1 /* 2 * BPF Jit compiler for s390. 3 * 4 * Copyright IBM Corp. 2012 5 * 6 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com> 7 */ 8 #include <linux/moduleloader.h> 9 #include <linux/netdevice.h> 10 #include <linux/if_vlan.h> 11 #include <linux/filter.h> 12 #include <linux/random.h> 13 #include <linux/init.h> 14 #include <asm/cacheflush.h> 15 #include <asm/processor.h> 16 #include <asm/facility.h> 17 18 /* 19 * Conventions: 20 * %r2 = skb pointer 21 * %r3 = offset parameter 22 * %r4 = scratch register / length parameter 23 * %r5 = BPF A accumulator 24 * %r8 = return address 25 * %r9 = save register for skb pointer 26 * %r10 = skb->data 27 * %r11 = skb->len - skb->data_len (headlen) 28 * %r12 = BPF X accumulator 29 * %r13 = literal pool pointer 30 * 0(%r15) - 63(%r15) scratch memory array with BPF_MEMWORDS 31 */ 32 int bpf_jit_enable __read_mostly; 33 34 /* 35 * assembly code in arch/x86/net/bpf_jit.S 36 */ 37 extern u8 sk_load_word[], sk_load_half[], sk_load_byte[], sk_load_byte_msh[]; 38 extern u8 sk_load_word_ind[], sk_load_half_ind[], sk_load_byte_ind[]; 39 40 struct bpf_jit { 41 unsigned int seen; 42 u8 *start; 43 u8 *prg; 44 u8 *mid; 45 u8 *lit; 46 u8 *end; 47 u8 *base_ip; 48 u8 *ret0_ip; 49 u8 *exit_ip; 50 unsigned int off_load_word; 51 unsigned int off_load_half; 52 unsigned int off_load_byte; 53 unsigned int off_load_bmsh; 54 unsigned int off_load_iword; 55 unsigned int off_load_ihalf; 56 unsigned int off_load_ibyte; 57 }; 58 59 #define BPF_SIZE_MAX 4096 /* Max size for program */ 60 61 #define SEEN_DATAREF 1 /* might call external helpers */ 62 #define SEEN_XREG 2 /* ebx is used */ 63 #define SEEN_MEM 4 /* use mem[] for temporary storage */ 64 #define SEEN_RET0 8 /* pc_ret0 points to a valid return 0 */ 65 #define SEEN_LITERAL 16 /* code uses literals */ 66 #define SEEN_LOAD_WORD 32 /* code uses sk_load_word */ 67 #define SEEN_LOAD_HALF 64 /* code uses sk_load_half */ 68 #define SEEN_LOAD_BYTE 128 /* code uses sk_load_byte */ 69 #define SEEN_LOAD_BMSH 256 /* code uses sk_load_byte_msh */ 70 #define SEEN_LOAD_IWORD 512 /* code uses sk_load_word_ind */ 71 #define SEEN_LOAD_IHALF 1024 /* code uses sk_load_half_ind */ 72 #define SEEN_LOAD_IBYTE 2048 /* code uses sk_load_byte_ind */ 73 74 #define EMIT2(op) \ 75 ({ \ 76 if (jit->prg + 2 <= jit->mid) \ 77 *(u16 *) jit->prg = op; \ 78 jit->prg += 2; \ 79 }) 80 81 #define EMIT4(op) \ 82 ({ \ 83 if (jit->prg + 4 <= jit->mid) \ 84 *(u32 *) jit->prg = op; \ 85 jit->prg += 4; \ 86 }) 87 88 #define EMIT4_DISP(op, disp) \ 89 ({ \ 90 unsigned int __disp = (disp) & 0xfff; \ 91 EMIT4(op | __disp); \ 92 }) 93 94 #define EMIT4_IMM(op, imm) \ 95 ({ \ 96 unsigned int __imm = (imm) & 0xffff; \ 97 EMIT4(op | __imm); \ 98 }) 99 100 #define EMIT4_PCREL(op, pcrel) \ 101 ({ \ 102 long __pcrel = ((pcrel) >> 1) & 0xffff; \ 103 EMIT4(op | __pcrel); \ 104 }) 105 106 #define EMIT6(op1, op2) \ 107 ({ \ 108 if (jit->prg + 6 <= jit->mid) { \ 109 *(u32 *) jit->prg = op1; \ 110 *(u16 *) (jit->prg + 4) = op2; \ 111 } \ 112 jit->prg += 6; \ 113 }) 114 115 #define EMIT6_DISP(op1, op2, disp) \ 116 ({ \ 117 unsigned int __disp = (disp) & 0xfff; \ 118 EMIT6(op1 | __disp, op2); \ 119 }) 120 121 #define EMIT6_IMM(op, imm) \ 122 ({ \ 123 unsigned int __imm = (imm); \ 124 EMIT6(op | (__imm >> 16), __imm & 0xffff); \ 125 }) 126 127 #define EMIT_CONST(val) \ 128 ({ \ 129 unsigned int ret; \ 130 ret = (unsigned int) (jit->lit - jit->base_ip); \ 131 jit->seen |= SEEN_LITERAL; \ 132 if (jit->lit + 4 <= jit->end) \ 133 *(u32 *) jit->lit = val; \ 134 jit->lit += 4; \ 135 ret; \ 136 }) 137 138 #define EMIT_FN_CONST(bit, fn) \ 139 ({ \ 140 unsigned int ret; \ 141 ret = (unsigned int) (jit->lit - jit->base_ip); \ 142 if (jit->seen & bit) { \ 143 jit->seen |= SEEN_LITERAL; \ 144 if (jit->lit + 8 <= jit->end) \ 145 *(void **) jit->lit = fn; \ 146 jit->lit += 8; \ 147 } \ 148 ret; \ 149 }) 150 151 static void bpf_jit_prologue(struct bpf_jit *jit) 152 { 153 /* Save registers and create stack frame if necessary */ 154 if (jit->seen & SEEN_DATAREF) { 155 /* stmg %r8,%r15,88(%r15) */ 156 EMIT6(0xeb8ff058, 0x0024); 157 /* lgr %r14,%r15 */ 158 EMIT4(0xb90400ef); 159 /* ahi %r15,<offset> */ 160 EMIT4_IMM(0xa7fa0000, (jit->seen & SEEN_MEM) ? -112 : -80); 161 /* stg %r14,152(%r15) */ 162 EMIT6(0xe3e0f098, 0x0024); 163 } else if ((jit->seen & SEEN_XREG) && (jit->seen & SEEN_LITERAL)) 164 /* stmg %r12,%r13,120(%r15) */ 165 EMIT6(0xebcdf078, 0x0024); 166 else if (jit->seen & SEEN_XREG) 167 /* stg %r12,120(%r15) */ 168 EMIT6(0xe3c0f078, 0x0024); 169 else if (jit->seen & SEEN_LITERAL) 170 /* stg %r13,128(%r15) */ 171 EMIT6(0xe3d0f080, 0x0024); 172 173 /* Setup literal pool */ 174 if (jit->seen & SEEN_LITERAL) { 175 /* basr %r13,0 */ 176 EMIT2(0x0dd0); 177 jit->base_ip = jit->prg; 178 } 179 jit->off_load_word = EMIT_FN_CONST(SEEN_LOAD_WORD, sk_load_word); 180 jit->off_load_half = EMIT_FN_CONST(SEEN_LOAD_HALF, sk_load_half); 181 jit->off_load_byte = EMIT_FN_CONST(SEEN_LOAD_BYTE, sk_load_byte); 182 jit->off_load_bmsh = EMIT_FN_CONST(SEEN_LOAD_BMSH, sk_load_byte_msh); 183 jit->off_load_iword = EMIT_FN_CONST(SEEN_LOAD_IWORD, sk_load_word_ind); 184 jit->off_load_ihalf = EMIT_FN_CONST(SEEN_LOAD_IHALF, sk_load_half_ind); 185 jit->off_load_ibyte = EMIT_FN_CONST(SEEN_LOAD_IBYTE, sk_load_byte_ind); 186 187 /* Filter needs to access skb data */ 188 if (jit->seen & SEEN_DATAREF) { 189 /* l %r11,<len>(%r2) */ 190 EMIT4_DISP(0x58b02000, offsetof(struct sk_buff, len)); 191 /* s %r11,<data_len>(%r2) */ 192 EMIT4_DISP(0x5bb02000, offsetof(struct sk_buff, data_len)); 193 /* lg %r10,<data>(%r2) */ 194 EMIT6_DISP(0xe3a02000, 0x0004, 195 offsetof(struct sk_buff, data)); 196 } 197 } 198 199 static void bpf_jit_epilogue(struct bpf_jit *jit) 200 { 201 /* Return 0 */ 202 if (jit->seen & SEEN_RET0) { 203 jit->ret0_ip = jit->prg; 204 /* lghi %r2,0 */ 205 EMIT4(0xa7290000); 206 } 207 jit->exit_ip = jit->prg; 208 /* Restore registers */ 209 if (jit->seen & SEEN_DATAREF) 210 /* lmg %r8,%r15,<offset>(%r15) */ 211 EMIT6_DISP(0xeb8ff000, 0x0004, 212 (jit->seen & SEEN_MEM) ? 200 : 168); 213 else if ((jit->seen & SEEN_XREG) && (jit->seen & SEEN_LITERAL)) 214 /* lmg %r12,%r13,120(%r15) */ 215 EMIT6(0xebcdf078, 0x0004); 216 else if (jit->seen & SEEN_XREG) 217 /* lg %r12,120(%r15) */ 218 EMIT6(0xe3c0f078, 0x0004); 219 else if (jit->seen & SEEN_LITERAL) 220 /* lg %r13,128(%r15) */ 221 EMIT6(0xe3d0f080, 0x0004); 222 /* br %r14 */ 223 EMIT2(0x07fe); 224 } 225 226 /* Helper to find the offset of pkt_type in sk_buff 227 * Make sure its still a 3bit field starting at the MSBs within a byte. 228 */ 229 #define PKT_TYPE_MAX 0xe0 230 static int pkt_type_offset; 231 232 static int __init bpf_pkt_type_offset_init(void) 233 { 234 struct sk_buff skb_probe = { 235 .pkt_type = ~0, 236 }; 237 char *ct = (char *)&skb_probe; 238 int off; 239 240 pkt_type_offset = -1; 241 for (off = 0; off < sizeof(struct sk_buff); off++) { 242 if (!ct[off]) 243 continue; 244 if (ct[off] == PKT_TYPE_MAX) 245 pkt_type_offset = off; 246 else { 247 /* Found non matching bit pattern, fix needed. */ 248 WARN_ON_ONCE(1); 249 pkt_type_offset = -1; 250 return -1; 251 } 252 } 253 return 0; 254 } 255 device_initcall(bpf_pkt_type_offset_init); 256 257 /* 258 * make sure we dont leak kernel information to user 259 */ 260 static void bpf_jit_noleaks(struct bpf_jit *jit, struct sock_filter *filter) 261 { 262 /* Clear temporary memory if (seen & SEEN_MEM) */ 263 if (jit->seen & SEEN_MEM) 264 /* xc 0(64,%r15),0(%r15) */ 265 EMIT6(0xd73ff000, 0xf000); 266 /* Clear X if (seen & SEEN_XREG) */ 267 if (jit->seen & SEEN_XREG) 268 /* lhi %r12,0 */ 269 EMIT4(0xa7c80000); 270 /* Clear A if the first register does not set it. */ 271 switch (filter[0].code) { 272 case BPF_S_LD_W_ABS: 273 case BPF_S_LD_H_ABS: 274 case BPF_S_LD_B_ABS: 275 case BPF_S_LD_W_LEN: 276 case BPF_S_LD_W_IND: 277 case BPF_S_LD_H_IND: 278 case BPF_S_LD_B_IND: 279 case BPF_S_LDX_B_MSH: 280 case BPF_S_LD_IMM: 281 case BPF_S_LD_MEM: 282 case BPF_S_MISC_TXA: 283 case BPF_S_ANC_PROTOCOL: 284 case BPF_S_ANC_PKTTYPE: 285 case BPF_S_ANC_IFINDEX: 286 case BPF_S_ANC_MARK: 287 case BPF_S_ANC_QUEUE: 288 case BPF_S_ANC_HATYPE: 289 case BPF_S_ANC_RXHASH: 290 case BPF_S_ANC_CPU: 291 case BPF_S_ANC_VLAN_TAG: 292 case BPF_S_ANC_VLAN_TAG_PRESENT: 293 case BPF_S_RET_K: 294 /* first instruction sets A register */ 295 break; 296 default: /* A = 0 */ 297 /* lhi %r5,0 */ 298 EMIT4(0xa7580000); 299 } 300 } 301 302 static int bpf_jit_insn(struct bpf_jit *jit, struct sock_filter *filter, 303 unsigned int *addrs, int i, int last) 304 { 305 unsigned int K; 306 int offset; 307 unsigned int mask; 308 309 K = filter->k; 310 switch (filter->code) { 311 case BPF_S_ALU_ADD_X: /* A += X */ 312 jit->seen |= SEEN_XREG; 313 /* ar %r5,%r12 */ 314 EMIT2(0x1a5c); 315 break; 316 case BPF_S_ALU_ADD_K: /* A += K */ 317 if (!K) 318 break; 319 if (K <= 16383) 320 /* ahi %r5,<K> */ 321 EMIT4_IMM(0xa75a0000, K); 322 else if (test_facility(21)) 323 /* alfi %r5,<K> */ 324 EMIT6_IMM(0xc25b0000, K); 325 else 326 /* a %r5,<d(K)>(%r13) */ 327 EMIT4_DISP(0x5a50d000, EMIT_CONST(K)); 328 break; 329 case BPF_S_ALU_SUB_X: /* A -= X */ 330 jit->seen |= SEEN_XREG; 331 /* sr %r5,%r12 */ 332 EMIT2(0x1b5c); 333 break; 334 case BPF_S_ALU_SUB_K: /* A -= K */ 335 if (!K) 336 break; 337 if (K <= 16384) 338 /* ahi %r5,-K */ 339 EMIT4_IMM(0xa75a0000, -K); 340 else if (test_facility(21)) 341 /* alfi %r5,-K */ 342 EMIT6_IMM(0xc25b0000, -K); 343 else 344 /* s %r5,<d(K)>(%r13) */ 345 EMIT4_DISP(0x5b50d000, EMIT_CONST(K)); 346 break; 347 case BPF_S_ALU_MUL_X: /* A *= X */ 348 jit->seen |= SEEN_XREG; 349 /* msr %r5,%r12 */ 350 EMIT4(0xb252005c); 351 break; 352 case BPF_S_ALU_MUL_K: /* A *= K */ 353 if (K <= 16383) 354 /* mhi %r5,K */ 355 EMIT4_IMM(0xa75c0000, K); 356 else if (test_facility(34)) 357 /* msfi %r5,<K> */ 358 EMIT6_IMM(0xc2510000, K); 359 else 360 /* ms %r5,<d(K)>(%r13) */ 361 EMIT4_DISP(0x7150d000, EMIT_CONST(K)); 362 break; 363 case BPF_S_ALU_DIV_X: /* A /= X */ 364 jit->seen |= SEEN_XREG | SEEN_RET0; 365 /* ltr %r12,%r12 */ 366 EMIT2(0x12cc); 367 /* jz <ret0> */ 368 EMIT4_PCREL(0xa7840000, (jit->ret0_ip - jit->prg)); 369 /* lhi %r4,0 */ 370 EMIT4(0xa7480000); 371 /* dr %r4,%r12 */ 372 EMIT2(0x1d4c); 373 break; 374 case BPF_S_ALU_DIV_K: /* A = reciprocal_divide(A, K) */ 375 /* m %r4,<d(K)>(%r13) */ 376 EMIT4_DISP(0x5c40d000, EMIT_CONST(K)); 377 /* lr %r5,%r4 */ 378 EMIT2(0x1854); 379 break; 380 case BPF_S_ALU_MOD_X: /* A %= X */ 381 jit->seen |= SEEN_XREG | SEEN_RET0; 382 /* ltr %r12,%r12 */ 383 EMIT2(0x12cc); 384 /* jz <ret0> */ 385 EMIT4_PCREL(0xa7840000, (jit->ret0_ip - jit->prg)); 386 /* lhi %r4,0 */ 387 EMIT4(0xa7480000); 388 /* dr %r4,%r12 */ 389 EMIT2(0x1d4c); 390 /* lr %r5,%r4 */ 391 EMIT2(0x1854); 392 break; 393 case BPF_S_ALU_MOD_K: /* A %= K */ 394 /* lhi %r4,0 */ 395 EMIT4(0xa7480000); 396 /* d %r4,<d(K)>(%r13) */ 397 EMIT4_DISP(0x5d40d000, EMIT_CONST(K)); 398 /* lr %r5,%r4 */ 399 EMIT2(0x1854); 400 break; 401 case BPF_S_ALU_AND_X: /* A &= X */ 402 jit->seen |= SEEN_XREG; 403 /* nr %r5,%r12 */ 404 EMIT2(0x145c); 405 break; 406 case BPF_S_ALU_AND_K: /* A &= K */ 407 if (test_facility(21)) 408 /* nilf %r5,<K> */ 409 EMIT6_IMM(0xc05b0000, K); 410 else 411 /* n %r5,<d(K)>(%r13) */ 412 EMIT4_DISP(0x5450d000, EMIT_CONST(K)); 413 break; 414 case BPF_S_ALU_OR_X: /* A |= X */ 415 jit->seen |= SEEN_XREG; 416 /* or %r5,%r12 */ 417 EMIT2(0x165c); 418 break; 419 case BPF_S_ALU_OR_K: /* A |= K */ 420 if (test_facility(21)) 421 /* oilf %r5,<K> */ 422 EMIT6_IMM(0xc05d0000, K); 423 else 424 /* o %r5,<d(K)>(%r13) */ 425 EMIT4_DISP(0x5650d000, EMIT_CONST(K)); 426 break; 427 case BPF_S_ANC_ALU_XOR_X: /* A ^= X; */ 428 case BPF_S_ALU_XOR_X: 429 jit->seen |= SEEN_XREG; 430 /* xr %r5,%r12 */ 431 EMIT2(0x175c); 432 break; 433 case BPF_S_ALU_XOR_K: /* A ^= K */ 434 if (!K) 435 break; 436 /* x %r5,<d(K)>(%r13) */ 437 EMIT4_DISP(0x5750d000, EMIT_CONST(K)); 438 break; 439 case BPF_S_ALU_LSH_X: /* A <<= X; */ 440 jit->seen |= SEEN_XREG; 441 /* sll %r5,0(%r12) */ 442 EMIT4(0x8950c000); 443 break; 444 case BPF_S_ALU_LSH_K: /* A <<= K */ 445 if (K == 0) 446 break; 447 /* sll %r5,K */ 448 EMIT4_DISP(0x89500000, K); 449 break; 450 case BPF_S_ALU_RSH_X: /* A >>= X; */ 451 jit->seen |= SEEN_XREG; 452 /* srl %r5,0(%r12) */ 453 EMIT4(0x8850c000); 454 break; 455 case BPF_S_ALU_RSH_K: /* A >>= K; */ 456 if (K == 0) 457 break; 458 /* srl %r5,K */ 459 EMIT4_DISP(0x88500000, K); 460 break; 461 case BPF_S_ALU_NEG: /* A = -A */ 462 /* lnr %r5,%r5 */ 463 EMIT2(0x1155); 464 break; 465 case BPF_S_JMP_JA: /* ip += K */ 466 offset = addrs[i + K] + jit->start - jit->prg; 467 EMIT4_PCREL(0xa7f40000, offset); 468 break; 469 case BPF_S_JMP_JGT_K: /* ip += (A > K) ? jt : jf */ 470 mask = 0x200000; /* jh */ 471 goto kbranch; 472 case BPF_S_JMP_JGE_K: /* ip += (A >= K) ? jt : jf */ 473 mask = 0xa00000; /* jhe */ 474 goto kbranch; 475 case BPF_S_JMP_JEQ_K: /* ip += (A == K) ? jt : jf */ 476 mask = 0x800000; /* je */ 477 kbranch: /* Emit compare if the branch targets are different */ 478 if (filter->jt != filter->jf) { 479 if (K <= 16383) 480 /* chi %r5,<K> */ 481 EMIT4_IMM(0xa75e0000, K); 482 else if (test_facility(21)) 483 /* clfi %r5,<K> */ 484 EMIT6_IMM(0xc25f0000, K); 485 else 486 /* c %r5,<d(K)>(%r13) */ 487 EMIT4_DISP(0x5950d000, EMIT_CONST(K)); 488 } 489 branch: if (filter->jt == filter->jf) { 490 if (filter->jt == 0) 491 break; 492 /* j <jt> */ 493 offset = addrs[i + filter->jt] + jit->start - jit->prg; 494 EMIT4_PCREL(0xa7f40000, offset); 495 break; 496 } 497 if (filter->jt != 0) { 498 /* brc <mask>,<jt> */ 499 offset = addrs[i + filter->jt] + jit->start - jit->prg; 500 EMIT4_PCREL(0xa7040000 | mask, offset); 501 } 502 if (filter->jf != 0) { 503 /* brc <mask^15>,<jf> */ 504 offset = addrs[i + filter->jf] + jit->start - jit->prg; 505 EMIT4_PCREL(0xa7040000 | (mask ^ 0xf00000), offset); 506 } 507 break; 508 case BPF_S_JMP_JSET_K: /* ip += (A & K) ? jt : jf */ 509 mask = 0x700000; /* jnz */ 510 /* Emit test if the branch targets are different */ 511 if (filter->jt != filter->jf) { 512 if (K > 65535) { 513 /* lr %r4,%r5 */ 514 EMIT2(0x1845); 515 /* n %r4,<d(K)>(%r13) */ 516 EMIT4_DISP(0x5440d000, EMIT_CONST(K)); 517 } else 518 /* tmll %r5,K */ 519 EMIT4_IMM(0xa7510000, K); 520 } 521 goto branch; 522 case BPF_S_JMP_JGT_X: /* ip += (A > X) ? jt : jf */ 523 mask = 0x200000; /* jh */ 524 goto xbranch; 525 case BPF_S_JMP_JGE_X: /* ip += (A >= X) ? jt : jf */ 526 mask = 0xa00000; /* jhe */ 527 goto xbranch; 528 case BPF_S_JMP_JEQ_X: /* ip += (A == X) ? jt : jf */ 529 mask = 0x800000; /* je */ 530 xbranch: /* Emit compare if the branch targets are different */ 531 if (filter->jt != filter->jf) { 532 jit->seen |= SEEN_XREG; 533 /* cr %r5,%r12 */ 534 EMIT2(0x195c); 535 } 536 goto branch; 537 case BPF_S_JMP_JSET_X: /* ip += (A & X) ? jt : jf */ 538 mask = 0x700000; /* jnz */ 539 /* Emit test if the branch targets are different */ 540 if (filter->jt != filter->jf) { 541 jit->seen |= SEEN_XREG; 542 /* lr %r4,%r5 */ 543 EMIT2(0x1845); 544 /* nr %r4,%r12 */ 545 EMIT2(0x144c); 546 } 547 goto branch; 548 case BPF_S_LD_W_ABS: /* A = *(u32 *) (skb->data+K) */ 549 jit->seen |= SEEN_DATAREF | SEEN_RET0 | SEEN_LOAD_WORD; 550 offset = jit->off_load_word; 551 goto load_abs; 552 case BPF_S_LD_H_ABS: /* A = *(u16 *) (skb->data+K) */ 553 jit->seen |= SEEN_DATAREF | SEEN_RET0 | SEEN_LOAD_HALF; 554 offset = jit->off_load_half; 555 goto load_abs; 556 case BPF_S_LD_B_ABS: /* A = *(u8 *) (skb->data+K) */ 557 jit->seen |= SEEN_DATAREF | SEEN_RET0 | SEEN_LOAD_BYTE; 558 offset = jit->off_load_byte; 559 load_abs: if ((int) K < 0) 560 goto out; 561 call_fn: /* lg %r1,<d(function)>(%r13) */ 562 EMIT6_DISP(0xe310d000, 0x0004, offset); 563 /* l %r3,<d(K)>(%r13) */ 564 EMIT4_DISP(0x5830d000, EMIT_CONST(K)); 565 /* basr %r8,%r1 */ 566 EMIT2(0x0d81); 567 /* jnz <ret0> */ 568 EMIT4_PCREL(0xa7740000, (jit->ret0_ip - jit->prg)); 569 break; 570 case BPF_S_LD_W_IND: /* A = *(u32 *) (skb->data+K+X) */ 571 jit->seen |= SEEN_DATAREF | SEEN_RET0 | SEEN_LOAD_IWORD; 572 offset = jit->off_load_iword; 573 goto call_fn; 574 case BPF_S_LD_H_IND: /* A = *(u16 *) (skb->data+K+X) */ 575 jit->seen |= SEEN_DATAREF | SEEN_RET0 | SEEN_LOAD_IHALF; 576 offset = jit->off_load_ihalf; 577 goto call_fn; 578 case BPF_S_LD_B_IND: /* A = *(u8 *) (skb->data+K+X) */ 579 jit->seen |= SEEN_DATAREF | SEEN_RET0 | SEEN_LOAD_IBYTE; 580 offset = jit->off_load_ibyte; 581 goto call_fn; 582 case BPF_S_LDX_B_MSH: 583 /* X = (*(u8 *)(skb->data+K) & 0xf) << 2 */ 584 jit->seen |= SEEN_RET0; 585 if ((int) K < 0) { 586 /* j <ret0> */ 587 EMIT4_PCREL(0xa7f40000, (jit->ret0_ip - jit->prg)); 588 break; 589 } 590 jit->seen |= SEEN_DATAREF | SEEN_LOAD_BMSH; 591 offset = jit->off_load_bmsh; 592 goto call_fn; 593 case BPF_S_LD_W_LEN: /* A = skb->len; */ 594 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4); 595 /* l %r5,<d(len)>(%r2) */ 596 EMIT4_DISP(0x58502000, offsetof(struct sk_buff, len)); 597 break; 598 case BPF_S_LDX_W_LEN: /* X = skb->len; */ 599 jit->seen |= SEEN_XREG; 600 /* l %r12,<d(len)>(%r2) */ 601 EMIT4_DISP(0x58c02000, offsetof(struct sk_buff, len)); 602 break; 603 case BPF_S_LD_IMM: /* A = K */ 604 if (K <= 16383) 605 /* lhi %r5,K */ 606 EMIT4_IMM(0xa7580000, K); 607 else if (test_facility(21)) 608 /* llilf %r5,<K> */ 609 EMIT6_IMM(0xc05f0000, K); 610 else 611 /* l %r5,<d(K)>(%r13) */ 612 EMIT4_DISP(0x5850d000, EMIT_CONST(K)); 613 break; 614 case BPF_S_LDX_IMM: /* X = K */ 615 jit->seen |= SEEN_XREG; 616 if (K <= 16383) 617 /* lhi %r12,<K> */ 618 EMIT4_IMM(0xa7c80000, K); 619 else if (test_facility(21)) 620 /* llilf %r12,<K> */ 621 EMIT6_IMM(0xc0cf0000, K); 622 else 623 /* l %r12,<d(K)>(%r13) */ 624 EMIT4_DISP(0x58c0d000, EMIT_CONST(K)); 625 break; 626 case BPF_S_LD_MEM: /* A = mem[K] */ 627 jit->seen |= SEEN_MEM; 628 /* l %r5,<K>(%r15) */ 629 EMIT4_DISP(0x5850f000, 630 (jit->seen & SEEN_DATAREF) ? 160 + K*4 : K*4); 631 break; 632 case BPF_S_LDX_MEM: /* X = mem[K] */ 633 jit->seen |= SEEN_XREG | SEEN_MEM; 634 /* l %r12,<K>(%r15) */ 635 EMIT4_DISP(0x58c0f000, 636 (jit->seen & SEEN_DATAREF) ? 160 + K*4 : K*4); 637 break; 638 case BPF_S_MISC_TAX: /* X = A */ 639 jit->seen |= SEEN_XREG; 640 /* lr %r12,%r5 */ 641 EMIT2(0x18c5); 642 break; 643 case BPF_S_MISC_TXA: /* A = X */ 644 jit->seen |= SEEN_XREG; 645 /* lr %r5,%r12 */ 646 EMIT2(0x185c); 647 break; 648 case BPF_S_RET_K: 649 if (K == 0) { 650 jit->seen |= SEEN_RET0; 651 if (last) 652 break; 653 /* j <ret0> */ 654 EMIT4_PCREL(0xa7f40000, jit->ret0_ip - jit->prg); 655 } else { 656 if (K <= 16383) 657 /* lghi %r2,K */ 658 EMIT4_IMM(0xa7290000, K); 659 else 660 /* llgf %r2,<K>(%r13) */ 661 EMIT6_DISP(0xe320d000, 0x0016, EMIT_CONST(K)); 662 /* j <exit> */ 663 if (last && !(jit->seen & SEEN_RET0)) 664 break; 665 EMIT4_PCREL(0xa7f40000, jit->exit_ip - jit->prg); 666 } 667 break; 668 case BPF_S_RET_A: 669 /* llgfr %r2,%r5 */ 670 EMIT4(0xb9160025); 671 /* j <exit> */ 672 EMIT4_PCREL(0xa7f40000, jit->exit_ip - jit->prg); 673 break; 674 case BPF_S_ST: /* mem[K] = A */ 675 jit->seen |= SEEN_MEM; 676 /* st %r5,<K>(%r15) */ 677 EMIT4_DISP(0x5050f000, 678 (jit->seen & SEEN_DATAREF) ? 160 + K*4 : K*4); 679 break; 680 case BPF_S_STX: /* mem[K] = X : mov %ebx,off8(%rbp) */ 681 jit->seen |= SEEN_XREG | SEEN_MEM; 682 /* st %r12,<K>(%r15) */ 683 EMIT4_DISP(0x50c0f000, 684 (jit->seen & SEEN_DATAREF) ? 160 + K*4 : K*4); 685 break; 686 case BPF_S_ANC_PROTOCOL: /* A = ntohs(skb->protocol); */ 687 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2); 688 /* lhi %r5,0 */ 689 EMIT4(0xa7580000); 690 /* icm %r5,3,<d(protocol)>(%r2) */ 691 EMIT4_DISP(0xbf532000, offsetof(struct sk_buff, protocol)); 692 break; 693 case BPF_S_ANC_IFINDEX: /* if (!skb->dev) return 0; 694 * A = skb->dev->ifindex */ 695 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4); 696 jit->seen |= SEEN_RET0; 697 /* lg %r1,<d(dev)>(%r2) */ 698 EMIT6_DISP(0xe3102000, 0x0004, offsetof(struct sk_buff, dev)); 699 /* ltgr %r1,%r1 */ 700 EMIT4(0xb9020011); 701 /* jz <ret0> */ 702 EMIT4_PCREL(0xa7840000, jit->ret0_ip - jit->prg); 703 /* l %r5,<d(ifindex)>(%r1) */ 704 EMIT4_DISP(0x58501000, offsetof(struct net_device, ifindex)); 705 break; 706 case BPF_S_ANC_MARK: /* A = skb->mark */ 707 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4); 708 /* l %r5,<d(mark)>(%r2) */ 709 EMIT4_DISP(0x58502000, offsetof(struct sk_buff, mark)); 710 break; 711 case BPF_S_ANC_QUEUE: /* A = skb->queue_mapping */ 712 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2); 713 /* lhi %r5,0 */ 714 EMIT4(0xa7580000); 715 /* icm %r5,3,<d(queue_mapping)>(%r2) */ 716 EMIT4_DISP(0xbf532000, offsetof(struct sk_buff, queue_mapping)); 717 break; 718 case BPF_S_ANC_HATYPE: /* if (!skb->dev) return 0; 719 * A = skb->dev->type */ 720 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, type) != 2); 721 jit->seen |= SEEN_RET0; 722 /* lg %r1,<d(dev)>(%r2) */ 723 EMIT6_DISP(0xe3102000, 0x0004, offsetof(struct sk_buff, dev)); 724 /* ltgr %r1,%r1 */ 725 EMIT4(0xb9020011); 726 /* jz <ret0> */ 727 EMIT4_PCREL(0xa7840000, jit->ret0_ip - jit->prg); 728 /* lhi %r5,0 */ 729 EMIT4(0xa7580000); 730 /* icm %r5,3,<d(type)>(%r1) */ 731 EMIT4_DISP(0xbf531000, offsetof(struct net_device, type)); 732 break; 733 case BPF_S_ANC_RXHASH: /* A = skb->rxhash */ 734 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, rxhash) != 4); 735 /* l %r5,<d(rxhash)>(%r2) */ 736 EMIT4_DISP(0x58502000, offsetof(struct sk_buff, rxhash)); 737 break; 738 case BPF_S_ANC_VLAN_TAG: 739 case BPF_S_ANC_VLAN_TAG_PRESENT: 740 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2); 741 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000); 742 /* lhi %r5,0 */ 743 EMIT4(0xa7580000); 744 /* icm %r5,3,<d(vlan_tci)>(%r2) */ 745 EMIT4_DISP(0xbf532000, offsetof(struct sk_buff, vlan_tci)); 746 if (filter->code == BPF_S_ANC_VLAN_TAG) { 747 /* nill %r5,0xefff */ 748 EMIT4_IMM(0xa5570000, ~VLAN_TAG_PRESENT); 749 } else { 750 /* nill %r5,0x1000 */ 751 EMIT4_IMM(0xa5570000, VLAN_TAG_PRESENT); 752 /* srl %r5,12 */ 753 EMIT4_DISP(0x88500000, 12); 754 } 755 break; 756 case BPF_S_ANC_PKTTYPE: 757 if (pkt_type_offset < 0) 758 goto out; 759 /* lhi %r5,0 */ 760 EMIT4(0xa7580000); 761 /* ic %r5,<d(pkt_type_offset)>(%r2) */ 762 EMIT4_DISP(0x43502000, pkt_type_offset); 763 /* srl %r5,5 */ 764 EMIT4_DISP(0x88500000, 5); 765 break; 766 case BPF_S_ANC_CPU: /* A = smp_processor_id() */ 767 #ifdef CONFIG_SMP 768 /* l %r5,<d(cpu_nr)> */ 769 EMIT4_DISP(0x58500000, offsetof(struct _lowcore, cpu_nr)); 770 #else 771 /* lhi %r5,0 */ 772 EMIT4(0xa7580000); 773 #endif 774 break; 775 default: /* too complex, give up */ 776 goto out; 777 } 778 addrs[i] = jit->prg - jit->start; 779 return 0; 780 out: 781 return -1; 782 } 783 784 /* 785 * Note: for security reasons, bpf code will follow a randomly 786 * sized amount of illegal instructions. 787 */ 788 struct bpf_binary_header { 789 unsigned int pages; 790 u8 image[]; 791 }; 792 793 static struct bpf_binary_header *bpf_alloc_binary(unsigned int bpfsize, 794 u8 **image_ptr) 795 { 796 struct bpf_binary_header *header; 797 unsigned int sz, hole; 798 799 /* Most BPF filters are really small, but if some of them fill a page, 800 * allow at least 128 extra bytes for illegal instructions. 801 */ 802 sz = round_up(bpfsize + sizeof(*header) + 128, PAGE_SIZE); 803 header = module_alloc(sz); 804 if (!header) 805 return NULL; 806 memset(header, 0, sz); 807 header->pages = sz / PAGE_SIZE; 808 hole = sz - bpfsize + sizeof(*header); 809 /* Insert random number of illegal instructions before BPF code 810 * and make sure the first instruction starts at an even address. 811 */ 812 *image_ptr = &header->image[(prandom_u32() % hole) & -2]; 813 return header; 814 } 815 816 void bpf_jit_compile(struct sk_filter *fp) 817 { 818 struct bpf_binary_header *header = NULL; 819 unsigned long size, prg_len, lit_len; 820 struct bpf_jit jit, cjit; 821 unsigned int *addrs; 822 int pass, i; 823 824 if (!bpf_jit_enable) 825 return; 826 addrs = kcalloc(fp->len, sizeof(*addrs), GFP_KERNEL); 827 if (addrs == NULL) 828 return; 829 memset(&jit, 0, sizeof(cjit)); 830 memset(&cjit, 0, sizeof(cjit)); 831 832 for (pass = 0; pass < 10; pass++) { 833 jit.prg = jit.start; 834 jit.lit = jit.mid; 835 836 bpf_jit_prologue(&jit); 837 bpf_jit_noleaks(&jit, fp->insns); 838 for (i = 0; i < fp->len; i++) { 839 if (bpf_jit_insn(&jit, fp->insns + i, addrs, i, 840 i == fp->len - 1)) 841 goto out; 842 } 843 bpf_jit_epilogue(&jit); 844 if (jit.start) { 845 WARN_ON(jit.prg > cjit.prg || jit.lit > cjit.lit); 846 if (memcmp(&jit, &cjit, sizeof(jit)) == 0) 847 break; 848 } else if (jit.prg == cjit.prg && jit.lit == cjit.lit) { 849 prg_len = jit.prg - jit.start; 850 lit_len = jit.lit - jit.mid; 851 size = prg_len + lit_len; 852 if (size >= BPF_SIZE_MAX) 853 goto out; 854 header = bpf_alloc_binary(size, &jit.start); 855 if (!header) 856 goto out; 857 jit.prg = jit.mid = jit.start + prg_len; 858 jit.lit = jit.end = jit.start + prg_len + lit_len; 859 jit.base_ip += (unsigned long) jit.start; 860 jit.exit_ip += (unsigned long) jit.start; 861 jit.ret0_ip += (unsigned long) jit.start; 862 } 863 cjit = jit; 864 } 865 if (bpf_jit_enable > 1) { 866 bpf_jit_dump(fp->len, jit.end - jit.start, pass, jit.start); 867 if (jit.start) 868 print_fn_code(jit.start, jit.mid - jit.start); 869 } 870 if (jit.start) { 871 set_memory_ro((unsigned long)header, header->pages); 872 fp->bpf_func = (void *) jit.start; 873 } 874 out: 875 kfree(addrs); 876 } 877 878 void bpf_jit_free(struct sk_filter *fp) 879 { 880 unsigned long addr = (unsigned long)fp->bpf_func & PAGE_MASK; 881 struct bpf_binary_header *header = (void *)addr; 882 883 if (fp->bpf_func == sk_run_filter) 884 return; 885 set_memory_rw(addr, header->pages); 886 module_free(NULL, header); 887 } 888