1 /* bpf_jit_comp.c: BPF JIT compiler 2 * 3 * Copyright 2011 Matt Evans <matt@ozlabs.org>, IBM Corporation 4 * 5 * Based on the x86 BPF compiler, by Eric Dumazet (eric.dumazet@gmail.com) 6 * Ported to ppc32 by Denis Kirjanov <kda@linux-powerpc.org> 7 * 8 * This program is free software; you can redistribute it and/or 9 * modify it under the terms of the GNU General Public License 10 * as published by the Free Software Foundation; version 2 11 * of the License. 12 */ 13 #include <linux/moduleloader.h> 14 #include <asm/cacheflush.h> 15 #include <linux/netdevice.h> 16 #include <linux/filter.h> 17 #include <linux/if_vlan.h> 18 19 #include "bpf_jit32.h" 20 21 static inline void bpf_flush_icache(void *start, void *end) 22 { 23 smp_wmb(); 24 flush_icache_range((unsigned long)start, (unsigned long)end); 25 } 26 27 static void bpf_jit_build_prologue(struct bpf_prog *fp, u32 *image, 28 struct codegen_context *ctx) 29 { 30 int i; 31 const struct sock_filter *filter = fp->insns; 32 33 if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) { 34 /* Make stackframe */ 35 if (ctx->seen & SEEN_DATAREF) { 36 /* If we call any helpers (for loads), save LR */ 37 EMIT(PPC_INST_MFLR | __PPC_RT(R0)); 38 PPC_BPF_STL(0, 1, PPC_LR_STKOFF); 39 40 /* Back up non-volatile regs. */ 41 PPC_BPF_STL(r_D, 1, -(REG_SZ*(32-r_D))); 42 PPC_BPF_STL(r_HL, 1, -(REG_SZ*(32-r_HL))); 43 } 44 if (ctx->seen & SEEN_MEM) { 45 /* 46 * Conditionally save regs r15-r31 as some will be used 47 * for M[] data. 48 */ 49 for (i = r_M; i < (r_M+16); i++) { 50 if (ctx->seen & (1 << (i-r_M))) 51 PPC_BPF_STL(i, 1, -(REG_SZ*(32-i))); 52 } 53 } 54 PPC_BPF_STLU(1, 1, -BPF_PPC_STACKFRAME); 55 } 56 57 if (ctx->seen & SEEN_DATAREF) { 58 /* 59 * If this filter needs to access skb data, 60 * prepare r_D and r_HL: 61 * r_HL = skb->len - skb->data_len 62 * r_D = skb->data 63 */ 64 PPC_LWZ_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff, 65 data_len)); 66 PPC_LWZ_OFFS(r_HL, r_skb, offsetof(struct sk_buff, len)); 67 PPC_SUB(r_HL, r_HL, r_scratch1); 68 PPC_LL_OFFS(r_D, r_skb, offsetof(struct sk_buff, data)); 69 } 70 71 if (ctx->seen & SEEN_XREG) { 72 /* 73 * TODO: Could also detect whether first instr. sets X and 74 * avoid this (as below, with A). 75 */ 76 PPC_LI(r_X, 0); 77 } 78 79 /* make sure we dont leak kernel information to user */ 80 if (bpf_needs_clear_a(&filter[0])) 81 PPC_LI(r_A, 0); 82 } 83 84 static void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx) 85 { 86 int i; 87 88 if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) { 89 PPC_ADDI(1, 1, BPF_PPC_STACKFRAME); 90 if (ctx->seen & SEEN_DATAREF) { 91 PPC_BPF_LL(0, 1, PPC_LR_STKOFF); 92 PPC_MTLR(0); 93 PPC_BPF_LL(r_D, 1, -(REG_SZ*(32-r_D))); 94 PPC_BPF_LL(r_HL, 1, -(REG_SZ*(32-r_HL))); 95 } 96 if (ctx->seen & SEEN_MEM) { 97 /* Restore any saved non-vol registers */ 98 for (i = r_M; i < (r_M+16); i++) { 99 if (ctx->seen & (1 << (i-r_M))) 100 PPC_BPF_LL(i, 1, -(REG_SZ*(32-i))); 101 } 102 } 103 } 104 /* The RETs have left a return value in R3. */ 105 106 PPC_BLR(); 107 } 108 109 #define CHOOSE_LOAD_FUNC(K, func) \ 110 ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset) 111 112 /* Assemble the body code between the prologue & epilogue. */ 113 static int bpf_jit_build_body(struct bpf_prog *fp, u32 *image, 114 struct codegen_context *ctx, 115 unsigned int *addrs) 116 { 117 const struct sock_filter *filter = fp->insns; 118 int flen = fp->len; 119 u8 *func; 120 unsigned int true_cond; 121 int i; 122 123 /* Start of epilogue code */ 124 unsigned int exit_addr = addrs[flen]; 125 126 for (i = 0; i < flen; i++) { 127 unsigned int K = filter[i].k; 128 u16 code = bpf_anc_helper(&filter[i]); 129 130 /* 131 * addrs[] maps a BPF bytecode address into a real offset from 132 * the start of the body code. 133 */ 134 addrs[i] = ctx->idx * 4; 135 136 switch (code) { 137 /*** ALU ops ***/ 138 case BPF_ALU | BPF_ADD | BPF_X: /* A += X; */ 139 ctx->seen |= SEEN_XREG; 140 PPC_ADD(r_A, r_A, r_X); 141 break; 142 case BPF_ALU | BPF_ADD | BPF_K: /* A += K; */ 143 if (!K) 144 break; 145 PPC_ADDI(r_A, r_A, IMM_L(K)); 146 if (K >= 32768) 147 PPC_ADDIS(r_A, r_A, IMM_HA(K)); 148 break; 149 case BPF_ALU | BPF_SUB | BPF_X: /* A -= X; */ 150 ctx->seen |= SEEN_XREG; 151 PPC_SUB(r_A, r_A, r_X); 152 break; 153 case BPF_ALU | BPF_SUB | BPF_K: /* A -= K */ 154 if (!K) 155 break; 156 PPC_ADDI(r_A, r_A, IMM_L(-K)); 157 if (K >= 32768) 158 PPC_ADDIS(r_A, r_A, IMM_HA(-K)); 159 break; 160 case BPF_ALU | BPF_MUL | BPF_X: /* A *= X; */ 161 ctx->seen |= SEEN_XREG; 162 PPC_MULW(r_A, r_A, r_X); 163 break; 164 case BPF_ALU | BPF_MUL | BPF_K: /* A *= K */ 165 if (K < 32768) 166 PPC_MULI(r_A, r_A, K); 167 else { 168 PPC_LI32(r_scratch1, K); 169 PPC_MULW(r_A, r_A, r_scratch1); 170 } 171 break; 172 case BPF_ALU | BPF_MOD | BPF_X: /* A %= X; */ 173 case BPF_ALU | BPF_DIV | BPF_X: /* A /= X; */ 174 ctx->seen |= SEEN_XREG; 175 PPC_CMPWI(r_X, 0); 176 if (ctx->pc_ret0 != -1) { 177 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]); 178 } else { 179 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12); 180 PPC_LI(r_ret, 0); 181 PPC_JMP(exit_addr); 182 } 183 if (code == (BPF_ALU | BPF_MOD | BPF_X)) { 184 PPC_DIVWU(r_scratch1, r_A, r_X); 185 PPC_MULW(r_scratch1, r_X, r_scratch1); 186 PPC_SUB(r_A, r_A, r_scratch1); 187 } else { 188 PPC_DIVWU(r_A, r_A, r_X); 189 } 190 break; 191 case BPF_ALU | BPF_MOD | BPF_K: /* A %= K; */ 192 PPC_LI32(r_scratch2, K); 193 PPC_DIVWU(r_scratch1, r_A, r_scratch2); 194 PPC_MULW(r_scratch1, r_scratch2, r_scratch1); 195 PPC_SUB(r_A, r_A, r_scratch1); 196 break; 197 case BPF_ALU | BPF_DIV | BPF_K: /* A /= K */ 198 if (K == 1) 199 break; 200 PPC_LI32(r_scratch1, K); 201 PPC_DIVWU(r_A, r_A, r_scratch1); 202 break; 203 case BPF_ALU | BPF_AND | BPF_X: 204 ctx->seen |= SEEN_XREG; 205 PPC_AND(r_A, r_A, r_X); 206 break; 207 case BPF_ALU | BPF_AND | BPF_K: 208 if (!IMM_H(K)) 209 PPC_ANDI(r_A, r_A, K); 210 else { 211 PPC_LI32(r_scratch1, K); 212 PPC_AND(r_A, r_A, r_scratch1); 213 } 214 break; 215 case BPF_ALU | BPF_OR | BPF_X: 216 ctx->seen |= SEEN_XREG; 217 PPC_OR(r_A, r_A, r_X); 218 break; 219 case BPF_ALU | BPF_OR | BPF_K: 220 if (IMM_L(K)) 221 PPC_ORI(r_A, r_A, IMM_L(K)); 222 if (K >= 65536) 223 PPC_ORIS(r_A, r_A, IMM_H(K)); 224 break; 225 case BPF_ANC | SKF_AD_ALU_XOR_X: 226 case BPF_ALU | BPF_XOR | BPF_X: /* A ^= X */ 227 ctx->seen |= SEEN_XREG; 228 PPC_XOR(r_A, r_A, r_X); 229 break; 230 case BPF_ALU | BPF_XOR | BPF_K: /* A ^= K */ 231 if (IMM_L(K)) 232 PPC_XORI(r_A, r_A, IMM_L(K)); 233 if (K >= 65536) 234 PPC_XORIS(r_A, r_A, IMM_H(K)); 235 break; 236 case BPF_ALU | BPF_LSH | BPF_X: /* A <<= X; */ 237 ctx->seen |= SEEN_XREG; 238 PPC_SLW(r_A, r_A, r_X); 239 break; 240 case BPF_ALU | BPF_LSH | BPF_K: 241 if (K == 0) 242 break; 243 else 244 PPC_SLWI(r_A, r_A, K); 245 break; 246 case BPF_ALU | BPF_RSH | BPF_X: /* A >>= X; */ 247 ctx->seen |= SEEN_XREG; 248 PPC_SRW(r_A, r_A, r_X); 249 break; 250 case BPF_ALU | BPF_RSH | BPF_K: /* A >>= K; */ 251 if (K == 0) 252 break; 253 else 254 PPC_SRWI(r_A, r_A, K); 255 break; 256 case BPF_ALU | BPF_NEG: 257 PPC_NEG(r_A, r_A); 258 break; 259 case BPF_RET | BPF_K: 260 PPC_LI32(r_ret, K); 261 if (!K) { 262 if (ctx->pc_ret0 == -1) 263 ctx->pc_ret0 = i; 264 } 265 /* 266 * If this isn't the very last instruction, branch to 267 * the epilogue if we've stuff to clean up. Otherwise, 268 * if there's nothing to tidy, just return. If we /are/ 269 * the last instruction, we're about to fall through to 270 * the epilogue to return. 271 */ 272 if (i != flen - 1) { 273 /* 274 * Note: 'seen' is properly valid only on pass 275 * #2. Both parts of this conditional are the 276 * same instruction size though, meaning the 277 * first pass will still correctly determine the 278 * code size/addresses. 279 */ 280 if (ctx->seen) 281 PPC_JMP(exit_addr); 282 else 283 PPC_BLR(); 284 } 285 break; 286 case BPF_RET | BPF_A: 287 PPC_MR(r_ret, r_A); 288 if (i != flen - 1) { 289 if (ctx->seen) 290 PPC_JMP(exit_addr); 291 else 292 PPC_BLR(); 293 } 294 break; 295 case BPF_MISC | BPF_TAX: /* X = A */ 296 PPC_MR(r_X, r_A); 297 break; 298 case BPF_MISC | BPF_TXA: /* A = X */ 299 ctx->seen |= SEEN_XREG; 300 PPC_MR(r_A, r_X); 301 break; 302 303 /*** Constant loads/M[] access ***/ 304 case BPF_LD | BPF_IMM: /* A = K */ 305 PPC_LI32(r_A, K); 306 break; 307 case BPF_LDX | BPF_IMM: /* X = K */ 308 PPC_LI32(r_X, K); 309 break; 310 case BPF_LD | BPF_MEM: /* A = mem[K] */ 311 PPC_MR(r_A, r_M + (K & 0xf)); 312 ctx->seen |= SEEN_MEM | (1<<(K & 0xf)); 313 break; 314 case BPF_LDX | BPF_MEM: /* X = mem[K] */ 315 PPC_MR(r_X, r_M + (K & 0xf)); 316 ctx->seen |= SEEN_MEM | (1<<(K & 0xf)); 317 break; 318 case BPF_ST: /* mem[K] = A */ 319 PPC_MR(r_M + (K & 0xf), r_A); 320 ctx->seen |= SEEN_MEM | (1<<(K & 0xf)); 321 break; 322 case BPF_STX: /* mem[K] = X */ 323 PPC_MR(r_M + (K & 0xf), r_X); 324 ctx->seen |= SEEN_XREG | SEEN_MEM | (1<<(K & 0xf)); 325 break; 326 case BPF_LD | BPF_W | BPF_LEN: /* A = skb->len; */ 327 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4); 328 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, len)); 329 break; 330 case BPF_LDX | BPF_W | BPF_LEN: /* X = skb->len; */ 331 PPC_LWZ_OFFS(r_X, r_skb, offsetof(struct sk_buff, len)); 332 break; 333 334 /*** Ancillary info loads ***/ 335 case BPF_ANC | SKF_AD_PROTOCOL: /* A = ntohs(skb->protocol); */ 336 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, 337 protocol) != 2); 338 PPC_NTOHS_OFFS(r_A, r_skb, offsetof(struct sk_buff, 339 protocol)); 340 break; 341 case BPF_ANC | SKF_AD_IFINDEX: 342 case BPF_ANC | SKF_AD_HATYPE: 343 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, 344 ifindex) != 4); 345 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, 346 type) != 2); 347 PPC_LL_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff, 348 dev)); 349 PPC_CMPDI(r_scratch1, 0); 350 if (ctx->pc_ret0 != -1) { 351 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]); 352 } else { 353 /* Exit, returning 0; first pass hits here. */ 354 PPC_BCC_SHORT(COND_NE, ctx->idx * 4 + 12); 355 PPC_LI(r_ret, 0); 356 PPC_JMP(exit_addr); 357 } 358 if (code == (BPF_ANC | SKF_AD_IFINDEX)) { 359 PPC_LWZ_OFFS(r_A, r_scratch1, 360 offsetof(struct net_device, ifindex)); 361 } else { 362 PPC_LHZ_OFFS(r_A, r_scratch1, 363 offsetof(struct net_device, type)); 364 } 365 366 break; 367 case BPF_ANC | SKF_AD_MARK: 368 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4); 369 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, 370 mark)); 371 break; 372 case BPF_ANC | SKF_AD_RXHASH: 373 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4); 374 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, 375 hash)); 376 break; 377 case BPF_ANC | SKF_AD_VLAN_TAG: 378 case BPF_ANC | SKF_AD_VLAN_TAG_PRESENT: 379 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2); 380 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000); 381 382 PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, 383 vlan_tci)); 384 if (code == (BPF_ANC | SKF_AD_VLAN_TAG)) { 385 PPC_ANDI(r_A, r_A, ~VLAN_TAG_PRESENT); 386 } else { 387 PPC_ANDI(r_A, r_A, VLAN_TAG_PRESENT); 388 PPC_SRWI(r_A, r_A, 12); 389 } 390 break; 391 case BPF_ANC | SKF_AD_QUEUE: 392 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, 393 queue_mapping) != 2); 394 PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, 395 queue_mapping)); 396 break; 397 case BPF_ANC | SKF_AD_PKTTYPE: 398 PPC_LBZ_OFFS(r_A, r_skb, PKT_TYPE_OFFSET()); 399 PPC_ANDI(r_A, r_A, PKT_TYPE_MAX); 400 PPC_SRWI(r_A, r_A, 5); 401 break; 402 case BPF_ANC | SKF_AD_CPU: 403 PPC_BPF_LOAD_CPU(r_A); 404 break; 405 /*** Absolute loads from packet header/data ***/ 406 case BPF_LD | BPF_W | BPF_ABS: 407 func = CHOOSE_LOAD_FUNC(K, sk_load_word); 408 goto common_load; 409 case BPF_LD | BPF_H | BPF_ABS: 410 func = CHOOSE_LOAD_FUNC(K, sk_load_half); 411 goto common_load; 412 case BPF_LD | BPF_B | BPF_ABS: 413 func = CHOOSE_LOAD_FUNC(K, sk_load_byte); 414 common_load: 415 /* Load from [K]. */ 416 ctx->seen |= SEEN_DATAREF; 417 PPC_FUNC_ADDR(r_scratch1, func); 418 PPC_MTLR(r_scratch1); 419 PPC_LI32(r_addr, K); 420 PPC_BLRL(); 421 /* 422 * Helper returns 'lt' condition on error, and an 423 * appropriate return value in r3 424 */ 425 PPC_BCC(COND_LT, exit_addr); 426 break; 427 428 /*** Indirect loads from packet header/data ***/ 429 case BPF_LD | BPF_W | BPF_IND: 430 func = sk_load_word; 431 goto common_load_ind; 432 case BPF_LD | BPF_H | BPF_IND: 433 func = sk_load_half; 434 goto common_load_ind; 435 case BPF_LD | BPF_B | BPF_IND: 436 func = sk_load_byte; 437 common_load_ind: 438 /* 439 * Load from [X + K]. Negative offsets are tested for 440 * in the helper functions. 441 */ 442 ctx->seen |= SEEN_DATAREF | SEEN_XREG; 443 PPC_FUNC_ADDR(r_scratch1, func); 444 PPC_MTLR(r_scratch1); 445 PPC_ADDI(r_addr, r_X, IMM_L(K)); 446 if (K >= 32768) 447 PPC_ADDIS(r_addr, r_addr, IMM_HA(K)); 448 PPC_BLRL(); 449 /* If error, cr0.LT set */ 450 PPC_BCC(COND_LT, exit_addr); 451 break; 452 453 case BPF_LDX | BPF_B | BPF_MSH: 454 func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh); 455 goto common_load; 456 break; 457 458 /*** Jump and branches ***/ 459 case BPF_JMP | BPF_JA: 460 if (K != 0) 461 PPC_JMP(addrs[i + 1 + K]); 462 break; 463 464 case BPF_JMP | BPF_JGT | BPF_K: 465 case BPF_JMP | BPF_JGT | BPF_X: 466 true_cond = COND_GT; 467 goto cond_branch; 468 case BPF_JMP | BPF_JGE | BPF_K: 469 case BPF_JMP | BPF_JGE | BPF_X: 470 true_cond = COND_GE; 471 goto cond_branch; 472 case BPF_JMP | BPF_JEQ | BPF_K: 473 case BPF_JMP | BPF_JEQ | BPF_X: 474 true_cond = COND_EQ; 475 goto cond_branch; 476 case BPF_JMP | BPF_JSET | BPF_K: 477 case BPF_JMP | BPF_JSET | BPF_X: 478 true_cond = COND_NE; 479 /* Fall through */ 480 cond_branch: 481 /* same targets, can avoid doing the test :) */ 482 if (filter[i].jt == filter[i].jf) { 483 if (filter[i].jt > 0) 484 PPC_JMP(addrs[i + 1 + filter[i].jt]); 485 break; 486 } 487 488 switch (code) { 489 case BPF_JMP | BPF_JGT | BPF_X: 490 case BPF_JMP | BPF_JGE | BPF_X: 491 case BPF_JMP | BPF_JEQ | BPF_X: 492 ctx->seen |= SEEN_XREG; 493 PPC_CMPLW(r_A, r_X); 494 break; 495 case BPF_JMP | BPF_JSET | BPF_X: 496 ctx->seen |= SEEN_XREG; 497 PPC_AND_DOT(r_scratch1, r_A, r_X); 498 break; 499 case BPF_JMP | BPF_JEQ | BPF_K: 500 case BPF_JMP | BPF_JGT | BPF_K: 501 case BPF_JMP | BPF_JGE | BPF_K: 502 if (K < 32768) 503 PPC_CMPLWI(r_A, K); 504 else { 505 PPC_LI32(r_scratch1, K); 506 PPC_CMPLW(r_A, r_scratch1); 507 } 508 break; 509 case BPF_JMP | BPF_JSET | BPF_K: 510 if (K < 32768) 511 /* PPC_ANDI is /only/ dot-form */ 512 PPC_ANDI(r_scratch1, r_A, K); 513 else { 514 PPC_LI32(r_scratch1, K); 515 PPC_AND_DOT(r_scratch1, r_A, 516 r_scratch1); 517 } 518 break; 519 } 520 /* Sometimes branches are constructed "backward", with 521 * the false path being the branch and true path being 522 * a fallthrough to the next instruction. 523 */ 524 if (filter[i].jt == 0) 525 /* Swap the sense of the branch */ 526 PPC_BCC(true_cond ^ COND_CMP_TRUE, 527 addrs[i + 1 + filter[i].jf]); 528 else { 529 PPC_BCC(true_cond, addrs[i + 1 + filter[i].jt]); 530 if (filter[i].jf != 0) 531 PPC_JMP(addrs[i + 1 + filter[i].jf]); 532 } 533 break; 534 default: 535 /* The filter contains something cruel & unusual. 536 * We don't handle it, but also there shouldn't be 537 * anything missing from our list. 538 */ 539 if (printk_ratelimit()) 540 pr_err("BPF filter opcode %04x (@%d) unsupported\n", 541 filter[i].code, i); 542 return -ENOTSUPP; 543 } 544 545 } 546 /* Set end-of-body-code address for exit. */ 547 addrs[i] = ctx->idx * 4; 548 549 return 0; 550 } 551 552 void bpf_jit_compile(struct bpf_prog *fp) 553 { 554 unsigned int proglen; 555 unsigned int alloclen; 556 u32 *image = NULL; 557 u32 *code_base; 558 unsigned int *addrs; 559 struct codegen_context cgctx; 560 int pass; 561 int flen = fp->len; 562 563 if (!bpf_jit_enable) 564 return; 565 566 addrs = kzalloc((flen+1) * sizeof(*addrs), GFP_KERNEL); 567 if (addrs == NULL) 568 return; 569 570 /* 571 * There are multiple assembly passes as the generated code will change 572 * size as it settles down, figuring out the max branch offsets/exit 573 * paths required. 574 * 575 * The range of standard conditional branches is +/- 32Kbytes. Since 576 * BPF_MAXINSNS = 4096, we can only jump from (worst case) start to 577 * finish with 8 bytes/instruction. Not feasible, so long jumps are 578 * used, distinct from short branches. 579 * 580 * Current: 581 * 582 * For now, both branch types assemble to 2 words (short branches padded 583 * with a NOP); this is less efficient, but assembly will always complete 584 * after exactly 3 passes: 585 * 586 * First pass: No code buffer; Program is "faux-generated" -- no code 587 * emitted but maximum size of output determined (and addrs[] filled 588 * in). Also, we note whether we use M[], whether we use skb data, etc. 589 * All generation choices assumed to be 'worst-case', e.g. branches all 590 * far (2 instructions), return path code reduction not available, etc. 591 * 592 * Second pass: Code buffer allocated with size determined previously. 593 * Prologue generated to support features we have seen used. Exit paths 594 * determined and addrs[] is filled in again, as code may be slightly 595 * smaller as a result. 596 * 597 * Third pass: Code generated 'for real', and branch destinations 598 * determined from now-accurate addrs[] map. 599 * 600 * Ideal: 601 * 602 * If we optimise this, near branches will be shorter. On the 603 * first assembly pass, we should err on the side of caution and 604 * generate the biggest code. On subsequent passes, branches will be 605 * generated short or long and code size will reduce. With smaller 606 * code, more branches may fall into the short category, and code will 607 * reduce more. 608 * 609 * Finally, if we see one pass generate code the same size as the 610 * previous pass we have converged and should now generate code for 611 * real. Allocating at the end will also save the memory that would 612 * otherwise be wasted by the (small) current code shrinkage. 613 * Preferably, we should do a small number of passes (e.g. 5) and if we 614 * haven't converged by then, get impatient and force code to generate 615 * as-is, even if the odd branch would be left long. The chances of a 616 * long jump are tiny with all but the most enormous of BPF filter 617 * inputs, so we should usually converge on the third pass. 618 */ 619 620 cgctx.idx = 0; 621 cgctx.seen = 0; 622 cgctx.pc_ret0 = -1; 623 /* Scouting faux-generate pass 0 */ 624 if (bpf_jit_build_body(fp, 0, &cgctx, addrs)) 625 /* We hit something illegal or unsupported. */ 626 goto out; 627 628 /* 629 * Pretend to build prologue, given the features we've seen. This will 630 * update ctgtx.idx as it pretends to output instructions, then we can 631 * calculate total size from idx. 632 */ 633 bpf_jit_build_prologue(fp, 0, &cgctx); 634 bpf_jit_build_epilogue(0, &cgctx); 635 636 proglen = cgctx.idx * 4; 637 alloclen = proglen + FUNCTION_DESCR_SIZE; 638 image = module_alloc(alloclen); 639 if (!image) 640 goto out; 641 642 code_base = image + (FUNCTION_DESCR_SIZE/4); 643 644 /* Code generation passes 1-2 */ 645 for (pass = 1; pass < 3; pass++) { 646 /* Now build the prologue, body code & epilogue for real. */ 647 cgctx.idx = 0; 648 bpf_jit_build_prologue(fp, code_base, &cgctx); 649 bpf_jit_build_body(fp, code_base, &cgctx, addrs); 650 bpf_jit_build_epilogue(code_base, &cgctx); 651 652 if (bpf_jit_enable > 1) 653 pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass, 654 proglen - (cgctx.idx * 4), cgctx.seen); 655 } 656 657 if (bpf_jit_enable > 1) 658 /* Note that we output the base address of the code_base 659 * rather than image, since opcodes are in code_base. 660 */ 661 bpf_jit_dump(flen, proglen, pass, code_base); 662 663 bpf_flush_icache(code_base, code_base + (proglen/4)); 664 665 #ifdef CONFIG_PPC64 666 /* Function descriptor nastiness: Address + TOC */ 667 ((u64 *)image)[0] = (u64)code_base; 668 ((u64 *)image)[1] = local_paca->kernel_toc; 669 #endif 670 671 fp->bpf_func = (void *)image; 672 fp->jited = 1; 673 674 out: 675 kfree(addrs); 676 return; 677 } 678 679 void bpf_jit_free(struct bpf_prog *fp) 680 { 681 if (fp->jited) 682 module_memfree(fp->bpf_func); 683 684 bpf_prog_unlock_free(fp); 685 } 686