1 /* bpf_jit_comp.c: BPF JIT compiler for PPC64 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 * 7 * This program is free software; you can redistribute it and/or 8 * modify it under the terms of the GNU General Public License 9 * as published by the Free Software Foundation; version 2 10 * of the License. 11 */ 12 #include <linux/moduleloader.h> 13 #include <asm/cacheflush.h> 14 #include <linux/netdevice.h> 15 #include <linux/filter.h> 16 #include <linux/if_vlan.h> 17 18 #include "bpf_jit.h" 19 20 int bpf_jit_enable __read_mostly; 21 22 static inline void bpf_flush_icache(void *start, void *end) 23 { 24 smp_wmb(); 25 flush_icache_range((unsigned long)start, (unsigned long)end); 26 } 27 28 static void bpf_jit_build_prologue(struct sk_filter *fp, u32 *image, 29 struct codegen_context *ctx) 30 { 31 int i; 32 const struct sock_filter *filter = fp->insns; 33 34 if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) { 35 /* Make stackframe */ 36 if (ctx->seen & SEEN_DATAREF) { 37 /* If we call any helpers (for loads), save LR */ 38 EMIT(PPC_INST_MFLR | __PPC_RT(R0)); 39 PPC_STD(0, 1, 16); 40 41 /* Back up non-volatile regs. */ 42 PPC_STD(r_D, 1, -(8*(32-r_D))); 43 PPC_STD(r_HL, 1, -(8*(32-r_HL))); 44 } 45 if (ctx->seen & SEEN_MEM) { 46 /* 47 * Conditionally save regs r15-r31 as some will be used 48 * for M[] data. 49 */ 50 for (i = r_M; i < (r_M+16); i++) { 51 if (ctx->seen & (1 << (i-r_M))) 52 PPC_STD(i, 1, -(8*(32-i))); 53 } 54 } 55 EMIT(PPC_INST_STDU | __PPC_RS(R1) | __PPC_RA(R1) | 56 (-BPF_PPC_STACKFRAME & 0xfffc)); 57 } 58 59 if (ctx->seen & SEEN_DATAREF) { 60 /* 61 * If this filter needs to access skb data, 62 * prepare r_D and r_HL: 63 * r_HL = skb->len - skb->data_len 64 * r_D = skb->data 65 */ 66 PPC_LWZ_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff, 67 data_len)); 68 PPC_LWZ_OFFS(r_HL, r_skb, offsetof(struct sk_buff, len)); 69 PPC_SUB(r_HL, r_HL, r_scratch1); 70 PPC_LD_OFFS(r_D, r_skb, offsetof(struct sk_buff, data)); 71 } 72 73 if (ctx->seen & SEEN_XREG) { 74 /* 75 * TODO: Could also detect whether first instr. sets X and 76 * avoid this (as below, with A). 77 */ 78 PPC_LI(r_X, 0); 79 } 80 81 switch (filter[0].code) { 82 case BPF_RET | BPF_K: 83 case BPF_LD | BPF_W | BPF_LEN: 84 case BPF_LD | BPF_W | BPF_ABS: 85 case BPF_LD | BPF_H | BPF_ABS: 86 case BPF_LD | BPF_B | BPF_ABS: 87 /* first instruction sets A register (or is RET 'constant') */ 88 break; 89 default: 90 /* make sure we dont leak kernel information to user */ 91 PPC_LI(r_A, 0); 92 } 93 } 94 95 static void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx) 96 { 97 int i; 98 99 if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) { 100 PPC_ADDI(1, 1, BPF_PPC_STACKFRAME); 101 if (ctx->seen & SEEN_DATAREF) { 102 PPC_LD(0, 1, 16); 103 PPC_MTLR(0); 104 PPC_LD(r_D, 1, -(8*(32-r_D))); 105 PPC_LD(r_HL, 1, -(8*(32-r_HL))); 106 } 107 if (ctx->seen & SEEN_MEM) { 108 /* Restore any saved non-vol registers */ 109 for (i = r_M; i < (r_M+16); i++) { 110 if (ctx->seen & (1 << (i-r_M))) 111 PPC_LD(i, 1, -(8*(32-i))); 112 } 113 } 114 } 115 /* The RETs have left a return value in R3. */ 116 117 PPC_BLR(); 118 } 119 120 #define CHOOSE_LOAD_FUNC(K, func) \ 121 ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset) 122 123 /* Assemble the body code between the prologue & epilogue. */ 124 static int bpf_jit_build_body(struct sk_filter *fp, u32 *image, 125 struct codegen_context *ctx, 126 unsigned int *addrs) 127 { 128 const struct sock_filter *filter = fp->insns; 129 int flen = fp->len; 130 u8 *func; 131 unsigned int true_cond; 132 int i; 133 134 /* Start of epilogue code */ 135 unsigned int exit_addr = addrs[flen]; 136 137 for (i = 0; i < flen; i++) { 138 unsigned int K = filter[i].k; 139 u16 code = bpf_anc_helper(&filter[i]); 140 141 /* 142 * addrs[] maps a BPF bytecode address into a real offset from 143 * the start of the body code. 144 */ 145 addrs[i] = ctx->idx * 4; 146 147 switch (code) { 148 /*** ALU ops ***/ 149 case BPF_ALU | BPF_ADD | BPF_X: /* A += X; */ 150 ctx->seen |= SEEN_XREG; 151 PPC_ADD(r_A, r_A, r_X); 152 break; 153 case BPF_ALU | BPF_ADD | 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_SUB | BPF_X: /* A -= X; */ 161 ctx->seen |= SEEN_XREG; 162 PPC_SUB(r_A, r_A, r_X); 163 break; 164 case BPF_ALU | BPF_SUB | BPF_K: /* A -= K */ 165 if (!K) 166 break; 167 PPC_ADDI(r_A, r_A, IMM_L(-K)); 168 if (K >= 32768) 169 PPC_ADDIS(r_A, r_A, IMM_HA(-K)); 170 break; 171 case BPF_ALU | BPF_MUL | BPF_X: /* A *= X; */ 172 ctx->seen |= SEEN_XREG; 173 PPC_MUL(r_A, r_A, r_X); 174 break; 175 case BPF_ALU | BPF_MUL | BPF_K: /* A *= K */ 176 if (K < 32768) 177 PPC_MULI(r_A, r_A, K); 178 else { 179 PPC_LI32(r_scratch1, K); 180 PPC_MUL(r_A, r_A, r_scratch1); 181 } 182 break; 183 case BPF_ALU | BPF_MOD | BPF_X: /* A %= X; */ 184 ctx->seen |= SEEN_XREG; 185 PPC_CMPWI(r_X, 0); 186 if (ctx->pc_ret0 != -1) { 187 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]); 188 } else { 189 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12); 190 PPC_LI(r_ret, 0); 191 PPC_JMP(exit_addr); 192 } 193 PPC_DIVWU(r_scratch1, r_A, r_X); 194 PPC_MUL(r_scratch1, r_X, r_scratch1); 195 PPC_SUB(r_A, r_A, r_scratch1); 196 break; 197 case BPF_ALU | BPF_MOD | BPF_K: /* A %= K; */ 198 PPC_LI32(r_scratch2, K); 199 PPC_DIVWU(r_scratch1, r_A, r_scratch2); 200 PPC_MUL(r_scratch1, r_scratch2, r_scratch1); 201 PPC_SUB(r_A, r_A, r_scratch1); 202 break; 203 case BPF_ALU | BPF_DIV | BPF_X: /* A /= X; */ 204 ctx->seen |= SEEN_XREG; 205 PPC_CMPWI(r_X, 0); 206 if (ctx->pc_ret0 != -1) { 207 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]); 208 } else { 209 /* 210 * Exit, returning 0; first pass hits here 211 * (longer worst-case code size). 212 */ 213 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12); 214 PPC_LI(r_ret, 0); 215 PPC_JMP(exit_addr); 216 } 217 PPC_DIVWU(r_A, r_A, r_X); 218 break; 219 case BPF_ALU | BPF_DIV | BPF_K: /* A /= K */ 220 if (K == 1) 221 break; 222 PPC_LI32(r_scratch1, K); 223 PPC_DIVWU(r_A, r_A, r_scratch1); 224 break; 225 case BPF_ALU | BPF_AND | BPF_X: 226 ctx->seen |= SEEN_XREG; 227 PPC_AND(r_A, r_A, r_X); 228 break; 229 case BPF_ALU | BPF_AND | BPF_K: 230 if (!IMM_H(K)) 231 PPC_ANDI(r_A, r_A, K); 232 else { 233 PPC_LI32(r_scratch1, K); 234 PPC_AND(r_A, r_A, r_scratch1); 235 } 236 break; 237 case BPF_ALU | BPF_OR | BPF_X: 238 ctx->seen |= SEEN_XREG; 239 PPC_OR(r_A, r_A, r_X); 240 break; 241 case BPF_ALU | BPF_OR | BPF_K: 242 if (IMM_L(K)) 243 PPC_ORI(r_A, r_A, IMM_L(K)); 244 if (K >= 65536) 245 PPC_ORIS(r_A, r_A, IMM_H(K)); 246 break; 247 case BPF_ANC | SKF_AD_ALU_XOR_X: 248 case BPF_ALU | BPF_XOR | BPF_X: /* A ^= X */ 249 ctx->seen |= SEEN_XREG; 250 PPC_XOR(r_A, r_A, r_X); 251 break; 252 case BPF_ALU | BPF_XOR | BPF_K: /* A ^= K */ 253 if (IMM_L(K)) 254 PPC_XORI(r_A, r_A, IMM_L(K)); 255 if (K >= 65536) 256 PPC_XORIS(r_A, r_A, IMM_H(K)); 257 break; 258 case BPF_ALU | BPF_LSH | BPF_X: /* A <<= X; */ 259 ctx->seen |= SEEN_XREG; 260 PPC_SLW(r_A, r_A, r_X); 261 break; 262 case BPF_ALU | BPF_LSH | BPF_K: 263 if (K == 0) 264 break; 265 else 266 PPC_SLWI(r_A, r_A, K); 267 break; 268 case BPF_ALU | BPF_RSH | BPF_X: /* A >>= X; */ 269 ctx->seen |= SEEN_XREG; 270 PPC_SRW(r_A, r_A, r_X); 271 break; 272 case BPF_ALU | BPF_RSH | BPF_K: /* A >>= K; */ 273 if (K == 0) 274 break; 275 else 276 PPC_SRWI(r_A, r_A, K); 277 break; 278 case BPF_ALU | BPF_NEG: 279 PPC_NEG(r_A, r_A); 280 break; 281 case BPF_RET | BPF_K: 282 PPC_LI32(r_ret, K); 283 if (!K) { 284 if (ctx->pc_ret0 == -1) 285 ctx->pc_ret0 = i; 286 } 287 /* 288 * If this isn't the very last instruction, branch to 289 * the epilogue if we've stuff to clean up. Otherwise, 290 * if there's nothing to tidy, just return. If we /are/ 291 * the last instruction, we're about to fall through to 292 * the epilogue to return. 293 */ 294 if (i != flen - 1) { 295 /* 296 * Note: 'seen' is properly valid only on pass 297 * #2. Both parts of this conditional are the 298 * same instruction size though, meaning the 299 * first pass will still correctly determine the 300 * code size/addresses. 301 */ 302 if (ctx->seen) 303 PPC_JMP(exit_addr); 304 else 305 PPC_BLR(); 306 } 307 break; 308 case BPF_RET | BPF_A: 309 PPC_MR(r_ret, r_A); 310 if (i != flen - 1) { 311 if (ctx->seen) 312 PPC_JMP(exit_addr); 313 else 314 PPC_BLR(); 315 } 316 break; 317 case BPF_MISC | BPF_TAX: /* X = A */ 318 PPC_MR(r_X, r_A); 319 break; 320 case BPF_MISC | BPF_TXA: /* A = X */ 321 ctx->seen |= SEEN_XREG; 322 PPC_MR(r_A, r_X); 323 break; 324 325 /*** Constant loads/M[] access ***/ 326 case BPF_LD | BPF_IMM: /* A = K */ 327 PPC_LI32(r_A, K); 328 break; 329 case BPF_LDX | BPF_IMM: /* X = K */ 330 PPC_LI32(r_X, K); 331 break; 332 case BPF_LD | BPF_MEM: /* A = mem[K] */ 333 PPC_MR(r_A, r_M + (K & 0xf)); 334 ctx->seen |= SEEN_MEM | (1<<(K & 0xf)); 335 break; 336 case BPF_LDX | BPF_MEM: /* X = mem[K] */ 337 PPC_MR(r_X, r_M + (K & 0xf)); 338 ctx->seen |= SEEN_MEM | (1<<(K & 0xf)); 339 break; 340 case BPF_ST: /* mem[K] = A */ 341 PPC_MR(r_M + (K & 0xf), r_A); 342 ctx->seen |= SEEN_MEM | (1<<(K & 0xf)); 343 break; 344 case BPF_STX: /* mem[K] = X */ 345 PPC_MR(r_M + (K & 0xf), r_X); 346 ctx->seen |= SEEN_XREG | SEEN_MEM | (1<<(K & 0xf)); 347 break; 348 case BPF_LD | BPF_W | BPF_LEN: /* A = skb->len; */ 349 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4); 350 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, len)); 351 break; 352 case BPF_LDX | BPF_W | BPF_LEN: /* X = skb->len; */ 353 PPC_LWZ_OFFS(r_X, r_skb, offsetof(struct sk_buff, len)); 354 break; 355 356 /*** Ancillary info loads ***/ 357 case BPF_ANC | SKF_AD_PROTOCOL: /* A = ntohs(skb->protocol); */ 358 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, 359 protocol) != 2); 360 PPC_NTOHS_OFFS(r_A, r_skb, offsetof(struct sk_buff, 361 protocol)); 362 break; 363 case BPF_ANC | SKF_AD_IFINDEX: 364 PPC_LD_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff, 365 dev)); 366 PPC_CMPDI(r_scratch1, 0); 367 if (ctx->pc_ret0 != -1) { 368 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]); 369 } else { 370 /* Exit, returning 0; first pass hits here. */ 371 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12); 372 PPC_LI(r_ret, 0); 373 PPC_JMP(exit_addr); 374 } 375 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, 376 ifindex) != 4); 377 PPC_LWZ_OFFS(r_A, r_scratch1, 378 offsetof(struct net_device, ifindex)); 379 break; 380 case BPF_ANC | SKF_AD_MARK: 381 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4); 382 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, 383 mark)); 384 break; 385 case BPF_ANC | SKF_AD_RXHASH: 386 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4); 387 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, 388 hash)); 389 break; 390 case BPF_ANC | SKF_AD_VLAN_TAG: 391 case BPF_ANC | SKF_AD_VLAN_TAG_PRESENT: 392 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2); 393 PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, 394 vlan_tci)); 395 if (code == (BPF_ANC | SKF_AD_VLAN_TAG)) 396 PPC_ANDI(r_A, r_A, VLAN_VID_MASK); 397 else 398 PPC_ANDI(r_A, r_A, VLAN_TAG_PRESENT); 399 break; 400 case BPF_ANC | SKF_AD_QUEUE: 401 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, 402 queue_mapping) != 2); 403 PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, 404 queue_mapping)); 405 break; 406 case BPF_ANC | SKF_AD_CPU: 407 #ifdef CONFIG_SMP 408 /* 409 * PACA ptr is r13: 410 * raw_smp_processor_id() = local_paca->paca_index 411 */ 412 BUILD_BUG_ON(FIELD_SIZEOF(struct paca_struct, 413 paca_index) != 2); 414 PPC_LHZ_OFFS(r_A, 13, 415 offsetof(struct paca_struct, paca_index)); 416 #else 417 PPC_LI(r_A, 0); 418 #endif 419 break; 420 421 /*** Absolute loads from packet header/data ***/ 422 case BPF_LD | BPF_W | BPF_ABS: 423 func = CHOOSE_LOAD_FUNC(K, sk_load_word); 424 goto common_load; 425 case BPF_LD | BPF_H | BPF_ABS: 426 func = CHOOSE_LOAD_FUNC(K, sk_load_half); 427 goto common_load; 428 case BPF_LD | BPF_B | BPF_ABS: 429 func = CHOOSE_LOAD_FUNC(K, sk_load_byte); 430 common_load: 431 /* Load from [K]. */ 432 ctx->seen |= SEEN_DATAREF; 433 PPC_LI64(r_scratch1, func); 434 PPC_MTLR(r_scratch1); 435 PPC_LI32(r_addr, K); 436 PPC_BLRL(); 437 /* 438 * Helper returns 'lt' condition on error, and an 439 * appropriate return value in r3 440 */ 441 PPC_BCC(COND_LT, exit_addr); 442 break; 443 444 /*** Indirect loads from packet header/data ***/ 445 case BPF_LD | BPF_W | BPF_IND: 446 func = sk_load_word; 447 goto common_load_ind; 448 case BPF_LD | BPF_H | BPF_IND: 449 func = sk_load_half; 450 goto common_load_ind; 451 case BPF_LD | BPF_B | BPF_IND: 452 func = sk_load_byte; 453 common_load_ind: 454 /* 455 * Load from [X + K]. Negative offsets are tested for 456 * in the helper functions. 457 */ 458 ctx->seen |= SEEN_DATAREF | SEEN_XREG; 459 PPC_LI64(r_scratch1, func); 460 PPC_MTLR(r_scratch1); 461 PPC_ADDI(r_addr, r_X, IMM_L(K)); 462 if (K >= 32768) 463 PPC_ADDIS(r_addr, r_addr, IMM_HA(K)); 464 PPC_BLRL(); 465 /* If error, cr0.LT set */ 466 PPC_BCC(COND_LT, exit_addr); 467 break; 468 469 case BPF_LDX | BPF_B | BPF_MSH: 470 func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh); 471 goto common_load; 472 break; 473 474 /*** Jump and branches ***/ 475 case BPF_JMP | BPF_JA: 476 if (K != 0) 477 PPC_JMP(addrs[i + 1 + K]); 478 break; 479 480 case BPF_JMP | BPF_JGT | BPF_K: 481 case BPF_JMP | BPF_JGT | BPF_X: 482 true_cond = COND_GT; 483 goto cond_branch; 484 case BPF_JMP | BPF_JGE | BPF_K: 485 case BPF_JMP | BPF_JGE | BPF_X: 486 true_cond = COND_GE; 487 goto cond_branch; 488 case BPF_JMP | BPF_JEQ | BPF_K: 489 case BPF_JMP | BPF_JEQ | BPF_X: 490 true_cond = COND_EQ; 491 goto cond_branch; 492 case BPF_JMP | BPF_JSET | BPF_K: 493 case BPF_JMP | BPF_JSET | BPF_X: 494 true_cond = COND_NE; 495 /* Fall through */ 496 cond_branch: 497 /* same targets, can avoid doing the test :) */ 498 if (filter[i].jt == filter[i].jf) { 499 if (filter[i].jt > 0) 500 PPC_JMP(addrs[i + 1 + filter[i].jt]); 501 break; 502 } 503 504 switch (code) { 505 case BPF_JMP | BPF_JGT | BPF_X: 506 case BPF_JMP | BPF_JGE | BPF_X: 507 case BPF_JMP | BPF_JEQ | BPF_X: 508 ctx->seen |= SEEN_XREG; 509 PPC_CMPLW(r_A, r_X); 510 break; 511 case BPF_JMP | BPF_JSET | BPF_X: 512 ctx->seen |= SEEN_XREG; 513 PPC_AND_DOT(r_scratch1, r_A, r_X); 514 break; 515 case BPF_JMP | BPF_JEQ | BPF_K: 516 case BPF_JMP | BPF_JGT | BPF_K: 517 case BPF_JMP | BPF_JGE | BPF_K: 518 if (K < 32768) 519 PPC_CMPLWI(r_A, K); 520 else { 521 PPC_LI32(r_scratch1, K); 522 PPC_CMPLW(r_A, r_scratch1); 523 } 524 break; 525 case BPF_JMP | BPF_JSET | BPF_K: 526 if (K < 32768) 527 /* PPC_ANDI is /only/ dot-form */ 528 PPC_ANDI(r_scratch1, r_A, K); 529 else { 530 PPC_LI32(r_scratch1, K); 531 PPC_AND_DOT(r_scratch1, r_A, 532 r_scratch1); 533 } 534 break; 535 } 536 /* Sometimes branches are constructed "backward", with 537 * the false path being the branch and true path being 538 * a fallthrough to the next instruction. 539 */ 540 if (filter[i].jt == 0) 541 /* Swap the sense of the branch */ 542 PPC_BCC(true_cond ^ COND_CMP_TRUE, 543 addrs[i + 1 + filter[i].jf]); 544 else { 545 PPC_BCC(true_cond, addrs[i + 1 + filter[i].jt]); 546 if (filter[i].jf != 0) 547 PPC_JMP(addrs[i + 1 + filter[i].jf]); 548 } 549 break; 550 default: 551 /* The filter contains something cruel & unusual. 552 * We don't handle it, but also there shouldn't be 553 * anything missing from our list. 554 */ 555 if (printk_ratelimit()) 556 pr_err("BPF filter opcode %04x (@%d) unsupported\n", 557 filter[i].code, i); 558 return -ENOTSUPP; 559 } 560 561 } 562 /* Set end-of-body-code address for exit. */ 563 addrs[i] = ctx->idx * 4; 564 565 return 0; 566 } 567 568 void bpf_jit_compile(struct sk_filter *fp) 569 { 570 unsigned int proglen; 571 unsigned int alloclen; 572 u32 *image = NULL; 573 u32 *code_base; 574 unsigned int *addrs; 575 struct codegen_context cgctx; 576 int pass; 577 int flen = fp->len; 578 579 if (!bpf_jit_enable) 580 return; 581 582 addrs = kzalloc((flen+1) * sizeof(*addrs), GFP_KERNEL); 583 if (addrs == NULL) 584 return; 585 586 /* 587 * There are multiple assembly passes as the generated code will change 588 * size as it settles down, figuring out the max branch offsets/exit 589 * paths required. 590 * 591 * The range of standard conditional branches is +/- 32Kbytes. Since 592 * BPF_MAXINSNS = 4096, we can only jump from (worst case) start to 593 * finish with 8 bytes/instruction. Not feasible, so long jumps are 594 * used, distinct from short branches. 595 * 596 * Current: 597 * 598 * For now, both branch types assemble to 2 words (short branches padded 599 * with a NOP); this is less efficient, but assembly will always complete 600 * after exactly 3 passes: 601 * 602 * First pass: No code buffer; Program is "faux-generated" -- no code 603 * emitted but maximum size of output determined (and addrs[] filled 604 * in). Also, we note whether we use M[], whether we use skb data, etc. 605 * All generation choices assumed to be 'worst-case', e.g. branches all 606 * far (2 instructions), return path code reduction not available, etc. 607 * 608 * Second pass: Code buffer allocated with size determined previously. 609 * Prologue generated to support features we have seen used. Exit paths 610 * determined and addrs[] is filled in again, as code may be slightly 611 * smaller as a result. 612 * 613 * Third pass: Code generated 'for real', and branch destinations 614 * determined from now-accurate addrs[] map. 615 * 616 * Ideal: 617 * 618 * If we optimise this, near branches will be shorter. On the 619 * first assembly pass, we should err on the side of caution and 620 * generate the biggest code. On subsequent passes, branches will be 621 * generated short or long and code size will reduce. With smaller 622 * code, more branches may fall into the short category, and code will 623 * reduce more. 624 * 625 * Finally, if we see one pass generate code the same size as the 626 * previous pass we have converged and should now generate code for 627 * real. Allocating at the end will also save the memory that would 628 * otherwise be wasted by the (small) current code shrinkage. 629 * Preferably, we should do a small number of passes (e.g. 5) and if we 630 * haven't converged by then, get impatient and force code to generate 631 * as-is, even if the odd branch would be left long. The chances of a 632 * long jump are tiny with all but the most enormous of BPF filter 633 * inputs, so we should usually converge on the third pass. 634 */ 635 636 cgctx.idx = 0; 637 cgctx.seen = 0; 638 cgctx.pc_ret0 = -1; 639 /* Scouting faux-generate pass 0 */ 640 if (bpf_jit_build_body(fp, 0, &cgctx, addrs)) 641 /* We hit something illegal or unsupported. */ 642 goto out; 643 644 /* 645 * Pretend to build prologue, given the features we've seen. This will 646 * update ctgtx.idx as it pretends to output instructions, then we can 647 * calculate total size from idx. 648 */ 649 bpf_jit_build_prologue(fp, 0, &cgctx); 650 bpf_jit_build_epilogue(0, &cgctx); 651 652 proglen = cgctx.idx * 4; 653 alloclen = proglen + FUNCTION_DESCR_SIZE; 654 image = module_alloc(alloclen); 655 if (!image) 656 goto out; 657 658 code_base = image + (FUNCTION_DESCR_SIZE/4); 659 660 /* Code generation passes 1-2 */ 661 for (pass = 1; pass < 3; pass++) { 662 /* Now build the prologue, body code & epilogue for real. */ 663 cgctx.idx = 0; 664 bpf_jit_build_prologue(fp, code_base, &cgctx); 665 bpf_jit_build_body(fp, code_base, &cgctx, addrs); 666 bpf_jit_build_epilogue(code_base, &cgctx); 667 668 if (bpf_jit_enable > 1) 669 pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass, 670 proglen - (cgctx.idx * 4), cgctx.seen); 671 } 672 673 if (bpf_jit_enable > 1) 674 /* Note that we output the base address of the code_base 675 * rather than image, since opcodes are in code_base. 676 */ 677 bpf_jit_dump(flen, proglen, pass, code_base); 678 679 if (image) { 680 bpf_flush_icache(code_base, code_base + (proglen/4)); 681 /* Function descriptor nastiness: Address + TOC */ 682 ((u64 *)image)[0] = (u64)code_base; 683 ((u64 *)image)[1] = local_paca->kernel_toc; 684 fp->bpf_func = (void *)image; 685 fp->jited = 1; 686 } 687 out: 688 kfree(addrs); 689 return; 690 } 691 692 void bpf_jit_free(struct sk_filter *fp) 693 { 694 if (fp->jited) 695 module_free(NULL, fp->bpf_func); 696 kfree(fp); 697 } 698