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 v4.12-rc5 456 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 <errno.h> 16#include <dwarf-regs.h> 17#include <linux/filter.h> 18 19#define BPF_REG_SIZE 8 20 21#define JMP_TO_ERROR_CODE -1 22#define JMP_TO_SUCCESS_CODE -2 23#define JMP_TO_USER_CODE -3 24 25struct bpf_insn_pos { 26 struct bpf_insn *begin; 27 struct bpf_insn *end; 28 struct bpf_insn *pos; 29}; 30 31static inline int 32pos_get_cnt(struct bpf_insn_pos *pos) 33{ 34 return pos->pos - pos->begin; 35} 36 37static int 38append_insn(struct bpf_insn new_insn, struct bpf_insn_pos *pos) 39{ 40 if (!pos->pos) 41 return -BPF_LOADER_ERRNO__PROLOGUE2BIG; 42 43 if (pos->pos + 1 >= pos->end) { 44 pr_err("bpf prologue: prologue too long\n"); 45 pos->pos = NULL; 46 return -BPF_LOADER_ERRNO__PROLOGUE2BIG; 47 } 48 49 *(pos->pos)++ = new_insn; 50 return 0; 51} 52 53static int 54check_pos(struct bpf_insn_pos *pos) 55{ 56 if (!pos->pos || pos->pos >= pos->end) 57 return -BPF_LOADER_ERRNO__PROLOGUE2BIG; 58 return 0; 59} 60 61/* Give it a shorter name */ 62#define ins(i, p) append_insn((i), (p)) 63 64/* 65 * Give a register name (in 'reg'), generate instruction to 66 * load register into an eBPF register rd: 67 * 'ldd target_reg, offset(ctx_reg)', where: 68 * ctx_reg is pre initialized to pointer of 'struct pt_regs'. 69 */ 70static int 71gen_ldx_reg_from_ctx(struct bpf_insn_pos *pos, int ctx_reg, 72 const char *reg, int target_reg) 73{ 74 int offset = regs_query_register_offset(reg); 75 76 if (offset < 0) { 77 pr_err("bpf: prologue: failed to get register %s\n", 78 reg); 79 return offset; 80 } 81 ins(BPF_LDX_MEM(BPF_DW, target_reg, ctx_reg, offset), pos); 82 83 return check_pos(pos); 84} 85 86/* 87 * Generate a BPF_FUNC_probe_read function call. 88 * 89 * src_base_addr_reg is a register holding base address, 90 * dst_addr_reg is a register holding dest address (on stack), 91 * result is: 92 * 93 * *[dst_addr_reg] = *([src_base_addr_reg] + offset) 94 * 95 * Arguments of BPF_FUNC_probe_read: 96 * ARG1: ptr to stack (dest) 97 * ARG2: size (8) 98 * ARG3: unsafe ptr (src) 99 */ 100static int 101gen_read_mem(struct bpf_insn_pos *pos, 102 int src_base_addr_reg, 103 int dst_addr_reg, 104 long offset) 105{ 106 /* mov arg3, src_base_addr_reg */ 107 if (src_base_addr_reg != BPF_REG_ARG3) 108 ins(BPF_MOV64_REG(BPF_REG_ARG3, src_base_addr_reg), pos); 109 /* add arg3, #offset */ 110 if (offset) 111 ins(BPF_ALU64_IMM(BPF_ADD, BPF_REG_ARG3, offset), pos); 112 113 /* mov arg2, #reg_size */ 114 ins(BPF_ALU64_IMM(BPF_MOV, BPF_REG_ARG2, BPF_REG_SIZE), pos); 115 116 /* mov arg1, dst_addr_reg */ 117 if (dst_addr_reg != BPF_REG_ARG1) 118 ins(BPF_MOV64_REG(BPF_REG_ARG1, dst_addr_reg), pos); 119 120 /* Call probe_read */ 121 ins(BPF_EMIT_CALL(BPF_FUNC_probe_read), pos); 122 /* 123 * Error processing: if read fail, goto error code, 124 * will be relocated. Target should be the start of 125 * error processing code. 126 */ 127 ins(BPF_JMP_IMM(BPF_JNE, BPF_REG_0, 0, JMP_TO_ERROR_CODE), 128 pos); 129 130 return check_pos(pos); 131} 132 133/* 134 * Each arg should be bare register. Fetch and save them into argument 135 * registers (r3 - r5). 136 * 137 * BPF_REG_1 should have been initialized with pointer to 138 * 'struct pt_regs'. 139 */ 140static int 141gen_prologue_fastpath(struct bpf_insn_pos *pos, 142 struct probe_trace_arg *args, int nargs) 143{ 144 int i, err = 0; 145 146 for (i = 0; i < nargs; i++) { 147 err = gen_ldx_reg_from_ctx(pos, BPF_REG_1, args[i].value, 148 BPF_PROLOGUE_START_ARG_REG + i); 149 if (err) 150 goto errout; 151 } 152 153 return check_pos(pos); 154errout: 155 return err; 156} 157 158/* 159 * Slow path: 160 * At least one argument has the form of 'offset($rx)'. 161 * 162 * Following code first stores them into stack, then loads all of then 163 * to r2 - r5. 164 * Before final loading, the final result should be: 165 * 166 * low address 167 * BPF_REG_FP - 24 ARG3 168 * BPF_REG_FP - 16 ARG2 169 * BPF_REG_FP - 8 ARG1 170 * BPF_REG_FP 171 * high address 172 * 173 * For each argument (described as: offn(...off2(off1(reg)))), 174 * generates following code: 175 * 176 * r7 <- fp 177 * r7 <- r7 - stack_offset // Ideal code should initialize r7 using 178 * // fp before generating args. However, 179 * // eBPF won't regard r7 as stack pointer 180 * // if it is generated by minus 8 from 181 * // another stack pointer except fp. 182 * // This is why we have to set r7 183 * // to fp for each variable. 184 * r3 <- value of 'reg'-> generated using gen_ldx_reg_from_ctx() 185 * (r7) <- r3 // skip following instructions for bare reg 186 * r3 <- r3 + off1 . // skip if off1 == 0 187 * r2 <- 8 \ 188 * r1 <- r7 |-> generated by gen_read_mem() 189 * call probe_read / 190 * jnei r0, 0, err ./ 191 * r3 <- (r7) 192 * r3 <- r3 + off2 . // skip if off2 == 0 193 * r2 <- 8 \ // r2 may be broken by probe_read, so set again 194 * r1 <- r7 |-> generated by gen_read_mem() 195 * call probe_read / 196 * jnei r0, 0, err ./ 197 * ... 198 */ 199static int 200gen_prologue_slowpath(struct bpf_insn_pos *pos, 201 struct probe_trace_arg *args, int nargs) 202{ 203 int err, i; 204 205 for (i = 0; i < nargs; i++) { 206 struct probe_trace_arg *arg = &args[i]; 207 const char *reg = arg->value; 208 struct probe_trace_arg_ref *ref = NULL; 209 int stack_offset = (i + 1) * -8; 210 211 pr_debug("prologue: fetch arg %d, base reg is %s\n", 212 i, reg); 213 214 /* value of base register is stored into ARG3 */ 215 err = gen_ldx_reg_from_ctx(pos, BPF_REG_CTX, reg, 216 BPF_REG_ARG3); 217 if (err) { 218 pr_err("prologue: failed to get offset of register %s\n", 219 reg); 220 goto errout; 221 } 222 223 /* Make r7 the stack pointer. */ 224 ins(BPF_MOV64_REG(BPF_REG_7, BPF_REG_FP), pos); 225 /* r7 += -8 */ 226 ins(BPF_ALU64_IMM(BPF_ADD, BPF_REG_7, stack_offset), pos); 227 /* 228 * Store r3 (base register) onto stack 229 * Ensure fp[offset] is set. 230 * fp is the only valid base register when storing 231 * into stack. We are not allowed to use r7 as base 232 * register here. 233 */ 234 ins(BPF_STX_MEM(BPF_DW, BPF_REG_FP, BPF_REG_ARG3, 235 stack_offset), pos); 236 237 ref = arg->ref; 238 while (ref) { 239 pr_debug("prologue: arg %d: offset %ld\n", 240 i, ref->offset); 241 err = gen_read_mem(pos, BPF_REG_3, BPF_REG_7, 242 ref->offset); 243 if (err) { 244 pr_err("prologue: failed to generate probe_read function call\n"); 245 goto errout; 246 } 247 248 ref = ref->next; 249 /* 250 * Load previous result into ARG3. Use 251 * BPF_REG_FP instead of r7 because verifier 252 * allows FP based addressing only. 253 */ 254 if (ref) 255 ins(BPF_LDX_MEM(BPF_DW, BPF_REG_ARG3, 256 BPF_REG_FP, stack_offset), pos); 257 } 258 } 259 260 /* Final pass: read to registers */ 261 for (i = 0; i < nargs; i++) 262 ins(BPF_LDX_MEM(BPF_DW, BPF_PROLOGUE_START_ARG_REG + i, 263 BPF_REG_FP, -BPF_REG_SIZE * (i + 1)), pos); 264 265 ins(BPF_JMP_IMM(BPF_JA, BPF_REG_0, 0, JMP_TO_SUCCESS_CODE), pos); 266 267 return check_pos(pos); 268errout: 269 return err; 270} 271 272static int 273prologue_relocate(struct bpf_insn_pos *pos, struct bpf_insn *error_code, 274 struct bpf_insn *success_code, struct bpf_insn *user_code) 275{ 276 struct bpf_insn *insn; 277 278 if (check_pos(pos)) 279 return -BPF_LOADER_ERRNO__PROLOGUE2BIG; 280 281 for (insn = pos->begin; insn < pos->pos; insn++) { 282 struct bpf_insn *target; 283 u8 class = BPF_CLASS(insn->code); 284 u8 opcode; 285 286 if (class != BPF_JMP) 287 continue; 288 opcode = BPF_OP(insn->code); 289 if (opcode == BPF_CALL) 290 continue; 291 292 switch (insn->off) { 293 case JMP_TO_ERROR_CODE: 294 target = error_code; 295 break; 296 case JMP_TO_SUCCESS_CODE: 297 target = success_code; 298 break; 299 case JMP_TO_USER_CODE: 300 target = user_code; 301 break; 302 default: 303 pr_err("bpf prologue: internal error: relocation failed\n"); 304 return -BPF_LOADER_ERRNO__PROLOGUE; 305 } 306 307 insn->off = target - (insn + 1); 308 } 309 return 0; 310} 311 312int bpf__gen_prologue(struct probe_trace_arg *args, int nargs, 313 struct bpf_insn *new_prog, size_t *new_cnt, 314 size_t cnt_space) 315{ 316 struct bpf_insn *success_code = NULL; 317 struct bpf_insn *error_code = NULL; 318 struct bpf_insn *user_code = NULL; 319 struct bpf_insn_pos pos; 320 bool fastpath = true; 321 int err = 0, i; 322 323 if (!new_prog || !new_cnt) 324 return -EINVAL; 325 326 if (cnt_space > BPF_MAXINSNS) 327 cnt_space = BPF_MAXINSNS; 328 329 pos.begin = new_prog; 330 pos.end = new_prog + cnt_space; 331 pos.pos = new_prog; 332 333 if (!nargs) { 334 ins(BPF_ALU64_IMM(BPF_MOV, BPF_PROLOGUE_FETCH_RESULT_REG, 0), 335 &pos); 336 337 if (check_pos(&pos)) 338 goto errout; 339 340 *new_cnt = pos_get_cnt(&pos); 341 return 0; 342 } 343 344 if (nargs > BPF_PROLOGUE_MAX_ARGS) { 345 pr_warning("bpf: prologue: %d arguments are dropped\n", 346 nargs - BPF_PROLOGUE_MAX_ARGS); 347 nargs = BPF_PROLOGUE_MAX_ARGS; 348 } 349 350 /* First pass: validation */ 351 for (i = 0; i < nargs; i++) { 352 struct probe_trace_arg_ref *ref = args[i].ref; 353 354 if (args[i].value[0] == '@') { 355 /* TODO: fetch global variable */ 356 pr_err("bpf: prologue: global %s%+ld not support\n", 357 args[i].value, ref ? ref->offset : 0); 358 return -ENOTSUP; 359 } 360 361 while (ref) { 362 /* fastpath is true if all args has ref == NULL */ 363 fastpath = false; 364 365 /* 366 * Instruction encodes immediate value using 367 * s32, ref->offset is long. On systems which 368 * can't fill long in s32, refuse to process if 369 * ref->offset too large (or small). 370 */ 371#ifdef __LP64__ 372#define OFFSET_MAX ((1LL << 31) - 1) 373#define OFFSET_MIN ((1LL << 31) * -1) 374 if (ref->offset > OFFSET_MAX || 375 ref->offset < OFFSET_MIN) { 376 pr_err("bpf: prologue: offset out of bound: %ld\n", 377 ref->offset); 378 return -BPF_LOADER_ERRNO__PROLOGUEOOB; 379 } 380#endif 381 ref = ref->next; 382 } 383 } 384 pr_debug("prologue: pass validation\n"); 385 386 if (fastpath) { 387 /* If all variables are registers... */ 388 pr_debug("prologue: fast path\n"); 389 err = gen_prologue_fastpath(&pos, args, nargs); 390 if (err) 391 goto errout; 392 } else { 393 pr_debug("prologue: slow path\n"); 394 395 /* Initialization: move ctx to a callee saved register. */ 396 ins(BPF_MOV64_REG(BPF_REG_CTX, BPF_REG_ARG1), &pos); 397 398 err = gen_prologue_slowpath(&pos, args, nargs); 399 if (err) 400 goto errout; 401 /* 402 * start of ERROR_CODE (only slow pass needs error code) 403 * mov r2 <- 1 // r2 is error number 404 * mov r3 <- 0 // r3, r4... should be touched or 405 * // verifier would complain 406 * mov r4 <- 0 407 * ... 408 * goto usercode 409 */ 410 error_code = pos.pos; 411 ins(BPF_ALU64_IMM(BPF_MOV, BPF_PROLOGUE_FETCH_RESULT_REG, 1), 412 &pos); 413 414 for (i = 0; i < nargs; i++) 415 ins(BPF_ALU64_IMM(BPF_MOV, 416 BPF_PROLOGUE_START_ARG_REG + i, 417 0), 418 &pos); 419 ins(BPF_JMP_IMM(BPF_JA, BPF_REG_0, 0, JMP_TO_USER_CODE), 420 &pos); 421 } 422 423 /* 424 * start of SUCCESS_CODE: 425 * mov r2 <- 0 426 * goto usercode // skip 427 */ 428 success_code = pos.pos; 429 ins(BPF_ALU64_IMM(BPF_MOV, BPF_PROLOGUE_FETCH_RESULT_REG, 0), &pos); 430 431 /* 432 * start of USER_CODE: 433 * Restore ctx to r1 434 */ 435 user_code = pos.pos; 436 if (!fastpath) { 437 /* 438 * Only slow path needs restoring of ctx. In fast path, 439 * register are loaded directly from r1. 440 */ 441 ins(BPF_MOV64_REG(BPF_REG_ARG1, BPF_REG_CTX), &pos); 442 err = prologue_relocate(&pos, error_code, success_code, 443 user_code); 444 if (err) 445 goto errout; 446 } 447 448 err = check_pos(&pos); 449 if (err) 450 goto errout; 451 452 *new_cnt = pos_get_cnt(&pos); 453 return 0; 454errout: 455 return err; 456}