Linux kernel mirror (for testing) git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel os linux
1
fork

Configure Feed

Select the types of activity you want to include in your feed.

at v4.7 455 lines 11 kB view raw
1/* 2 * bpf-prologue.c 3 * 4 * Copyright (C) 2015 He Kuang <hekuang@huawei.com> 5 * Copyright (C) 2015 Wang Nan <wangnan0@huawei.com> 6 * Copyright (C) 2015 Huawei Inc. 7 */ 8 9#include <bpf/libbpf.h> 10#include "perf.h" 11#include "debug.h" 12#include "bpf-loader.h" 13#include "bpf-prologue.h" 14#include "probe-finder.h" 15#include <dwarf-regs.h> 16#include <linux/filter.h> 17 18#define BPF_REG_SIZE 8 19 20#define JMP_TO_ERROR_CODE -1 21#define JMP_TO_SUCCESS_CODE -2 22#define JMP_TO_USER_CODE -3 23 24struct bpf_insn_pos { 25 struct bpf_insn *begin; 26 struct bpf_insn *end; 27 struct bpf_insn *pos; 28}; 29 30static inline int 31pos_get_cnt(struct bpf_insn_pos *pos) 32{ 33 return pos->pos - pos->begin; 34} 35 36static int 37append_insn(struct bpf_insn new_insn, struct bpf_insn_pos *pos) 38{ 39 if (!pos->pos) 40 return -BPF_LOADER_ERRNO__PROLOGUE2BIG; 41 42 if (pos->pos + 1 >= pos->end) { 43 pr_err("bpf prologue: prologue too long\n"); 44 pos->pos = NULL; 45 return -BPF_LOADER_ERRNO__PROLOGUE2BIG; 46 } 47 48 *(pos->pos)++ = new_insn; 49 return 0; 50} 51 52static int 53check_pos(struct bpf_insn_pos *pos) 54{ 55 if (!pos->pos || pos->pos >= pos->end) 56 return -BPF_LOADER_ERRNO__PROLOGUE2BIG; 57 return 0; 58} 59 60/* Give it a shorter name */ 61#define ins(i, p) append_insn((i), (p)) 62 63/* 64 * Give a register name (in 'reg'), generate instruction to 65 * load register into an eBPF register rd: 66 * 'ldd target_reg, offset(ctx_reg)', where: 67 * ctx_reg is pre initialized to pointer of 'struct pt_regs'. 68 */ 69static int 70gen_ldx_reg_from_ctx(struct bpf_insn_pos *pos, int ctx_reg, 71 const char *reg, int target_reg) 72{ 73 int offset = regs_query_register_offset(reg); 74 75 if (offset < 0) { 76 pr_err("bpf: prologue: failed to get register %s\n", 77 reg); 78 return offset; 79 } 80 ins(BPF_LDX_MEM(BPF_DW, target_reg, ctx_reg, offset), pos); 81 82 return check_pos(pos); 83} 84 85/* 86 * Generate a BPF_FUNC_probe_read function call. 87 * 88 * src_base_addr_reg is a register holding base address, 89 * dst_addr_reg is a register holding dest address (on stack), 90 * result is: 91 * 92 * *[dst_addr_reg] = *([src_base_addr_reg] + offset) 93 * 94 * Arguments of BPF_FUNC_probe_read: 95 * ARG1: ptr to stack (dest) 96 * ARG2: size (8) 97 * ARG3: unsafe ptr (src) 98 */ 99static int 100gen_read_mem(struct bpf_insn_pos *pos, 101 int src_base_addr_reg, 102 int dst_addr_reg, 103 long offset) 104{ 105 /* mov arg3, src_base_addr_reg */ 106 if (src_base_addr_reg != BPF_REG_ARG3) 107 ins(BPF_MOV64_REG(BPF_REG_ARG3, src_base_addr_reg), pos); 108 /* add arg3, #offset */ 109 if (offset) 110 ins(BPF_ALU64_IMM(BPF_ADD, BPF_REG_ARG3, offset), pos); 111 112 /* mov arg2, #reg_size */ 113 ins(BPF_ALU64_IMM(BPF_MOV, BPF_REG_ARG2, BPF_REG_SIZE), pos); 114 115 /* mov arg1, dst_addr_reg */ 116 if (dst_addr_reg != BPF_REG_ARG1) 117 ins(BPF_MOV64_REG(BPF_REG_ARG1, dst_addr_reg), pos); 118 119 /* Call probe_read */ 120 ins(BPF_EMIT_CALL(BPF_FUNC_probe_read), pos); 121 /* 122 * Error processing: if read fail, goto error code, 123 * will be relocated. Target should be the start of 124 * error processing code. 125 */ 126 ins(BPF_JMP_IMM(BPF_JNE, BPF_REG_0, 0, JMP_TO_ERROR_CODE), 127 pos); 128 129 return check_pos(pos); 130} 131 132/* 133 * Each arg should be bare register. Fetch and save them into argument 134 * registers (r3 - r5). 135 * 136 * BPF_REG_1 should have been initialized with pointer to 137 * 'struct pt_regs'. 138 */ 139static int 140gen_prologue_fastpath(struct bpf_insn_pos *pos, 141 struct probe_trace_arg *args, int nargs) 142{ 143 int i, err = 0; 144 145 for (i = 0; i < nargs; i++) { 146 err = gen_ldx_reg_from_ctx(pos, BPF_REG_1, args[i].value, 147 BPF_PROLOGUE_START_ARG_REG + i); 148 if (err) 149 goto errout; 150 } 151 152 return check_pos(pos); 153errout: 154 return err; 155} 156 157/* 158 * Slow path: 159 * At least one argument has the form of 'offset($rx)'. 160 * 161 * Following code first stores them into stack, then loads all of then 162 * to r2 - r5. 163 * Before final loading, the final result should be: 164 * 165 * low address 166 * BPF_REG_FP - 24 ARG3 167 * BPF_REG_FP - 16 ARG2 168 * BPF_REG_FP - 8 ARG1 169 * BPF_REG_FP 170 * high address 171 * 172 * For each argument (described as: offn(...off2(off1(reg)))), 173 * generates following code: 174 * 175 * r7 <- fp 176 * r7 <- r7 - stack_offset // Ideal code should initialize r7 using 177 * // fp before generating args. However, 178 * // eBPF won't regard r7 as stack pointer 179 * // if it is generated by minus 8 from 180 * // another stack pointer except fp. 181 * // This is why we have to set r7 182 * // to fp for each variable. 183 * r3 <- value of 'reg'-> generated using gen_ldx_reg_from_ctx() 184 * (r7) <- r3 // skip following instructions for bare reg 185 * r3 <- r3 + off1 . // skip if off1 == 0 186 * r2 <- 8 \ 187 * r1 <- r7 |-> generated by gen_read_mem() 188 * call probe_read / 189 * jnei r0, 0, err ./ 190 * r3 <- (r7) 191 * r3 <- r3 + off2 . // skip if off2 == 0 192 * r2 <- 8 \ // r2 may be broken by probe_read, so set again 193 * r1 <- r7 |-> generated by gen_read_mem() 194 * call probe_read / 195 * jnei r0, 0, err ./ 196 * ... 197 */ 198static int 199gen_prologue_slowpath(struct bpf_insn_pos *pos, 200 struct probe_trace_arg *args, int nargs) 201{ 202 int err, i; 203 204 for (i = 0; i < nargs; i++) { 205 struct probe_trace_arg *arg = &args[i]; 206 const char *reg = arg->value; 207 struct probe_trace_arg_ref *ref = NULL; 208 int stack_offset = (i + 1) * -8; 209 210 pr_debug("prologue: fetch arg %d, base reg is %s\n", 211 i, reg); 212 213 /* value of base register is stored into ARG3 */ 214 err = gen_ldx_reg_from_ctx(pos, BPF_REG_CTX, reg, 215 BPF_REG_ARG3); 216 if (err) { 217 pr_err("prologue: failed to get offset of register %s\n", 218 reg); 219 goto errout; 220 } 221 222 /* Make r7 the stack pointer. */ 223 ins(BPF_MOV64_REG(BPF_REG_7, BPF_REG_FP), pos); 224 /* r7 += -8 */ 225 ins(BPF_ALU64_IMM(BPF_ADD, BPF_REG_7, stack_offset), pos); 226 /* 227 * Store r3 (base register) onto stack 228 * Ensure fp[offset] is set. 229 * fp is the only valid base register when storing 230 * into stack. We are not allowed to use r7 as base 231 * register here. 232 */ 233 ins(BPF_STX_MEM(BPF_DW, BPF_REG_FP, BPF_REG_ARG3, 234 stack_offset), pos); 235 236 ref = arg->ref; 237 while (ref) { 238 pr_debug("prologue: arg %d: offset %ld\n", 239 i, ref->offset); 240 err = gen_read_mem(pos, BPF_REG_3, BPF_REG_7, 241 ref->offset); 242 if (err) { 243 pr_err("prologue: failed to generate probe_read function call\n"); 244 goto errout; 245 } 246 247 ref = ref->next; 248 /* 249 * Load previous result into ARG3. Use 250 * BPF_REG_FP instead of r7 because verifier 251 * allows FP based addressing only. 252 */ 253 if (ref) 254 ins(BPF_LDX_MEM(BPF_DW, BPF_REG_ARG3, 255 BPF_REG_FP, stack_offset), pos); 256 } 257 } 258 259 /* Final pass: read to registers */ 260 for (i = 0; i < nargs; i++) 261 ins(BPF_LDX_MEM(BPF_DW, BPF_PROLOGUE_START_ARG_REG + i, 262 BPF_REG_FP, -BPF_REG_SIZE * (i + 1)), pos); 263 264 ins(BPF_JMP_IMM(BPF_JA, BPF_REG_0, 0, JMP_TO_SUCCESS_CODE), pos); 265 266 return check_pos(pos); 267errout: 268 return err; 269} 270 271static int 272prologue_relocate(struct bpf_insn_pos *pos, struct bpf_insn *error_code, 273 struct bpf_insn *success_code, struct bpf_insn *user_code) 274{ 275 struct bpf_insn *insn; 276 277 if (check_pos(pos)) 278 return -BPF_LOADER_ERRNO__PROLOGUE2BIG; 279 280 for (insn = pos->begin; insn < pos->pos; insn++) { 281 struct bpf_insn *target; 282 u8 class = BPF_CLASS(insn->code); 283 u8 opcode; 284 285 if (class != BPF_JMP) 286 continue; 287 opcode = BPF_OP(insn->code); 288 if (opcode == BPF_CALL) 289 continue; 290 291 switch (insn->off) { 292 case JMP_TO_ERROR_CODE: 293 target = error_code; 294 break; 295 case JMP_TO_SUCCESS_CODE: 296 target = success_code; 297 break; 298 case JMP_TO_USER_CODE: 299 target = user_code; 300 break; 301 default: 302 pr_err("bpf prologue: internal error: relocation failed\n"); 303 return -BPF_LOADER_ERRNO__PROLOGUE; 304 } 305 306 insn->off = target - (insn + 1); 307 } 308 return 0; 309} 310 311int bpf__gen_prologue(struct probe_trace_arg *args, int nargs, 312 struct bpf_insn *new_prog, size_t *new_cnt, 313 size_t cnt_space) 314{ 315 struct bpf_insn *success_code = NULL; 316 struct bpf_insn *error_code = NULL; 317 struct bpf_insn *user_code = NULL; 318 struct bpf_insn_pos pos; 319 bool fastpath = true; 320 int err = 0, i; 321 322 if (!new_prog || !new_cnt) 323 return -EINVAL; 324 325 if (cnt_space > BPF_MAXINSNS) 326 cnt_space = BPF_MAXINSNS; 327 328 pos.begin = new_prog; 329 pos.end = new_prog + cnt_space; 330 pos.pos = new_prog; 331 332 if (!nargs) { 333 ins(BPF_ALU64_IMM(BPF_MOV, BPF_PROLOGUE_FETCH_RESULT_REG, 0), 334 &pos); 335 336 if (check_pos(&pos)) 337 goto errout; 338 339 *new_cnt = pos_get_cnt(&pos); 340 return 0; 341 } 342 343 if (nargs > BPF_PROLOGUE_MAX_ARGS) { 344 pr_warning("bpf: prologue: %d arguments are dropped\n", 345 nargs - BPF_PROLOGUE_MAX_ARGS); 346 nargs = BPF_PROLOGUE_MAX_ARGS; 347 } 348 349 /* First pass: validation */ 350 for (i = 0; i < nargs; i++) { 351 struct probe_trace_arg_ref *ref = args[i].ref; 352 353 if (args[i].value[0] == '@') { 354 /* TODO: fetch global variable */ 355 pr_err("bpf: prologue: global %s%+ld not support\n", 356 args[i].value, ref ? ref->offset : 0); 357 return -ENOTSUP; 358 } 359 360 while (ref) { 361 /* fastpath is true if all args has ref == NULL */ 362 fastpath = false; 363 364 /* 365 * Instruction encodes immediate value using 366 * s32, ref->offset is long. On systems which 367 * can't fill long in s32, refuse to process if 368 * ref->offset too large (or small). 369 */ 370#ifdef __LP64__ 371#define OFFSET_MAX ((1LL << 31) - 1) 372#define OFFSET_MIN ((1LL << 31) * -1) 373 if (ref->offset > OFFSET_MAX || 374 ref->offset < OFFSET_MIN) { 375 pr_err("bpf: prologue: offset out of bound: %ld\n", 376 ref->offset); 377 return -BPF_LOADER_ERRNO__PROLOGUEOOB; 378 } 379#endif 380 ref = ref->next; 381 } 382 } 383 pr_debug("prologue: pass validation\n"); 384 385 if (fastpath) { 386 /* If all variables are registers... */ 387 pr_debug("prologue: fast path\n"); 388 err = gen_prologue_fastpath(&pos, args, nargs); 389 if (err) 390 goto errout; 391 } else { 392 pr_debug("prologue: slow path\n"); 393 394 /* Initialization: move ctx to a callee saved register. */ 395 ins(BPF_MOV64_REG(BPF_REG_CTX, BPF_REG_ARG1), &pos); 396 397 err = gen_prologue_slowpath(&pos, args, nargs); 398 if (err) 399 goto errout; 400 /* 401 * start of ERROR_CODE (only slow pass needs error code) 402 * mov r2 <- 1 // r2 is error number 403 * mov r3 <- 0 // r3, r4... should be touched or 404 * // verifier would complain 405 * mov r4 <- 0 406 * ... 407 * goto usercode 408 */ 409 error_code = pos.pos; 410 ins(BPF_ALU64_IMM(BPF_MOV, BPF_PROLOGUE_FETCH_RESULT_REG, 1), 411 &pos); 412 413 for (i = 0; i < nargs; i++) 414 ins(BPF_ALU64_IMM(BPF_MOV, 415 BPF_PROLOGUE_START_ARG_REG + i, 416 0), 417 &pos); 418 ins(BPF_JMP_IMM(BPF_JA, BPF_REG_0, 0, JMP_TO_USER_CODE), 419 &pos); 420 } 421 422 /* 423 * start of SUCCESS_CODE: 424 * mov r2 <- 0 425 * goto usercode // skip 426 */ 427 success_code = pos.pos; 428 ins(BPF_ALU64_IMM(BPF_MOV, BPF_PROLOGUE_FETCH_RESULT_REG, 0), &pos); 429 430 /* 431 * start of USER_CODE: 432 * Restore ctx to r1 433 */ 434 user_code = pos.pos; 435 if (!fastpath) { 436 /* 437 * Only slow path needs restoring of ctx. In fast path, 438 * register are loaded directly from r1. 439 */ 440 ins(BPF_MOV64_REG(BPF_REG_ARG1, BPF_REG_CTX), &pos); 441 err = prologue_relocate(&pos, error_code, success_code, 442 user_code); 443 if (err) 444 goto errout; 445 } 446 447 err = check_pos(&pos); 448 if (err) 449 goto errout; 450 451 *new_cnt = pos_get_cnt(&pos); 452 return 0; 453errout: 454 return err; 455}