bpf-prologue.c (12225B)
1// SPDX-License-Identifier: GPL-2.0 2/* 3 * bpf-prologue.c 4 * 5 * Copyright (C) 2015 He Kuang <hekuang@huawei.com> 6 * Copyright (C) 2015 Wang Nan <wangnan0@huawei.com> 7 * Copyright (C) 2015 Huawei Inc. 8 */ 9 10#include <bpf/libbpf.h> 11#include "debug.h" 12#include "bpf-loader.h" 13#include "bpf-prologue.h" 14#include "probe-finder.h" 15#include <errno.h> 16#include <stdlib.h> 17#include <dwarf-regs.h> 18#include <linux/filter.h> 19 20#define BPF_REG_SIZE 8 21 22#define JMP_TO_ERROR_CODE -1 23#define JMP_TO_SUCCESS_CODE -2 24#define JMP_TO_USER_CODE -3 25 26struct bpf_insn_pos { 27 struct bpf_insn *begin; 28 struct bpf_insn *end; 29 struct bpf_insn *pos; 30}; 31 32static inline int 33pos_get_cnt(struct bpf_insn_pos *pos) 34{ 35 return pos->pos - pos->begin; 36} 37 38static int 39append_insn(struct bpf_insn new_insn, struct bpf_insn_pos *pos) 40{ 41 if (!pos->pos) 42 return -BPF_LOADER_ERRNO__PROLOGUE2BIG; 43 44 if (pos->pos + 1 >= pos->end) { 45 pr_err("bpf prologue: prologue too long\n"); 46 pos->pos = NULL; 47 return -BPF_LOADER_ERRNO__PROLOGUE2BIG; 48 } 49 50 *(pos->pos)++ = new_insn; 51 return 0; 52} 53 54static int 55check_pos(struct bpf_insn_pos *pos) 56{ 57 if (!pos->pos || pos->pos >= pos->end) 58 return -BPF_LOADER_ERRNO__PROLOGUE2BIG; 59 return 0; 60} 61 62/* 63 * Convert type string (u8/u16/u32/u64/s8/s16/s32/s64 ..., see 64 * Documentation/trace/kprobetrace.rst) to size field of BPF_LDX_MEM 65 * instruction (BPF_{B,H,W,DW}). 66 */ 67static int 68argtype_to_ldx_size(const char *type) 69{ 70 int arg_size = type ? atoi(&type[1]) : 64; 71 72 switch (arg_size) { 73 case 8: 74 return BPF_B; 75 case 16: 76 return BPF_H; 77 case 32: 78 return BPF_W; 79 case 64: 80 default: 81 return BPF_DW; 82 } 83} 84 85static const char * 86insn_sz_to_str(int insn_sz) 87{ 88 switch (insn_sz) { 89 case BPF_B: 90 return "BPF_B"; 91 case BPF_H: 92 return "BPF_H"; 93 case BPF_W: 94 return "BPF_W"; 95 case BPF_DW: 96 return "BPF_DW"; 97 default: 98 return "UNKNOWN"; 99 } 100} 101 102/* Give it a shorter name */ 103#define ins(i, p) append_insn((i), (p)) 104 105/* 106 * Give a register name (in 'reg'), generate instruction to 107 * load register into an eBPF register rd: 108 * 'ldd target_reg, offset(ctx_reg)', where: 109 * ctx_reg is pre initialized to pointer of 'struct pt_regs'. 110 */ 111static int 112gen_ldx_reg_from_ctx(struct bpf_insn_pos *pos, int ctx_reg, 113 const char *reg, int target_reg) 114{ 115 int offset = regs_query_register_offset(reg); 116 117 if (offset < 0) { 118 pr_err("bpf: prologue: failed to get register %s\n", 119 reg); 120 return offset; 121 } 122 ins(BPF_LDX_MEM(BPF_DW, target_reg, ctx_reg, offset), pos); 123 124 return check_pos(pos); 125} 126 127/* 128 * Generate a BPF_FUNC_probe_read function call. 129 * 130 * src_base_addr_reg is a register holding base address, 131 * dst_addr_reg is a register holding dest address (on stack), 132 * result is: 133 * 134 * *[dst_addr_reg] = *([src_base_addr_reg] + offset) 135 * 136 * Arguments of BPF_FUNC_probe_read: 137 * ARG1: ptr to stack (dest) 138 * ARG2: size (8) 139 * ARG3: unsafe ptr (src) 140 */ 141static int 142gen_read_mem(struct bpf_insn_pos *pos, 143 int src_base_addr_reg, 144 int dst_addr_reg, 145 long offset, 146 int probeid) 147{ 148 /* mov arg3, src_base_addr_reg */ 149 if (src_base_addr_reg != BPF_REG_ARG3) 150 ins(BPF_MOV64_REG(BPF_REG_ARG3, src_base_addr_reg), pos); 151 /* add arg3, #offset */ 152 if (offset) 153 ins(BPF_ALU64_IMM(BPF_ADD, BPF_REG_ARG3, offset), pos); 154 155 /* mov arg2, #reg_size */ 156 ins(BPF_ALU64_IMM(BPF_MOV, BPF_REG_ARG2, BPF_REG_SIZE), pos); 157 158 /* mov arg1, dst_addr_reg */ 159 if (dst_addr_reg != BPF_REG_ARG1) 160 ins(BPF_MOV64_REG(BPF_REG_ARG1, dst_addr_reg), pos); 161 162 /* Call probe_read */ 163 ins(BPF_EMIT_CALL(probeid), pos); 164 /* 165 * Error processing: if read fail, goto error code, 166 * will be relocated. Target should be the start of 167 * error processing code. 168 */ 169 ins(BPF_JMP_IMM(BPF_JNE, BPF_REG_0, 0, JMP_TO_ERROR_CODE), 170 pos); 171 172 return check_pos(pos); 173} 174 175/* 176 * Each arg should be bare register. Fetch and save them into argument 177 * registers (r3 - r5). 178 * 179 * BPF_REG_1 should have been initialized with pointer to 180 * 'struct pt_regs'. 181 */ 182static int 183gen_prologue_fastpath(struct bpf_insn_pos *pos, 184 struct probe_trace_arg *args, int nargs) 185{ 186 int i, err = 0; 187 188 for (i = 0; i < nargs; i++) { 189 err = gen_ldx_reg_from_ctx(pos, BPF_REG_1, args[i].value, 190 BPF_PROLOGUE_START_ARG_REG + i); 191 if (err) 192 goto errout; 193 } 194 195 return check_pos(pos); 196errout: 197 return err; 198} 199 200/* 201 * Slow path: 202 * At least one argument has the form of 'offset($rx)'. 203 * 204 * Following code first stores them into stack, then loads all of then 205 * to r2 - r5. 206 * Before final loading, the final result should be: 207 * 208 * low address 209 * BPF_REG_FP - 24 ARG3 210 * BPF_REG_FP - 16 ARG2 211 * BPF_REG_FP - 8 ARG1 212 * BPF_REG_FP 213 * high address 214 * 215 * For each argument (described as: offn(...off2(off1(reg)))), 216 * generates following code: 217 * 218 * r7 <- fp 219 * r7 <- r7 - stack_offset // Ideal code should initialize r7 using 220 * // fp before generating args. However, 221 * // eBPF won't regard r7 as stack pointer 222 * // if it is generated by minus 8 from 223 * // another stack pointer except fp. 224 * // This is why we have to set r7 225 * // to fp for each variable. 226 * r3 <- value of 'reg'-> generated using gen_ldx_reg_from_ctx() 227 * (r7) <- r3 // skip following instructions for bare reg 228 * r3 <- r3 + off1 . // skip if off1 == 0 229 * r2 <- 8 \ 230 * r1 <- r7 |-> generated by gen_read_mem() 231 * call probe_read / 232 * jnei r0, 0, err ./ 233 * r3 <- (r7) 234 * r3 <- r3 + off2 . // skip if off2 == 0 235 * r2 <- 8 \ // r2 may be broken by probe_read, so set again 236 * r1 <- r7 |-> generated by gen_read_mem() 237 * call probe_read / 238 * jnei r0, 0, err ./ 239 * ... 240 */ 241static int 242gen_prologue_slowpath(struct bpf_insn_pos *pos, 243 struct probe_trace_arg *args, int nargs) 244{ 245 int err, i, probeid; 246 247 for (i = 0; i < nargs; i++) { 248 struct probe_trace_arg *arg = &args[i]; 249 const char *reg = arg->value; 250 struct probe_trace_arg_ref *ref = NULL; 251 int stack_offset = (i + 1) * -8; 252 253 pr_debug("prologue: fetch arg %d, base reg is %s\n", 254 i, reg); 255 256 /* value of base register is stored into ARG3 */ 257 err = gen_ldx_reg_from_ctx(pos, BPF_REG_CTX, reg, 258 BPF_REG_ARG3); 259 if (err) { 260 pr_err("prologue: failed to get offset of register %s\n", 261 reg); 262 goto errout; 263 } 264 265 /* Make r7 the stack pointer. */ 266 ins(BPF_MOV64_REG(BPF_REG_7, BPF_REG_FP), pos); 267 /* r7 += -8 */ 268 ins(BPF_ALU64_IMM(BPF_ADD, BPF_REG_7, stack_offset), pos); 269 /* 270 * Store r3 (base register) onto stack 271 * Ensure fp[offset] is set. 272 * fp is the only valid base register when storing 273 * into stack. We are not allowed to use r7 as base 274 * register here. 275 */ 276 ins(BPF_STX_MEM(BPF_DW, BPF_REG_FP, BPF_REG_ARG3, 277 stack_offset), pos); 278 279 ref = arg->ref; 280 probeid = BPF_FUNC_probe_read_kernel; 281 while (ref) { 282 pr_debug("prologue: arg %d: offset %ld\n", 283 i, ref->offset); 284 285 if (ref->user_access) 286 probeid = BPF_FUNC_probe_read_user; 287 288 err = gen_read_mem(pos, BPF_REG_3, BPF_REG_7, 289 ref->offset, probeid); 290 if (err) { 291 pr_err("prologue: failed to generate probe_read function call\n"); 292 goto errout; 293 } 294 295 ref = ref->next; 296 /* 297 * Load previous result into ARG3. Use 298 * BPF_REG_FP instead of r7 because verifier 299 * allows FP based addressing only. 300 */ 301 if (ref) 302 ins(BPF_LDX_MEM(BPF_DW, BPF_REG_ARG3, 303 BPF_REG_FP, stack_offset), pos); 304 } 305 } 306 307 /* Final pass: read to registers */ 308 for (i = 0; i < nargs; i++) { 309 int insn_sz = (args[i].ref) ? argtype_to_ldx_size(args[i].type) : BPF_DW; 310 311 pr_debug("prologue: load arg %d, insn_sz is %s\n", 312 i, insn_sz_to_str(insn_sz)); 313 ins(BPF_LDX_MEM(insn_sz, BPF_PROLOGUE_START_ARG_REG + i, 314 BPF_REG_FP, -BPF_REG_SIZE * (i + 1)), pos); 315 } 316 317 ins(BPF_JMP_IMM(BPF_JA, BPF_REG_0, 0, JMP_TO_SUCCESS_CODE), pos); 318 319 return check_pos(pos); 320errout: 321 return err; 322} 323 324static int 325prologue_relocate(struct bpf_insn_pos *pos, struct bpf_insn *error_code, 326 struct bpf_insn *success_code, struct bpf_insn *user_code) 327{ 328 struct bpf_insn *insn; 329 330 if (check_pos(pos)) 331 return -BPF_LOADER_ERRNO__PROLOGUE2BIG; 332 333 for (insn = pos->begin; insn < pos->pos; insn++) { 334 struct bpf_insn *target; 335 u8 class = BPF_CLASS(insn->code); 336 u8 opcode; 337 338 if (class != BPF_JMP) 339 continue; 340 opcode = BPF_OP(insn->code); 341 if (opcode == BPF_CALL) 342 continue; 343 344 switch (insn->off) { 345 case JMP_TO_ERROR_CODE: 346 target = error_code; 347 break; 348 case JMP_TO_SUCCESS_CODE: 349 target = success_code; 350 break; 351 case JMP_TO_USER_CODE: 352 target = user_code; 353 break; 354 default: 355 pr_err("bpf prologue: internal error: relocation failed\n"); 356 return -BPF_LOADER_ERRNO__PROLOGUE; 357 } 358 359 insn->off = target - (insn + 1); 360 } 361 return 0; 362} 363 364int bpf__gen_prologue(struct probe_trace_arg *args, int nargs, 365 struct bpf_insn *new_prog, size_t *new_cnt, 366 size_t cnt_space) 367{ 368 struct bpf_insn *success_code = NULL; 369 struct bpf_insn *error_code = NULL; 370 struct bpf_insn *user_code = NULL; 371 struct bpf_insn_pos pos; 372 bool fastpath = true; 373 int err = 0, i; 374 375 if (!new_prog || !new_cnt) 376 return -EINVAL; 377 378 if (cnt_space > BPF_MAXINSNS) 379 cnt_space = BPF_MAXINSNS; 380 381 pos.begin = new_prog; 382 pos.end = new_prog + cnt_space; 383 pos.pos = new_prog; 384 385 if (!nargs) { 386 ins(BPF_ALU64_IMM(BPF_MOV, BPF_PROLOGUE_FETCH_RESULT_REG, 0), 387 &pos); 388 389 if (check_pos(&pos)) 390 goto errout; 391 392 *new_cnt = pos_get_cnt(&pos); 393 return 0; 394 } 395 396 if (nargs > BPF_PROLOGUE_MAX_ARGS) { 397 pr_warning("bpf: prologue: %d arguments are dropped\n", 398 nargs - BPF_PROLOGUE_MAX_ARGS); 399 nargs = BPF_PROLOGUE_MAX_ARGS; 400 } 401 402 /* First pass: validation */ 403 for (i = 0; i < nargs; i++) { 404 struct probe_trace_arg_ref *ref = args[i].ref; 405 406 if (args[i].value[0] == '@') { 407 /* TODO: fetch global variable */ 408 pr_err("bpf: prologue: global %s%+ld not support\n", 409 args[i].value, ref ? ref->offset : 0); 410 return -ENOTSUP; 411 } 412 413 while (ref) { 414 /* fastpath is true if all args has ref == NULL */ 415 fastpath = false; 416 417 /* 418 * Instruction encodes immediate value using 419 * s32, ref->offset is long. On systems which 420 * can't fill long in s32, refuse to process if 421 * ref->offset too large (or small). 422 */ 423#ifdef __LP64__ 424#define OFFSET_MAX ((1LL << 31) - 1) 425#define OFFSET_MIN ((1LL << 31) * -1) 426 if (ref->offset > OFFSET_MAX || 427 ref->offset < OFFSET_MIN) { 428 pr_err("bpf: prologue: offset out of bound: %ld\n", 429 ref->offset); 430 return -BPF_LOADER_ERRNO__PROLOGUEOOB; 431 } 432#endif 433 ref = ref->next; 434 } 435 } 436 pr_debug("prologue: pass validation\n"); 437 438 if (fastpath) { 439 /* If all variables are registers... */ 440 pr_debug("prologue: fast path\n"); 441 err = gen_prologue_fastpath(&pos, args, nargs); 442 if (err) 443 goto errout; 444 } else { 445 pr_debug("prologue: slow path\n"); 446 447 /* Initialization: move ctx to a callee saved register. */ 448 ins(BPF_MOV64_REG(BPF_REG_CTX, BPF_REG_ARG1), &pos); 449 450 err = gen_prologue_slowpath(&pos, args, nargs); 451 if (err) 452 goto errout; 453 /* 454 * start of ERROR_CODE (only slow pass needs error code) 455 * mov r2 <- 1 // r2 is error number 456 * mov r3 <- 0 // r3, r4... should be touched or 457 * // verifier would complain 458 * mov r4 <- 0 459 * ... 460 * goto usercode 461 */ 462 error_code = pos.pos; 463 ins(BPF_ALU64_IMM(BPF_MOV, BPF_PROLOGUE_FETCH_RESULT_REG, 1), 464 &pos); 465 466 for (i = 0; i < nargs; i++) 467 ins(BPF_ALU64_IMM(BPF_MOV, 468 BPF_PROLOGUE_START_ARG_REG + i, 469 0), 470 &pos); 471 ins(BPF_JMP_IMM(BPF_JA, BPF_REG_0, 0, JMP_TO_USER_CODE), 472 &pos); 473 } 474 475 /* 476 * start of SUCCESS_CODE: 477 * mov r2 <- 0 478 * goto usercode // skip 479 */ 480 success_code = pos.pos; 481 ins(BPF_ALU64_IMM(BPF_MOV, BPF_PROLOGUE_FETCH_RESULT_REG, 0), &pos); 482 483 /* 484 * start of USER_CODE: 485 * Restore ctx to r1 486 */ 487 user_code = pos.pos; 488 if (!fastpath) { 489 /* 490 * Only slow path needs restoring of ctx. In fast path, 491 * register are loaded directly from r1. 492 */ 493 ins(BPF_MOV64_REG(BPF_REG_ARG1, BPF_REG_CTX), &pos); 494 err = prologue_relocate(&pos, error_code, success_code, 495 user_code); 496 if (err) 497 goto errout; 498 } 499 500 err = check_pos(&pos); 501 if (err) 502 goto errout; 503 504 *new_cnt = pos_get_cnt(&pos); 505 return 0; 506errout: 507 return err; 508}