
| 280 | if ((t == SCALAR_VALUE || t == PTR_TO_STACK) && |
| 281 | tnum_is_const(reg->var_off)) { |
| 282 | /* reg->off should be 0 for SCALAR_VALUE */ |
| 283 | verbose(env, "%lld", reg->var_off.value + reg->off); |
| 284 | if (t == PTR_TO_STACK) |
| 285 | verbose(env, ",call_%d", func(env, reg)->callsite); |
| 286 | } else { |
| 287 | verbose(env, "(id=%d", reg->id); |
| 288 | if (t != SCALAR_VALUE) |
| 289 | verbose(env, ",off=%d", reg->off); |
| 290 | if (type_is_pkt_pointer(t)) |
| 291 | verbose(env, ",r=%d", reg->range); |
| 292 | else if (t == CONST_PTR_TO_MAP || |
| 293 | t == PTR_TO_MAP_VALUE || |
| 294 | t == PTR_TO_MAP_VALUE_OR_NULL) |
| 295 | verbose(env, ",ks=%d,vs=%d", |
| 296 | reg->map_ptr->key_size, |
| 297 | reg->map_ptr->value_size); |
| 298 | if (tnum_is_const(reg->var_off)) { |
| 299 | /* Typically an immediate SCALAR_VALUE, but |
| 300 | * could be a pointer whose offset is too big |
| 301 | * for reg->off |
| 302 | */ |
| 303 | verbose(env, ",imm=%llx", reg->var_off.value); |
| 304 | } else { |
| 305 | if (reg->smin_value != reg->umin_value && |
| 306 | reg->smin_value != S64_MIN) |
| 307 | verbose(env, ",smin_value=%lld", |
| 308 | (long long)reg->smin_value); |
| 309 | if (reg->smax_value != reg->umax_value && |
| 310 | reg->smax_value != S64_MAX) |
| 311 | verbose(env, ",smax_value=%lld", |
| 312 | (long long)reg->smax_value); |
| 313 | if (reg->umin_value != 0) |
| 314 | verbose(env, ",umin_value=%llu", |
| 315 | (unsigned long long)reg->umin_value); |
| 316 | if (reg->umax_value != U64_MAX) |
| 317 | verbose(env, ",umax_value=%llu", |
| 318 | (unsigned long long)reg->umax_value); |
| 319 | if (!tnum_is_unknown(reg->var_off)) { |
| 320 | char tn_buf[48]; |
| 321 | |
| 322 | tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off); |
| 323 | verbose(env, ",var_off=%s", tn_buf); |
| 324 | } |
| 325 | } |
| 326 | verbose(env, ")"); |
| 327 | } |
| 328 | } |
| 329 | for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) { |
| 330 | if (state->stack[i].slot_type[0] == STACK_SPILL) { |
| 331 | verbose(env, " fp%d", |
| 332 | (-i - 1) * BPF_REG_SIZE); |
| 333 | print_liveness(env, state->stack[i].spilled_ptr.live); |
| 334 | verbose(env, "=%s", |
| 335 | reg_type_str[state->stack[i].spilled_ptr.type]); |
| 336 | } |
| 337 | if (state->stack[i].slot_type[0] == STACK_ZERO) |
| 338 | verbose(env, " fp%d=0", (-i - 1) * BPF_REG_SIZE); |
| 339 | } |
| 340 | verbose(env, "\n"); |
| 341 | } |
| 342 | |
| 343 | static int copy_stack_state(struct bpf_func_state *dst, |
| 344 | const struct bpf_func_state *src) |
| 345 | { |
| 346 | if (!src->stack) |
| 347 | return 0; |
| 348 | if (WARN_ON_ONCE(dst->allocated_stack < src->allocated_stack)) { |
| 349 | /* internal bug, make state invalid to reject the program */ |
| 350 | memset(dst, 0, sizeof(*dst)); |
| 351 | return -EFAULT; |
| 352 | } |
| 353 | memcpy(dst->stack, src->stack, |
| 354 | sizeof(*src->stack) * (src->allocated_stack / BPF_REG_SIZE)); |
| 355 | return 0; |
| 356 | } |
| 357 | |
| 358 | /* do_check() starts with zero-sized stack in struct bpf_verifier_state to |
| 359 | * make it consume minimal amount of memory. check_stack_write() access from |
| 360 | * the program calls into realloc_func_state() to grow the stack size. |
| 361 | * Note there is a non-zero 'parent' pointer inside bpf_verifier_state |
| 362 | * which this function copies over. It points to previous bpf_verifier_state |
| 363 | * which is never reallocated |
| 364 | */ |
| 365 | static int realloc_func_state(struct bpf_func_state *state, int size, |
| 366 | bool copy_old) |
| 367 | { |
| 368 | u32 old_size = state->allocated_stack; |
| 369 | struct bpf_stack_state *new_stack; |
| 370 | int slot = size / BPF_REG_SIZE; |
| 371 | |
| 372 | if (size <= old_size || !size) { |
| 373 | if (copy_old) |
| 374 | return 0; |
| 375 | state->allocated_stack = slot * BPF_REG_SIZE; |
| 376 | if (!size && old_size) { |
| 377 | kfree(state->stack); |
| 378 | state->stack = NULL; |
| 379 | } |
| 380 | return 0; |
| 381 | } |
| 382 | new_stack = kmalloc_array(slot, sizeof(struct bpf_stack_state), |
| 383 | GFP_KERNEL); |
| 384 | if (!new_stack) |
| 385 | return -ENOMEM; |
| 386 | if (copy_old) { |
| 387 | if (state->stack) |
| 388 | memcpy(new_stack, state->stack, |
| 389 | sizeof(*new_stack) * (old_size / BPF_REG_SIZE)); |
| 390 | memset(new_stack + old_size / BPF_REG_SIZE, 0, |
| 391 | sizeof(*new_stack) * (size - old_size) / BPF_REG_SIZE); |
| 392 | } |
| 393 | state->allocated_stack = slot * BPF_REG_SIZE; |
| 394 | kfree(state->stack); |
| 395 | state->stack = new_stack; |
| 396 | return 0; |
| 397 | } |
| 398 | |
| 399 | static void free_func_state(struct bpf_func_state *state) |
| 400 | { |
| 401 | if (!state) |
| 402 | return; |
| 403 | kfree(state->stack); |
| 404 | kfree(state); |
| 405 | } |
| 406 | |
| 407 | static void free_verifier_state(struct bpf_verifier_state *state, |
| 408 | bool free_self) |
| 409 | { |
| 410 | int i; |
| 411 | |
| 412 | for (i = 0; i <= state->curframe; i++) { |
| 413 | free_func_state(state->frame[i]); |
| 414 | state->frame[i] = NULL; |
| 415 | } |
| 416 | if (free_self) |
| 417 | kfree(state); |
| 418 | } |
| 419 | |
| 420 | /* copy verifier state from src to dst growing dst stack space |
| 421 | * when necessary to accommodate larger src stack |
| 422 | */ |
| 423 | static int copy_func_state(struct bpf_func_state *dst, |
| 424 | const struct bpf_func_state *src) |
| 425 | { |
| 426 | int err; |
| 427 | |
| 428 | err = realloc_func_state(dst, src->allocated_stack, false); |
| 429 | if (err) |
| 430 | return err; |
| 431 | memcpy(dst, src, offsetof(struct bpf_func_state, allocated_stack)); |
| 432 | return copy_stack_state(dst, src); |
| 433 | } |
| 434 | |
| 435 | static int copy_verifier_state(struct bpf_verifier_state *dst_state, |
| 436 | const struct bpf_verifier_state *src) |
| 437 | { |
| 438 | struct bpf_func_state *dst; |
| 439 | int i, err; |
| 440 | |
| 441 | /* if dst has more stack frames then src frame, free them */ |
| 442 | for (i = src->curframe + 1; i <= dst_state->curframe; i++) { |
| 443 | free_func_state(dst_state->frame[i]); |
| 444 | dst_state->frame[i] = NULL; |
| 445 | } |
| 446 | dst_state->curframe = src->curframe; |
| 447 | dst_state->parent = src->parent; |
| 448 | for (i = 0; i <= src->curframe; i++) { |
| 449 | dst = dst_state->frame[i]; |
| 450 | if (!dst) { |
| 451 | dst = kzalloc(sizeof(*dst), GFP_KERNEL); |
| 452 | if (!dst) |
| 453 | return -ENOMEM; |
| 454 | dst_state->frame[i] = dst; |
| 455 | } |
| 456 | err = copy_func_state(dst, src->frame[i]); |
| 457 | if (err) |
| 458 | return err; |
| 459 | } |
| 460 | return 0; |
| 461 | } |
| 462 | |
| 463 | static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx, |
| 464 | int *insn_idx) |
| 465 | { |
| 466 | struct bpf_verifier_state *cur = env->cur_state; |
| 467 | struct bpf_verifier_stack_elem *elem, *head = env->head; |
| 468 | int err; |
| 469 | |
| 470 | if (env->head == NULL) |
| 471 | return -ENOENT; |
| 472 | |
| 473 | if (cur) { |
| 474 | err = copy_verifier_state(cur, &head->st); |
| 475 | if (err) |
| 476 | return err; |
| 477 | } |
| 478 | if (insn_idx) |
| 479 | *insn_idx = head->insn_idx; |
| 480 | if (prev_insn_idx) |
| 481 | *prev_insn_idx = head->prev_insn_idx; |
| 482 | elem = head->next; |
| 483 | free_verifier_state(&head->st, false); |
| 484 | kfree(head); |
| 485 | env->head = elem; |
| 486 | env->stack_size--; |
| 487 | return 0; |
| 488 | } |
| 489 | |
| 490 | static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env, |
| 491 | int insn_idx, int prev_insn_idx) |
| 492 | { |
| 493 | struct bpf_verifier_state *cur = env->cur_state; |
| 494 | struct bpf_verifier_stack_elem *elem; |
| 495 | int err; |
| 496 | |
| 497 | elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL); |
| 498 | if (!elem) |
| 499 | goto err; |
| 500 | |
| 501 | elem->insn_idx = insn_idx; |
| 502 | elem->prev_insn_idx = prev_insn_idx; |
| 503 | elem->next = env->head; |
| 504 | env->head = elem; |
| 505 | env->stack_size++; |
| 506 | err = copy_verifier_state(&elem->st, cur); |
| 507 | if (err) |
| 508 | goto err; |
| 509 | if (env->stack_size > BPF_COMPLEXITY_LIMIT_STACK) { |
| 510 | verbose(env, "BPF program is too complex\n"); |
| 511 | goto err; |
| 512 | } |
| 513 | return &elem->st; |
| 514 | err: |
| 515 | free_verifier_state(env->cur_state, true); |
| 516 | env->cur_state = NULL; |
| 517 | /* pop all elements and return */ |
| 518 | while (!pop_stack(env, NULL, NULL)); |
| 519 | return NULL; |
| 520 | } |
| 521 | |
| 522 | #define CALLER_SAVED_REGS 6 |
| 523 | static const int caller_saved[CALLER_SAVED_REGS] = { |
| 524 | BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5 |
| 525 | }; |
| 526 | |
| 527 | static void __mark_reg_not_init(struct bpf_reg_state *reg); |
| 528 | |
| 529 | /* Mark the unknown part of a register (variable offset or scalar value) as |
| 530 | * known to have the value @imm. |
| 531 | */ |
| 532 | static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm) |
| 533 | { |
| 534 | reg->id = 0; |
| 535 | reg->var_off = tnum_const(imm); |
| 536 | reg->smin_value = (s64)imm; |
| 537 | reg->smax_value = (s64)imm; |
| 538 | reg->umin_value = imm; |
| 539 | reg->umax_value = imm; |
| 540 | } |
| 541 | |
| 542 | /* Mark the 'variable offset' part of a register as zero. This should be |
| 543 | * used only on registers holding a pointer type. |
| 544 | */ |
| 545 | static void __mark_reg_known_zero(struct bpf_reg_state *reg) |
| 546 | { |
| 547 | __mark_reg_known(reg, 0); |
| 548 | } |
| 549 | |
| 550 | static void __mark_reg_const_zero(struct bpf_reg_state *reg) |
| 551 | { |
| 552 | __mark_reg_known(reg, 0); |
| 553 | reg->off = 0; |
| 554 | reg->type = SCALAR_VALUE; |
| 555 | } |
| 556 | |
| 557 | static void mark_reg_known_zero(struct bpf_verifier_env *env, |
| 558 | struct bpf_reg_state *regs, u32 regno) |
| 559 | { |
| 560 | if (WARN_ON(regno >= MAX_BPF_REG)) { |
| 561 | verbose(env, "mark_reg_known_zero(regs, %u)\n", regno); |
| 562 | /* Something bad happened, let's kill all regs */ |
| 563 | for (regno = 0; regno < MAX_BPF_REG; regno++) |
| 564 | __mark_reg_not_init(regs + regno); |
| 565 | return; |
| 566 | } |
| 567 | __mark_reg_known_zero(regs + regno); |
| 568 | } |
| 569 | |
| 570 | static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg) |
| 571 | { |
| 572 | return type_is_pkt_pointer(reg->type); |
| 573 | } |
| 574 | |
| 575 | static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg) |
| 576 | { |
| 577 | return reg_is_pkt_pointer(reg) || |
| 578 | reg->type == PTR_TO_PACKET_END; |
| 579 | } |
| 580 | |
| 581 | /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */ |
| 582 | static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg, |
| 583 | enum bpf_reg_type which) |
| 584 | { |
| 585 | /* The register can already have a range from prior markings. |
| 586 | * This is fine as long as it hasn't been advanced from its |
| 587 | * origin. |
| 588 | */ |
| 589 | return reg->type == which && |
| 590 | reg->id == 0 && |
| 591 | reg->off == 0 && |
| 592 | tnum_equals_const(reg->var_off, 0); |
| 593 | } |
| 594 | |
| 595 | /* Attempts to improve min/max values based on var_off information */ |
| 596 | static void __update_reg_bounds(struct bpf_reg_state *reg) |
| 597 | { |
| 598 | /* min signed is max(sign bit) | min(other bits) */ |
| 599 | reg->smin_value = max_t(s64, reg->smin_value, |
| 600 | reg->var_off.value | (reg->var_off.mask & S64_MIN)); |
| 601 | /* max signed is min(sign bit) | max(other bits) */ |
| 602 | reg->smax_value = min_t(s64, reg->smax_value, |
| 603 | reg->var_off.value | (reg->var_off.mask & S64_MAX)); |
| 604 | reg->umin_value = max(reg->umin_value, reg->var_off.value); |
| 605 | reg->umax_value = min(reg->umax_value, |
| 606 | reg->var_off.value | reg->var_off.mask); |
| 607 | } |
| 608 | |
| 609 | /* Uses signed min/max values to inform unsigned, and vice-versa */ |
| 610 | static void __reg_deduce_bounds(struct bpf_reg_state *reg) |
| 611 | { |
| 612 | /* Learn sign from signed bounds. |
| 613 | * If we cannot cross the sign boundary, then signed and unsigned bounds |
| 614 | * are the same, so combine. This works even in the negative case, e.g. |
| 615 | * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff. |
| 616 | */ |
| 617 | if (reg->smin_value >= 0 || reg->smax_value < 0) { |
| 618 | reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value, |
| 619 | reg->umin_value); |
| 620 | reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value, |
| 621 | reg->umax_value); |
| 622 | return; |
| 623 | } |
| 624 | /* Learn sign from unsigned bounds. Signed bounds cross the sign |
| 625 | * boundary, so we must be careful. |
| 626 | */ |
| 627 | if ((s64)reg->umax_value >= 0) { |
| 628 | /* Positive. We can't learn anything from the smin, but smax |
| 629 | * is positive, hence safe. |
| 630 | */ |
| 631 | reg->smin_value = reg->umin_value; |
| 632 | reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value, |
| 633 | reg->umax_value); |
| 634 | } else if ((s64)reg->umin_value < 0) { |
| 635 | /* Negative. We can't learn anything from the smax, but smin |
| 636 | * is negative, hence safe. |
| 637 | */ |
| 638 | reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value, |
| 639 | reg->umin_value); |
| 640 | reg->smax_value = reg->umax_value; |
| 641 | } |
| 642 | } |
| 643 | |
| 644 | /* Attempts to improve var_off based on unsigned min/max information */ |
| 645 | static void __reg_bound_offset(struct bpf_reg_state *reg) |
| 646 | { |
| 647 | reg->var_off = tnum_intersect(reg->var_off, |
| 648 | tnum_range(reg->umin_value, |
| 649 | reg->umax_value)); |
| 650 | } |
| 651 | |
| 652 | /* Reset the min/max bounds of a register */ |
| 653 | static void __mark_reg_unbounded(struct bpf_reg_state *reg) |
| 654 | { |
| 655 | reg->smin_value = S64_MIN; |
| 656 | reg->smax_value = S64_MAX; |
| 657 | reg->umin_value = 0; |
| 658 | reg->umax_value = U64_MAX; |
| 659 | } |
| 660 | |
| 661 | /* Mark a register as having a completely unknown (scalar) value. */ |
| 662 | static void __mark_reg_unknown(struct bpf_reg_state *reg) |
| 663 | { |
| 664 | reg->type = SCALAR_VALUE; |
| 665 | reg->id = 0; |
| 666 | reg->off = 0; |
| 667 | reg->var_off = tnum_unknown; |
| 668 | reg->frameno = 0; |
| 669 | __mark_reg_unbounded(reg); |
| 670 | } |
| 671 | |
| 672 | static void mark_reg_unknown(struct bpf_verifier_env *env, |
| 673 | struct bpf_reg_state *regs, u32 regno) |
| 674 | { |
| 675 | if (WARN_ON(regno >= MAX_BPF_REG)) { |
| 676 | verbose(env, "mark_reg_unknown(regs, %u)\n", regno); |
| 677 | /* Something bad happened, let's kill all regs except FP */ |
| 678 | for (regno = 0; regno < BPF_REG_FP; regno++) |
| 679 | __mark_reg_not_init(regs + regno); |
| 680 | return; |
| 681 | } |
| 682 | __mark_reg_unknown(regs + regno); |
| 683 | } |
| 684 | |
| 685 | static void __mark_reg_not_init(struct bpf_reg_state *reg) |
| 686 | { |
| 687 | __mark_reg_unknown(reg); |
| 688 | reg->type = NOT_INIT; |
| 689 | } |
| 690 | |
| 691 | static void mark_reg_not_init(struct bpf_verifier_env *env, |
| 692 | struct bpf_reg_state *regs, u32 regno) |
| 693 | { |
| 694 | if (WARN_ON(regno >= MAX_BPF_REG)) { |
| 695 | verbose(env, "mark_reg_not_init(regs, %u)\n", regno); |
| 696 | /* Something bad happened, let's kill all regs except FP */ |
| 697 | for (regno = 0; regno < BPF_REG_FP; regno++) |
| 698 | __mark_reg_not_init(regs + regno); |
| 699 | return; |
| 700 | } |
| 701 | __mark_reg_not_init(regs + regno); |
| 702 | } |
| 703 | |
| 704 | static void init_reg_state(struct bpf_verifier_env *env, |
| 705 | struct bpf_func_state *state) |
| 706 | { |
| 707 | struct bpf_reg_state *regs = state->regs; |
| 708 | int i; |
| 709 | |
| 710 | for (i = 0; i < MAX_BPF_REG; i++) { |
| 711 | mark_reg_not_init(env, regs, i); |
| 712 | regs[i].live = REG_LIVE_NONE; |
| 713 | } |
| 714 | |
| 715 | /* frame pointer */ |
| 716 | regs[BPF_REG_FP].type = PTR_TO_STACK; |
| 717 | mark_reg_known_zero(env, regs, BPF_REG_FP); |
| 718 | regs[BPF_REG_FP].frameno = state->frameno; |
| 719 | |
| 720 | /* 1st arg to a function */ |
| 721 | regs[BPF_REG_1].type = PTR_TO_CTX; |
| 722 | mark_reg_known_zero(env, regs, BPF_REG_1); |
| 723 | } |
| 724 | |
| 725 | #define BPF_MAIN_FUNC (-1) |
| 726 | static void init_func_state(struct bpf_verifier_env *env, |
| 727 | struct bpf_func_state *state, |
| 728 | int callsite, int frameno, int subprogno) |
| 729 | { |
| 730 | state->callsite = callsite; |
| 731 | state->frameno = frameno; |
| 732 | state->subprogno = subprogno; |
| 733 | init_reg_state(env, state); |
| 734 | } |
| 735 | |
| 736 | enum reg_arg_type { |
| 737 | SRC_OP, /* register is used as source operand */ |
| 738 | DST_OP, /* register is used as destination operand */ |
| 739 | DST_OP_NO_MARK /* same as above, check only, don't mark */ |
| 740 | }; |
| 741 | |
| 742 | static int cmp_subprogs(const void *a, const void *b) |
| 743 | { |
| 744 | return ((struct bpf_subprog_info *)a)->start - |
| 745 | ((struct bpf_subprog_info *)b)->start; |
| 746 | } |
| 747 | |
| 748 | static int find_subprog(struct bpf_verifier_env *env, int off) |
| 749 | { |
| 750 | struct bpf_subprog_info *p; |
| 751 | |
| 752 | p = bsearch(&off, env->subprog_info, env->subprog_cnt, |
| 753 | sizeof(env->subprog_info[0]), cmp_subprogs); |
| 754 | if (!p) |
| 755 | return -ENOENT; |
| 756 | return p - env->subprog_info; |
| 757 | |
| 758 | } |
| 759 | |
| 760 | static int add_subprog(struct bpf_verifier_env *env, int off) |
| 761 | { |
| 762 | int insn_cnt = env->prog->len; |
| 763 | int ret; |
| 764 | |
| 765 | if (off >= insn_cnt || off < 0) { |
| 766 | verbose(env, "call to invalid destination\n"); |
| 767 | return -EINVAL; |
| 768 | } |
| 769 | ret = find_subprog(env, off); |
| 770 | if (ret >= 0) |
| 771 | return 0; |
| 772 | if (env->subprog_cnt >= BPF_MAX_SUBPROGS) { |
| 773 | verbose(env, "too many subprograms\n"); |
| 774 | return -E2BIG; |
| 775 | } |
| 776 | env->subprog_info[env->subprog_cnt++].start = off; |
| 777 | sort(env->subprog_info, env->subprog_cnt, |
| 778 | sizeof(env->subprog_info[0]), cmp_subprogs, NULL); |
| 779 | return 0; |
| 780 | } |
| 781 | |
| 782 | static int check_subprogs(struct bpf_verifier_env *env) |
| 783 | { |
| 784 | int i, ret, subprog_start, subprog_end, off, cur_subprog = 0; |
| 785 | struct bpf_subprog_info *subprog = env->subprog_info; |
| 786 | struct bpf_insn *insn = env->prog->insnsi; |
| 787 | int insn_cnt = env->prog->len; |
| 788 | |
| 789 | /* Add entry function. */ |
| 790 | ret = add_subprog(env, 0); |
| 791 | if (ret < 0) |
| 792 | return ret; |
| 793 | |
| 794 | /* determine subprog starts. The end is one before the next starts */ |
| 795 | for (i = 0; i < insn_cnt; i++) { |
| 796 | if (insn[i].code != (BPF_JMP | BPF_CALL)) |
| 797 | continue; |
| 798 | if (insn[i].src_reg != BPF_PSEUDO_CALL) |
| 799 | continue; |
| 800 | if (!env->allow_ptr_leaks) { |
| 801 | verbose(env, "function calls to other bpf functions are allowed for root only\n"); |
| 802 | return -EPERM; |
| 803 | } |
| 804 | if (bpf_prog_is_dev_bound(env->prog->aux)) { |
| 805 | verbose(env, "function calls in offloaded programs are not supported yet\n"); |
| 806 | return -EINVAL; |
| 807 | } |
| 808 | ret = add_subprog(env, i + insn[i].imm + 1); |
| 809 | if (ret < 0) |
| 810 | return ret; |
| 811 | } |
| 812 | |
| 813 | /* Add a fake 'exit' subprog which could simplify subprog iteration |
| 814 | * logic. 'subprog_cnt' should not be increased. |
| 815 | */ |
| 816 | subprog[env->subprog_cnt].start = insn_cnt; |
| 817 | |
| 818 | if (env->log.level > 1) |
| 819 | for (i = 0; i < env->subprog_cnt; i++) |
| 820 | verbose(env, "func#%d @%d\n", i, subprog[i].start); |
| 821 | |
| 822 | /* now check that all jumps are within the same subprog */ |
| 823 | subprog_start = subprog[cur_subprog].start; |
| 824 | subprog_end = subprog[cur_subprog + 1].start; |
| 825 | for (i = 0; i < insn_cnt; i++) { |
| 826 | u8 code = insn[i].code; |
| 827 | |
| 828 | if (BPF_CLASS(code) != BPF_JMP) |
| 829 | goto next; |
| 830 | if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL) |
| 831 | goto next; |
| 832 | off = i + insn[i].off + 1; |
| 833 | if (off < subprog_start || off >= subprog_end) { |
| 834 | verbose(env, "jump out of range from insn %d to %d\n", i, off); |
| 835 | return -EINVAL; |
| 836 | } |
| 837 | next: |
| 838 | if (i == subprog_end - 1) { |
| 839 | /* to avoid fall-through from one subprog into another |
| 840 | * the last insn of the subprog should be either exit |
| 841 | * or unconditional jump back |
| 842 | */ |
| 843 | if (code != (BPF_JMP | BPF_EXIT) && |
| 844 | code != (BPF_JMP | BPF_JA)) { |
| 845 | verbose(env, "last insn is not an exit or jmp\n"); |
| 846 | return -EINVAL; |
| 847 | } |
| 848 | subprog_start = subprog_end; |
| 849 | cur_subprog++; |
| 850 | if (cur_subprog < env->subprog_cnt) |
| 851 | subprog_end = subprog[cur_subprog + 1].start; |
| 852 | } |
| 853 | } |
| 854 | return 0; |
| 855 | } |
| 856 | |
| 857 | static |
| 858 | struct bpf_verifier_state *skip_callee(struct bpf_verifier_env *env, |
| 859 | const struct bpf_verifier_state *state, |
| 860 | struct bpf_verifier_state *parent, |
| 861 | u32 regno) |
| 862 | { |
| 863 | struct bpf_verifier_state *tmp = NULL; |
| 864 | |
| 865 | /* 'parent' could be a state of caller and |
| 866 | * 'state' could be a state of callee. In such case |
| 867 | * parent->curframe < state->curframe |
| 868 | * and it's ok for r1 - r5 registers |
| 869 | * |
| 870 | * 'parent' could be a callee's state after it bpf_exit-ed. |
| 871 | * In such case parent->curframe > state->curframe |
| 872 | * and it's ok for r0 only |
| 873 | */ |
| 874 | if (parent->curframe == state->curframe || |
| 875 | (parent->curframe < state->curframe && |
| 876 | regno >= BPF_REG_1 && regno <= BPF_REG_5) || |
| 877 | (parent->curframe > state->curframe && |
| 878 | regno == BPF_REG_0)) |
| 879 | return parent; |
| 880 | |
| 881 | if (parent->curframe > state->curframe && |
| 882 | regno >= BPF_REG_6) { |
| 883 | /* for callee saved regs we have to skip the whole chain |
| 884 | * of states that belong to callee and mark as LIVE_READ |
| 885 | * the registers before the call |
| 886 | */ |
| 887 | tmp = parent; |
| 888 | while (tmp && tmp->curframe != state->curframe) { |
| 889 | tmp = tmp->parent; |
| 890 | } |
| 891 | if (!tmp) |
| 892 | goto bug; |
| 893 | parent = tmp; |
| 894 | } else { |
| 895 | goto bug; |
| 896 | } |
| 897 | return parent; |
| 898 | bug: |
| 899 | verbose(env, "verifier bug regno %d tmp %p\n", regno, tmp); |
| 900 | verbose(env, "regno %d parent frame %d current frame %d\n", |
| 901 | regno, parent->curframe, state->curframe); |
| 902 | return NULL; |
| 903 | } |
| 904 | |
| 905 | static int mark_reg_read(struct bpf_verifier_env *env, |
| 906 | const struct bpf_verifier_state *state, |
| 907 | struct bpf_verifier_state *parent, |
| 908 | u32 regno) |
| 909 | { |
| 910 | bool writes = parent == state->parent; /* Observe write marks */ |
| 911 | |
| 912 | if (regno == BPF_REG_FP) |
| 913 | /* We don't need to worry about FP liveness because it's read-only */ |
| 914 | return 0; |
| 915 | |
| 916 | while (parent) { |
| 917 | /* if read wasn't screened by an earlier write ... */ |
| 918 | if (writes && state->frame[state->curframe]->regs[regno].live & REG_LIVE_WRITTEN) |
| 919 | break; |
| 920 | parent = skip_callee(env, state, parent, regno); |
| 921 | if (!parent) |
| 922 | return -EFAULT; |
| 923 | /* ... then we depend on parent's value */ |
| 924 | parent->frame[parent->curframe]->regs[regno].live |= REG_LIVE_READ; |
| 925 | state = parent; |
| 926 | parent = state->parent; |
| 927 | writes = true; |
| 928 | } |
| 929 | return 0; |
| 930 | } |
| 931 | |
| 932 | static int check_reg_arg(struct bpf_verifier_env *env, u32 regno, |
| 933 | enum reg_arg_type t) |
| 934 | { |
| 935 | struct bpf_verifier_state *vstate = env->cur_state; |
| 936 | struct bpf_func_state *state = vstate->frame[vstate->curframe]; |
| 937 | struct bpf_reg_state *regs = state->regs; |
| 938 | |
| 939 | if (regno >= MAX_BPF_REG) { |
| 940 | verbose(env, "R%d is invalid\n", regno); |
| 941 | return -EINVAL; |
| 942 | } |
| 943 | |
| 944 | if (t == SRC_OP) { |
| 945 | /* check whether register used as source operand can be read */ |
| 946 | if (regs[regno].type == NOT_INIT) { |
| 947 | verbose(env, "R%d !read_ok\n", regno); |
| 948 | return -EACCES; |
| 949 | } |
| 950 | return mark_reg_read(env, vstate, vstate->parent, regno); |
| 951 | } else { |
| 952 | /* check whether register used as dest operand can be written to */ |
| 953 | if (regno == BPF_REG_FP) { |
| 954 | verbose(env, "frame pointer is read only\n"); |
| 955 | return -EACCES; |
| 956 | } |
| 957 | regs[regno].live |= REG_LIVE_WRITTEN; |
| 958 | if (t == DST_OP) |
| 959 | mark_reg_unknown(env, regs, regno); |
| 960 | } |
| 961 | return 0; |
| 962 | } |
| 963 | |
| 964 | static bool is_spillable_regtype(enum bpf_reg_type type) |
| 965 | { |
| 966 | switch (type) { |
| 967 | case PTR_TO_MAP_VALUE: |
| 968 | case PTR_TO_MAP_VALUE_OR_NULL: |
| 969 | case PTR_TO_STACK: |
| 970 | case PTR_TO_CTX: |
| 971 | case PTR_TO_PACKET: |
| 972 | case PTR_TO_PACKET_META: |
| 973 | case PTR_TO_PACKET_END: |
| 974 | case CONST_PTR_TO_MAP: |
| 975 | return true; |
| 976 | default: |
| 977 | return false; |
| 978 | } |
| 979 | } |
| 980 | |
| 981 | /* Does this register contain a constant zero? */ |
| 982 | static bool register_is_null(struct bpf_reg_state *reg) |
| 983 | { |
| 984 | return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0); |
| 985 | } |
| 986 | |
| 987 | /* check_stack_read/write functions track spill/fill of registers, |
| 988 | * stack boundary and alignment are checked in check_mem_access() |
| 989 | */ |
| 990 | static int check_stack_write(struct bpf_verifier_env *env, |
| 991 | struct bpf_func_state *state, /* func where register points to */ |
| 992 | int off, int size, int value_regno) |
| 993 | { |
| 994 | struct bpf_func_state *cur; /* state of the current function */ |
| 995 | int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err; |
| 996 | enum bpf_reg_type type; |
| 997 | |
| 998 | err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE), |
| 999 | true); |
| 1000 | if (err) |
| 1001 | return err; |
| 1002 | /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0, |
| 1003 | * so it's aligned access and [off, off + size) are within stack limits |
| 1004 | */ |
| 1005 | if (!env->allow_ptr_leaks && |
| 1006 | state->stack[spi].slot_type[0] == STACK_SPILL && |
| 1007 | size != BPF_REG_SIZE) { |
| 1008 | verbose(env, "attempt to corrupt spilled pointer on stack\n"); |
| 1009 | return -EACCES; |
| 1010 | } |
| 1011 | |
| 1012 | cur = env->cur_state->frame[env->cur_state->curframe]; |
| 1013 | if (value_regno >= 0 && |
| 1014 | is_spillable_regtype((type = cur->regs[value_regno].type))) { |
| 1015 | |
| 1016 | /* register containing pointer is being spilled into stack */ |
| 1017 | if (size != BPF_REG_SIZE) { |
| 1018 | verbose(env, "invalid size of register spill\n"); |
| 1019 | return -EACCES; |
| 1020 | } |
| 1021 | |
| 1022 | if (state != cur && type == PTR_TO_STACK) { |
| 1023 | verbose(env, "cannot spill pointers to stack into stack frame of the caller\n"); |
| 1024 | return -EINVAL; |
| 1025 | } |
| 1026 | |
| 1027 | /* save register state */ |
| 1028 | state->stack[spi].spilled_ptr = cur->regs[value_regno]; |
| 1029 | state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN; |
| 1030 | |
| 1031 | for (i = 0; i < BPF_REG_SIZE; i++) |
| 1032 | state->stack[spi].slot_type[i] = STACK_SPILL; |
| 1033 | } else { |
| 1034 | u8 type = STACK_MISC; |
| 1035 | |
| 1036 | /* regular write of data into stack */ |
| 1037 | state->stack[spi].spilled_ptr = (struct bpf_reg_state) {}; |
| 1038 | |
| 1039 | /* only mark the slot as written if all 8 bytes were written |
| 1040 | * otherwise read propagation may incorrectly stop too soon |
| 1041 | * when stack slots are partially written. |
| 1042 | * This heuristic means that read propagation will be |
| 1043 | * conservative, since it will add reg_live_read marks |
| 1044 | * to stack slots all the way to first state when programs |
| 1045 | * writes+reads less than 8 bytes |
| 1046 | */ |
| 1047 | if (size == BPF_REG_SIZE) |
| 1048 | state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN; |
| 1049 | |
| 1050 | /* when we zero initialize stack slots mark them as such */ |
| 1051 | if (value_regno >= 0 && |
| 1052 | register_is_null(&cur->regs[value_regno])) |
| 1053 | type = STACK_ZERO; |
| 1054 | |
| 1055 | for (i = 0; i < size; i++) |
| 1056 | state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] = |
| 1057 | type; |
| 1058 | } |
| 1059 | return 0; |
| 1060 | } |
| 1061 | |
| 1062 | /* registers of every function are unique and mark_reg_read() propagates |
| 1063 | * the liveness in the following cases: |
| 1064 | * - from callee into caller for R1 - R5 that were used as arguments |
| 1065 | * - from caller into callee for R0 that used as result of the call |
| 1066 | * - from caller to the same caller skipping states of the callee for R6 - R9, |
| 1067 | * since R6 - R9 are callee saved by implicit function prologue and |
| 1068 | * caller's R6 != callee's R6, so when we propagate liveness up to |
| 1069 | * parent states we need to skip callee states for R6 - R9. |
| 1070 | * |
| 1071 | * stack slot marking is different, since stacks of caller and callee are |
| 1072 | * accessible in both (since caller can pass a pointer to caller's stack to |
| 1073 | * callee which can pass it to another function), hence mark_stack_slot_read() |
| 1074 | * has to propagate the stack liveness to all parent states at given frame number. |
| 1075 | * Consider code: |
| 1076 | * f1() { |
| 1077 | * ptr = fp - 8; |
| 1078 | * *ptr = ctx; |
| 1079 | * call f2 { |
| 1080 | * .. = *ptr; |
| 1081 | * } |
| 1082 | * .. = *ptr; |
| 1083 | * } |
| 1084 | * First *ptr is reading from f1's stack and mark_stack_slot_read() has |
| 1085 | * to mark liveness at the f1's frame and not f2's frame. |
| 1086 | * Second *ptr is also reading from f1's stack and mark_stack_slot_read() has |
| 1087 | * to propagate liveness to f2 states at f1's frame level and further into |
| 1088 | * f1 states at f1's frame level until write into that stack slot |
| 1089 | */ |
| 1090 | static void mark_stack_slot_read(struct bpf_verifier_env *env, |
| 1091 | const struct bpf_verifier_state *state, |
| 1092 | struct bpf_verifier_state *parent, |
| 1093 | int slot, int frameno) |
| 1094 | { |
| 1095 | bool writes = parent == state->parent; /* Observe write marks */ |
| 1096 | |
| 1097 | while (parent) { |
| 1098 | if (parent->frame[frameno]->allocated_stack <= slot * BPF_REG_SIZE) |
| 1099 | /* since LIVE_WRITTEN mark is only done for full 8-byte |
| 1100 | * write the read marks are conservative and parent |
| 1101 | * state may not even have the stack allocated. In such case |
| 1102 | * end the propagation, since the loop reached beginning |
| 1103 | * of the function |
| 1104 | */ |
| 1105 | break; |
| 1106 | /* if read wasn't screened by an earlier write ... */ |
| 1107 | if (writes && state->frame[frameno]->stack[slot].spilled_ptr.live & REG_LIVE_WRITTEN) |
| 1108 | break; |
| 1109 | /* ... then we depend on parent's value */ |
| 1110 | parent->frame[frameno]->stack[slot].spilled_ptr.live |= REG_LIVE_READ; |
| 1111 | state = parent; |
| 1112 | parent = state->parent; |
| 1113 | writes = true; |
| 1114 | } |
| 1115 | } |
| 1116 | |
| 1117 | static int check_stack_read(struct bpf_verifier_env *env, |
| 1118 | struct bpf_func_state *reg_state /* func where register points to */, |
| 1119 | int off, int size, int value_regno) |
| 1120 | { |
| 1121 | struct bpf_verifier_state *vstate = env->cur_state; |
| 1122 | struct bpf_func_state *state = vstate->frame[vstate->curframe]; |
| 1123 | int i, slot = -off - 1, spi = slot / BPF_REG_SIZE; |
| 1124 | u8 *stype; |
| 1125 | |
| 1126 | if (reg_state->allocated_stack <= slot) { |
| 1127 | verbose(env, "invalid read from stack off %d+0 size %d\n", |
| 1128 | off, size); |
| 1129 | return -EACCES; |
| 1130 | } |
| 1131 | stype = reg_state->stack[spi].slot_type; |
| 1132 | |
| 1133 | if (stype[0] == STACK_SPILL) { |
| 1134 | if (size != BPF_REG_SIZE) { |
| 1135 | verbose(env, "invalid size of register spill\n"); |
| 1136 | return -EACCES; |
| 1137 | } |
| 1138 | for (i = 1; i < BPF_REG_SIZE; i++) { |
| 1139 | if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) { |
| 1140 | verbose(env, "corrupted spill memory\n"); |
| 1141 | return -EACCES; |
| 1142 | } |
| 1143 | } |
| 1144 | |
| 1145 | if (value_regno >= 0) { |
| 1146 | /* restore register state from stack */ |
| 1147 | state->regs[value_regno] = reg_state->stack[spi].spilled_ptr; |
| 1148 | /* mark reg as written since spilled pointer state likely |
| 1149 | * has its liveness marks cleared by is_state_visited() |
| 1150 | * which resets stack/reg liveness for state transitions |
| 1151 | */ |
| 1152 | state->regs[value_regno].live |= REG_LIVE_WRITTEN; |
| 1153 | } |
| 1154 | mark_stack_slot_read(env, vstate, vstate->parent, spi, |
| 1155 | reg_state->frameno); |
| 1156 | return 0; |
| 1157 | } else { |
| 1158 | int zeros = 0; |
| 1159 | |
| 1160 | for (i = 0; i < size; i++) { |
| 1161 | if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC) |
| 1162 | continue; |
| 1163 | if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) { |
| 1164 | zeros++; |
| 1165 | continue; |
| 1166 | } |
| 1167 | verbose(env, "invalid read from stack off %d+%d size %d\n", |
| 1168 | off, i, size); |
| 1169 | return -EACCES; |
| 1170 | } |
| 1171 | mark_stack_slot_read(env, vstate, vstate->parent, spi, |
| 1172 | reg_state->frameno); |
| 1173 | if (value_regno >= 0) { |
| 1174 | if (zeros == size) { |
| 1175 | /* any size read into register is zero extended, |
| 1176 | * so the whole register == const_zero |
| 1177 | */ |
| 1178 | __mark_reg_const_zero(&state->regs[value_regno]); |
| 1179 | } else { |
| 1180 | /* have read misc data from the stack */ |
| 1181 | mark_reg_unknown(env, state->regs, value_regno); |
| 1182 | } |
| 1183 | state->regs[value_regno].live |= REG_LIVE_WRITTEN; |
| 1184 | } |
| 1185 | return 0; |
| 1186 | } |
| 1187 | } |
| 1188 | |
| 1189 | /* check read/write into map element returned by bpf_map_lookup_elem() */ |
| 1190 | static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off, |
| 1191 | int size, bool zero_size_allowed) |
| 1192 | { |
| 1193 | struct bpf_reg_state *regs = cur_regs(env); |
| 1194 | struct bpf_map *map = regs[regno].map_ptr; |
| 1195 | |
| 1196 | if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) || |
| 1197 | off + size > map->value_size) { |
| 1198 | verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n", |
| 1199 | map->value_size, off, size); |
| 1200 | return -EACCES; |
| 1201 | } |
| 1202 | return 0; |
| 1203 | } |
| 1204 | |
| 1205 | /* check read/write into a map element with possible variable offset */ |
| 1206 | static int check_map_access(struct bpf_verifier_env *env, u32 regno, |
| 1207 | int off, int size, bool zero_size_allowed) |
| 1208 | { |
| 1209 | struct bpf_verifier_state *vstate = env->cur_state; |
| 1210 | struct bpf_func_state *state = vstate->frame[vstate->curframe]; |
| 1211 | struct bpf_reg_state *reg = &state->regs[regno]; |
| 1212 | int err; |
| 1213 | |
| 1214 | /* We may have adjusted the register to this map value, so we |
| 1215 | * need to try adding each of min_value and max_value to off |
| 1216 | * to make sure our theoretical access will be safe. |
| 1217 | */ |
| 1218 | if (env->log.level) |
| 1219 | print_verifier_state(env, state); |
| 1220 | /* The minimum value is only important with signed |
| 1221 | * comparisons where we can't assume the floor of a |
| 1222 | * value is 0. If we are using signed variables for our |
| 1223 | * index'es we need to make sure that whatever we use |
| 1224 | * will have a set floor within our range. |
| 1225 | */ |
| 1226 | if (reg->smin_value < 0) { |
| 1227 | verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n", |
| 1228 | regno); |
| 1229 | return -EACCES; |
| 1230 | } |
| 1231 | err = __check_map_access(env, regno, reg->smin_value + off, size, |
| 1232 | zero_size_allowed); |
| 1233 | if (err) { |
| 1234 | verbose(env, "R%d min value is outside of the array range\n", |
| 1235 | regno); |
| 1236 | return err; |
| 1237 | } |
| 1238 | |
| 1239 | /* If we haven't set a max value then we need to bail since we can't be |
| 1240 | * sure we won't do bad things. |
| 1241 | * If reg->umax_value + off could overflow, treat that as unbounded too. |
| 1242 | */ |
| 1243 | if (reg->umax_value >= BPF_MAX_VAR_OFF) { |
| 1244 | verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n", |
| 1245 | regno); |
| 1246 | return -EACCES; |
| 1247 | } |
| 1248 | err = __check_map_access(env, regno, reg->umax_value + off, size, |
| 1249 | zero_size_allowed); |
| 1250 | if (err) |
| 1251 | verbose(env, "R%d max value is outside of the array range\n", |
| 1252 | regno); |
| 1253 | return err; |
| 1254 | } |
| 1255 | |
| 1256 | #define MAX_PACKET_OFF 0xffff |
| 1257 | |
| 1258 | static bool may_access_direct_pkt_data(struct bpf_verifier_env *env, |
| 1259 | const struct bpf_call_arg_meta *meta, |
| 1260 | enum bpf_access_type t) |
| 1261 | { |
| 1262 | switch (env->prog->type) { |
| 1263 | case BPF_PROG_TYPE_LWT_IN: |
| 1264 | case BPF_PROG_TYPE_LWT_OUT: |
| 1265 | /* dst_input() and dst_output() can't write for now */ |
| 1266 | if (t == BPF_WRITE) |
| 1267 | return false; |
| 1268 | /* fallthrough */ |
| 1269 | case BPF_PROG_TYPE_SCHED_CLS: |
| 1270 | case BPF_PROG_TYPE_SCHED_ACT: |
| 1271 | case BPF_PROG_TYPE_XDP: |
| 1272 | case BPF_PROG_TYPE_LWT_XMIT: |
| 1273 | case BPF_PROG_TYPE_SK_SKB: |
| 1274 | case BPF_PROG_TYPE_SK_MSG: |
| 1275 | if (meta) |
| 1276 | return meta->pkt_access; |
| 1277 | |
| 1278 | env->seen_direct_write = true; |
| 1279 | return true; |
| 1280 | default: |
| 1281 | return false; |
| 1282 | } |
| 1283 | } |
| 1284 | |
| 1285 | static int __check_packet_access(struct bpf_verifier_env *env, u32 regno, |
| 1286 | int off, int size, bool zero_size_allowed) |
| 1287 | { |
| 1288 | struct bpf_reg_state *regs = cur_regs(env); |
| 1289 | struct bpf_reg_state *reg = ®s[regno]; |
| 1290 | |
| 1291 | if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) || |
| 1292 | (u64)off + size > reg->range) { |
| 1293 | verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n", |
| 1294 | off, size, regno, reg->id, reg->off, reg->range); |
| 1295 | return -EACCES; |
| 1296 | } |
| 1297 | return 0; |
| 1298 | } |
| 1299 | |
| 1300 | static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off, |
| 1301 | int size, bool zero_size_allowed) |
| 1302 | { |
| 1303 | struct bpf_reg_state *regs = cur_regs(env); |
| 1304 | struct bpf_reg_state *reg = ®s[regno]; |
| 1305 | int err; |
| 1306 | |
| 1307 | /* We may have added a variable offset to the packet pointer; but any |
| 1308 | * reg->range we have comes after that. We are only checking the fixed |
| 1309 | * offset. |
| 1310 | */ |
| 1311 | |
| 1312 | /* We don't allow negative numbers, because we aren't tracking enough |
| 1313 | * detail to prove they're safe. |
| 1314 | */ |
| 1315 | if (reg->smin_value < 0) { |
| 1316 | verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n", |
| 1317 | regno); |
| 1318 | return -EACCES; |
| 1319 | } |
| 1320 | err = __check_packet_access(env, regno, off, size, zero_size_allowed); |
| 1321 | if (err) { |
| 1322 | verbose(env, "R%d offset is outside of the packet\n", regno); |
| 1323 | return err; |
| 1324 | } |
| 1325 | return err; |
| 1326 | } |
| 1327 | |
| 1328 | /* check access to 'struct bpf_context' fields. Supports fixed offsets only */ |
| 1329 | static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size, |
| 1330 | enum bpf_access_type t, enum bpf_reg_type *reg_type) |
| 1331 | { |
| 1332 | struct bpf_insn_access_aux info = { |
| 1333 | .reg_type = *reg_type, |
| 1334 | }; |
| 1335 | |
| 1336 | if (env->ops->is_valid_access && |
| 1337 | env->ops->is_valid_access(off, size, t, env->prog, &info)) { |
| 1338 | /* A non zero info.ctx_field_size indicates that this field is a |
| 1339 | * candidate for later verifier transformation to load the whole |
| 1340 | * field and then apply a mask when accessed with a narrower |
| 1341 | * access than actual ctx access size. A zero info.ctx_field_size |
| 1342 | * will only allow for whole field access and rejects any other |
| 1343 | * type of narrower access. |
| 1344 | */ |
| 1345 | *reg_type = info.reg_type; |
| 1346 | |
| 1347 | env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size; |
| 1348 | /* remember the offset of last byte accessed in ctx */ |
| 1349 | if (env->prog->aux->max_ctx_offset < off + size) |
| 1350 | env->prog->aux->max_ctx_offset = off + size; |
| 1351 | return 0; |
| 1352 | } |
| 1353 | |
| 1354 | verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size); |
| 1355 | return -EACCES; |
| 1356 | } |
| 1357 | |
| 1358 | static bool __is_pointer_value(bool allow_ptr_leaks, |
| 1359 | const struct bpf_reg_state *reg) |
| 1360 | { |
| 1361 | if (allow_ptr_leaks) |
| 1362 | return false; |
| 1363 | |
| 1364 | return reg->type != SCALAR_VALUE; |
| 1365 | } |
| 1366 | |
| 1367 | static bool is_pointer_value(struct bpf_verifier_env *env, int regno) |
| 1368 | { |
| 1369 | return __is_pointer_value(env->allow_ptr_leaks, cur_regs(env) + regno); |
| 1370 | } |
| 1371 | |
| 1372 | static bool is_ctx_reg(struct bpf_verifier_env *env, int regno) |
| 1373 | { |
| 1374 | const struct bpf_reg_state *reg = cur_regs(env) + regno; |
| 1375 | |
| 1376 | return reg->type == PTR_TO_CTX; |
| 1377 | } |
| 1378 | |
| 1379 | static bool is_pkt_reg(struct bpf_verifier_env *env, int regno) |
| 1380 | { |
| 1381 | const struct bpf_reg_state *reg = cur_regs(env) + regno; |
| 1382 | |
| 1383 | return type_is_pkt_pointer(reg->type); |
| 1384 | } |
| 1385 | |
| 1386 | static int check_pkt_ptr_alignment(struct bpf_verifier_env *env, |
| 1387 | const struct bpf_reg_state *reg, |
| 1388 | int off, int size, bool strict) |
| 1389 | { |
| 1390 | struct tnum reg_off; |
| 1391 | int ip_align; |
| 1392 | |
| 1393 | /* Byte size accesses are always allowed. */ |
| 1394 | if (!strict || size == 1) |
| 1395 | return 0; |
| 1396 | |
| 1397 | /* For platforms that do not have a Kconfig enabling |
| 1398 | * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of |
| 1399 | * NET_IP_ALIGN is universally set to '2'. And on platforms |
| 1400 | * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get |
| 1401 | * to this code only in strict mode where we want to emulate |
| 1402 | * the NET_IP_ALIGN==2 checking. Therefore use an |
| 1403 | * unconditional IP align value of '2'. |
| 1404 | */ |
| 1405 | ip_align = 2; |
| 1406 | |
| 1407 | reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off)); |
| 1408 | if (!tnum_is_aligned(reg_off, size)) { |
| 1409 | char tn_buf[48]; |
| 1410 | |
| 1411 | tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off); |
| 1412 | verbose(env, |
| 1413 | "misaligned packet access off %d+%s+%d+%d size %d\n", |
| 1414 | ip_align, tn_buf, reg->off, off, size); |
| 1415 | return -EACCES; |
| 1416 | } |
| 1417 | |
| 1418 | return 0; |
| 1419 | } |
| 1420 | |
| 1421 | static int check_generic_ptr_alignment(struct bpf_verifier_env *env, |
| 1422 | const struct bpf_reg_state *reg, |
| 1423 | const char *pointer_desc, |
| 1424 | int off, int size, bool strict) |
| 1425 | { |
| 1426 | struct tnum reg_off; |
| 1427 | |
| 1428 | /* Byte size accesses are always allowed. */ |
| 1429 | if (!strict || size == 1) |
| 1430 | return 0; |
| 1431 | |
| 1432 | reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off)); |
| 1433 | if (!tnum_is_aligned(reg_off, size)) { |
| 1434 | char tn_buf[48]; |
| 1435 | |
| 1436 | tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off); |
| 1437 | verbose(env, "misaligned %saccess off %s+%d+%d size %d\n", |
| 1438 | pointer_desc, tn_buf, reg->off, off, size); |
| 1439 | return -EACCES; |
| 1440 | } |
| 1441 | |
| 1442 | return 0; |
| 1443 | } |
| 1444 | |
| 1445 | static int check_ptr_alignment(struct bpf_verifier_env *env, |
| 1446 | const struct bpf_reg_state *reg, int off, |
| 1447 | int size, bool strict_alignment_once) |
| 1448 | { |
| 1449 | bool strict = env->strict_alignment || strict_alignment_once; |
| 1450 | const char *pointer_desc = ""; |
| 1451 | |
| 1452 | switch (reg->type) { |
| 1453 | case PTR_TO_PACKET: |
| 1454 | case PTR_TO_PACKET_META: |
| 1455 | /* Special case, because of NET_IP_ALIGN. Given metadata sits |
| 1456 | * right in front, treat it the very same way. |
| 1457 | */ |
| 1458 | return check_pkt_ptr_alignment(env, reg, off, size, strict); |
| 1459 | case PTR_TO_MAP_VALUE: |
| 1460 | pointer_desc = "value "; |
| 1461 | break; |
| 1462 | case PTR_TO_CTX: |
| 1463 | pointer_desc = "context "; |
| 1464 | break; |
| 1465 | case PTR_TO_STACK: |
| 1466 | pointer_desc = "stack "; |
| 1467 | /* The stack spill tracking logic in check_stack_write() |
| 1468 | * and check_stack_read() relies on stack accesses being |
| 1469 | * aligned. |
| 1470 | */ |
| 1471 | strict = true; |
| 1472 | break; |
| 1473 | default: |
| 1474 | break; |
| 1475 | } |
| 1476 | return check_generic_ptr_alignment(env, reg, pointer_desc, off, size, |
| 1477 | strict); |
| 1478 | } |
| 1479 | |
| 1480 | static int update_stack_depth(struct bpf_verifier_env *env, |
| 1481 | const struct bpf_func_state *func, |
| 1482 | int off) |
| 1483 | { |
| 1484 | u16 stack = env->subprog_info[func->subprogno].stack_depth; |
| 1485 | |
| 1486 | if (stack >= -off) |
| 1487 | return 0; |
| 1488 | |
| 1489 | /* update known max for given subprogram */ |
| 1490 | env->subprog_info[func->subprogno].stack_depth = -off; |
| 1491 | return 0; |
| 1492 | } |
| 1493 | |
| 1494 | /* starting from main bpf function walk all instructions of the function |
| 1495 | * and recursively walk all callees that given function can call. |
| 1496 | * Ignore jump and exit insns. |
| 1497 | * Since recursion is prevented by check_cfg() this algorithm |
| 1498 | * only needs a local stack of MAX_CALL_FRAMES to remember callsites |
| 1499 | */ |
| 1500 | static int check_max_stack_depth(struct bpf_verifier_env *env) |
| 1501 | { |
| 1502 | int depth = 0, frame = 0, idx = 0, i = 0, subprog_end; |
| 1503 | struct bpf_subprog_info *subprog = env->subprog_info; |
| 1504 | struct bpf_insn *insn = env->prog->insnsi; |
| 1505 | int ret_insn[MAX_CALL_FRAMES]; |
| 1506 | int ret_prog[MAX_CALL_FRAMES]; |
| 1507 | |
| 1508 | process_func: |
| 1509 | /* round up to 32-bytes, since this is granularity |
| 1510 | * of interpreter stack size |
| 1511 | */ |
| 1512 | depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32); |
| 1513 | if (depth > MAX_BPF_STACK) { |
| 1514 | verbose(env, "combined stack size of %d calls is %d. Too large\n", |
| 1515 | frame + 1, depth); |
| 1516 | return -EACCES; |
| 1517 | } |
| 1518 | continue_func: |
| 1519 | subprog_end = subprog[idx + 1].start; |
| 1520 | for (; i < subprog_end; i++) { |
| 1521 | if (insn[i].code != (BPF_JMP | BPF_CALL)) |
| 1522 | continue; |
| 1523 | if (insn[i].src_reg != BPF_PSEUDO_CALL) |
| 1524 | continue; |
| 1525 | /* remember insn and function to return to */ |
| 1526 | ret_insn[frame] = i + 1; |
| 1527 | ret_prog[frame] = idx; |
| 1528 | |
| 1529 | /* find the callee */ |
| 1530 | i = i + insn[i].imm + 1; |
| 1531 | idx = find_subprog(env, i); |
| 1532 | if (idx < 0) { |
| 1533 | WARN_ONCE(1, "verifier bug. No program starts at insn %d\n", |
| 1534 | i); |
| 1535 | return -EFAULT; |
| 1536 | } |
| 1537 | frame++; |
| 1538 | if (frame >= MAX_CALL_FRAMES) { |
| 1539 | WARN_ONCE(1, "verifier bug. Call stack is too deep\n"); |
| 1540 | return -EFAULT; |
| 1541 | } |
| 1542 | goto process_func; |
| 1543 | } |
| 1544 | /* end of for() loop means the last insn of the 'subprog' |
| 1545 | * was reached. Doesn't matter whether it was JA or EXIT |
| 1546 | */ |
| 1547 | if (frame == 0) |
| 1548 | return 0; |
| 1549 | depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32); |
| 1550 | frame--; |
| 1551 | i = ret_insn[frame]; |
| 1552 | idx = ret_prog[frame]; |
| 1553 | goto continue_func; |
| 1554 | } |
| 1555 | |
| 1556 | #ifndef CONFIG_BPF_JIT_ALWAYS_ON |
| 1557 | static int get_callee_stack_depth(struct bpf_verifier_env *env, |
| 1558 | const struct bpf_insn *insn, int idx) |
| 1559 | { |
| 1560 | int start = idx + insn->imm + 1, subprog; |
| 1561 | |
| 1562 | subprog = find_subprog(env, start); |
| 1563 | if (subprog < 0) { |
| 1564 | WARN_ONCE(1, "verifier bug. No program starts at insn %d\n", |
| 1565 | start); |
| 1566 | return -EFAULT; |
| 1567 | } |
| 1568 | return env->subprog_info[subprog].stack_depth; |
| 1569 | } |
| 1570 | #endif |
| 1571 | |
| 1572 | /* truncate register to smaller size (in bytes) |
| 1573 | * must be called with size < BPF_REG_SIZE |
| 1574 | */ |
| 1575 | static void coerce_reg_to_size(struct bpf_reg_state *reg, int size) |
| 1576 | { |
| 1577 | u64 mask; |
| 1578 | |
| 1579 | /* clear high bits in bit representation */ |
| 1580 | reg->var_off = tnum_cast(reg->var_off, size); |
| 1581 | |
| 1582 | /* fix arithmetic bounds */ |
| 1583 | mask = ((u64)1 << (size * 8)) - 1; |
| 1584 | if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) { |
| 1585 | reg->umin_value &= mask; |
| 1586 | reg->umax_value &= mask; |
| 1587 | } else { |
| 1588 | reg->umin_value = 0; |
| 1589 | reg->umax_value = mask; |
| 1590 | } |
| 1591 | reg->smin_value = reg->umin_value; |
| 1592 | reg->smax_value = reg->umax_value; |
| 1593 | } |
| 1594 | |
| 1595 | /* check whether memory at (regno + off) is accessible for t = (read | write) |
| 1596 | * if t==write, value_regno is a register which value is stored into memory |
| 1597 | * if t==read, value_regno is a register which will receive the value from memory |
| 1598 | * if t==write && value_regno==-1, some unknown value is stored into memory |
| 1599 | * if t==read && value_regno==-1, don't care what we read from memory |
| 1600 | */ |
| 1601 | static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno, |
| 1602 | int off, int bpf_size, enum bpf_access_type t, |
| 1603 | int value_regno, bool strict_alignment_once) |
| 1604 | { |
| 1605 | struct bpf_reg_state *regs = cur_regs(env); |
| 1606 | struct bpf_reg_state *reg = regs + regno; |
| 1607 | struct bpf_func_state *state; |
| 1608 | int size, err = 0; |
| 1609 | |
| 1610 | size = bpf_size_to_bytes(bpf_size); |
| 1611 | if (size < 0) |
| 1612 | return size; |
| 1613 | |
| 1614 | /* alignment checks will add in reg->off themselves */ |
| 1615 | err = check_ptr_alignment(env, reg, off, size, strict_alignment_once); |
| 1616 | if (err) |
| 1617 | return err; |
| 1618 | |
| 1619 | /* for access checks, reg->off is just part of off */ |
| 1620 | off += reg->off; |
| 1621 | |
| 1622 | if (reg->type == PTR_TO_MAP_VALUE) { |
| 1623 | if (t == BPF_WRITE && value_regno >= 0 && |
| 1624 | is_pointer_value(env, value_regno)) { |
| 1625 | verbose(env, "R%d leaks addr into map\n", value_regno); |
| 1626 | return -EACCES; |
| 1627 | } |
| 1628 | |
| 1629 | err = check_map_access(env, regno, off, size, false); |
| 1630 | if (!err && t == BPF_READ && value_regno >= 0) |
| 1631 | mark_reg_unknown(env, regs, value_regno); |
| 1632 | |
| 1633 | } else if (reg->type == PTR_TO_CTX) { |
| 1634 | enum bpf_reg_type reg_type = SCALAR_VALUE; |
| 1635 | |
| 1636 | if (t == BPF_WRITE && value_regno >= 0 && |
| 1637 | is_pointer_value(env, value_regno)) { |
| 1638 | verbose(env, "R%d leaks addr into ctx\n", value_regno); |
| 1639 | return -EACCES; |
| 1640 | } |
| 1641 | /* ctx accesses must be at a fixed offset, so that we can |
| 1642 | * determine what type of data were returned. |
| 1643 | */ |
| 1644 | if (reg->off) { |
| 1645 | verbose(env, |
| 1646 | "dereference of modified ctx ptr R%d off=%d+%d, ctx+const is allowed, ctx+const+const is not\n", |
| 1647 | regno, reg->off, off - reg->off); |
| 1648 | return -EACCES; |
| 1649 | } |
| 1650 | if (!tnum_is_const(reg->var_off) || reg->var_off.value) { |
| 1651 | char tn_buf[48]; |
| 1652 | |
| 1653 | tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off); |
| 1654 | verbose(env, |
| 1655 | "variable ctx access var_off=%s off=%d size=%d", |
| 1656 | tn_buf, off, size); |
| 1657 | return -EACCES; |
| 1658 | } |
| 1659 | err = check_ctx_access(env, insn_idx, off, size, t, ®_type); |
| 1660 | if (!err && t == BPF_READ && value_regno >= 0) { |
| 1661 | /* ctx access returns either a scalar, or a |
| 1662 | * PTR_TO_PACKET[_META,_END]. In the latter |
| 1663 | * case, we know the offset is zero. |
| 1664 | */ |
| 1665 | if (reg_type == SCALAR_VALUE) |
| 1666 | mark_reg_unknown(env, regs, value_regno); |
| 1667 | else |
| 1668 | mark_reg_known_zero(env, regs, |
| 1669 | value_regno); |
| 1670 | regs[value_regno].id = 0; |
| 1671 | regs[value_regno].off = 0; |
| 1672 | regs[value_regno].range = 0; |
| 1673 | regs[value_regno].type = reg_type; |
| 1674 | } |
| 1675 | |
| 1676 | } else if (reg->type == PTR_TO_STACK) { |
| 1677 | /* stack accesses must be at a fixed offset, so that we can |
| 1678 | * determine what type of data were returned. |
| 1679 | * See check_stack_read(). |
| 1680 | */ |
| 1681 | if (!tnum_is_const(reg->var_off)) { |
| 1682 | char tn_buf[48]; |
| 1683 | |
| 1684 | tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off); |
| 1685 | verbose(env, "variable stack access var_off=%s off=%d size=%d", |
| 1686 | tn_buf, off, size); |
| 1687 | return -EACCES; |
| 1688 | } |
| 1689 | off += reg->var_off.value; |
| 1690 | if (off >= 0 || off < -MAX_BPF_STACK) { |
| 1691 | verbose(env, "invalid stack off=%d size=%d\n", off, |
| 1692 | size); |
| 1693 | return -EACCES; |
| 1694 | } |
| 1695 | |
| 1696 | state = func(env, reg); |
| 1697 | err = update_stack_depth(env, state, off); |
| 1698 | if (err) |
| 1699 | return err; |
| 1700 | |
| 1701 | if (t == BPF_WRITE) |
| 1702 | err = check_stack_write(env, state, off, size, |
| 1703 | value_regno); |
| 1704 | else |
| 1705 | err = check_stack_read(env, state, off, size, |
| 1706 | value_regno); |
| 1707 | } else if (reg_is_pkt_pointer(reg)) { |
| 1708 | if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) { |
| 1709 | verbose(env, "cannot write into packet\n"); |
| 1710 | return -EACCES; |
| 1711 | } |
| 1712 | if (t == BPF_WRITE && value_regno >= 0 && |
| 1713 | is_pointer_value(env, value_regno)) { |
| 1714 | verbose(env, "R%d leaks addr into packet\n", |
| 1715 | value_regno); |
| 1716 | return -EACCES; |
| 1717 | } |
| 1718 | err = check_packet_access(env, regno, off, size, false); |
| 1719 | if (!err && t == BPF_READ && value_regno >= 0) |
| 1720 | mark_reg_unknown(env, regs, value_regno); |
| 1721 | } else { |
| 1722 | verbose(env, "R%d invalid mem access '%s'\n", regno, |
| 1723 | reg_type_str[reg->type]); |
| 1724 | return -EACCES; |
| 1725 | } |
| 1726 | |
| 1727 | if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ && |
| 1728 | regs[value_regno].type == SCALAR_VALUE) { |
| 1729 | /* b/h/w load zero-extends, mark upper bits as known 0 */ |
| 1730 | coerce_reg_to_size(®s[value_regno], size); |
| 1731 | } |
| 1732 | return err; |
| 1733 | } |
| 1734 | |
| 1735 | static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn) |
| 1736 | { |
| 1737 | int err; |
| 1738 | |
| 1739 | if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) || |
| 1740 | insn->imm != 0) { |
| 1741 | verbose(env, "BPF_XADD uses reserved fields\n"); |
| 1742 | return -EINVAL; |
| 1743 | } |
| 1744 | |
| 1745 | /* check src1 operand */ |
| 1746 | err = check_reg_arg(env, insn->src_reg, SRC_OP); |
| 1747 | if (err) |
| 1748 | return err; |
| 1749 | |
| 1750 | /* check src2 operand */ |
| 1751 | err = check_reg_arg(env, insn->dst_reg, SRC_OP); |
| 1752 | if (err) |
| 1753 | return err; |
| 1754 | |
| 1755 | if (is_pointer_value(env, insn->src_reg)) { |
| 1756 | verbose(env, "R%d leaks addr into mem\n", insn->src_reg); |
| 1757 | return -EACCES; |
| 1758 | } |
| 1759 | |
| 1760 | if (is_ctx_reg(env, insn->dst_reg) || |
| 1761 | is_pkt_reg(env, insn->dst_reg)) { |
| 1762 | verbose(env, "BPF_XADD stores into R%d %s is not allowed\n", |
| 1763 | insn->dst_reg, is_ctx_reg(env, insn->dst_reg) ? |
| 1764 | "context" : "packet"); |
| 1765 | return -EACCES; |
| 1766 | } |
| 1767 | |
| 1768 | /* check whether atomic_add can read the memory */ |
| 1769 | err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off, |
| 1770 | BPF_SIZE(insn->code), BPF_READ, -1, true); |
| 1771 | if (err) |
| 1772 | return err; |
| 1773 | |
| 1774 | /* check whether atomic_add can write into the same memory */ |
| 1775 | return check_mem_access(env, insn_idx, insn->dst_reg, insn->off, |
| 1776 | BPF_SIZE(insn->code), BPF_WRITE, -1, true); |
| 1777 | } |
| 1778 | |
| 1779 | /* when register 'regno' is passed into function that will read 'access_size' |
| 1780 | * bytes from that pointer, make sure that it's within stack boundary |
| 1781 | * and all elements of stack are initialized. |
| 1782 | * Unlike most pointer bounds-checking functions, this one doesn't take an |
| 1783 | * 'off' argument, so it has to add in reg->off itself. |
| 1784 | */ |
| 1785 | static int check_stack_boundary(struct bpf_verifier_env *env, int regno, |
| 1786 | int access_size, bool zero_size_allowed, |
| 1787 | struct bpf_call_arg_meta *meta) |
| 1788 | { |
| 1789 | struct bpf_reg_state *reg = cur_regs(env) + regno; |
| 1790 | struct bpf_func_state *state = func(env, reg); |
| 1791 | int off, i, slot, spi; |
| 1792 | |
| 1793 | if (reg->type != PTR_TO_STACK) { |
| 1794 | /* Allow zero-byte read from NULL, regardless of pointer type */ |
| 1795 | if (zero_size_allowed && access_size == 0 && |
| 1796 | register_is_null(reg)) |
| 1797 | return 0; |
| 1798 | |
| 1799 | verbose(env, "R%d type=%s expected=%s\n", regno, |
| 1800 | reg_type_str[reg->type], |
| 1801 | reg_type_str[PTR_TO_STACK]); |
| 1802 | return -EACCES; |
| 1803 | } |
| 1804 | |
| 1805 | /* Only allow fixed-offset stack reads */ |
| 1806 | if (!tnum_is_const(reg->var_off)) { |
| 1807 | char tn_buf[48]; |
| 1808 | |
| 1809 | tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off); |
| 1810 | verbose(env, "invalid variable stack read R%d var_off=%s\n", |
| 1811 | regno, tn_buf); |
| 1812 | return -EACCES; |
| 1813 | } |
| 1814 | off = reg->off + reg->var_off.value; |
| 1815 | if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 || |
| 1816 | access_size < 0 || (access_size == 0 && !zero_size_allowed)) { |
| 1817 | verbose(env, "invalid stack type R%d off=%d access_size=%d\n", |
| 1818 | regno, off, access_size); |
| 1819 | return -EACCES; |
| 1820 | } |
| 1821 | |
| 1822 | if (meta && meta->raw_mode) { |
| 1823 | meta->access_size = access_size; |
| 1824 | meta->regno = regno; |
| 1825 | return 0; |
| 1826 | } |
| 1827 | |
| 1828 | for (i = 0; i < access_size; i++) { |
| 1829 | u8 *stype; |
| 1830 | |
| 1831 | slot = -(off + i) - 1; |
| 1832 | spi = slot / BPF_REG_SIZE; |
| 1833 | if (state->allocated_stack <= slot) |
| 1834 | goto err; |
| 1835 | stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE]; |
| 1836 | if (*stype == STACK_MISC) |
| 1837 | goto mark; |
| 1838 | if (*stype == STACK_ZERO) { |
| 1839 | /* helper can write anything into the stack */ |
| 1840 | *stype = STACK_MISC; |
| 1841 | goto mark; |
| 1842 | } |
| 1843 | err: |
| 1844 | verbose(env, "invalid indirect read from stack off %d+%d size %d\n", |
| 1845 | off, i, access_size); |
| 1846 | return -EACCES; |
| 1847 | mark: |
| 1848 | /* reading any byte out of 8-byte 'spill_slot' will cause |
| 1849 | * the whole slot to be marked as 'read' |
| 1850 | */ |
| 1851 | mark_stack_slot_read(env, env->cur_state, env->cur_state->parent, |
| 1852 | spi, state->frameno); |
| 1853 | } |
| 1854 | return update_stack_depth(env, state, off); |
| 1855 | } |
| 1856 | |
| 1857 | static int check_helper_mem_access(struct bpf_verifier_env *env, int regno, |
| 1858 | int access_size, bool zero_size_allowed, |
| 1859 | struct bpf_call_arg_meta *meta) |
| 1860 | { |
| 1861 | struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno]; |
| 1862 | |
| 1863 | switch (reg->type) { |
| 1864 | case PTR_TO_PACKET: |
| 1865 | case PTR_TO_PACKET_META: |
| 1866 | return check_packet_access(env, regno, reg->off, access_size, |
| 1867 | zero_size_allowed); |
| 1868 | case PTR_TO_MAP_VALUE: |
| 1869 | return check_map_access(env, regno, reg->off, access_size, |
| 1870 | zero_size_allowed); |
| 1871 | default: /* scalar_value|ptr_to_stack or invalid ptr */ |
| 1872 | return check_stack_boundary(env, regno, access_size, |
| 1873 | zero_size_allowed, meta); |
| 1874 | } |
| 1875 | } |
| 1876 | |
| 1877 | static bool arg_type_is_mem_ptr(enum bpf_arg_type type) |
| 1878 | { |
| 1879 | return type == ARG_PTR_TO_MEM || |
| 1880 | type == ARG_PTR_TO_MEM_OR_NULL || |
| 1881 | type == ARG_PTR_TO_UNINIT_MEM; |
| 1882 | } |
| 1883 | |
| 1884 | static bool arg_type_is_mem_size(enum bpf_arg_type type) |
| 1885 | { |
| 1886 | return type == ARG_CONST_SIZE || |
| 1887 | type == ARG_CONST_SIZE_OR_ZERO; |
| 1888 | } |
| 1889 | |
| 1890 | static int check_func_arg(struct bpf_verifier_env *env, u32 regno, |
| 1891 | enum bpf_arg_type arg_type, |
| 1892 | struct bpf_call_arg_meta *meta) |
| 1893 | { |
| 1894 | struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno]; |
| 1895 | enum bpf_reg_type expected_type, type = reg->type; |
| 1896 | int err = 0; |
| 1897 | |
| 1898 | if (arg_type == ARG_DONTCARE) |
| 1899 | return 0; |
| 1900 | |
| 1901 | err = check_reg_arg(env, regno, SRC_OP); |
| 1902 | if (err) |
| 1903 | return err; |
| 1904 | |
| 1905 | if (arg_type == ARG_ANYTHING) { |
| 1906 | if (is_pointer_value(env, regno)) { |
| 1907 | verbose(env, "R%d leaks addr into helper function\n", |
| 1908 | regno); |
| 1909 | return -EACCES; |
| 1910 | } |
| 1911 | return 0; |
| 1912 | } |
| 1913 | |
| 1914 | if (type_is_pkt_pointer(type) && |
| 1915 | !may_access_direct_pkt_data(env, meta, BPF_READ)) { |
| 1916 | verbose(env, "helper access to the packet is not allowed\n"); |
| 1917 | return -EACCES; |
| 1918 | } |
| 1919 | |
| 1920 | if (arg_type == ARG_PTR_TO_MAP_KEY || |
| 1921 | arg_type == ARG_PTR_TO_MAP_VALUE) { |
| 1922 | expected_type = PTR_TO_STACK; |
| 1923 | if (!type_is_pkt_pointer(type) && type != PTR_TO_MAP_VALUE && |
| 1924 | type != expected_type) |
| 1925 | goto err_type; |
| 1926 | } else if (arg_type == ARG_CONST_SIZE || |
| 1927 | arg_type == ARG_CONST_SIZE_OR_ZERO) { |
| 1928 | expected_type = SCALAR_VALUE; |
| 1929 | if (type != expected_type) |
| 1930 | goto err_type; |
| 1931 | } else if (arg_type == ARG_CONST_MAP_PTR) { |
| 1932 | expected_type = CONST_PTR_TO_MAP; |
| 1933 | if (type != expected_type) |
| 1934 | goto err_type; |
| 1935 | } else if (arg_type == ARG_PTR_TO_CTX) { |
| 1936 | expected_type = PTR_TO_CTX; |
| 1937 | if (type != expected_type) |
| 1938 | goto err_type; |
| 1939 | } else if (arg_type_is_mem_ptr(arg_type)) { |
| 1940 | expected_type = PTR_TO_STACK; |
| 1941 | /* One exception here. In case function allows for NULL to be |
| 1942 | * passed in as argument, it's a SCALAR_VALUE type. Final test |
| 1943 | * happens during stack boundary checking. |
| 1944 | */ |
| 1945 | if (register_is_null(reg) && |
| 1946 | arg_type == ARG_PTR_TO_MEM_OR_NULL) |
| 1947 | /* final test in check_stack_boundary() */; |
| 1948 | else if (!type_is_pkt_pointer(type) && |
| 1949 | type != PTR_TO_MAP_VALUE && |
| 1950 | type != expected_type) |
| 1951 | goto err_type; |
| 1952 | meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM; |
| 1953 | } else { |
| 1954 | verbose(env, "unsupported arg_type %d\n", arg_type); |
| 1955 | return -EFAULT; |
| 1956 | } |
| 1957 | |
| 1958 | if (arg_type == ARG_CONST_MAP_PTR) { |
| 1959 | /* bpf_map_xxx(map_ptr) call: remember that map_ptr */ |
| 1960 | meta->map_ptr = reg->map_ptr; |
| 1961 | } else if (arg_type == ARG_PTR_TO_MAP_KEY) { |
| 1962 | /* bpf_map_xxx(..., map_ptr, ..., key) call: |
| 1963 | * check that [key, key + map->key_size) are within |
| 1964 | * stack limits and initialized |
| 1965 | */ |
| 1966 | if (!meta->map_ptr) { |
| 1967 | /* in function declaration map_ptr must come before |
| 1968 | * map_key, so that it's verified and known before |
| 1969 | * we have to check map_key here. Otherwise it means |
| 1970 | * that kernel subsystem misconfigured verifier |
| 1971 | */ |
| 1972 | verbose(env, "invalid map_ptr to access map->key\n"); |
| 1973 | return -EACCES; |
| 1974 | } |
| 1975 | err = check_helper_mem_access(env, regno, |
| 1976 | meta->map_ptr->key_size, false, |
| 1977 | NULL); |
| 1978 | } else if (arg_type == ARG_PTR_TO_MAP_VALUE) { |
| 1979 | /* bpf_map_xxx(..., map_ptr, ..., value) call: |
| 1980 | * check [value, value + map->value_size) validity |
| 1981 | */ |
| 1982 | if (!meta->map_ptr) { |
| 1983 | /* kernel subsystem misconfigured verifier */ |
| 1984 | verbose(env, "invalid map_ptr to access map->value\n"); |
| 1985 | return -EACCES; |
| 1986 | } |
| 1987 | err = check_helper_mem_access(env, regno, |
| 1988 | meta->map_ptr->value_size, false, |
| 1989 | NULL); |
| 1990 | } else if (arg_type_is_mem_size(arg_type)) { |
| 1991 | bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO); |
| 1992 | |
| 1993 | /* remember the mem_size which may be used later |
| 1994 | * to refine return values. |
| 1995 | */ |
| 1996 | meta->msize_smax_value = reg->smax_value; |
| 1997 | meta->msize_umax_value = reg->umax_value; |
| 1998 | |
| 1999 | /* The register is SCALAR_VALUE; the access check |
| 2000 | * happens using its boundaries. |
| 2001 | */ |
| 2002 | if (!tnum_is_const(reg->var_off)) |
| 2003 | /* For unprivileged variable accesses, disable raw |
| 2004 | * mode so that the program is required to |
| 2005 | * initialize all the memory that the helper could |
| 2006 | * just partially fill up. |
| 2007 | */ |
| 2008 | meta = NULL; |
| 2009 | |
| 2010 | if (reg->smin_value < 0) { |
| 2011 | verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n", |
| 2012 | regno); |
| 2013 | return -EACCES; |
| 2014 | } |
| 2015 | |
| 2016 | if (reg->umin_value == 0) { |
| 2017 | err = check_helper_mem_access(env, regno - 1, 0, |
| 2018 | zero_size_allowed, |
| 2019 | meta); |
| 2020 | if (err) |
| 2021 | return err; |
| 2022 | } |
| 2023 | |
| 2024 | if (reg->umax_value >= BPF_MAX_VAR_SIZ) { |
| 2025 | verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n", |
| 2026 | regno); |
| 2027 | return -EACCES; |
| 2028 | } |
| 2029 | err = check_helper_mem_access(env, regno - 1, |
| 2030 | reg->umax_value, |
| 2031 | zero_size_allowed, meta); |
| 2032 | } |
| 2033 | |
| 2034 | return err; |
| 2035 | err_type: |
| 2036 | verbose(env, "R%d type=%s expected=%s\n", regno, |
| 2037 | reg_type_str[type], reg_type_str[expected_type]); |
| 2038 | return -EACCES; |
| 2039 | } |
| 2040 | |
| 2041 | static int check_map_func_compatibility(struct bpf_verifier_env *env, |
| 2042 | struct bpf_map *map, int func_id) |
| 2043 | { |
| 2044 | if (!map) |
| 2045 | return 0; |
| 2046 | |
| 2047 | /* We need a two way check, first is from map perspective ... */ |
| 2048 | switch (map->map_type) { |
| 2049 | case BPF_MAP_TYPE_PROG_ARRAY: |
| 2050 | if (func_id != BPF_FUNC_tail_call) |
| 2051 | goto error; |
| 2052 | break; |
| 2053 | case BPF_MAP_TYPE_PERF_EVENT_ARRAY: |
| 2054 | if (func_id != BPF_FUNC_perf_event_read && |
| 2055 | func_id != BPF_FUNC_perf_event_output && |
| 2056 | func_id != BPF_FUNC_perf_event_read_value) |
| 2057 | goto error; |
| 2058 | break; |
| 2059 | case BPF_MAP_TYPE_STACK_TRACE: |
| 2060 | if (func_id != BPF_FUNC_get_stackid) |
| 2061 | goto error; |
| 2062 | break; |
| 2063 | case BPF_MAP_TYPE_CGROUP_ARRAY: |
| 2064 | if (func_id != BPF_FUNC_skb_under_cgroup && |
| 2065 | func_id != BPF_FUNC_current_task_under_cgroup) |
| 2066 | goto error; |
| 2067 | break; |
| 2068 | /* devmap returns a pointer to a live net_device ifindex that we cannot |
| 2069 | * allow to be modified from bpf side. So do not allow lookup elements |
| 2070 | * for now. |
| 2071 | */ |
| 2072 | case BPF_MAP_TYPE_DEVMAP: |
| 2073 | if (func_id != BPF_FUNC_redirect_map) |
| 2074 | goto error; |
| 2075 | break; |
| 2076 | /* Restrict bpf side of cpumap and xskmap, open when use-cases |
| 2077 | * appear. |
| 2078 | */ |
| 2079 | case BPF_MAP_TYPE_CPUMAP: |
| 2080 | case BPF_MAP_TYPE_XSKMAP: |
| 2081 | if (func_id != BPF_FUNC_redirect_map) |
| 2082 | goto error; |
| 2083 | break; |
| 2084 | case BPF_MAP_TYPE_ARRAY_OF_MAPS: |
| 2085 | case BPF_MAP_TYPE_HASH_OF_MAPS: |
| 2086 | if (func_id != BPF_FUNC_map_lookup_elem) |
| 2087 | goto error; |
| 2088 | break; |
| 2089 | case BPF_MAP_TYPE_SOCKMAP: |
| 2090 | if (func_id != BPF_FUNC_sk_redirect_map && |
| 2091 | func_id != BPF_FUNC_sock_map_update && |
| 2092 | func_id != BPF_FUNC_map_delete_elem && |
| 2093 | func_id != BPF_FUNC_msg_redirect_map) |
| 2094 | goto error; |
| 2095 | break; |
| 2096 | default: |
| 2097 | break; |
| 2098 | } |
| 2099 | |
| 2100 | /* ... and second from the function itself. */ |
| 2101 | switch (func_id) { |
| 2102 | case BPF_FUNC_tail_call: |
| 2103 | if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY) |
| 2104 | goto error; |
| 2105 | if (env->subprog_cnt > 1) { |
| 2106 | verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n"); |
| 2107 | return -EINVAL; |
| 2108 | } |
| 2109 | break; |
| 2110 | case BPF_FUNC_perf_event_read: |
| 2111 | case BPF_FUNC_perf_event_output: |
| 2112 | case BPF_FUNC_perf_event_read_value: |
| 2113 | if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY) |
| 2114 | goto error; |
| 2115 | break; |
| 2116 | case BPF_FUNC_get_stackid: |
| 2117 | if (map->map_type != BPF_MAP_TYPE_STACK_TRACE) |
| 2118 | goto error; |
| 2119 | break; |
| 2120 | case BPF_FUNC_current_task_under_cgroup: |
| 2121 | case BPF_FUNC_skb_under_cgroup: |
| 2122 | if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY) |
| 2123 | goto error; |
| 2124 | break; |
| 2125 | case BPF_FUNC_redirect_map: |
| 2126 | if (map->map_type != BPF_MAP_TYPE_DEVMAP && |
| 2127 | map->map_type != BPF_MAP_TYPE_CPUMAP && |
| 2128 | map->map_type != BPF_MAP_TYPE_XSKMAP) |
| 2129 | goto error; |
| 2130 | break; |
| 2131 | case BPF_FUNC_sk_redirect_map: |
| 2132 | case BPF_FUNC_msg_redirect_map: |
| 2133 | if (map->map_type != BPF_MAP_TYPE_SOCKMAP) |
| 2134 | goto error; |
| 2135 | break; |
| 2136 | case BPF_FUNC_sock_map_update: |
| 2137 | if (map->map_type != BPF_MAP_TYPE_SOCKMAP) |
| 2138 | goto error; |
| 2139 | break; |
| 2140 | default: |
| 2141 | break; |
| 2142 | } |
| 2143 | |
| 2144 | return 0; |
| 2145 | error: |
| 2146 | verbose(env, "cannot pass map_type %d into func %s#%d\n", |
| 2147 | map->map_type, func_id_name(func_id), func_id); |
| 2148 | return -EINVAL; |
| 2149 | } |
| 2150 | |
| 2151 | static bool check_raw_mode_ok(const struct bpf_func_proto *fn) |
| 2152 | { |
| 2153 | int count = 0; |
| 2154 | |
| 2155 | if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM) |
| 2156 | count++; |
| 2157 | if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM) |
| 2158 | count++; |
| 2159 | if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM) |
| 2160 | count++; |
| 2161 | if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM) |
| 2162 | count++; |
| 2163 | if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM) |
| 2164 | count++; |
| 2165 | |
| 2166 | /* We only support one arg being in raw mode at the moment, |
| 2167 | * which is sufficient for the helper functions we have |
| 2168 | * right now. |
| 2169 | */ |
| 2170 | return count <= 1; |
| 2171 | } |
| 2172 | |
| 2173 | static bool check_args_pair_invalid(enum bpf_arg_type arg_curr, |
| 2174 | enum bpf_arg_type arg_next) |
| 2175 | { |
| 2176 | return (arg_type_is_mem_ptr(arg_curr) && |
| 2177 | !arg_type_is_mem_size(arg_next)) || |
| 2178 | (!arg_type_is_mem_ptr(arg_curr) && |
| 2179 | arg_type_is_mem_size(arg_next)); |
| 2180 | } |
| 2181 | |
| 2182 | static bool check_arg_pair_ok(const struct bpf_func_proto *fn) |
| 2183 | { |
| 2184 | /* bpf_xxx(..., buf, len) call will access 'len' |
| 2185 | * bytes from memory 'buf'. Both arg types need |
| 2186 | * to be paired, so make sure there's no buggy |
| 2187 | * helper function specification. |
| 2188 | */ |
| 2189 | if (arg_type_is_mem_size(fn->arg1_type) || |
| 2190 | arg_type_is_mem_ptr(fn->arg5_type) || |
| 2191 | check_args_pair_invalid(fn->arg1_type, fn->arg2_type) || |
| 2192 | check_args_pair_invalid(fn->arg2_type, fn->arg3_type) || |
| 2193 | check_args_pair_invalid(fn->arg3_type, fn->arg4_type) || |
| 2194 | check_args_pair_invalid(fn->arg4_type, fn->arg5_type)) |
| 2195 | return false; |
| 2196 | |
| 2197 | return true; |
| 2198 | } |
| 2199 | |
| 2200 | static int check_func_proto(const struct bpf_func_proto *fn) |
| 2201 | { |
| 2202 | return check_raw_mode_ok(fn) && |
| 2203 | check_arg_pair_ok(fn) ? 0 : -EINVAL; |
| 2204 | } |
| 2205 | |
| 2206 | /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END] |
| 2207 | * are now invalid, so turn them into unknown SCALAR_VALUE. |
| 2208 | */ |
| 2209 | static void __clear_all_pkt_pointers(struct bpf_verifier_env *env, |
| 2210 | struct bpf_func_state *state) |
| 2211 | { |
| 2212 | struct bpf_reg_state *regs = state->regs, *reg; |
| 2213 | int i; |
| 2214 | |
| 2215 | for (i = 0; i < MAX_BPF_REG; i++) |
| 2216 | if (reg_is_pkt_pointer_any(®s[i])) |
| 2217 | mark_reg_unknown(env, regs, i); |
| 2218 | |
| 2219 | for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) { |
| 2220 | if (state->stack[i].slot_type[0] != STACK_SPILL) |
| 2221 | continue; |
| 2222 | reg = &state->stack[i].spilled_ptr; |
| 2223 | if (reg_is_pkt_pointer_any(reg)) |
| 2224 | __mark_reg_unknown(reg); |
| 2225 | } |
| 2226 | } |
| 2227 | |
| 2228 | static void clear_all_pkt_pointers(struct bpf_verifier_env *env) |
| 2229 | { |
| 2230 | struct bpf_verifier_state *vstate = env->cur_state; |
| 2231 | int i; |
| 2232 | |
| 2233 | for (i = 0; i <= vstate->curframe; i++) |
| 2234 | __clear_all_pkt_pointers(env, vstate->frame[i]); |
| 2235 | } |
| 2236 | |
| 2237 | static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn, |
| 2238 | int *insn_idx) |
| 2239 | { |
| 2240 | struct bpf_verifier_state *state = env->cur_state; |
| 2241 | struct bpf_func_state *caller, *callee; |
| 2242 | int i, subprog, target_insn; |
| 2243 | |
| 2244 | if (state->curframe + 1 >= MAX_CALL_FRAMES) { |
| 2245 | verbose(env, "the call stack of %d frames is too deep\n", |
| 2246 | state->curframe + 2); |
| 2247 | return -E2BIG; |
| 2248 | } |
| 2249 | |
| 2250 | target_insn = *insn_idx + insn->imm; |
| 2251 | subprog = find_subprog(env, target_insn + 1); |
| 2252 | if (subprog < 0) { |
| 2253 | verbose(env, "verifier bug. No program starts at insn %d\n", |
| 2254 | target_insn + 1); |
| 2255 | return -EFAULT; |
| 2256 | } |
| 2257 | |
| 2258 | caller = state->frame[state->curframe]; |
| 2259 | if (state->frame[state->curframe + 1]) { |
| 2260 | verbose(env, "verifier bug. Frame %d already allocated\n", |
| 2261 | state->curframe + 1); |
| 2262 | return -EFAULT; |
| 2263 | } |
| 2264 | |
| 2265 | callee = kzalloc(sizeof(*callee), GFP_KERNEL); |
| 2266 | if (!callee) |
| 2267 | return -ENOMEM; |
| 2268 | state->frame[state->curframe + 1] = callee; |
| 2269 | |
| 2270 | /* callee cannot access r0, r6 - r9 for reading and has to write |
| 2271 | * into its own stack before reading from it. |
| 2272 | * callee can read/write into caller's stack |
| 2273 | */ |
| 2274 | init_func_state(env, callee, |
| 2275 | /* remember the callsite, it will be used by bpf_exit */ |
| 2276 | *insn_idx /* callsite */, |
| 2277 | state->curframe + 1 /* frameno within this callchain */, |
| 2278 | subprog /* subprog number within this prog */); |
| 2279 | |
| 2280 | /* copy r1 - r5 args that callee can access */ |
| 2281 | for (i = BPF_REG_1; i <= BPF_REG_5; i++) |
| 2282 | callee->regs[i] = caller->regs[i]; |
| 2283 | |
| 2284 | /* after the call regsiters r0 - r5 were scratched */ |
| 2285 | for (i = 0; i < CALLER_SAVED_REGS; i++) { |
| 2286 | mark_reg_not_init(env, caller->regs, caller_saved[i]); |
| 2287 | check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK); |
| 2288 | } |
| 2289 | |
| 2290 | /* only increment it after check_reg_arg() finished */ |
| 2291 | state->curframe++; |
| 2292 | |
| 2293 | /* and go analyze first insn of the callee */ |
| 2294 | *insn_idx = target_insn; |
| 2295 | |
| 2296 | if (env->log.level) { |
| 2297 | verbose(env, "caller:\n"); |
| 2298 | print_verifier_state(env, caller); |
| 2299 | verbose(env, "callee:\n"); |
| 2300 | print_verifier_state(env, callee); |
| 2301 | } |
| 2302 | return 0; |
| 2303 | } |
| 2304 | |
| 2305 | static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx) |
| 2306 | { |
| 2307 | struct bpf_verifier_state *state = env->cur_state; |
| 2308 | struct bpf_func_state *caller, *callee; |
| 2309 | struct bpf_reg_state *r0; |
| 2310 | |
| 2311 | callee = state->frame[state->curframe]; |
| 2312 | r0 = &callee->regs[BPF_REG_0]; |
| 2313 | if (r0->type == PTR_TO_STACK) { |
| 2314 | /* technically it's ok to return caller's stack pointer |
| 2315 | * (or caller's caller's pointer) back to the caller, |
| 2316 | * since these pointers are valid. Only current stack |
| 2317 | * pointer will be invalid as soon as function exits, |
| 2318 | * but let's be conservative |
| 2319 | */ |
| 2320 | verbose(env, "cannot return stack pointer to the caller\n"); |
| 2321 | return -EINVAL; |
| 2322 | } |
| 2323 | |
| 2324 | state->curframe--; |
| 2325 | caller = state->frame[state->curframe]; |
| 2326 | /* return to the caller whatever r0 had in the callee */ |
| 2327 | caller->regs[BPF_REG_0] = *r0; |
| 2328 | |
| 2329 | *insn_idx = callee->callsite + 1; |
| 2330 | if (env->log.level) { |
| 2331 | verbose(env, "returning from callee:\n"); |
| 2332 | print_verifier_state(env, callee); |
| 2333 | verbose(env, "to caller at %d:\n", *insn_idx); |
| 2334 | print_verifier_state(env, caller); |
| 2335 | } |
| 2336 | /* clear everything in the callee */ |
| 2337 | free_func_state(callee); |
| 2338 | state->frame[state->curframe + 1] = NULL; |
| 2339 | return 0; |
| 2340 | } |
| 2341 | |
| 2342 | static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type, |
| 2343 | int func_id, |
| 2344 | struct bpf_call_arg_meta *meta) |
| 2345 | { |
| 2346 | struct bpf_reg_state *ret_reg = ®s[BPF_REG_0]; |
| 2347 | |
| 2348 | if (ret_type != RET_INTEGER || |
| 2349 | (func_id != BPF_FUNC_get_stack && |
| 2350 | func_id != BPF_FUNC_probe_read_str)) |
| 2351 | return; |
| 2352 | |
| 2353 | ret_reg->smax_value = meta->msize_smax_value; |
| 2354 | ret_reg->umax_value = meta->msize_umax_value; |
| 2355 | __reg_deduce_bounds(ret_reg); |
| 2356 | __reg_bound_offset(ret_reg); |
| 2357 | } |
| 2358 | |
| 2359 | static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx) |
| 2360 | { |
| 2361 | const struct bpf_func_proto *fn = NULL; |
| 2362 | struct bpf_reg_state *regs; |
| 2363 | struct bpf_call_arg_meta meta; |
| 2364 | bool changes_data; |
| 2365 | int i, err; |
| 2366 | |
| 2367 | /* find function prototype */ |
| 2368 | if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) { |
| 2369 | verbose(env, "invalid func %s#%d\n", func_id_name(func_id), |
| 2370 | func_id); |
| 2371 | return -EINVAL; |
| 2372 | } |
| 2373 | |
| 2374 | if (env->ops->get_func_proto) |
| 2375 | fn = env->ops->get_func_proto(func_id, env->prog); |
| 2376 | if (!fn) { |
| 2377 | verbose(env, "unknown func %s#%d\n", func_id_name(func_id), |
| 2378 | func_id); |
| 2379 | return -EINVAL; |
| 2380 | } |
| 2381 | |
| 2382 | /* eBPF programs must be GPL compatible to use GPL-ed functions */ |
| 2383 | if (!env->prog->gpl_compatible && fn->gpl_only) { |
| 2384 | verbose(env, "cannot call GPL only function from proprietary program\n"); |
| 2385 | return -EINVAL; |
| 2386 | } |
| 2387 | |
| 2388 | /* With LD_ABS/IND some JITs save/restore skb from r1. */ |
| 2389 | changes_data = bpf_helper_changes_pkt_data(fn->func); |
| 2390 | if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) { |
| 2391 | verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n", |
| 2392 | func_id_name(func_id), func_id); |
| 2393 | return -EINVAL; |
| 2394 | } |
| 2395 | |
| 2396 | memset(&meta, 0, sizeof(meta)); |
| 2397 | meta.pkt_access = fn->pkt_access; |
| 2398 | |
| 2399 | err = check_func_proto(fn); |
| 2400 | if (err) { |
| 2401 | verbose(env, "kernel subsystem misconfigured func %s#%d\n", |
| 2402 | func_id_name(func_id), func_id); |
| 2403 | return err; |
| 2404 | } |
| 2405 | |
| 2406 | /* check args */ |
| 2407 | err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta); |
| 2408 | if (err) |
| 2409 | return err; |
| 2410 | err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta); |
| 2411 | if (err) |
| 2412 | return err; |
| 2413 | if (func_id == BPF_FUNC_tail_call) { |
| 2414 | if (meta.map_ptr == NULL) { |
| 2415 | verbose(env, "verifier bug\n"); |
| 2416 | return -EINVAL; |
| 2417 | } |
| 2418 | env->insn_aux_data[insn_idx].map_ptr = meta.map_ptr; |
| 2419 | } |
| 2420 | err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta); |
| 2421 | if (err) |
| 2422 | return err; |
| 2423 | err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta); |
| 2424 | if (err) |
| 2425 | return err; |
| 2426 | err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta); |
| 2427 | if (err) |
| 2428 | return err; |
| 2429 | |
| 2430 | /* Mark slots with STACK_MISC in case of raw mode, stack offset |
| 2431 | * is inferred from register state. |
| 2432 | */ |
| 2433 | for (i = 0; i < meta.access_size; i++) { |
| 2434 | err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B, |
| 2435 | BPF_WRITE, -1, false); |
| 2436 | if (err) |
| 2437 | return err; |
| 2438 | } |
| 2439 | |
| 2440 | regs = cur_regs(env); |
| 2441 | /* reset caller saved regs */ |
| 2442 | for (i = 0; i < CALLER_SAVED_REGS; i++) { |
| 2443 | mark_reg_not_init(env, regs, caller_saved[i]); |
| 2444 | check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK); |
| 2445 | } |
| 2446 | |
| 2447 | /* update return register (already marked as written above) */ |
| 2448 | if (fn->ret_type == RET_INTEGER) { |
| 2449 | /* sets type to SCALAR_VALUE */ |
| 2450 | mark_reg_unknown(env, regs, BPF_REG_0); |
| 2451 | } else if (fn->ret_type == RET_VOID) { |
| 2452 | regs[BPF_REG_0].type = NOT_INIT; |
| 2453 | } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL) { |
| 2454 | struct bpf_insn_aux_data *insn_aux; |
| 2455 | |
| 2456 | regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL; |
| 2457 | /* There is no offset yet applied, variable or fixed */ |
| 2458 | mark_reg_known_zero(env, regs, BPF_REG_0); |
| 2459 | regs[BPF_REG_0].off = 0; |
| 2460 | /* remember map_ptr, so that check_map_access() |
| 2461 | * can check 'value_size' boundary of memory access |
| 2462 | * to map element returned from bpf_map_lookup_elem() |
| 2463 | */ |
| 2464 | if (meta.map_ptr == NULL) { |
| 2465 | verbose(env, |
| 2466 | "kernel subsystem misconfigured verifier\n"); |
| 2467 | return -EINVAL; |
| 2468 | } |
| 2469 | regs[BPF_REG_0].map_ptr = meta.map_ptr; |
| 2470 | regs[BPF_REG_0].id = ++env->id_gen; |
| 2471 | insn_aux = &env->insn_aux_data[insn_idx]; |
| 2472 | if (!insn_aux->map_ptr) |
| 2473 | insn_aux->map_ptr = meta.map_ptr; |
| 2474 | else if (insn_aux->map_ptr != meta.map_ptr) |
| 2475 | insn_aux->map_ptr = BPF_MAP_PTR_POISON; |
| 2476 | } else { |
| 2477 | verbose(env, "unknown return type %d of func %s#%d\n", |
| 2478 | fn->ret_type, func_id_name(func_id), func_id); |
| 2479 | return -EINVAL; |
| 2480 | } |
| 2481 | |
| 2482 | do_refine_retval_range(regs, fn->ret_type, func_id, &meta); |
| 2483 | |
| 2484 | err = check_map_func_compatibility(env, meta.map_ptr, func_id); |
| 2485 | if (err) |
| 2486 | return err; |
| 2487 | |
| 2488 | if (func_id == BPF_FUNC_get_stack && !env->prog->has_callchain_buf) { |
| 2489 | const char *err_str; |
| 2490 | |
| 2491 | #ifdef CONFIG_PERF_EVENTS |
| 2492 | err = get_callchain_buffers(sysctl_perf_event_max_stack); |
| 2493 | err_str = "cannot get callchain buffer for func %s#%d\n"; |
| 2494 | #else |
| 2495 | err = -ENOTSUPP; |
| 2496 | err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n"; |
| 2497 | #endif |
| 2498 | if (err) { |
| 2499 | verbose(env, err_str, func_id_name(func_id), func_id); |
| 2500 | return err; |
| 2501 | } |
| 2502 | |
| 2503 | env->prog->has_callchain_buf = true; |
| 2504 | } |
| 2505 | |
| 2506 | if (changes_data) |
| 2507 | clear_all_pkt_pointers(env); |
| 2508 | return 0; |
| 2509 | } |
| 2510 | |
| 2511 | static bool signed_add_overflows(s64 a, s64 b) |
| 2512 | { |
| 2513 | /* Do the add in u64, where overflow is well-defined */ |
| 2514 | s64 res = (s64)((u64)a + (u64)b); |
| 2515 | |
| 2516 | if (b < 0) |
| 2517 | return res > a; |
| 2518 | return res < a; |
| 2519 | } |
| 2520 | |
| 2521 | static bool signed_sub_overflows(s64 a, s64 b) |
| 2522 | { |
| 2523 | /* Do the sub in u64, where overflow is well-defined */ |
| 2524 | s64 res = (s64)((u64)a - (u64)b); |
| 2525 | |
| 2526 | if (b < 0) |
| 2527 | return res < a; |
| 2528 | return res > a; |
| 2529 | } |
| 2530 | |
| 2531 | static bool check_reg_sane_offset(struct bpf_verifier_env *env, |
| 2532 | const struct bpf_reg_state *reg, |
| 2533 | enum bpf_reg_type type) |
| 2534 | { |
| 2535 | bool known = tnum_is_const(reg->var_off); |
| 2536 | s64 val = reg->var_off.value; |
| 2537 | s64 smin = reg->smin_value; |
| 2538 | |
| 2539 | if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) { |
| 2540 | verbose(env, "math between %s pointer and %lld is not allowed\n", |
| 2541 | reg_type_str[type], val); |
| 2542 | return false; |
| 2543 | } |
| 2544 | |
| 2545 | if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) { |
| 2546 | verbose(env, "%s pointer offset %d is not allowed\n", |
| 2547 | reg_type_str[type], reg->off); |
| 2548 | return false; |
| 2549 | } |
| 2550 | |
| 2551 | if (smin == S64_MIN) { |
| 2552 | verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n", |
| 2553 | reg_type_str[type]); |
| 2554 | return false; |
| 2555 | } |
| 2556 | |
| 2557 | if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) { |
| 2558 | verbose(env, "value %lld makes %s pointer be out of bounds\n", |
| 2559 | smin, reg_type_str[type]); |
| 2560 | return false; |
| 2561 | } |
| 2562 | |
| 2563 | return true; |
| 2564 | } |
| 2565 | |
| 2566 | /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off. |
| 2567 | * Caller should also handle BPF_MOV case separately. |
| 2568 | * If we return -EACCES, caller may want to try again treating pointer as a |
| 2569 | * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks. |
| 2570 | */ |
| 2571 | static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env, |
| 2572 | struct bpf_insn *insn, |
| 2573 | const struct bpf_reg_state *ptr_reg, |
| 2574 | const struct bpf_reg_state *off_reg) |
| 2575 | { |
| 2576 | struct bpf_verifier_state *vstate = env->cur_state; |
| 2577 | struct bpf_func_state *state = vstate->frame[vstate->curframe]; |
| 2578 | struct bpf_reg_state *regs = state->regs, *dst_reg; |
| 2579 | bool known = tnum_is_const(off_reg->var_off); |
| 2580 | s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value, |
| 2581 | smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value; |
| 2582 | u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value, |
| 2583 | umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value; |
| 2584 | u8 opcode = BPF_OP(insn->code); |
| 2585 | u32 dst = insn->dst_reg; |
| 2586 | |
| 2587 | dst_reg = ®s[dst]; |
| 2588 | |
| 2589 | if ((known && (smin_val != smax_val || umin_val != umax_val)) || |
| 2590 | smin_val > smax_val || umin_val > umax_val) { |
| 2591 | /* Taint dst register if offset had invalid bounds derived from |
| 2592 | * e.g. dead branches. |
| 2593 | */ |
| 2594 | __mark_reg_unknown(dst_reg); |
| 2595 | return 0; |
| 2596 | } |
| 2597 | |
| 2598 | if (BPF_CLASS(insn->code) != BPF_ALU64) { |
| 2599 | /* 32-bit ALU ops on pointers produce (meaningless) scalars */ |
| 2600 | verbose(env, |
| 2601 | "R%d 32-bit pointer arithmetic prohibited\n", |
| 2602 | dst); |
| 2603 | return -EACCES; |
| 2604 | } |
| 2605 | |
| 2606 | if (ptr_reg->type == PTR_TO_MAP_VALUE_OR_NULL) { |
| 2607 | verbose(env, "R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n", |
| 2608 | dst); |
| 2609 | return -EACCES; |
| 2610 | } |
| 2611 | if (ptr_reg->type == CONST_PTR_TO_MAP) { |
| 2612 | verbose(env, "R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n", |
| 2613 | dst); |
| 2614 | return -EACCES; |
| 2615 | } |
| 2616 | if (ptr_reg->type == PTR_TO_PACKET_END) { |
| 2617 | verbose(env, "R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n", |
| 2618 | dst); |
| 2619 | return -EACCES; |
| 2620 | } |
| 2621 | |
| 2622 | /* In case of 'scalar += pointer', dst_reg inherits pointer type and id. |
| 2623 | * The id may be overwritten later if we create a new variable offset. |
| 2624 | */ |
| 2625 | dst_reg->type = ptr_reg->type; |
| 2626 | dst_reg->id = ptr_reg->id; |
| 2627 | |
| 2628 | if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) || |
| 2629 | !check_reg_sane_offset(env, ptr_reg, ptr_reg->type)) |
| 2630 | return -EINVAL; |
| 2631 | |
| 2632 | switch (opcode) { |
| 2633 | case BPF_ADD: |
| 2634 | /* We can take a fixed offset as long as it doesn't overflow |
| 2635 | * the s32 'off' field |
| 2636 | */ |
| 2637 | if (known && (ptr_reg->off + smin_val == |
| 2638 | (s64)(s32)(ptr_reg->off + smin_val))) { |
| 2639 | /* pointer += K. Accumulate it into fixed offset */ |
| 2640 | dst_reg->smin_value = smin_ptr; |
| 2641 | dst_reg->smax_value = smax_ptr; |
| 2642 | dst_reg->umin_value = umin_ptr; |
| 2643 | dst_reg->umax_value = umax_ptr; |
| 2644 | dst_reg->var_off = ptr_reg->var_off; |
| 2645 | dst_reg->off = ptr_reg->off + smin_val; |
| 2646 | dst_reg->range = ptr_reg->range; |
| 2647 | break; |
| 2648 | } |
| 2649 | /* A new variable offset is created. Note that off_reg->off |
| 2650 | * == 0, since it's a scalar. |
| 2651 | * dst_reg gets the pointer type and since some positive |
| 2652 | * integer value was added to the pointer, give it a new 'id' |
| 2653 | * if it's a PTR_TO_PACKET. |
| 2654 | * this creates a new 'base' pointer, off_reg (variable) gets |
| 2655 | * added into the variable offset, and we copy the fixed offset |
| 2656 | * from ptr_reg. |
| 2657 | */ |
| 2658 | if (signed_add_overflows(smin_ptr, smin_val) || |
| 2659 | signed_add_overflows(smax_ptr, smax_val)) { |
| 2660 | dst_reg->smin_value = S64_MIN; |
| 2661 | dst_reg->smax_value = S64_MAX; |
| 2662 | } else { |
| 2663 | dst_reg->smin_value = smin_ptr + smin_val; |
| 2664 | dst_reg->smax_value = smax_ptr + smax_val; |
| 2665 | } |
| 2666 | if (umin_ptr + umin_val < umin_ptr || |
| 2667 | umax_ptr + umax_val < umax_ptr) { |
| 2668 | dst_reg->umin_value = 0; |
| 2669 | dst_reg->umax_value = U64_MAX; |
| 2670 | } else { |
| 2671 | dst_reg->umin_value = umin_ptr + umin_val; |
| 2672 | dst_reg->umax_value = umax_ptr + umax_val; |
| 2673 | } |
| 2674 | dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off); |
| 2675 | dst_reg->off = ptr_reg->off; |
| 2676 | if (reg_is_pkt_pointer(ptr_reg)) { |
| 2677 | dst_reg->id = ++env->id_gen; |
| 2678 | /* something was added to pkt_ptr, set range to zero */ |
| 2679 | dst_reg->range = 0; |
| 2680 | } |
| 2681 | break; |
| 2682 | case BPF_SUB: |
| 2683 | if (dst_reg == off_reg) { |
| 2684 | /* scalar -= pointer. Creates an unknown scalar */ |
| 2685 | verbose(env, "R%d tried to subtract pointer from scalar\n", |
| 2686 | dst); |
| 2687 | return -EACCES; |
| 2688 | } |
| 2689 | /* We don't allow subtraction from FP, because (according to |
| 2690 | * test_verifier.c test "invalid fp arithmetic", JITs might not |
| 2691 | * be able to deal with it. |
| 2692 | */ |
| 2693 | if (ptr_reg->type == PTR_TO_STACK) { |
| 2694 | verbose(env, "R%d subtraction from stack pointer prohibited\n", |
| 2695 | dst); |
| 2696 | return -EACCES; |
| 2697 | } |
| 2698 | if (known && (ptr_reg->off - smin_val == |
| 2699 | (s64)(s32)(ptr_reg->off - smin_val))) { |
| 2700 | /* pointer -= K. Subtract it from fixed offset */ |
| 2701 | dst_reg->smin_value = smin_ptr; |
| 2702 | dst_reg->smax_value = smax_ptr; |
| 2703 | dst_reg->umin_value = umin_ptr; |
| 2704 | dst_reg->umax_value = umax_ptr; |
| 2705 | dst_reg->var_off = ptr_reg->var_off; |
| 2706 | dst_reg->id = ptr_reg->id; |
| 2707 | dst_reg->off = ptr_reg->off - smin_val; |
| 2708 | dst_reg->range = ptr_reg->range; |
| 2709 | break; |
| 2710 | } |
| 2711 | /* A new variable offset is created. If the subtrahend is known |
| 2712 | * nonnegative, then any reg->range we had before is still good. |
| 2713 | */ |
| 2714 | if (signed_sub_overflows(smin_ptr, smax_val) || |
| 2715 | signed_sub_overflows(smax_ptr, smin_val)) { |
| 2716 | /* Overflow possible, we know nothing */ |
| 2717 | dst_reg->smin_value = S64_MIN; |
| 2718 | dst_reg->smax_value = S64_MAX; |
| 2719 | } else { |
| 2720 | dst_reg->smin_value = smin_ptr - smax_val; |
| 2721 | dst_reg->smax_value = smax_ptr - smin_val; |
| 2722 | } |
| 2723 | if (umin_ptr < umax_val) { |
| 2724 | /* Overflow possible, we know nothing */ |
| 2725 | dst_reg->umin_value = 0; |
| 2726 | dst_reg->umax_value = U64_MAX; |
| 2727 | } else { |
| 2728 | /* Cannot overflow (as long as bounds are consistent) */ |
| 2729 | dst_reg->umin_value = umin_ptr - umax_val; |
| 2730 | dst_reg->umax_value = umax_ptr - umin_val; |
| 2731 | } |
| 2732 | dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off); |
| 2733 | dst_reg->off = ptr_reg->off; |
| 2734 | if (reg_is_pkt_pointer(ptr_reg)) { |
| 2735 | dst_reg->id = ++env->id_gen; |
| 2736 | /* something was added to pkt_ptr, set range to zero */ |
| 2737 | if (smin_val < 0) |
| 2738 | dst_reg->range = 0; |
| 2739 | } |
| 2740 | break; |
| 2741 | case BPF_AND: |
| 2742 | case BPF_OR: |
| 2743 | case BPF_XOR: |
| 2744 | /* bitwise ops on pointers are troublesome, prohibit. */ |
| 2745 | verbose(env, "R%d bitwise operator %s on pointer prohibited\n", |
| 2746 | dst, bpf_alu_string[opcode >> 4]); |
| 2747 | return -EACCES; |
| 2748 | default: |
| 2749 | /* other operators (e.g. MUL,LSH) produce non-pointer results */ |
| 2750 | verbose(env, "R%d pointer arithmetic with %s operator prohibited\n", |
| 2751 | dst, bpf_alu_string[opcode >> 4]); |
| 2752 | return -EACCES; |
| 2753 | } |
| 2754 | |
| 2755 | if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type)) |
| 2756 | return -EINVAL; |
| 2757 | |
| 2758 | __update_reg_bounds(dst_reg); |
| 2759 | __reg_deduce_bounds(dst_reg); |
| 2760 | __reg_bound_offset(dst_reg); |
| 2761 | return 0; |
| 2762 | } |
| 2763 | |
| 2764 | /* WARNING: This function does calculations on 64-bit values, but the actual |
| 2765 | * execution may occur on 32-bit values. Therefore, things like bitshifts |
| 2766 | * need extra checks in the 32-bit case. |
| 2767 | */ |
| 2768 | static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env, |
| 2769 | struct bpf_insn *insn, |
| 2770 | struct bpf_reg_state *dst_reg, |
| 2771 | struct bpf_reg_state src_reg) |
| 2772 | { |
| 2773 | struct bpf_reg_state *regs = cur_regs(env); |
| 2774 | u8 opcode = BPF_OP(insn->code); |
| 2775 | bool src_known, dst_known; |
| 2776 | s64 smin_val, smax_val; |
| 2777 | u64 umin_val, umax_val; |
| 2778 | u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32; |
| 2779 | |
| 2780 | smin_val = src_reg.smin_value; |
| 2781 | smax_val = src_reg.smax_value; |
| 2782 | umin_val = src_reg.umin_value; |
| 2783 | umax_val = src_reg.umax_value; |
| 2784 | src_known = tnum_is_const(src_reg.var_off); |
| 2785 | dst_known = tnum_is_const(dst_reg->var_off); |
| 2786 | |
| 2787 | if ((src_known && (smin_val != smax_val || umin_val != umax_val)) || |
| 2788 | smin_val > smax_val || umin_val > umax_val) { |
| 2789 | /* Taint dst register if offset had invalid bounds derived from |
| 2790 | * e.g. dead branches. |
| 2791 | */ |
| 2792 | __mark_reg_unknown(dst_reg); |
| 2793 | return 0; |
| 2794 | } |
| 2795 | |
| 2796 | if (!src_known && |
| 2797 | opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) { |
| 2798 | __mark_reg_unknown(dst_reg); |
| 2799 | return 0; |
| 2800 | } |
| 2801 | |
| 2802 | switch (opcode) { |
| 2803 | case BPF_ADD: |
| 2804 | if (signed_add_overflows(dst_reg->smin_value, smin_val) || |
| 2805 | signed_add_overflows(dst_reg->smax_value, smax_val)) { |
| 2806 | dst_reg->smin_value = S64_MIN; |
| 2807 | dst_reg->smax_value = S64_MAX; |
| 2808 | } else { |
| 2809 | dst_reg->smin_value += smin_val; |
| 2810 | dst_reg->smax_value += smax_val; |
| 2811 | } |
| 2812 | if (dst_reg->umin_value + umin_val < umin_val || |
| 2813 | dst_reg->umax_value + umax_val < umax_val) { |
| 2814 | dst_reg->umin_value = 0; |
| 2815 | dst_reg->umax_value = U64_MAX; |
| 2816 | } else { |
| 2817 | dst_reg->umin_value += umin_val; |
| 2818 | dst_reg->umax_value += umax_val; |
| 2819 | } |
| 2820 | dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off); |
| 2821 | break; |
| 2822 | case BPF_SUB: |
| 2823 | if (signed_sub_overflows(dst_reg->smin_value, smax_val) || |
| 2824 | signed_sub_overflows(dst_reg->smax_value, smin_val)) { |
| 2825 | /* Overflow possible, we know nothing */ |
| 2826 | dst_reg->smin_value = S64_MIN; |
| 2827 | dst_reg->smax_value = S64_MAX; |
| 2828 | } else { |
| 2829 | dst_reg->smin_value -= smax_val; |
| 2830 | dst_reg->smax_value -= smin_val; |
| 2831 | } |
| 2832 | if (dst_reg->umin_value < umax_val) { |
| 2833 | /* Overflow possible, we know nothing */ |
| 2834 | dst_reg->umin_value = 0; |
| 2835 | dst_reg->umax_value = U64_MAX; |
| 2836 | } else { |
| 2837 | /* Cannot overflow (as long as bounds are consistent) */ |
| 2838 | dst_reg->umin_value -= umax_val; |
| 2839 | dst_reg->umax_value -= umin_val; |
| 2840 | } |
| 2841 | dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off); |
| 2842 | break; |
| 2843 | case BPF_MUL: |
| 2844 | dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off); |
| 2845 | if (smin_val < 0 || dst_reg->smin_value < 0) { |
| 2846 | /* Ain't nobody got time to multiply that sign */ |
| 2847 | __mark_reg_unbounded(dst_reg); |
| 2848 | __update_reg_bounds(dst_reg); |
| 2849 | break; |
| 2850 | } |
| 2851 | /* Both values are positive, so we can work with unsigned and |
| 2852 | * copy the result to signed (unless it exceeds S64_MAX). |
| 2853 | */ |
| 2854 | if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) { |
| 2855 | /* Potential overflow, we know nothing */ |
| 2856 | __mark_reg_unbounded(dst_reg); |
| 2857 | /* (except what we can learn from the var_off) */ |
| 2858 | __update_reg_bounds(dst_reg); |
| 2859 | break; |
| 2860 | } |
| 2861 | dst_reg->umin_value *= umin_val; |
| 2862 | dst_reg->umax_value *= umax_val; |
| 2863 | if (dst_reg->umax_value > S64_MAX) { |
| 2864 | /* Overflow possible, we know nothing */ |
| 2865 | dst_reg->smin_value = S64_MIN; |
| 2866 | dst_reg->smax_value = S64_MAX; |
| 2867 | } else { |
| 2868 | dst_reg->smin_value = dst_reg->umin_value; |
| 2869 | dst_reg->smax_value = dst_reg->umax_value; |
| 2870 | } |
| 2871 | break; |
| 2872 | case BPF_AND: |
| 2873 | if (src_known && dst_known) { |
| 2874 | __mark_reg_known(dst_reg, dst_reg->var_off.value & |
| 2875 | src_reg.var_off.value); |
| 2876 | break; |
| 2877 | } |
| 2878 | /* We get our minimum from the var_off, since that's inherently |
| 2879 | * bitwise. Our maximum is the minimum of the operands' maxima. |
| 2880 | */ |
| 2881 | dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off); |
| 2882 | dst_reg->umin_value = dst_reg->var_off.value; |
| 2883 | dst_reg->umax_value = min(dst_reg->umax_value, umax_val); |
| 2884 | if (dst_reg->smin_value < 0 || smin_val < 0) { |
| 2885 | /* Lose signed bounds when ANDing negative numbers, |
| 2886 | * ain't nobody got time for that. |
| 2887 | */ |
| 2888 | dst_reg->smin_value = S64_MIN; |
| 2889 | dst_reg->smax_value = S64_MAX; |
| 2890 | } else { |
| 2891 | /* ANDing two positives gives a positive, so safe to |
| 2892 | * cast result into s64. |
| 2893 | */ |
| 2894 | dst_reg->smin_value = dst_reg->umin_value; |
| 2895 | dst_reg->smax_value = dst_reg->umax_value; |
| 2896 | } |
| 2897 | /* We may learn something more from the var_off */ |
| 2898 | __update_reg_bounds(dst_reg); |
| 2899 | break; |
| 2900 | case BPF_OR: |
| 2901 | if (src_known && dst_known) { |
| 2902 | __mark_reg_known(dst_reg, dst_reg->var_off.value | |
| 2903 | src_reg.var_off.value); |
| 2904 | break; |
| 2905 | } |
| 2906 | /* We get our maximum from the var_off, and our minimum is the |
| 2907 | * maximum of the operands' minima |
| 2908 | */ |
| 2909 | dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off); |
| 2910 | dst_reg->umin_value = max(dst_reg->umin_value, umin_val); |
| 2911 | dst_reg->umax_value = dst_reg->var_off.value | |
| 2912 | dst_reg->var_off.mask; |
| 2913 | if (dst_reg->smin_value < 0 || smin_val < 0) { |
| 2914 | /* Lose signed bounds when ORing negative numbers, |
| 2915 | * ain't nobody got time for that. |
| 2916 | */ |
| 2917 | dst_reg->smin_value = S64_MIN; |
| 2918 | dst_reg->smax_value = S64_MAX; |
| 2919 | } else { |
| 2920 | /* ORing two positives gives a positive, so safe to |
| 2921 | * cast result into s64. |
| 2922 | */ |
| 2923 | dst_reg->smin_value = dst_reg->umin_value; |
| 2924 | dst_reg->smax_value = dst_reg->umax_value; |
| 2925 | } |
| 2926 | /* We may learn something more from the var_off */ |
| 2927 | __update_reg_bounds(dst_reg); |
| 2928 | break; |
| 2929 | case BPF_LSH: |
| 2930 | if (umax_val >= insn_bitness) { |
| 2931 | /* Shifts greater than 31 or 63 are undefined. |
| 2932 | * This includes shifts by a negative number. |
| 2933 | */ |
| 2934 | mark_reg_unknown(env, regs, insn->dst_reg); |
| 2935 | break; |
| 2936 | } |
| 2937 | /* We lose all sign bit information (except what we can pick |
| 2938 | * up from var_off) |
| 2939 | */ |
| 2940 | dst_reg->smin_value = S64_MIN; |
| 2941 | dst_reg->smax_value = S64_MAX; |
| 2942 | /* If we might shift our top bit out, then we know nothing */ |
| 2943 | if (dst_reg->umax_value > 1ULL << (63 - umax_val)) { |
| 2944 | dst_reg->umin_value = 0; |
| 2945 | dst_reg->umax_value = U64_MAX; |
| 2946 | } else { |
| 2947 | dst_reg->umin_value <<= umin_val; |
| 2948 | dst_reg->umax_value <<= umax_val; |
| 2949 | } |
| 2950 | dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val); |
| 2951 | /* We may learn something more from the var_off */ |
| 2952 | __update_reg_bounds(dst_reg); |
| 2953 | break; |
| 2954 | case BPF_RSH: |
| 2955 | if (umax_val >= insn_bitness) { |
| 2956 | /* Shifts greater than 31 or 63 are undefined. |
| 2957 | * This includes shifts by a negative number. |
| 2958 | */ |
| 2959 | mark_reg_unknown(env, regs, insn->dst_reg); |
| 2960 | break; |
| 2961 | } |
| 2962 | /* BPF_RSH is an unsigned shift. If the value in dst_reg might |
| 2963 | * be negative, then either: |
| 2964 | * 1) src_reg might be zero, so the sign bit of the result is |
| 2965 | * unknown, so we lose our signed bounds |
| 2966 | * 2) it's known negative, thus the unsigned bounds capture the |
| 2967 | * signed bounds |
| 2968 | * 3) the signed bounds cross zero, so they tell us nothing |
| 2969 | * about the result |
| 2970 | * If the value in dst_reg is known nonnegative, then again the |
| 2971 | * unsigned bounts capture the signed bounds. |
| 2972 | * Thus, in all cases it suffices to blow away our signed bounds |
| 2973 | * and rely on inferring new ones from the unsigned bounds and |
| 2974 | * var_off of the result. |
| 2975 | */ |
| 2976 | dst_reg->smin_value = S64_MIN; |
| 2977 | dst_reg->smax_value = S64_MAX; |
| 2978 | dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val); |
| 2979 | dst_reg->umin_value >>= umax_val; |
| 2980 | dst_reg->umax_value >>= umin_val; |
| 2981 | /* We may learn something more from the var_off */ |
| 2982 | __update_reg_bounds(dst_reg); |
| 2983 | break; |
| 2984 | case BPF_ARSH: |
| 2985 | if (umax_val >= insn_bitness) { |
| 2986 | /* Shifts greater than 31 or 63 are undefined. |
| 2987 | * This includes shifts by a negative number. |
| 2988 | */ |
| 2989 | mark_reg_unknown(env, regs, insn->dst_reg); |
| 2990 | break; |
| 2991 | } |
| 2992 | |
| 2993 | /* Upon reaching here, src_known is true and |
| 2994 | * umax_val is equal to umin_val. |
| 2995 | */ |
| 2996 | dst_reg->smin_value >>= umin_val; |
| 2997 | dst_reg->smax_value >>= umin_val; |
| 2998 | dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val); |
| 2999 | |
| 3000 | /* blow away the dst_reg umin_value/umax_value and rely on |
| 3001 | * dst_reg var_off to refine the result. |
| 3002 | */ |
| 3003 | dst_reg->umin_value = 0; |
| 3004 | dst_reg->umax_value = U64_MAX; |
| 3005 | __update_reg_bounds(dst_reg); |
| 3006 | break; |
| 3007 | default: |
| 3008 | mark_reg_unknown(env, regs, insn->dst_reg); |
| 3009 | break; |
| 3010 | } |
| 3011 | |
| 3012 | if (BPF_CLASS(insn->code) != BPF_ALU64) { |
| 3013 | /* 32-bit ALU ops are (32,32)->32 */ |
| 3014 | coerce_reg_to_size(dst_reg, 4); |
| 3015 | coerce_reg_to_size(&src_reg, 4); |
| 3016 | } |
| 3017 | |
| 3018 | __reg_deduce_bounds(dst_reg); |
| 3019 | __reg_bound_offset(dst_reg); |
| 3020 | return 0; |
| 3021 | } |
| 3022 | |
| 3023 | /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max |
| 3024 | * and var_off. |
| 3025 | */ |
| 3026 | static int adjust_reg_min_max_vals(struct bpf_verifier_env *env, |
| 3027 | struct bpf_insn *insn) |
| 3028 | { |
| 3029 | struct bpf_verifier_state *vstate = env->cur_state; |
| 3030 | struct bpf_func_state *state = vstate->frame[vstate->curframe]; |
| 3031 | struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg; |
| 3032 | struct bpf_reg_state *ptr_reg = NULL, off_reg = {0}; |
| 3033 | u8 opcode = BPF_OP(insn->code); |
| 3034 | |
| 3035 | dst_reg = ®s[insn->dst_reg]; |
| 3036 | src_reg = NULL; |
| 3037 | if (dst_reg->type != SCALAR_VALUE) |
| 3038 | ptr_reg = dst_reg; |
| 3039 | if (BPF_SRC(insn->code) == BPF_X) { |
| 3040 | src_reg = ®s[insn->src_reg]; |
| 3041 | if (src_reg->type != SCALAR_VALUE) { |
| 3042 | if (dst_reg->type != SCALAR_VALUE) { |
| 3043 | /* Combining two pointers by any ALU op yields |
| 3044 | * an arbitrary scalar. Disallow all math except |
| 3045 | * pointer subtraction |
| 3046 | */ |
| 3047 | if (opcode == BPF_SUB){ |
| 3048 | mark_reg_unknown(env, regs, insn->dst_reg); |
| 3049 | return 0; |
| 3050 | } |
| 3051 | verbose(env, "R%d pointer %s pointer prohibited\n", |
| 3052 | insn->dst_reg, |
| 3053 | bpf_alu_string[opcode >> 4]); |
| 3054 | return -EACCES; |
| 3055 | } else { |
| 3056 | /* scalar += pointer |
| 3057 | * This is legal, but we have to reverse our |
| 3058 | * src/dest handling in computing the range |
| 3059 | */ |
| 3060 | return adjust_ptr_min_max_vals(env, insn, |
| 3061 | src_reg, dst_reg); |
| 3062 | } |
| 3063 | } else if (ptr_reg) { |
| 3064 | /* pointer += scalar */ |
| 3065 | return adjust_ptr_min_max_vals(env, insn, |
| 3066 | dst_reg, src_reg); |
| 3067 | } |
| 3068 | } else { |
| 3069 | /* Pretend the src is a reg with a known value, since we only |
| 3070 | * need to be able to read from this state. |
| 3071 | */ |
| 3072 | off_reg.type = SCALAR_VALUE; |
| 3073 | __mark_reg_known(&off_reg, insn->imm); |
| 3074 | src_reg = &off_reg; |
| 3075 | if (ptr_reg) /* pointer += K */ |
| 3076 | return adjust_ptr_min_max_vals(env, insn, |
| 3077 | ptr_reg, src_reg); |
| 3078 | } |
| 3079 | |
| 3080 | /* Got here implies adding two SCALAR_VALUEs */ |
| 3081 | if (WARN_ON_ONCE(ptr_reg)) { |
| 3082 | print_verifier_state(env, state); |
| 3083 | verbose(env, "verifier internal error: unexpected ptr_reg\n"); |
| 3084 | return -EINVAL; |
| 3085 | } |
| 3086 | if (WARN_ON(!src_reg)) { |
| 3087 | print_verifier_state(env, state); |
| 3088 | verbose(env, "verifier internal error: no src_reg\n"); |
| 3089 | return -EINVAL; |
| 3090 | } |
| 3091 | return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg); |
| 3092 | } |
| 3093 | |
| 3094 | /* check validity of 32-bit and 64-bit arithmetic operations */ |
| 3095 | static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn) |
| 3096 | { |
| 3097 | struct bpf_reg_state *regs = cur_regs(env); |
| 3098 | u8 opcode = BPF_OP(insn->code); |
| 3099 | int err; |
| 3100 | |
| 3101 | if (opcode == BPF_END || opcode == BPF_NEG) { |
| 3102 | if (opcode == BPF_NEG) { |
| 3103 | if (BPF_SRC(insn->code) != 0 || |
| 3104 | insn->src_reg != BPF_REG_0 || |
| 3105 | insn->off != 0 || insn->imm != 0) { |
| 3106 | verbose(env, "BPF_NEG uses reserved fields\n"); |
| 3107 | return -EINVAL; |
| 3108 | } |
| 3109 | } else { |
| 3110 | if (insn->src_reg != BPF_REG_0 || insn->off != 0 || |
| 3111 | (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) || |
| 3112 | BPF_CLASS(insn->code) == BPF_ALU64) { |
| 3113 | verbose(env, "BPF_END uses reserved fields\n"); |
| 3114 | return -EINVAL; |
| 3115 | } |
| 3116 | } |
| 3117 | |
| 3118 | /* check src operand */ |
| 3119 | err = check_reg_arg(env, insn->dst_reg, SRC_OP); |
| 3120 | if (err) |
| 3121 | return err; |
| 3122 | |
| 3123 | if (is_pointer_value(env, insn->dst_reg)) { |
| 3124 | verbose(env, "R%d pointer arithmetic prohibited\n", |
| 3125 | insn->dst_reg); |
| 3126 | return -EACCES; |
| 3127 | } |
| 3128 | |
| 3129 | /* check dest operand */ |
| 3130 | err = check_reg_arg(env, insn->dst_reg, DST_OP); |
| 3131 | if (err) |
| 3132 | return err; |
| 3133 | |
| 3134 | } else if (opcode == BPF_MOV) { |
| 3135 | |
| 3136 | if (BPF_SRC(insn->code) == BPF_X) { |
| 3137 | if (insn->imm != 0 || insn->off != 0) { |
| 3138 | verbose(env, "BPF_MOV uses reserved fields\n"); |
| 3139 | return -EINVAL; |
| 3140 | } |
| 3141 | |
| 3142 | /* check src operand */ |
| 3143 | err = check_reg_arg(env, insn->src_reg, SRC_OP); |
| 3144 | if (err) |
| 3145 | return err; |
| 3146 | } else { |
| 3147 | if (insn->src_reg != BPF_REG_0 || insn->off != 0) { |
| 3148 | verbose(env, "BPF_MOV uses reserved fields\n"); |
| 3149 | return -EINVAL; |
| 3150 | } |
| 3151 | } |
| 3152 | |
| 3153 | /* check dest operand */ |
| 3154 | err = check_reg_arg(env, insn->dst_reg, DST_OP); |
| 3155 | if (err) |
| 3156 | return err; |
| 3157 | |
| 3158 | if (BPF_SRC(insn->code) == BPF_X) { |
| 3159 | if (BPF_CLASS(insn->code) == BPF_ALU64) { |
| 3160 | /* case: R1 = R2 |
| 3161 | * copy register state to dest reg |
| 3162 | */ |
| 3163 | regs[insn->dst_reg] = regs[insn->src_reg]; |
| 3164 | regs[insn->dst_reg].live |= REG_LIVE_WRITTEN; |
| 3165 | } else { |
| 3166 | /* R1 = (u32) R2 */ |
| 3167 | if (is_pointer_value(env, insn->src_reg)) { |
| 3168 | verbose(env, |
| 3169 | "R%d partial copy of pointer\n", |
| 3170 | insn->src_reg); |
| 3171 | return -EACCES; |
| 3172 | } |
| 3173 | mark_reg_unknown(env, regs, insn->dst_reg); |
| 3174 | coerce_reg_to_size(®s[insn->dst_reg], 4); |
| 3175 | } |
| 3176 | } else { |
| 3177 | /* case: R = imm |
| 3178 | * remember the value we stored into this reg |
| 3179 | */ |
| 3180 | regs[insn->dst_reg].type = SCALAR_VALUE; |
| 3181 | if (BPF_CLASS(insn->code) == BPF_ALU64) { |
| 3182 | __mark_reg_known(regs + insn->dst_reg, |
| 3183 | insn->imm); |
| 3184 | } else { |
| 3185 | __mark_reg_known(regs + insn->dst_reg, |
| 3186 | (u32)insn->imm); |
| 3187 | } |
| 3188 | } |
| 3189 | |
| 3190 | } else if (opcode > BPF_END) { |
| 3191 | verbose(env, "invalid BPF_ALU opcode %x\n", opcode); |
| 3192 | return -EINVAL; |
| 3193 | |
| 3194 | } else { /* all other ALU ops: and, sub, xor, add, ... */ |
| 3195 | |
| 3196 | if (BPF_SRC(insn->code) == BPF_X) { |
| 3197 | if (insn->imm != 0 || insn->off != 0) { |
| 3198 | verbose(env, "BPF_ALU uses reserved fields\n"); |
| 3199 | return -EINVAL; |
| 3200 | } |
| 3201 | /* check src1 operand */ |
| 3202 | err = check_reg_arg(env, insn->src_reg, SRC_OP); |
| 3203 | if (err) |
| 3204 | return err; |
| 3205 | } else { |
| 3206 | if (insn->src_reg != BPF_REG_0 || insn->off != 0) { |
| 3207 | verbose(env, "BPF_ALU uses reserved fields\n"); |
| 3208 | return -EINVAL; |
| 3209 | } |
| 3210 | } |
| 3211 | |
| 3212 | /* check src2 operand */ |
| 3213 | err = check_reg_arg(env, insn->dst_reg, SRC_OP); |
| 3214 | if (err) |
| 3215 | return err; |
| 3216 | |
| 3217 | if ((opcode == BPF_MOD || opcode == BPF_DIV) && |
| 3218 | BPF_SRC(insn->code) == BPF_K && insn->imm == 0) { |
| 3219 | verbose(env, "div by zero\n"); |
| 3220 | return -EINVAL; |
| 3221 | } |
| 3222 | |
| 3223 | if (opcode == BPF_ARSH && BPF_CLASS(insn->code) != BPF_ALU64) { |
| 3224 | verbose(env, "BPF_ARSH not supported for 32 bit ALU\n"); |
| 3225 | return -EINVAL; |
| 3226 | } |
| 3227 | |
| 3228 | if ((opcode == BPF_LSH || opcode == BPF_RSH || |
| 3229 | opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) { |
| 3230 | int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32; |
| 3231 | |
| 3232 | if (insn->imm < 0 || insn->imm >= size) { |
| 3233 | verbose(env, "invalid shift %d\n", insn->imm); |
| 3234 | return -EINVAL; |
| 3235 | } |
| 3236 | } |
| 3237 | |
| 3238 | /* check dest operand */ |
| 3239 | err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK); |
| 3240 | if (err) |
| 3241 | return err; |
| 3242 | |
| 3243 | return adjust_reg_min_max_vals(env, insn); |
| 3244 | } |
| 3245 | |
| 3246 | return 0; |
| 3247 | } |
| 3248 | |
| 3249 | static void find_good_pkt_pointers(struct bpf_verifier_state *vstate, |
| 3250 | struct bpf_reg_state *dst_reg, |
| 3251 | enum bpf_reg_type type, |
| 3252 | bool range_right_open) |
| 3253 | { |
| 3254 | struct bpf_func_state *state = vstate->frame[vstate->curframe]; |
| 3255 | struct bpf_reg_state *regs = state->regs, *reg; |
| 3256 | u16 new_range; |
| 3257 | int i, j; |
| 3258 | |
| 3259 | if (dst_reg->off < 0 || |
| 3260 | (dst_reg->off == 0 && range_right_open)) |
| 3261 | /* This doesn't give us any range */ |
| 3262 | return; |
| 3263 | |
| 3264 | if (dst_reg->umax_value > MAX_PACKET_OFF || |
| 3265 | dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF) |
| 3266 | /* Risk of overflow. For instance, ptr + (1<<63) may be less |
| 3267 | * than pkt_end, but that's because it's also less than pkt. |
| 3268 | */ |
| 3269 | return; |
| 3270 | |
| 3271 | new_range = dst_reg->off; |
| 3272 | if (range_right_open) |
| 3273 | new_range--; |
| 3274 | |
| 3275 | /* Examples for register markings: |
| 3276 | * |
| 3277 | * pkt_data in dst register: |
| 3278 | * |
| 3279 | * r2 = r3; |
| 3280 | * r2 += 8; |
| 3281 | * if (r2 > pkt_end) goto <handle exception> |
| 3282 | * <access okay> |
| 3283 | * |
| 3284 | * r2 = r3; |
| 3285 | * r2 += 8; |
| 3286 | * if (r2 < pkt_end) goto <access okay> |
| 3287 | * <handle exception> |
| 3288 | * |
| 3289 | * Where: |
| 3290 | * r2 == dst_reg, pkt_end == src_reg |
| 3291 | * r2=pkt(id=n,off=8,r=0) |
| 3292 | * r3=pkt(id=n,off=0,r=0) |
| 3293 | * |
| 3294 | * pkt_data in src register: |
| 3295 | * |
| 3296 | * r2 = r3; |
| 3297 | * r2 += 8; |
| 3298 | * if (pkt_end >= r2) goto <access okay> |
| 3299 | * <handle exception> |
| 3300 | * |
| 3301 | * r2 = r3; |
| 3302 | * r2 += 8; |
| 3303 | * if (pkt_end <= r2) goto <handle exception> |
| 3304 | * <access okay> |
| 3305 | * |
| 3306 | * Where: |
| 3307 | * pkt_end == dst_reg, r2 == src_reg |
| 3308 | * r2=pkt(id=n,off=8,r=0) |
| 3309 | * r3=pkt(id=n,off=0,r=0) |
| 3310 | * |
| 3311 | * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8) |
| 3312 | * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8) |
| 3313 | * and [r3, r3 + 8-1) respectively is safe to access depending on |
| 3314 | * the check. |
| 3315 | */ |
| 3316 | |
| 3317 | /* If our ids match, then we must have the same max_value. And we |
| 3318 | * don't care about the other reg's fixed offset, since if it's too big |
| 3319 | * the range won't allow anything. |
| 3320 | * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16. |
| 3321 | */ |
| 3322 | for (i = 0; i < MAX_BPF_REG; i++) |
| 3323 | if (regs[i].type == type && regs[i].id == dst_reg->id) |
| 3324 | /* keep the maximum range already checked */ |
| 3325 | regs[i].range = max(regs[i].range, new_range); |
| 3326 | |
| 3327 | for (j = 0; j <= vstate->curframe; j++) { |
| 3328 | state = vstate->frame[j]; |
| 3329 | for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) { |
| 3330 | if (state->stack[i].slot_type[0] != STACK_SPILL) |
| 3331 | continue; |
| 3332 | reg = &state->stack[i].spilled_ptr; |
| 3333 | if (reg->type == type && reg->id == dst_reg->id) |
| 3334 | reg->range = max(reg->range, new_range); |
| 3335 | } |
| 3336 | } |
| 3337 | } |
| 3338 | |
| 3339 | /* Adjusts the register min/max values in the case that the dst_reg is the |
| 3340 | * variable register that we are working on, and src_reg is a constant or we're |
| 3341 | * simply doing a BPF_K check. |
| 3342 | * In JEQ/JNE cases we also adjust the var_off values. |
| 3343 | */ |
| 3344 | static void reg_set_min_max(struct bpf_reg_state *true_reg, |
| 3345 | struct bpf_reg_state *false_reg, u64 val, |
| 3346 | u8 opcode) |
| 3347 | { |
| 3348 | /* If the dst_reg is a pointer, we can't learn anything about its |
| 3349 | * variable offset from the compare (unless src_reg were a pointer into |
| 3350 | * the same object, but we don't bother with that. |
| 3351 | * Since false_reg and true_reg have the same type by construction, we |
| 3352 | * only need to check one of them for pointerness. |
| 3353 | */ |
| 3354 | if (__is_pointer_value(false, false_reg)) |
| 3355 | return; |
| 3356 | |
| 3357 | switch (opcode) { |
| 3358 | case BPF_JEQ: |
| 3359 | /* If this is false then we know nothing Jon Snow, but if it is |
| 3360 | * true then we know for sure. |
| 3361 | */ |
| 3362 | __mark_reg_known(true_reg, val); |
| 3363 | break; |
| 3364 | case BPF_JNE: |
| 3365 | /* If this is true we know nothing Jon Snow, but if it is false |
| 3366 | * we know the value for sure; |
| 3367 | */ |
| 3368 | __mark_reg_known(false_reg, val); |
| 3369 | break; |
| 3370 | case BPF_JGT: |
| 3371 | false_reg->umax_value = min(false_reg->umax_value, val); |
| 3372 | true_reg->umin_value = max(true_reg->umin_value, val + 1); |
| 3373 | break; |
| 3374 | case BPF_JSGT: |
| 3375 | false_reg->smax_value = min_t(s64, false_reg->smax_value, val); |
| 3376 | true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1); |
| 3377 | break; |
| 3378 | case BPF_JLT: |
| 3379 | false_reg->umin_value = max(false_reg->umin_value, val); |
| 3380 | true_reg->umax_value = min(true_reg->umax_value, val - 1); |
| 3381 | break; |
| 3382 | case BPF_JSLT: |
| 3383 | false_reg->smin_value = max_t(s64, false_reg->smin_value, val); |
| 3384 | true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1); |
| 3385 | break; |
| 3386 | case BPF_JGE: |
| 3387 | false_reg->umax_value = min(false_reg->umax_value, val - 1); |
| 3388 | true_reg->umin_value = max(true_reg->umin_value, val); |
| 3389 | break; |
| 3390 | case BPF_JSGE: |
| 3391 | false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1); |
| 3392 | true_reg->smin_value = max_t(s64, true_reg->smin_value, val); |
| 3393 | break; |
| 3394 | case BPF_JLE: |
| 3395 | false_reg->umin_value = max(false_reg->umin_value, val + 1); |
| 3396 | true_reg->umax_value = min(true_reg->umax_value, val); |
| 3397 | break; |
| 3398 | case BPF_JSLE: |
| 3399 | false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1); |
| 3400 | true_reg->smax_value = min_t(s64, true_reg->smax_value, val); |
| 3401 | break; |
| 3402 | default: |
| 3403 | break; |
| 3404 | } |
| 3405 | |
| 3406 | __reg_deduce_bounds(false_reg); |
| 3407 | __reg_deduce_bounds(true_reg); |
| 3408 | /* We might have learned some bits from the bounds. */ |
| 3409 | __reg_bound_offset(false_reg); |
| 3410 | __reg_bound_offset(true_reg); |
| 3411 | /* Intersecting with the old var_off might have improved our bounds |
| 3412 | * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc), |
| 3413 | * then new var_off is (0; 0x7f...fc) which improves our umax. |
| 3414 | */ |
| 3415 | __update_reg_bounds(false_reg); |
| 3416 | __update_reg_bounds(true_reg); |
| 3417 | } |
| 3418 | |
| 3419 | /* Same as above, but for the case that dst_reg holds a constant and src_reg is |
| 3420 | * the variable reg. |
| 3421 | */ |
| 3422 | static void reg_set_min_max_inv(struct bpf_reg_state *true_reg, |
| 3423 | struct bpf_reg_state *false_reg, u64 val, |
| 3424 | u8 opcode) |
| 3425 | { |
| 3426 | if (__is_pointer_value(false, false_reg)) |
| 3427 | return; |
| 3428 | |
| 3429 | switch (opcode) { |
| 3430 | case BPF_JEQ: |
| 3431 | /* If this is false then we know nothing Jon Snow, but if it is |
| 3432 | * true then we know for sure. |
| 3433 | */ |
| 3434 | __mark_reg_known(true_reg, val); |
| 3435 | break; |
| 3436 | case BPF_JNE: |
| 3437 | /* If this is true we know nothing Jon Snow, but if it is false |
| 3438 | * we know the value for sure; |
| 3439 | */ |
| 3440 | __mark_reg_known(false_reg, val); |
| 3441 | break; |
| 3442 | case BPF_JGT: |
| 3443 | true_reg->umax_value = min(true_reg->umax_value, val - 1); |
| 3444 | false_reg->umin_value = max(false_reg->umin_value, val); |
| 3445 | break; |
| 3446 | case BPF_JSGT: |
| 3447 | true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1); |
| 3448 | false_reg->smin_value = max_t(s64, false_reg->smin_value, val); |
| 3449 | break; |
| 3450 | case BPF_JLT: |
| 3451 | true_reg->umin_value = max(true_reg->umin_value, val + 1); |
| 3452 | false_reg->umax_value = min(false_reg->umax_value, val); |
| 3453 | break; |
| 3454 | case BPF_JSLT: |
| 3455 | true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1); |
| 3456 | false_reg->smax_value = min_t(s64, false_reg->smax_value, val); |
| 3457 | break; |
| 3458 | case BPF_JGE: |
| 3459 | true_reg->umax_value = min(true_reg->umax_value, val); |
| 3460 | false_reg->umin_value = max(false_reg->umin_value, val + 1); |
| 3461 | break; |
| 3462 | case BPF_JSGE: |
| 3463 | true_reg->smax_value = min_t(s64, true_reg->smax_value, val); |
| 3464 | false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1); |
| 3465 | break; |
| 3466 | case BPF_JLE: |
| 3467 | true_reg->umin_value = max(true_reg->umin_value, val); |
| 3468 | false_reg->umax_value = min(false_reg->umax_value, val - 1); |
| 3469 | break; |
| 3470 | case BPF_JSLE: |
| 3471 | true_reg->smin_value = max_t(s64, true_reg->smin_value, val); |
| 3472 | false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1); |
| 3473 | break; |
| 3474 | default: |
| 3475 | break; |
| 3476 | } |
| 3477 | |
| 3478 | __reg_deduce_bounds(false_reg); |
| 3479 | __reg_deduce_bounds(true_reg); |
| 3480 | /* We might have learned some bits from the bounds. */ |
| 3481 | __reg_bound_offset(false_reg); |
| 3482 | __reg_bound_offset(true_reg); |
| 3483 | /* Intersecting with the old var_off might have improved our bounds |
| 3484 | * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc), |
| 3485 | * then new var_off is (0; 0x7f...fc) which improves our umax. |
| 3486 | */ |
| 3487 | __update_reg_bounds(false_reg); |
| 3488 | __update_reg_bounds(true_reg); |
| 3489 | } |
| 3490 | |
| 3491 | /* Regs are known to be equal, so intersect their min/max/var_off */ |
| 3492 | static void __reg_combine_min_max(struct bpf_reg_state *src_reg, |
| 3493 | struct bpf_reg_state *dst_reg) |
| 3494 | { |
| 3495 | src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value, |
| 3496 | dst_reg->umin_value); |
| 3497 | src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value, |
| 3498 | dst_reg->umax_value); |
| 3499 | src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value, |
| 3500 | dst_reg->smin_value); |
| 3501 | src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value, |
| 3502 | dst_reg->smax_value); |
| 3503 | src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off, |
| 3504 | dst_reg->var_off); |
| 3505 | /* We might have learned new bounds from the var_off. */ |
| 3506 | __update_reg_bounds(src_reg); |
| 3507 | __update_reg_bounds(dst_reg); |
| 3508 | /* We might have learned something about the sign bit. */ |
| 3509 | __reg_deduce_bounds(src_reg); |
| 3510 | __reg_deduce_bounds(dst_reg); |
| 3511 | /* We might have learned some bits from the bounds. */ |
| 3512 | __reg_bound_offset(src_reg); |
| 3513 | __reg_bound_offset(dst_reg); |
| 3514 | /* Intersecting with the old var_off might have improved our bounds |
| 3515 | * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc), |
| 3516 | * then new var_off is (0; 0x7f...fc) which improves our umax. |
| 3517 | */ |
| 3518 | __update_reg_bounds(src_reg); |
| 3519 | __update_reg_bounds(dst_reg); |
| 3520 | } |
| 3521 | |
| 3522 | static void reg_combine_min_max(struct bpf_reg_state *true_src, |
| 3523 | struct bpf_reg_state *true_dst, |
| 3524 | struct bpf_reg_state *false_src, |
| 3525 | struct bpf_reg_state *false_dst, |
| 3526 | u8 opcode) |
| 3527 | { |
| 3528 | switch (opcode) { |
| 3529 | case BPF_JEQ: |
| 3530 | __reg_combine_min_max(true_src, true_dst); |
| 3531 | break; |
| 3532 | case BPF_JNE: |
| 3533 | __reg_combine_min_max(false_src, false_dst); |
| 3534 | break; |
| 3535 | } |
| 3536 | } |
| 3537 | |
| 3538 | static void mark_map_reg(struct bpf_reg_state *regs, u32 regno, u32 id, |
| 3539 | bool is_null) |
| 3540 | { |
| 3541 | struct bpf_reg_state *reg = ®s[regno]; |
| 3542 | |
| 3543 | if (reg->type == PTR_TO_MAP_VALUE_OR_NULL && reg->id == id) { |
| 3544 | /* Old offset (both fixed and variable parts) should |
| 3545 | * have been known-zero, because we don't allow pointer |
| 3546 | * arithmetic on pointers that might be NULL. |
| 3547 | */ |
| 3548 | if (WARN_ON_ONCE(reg->smin_value || reg->smax_value || |
| 3549 | !tnum_equals_const(reg->var_off, 0) || |
| 3550 | reg->off)) { |
| 3551 | __mark_reg_known_zero(reg); |
| 3552 | reg->off = 0; |
| 3553 | } |
| 3554 | if (is_null) { |
| 3555 | reg->type = SCALAR_VALUE; |
| 3556 | } else if (reg->map_ptr->inner_map_meta) { |
| 3557 | reg->type = CONST_PTR_TO_MAP; |
| 3558 | reg->map_ptr = reg->map_ptr->inner_map_meta; |
| 3559 | } else { |
| 3560 | reg->type = PTR_TO_MAP_VALUE; |
| 3561 | } |
| 3562 | /* We don't need id from this point onwards anymore, thus we |
| 3563 | * should better reset it, so that state pruning has chances |
| 3564 | * to take effect. |
| 3565 | */ |
| 3566 | reg->id = 0; |
| 3567 | } |
| 3568 | } |
| 3569 | |
| 3570 | /* The logic is similar to find_good_pkt_pointers(), both could eventually |
| 3571 | * be folded together at some point. |
| 3572 | */ |
| 3573 | static void mark_map_regs(struct bpf_verifier_state *vstate, u32 regno, |
| 3574 | bool is_null) |
| 3575 | { |
| 3576 | struct bpf_func_state *state = vstate->frame[vstate->curframe]; |
| 3577 | struct bpf_reg_state *regs = state->regs; |
| 3578 | u32 id = regs[regno].id; |
| 3579 | int i, j; |
| 3580 | |
| 3581 | for (i = 0; i < MAX_BPF_REG; i++) |
| 3582 | mark_map_reg(regs, i, id, is_null); |
| 3583 | |
| 3584 | for (j = 0; j <= vstate->curframe; j++) { |
| 3585 | state = vstate->frame[j]; |
| 3586 | for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) { |
| 3587 | if (state->stack[i].slot_type[0] != STACK_SPILL) |
| 3588 | continue; |
| 3589 | mark_map_reg(&state->stack[i].spilled_ptr, 0, id, is_null); |
| 3590 | } |
| 3591 | } |
| 3592 | } |
| 3593 | |
| 3594 | static bool try_match_pkt_pointers(const struct bpf_insn *insn, |
| 3595 | struct bpf_reg_state *dst_reg, |
| 3596 | struct bpf_reg_state *src_reg, |
| 3597 | struct bpf_verifier_state *this_branch, |
| 3598 | struct bpf_verifier_state *other_branch) |
| 3599 | { |
| 3600 | if (BPF_SRC(insn->code) != BPF_X) |
| 3601 | return false; |
| 3602 | |
| 3603 | switch (BPF_OP(insn->code)) { |
| 3604 | case BPF_JGT: |
| 3605 | if ((dst_reg->type == PTR_TO_PACKET && |
| 3606 | src_reg->type == PTR_TO_PACKET_END) || |
| 3607 | (dst_reg->type == PTR_TO_PACKET_META && |
| 3608 | reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) { |
| 3609 | /* pkt_data' > pkt_end, pkt_meta' > pkt_data */ |
| 3610 | find_good_pkt_pointers(this_branch, dst_reg, |
| 3611 | dst_reg->type, false); |
| 3612 | } else if ((dst_reg->type == PTR_TO_PACKET_END && |
| 3613 | src_reg->type == PTR_TO_PACKET) || |
| 3614 | (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) && |
| 3615 | src_reg->type == PTR_TO_PACKET_META)) { |
| 3616 | /* pkt_end > pkt_data', pkt_data > pkt_meta' */ |
| 3617 | find_good_pkt_pointers(other_branch, src_reg, |
| 3618 | src_reg->type, true); |
| 3619 | } else { |
| 3620 | return false; |
| 3621 | } |
| 3622 | break; |
| 3623 | case BPF_JLT: |
| 3624 | if ((dst_reg->type == PTR_TO_PACKET && |
| 3625 | src_reg->type == PTR_TO_PACKET_END) || |
| 3626 | (dst_reg->type == PTR_TO_PACKET_META && |
| 3627 | reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) { |
| 3628 | /* pkt_data' < pkt_end, pkt_meta' < pkt_data */ |
| 3629 | find_good_pkt_pointers(other_branch, dst_reg, |
| 3630 | dst_reg->type, true); |
| 3631 | } else if ((dst_reg->type == PTR_TO_PACKET_END && |
| 3632 | src_reg->type == PTR_TO_PACKET) || |
| 3633 | (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) && |
| 3634 | src_reg->type == PTR_TO_PACKET_META)) { |
| 3635 | /* pkt_end < pkt_data', pkt_data > pkt_meta' */ |
| 3636 | find_good_pkt_pointers(this_branch, src_reg, |
| 3637 | src_reg->type, false); |
| 3638 | } else { |
| 3639 | return false; |
| 3640 | } |
| 3641 | break; |
| 3642 | case BPF_JGE: |
| 3643 | if ((dst_reg->type == PTR_TO_PACKET && |
| 3644 | src_reg->type == PTR_TO_PACKET_END) || |
| 3645 | (dst_reg->type == PTR_TO_PACKET_META && |
| 3646 | reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) { |
| 3647 | /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */ |
| 3648 | find_good_pkt_pointers(this_branch, dst_reg, |
| 3649 | dst_reg->type, true); |
| 3650 | } else if ((dst_reg->type == PTR_TO_PACKET_END && |
| 3651 | src_reg->type == PTR_TO_PACKET) || |
| 3652 | (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) && |
| 3653 | src_reg->type == PTR_TO_PACKET_META)) { |
| 3654 | /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */ |
| 3655 | find_good_pkt_pointers(other_branch, src_reg, |
| 3656 | src_reg->type, false); |
| 3657 | } else { |
| 3658 | return false; |
| 3659 | } |
| 3660 | break; |
| 3661 | case BPF_JLE: |
| 3662 | if ((dst_reg->type == PTR_TO_PACKET && |
| 3663 | src_reg->type == PTR_TO_PACKET_END) || |
| 3664 | (dst_reg->type == PTR_TO_PACKET_META && |
| 3665 | reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) { |
| 3666 | /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */ |
| 3667 | find_good_pkt_pointers(other_branch, dst_reg, |
| 3668 | dst_reg->type, false); |
| 3669 | } else if ((dst_reg->type == PTR_TO_PACKET_END && |
| 3670 | src_reg->type == PTR_TO_PACKET) || |
| 3671 | (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) && |
| 3672 | src_reg->type == PTR_TO_PACKET_META)) { |
| 3673 | /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */ |
| 3674 | find_good_pkt_pointers(this_branch, src_reg, |
| 3675 | src_reg->type, true); |
| 3676 | } else { |
| 3677 | return false; |
| 3678 | } |
| 3679 | break; |
| 3680 | default: |
| 3681 | return false; |
| 3682 | } |
| 3683 | |
| 3684 | return true; |
| 3685 | } |
| 3686 | |
| 3687 | static int check_cond_jmp_op(struct bpf_verifier_env *env, |
| 3688 | struct bpf_insn *insn, int *insn_idx) |
| 3689 | { |
| 3690 | struct bpf_verifier_state *this_branch = env->cur_state; |
| 3691 | struct bpf_verifier_state *other_branch; |
| 3692 | struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs; |
| 3693 | struct bpf_reg_state *dst_reg, *other_branch_regs; |
| 3694 | u8 opcode = BPF_OP(insn->code); |
| 3695 | int err; |
| 3696 | |
| 3697 | if (opcode > BPF_JSLE) { |
| 3698 | verbose(env, "invalid BPF_JMP opcode %x\n", opcode); |
| 3699 | return -EINVAL; |
| 3700 | } |
| 3701 | |
| 3702 | if (BPF_SRC(insn->code) == BPF_X) { |
| 3703 | if (insn->imm != 0) { |
| 3704 | verbose(env, "BPF_JMP uses reserved fields\n"); |
| 3705 | return -EINVAL; |
| 3706 | } |
| 3707 | |
| 3708 | /* check src1 operand */ |
| 3709 | err = check_reg_arg(env, insn->src_reg, SRC_OP); |
| 3710 | if (err) |
| 3711 | return err; |
| 3712 | |
| 3713 | if (is_pointer_value(env, insn->src_reg)) { |
| 3714 | verbose(env, "R%d pointer comparison prohibited\n", |
| 3715 | insn->src_reg); |
| 3716 | return -EACCES; |
| 3717 | } |
| 3718 | } else { |
| 3719 | if (insn->src_reg != BPF_REG_0) { |
| 3720 | verbose(env, "BPF_JMP uses reserved fields\n"); |
| 3721 | return -EINVAL; |
| 3722 | } |
| 3723 | } |
| 3724 | |
| 3725 | /* check src2 operand */ |
| 3726 | err = check_reg_arg(env, insn->dst_reg, SRC_OP); |
| 3727 | if (err) |
| 3728 | return err; |
| 3729 | |
| 3730 | dst_reg = ®s[insn->dst_reg]; |
| 3731 | |
| 3732 | /* detect if R == 0 where R was initialized to zero earlier */ |
| 3733 | if (BPF_SRC(insn->code) == BPF_K && |
| 3734 | (opcode == BPF_JEQ || opcode == BPF_JNE) && |
| 3735 | dst_reg->type == SCALAR_VALUE && |
| 3736 | tnum_is_const(dst_reg->var_off)) { |
| 3737 | if ((opcode == BPF_JEQ && dst_reg->var_off.value == insn->imm) || |
| 3738 | (opcode == BPF_JNE && dst_reg->var_off.value != insn->imm)) { |
| 3739 | /* if (imm == imm) goto pc+off; |
| 3740 | * only follow the goto, ignore fall-through |
| 3741 | */ |
| 3742 | *insn_idx += insn->off; |
| 3743 | return 0; |
| 3744 | } else { |
| 3745 | /* if (imm != imm) goto pc+off; |
| 3746 | * only follow fall-through branch, since |
| 3747 | * that's where the program will go |
| 3748 | */ |
| 3749 | return 0; |
| 3750 | } |
| 3751 | } |
| 3752 | |
| 3753 | other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx); |
| 3754 | if (!other_branch) |
| 3755 | return -EFAULT; |
| 3756 | other_branch_regs = other_branch->frame[other_branch->curframe]->regs; |
| 3757 | |
| 3758 | /* detect if we are comparing against a constant value so we can adjust |
| 3759 | * our min/max values for our dst register. |
| 3760 | * this is only legit if both are scalars (or pointers to the same |
| 3761 | * object, I suppose, but we don't support that right now), because |
| 3762 | * otherwise the different base pointers mean the offsets aren't |
| 3763 | * comparable. |
| 3764 | */ |
| 3765 | if (BPF_SRC(insn->code) == BPF_X) { |
| 3766 | if (dst_reg->type == SCALAR_VALUE && |
| 3767 | regs[insn->src_reg].type == SCALAR_VALUE) { |
| 3768 | if (tnum_is_const(regs[insn->src_reg].var_off)) |
| 3769 | reg_set_min_max(&other_branch_regs[insn->dst_reg], |
| 3770 | dst_reg, regs[insn->src_reg].var_off.value, |
| 3771 | opcode); |
| 3772 | else if (tnum_is_const(dst_reg->var_off)) |
| 3773 | reg_set_min_max_inv(&other_branch_regs[insn->src_reg], |
| 3774 | ®s[insn->src_reg], |
| 3775 | dst_reg->var_off.value, opcode); |
| 3776 | else if (opcode == BPF_JEQ || opcode == BPF_JNE) |
| 3777 | /* Comparing for equality, we can combine knowledge */ |
| 3778 | reg_combine_min_max(&other_branch_regs[insn->src_reg], |
| 3779 | &other_branch_regs[insn->dst_reg], |
| 3780 | ®s[insn->src_reg], |
| 3781 | ®s[insn->dst_reg], opcode); |
| 3782 | } |
| 3783 | } else if (dst_reg->type == SCALAR_VALUE) { |
| 3784 | reg_set_min_max(&other_branch_regs[insn->dst_reg], |
| 3785 | dst_reg, insn->imm, opcode); |
| 3786 | } |
| 3787 | |
| 3788 | /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */ |
| 3789 | if (BPF_SRC(insn->code) == BPF_K && |
| 3790 | insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) && |
| 3791 | dst_reg->type == PTR_TO_MAP_VALUE_OR_NULL) { |
| 3792 | /* Mark all identical map registers in each branch as either |
| 3793 | * safe or unknown depending R == 0 or R != 0 conditional. |
| 3794 | */ |
| 3795 | mark_map_regs(this_branch, insn->dst_reg, opcode == BPF_JNE); |
| 3796 | mark_map_regs(other_branch, insn->dst_reg, opcode == BPF_JEQ); |
| 3797 | } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg], |
| 3798 | this_branch, other_branch) && |
| 3799 | is_pointer_value(env, insn->dst_reg)) { |
| 3800 | verbose(env, "R%d pointer comparison prohibited\n", |
| 3801 | insn->dst_reg); |
| 3802 | return -EACCES; |
| 3803 | } |
| 3804 | if (env->log.level) |
| 3805 | print_verifier_state(env, this_branch->frame[this_branch->curframe]); |
| 3806 | return 0; |
| 3807 | } |
| 3808 | |
| 3809 | /* return the map pointer stored inside BPF_LD_IMM64 instruction */ |
| 3810 | static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn) |
| 3811 | { |
| 3812 | u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32; |
| 3813 | |
| 3814 | return (struct bpf_map *) (unsigned long) imm64; |
| 3815 | } |
| 3816 | |
| 3817 | /* verify BPF_LD_IMM64 instruction */ |
| 3818 | static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn) |
| 3819 | { |
| 3820 | struct bpf_reg_state *regs = cur_regs(env); |
| 3821 | int err; |
| 3822 | |
| 3823 | if (BPF_SIZE(insn->code) != BPF_DW) { |
| 3824 | verbose(env, "invalid BPF_LD_IMM insn\n"); |
| 3825 | return -EINVAL; |
| 3826 | } |
| 3827 | if (insn->off != 0) { |
| 3828 | verbose(env, "BPF_LD_IMM64 uses reserved fields\n"); |
| 3829 | return -EINVAL; |
| 3830 | } |
| 3831 | |
| 3832 | err = check_reg_arg(env, insn->dst_reg, DST_OP); |
| 3833 | if (err) |
| 3834 | return err; |
| 3835 | |
| 3836 | if (insn->src_reg == 0) { |
| 3837 | u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm; |
| 3838 | |
| 3839 | regs[insn->dst_reg].type = SCALAR_VALUE; |
| 3840 | __mark_reg_known(®s[insn->dst_reg], imm); |
| 3841 | return 0; |
| 3842 | } |
| 3843 | |
| 3844 | /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */ |
| 3845 | BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD); |
| 3846 | |
| 3847 | regs[insn->dst_reg].type = CONST_PTR_TO_MAP; |
| 3848 | regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn); |
| 3849 | return 0; |
| 3850 | } |
| 3851 | |
| 3852 | static bool may_access_skb(enum bpf_prog_type type) |
| 3853 | { |
| 3854 | switch (type) { |
| 3855 | case BPF_PROG_TYPE_SOCKET_FILTER: |
| 3856 | case BPF_PROG_TYPE_SCHED_CLS: |
| 3857 | case BPF_PROG_TYPE_SCHED_ACT: |
| 3858 | return true; |
| 3859 | default: |
| 3860 | return false; |
| 3861 | } |
| 3862 | } |
| 3863 | |
| 3864 | /* verify safety of LD_ABS|LD_IND instructions: |
| 3865 | * - they can only appear in the programs where ctx == skb |
| 3866 | * - since they are wrappers of function calls, they scratch R1-R5 registers, |
| 3867 | * preserve R6-R9, and store return value into R0 |
| 3868 | * |
| 3869 | * Implicit input: |
| 3870 | * ctx == skb == R6 == CTX |
| 3871 | * |
| 3872 | * Explicit input: |
| 3873 | * SRC == any register |
| 3874 | * IMM == 32-bit immediate |
| 3875 | * |
| 3876 | * Output: |
| 3877 | * R0 - 8/16/32-bit skb data converted to cpu endianness |
| 3878 | */ |
| 3879 | static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn) |
| 3880 | { |
| 3881 | struct bpf_reg_state *regs = cur_regs(env); |
| 3882 | u8 mode = BPF_MODE(insn->code); |
| 3883 | int i, err; |
| 3884 | |
| 3885 | if (!may_access_skb(env->prog->type)) { |
| 3886 | verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n"); |
| 3887 | return -EINVAL; |
| 3888 | } |
| 3889 | |
| 3890 | if (!env->ops->gen_ld_abs) { |
| 3891 | verbose(env, "bpf verifier is misconfigured\n"); |
| 3892 | return -EINVAL; |
| 3893 | } |
| 3894 | |
| 3895 | if (env->subprog_cnt > 1) { |
| 3896 | /* when program has LD_ABS insn JITs and interpreter assume |
| 3897 | * that r1 == ctx == skb which is not the case for callees |
| 3898 | * that can have arbitrary arguments. It's problematic |
| 3899 | * for main prog as well since JITs would need to analyze |
| 3900 | * all functions in order to make proper register save/restore |
| 3901 | * decisions in the main prog. Hence disallow LD_ABS with calls |
| 3902 | */ |
| 3903 | verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n"); |
| 3904 | return -EINVAL; |
| 3905 | } |
| 3906 | |
| 3907 | if (insn->dst_reg != BPF_REG_0 || insn->off != 0 || |
| 3908 | BPF_SIZE(insn->code) == BPF_DW || |
| 3909 | (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) { |
| 3910 | verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n"); |
| 3911 | return -EINVAL; |
| 3912 | } |
| 3913 | |
| 3914 | /* check whether implicit source operand (register R6) is readable */ |
| 3915 | err = check_reg_arg(env, BPF_REG_6, SRC_OP); |
| 3916 | if (err) |
| 3917 | return err; |
| 3918 | |
| 3919 | if (regs[BPF_REG_6].type != PTR_TO_CTX) { |
| 3920 | verbose(env, |
| 3921 | "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n"); |
| 3922 | return -EINVAL; |
| 3923 | } |
| 3924 | |
| 3925 | if (mode == BPF_IND) { |
| 3926 | /* check explicit source operand */ |
| 3927 | err = check_reg_arg(env, insn->src_reg, SRC_OP); |
| 3928 | if (err) |
| 3929 | return err; |
| 3930 | } |
| 3931 | |
| 3932 | /* reset caller saved regs to unreadable */ |
| 3933 | for (i = 0; i < CALLER_SAVED_REGS; i++) { |
| 3934 | mark_reg_not_init(env, regs, caller_saved[i]); |
| 3935 | check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK); |
| 3936 | } |
| 3937 | |
| 3938 | /* mark destination R0 register as readable, since it contains |
| 3939 | * the value fetched from the packet. |
| 3940 | * Already marked as written above. |
| 3941 | */ |
| 3942 | mark_reg_unknown(env, regs, BPF_REG_0); |
| 3943 | return 0; |
| 3944 | } |
| 3945 | |
| 3946 | static int check_return_code(struct bpf_verifier_env *env) |
| 3947 | { |
| 3948 | struct bpf_reg_state *reg; |
| 3949 | struct tnum range = tnum_range(0, 1); |
| 3950 | |
| 3951 | switch (env->prog->type) { |
| 3952 | case BPF_PROG_TYPE_CGROUP_SKB: |
| 3953 | case BPF_PROG_TYPE_CGROUP_SOCK: |
| 3954 | case BPF_PROG_TYPE_CGROUP_SOCK_ADDR: |
| 3955 | case BPF_PROG_TYPE_SOCK_OPS: |
| 3956 | case BPF_PROG_TYPE_CGROUP_DEVICE: |
| 3957 | break; |
| 3958 | default: |
| 3959 | return 0; |
| 3960 | } |
| 3961 | |
| 3962 | reg = cur_regs(env) + BPF_REG_0; |
| 3963 | if (reg->type != SCALAR_VALUE) { |
| 3964 | verbose(env, "At program exit the register R0 is not a known value (%s)\n", |
| 3965 | reg_type_str[reg->type]); |
| 3966 | return -EINVAL; |
| 3967 | } |
| 3968 | |
| 3969 | if (!tnum_in(range, reg->var_off)) { |
| 3970 | verbose(env, "At program exit the register R0 "); |
| 3971 | if (!tnum_is_unknown(reg->var_off)) { |
| 3972 | char tn_buf[48]; |
| 3973 | |
| 3974 | tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off); |
| 3975 | verbose(env, "has value %s", tn_buf); |
| 3976 | } else { |
| 3977 | verbose(env, "has unknown scalar value"); |
| 3978 | } |
| 3979 | verbose(env, " should have been 0 or 1\n"); |
| 3980 | return -EINVAL; |
| 3981 | } |
| 3982 | return 0; |
| 3983 | } |
| 3984 | |
| 3985 | /* non-recursive DFS pseudo code |
| 3986 | * 1 procedure DFS-iterative(G,v): |
| 3987 | * 2 label v as discovered |
| 3988 | * 3 let S be a stack |
| 3989 | * 4 S.push(v) |
| 3990 | * 5 while S is not empty |
| 3991 | * 6 t <- S.pop() |
| 3992 | * 7 if t is what we're looking for: |
| 3993 | * 8 return t |
| 3994 | * 9 for all edges e in G.adjacentEdges(t) do |
| 3995 | * 10 if edge e is already labelled |
| 3996 | * 11 continue with the next edge |
| 3997 | * 12 w <- G.adjacentVertex(t,e) |
| 3998 | * 13 if vertex w is not discovered and not explored |
| 3999 | * 14 label e as tree-edge |
| 4000 | * 15 label w as discovered |
| 4001 | * 16 S.push(w) |
| 4002 | * 17 continue at 5 |
| 4003 | * 18 else if vertex w is discovered |
| 4004 | * 19 label e as back-edge |
| 4005 | * 20 else |
| 4006 | * 21 // vertex w is explored |
| 4007 | * 22 label e as forward- or cross-edge |
| 4008 | * 23 label t as explored |
| 4009 | * 24 S.pop() |
| 4010 | * |
| 4011 | * convention: |
| 4012 | * 0x10 - discovered |
| 4013 | * 0x11 - discovered and fall-through edge labelled |
| 4014 | * 0x12 - discovered and fall-through and branch edges labelled |
| 4015 | * 0x20 - explored |
| 4016 | */ |
| 4017 | |
| 4018 | enum { |
| 4019 | DISCOVERED = 0x10, |
| 4020 | EXPLORED = 0x20, |
| 4021 | FALLTHROUGH = 1, |
| 4022 | BRANCH = 2, |
| 4023 | }; |
| 4024 | |
| 4025 | #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L) |
| 4026 | |
| 4027 | static int *insn_stack; /* stack of insns to process */ |
| 4028 | static int cur_stack; /* current stack index */ |
| 4029 | static int *insn_state; |
| 4030 | |
| 4031 | /* t, w, e - match pseudo-code above: |
| 4032 | * t - index of current instruction |
| 4033 | * w - next instruction |
| 4034 | * e - edge |
| 4035 | */ |
| 4036 | static int push_insn(int t, int w, int e, struct bpf_verifier_env *env) |
| 4037 | { |
| 4038 | if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH)) |
| 4039 | return 0; |
| 4040 | |
| 4041 | if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH)) |
| 4042 | return 0; |
| 4043 | |
| 4044 | if (w < 0 || w >= env->prog->len) { |
| 4045 | verbose(env, "jump out of range from insn %d to %d\n", t, w); |
| 4046 | return -EINVAL; |
| 4047 | } |
| 4048 | |
| 4049 | if (e == BRANCH) |
| 4050 | /* mark branch target for state pruning */ |
| 4051 | env->explored_states[w] = STATE_LIST_MARK; |
| 4052 | |
| 4053 | if (insn_state[w] == 0) { |
| 4054 | /* tree-edge */ |
| 4055 | insn_state[t] = DISCOVERED | e; |
| 4056 | insn_state[w] = DISCOVERED; |
| 4057 | if (cur_stack >= env->prog->len) |
| 4058 | return -E2BIG; |
| 4059 | insn_stack[cur_stack++] = w; |
| 4060 | return 1; |
| 4061 | } else if ((insn_state[w] & 0xF0) == DISCOVERED) { |
| 4062 | verbose(env, "back-edge from insn %d to %d\n", t, w); |
| 4063 | return -EINVAL; |
| 4064 | } else if (insn_state[w] == EXPLORED) { |
| 4065 | /* forward- or cross-edge */ |
| 4066 | insn_state[t] = DISCOVERED | e; |
| 4067 | } else { |
| 4068 | verbose(env, "insn state internal bug\n"); |
| 4069 | return -EFAULT; |
| 4070 | } |
| 4071 | return 0; |
| 4072 | } |
| 4073 | |
| 4074 | /* non-recursive depth-first-search to detect loops in BPF program |
| 4075 | * loop == back-edge in directed graph |
| 4076 | */ |
| 4077 | static int check_cfg(struct bpf_verifier_env *env) |
| 4078 | { |
| 4079 | struct bpf_insn *insns = env->prog->insnsi; |
| 4080 | int insn_cnt = env->prog->len; |
| 4081 | int ret = 0; |
| 4082 | int i, t; |
| 4083 | |
| 4084 | ret = check_subprogs(env); |
| 4085 | if (ret < 0) |
| 4086 | return ret; |
| 4087 | |
| 4088 | insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL); |
| 4089 | if (!insn_state) |
| 4090 | return -ENOMEM; |
| 4091 | |
| 4092 | insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL); |
| 4093 | if (!insn_stack) { |
| 4094 | kfree(insn_state); |
| 4095 | return -ENOMEM; |
| 4096 | } |
| 4097 | |
| 4098 | insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */ |
| 4099 | insn_stack[0] = 0; /* 0 is the first instruction */ |
| 4100 | cur_stack = 1; |
| 4101 | |
| 4102 | peek_stack: |
| 4103 | if (cur_stack == 0) |
| 4104 | goto check_state; |
| 4105 | t = insn_stack[cur_stack - 1]; |
| 4106 | |
| 4107 | if (BPF_CLASS(insns[t].code) == BPF_JMP) { |
| 4108 | u8 opcode = BPF_OP(insns[t].code); |
| 4109 | |
| 4110 | if (opcode == BPF_EXIT) { |
| 4111 | goto mark_explored; |
| 4112 | } else if (opcode == BPF_CALL) { |
| 4113 | ret = push_insn(t, t + 1, FALLTHROUGH, env); |
| 4114 | if (ret == 1) |
| 4115 | goto peek_stack; |
| 4116 | else if (ret < 0) |
| 4117 | goto err_free; |
| 4118 | if (t + 1 < insn_cnt) |
| 4119 | env->explored_states[t + 1] = STATE_LIST_MARK; |
| 4120 | if (insns[t].src_reg == BPF_PSEUDO_CALL) { |
| 4121 | env->explored_states[t] = STATE_LIST_MARK; |
| 4122 | ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env); |
| 4123 | if (ret == 1) |
| 4124 | goto peek_stack; |
| 4125 | else if (ret < 0) |
| 4126 | goto err_free; |
| 4127 | } |
| 4128 | } else if (opcode == BPF_JA) { |
| 4129 | if (BPF_SRC(insns[t].code) != BPF_K) { |
| 4130 | ret = -EINVAL; |
| 4131 | goto err_free; |
| 4132 | } |
| 4133 | /* unconditional jump with single edge */ |
| 4134 | ret = push_insn(t, t + insns[t].off + 1, |
| 4135 | FALLTHROUGH, env); |
| 4136 | if (ret == 1) |
| 4137 | goto peek_stack; |
| 4138 | else if (ret < 0) |
| 4139 | goto err_free; |
| 4140 | /* tell verifier to check for equivalent states |
| 4141 | * after every call and jump |
| 4142 | */ |
| 4143 | if (t + 1 < insn_cnt) |
| 4144 | env->explored_states[t + 1] = STATE_LIST_MARK; |
| 4145 | } else { |
| 4146 | /* conditional jump with two edges */ |
| 4147 | env->explored_states[t] = STATE_LIST_MARK; |
| 4148 | ret = push_insn(t, t + 1, FALLTHROUGH, env); |
| 4149 | if (ret == 1) |
| 4150 | goto peek_stack; |
| 4151 | else if (ret < 0) |
| 4152 | goto err_free; |
| 4153 | |
| 4154 | ret = push_insn(t, t + insns[t].off + 1, BRANCH, env); |
| 4155 | if (ret == 1) |
| 4156 | goto peek_stack; |
| 4157 | else if (ret < 0) |
| 4158 | goto err_free; |
| 4159 | } |
| 4160 | } else { |
| 4161 | /* all other non-branch instructions with single |
| 4162 | * fall-through edge |
| 4163 | */ |
| 4164 | ret = push_insn(t, t + 1, FALLTHROUGH, env); |
| 4165 | if (ret == 1) |
| 4166 | goto peek_stack; |
| 4167 | else if (ret < 0) |
| 4168 | goto err_free; |
| 4169 | } |
| 4170 | |
| 4171 | mark_explored: |
| 4172 | insn_state[t] = EXPLORED; |
| 4173 | if (cur_stack-- <= 0) { |
| 4174 | verbose(env, "pop stack internal bug\n"); |
| 4175 | ret = -EFAULT; |
| 4176 | goto err_free; |
| 4177 | } |
| 4178 | goto peek_stack; |
| 4179 | |
| 4180 | check_state: |
| 4181 | for (i = 0; i < insn_cnt; i++) { |
| 4182 | if (insn_state[i] != EXPLORED) { |
| 4183 | verbose(env, "unreachable insn %d\n", i); |
| 4184 | ret = -EINVAL; |
| 4185 | goto err_free; |
| 4186 | } |
| 4187 | } |
| 4188 | ret = 0; /* cfg looks good */ |
| 4189 | |
| 4190 | err_free: |
| 4191 | kfree(insn_state); |
| 4192 | kfree(insn_stack); |
| 4193 | return ret; |
| 4194 | } |
| 4195 | |
| 4196 | /* check %cur's range satisfies %old's */ |
| 4197 | static bool range_within(struct bpf_reg_state *old, |
| 4198 | struct bpf_reg_state *cur) |
| 4199 | { |
| 4200 | return old->umin_value <= cur->umin_value && |
| 4201 | old->umax_value >= cur->umax_value && |
| 4202 | old->smin_value <= cur->smin_value && |
| 4203 | old->smax_value >= cur->smax_value; |
| 4204 | } |
| 4205 | |
| 4206 | /* Maximum number of register states that can exist at once */ |
| 4207 | #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE) |
| 4208 | struct idpair { |
| 4209 | u32 old; |
| 4210 | u32 cur; |
| 4211 | }; |
| 4212 | |
| 4213 | /* If in the old state two registers had the same id, then they need to have |
| 4214 | * the same id in the new state as well. But that id could be different from |
| 4215 | * the old state, so we need to track the mapping from old to new ids. |
| 4216 | * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent |
| 4217 | * regs with old id 5 must also have new id 9 for the new state to be safe. But |
| 4218 | * regs with a different old id could still have new id 9, we don't care about |
| 4219 | * that. |
| 4220 | * So we look through our idmap to see if this old id has been seen before. If |
| 4221 | * so, we require the new id to match; otherwise, we add the id pair to the map. |
| 4222 | */ |
| 4223 | static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap) |
| 4224 | { |
| 4225 | unsigned int i; |
| 4226 | |
| 4227 | for (i = 0; i < ID_MAP_SIZE; i++) { |
| 4228 | if (!idmap[i].old) { |
| 4229 | /* Reached an empty slot; haven't seen this id before */ |
| 4230 | idmap[i].old = old_id; |
| 4231 | idmap[i].cur = cur_id; |
| 4232 | return true; |
| 4233 | } |
| 4234 | if (idmap[i].old == old_id) |
| 4235 | return idmap[i].cur == cur_id; |
| 4236 | } |
| 4237 | /* We ran out of idmap slots, which should be impossible */ |
| 4238 | WARN_ON_ONCE(1); |
| 4239 | return false; |
| 4240 | } |
| 4241 | |
| 4242 | /* Returns true if (rold safe implies rcur safe) */ |
| 4243 | static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur, |
| 4244 | struct idpair *idmap) |
| 4245 | { |
| 4246 | bool equal; |
| 4247 | |
| 4248 | if (!(rold->live & REG_LIVE_READ)) |
| 4249 | /* explored state didn't use this */ |
| 4250 | return true; |
| 4251 | |
| 4252 | equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, frameno)) == 0; |
| 4253 | |
| 4254 | if (rold->type == PTR_TO_STACK) |
| 4255 | /* two stack pointers are equal only if they're pointing to |
| 4256 | * the same stack frame, since fp-8 in foo != fp-8 in bar |
| 4257 | */ |
| 4258 | return equal && rold->frameno == rcur->frameno; |
| 4259 | |
| 4260 | if (equal) |
| 4261 | return true; |
| 4262 | |
| 4263 | if (rold->type == NOT_INIT) |
| 4264 | /* explored state can't have used this */ |
| 4265 | return true; |
| 4266 | if (rcur->type == NOT_INIT) |
| 4267 | return false; |
| 4268 | switch (rold->type) { |
| 4269 | case SCALAR_VALUE: |
| 4270 | if (rcur->type == SCALAR_VALUE) { |
| 4271 | /* new val must satisfy old val knowledge */ |
| 4272 | return range_within(rold, rcur) && |
| 4273 | tnum_in(rold->var_off, rcur->var_off); |
| 4274 | } else { |
| 4275 | /* We're trying to use a pointer in place of a scalar. |
| 4276 | * Even if the scalar was unbounded, this could lead to |
| 4277 | * pointer leaks because scalars are allowed to leak |
| 4278 | * while pointers are not. We could make this safe in |
| 4279 | * special cases if root is calling us, but it's |
| 4280 | * probably not worth the hassle. |
| 4281 | */ |
| 4282 | return false; |
| 4283 | } |
| 4284 | case PTR_TO_MAP_VALUE: |
| 4285 | /* If the new min/max/var_off satisfy the old ones and |
| 4286 | * everything else matches, we are OK. |
| 4287 | * We don't care about the 'id' value, because nothing |
| 4288 | * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL) |
| 4289 | */ |
| 4290 | return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 && |
| 4291 | range_within(rold, rcur) && |
| 4292 | tnum_in(rold->var_off, rcur->var_off); |
| 4293 | case PTR_TO_MAP_VALUE_OR_NULL: |
| 4294 | /* a PTR_TO_MAP_VALUE could be safe to use as a |
| 4295 | * PTR_TO_MAP_VALUE_OR_NULL into the same map. |
| 4296 | * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL- |
| 4297 | * checked, doing so could have affected others with the same |
| 4298 | * id, and we can't check for that because we lost the id when |
| 4299 | * we converted to a PTR_TO_MAP_VALUE. |
| 4300 | */ |
| 4301 | if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL) |
| 4302 | return false; |
| 4303 | if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id))) |
| 4304 | return false; |
| 4305 | /* Check our ids match any regs they're supposed to */ |
| 4306 | return check_ids(rold->id, rcur->id, idmap); |
| 4307 | case PTR_TO_PACKET_META: |
| 4308 | case PTR_TO_PACKET: |
| 4309 | if (rcur->type != rold->type) |
| 4310 | return false; |
| 4311 | /* We must have at least as much range as the old ptr |
| 4312 | * did, so that any accesses which were safe before are |
| 4313 | * still safe. This is true even if old range < old off, |
| 4314 | * since someone could have accessed through (ptr - k), or |
| 4315 | * even done ptr -= k in a register, to get a safe access. |
| 4316 | */ |
| 4317 | if (rold->range > rcur->range) |
| 4318 | return false; |
| 4319 | /* If the offsets don't match, we can't trust our alignment; |
| 4320 | * nor can we be sure that we won't fall out of range. |
| 4321 | */ |
| 4322 | if (rold->off != rcur->off) |
| 4323 | return false; |
| 4324 | /* id relations must be preserved */ |
| 4325 | if (rold->id && !check_ids(rold->id, rcur->id, idmap)) |
| 4326 | return false; |
| 4327 | /* new val must satisfy old val knowledge */ |
| 4328 | return range_within(rold, rcur) && |
| 4329 | tnum_in(rold->var_off, rcur->var_off); |
| 4330 | case PTR_TO_CTX: |
| 4331 | case CONST_PTR_TO_MAP: |
| 4332 | case PTR_TO_PACKET_END: |
| 4333 | /* Only valid matches are exact, which memcmp() above |
| 4334 | * would have accepted |
| 4335 | */ |
| 4336 | default: |
| 4337 | /* Don't know what's going on, just say it's not safe */ |
| 4338 | return false; |
| 4339 | } |
| 4340 | |
| 4341 | /* Shouldn't get here; if we do, say it's not safe */ |
| 4342 | WARN_ON_ONCE(1); |
| 4343 | return false; |
| 4344 | } |
| 4345 | |
| 4346 | static bool stacksafe(struct bpf_func_state *old, |
| 4347 | struct bpf_func_state *cur, |
| 4348 | struct idpair *idmap) |
| 4349 | { |
| 4350 | int i, spi; |
| 4351 | |
| 4352 | /* if explored stack has more populated slots than current stack |
| 4353 | * such stacks are not equivalent |
| 4354 | */ |
| 4355 | if (old->allocated_stack > cur->allocated_stack) |
| 4356 | return false; |
| 4357 | |
| 4358 | /* walk slots of the explored stack and ignore any additional |
| 4359 | * slots in the current stack, since explored(safe) state |
| 4360 | * didn't use them |
| 4361 | */ |
| 4362 | for (i = 0; i < old->allocated_stack; i++) { |
| 4363 | spi = i / BPF_REG_SIZE; |
| 4364 | |
| 4365 | if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) |
| 4366 | /* explored state didn't use this */ |
| 4367 | continue; |
| 4368 | |
| 4369 | if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID) |
| 4370 | continue; |
| 4371 | /* if old state was safe with misc data in the stack |
| 4372 | * it will be safe with zero-initialized stack. |
| 4373 | * The opposite is not true |
| 4374 | */ |
| 4375 | if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC && |
| 4376 | cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO) |
| 4377 | continue; |
| 4378 | if (old->stack[spi].slot_type[i % BPF_REG_SIZE] != |
| 4379 | cur->stack[spi].slot_type[i % BPF_REG_SIZE]) |
| 4380 | /* Ex: old explored (safe) state has STACK_SPILL in |
| 4381 | * this stack slot, but current has has STACK_MISC -> |
| 4382 | * this verifier states are not equivalent, |
| 4383 | * return false to continue verification of this path |
| 4384 | */ |
| 4385 | return false; |
| 4386 | if (i % BPF_REG_SIZE) |
| 4387 | continue; |
| 4388 | if (old->stack[spi].slot_type[0] != STACK_SPILL) |
| 4389 | continue; |
| 4390 | if (!regsafe(&old->stack[spi].spilled_ptr, |
| 4391 | &cur->stack[spi].spilled_ptr, |
| 4392 | idmap)) |
| 4393 | /* when explored and current stack slot are both storing |
| 4394 | * spilled registers, check that stored pointers types |
| 4395 | * are the same as well. |
| 4396 | * Ex: explored safe path could have stored |
| 4397 | * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8} |
| 4398 | * but current path has stored: |
| 4399 | * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16} |
| 4400 | * such verifier states are not equivalent. |
| 4401 | * return false to continue verification of this path |
| 4402 | */ |
| 4403 | return false; |
| 4404 | } |
| 4405 | return true; |
| 4406 | } |
| 4407 | |
| 4408 | /* compare two verifier states |
| 4409 | * |
| 4410 | * all states stored in state_list are known to be valid, since |
| 4411 | * verifier reached 'bpf_exit' instruction through them |
| 4412 | * |
| 4413 | * this function is called when verifier exploring different branches of |
| 4414 | * execution popped from the state stack. If it sees an old state that has |
| 4415 | * more strict register state and more strict stack state then this execution |
| 4416 | * branch doesn't need to be explored further, since verifier already |
| 4417 | * concluded that more strict state leads to valid finish. |
| 4418 | * |
| 4419 | * Therefore two states are equivalent if register state is more conservative |
| 4420 | * and explored stack state is more conservative than the current one. |
| 4421 | * Example: |
| 4422 | * explored current |
| 4423 | * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC) |
| 4424 | * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC) |
| 4425 | * |
| 4426 | * In other words if current stack state (one being explored) has more |
| 4427 | * valid slots than old one that already passed validation, it means |
| 4428 | * the verifier can stop exploring and conclude that current state is valid too |
| 4429 | * |
| 4430 | * Similarly with registers. If explored state has register type as invalid |
| 4431 | * whereas register type in current state is meaningful, it means that |
| 4432 | * the current state will reach 'bpf_exit' instruction safely |
| 4433 | */ |
| 4434 | static bool func_states_equal(struct bpf_func_state *old, |
| 4435 | struct bpf_func_state *cur) |
| 4436 | { |
| 4437 | struct idpair *idmap; |
| 4438 | bool ret = false; |
| 4439 | int i; |
| 4440 | |
| 4441 | idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL); |
| 4442 | /* If we failed to allocate the idmap, just say it's not safe */ |
| 4443 | if (!idmap) |
| 4444 | return false; |
| 4445 | |
| 4446 | for (i = 0; i < MAX_BPF_REG; i++) { |
| 4447 | if (!regsafe(&old->regs[i], &cur->regs[i], idmap)) |
| 4448 | goto out_free; |
| 4449 | } |
| 4450 | |
| 4451 | if (!stacksafe(old, cur, idmap)) |
| 4452 | goto out_free; |
| 4453 | ret = true; |
| 4454 | out_free: |
| 4455 | kfree(idmap); |
| 4456 | return ret; |
| 4457 | } |
| 4458 | |
| 4459 | static bool states_equal(struct bpf_verifier_env *env, |
| 4460 | struct bpf_verifier_state *old, |
| 4461 | struct bpf_verifier_state *cur) |
| 4462 | { |
| 4463 | int i; |
| 4464 | |
| 4465 | if (old->curframe != cur->curframe) |
| 4466 | return false; |
| 4467 | |
| 4468 | /* for states to be equal callsites have to be the same |
| 4469 | * and all frame states need to be equivalent |
| 4470 | */ |
| 4471 | for (i = 0; i <= old->curframe; i++) { |
| 4472 | if (old->frame[i]->callsite != cur->frame[i]->callsite) |
| 4473 | return false; |
| 4474 | if (!func_states_equal(old->frame[i], cur->frame[i])) |
| 4475 | return false; |
| 4476 | } |
| 4477 | return true; |
| 4478 | } |
| 4479 | |
| 4480 | /* A write screens off any subsequent reads; but write marks come from the |
| 4481 | * straight-line code between a state and its parent. When we arrive at an |
| 4482 | * equivalent state (jump target or such) we didn't arrive by the straight-line |
| 4483 | * code, so read marks in the state must propagate to the parent regardless |
| 4484 | * of the state's write marks. That's what 'parent == state->parent' comparison |
| 4485 | * in mark_reg_read() and mark_stack_slot_read() is for. |
| 4486 | */ |
| 4487 | static int propagate_liveness(struct bpf_verifier_env *env, |
| 4488 | const struct bpf_verifier_state *vstate, |
| 4489 | struct bpf_verifier_state *vparent) |
| 4490 | { |
| 4491 | int i, frame, err = 0; |
| 4492 | struct bpf_func_state *state, *parent; |
| 4493 | |
| 4494 | if (vparent->curframe != vstate->curframe) { |
| 4495 | WARN(1, "propagate_live: parent frame %d current frame %d\n", |
| 4496 | vparent->curframe, vstate->curframe); |
| 4497 | return -EFAULT; |
| 4498 | } |
| 4499 | /* Propagate read liveness of registers... */ |
| 4500 | BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG); |
| 4501 | /* We don't need to worry about FP liveness because it's read-only */ |
| 4502 | for (i = 0; i < BPF_REG_FP; i++) { |
| 4503 | if (vparent->frame[vparent->curframe]->regs[i].live & REG_LIVE_READ) |
| 4504 | continue; |
| 4505 | if (vstate->frame[vstate->curframe]->regs[i].live & REG_LIVE_READ) { |
| 4506 | err = mark_reg_read(env, vstate, vparent, i); |
| 4507 | if (err) |
| 4508 | return err; |
| 4509 | } |
| 4510 | } |
| 4511 | |
| 4512 | /* ... and stack slots */ |
| 4513 | for (frame = 0; frame <= vstate->curframe; frame++) { |
| 4514 | state = vstate->frame[frame]; |
| 4515 | parent = vparent->frame[frame]; |
| 4516 | for (i = 0; i < state->allocated_stack / BPF_REG_SIZE && |
| 4517 | i < parent->allocated_stack / BPF_REG_SIZE; i++) { |
| 4518 | if (parent->stack[i].spilled_ptr.live & REG_LIVE_READ) |
| 4519 | continue; |
| 4520 | if (state->stack[i].spilled_ptr.live & REG_LIVE_READ) |
| 4521 | mark_stack_slot_read(env, vstate, vparent, i, frame); |
| 4522 | } |
| 4523 | } |
| 4524 | return err; |
| 4525 | } |
| 4526 | |
| 4527 | static int is_state_visited(struct bpf_verifier_env *env, int insn_idx) |
| 4528 | { |
| 4529 | struct bpf_verifier_state_list *new_sl; |
| 4530 | struct bpf_verifier_state_list *sl; |
| 4531 | struct bpf_verifier_state *cur = env->cur_state; |
| 4532 | int i, j, err; |
| 4533 | |
| 4534 | sl = env->explored_states[insn_idx]; |
| 4535 | if (!sl) |
| 4536 | /* this 'insn_idx' instruction wasn't marked, so we will not |
| 4537 | * be doing state search here |
| 4538 | */ |
| 4539 | return 0; |
| 4540 | |
| 4541 | while (sl != STATE_LIST_MARK) { |
| 4542 | if (states_equal(env, &sl->state, cur)) { |
| 4543 | /* reached equivalent register/stack state, |
| 4544 | * prune the search. |
| 4545 | * Registers read by the continuation are read by us. |
| 4546 | * If we have any write marks in env->cur_state, they |
| 4547 | * will prevent corresponding reads in the continuation |
| 4548 | * from reaching our parent (an explored_state). Our |
| 4549 | * own state will get the read marks recorded, but |
| 4550 | * they'll be immediately forgotten as we're pruning |
| 4551 | * this state and will pop a new one. |
| 4552 | */ |
| 4553 | err = propagate_liveness(env, &sl->state, cur); |
| 4554 | if (err) |
| 4555 | return err; |
| 4556 | return 1; |
| 4557 | } |
| 4558 | sl = sl->next; |
| 4559 | } |
| 4560 | |
| 4561 | /* there were no equivalent states, remember current one. |
| 4562 | * technically the current state is not proven to be safe yet, |
| 4563 | * but it will either reach outer most bpf_exit (which means it's safe) |
| 4564 | * or it will be rejected. Since there are no loops, we won't be |
| 4565 | * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx) |
| 4566 | * again on the way to bpf_exit |
| 4567 | */ |
| 4568 | new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL); |
| 4569 | if (!new_sl) |
| 4570 | return -ENOMEM; |
| 4571 | |
| 4572 | /* add new state to the head of linked list */ |
| 4573 | err = copy_verifier_state(&new_sl->state, cur); |
| 4574 | if (err) { |
| 4575 | free_verifier_state(&new_sl->state, false); |
| 4576 | kfree(new_sl); |
| 4577 | return err; |
| 4578 | } |
| 4579 | new_sl->next = env->explored_states[insn_idx]; |
| 4580 | env->explored_states[insn_idx] = new_sl; |
| 4581 | /* connect new state to parentage chain */ |
| 4582 | cur->parent = &new_sl->state; |
| 4583 | /* clear write marks in current state: the writes we did are not writes |
| 4584 | * our child did, so they don't screen off its reads from us. |
| 4585 | * (There are no read marks in current state, because reads always mark |
| 4586 | * their parent and current state never has children yet. Only |
| 4587 | * explored_states can get read marks.) |
| 4588 | */ |
| 4589 | for (i = 0; i < BPF_REG_FP; i++) |
| 4590 | cur->frame[cur->curframe]->regs[i].live = REG_LIVE_NONE; |
| 4591 | |
| 4592 | /* all stack frames are accessible from callee, clear them all */ |
| 4593 | for (j = 0; j <= cur->curframe; j++) { |
| 4594 | struct bpf_func_state *frame = cur->frame[j]; |
| 4595 | |
| 4596 | for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) |
| 4597 | frame->stack[i].spilled_ptr.live = REG_LIVE_NONE; |
| 4598 | } |
| 4599 | return 0; |
| 4600 | } |
| 4601 | |
| 4602 | static int do_check(struct bpf_verifier_env *env) |
| 4603 | { |
| 4604 | struct bpf_verifier_state *state; |
| 4605 | struct bpf_insn *insns = env->prog->insnsi; |
| 4606 | struct bpf_reg_state *regs; |
| 4607 | int insn_cnt = env->prog->len, i; |
| 4608 | int insn_idx, prev_insn_idx = 0; |
| 4609 | int insn_processed = 0; |
| 4610 | bool do_print_state = false; |
| 4611 | |
| 4612 | state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL); |
| 4613 | if (!state) |
| 4614 | return -ENOMEM; |
| 4615 | state->curframe = 0; |
| 4616 | state->parent = NULL; |
| 4617 | state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL); |
| 4618 | if (!state->frame[0]) { |
| 4619 | kfree(state); |
| 4620 | return -ENOMEM; |
| 4621 | } |
| 4622 | env->cur_state = state; |
| 4623 | init_func_state(env, state->frame[0], |
| 4624 | BPF_MAIN_FUNC /* callsite */, |
| 4625 | 0 /* frameno */, |
| 4626 | 0 /* subprogno, zero == main subprog */); |
| 4627 | insn_idx = 0; |
| 4628 | for (;;) { |
| 4629 | struct bpf_insn *insn; |
| 4630 | u8 class; |
| 4631 | int err; |
| 4632 | |
| 4633 | if (insn_idx >= insn_cnt) { |
| 4634 | verbose(env, "invalid insn idx %d insn_cnt %d\n", |
| 4635 | insn_idx, insn_cnt); |
| 4636 | return -EFAULT; |
| 4637 | } |
| 4638 | |
| 4639 | insn = &insns[insn_idx]; |
| 4640 | class = BPF_CLASS(insn->code); |
| 4641 | |
| 4642 | if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) { |
| 4643 | verbose(env, |
| 4644 | "BPF program is too large. Processed %d insn\n", |
| 4645 | insn_processed); |
| 4646 | return -E2BIG; |
| 4647 | } |
| 4648 | |
| 4649 | err = is_state_visited(env, insn_idx); |
| 4650 | if (err < 0) |
| 4651 | return err; |
| 4652 | if (err == 1) { |
| 4653 | /* found equivalent state, can prune the search */ |
| 4654 | if (env->log.level) { |
| 4655 | if (do_print_state) |
| 4656 | verbose(env, "\nfrom %d to %d: safe\n", |
| 4657 | prev_insn_idx, insn_idx); |
| 4658 | else |
| 4659 | verbose(env, "%d: safe\n", insn_idx); |
| 4660 | } |
| 4661 | goto process_bpf_exit; |
| 4662 | } |
| 4663 | |
| 4664 | if (need_resched()) |
| 4665 | cond_resched(); |
| 4666 | |
| 4667 | if (env->log.level > 1 || (env->log.level && do_print_state)) { |
| 4668 | if (env->log.level > 1) |
| 4669 | verbose(env, "%d:", insn_idx); |
| 4670 | else |
| 4671 | verbose(env, "\nfrom %d to %d:", |
| 4672 | prev_insn_idx, insn_idx); |
| 4673 | print_verifier_state(env, state->frame[state->curframe]); |
| 4674 | do_print_state = false; |
| 4675 | } |
| 4676 | |
| 4677 | if (env->log.level) { |
| 4678 | const struct bpf_insn_cbs cbs = { |
| 4679 | .cb_print = verbose, |
| 4680 | .private_data = env, |
| 4681 | }; |
| 4682 | |
| 4683 | verbose(env, "%d: ", insn_idx); |
| 4684 | print_bpf_insn(&cbs, insn, env->allow_ptr_leaks); |
| 4685 | } |
| 4686 | |
| 4687 | if (bpf_prog_is_dev_bound(env->prog->aux)) { |
| 4688 | err = bpf_prog_offload_verify_insn(env, insn_idx, |
| 4689 | prev_insn_idx); |
| 4690 | if (err) |
| 4691 | return err; |
| 4692 | } |
| 4693 | |
| 4694 | regs = cur_regs(env); |
| 4695 | env->insn_aux_data[insn_idx].seen = true; |
| 4696 | if (class == BPF_ALU || class == BPF_ALU64) { |
| 4697 | err = check_alu_op(env, insn); |
| 4698 | if (err) |
| 4699 | return err; |
| 4700 | |
| 4701 | } else if (class == BPF_LDX) { |
| 4702 | enum bpf_reg_type *prev_src_type, src_reg_type; |
| 4703 | |
| 4704 | /* check for reserved fields is already done */ |
| 4705 | |
| 4706 | /* check src operand */ |
| 4707 | err = check_reg_arg(env, insn->src_reg, SRC_OP); |
| 4708 | if (err) |
| 4709 | return err; |
| 4710 | |
| 4711 | err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK); |
| 4712 | if (err) |
| 4713 | return err; |
| 4714 | |
| 4715 | src_reg_type = regs[insn->src_reg].type; |
| 4716 | |
| 4717 | /* check that memory (src_reg + off) is readable, |
| 4718 | * the state of dst_reg will be updated by this func |
| 4719 | */ |
| 4720 | err = check_mem_access(env, insn_idx, insn->src_reg, insn->off, |
| 4721 | BPF_SIZE(insn->code), BPF_READ, |
| 4722 | insn->dst_reg, false); |
| 4723 | if (err) |
| 4724 | return err; |
| 4725 | |
| 4726 | prev_src_type = &env->insn_aux_data[insn_idx].ptr_type; |
| 4727 | |
| 4728 | if (*prev_src_type == NOT_INIT) { |
| 4729 | /* saw a valid insn |
| 4730 | * dst_reg = *(u32 *)(src_reg + off) |
| 4731 | * save type to validate intersecting paths |
| 4732 | */ |
| 4733 | *prev_src_type = src_reg_type; |
| 4734 | |
| 4735 | } else if (src_reg_type != *prev_src_type && |
| 4736 | (src_reg_type == PTR_TO_CTX || |
| 4737 | *prev_src_type == PTR_TO_CTX)) { |
| 4738 | /* ABuser program is trying to use the same insn |
| 4739 | * dst_reg = *(u32*) (src_reg + off) |
| 4740 | * with different pointer types: |
| 4741 | * src_reg == ctx in one branch and |
| 4742 | * src_reg == stack|map in some other branch. |
| 4743 | * Reject it. |
| 4744 | */ |
| 4745 | verbose(env, "same insn cannot be used with different pointers\n"); |
| 4746 | return -EINVAL; |
| 4747 | } |
| 4748 | |
| 4749 | } else if (class == BPF_STX) { |
| 4750 | enum bpf_reg_type *prev_dst_type, dst_reg_type; |
| 4751 | |
| 4752 | if (BPF_MODE(insn->code) == BPF_XADD) { |
| 4753 | err = check_xadd(env, insn_idx, insn); |
| 4754 | if (err) |
| 4755 | return err; |
| 4756 | insn_idx++; |
| 4757 | continue; |
| 4758 | } |
| 4759 | |
| 4760 | /* check src1 operand */ |
| 4761 | err = check_reg_arg(env, insn->src_reg, SRC_OP); |
| 4762 | if (err) |
| 4763 | return err; |
| 4764 | /* check src2 operand */ |
| 4765 | err = check_reg_arg(env, insn->dst_reg, SRC_OP); |
| 4766 | if (err) |
| 4767 | return err; |
| 4768 | |
| 4769 | dst_reg_type = regs[insn->dst_reg].type; |
| 4770 | |
| 4771 | /* check that memory (dst_reg + off) is writeable */ |
| 4772 | err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off, |
| 4773 | BPF_SIZE(insn->code), BPF_WRITE, |
| 4774 | insn->src_reg, false); |
| 4775 | if (err) |
| 4776 | return err; |
| 4777 | |
| 4778 | prev_dst_type = &env->insn_aux_data[insn_idx].ptr_type; |
| 4779 | |
| 4780 | if (*prev_dst_type == NOT_INIT) { |
| 4781 | *prev_dst_type = dst_reg_type; |
| 4782 | } else if (dst_reg_type != *prev_dst_type && |
| 4783 | (dst_reg_type == PTR_TO_CTX || |
| 4784 | *prev_dst_type == PTR_TO_CTX)) { |
| 4785 | verbose(env, "same insn cannot be used with different pointers\n"); |
| 4786 | return -EINVAL; |
| 4787 | } |
| 4788 | |
| 4789 | } else if (class == BPF_ST) { |
| 4790 | if (BPF_MODE(insn->code) != BPF_MEM || |
| 4791 | insn->src_reg != BPF_REG_0) { |
| 4792 | verbose(env, "BPF_ST uses reserved fields\n"); |
| 4793 | return -EINVAL; |
| 4794 | } |
| 4795 | /* check src operand */ |
| 4796 | err = check_reg_arg(env, insn->dst_reg, SRC_OP); |
| 4797 | if (err) |
| 4798 | return err; |
| 4799 | |
| 4800 | if (is_ctx_reg(env, insn->dst_reg)) { |
| 4801 | verbose(env, "BPF_ST stores into R%d context is not allowed\n", |
| 4802 | insn->dst_reg); |
| 4803 | return -EACCES; |
| 4804 | } |
| 4805 | |
| 4806 | /* check that memory (dst_reg + off) is writeable */ |
| 4807 | err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off, |
| 4808 | BPF_SIZE(insn->code), BPF_WRITE, |
| 4809 | -1, false); |
| 4810 | if (err) |
| 4811 | return err; |
| 4812 | |
| 4813 | } else if (class == BPF_JMP) { |
| 4814 | u8 opcode = BPF_OP(insn->code); |
| 4815 | |
| 4816 | if (opcode == BPF_CALL) { |
| 4817 | if (BPF_SRC(insn->code) != BPF_K || |
| 4818 | insn->off != 0 || |
| 4819 | (insn->src_reg != BPF_REG_0 && |
| 4820 | insn->src_reg != BPF_PSEUDO_CALL) || |
| 4821 | insn->dst_reg != BPF_REG_0) { |
| 4822 | verbose(env, "BPF_CALL uses reserved fields\n"); |
| 4823 | return -EINVAL; |
| 4824 | } |
| 4825 | |
| 4826 | if (insn->src_reg == BPF_PSEUDO_CALL) |
| 4827 | err = check_func_call(env, insn, &insn_idx); |
| 4828 | else |
| 4829 | err = check_helper_call(env, insn->imm, insn_idx); |
| 4830 | if (err) |
| 4831 | return err; |
| 4832 | |
| 4833 | } else if (opcode == BPF_JA) { |
| 4834 | if (BPF_SRC(insn->code) != BPF_K || |
| 4835 | insn->imm != 0 || |
| 4836 | insn->src_reg != BPF_REG_0 || |
| 4837 | insn->dst_reg != BPF_REG_0) { |
| 4838 | verbose(env, "BPF_JA uses reserved fields\n"); |
| 4839 | return -EINVAL; |
| 4840 | } |
| 4841 | |
| 4842 | insn_idx += insn->off + 1; |
| 4843 | continue; |
| 4844 | |
| 4845 | } else if (opcode == BPF_EXIT) { |
| 4846 | if (BPF_SRC(insn->code) != BPF_K || |
| 4847 | insn->imm != 0 || |
| 4848 | insn->src_reg != BPF_REG_0 || |
| 4849 | insn->dst_reg != BPF_REG_0) { |
| 4850 | verbose(env, "BPF_EXIT uses reserved fields\n"); |
| 4851 | return -EINVAL; |
| 4852 | } |
| 4853 | |
| 4854 | if (state->curframe) { |
| 4855 | /* exit from nested function */ |
| 4856 | prev_insn_idx = insn_idx; |
| 4857 | err = prepare_func_exit(env, &insn_idx); |
| 4858 | if (err) |
| 4859 | return err; |
| 4860 | do_print_state = true; |
| 4861 | continue; |
| 4862 | } |
| 4863 | |
| 4864 | /* eBPF calling convetion is such that R0 is used |
| 4865 | * to return the value from eBPF program. |
| 4866 | * Make sure that it's readable at this time |
| 4867 | * of bpf_exit, which means that program wrote |
| 4868 | * something into it earlier |
| 4869 | */ |
| 4870 | err = check_reg_arg(env, BPF_REG_0, SRC_OP); |
| 4871 | if (err) |
| 4872 | return err; |
| 4873 | |
| 4874 | if (is_pointer_value(env, BPF_REG_0)) { |
| 4875 | verbose(env, "R0 leaks addr as return value\n"); |
| 4876 | return -EACCES; |
| 4877 | } |
| 4878 | |
| 4879 | err = check_return_code(env); |
| 4880 | if (err) |
| 4881 | return err; |
| 4882 | process_bpf_exit: |
| 4883 | err = pop_stack(env, &prev_insn_idx, &insn_idx); |
| 4884 | if (err < 0) { |
| 4885 | if (err != -ENOENT) |
| 4886 | return err; |
| 4887 | break; |
| 4888 | } else { |
| 4889 | do_print_state = true; |
| 4890 | continue; |
| 4891 | } |
| 4892 | } else { |
| 4893 | err = check_cond_jmp_op(env, insn, &insn_idx); |
| 4894 | if (err) |
| 4895 | return err; |
| 4896 | } |
| 4897 | } else if (class == BPF_LD) { |
| 4898 | u8 mode = BPF_MODE(insn->code); |
| 4899 | |
| 4900 | if (mode == BPF_ABS || mode == BPF_IND) { |
| 4901 | err = check_ld_abs(env, insn); |
| 4902 | if (err) |
| 4903 | return err; |
| 4904 | |
| 4905 | } else if (mode == BPF_IMM) { |
| 4906 | err = check_ld_imm(env, insn); |
| 4907 | if (err) |
| 4908 | return err; |
| 4909 | |
| 4910 | insn_idx++; |
| 4911 | env->insn_aux_data[insn_idx].seen = true; |
| 4912 | } else { |
| 4913 | verbose(env, "invalid BPF_LD mode\n"); |
| 4914 | return -EINVAL; |
| 4915 | } |
| 4916 | } else { |
| 4917 | verbose(env, "unknown insn class %d\n", class); |
| 4918 | return -EINVAL; |
| 4919 | } |
| 4920 | |
| 4921 | insn_idx++; |
| 4922 | } |
| 4923 | |
| 4924 | verbose(env, "processed %d insns (limit %d), stack depth ", |
| 4925 | insn_processed, BPF_COMPLEXITY_LIMIT_INSNS); |
| 4926 | for (i = 0; i < env->subprog_cnt; i++) { |
| 4927 | u32 depth = env->subprog_info[i].stack_depth; |
| 4928 | |
| 4929 | verbose(env, "%d", depth); |
| 4930 | if (i + 1 < env->subprog_cnt) |
| 4931 | verbose(env, "+"); |
| 4932 | } |
| 4933 | verbose(env, "\n"); |
| 4934 | env->prog->aux->stack_depth = env->subprog_info[0].stack_depth; |
| 4935 | return 0; |
| 4936 | } |
| 4937 | |
| 4938 | static int check_map_prealloc(struct bpf_map *map) |
| 4939 | { |
| 4940 | return (map->map_type != BPF_MAP_TYPE_HASH && |
| 4941 | map->map_type != BPF_MAP_TYPE_PERCPU_HASH && |
| 4942 | map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) || |
| 4943 | !(map->map_flags & BPF_F_NO_PREALLOC); |
| 4944 | } |
| 4945 | |
| 4946 | static int check_map_prog_compatibility(struct bpf_verifier_env *env, |
| 4947 | struct bpf_map *map, |
| 4948 | struct bpf_prog *prog) |
| 4949 | |
| 4950 | { |
| 4951 | /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use |
| 4952 | * preallocated hash maps, since doing memory allocation |
| 4953 | * in overflow_handler can crash depending on where nmi got |
| 4954 | * triggered. |
| 4955 | */ |
| 4956 | if (prog->type == BPF_PROG_TYPE_PERF_EVENT) { |
| 4957 | if (!check_map_prealloc(map)) { |
| 4958 | verbose(env, "perf_event programs can only use preallocated hash map\n"); |
| 4959 | return -EINVAL; |
| 4960 | } |
| 4961 | if (map->inner_map_meta && |
| 4962 | !check_map_prealloc(map->inner_map_meta)) { |
| 4963 | verbose(env, "perf_event programs can only use preallocated inner hash map\n"); |
| 4964 | return -EINVAL; |
| 4965 | } |
| 4966 | } |
| 4967 | |
| 4968 | if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) && |
| 4969 | !bpf_offload_dev_match(prog, map)) { |
| 4970 | verbose(env, "offload device mismatch between prog and map\n"); |
| 4971 | return -EINVAL; |
| 4972 | } |
| 4973 | |
| 4974 | return 0; |
| 4975 | } |
| 4976 | |
| 4977 | /* look for pseudo eBPF instructions that access map FDs and |
| 4978 | * replace them with actual map pointers |
| 4979 | */ |
| 4980 | static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env) |
| 4981 | { |
| 4982 | struct bpf_insn *insn = env->prog->insnsi; |
| 4983 | int insn_cnt = env->prog->len; |
| 4984 | int i, j, err; |
| 4985 | |
| 4986 | err = bpf_prog_calc_tag(env->prog); |
| 4987 | if (err) |
| 4988 | return err; |
| 4989 | |
| 4990 | for (i = 0; i < insn_cnt; i++, insn++) { |
| 4991 | if (BPF_CLASS(insn->code) == BPF_LDX && |
| 4992 | (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) { |
| 4993 | verbose(env, "BPF_LDX uses reserved fields\n"); |
| 4994 | return -EINVAL; |
| 4995 | } |
| 4996 | |
| 4997 | if (BPF_CLASS(insn->code) == BPF_STX && |
| 4998 | ((BPF_MODE(insn->code) != BPF_MEM && |
| 4999 | BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) { |
| 5000 | verbose(env, "BPF_STX uses reserved fields\n"); |
| 5001 | return -EINVAL; |
| 5002 | } |
| 5003 | |
| 5004 | if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) { |
| 5005 | struct bpf_map *map; |
| 5006 | struct fd f; |
| 5007 | |
| 5008 | if (i == insn_cnt - 1 || insn[1].code != 0 || |
| 5009 | insn[1].dst_reg != 0 || insn[1].src_reg != 0 || |
| 5010 | insn[1].off != 0) { |
| 5011 | verbose(env, "invalid bpf_ld_imm64 insn\n"); |
| 5012 | return -EINVAL; |
| 5013 | } |
| 5014 | |
| 5015 | if (insn->src_reg == 0) |
| 5016 | /* valid generic load 64-bit imm */ |
| 5017 | goto next_insn; |
| 5018 | |
| 5019 | if (insn->src_reg != BPF_PSEUDO_MAP_FD) { |
| 5020 | verbose(env, |
| 5021 | "unrecognized bpf_ld_imm64 insn\n"); |
| 5022 | return -EINVAL; |
| 5023 | } |
| 5024 | |
| 5025 | f = fdget(insn->imm); |
| 5026 | map = __bpf_map_get(f); |
| 5027 | if (IS_ERR(map)) { |
| 5028 | verbose(env, "fd %d is not pointing to valid bpf_map\n", |
| 5029 | insn->imm); |
| 5030 | return PTR_ERR(map); |
| 5031 | } |
| 5032 | |
| 5033 | err = check_map_prog_compatibility(env, map, env->prog); |
| 5034 | if (err) { |
| 5035 | fdput(f); |
| 5036 | return err; |
| 5037 | } |
| 5038 | |
| 5039 | /* store map pointer inside BPF_LD_IMM64 instruction */ |
| 5040 | insn[0].imm = (u32) (unsigned long) map; |
| 5041 | insn[1].imm = ((u64) (unsigned long) map) >> 32; |
| 5042 | |
| 5043 | /* check whether we recorded this map already */ |
| 5044 | for (j = 0; j < env->used_map_cnt; j++) |
| 5045 | if (env->used_maps[j] == map) { |
| 5046 | fdput(f); |
| 5047 | goto next_insn; |
| 5048 | } |
| 5049 | |
| 5050 | if (env->used_map_cnt >= MAX_USED_MAPS) { |
| 5051 | fdput(f); |
| 5052 | return -E2BIG; |
| 5053 | } |
| 5054 | |
| 5055 | /* hold the map. If the program is rejected by verifier, |
| 5056 | * the map will be released by release_maps() or it |
| 5057 | * will be used by the valid program until it's unloaded |
| 5058 | * and all maps are released in free_used_maps() |
| 5059 | */ |
| 5060 | map = bpf_map_inc(map, false); |
| 5061 | if (IS_ERR(map)) { |
| 5062 | fdput(f); |
| 5063 | return PTR_ERR(map); |
| 5064 | } |
| 5065 | env->used_maps[env->used_map_cnt++] = map; |
| 5066 | |
| 5067 | fdput(f); |
| 5068 | next_insn: |
| 5069 | insn++; |
| 5070 | i++; |
| 5071 | continue; |
| 5072 | } |
| 5073 | |
| 5074 | /* Basic sanity check before we invest more work here. */ |
| 5075 | if (!bpf_opcode_in_insntable(insn->code)) { |
| 5076 | verbose(env, "unknown opcode %02x\n", insn->code); |
| 5077 | return -EINVAL; |
| 5078 | } |
| 5079 | } |
| 5080 | |
| 5081 | /* now all pseudo BPF_LD_IMM64 instructions load valid |
| 5082 | * 'struct bpf_map *' into a register instead of user map_fd. |
| 5083 | * These pointers will be used later by verifier to validate map access. |
| 5084 | */ |
| 5085 | return 0; |
| 5086 | } |
| 5087 | |
| 5088 | /* drop refcnt of maps used by the rejected program */ |
| 5089 | static void release_maps(struct bpf_verifier_env *env) |
| 5090 | { |
| 5091 | int i; |
| 5092 | |
| 5093 | for (i = 0; i < env->used_map_cnt; i++) |
| 5094 | bpf_map_put(env->used_maps[i]); |
| 5095 | } |
| 5096 | |
| 5097 | /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */ |
| 5098 | static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env) |
| 5099 | { |
| 5100 | struct bpf_insn *insn = env->prog->insnsi; |
| 5101 | int insn_cnt = env->prog->len; |
| 5102 | int i; |
| 5103 | |
| 5104 | for (i = 0; i < insn_cnt; i++, insn++) |
| 5105 | if (insn->code == (BPF_LD | BPF_IMM | BPF_DW)) |
| 5106 | insn->src_reg = 0; |
| 5107 | } |
| 5108 | |
| 5109 | /* single env->prog->insni[off] instruction was replaced with the range |
| 5110 | * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying |
| 5111 | * [0, off) and [off, end) to new locations, so the patched range stays zero |
| 5112 | */ |
| 5113 | static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len, |
| 5114 | u32 off, u32 cnt) |
| 5115 | { |
| 5116 | struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data; |
| 5117 | int i; |
| 5118 | |
| 5119 | if (cnt == 1) |
| 5120 | return 0; |
| 5121 | new_data = vzalloc(sizeof(struct bpf_insn_aux_data) * prog_len); |
| 5122 | if (!new_data) |
| 5123 | return -ENOMEM; |
| 5124 | memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off); |
| 5125 | memcpy(new_data + off + cnt - 1, old_data + off, |
| 5126 | sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1)); |
| 5127 | for (i = off; i < off + cnt - 1; i++) |
| 5128 | new_data[i].seen = true; |
| 5129 | env->insn_aux_data = new_data; |
| 5130 | vfree(old_data); |
| 5131 | return 0; |
| 5132 | } |
| 5133 | |
| 5134 | static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len) |
| 5135 | { |
| 5136 | int i; |
| 5137 | |
| 5138 | if (len == 1) |
| 5139 | return; |
| 5140 | /* NOTE: fake 'exit' subprog should be updated as well. */ |
| 5141 | for (i = 0; i <= env->subprog_cnt; i++) { |
| 5142 | if (env->subprog_info[i].start < off) |
| 5143 | continue; |
| 5144 | env->subprog_info[i].start += len - 1; |
| 5145 | } |
| 5146 | } |
| 5147 | |
| 5148 | static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off, |
| 5149 | const struct bpf_insn *patch, u32 len) |
| 5150 | { |
| 5151 | struct bpf_prog *new_prog; |
| 5152 | |
| 5153 | new_prog = bpf_patch_insn_single(env->prog, off, patch, len); |
| 5154 | if (!new_prog) |
| 5155 | return NULL; |
| 5156 | if (adjust_insn_aux_data(env, new_prog->len, off, len)) |
| 5157 | return NULL; |
| 5158 | adjust_subprog_starts(env, off, len); |
| 5159 | return new_prog; |
| 5160 | } |
| 5161 | |
| 5162 | /* The verifier does more data flow analysis than llvm and will not |
| 5163 | * explore branches that are dead at run time. Malicious programs can |
| 5164 | * have dead code too. Therefore replace all dead at-run-time code |
| 5165 | * with 'ja -1'. |
| 5166 | * |
| 5167 | * Just nops are not optimal, e.g. if they would sit at the end of the |
| 5168 | * program and through another bug we would manage to jump there, then |
| 5169 | * we'd execute beyond program memory otherwise. Returning exception |
| 5170 | * code also wouldn't work since we can have subprogs where the dead |
| 5171 | * code could be located. |
| 5172 | */ |
| 5173 | static void sanitize_dead_code(struct bpf_verifier_env *env) |
| 5174 | { |
| 5175 | struct bpf_insn_aux_data *aux_data = env->insn_aux_data; |
| 5176 | struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1); |
| 5177 | struct bpf_insn *insn = env->prog->insnsi; |
| 5178 | const int insn_cnt = env->prog->len; |
| 5179 | int i; |
| 5180 | |
| 5181 | for (i = 0; i < insn_cnt; i++) { |
| 5182 | if (aux_data[i].seen) |
| 5183 | continue; |
| 5184 | memcpy(insn + i, &trap, sizeof(trap)); |
| 5185 | } |
| 5186 | } |
| 5187 | |
| 5188 | /* convert load instructions that access fields of 'struct __sk_buff' |
| 5189 | * into sequence of instructions that access fields of 'struct sk_buff' |
| 5190 | */ |
| 5191 | static int convert_ctx_accesses(struct bpf_verifier_env *env) |
| 5192 | { |
| 5193 | const struct bpf_verifier_ops *ops = env->ops; |
| 5194 | int i, cnt, size, ctx_field_size, delta = 0; |
| 5195 | const int insn_cnt = env->prog->len; |
| 5196 | struct bpf_insn insn_buf[16], *insn; |
| 5197 | struct bpf_prog *new_prog; |
| 5198 | enum bpf_access_type type; |
| 5199 | bool is_narrower_load; |
| 5200 | u32 target_size; |
| 5201 | |
| 5202 | if (ops->gen_prologue) { |
| 5203 | cnt = ops->gen_prologue(insn_buf, env->seen_direct_write, |
| 5204 | env->prog); |
| 5205 | if (cnt >= ARRAY_SIZE(insn_buf)) { |
| 5206 | verbose(env, "bpf verifier is misconfigured\n"); |
| 5207 | return -EINVAL; |
| 5208 | } else if (cnt) { |
| 5209 | new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt); |
| 5210 | if (!new_prog) |
| 5211 | return -ENOMEM; |
| 5212 | |
| 5213 | env->prog = new_prog; |
| 5214 | delta += cnt - 1; |
| 5215 | } |
| 5216 | } |
| 5217 | |
| 5218 | if (!ops->convert_ctx_access || bpf_prog_is_dev_bound(env->prog->aux)) |
| 5219 | return 0; |
| 5220 | |
| 5221 | insn = env->prog->insnsi + delta; |
| 5222 | |
| 5223 | for (i = 0; i < insn_cnt; i++, insn++) { |
| 5224 | if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) || |
| 5225 | insn->code == (BPF_LDX | BPF_MEM | BPF_H) || |
| 5226 | insn->code == (BPF_LDX | BPF_MEM | BPF_W) || |
| 5227 | insn->code == (BPF_LDX | BPF_MEM | BPF_DW)) |
| 5228 | type = BPF_READ; |
| 5229 | else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) || |
| 5230 | insn->code == (BPF_STX | BPF_MEM | BPF_H) || |
| 5231 | insn->code == (BPF_STX | BPF_MEM | BPF_W) || |
| 5232 | insn->code == (BPF_STX | BPF_MEM | BPF_DW)) |
| 5233 | type = BPF_WRITE; |
| 5234 | else |
| 5235 | continue; |
| 5236 | |
| 5237 | if (env->insn_aux_data[i + delta].ptr_type != PTR_TO_CTX) |
| 5238 | continue; |
| 5239 | |
| 5240 | ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size; |
| 5241 | size = BPF_LDST_BYTES(insn); |
| 5242 | |
| 5243 | /* If the read access is a narrower load of the field, |
| 5244 | * convert to a 4/8-byte load, to minimum program type specific |
| 5245 | * convert_ctx_access changes. If conversion is successful, |
| 5246 | * we will apply proper mask to the result. |
| 5247 | */ |
| 5248 | is_narrower_load = size < ctx_field_size; |
| 5249 | if (is_narrower_load) { |
| 5250 | u32 off = insn->off; |
| 5251 | u8 size_code; |
| 5252 | |
| 5253 | if (type == BPF_WRITE) { |
| 5254 | verbose(env, "bpf verifier narrow ctx access misconfigured\n"); |
| 5255 | return -EINVAL; |
| 5256 | } |
| 5257 | |
| 5258 | size_code = BPF_H; |
| 5259 | if (ctx_field_size == 4) |
| 5260 | size_code = BPF_W; |
| 5261 | else if (ctx_field_size == 8) |
| 5262 | size_code = BPF_DW; |
| 5263 | |
| 5264 | insn->off = off & ~(ctx_field_size - 1); |
| 5265 | insn->code = BPF_LDX | BPF_MEM | size_code; |
| 5266 | } |
| 5267 | |
| 5268 | target_size = 0; |
| 5269 | cnt = ops->convert_ctx_access(type, insn, insn_buf, env->prog, |
| 5270 | &target_size); |
| 5271 | if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) || |
| 5272 | (ctx_field_size && !target_size)) { |
| 5273 | verbose(env, "bpf verifier is misconfigured\n"); |
| 5274 | return -EINVAL; |
| 5275 | } |
| 5276 | |
| 5277 | if (is_narrower_load && size < target_size) { |
| 5278 | if (ctx_field_size <= 4) |
| 5279 | insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg, |
| 5280 | (1 << size * 8) - 1); |
| 5281 | else |
| 5282 | insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg, |
| 5283 | (1 << size * 8) - 1); |
| 5284 | } |
| 5285 | |
| 5286 | new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); |
| 5287 | if (!new_prog) |
| 5288 | return -ENOMEM; |
| 5289 | |
| 5290 | delta += cnt - 1; |
| 5291 | |
| 5292 | /* keep walking new program and skip insns we just inserted */ |
| 5293 | env->prog = new_prog; |
| 5294 | insn = new_prog->insnsi + i + delta; |
| 5295 | } |
| 5296 | |
| 5297 | return 0; |
| 5298 | } |
| 5299 | |
| 5300 | static int jit_subprogs(struct bpf_verifier_env *env) |
| 5301 | { |
| 5302 | struct bpf_prog *prog = env->prog, **func, *tmp; |
| 5303 | int i, j, subprog_start, subprog_end = 0, len, subprog; |
| 5304 | struct bpf_insn *insn; |
| 5305 | void *old_bpf_func; |
| 5306 | int err = -ENOMEM; |
| 5307 | |
| 5308 | if (env->subprog_cnt <= 1) |
| 5309 | return 0; |
| 5310 | |
| 5311 | for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) { |
| 5312 | if (insn->code != (BPF_JMP | BPF_CALL) || |
| 5313 | insn->src_reg != BPF_PSEUDO_CALL) |
| 5314 | continue; |
| 5315 | subprog = find_subprog(env, i + insn->imm + 1); |
| 5316 | if (subprog < 0) { |
| 5317 | WARN_ONCE(1, "verifier bug. No program starts at insn %d\n", |
| 5318 | i + insn->imm + 1); |
| 5319 | return -EFAULT; |
| 5320 | } |
| 5321 | /* temporarily remember subprog id inside insn instead of |
| 5322 | * aux_data, since next loop will split up all insns into funcs |
| 5323 | */ |
| 5324 | insn->off = subprog; |
| 5325 | /* remember original imm in case JIT fails and fallback |
| 5326 | * to interpreter will be needed |
| 5327 | */ |
| 5328 | env->insn_aux_data[i].call_imm = insn->imm; |
| 5329 | /* point imm to __bpf_call_base+1 from JITs point of view */ |
| 5330 | insn->imm = 1; |
| 5331 | } |
| 5332 | |
| 5333 | func = kzalloc(sizeof(prog) * env->subprog_cnt, GFP_KERNEL); |
| 5334 | if (!func) |
| 5335 | return -ENOMEM; |
| 5336 | |
| 5337 | for (i = 0; i < env->subprog_cnt; i++) { |
| 5338 | subprog_start = subprog_end; |
| 5339 | subprog_end = env->subprog_info[i + 1].start; |
| 5340 | |
| 5341 | len = subprog_end - subprog_start; |
| 5342 | func[i] = bpf_prog_alloc(bpf_prog_size(len), GFP_USER); |
| 5343 | if (!func[i]) |
| 5344 | goto out_free; |
| 5345 | memcpy(func[i]->insnsi, &prog->insnsi[subprog_start], |
| 5346 | len * sizeof(struct bpf_insn)); |
| 5347 | func[i]->type = prog->type; |
| 5348 | func[i]->len = len; |
| 5349 | if (bpf_prog_calc_tag(func[i])) |
| 5350 | goto out_free; |
| 5351 | func[i]->is_func = 1; |
| 5352 | /* Use bpf_prog_F_tag to indicate functions in stack traces. |
| 5353 | * Long term would need debug info to populate names |
| 5354 | */ |
| 5355 | func[i]->aux->name[0] = 'F'; |
| 5356 | func[i]->aux->stack_depth = env->subprog_info[i].stack_depth; |
| 5357 | func[i]->jit_requested = 1; |
| 5358 | func[i] = bpf_int_jit_compile(func[i]); |
| 5359 | if (!func[i]->jited) { |
| 5360 | err = -ENOTSUPP; |
| 5361 | goto out_free; |
| 5362 | } |
| 5363 | cond_resched(); |
| 5364 | } |
| 5365 | /* at this point all bpf functions were successfully JITed |
| 5366 | * now populate all bpf_calls with correct addresses and |
| 5367 | * run last pass of JIT |
| 5368 | */ |
| 5369 | for (i = 0; i < env->subprog_cnt; i++) { |
| 5370 | insn = func[i]->insnsi; |
| 5371 | for (j = 0; j < func[i]->len; j++, insn++) { |
| 5372 | if (insn->code != (BPF_JMP | BPF_CALL) || |
| 5373 | insn->src_reg != BPF_PSEUDO_CALL) |
| 5374 | continue; |
| 5375 | subprog = insn->off; |
| 5376 | insn->off = 0; |
| 5377 | insn->imm = (u64 (*)(u64, u64, u64, u64, u64)) |
| 5378 | func[subprog]->bpf_func - |
| 5379 | __bpf_call_base; |
| 5380 | } |
| 5381 | } |
| 5382 | for (i = 0; i < env->subprog_cnt; i++) { |
| 5383 | old_bpf_func = func[i]->bpf_func; |
| 5384 | tmp = bpf_int_jit_compile(func[i]); |
| 5385 | if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) { |
| 5386 | verbose(env, "JIT doesn't support bpf-to-bpf calls\n"); |
| 5387 | err = -EFAULT; |
| 5388 | goto out_free; |
| 5389 | } |
| 5390 | cond_resched(); |
| 5391 | } |
| 5392 | |
| 5393 | /* finally lock prog and jit images for all functions and |
| 5394 | * populate kallsysm |
| 5395 | */ |
| 5396 | for (i = 0; i < env->subprog_cnt; i++) { |
| 5397 | bpf_prog_lock_ro(func[i]); |
| 5398 | bpf_prog_kallsyms_add(func[i]); |
| 5399 | } |
| 5400 | |
| 5401 | /* Last step: make now unused interpreter insns from main |
| 5402 | * prog consistent for later dump requests, so they can |
| 5403 | * later look the same as if they were interpreted only. |
| 5404 | */ |
| 5405 | for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) { |
| 5406 | unsigned long addr; |
| 5407 | |
| 5408 | if (insn->code != (BPF_JMP | BPF_CALL) || |
| 5409 | insn->src_reg != BPF_PSEUDO_CALL) |
| 5410 | continue; |
| 5411 | insn->off = env->insn_aux_data[i].call_imm; |
| 5412 | subprog = find_subprog(env, i + insn->off + 1); |
| 5413 | addr = (unsigned long)func[subprog]->bpf_func; |
| 5414 | addr &= PAGE_MASK; |
| 5415 | insn->imm = (u64 (*)(u64, u64, u64, u64, u64)) |
| 5416 | addr - __bpf_call_base; |
| 5417 | } |
| 5418 | |
| 5419 | prog->jited = 1; |
| 5420 | prog->bpf_func = func[0]->bpf_func; |
| 5421 | prog->aux->func = func; |
| 5422 | prog->aux->func_cnt = env->subprog_cnt; |
| 5423 | return 0; |
| 5424 | out_free: |
| 5425 | for (i = 0; i < env->subprog_cnt; i++) |
| 5426 | if (func[i]) |
| 5427 | bpf_jit_free(func[i]); |
| 5428 | kfree(func); |
| 5429 | /* cleanup main prog to be interpreted */ |
| 5430 | prog->jit_requested = 0; |
| 5431 | for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) { |
| 5432 | if (insn->code != (BPF_JMP | BPF_CALL) || |
| 5433 | insn->src_reg != BPF_PSEUDO_CALL) |
| 5434 | continue; |
| 5435 | insn->off = 0; |
| 5436 | insn->imm = env->insn_aux_data[i].call_imm; |
| 5437 | } |
| 5438 | return err; |
| 5439 | } |
| 5440 | |
| 5441 | static int fixup_call_args(struct bpf_verifier_env *env) |
| 5442 | { |
| 5443 | #ifndef CONFIG_BPF_JIT_ALWAYS_ON |
| 5444 | struct bpf_prog *prog = env->prog; |
| 5445 | struct bpf_insn *insn = prog->insnsi; |
| 5446 | int i, depth; |
| 5447 | #endif |
| 5448 | int err; |
| 5449 | |
| 5450 | err = 0; |
| 5451 | if (env->prog->jit_requested) { |
| 5452 | err = jit_subprogs(env); |
| 5453 | if (err == 0) |
| 5454 | return 0; |
| 5455 | } |
| 5456 | #ifndef CONFIG_BPF_JIT_ALWAYS_ON |
| 5457 | for (i = 0; i < prog->len; i++, insn++) { |
| 5458 | if (insn->code != (BPF_JMP | BPF_CALL) || |
| 5459 | insn->src_reg != BPF_PSEUDO_CALL) |
| 5460 | continue; |
| 5461 | depth = get_callee_stack_depth(env, insn, i); |
| 5462 | if (depth < 0) |
| 5463 | return depth; |
| 5464 | bpf_patch_call_args(insn, depth); |
| 5465 | } |
| 5466 | err = 0; |
| 5467 | #endif |
| 5468 | return err; |
| 5469 | } |
| 5470 | |
| 5471 | /* fixup insn->imm field of bpf_call instructions |
| 5472 | * and inline eligible helpers as explicit sequence of BPF instructions |
| 5473 | * |
| 5474 | * this function is called after eBPF program passed verification |
| 5475 | */ |
| 5476 | static int fixup_bpf_calls(struct bpf_verifier_env *env) |
| 5477 | { |
| 5478 | struct bpf_prog *prog = env->prog; |
| 5479 | struct bpf_insn *insn = prog->insnsi; |
| 5480 | const struct bpf_func_proto *fn; |
| 5481 | const int insn_cnt = prog->len; |
| 5482 | struct bpf_insn insn_buf[16]; |
| 5483 | struct bpf_prog *new_prog; |
| 5484 | struct bpf_map *map_ptr; |
| 5485 | int i, cnt, delta = 0; |
| 5486 | |
| 5487 | for (i = 0; i < insn_cnt; i++, insn++) { |
| 5488 | if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) || |
| 5489 | insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) || |
| 5490 | insn->code == (BPF_ALU | BPF_MOD | BPF_X) || |
| 5491 | insn->code == (BPF_ALU | BPF_DIV | BPF_X)) { |
| 5492 | bool is64 = BPF_CLASS(insn->code) == BPF_ALU64; |
| 5493 | struct bpf_insn mask_and_div[] = { |
| 5494 | BPF_MOV32_REG(insn->src_reg, insn->src_reg), |
| 5495 | /* Rx div 0 -> 0 */ |
| 5496 | BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2), |
| 5497 | BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg), |
| 5498 | BPF_JMP_IMM(BPF_JA, 0, 0, 1), |
| 5499 | *insn, |
| 5500 | }; |
| 5501 | struct bpf_insn mask_and_mod[] = { |
| 5502 | BPF_MOV32_REG(insn->src_reg, insn->src_reg), |
| 5503 | /* Rx mod 0 -> Rx */ |
| 5504 | BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1), |
| 5505 | *insn, |
| 5506 | }; |
| 5507 | struct bpf_insn *patchlet; |
| 5508 | |
| 5509 | if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) || |
| 5510 | insn->code == (BPF_ALU | BPF_DIV | BPF_X)) { |
| 5511 | patchlet = mask_and_div + (is64 ? 1 : 0); |
| 5512 | cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0); |
| 5513 | } else { |
| 5514 | patchlet = mask_and_mod + (is64 ? 1 : 0); |
| 5515 | cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0); |
| 5516 | } |
| 5517 | |
| 5518 | new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt); |
| 5519 | if (!new_prog) |
| 5520 | return -ENOMEM; |
| 5521 | |
| 5522 | delta += cnt - 1; |
| 5523 | env->prog = prog = new_prog; |
| 5524 | insn = new_prog->insnsi + i + delta; |
| 5525 | continue; |
| 5526 | } |
| 5527 | |
| 5528 | if (BPF_CLASS(insn->code) == BPF_LD && |
| 5529 | (BPF_MODE(insn->code) == BPF_ABS || |
| 5530 | BPF_MODE(insn->code) == BPF_IND)) { |
| 5531 | cnt = env->ops->gen_ld_abs(insn, insn_buf); |
| 5532 | if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) { |
| 5533 | verbose(env, "bpf verifier is misconfigured\n"); |
| 5534 | return -EINVAL; |
| 5535 | } |
| 5536 | |
| 5537 | new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); |
| 5538 | if (!new_prog) |
| 5539 | return -ENOMEM; |
| 5540 | |
| 5541 | delta += cnt - 1; |
| 5542 | env->prog = prog = new_prog; |
| 5543 | insn = new_prog->insnsi + i + delta; |
| 5544 | continue; |
| 5545 | } |
| 5546 | |
| 5547 | if (insn->code != (BPF_JMP | BPF_CALL)) |
| 5548 | continue; |
| 5549 | if (insn->src_reg == BPF_PSEUDO_CALL) |
| 5550 | continue; |
| 5551 | |
| 5552 | if (insn->imm == BPF_FUNC_get_route_realm) |
| 5553 | prog->dst_needed = 1; |
| 5554 | if (insn->imm == BPF_FUNC_get_prandom_u32) |
| 5555 | bpf_user_rnd_init_once(); |
| 5556 | if (insn->imm == BPF_FUNC_override_return) |
| 5557 | prog->kprobe_override = 1; |
| 5558 | if (insn->imm == BPF_FUNC_tail_call) { |
| 5559 | /* If we tail call into other programs, we |
| 5560 | * cannot make any assumptions since they can |
| 5561 | * be replaced dynamically during runtime in |
| 5562 | * the program array. |
| 5563 | */ |
| 5564 | prog->cb_access = 1; |
| 5565 | env->prog->aux->stack_depth = MAX_BPF_STACK; |
| 5566 | |
| 5567 | /* mark bpf_tail_call as different opcode to avoid |
| 5568 | * conditional branch in the interpeter for every normal |
| 5569 | * call and to prevent accidental JITing by JIT compiler |
| 5570 | * that doesn't support bpf_tail_call yet |
| 5571 | */ |
| 5572 | insn->imm = 0; |
| 5573 | insn->code = BPF_JMP | BPF_TAIL_CALL; |
| 5574 | |
| 5575 | /* instead of changing every JIT dealing with tail_call |
| 5576 | * emit two extra insns: |
| 5577 | * if (index >= max_entries) goto out; |
| 5578 | * index &= array->index_mask; |
| 5579 | * to avoid out-of-bounds cpu speculation |
| 5580 | */ |
| 5581 | map_ptr = env->insn_aux_data[i + delta].map_ptr; |
| 5582 | if (map_ptr == BPF_MAP_PTR_POISON) { |
| 5583 | verbose(env, "tail_call abusing map_ptr\n"); |
| 5584 | return -EINVAL; |
| 5585 | } |
| 5586 | if (!map_ptr->unpriv_array) |
| 5587 | continue; |
| 5588 | insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3, |
| 5589 | map_ptr->max_entries, 2); |
| 5590 | insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3, |
| 5591 | container_of(map_ptr, |
| 5592 | struct bpf_array, |
| 5593 | map)->index_mask); |
| 5594 | insn_buf[2] = *insn; |
| 5595 | cnt = 3; |
| 5596 | new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); |
| 5597 | if (!new_prog) |
| 5598 | return -ENOMEM; |
| 5599 | |
| 5600 | delta += cnt - 1; |
| 5601 | env->prog = prog = new_prog; |
| 5602 | insn = new_prog->insnsi + i + delta; |
| 5603 | continue; |
| 5604 | } |
| 5605 | |
| 5606 | /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup |
| 5607 | * handlers are currently limited to 64 bit only. |
| 5608 | */ |
| 5609 | if (prog->jit_requested && BITS_PER_LONG == 64 && |
| 5610 | insn->imm == BPF_FUNC_map_lookup_elem) { |
| 5611 | map_ptr = env->insn_aux_data[i + delta].map_ptr; |
| 5612 | if (map_ptr == BPF_MAP_PTR_POISON || |
| 5613 | !map_ptr->ops->map_gen_lookup) |
| 5614 | goto patch_call_imm; |
| 5615 | |
| 5616 | cnt = map_ptr->ops->map_gen_lookup(map_ptr, insn_buf); |
| 5617 | if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) { |
| 5618 | verbose(env, "bpf verifier is misconfigured\n"); |
| 5619 | return -EINVAL; |
| 5620 | } |
| 5621 | |
| 5622 | new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, |
| 5623 | cnt); |
| 5624 | if (!new_prog) |
| 5625 | return -ENOMEM; |
| 5626 | |
| 5627 | delta += cnt - 1; |
| 5628 | |
| 5629 | /* keep walking new program and skip insns we just inserted */ |
| 5630 | env->prog = prog = new_prog; |
| 5631 | insn = new_prog->insnsi + i + delta; |
| 5632 | continue; |
| 5633 | } |
| 5634 | |
| 5635 | if (insn->imm == BPF_FUNC_redirect_map) { |
| 5636 | /* Note, we cannot use prog directly as imm as subsequent |
| 5637 | * rewrites would still change the prog pointer. The only |
| 5638 | * stable address we can use is aux, which also works with |
| 5639 | * prog clones during blinding. |
| 5640 | */ |
| 5641 | u64 addr = (unsigned long)prog->aux; |
| 5642 | struct bpf_insn r4_ld[] = { |
| 5643 | BPF_LD_IMM64(BPF_REG_4, addr), |
| 5644 | *insn, |
| 5645 | }; |
| 5646 | cnt = ARRAY_SIZE(r4_ld); |
| 5647 | |
| 5648 | new_prog = bpf_patch_insn_data(env, i + delta, r4_ld, cnt); |
| 5649 | if (!new_prog) |
| 5650 | return -ENOMEM; |
| 5651 | |
| 5652 | delta += cnt - 1; |
| 5653 | env->prog = prog = new_prog; |
| 5654 | insn = new_prog->insnsi + i + delta; |
| 5655 | } |
| 5656 | patch_call_imm: |
| 5657 | fn = env->ops->get_func_proto(insn->imm, env->prog); |
| 5658 | /* all functions that have prototype and verifier allowed |
| 5659 | * programs to call them, must be real in-kernel functions |
| 5660 | */ |
| 5661 | if (!fn->func) { |
| 5662 | verbose(env, |
| 5663 | "kernel subsystem misconfigured func %s#%d\n", |
| 5664 | func_id_name(insn->imm), insn->imm); |
| 5665 | return -EFAULT; |
| 5666 | } |
| 5667 | insn->imm = fn->func - __bpf_call_base; |
| 5668 | } |
| 5669 | |
| 5670 | return 0; |
| 5671 | } |
| 5672 | |
| 5673 | static void free_states(struct bpf_verifier_env *env) |
| 5674 | { |
| 5675 | struct bpf_verifier_state_list *sl, *sln; |
| 5676 | int i; |
| 5677 | |
| 5678 | if (!env->explored_states) |
| 5679 | return; |
| 5680 | |
| 5681 | for (i = 0; i < env->prog->len; i++) { |
| 5682 | sl = env->explored_states[i]; |
| 5683 | |
| 5684 | if (sl) |
| 5685 | while (sl != STATE_LIST_MARK) { |
| 5686 | sln = sl->next; |
| 5687 | free_verifier_state(&sl->state, false); |
| 5688 | kfree(sl); |
| 5689 | sl = sln; |
| 5690 | } |
| 5691 | } |
| 5692 | |
| 5693 | kfree(env->explored_states); |
| 5694 | } |
| 5695 | |
| 5696 | int bpf_check(struct bpf_prog **prog, union bpf_attr *attr) |
| 5697 | { |
| 5698 | struct bpf_verifier_env *env; |
| 5699 | struct bpf_verifier_log *log; |
| 5700 | int ret = -EINVAL; |
| 5701 | |
| 5702 | /* no program is valid */ |
| 5703 | if (ARRAY_SIZE(bpf_verifier_ops) == 0) |
| 5704 | return -EINVAL; |
| 5705 | |
| 5706 | /* 'struct bpf_verifier_env' can be global, but since it's not small, |
| 5707 | * allocate/free it every time bpf_check() is called |
| 5708 | */ |
| 5709 | env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL); |
| 5710 | if (!env) |
| 5711 | return -ENOMEM; |
| 5712 | log = &env->log; |
| 5713 | |
| 5714 | env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) * |
| 5715 | (*prog)->len); |
| 5716 | ret = -ENOMEM; |
| 5717 | if (!env->insn_aux_data) |
| 5718 | goto err_free_env; |
| 5719 | env->prog = *prog; |
| 5720 | env->ops = bpf_verifier_ops[env->prog->type]; |
| 5721 | |
| 5722 | /* grab the mutex to protect few globals used by verifier */ |
| 5723 | mutex_lock(&bpf_verifier_lock); |
| 5724 | |
| 5725 | if (attr->log_level || attr->log_buf || attr->log_size) { |
| 5726 | /* user requested verbose verifier output |
| 5727 | * and supplied buffer to store the verification trace |
| 5728 | */ |
| 5729 | log->level = attr->log_level; |
| 5730 | log->ubuf = (char __user *) (unsigned long) attr->log_buf; |
| 5731 | log->len_total = attr->log_size; |
| 5732 | |
| 5733 | ret = -EINVAL; |
| 5734 | /* log attributes have to be sane */ |
| 5735 | if (log->len_total < 128 || log->len_total > UINT_MAX >> 8 || |
| 5736 | !log->level || !log->ubuf) |
| 5737 | goto err_unlock; |
| 5738 | } |
| 5739 | |
| 5740 | env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT); |
| 5741 | if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)) |
| 5742 | env->strict_alignment = true; |
| 5743 | |
| 5744 | ret = replace_map_fd_with_map_ptr(env); |
| 5745 | if (ret < 0) |
| 5746 | goto skip_full_check; |
| 5747 | |
| 5748 | if (bpf_prog_is_dev_bound(env->prog->aux)) { |
| 5749 | ret = bpf_prog_offload_verifier_prep(env); |
| 5750 | if (ret) |
| 5751 | goto skip_full_check; |
| 5752 | } |
| 5753 | |
| 5754 | env->explored_states = kcalloc(env->prog->len, |
| 5755 | sizeof(struct bpf_verifier_state_list *), |
| 5756 | GFP_USER); |
| 5757 | ret = -ENOMEM; |
| 5758 | if (!env->explored_states) |
| 5759 | goto skip_full_check; |
| 5760 | |
| 5761 | env->allow_ptr_leaks = capable(CAP_SYS_ADMIN); |
| 5762 | |
| 5763 | ret = check_cfg(env); |
| 5764 | if (ret < 0) |
| 5765 | goto skip_full_check; |
| 5766 | |
| 5767 | ret = do_check(env); |
| 5768 | if (env->cur_state) { |
| 5769 | free_verifier_state(env->cur_state, true); |
| 5770 | env->cur_state = NULL; |
| 5771 | } |
| 5772 | |
| 5773 | skip_full_check: |
| 5774 | while (!pop_stack(env, NULL, NULL)); |
| 5775 | free_states(env); |
| 5776 | |
| 5777 | if (ret == 0) |
| 5778 | sanitize_dead_code(env); |
| 5779 | |
| 5780 | if (ret == 0) |
| 5781 | ret = check_max_stack_depth(env); |
| 5782 | |
| 5783 | if (ret == 0) |
| 5784 | /* program is valid, convert *(u32*)(ctx + off) accesses */ |
| 5785 | ret = convert_ctx_accesses(env); |
| 5786 | |
| 5787 | if (ret == 0) |
| 5788 | ret = fixup_bpf_calls(env); |
| 5789 | |
| 5790 | if (ret == 0) |
| 5791 | ret = fixup_call_args(env); |
| 5792 | |
| 5793 | if (log->level && bpf_verifier_log_full(log)) |
| 5794 | ret = -ENOSPC; |
| 5795 | if (log->level && !log->ubuf) { |
| 5796 | ret = -EFAULT; |
| 5797 | goto err_release_maps; |
| 5798 | } |
| 5799 | |
| 5800 | if (ret == 0 && env->used_map_cnt) { |
| 5801 | /* if program passed verifier, update used_maps in bpf_prog_info */ |
| 5802 | env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt, |
| 5803 | sizeof(env->used_maps[0]), |
| 5804 | GFP_KERNEL); |
| 5805 | |
| 5806 | if (!env->prog->aux->used_maps) { |
| 5807 | ret = -ENOMEM; |
| 5808 | goto err_release_maps; |
| 5809 | } |
| 5810 | |
| 5811 | memcpy(env->prog->aux->used_maps, env->used_maps, |
| 5812 | sizeof(env->used_maps[0]) * env->used_map_cnt); |
| 5813 | env->prog->aux->used_map_cnt = env->used_map_cnt; |
| 5814 | |
| 5815 | /* program is valid. Convert pseudo bpf_ld_imm64 into generic |
| 5816 | * bpf_ld_imm64 instructions |
| 5817 | */ |
| 5818 | convert_pseudo_ld_imm64(env); |
| 5819 | } |
| 5820 | |
| 5821 | err_release_maps: |
| 5822 | if (!env->prog->aux->used_maps) |
| 5823 | /* if we didn't copy map pointers into bpf_prog_info, release |
| 5824 | * them now. Otherwise free_used_maps() will release them. |
| 5825 | */ |
| 5826 | release_maps(env); |
| 5827 | *prog = env->prog; |
| 5828 | err_unlock: |
| 5829 | mutex_unlock(&bpf_verifier_lock); |
| 5830 | vfree(env->insn_aux_data); |
| 5831 | err_free_env: |
| 5832 | kfree(env); |
| 5833 | return ret; |
| 5834 | } |