1 /* 2 * Copyright (C) 2016 Netronome Systems, Inc. 3 * 4 * This software is dual licensed under the GNU General License Version 2, 5 * June 1991 as shown in the file COPYING in the top-level directory of this 6 * source tree or the BSD 2-Clause License provided below. You have the 7 * option to license this software under the complete terms of either license. 8 * 9 * The BSD 2-Clause License: 10 * 11 * Redistribution and use in source and binary forms, with or 12 * without modification, are permitted provided that the following 13 * conditions are met: 14 * 15 * 1. Redistributions of source code must retain the above 16 * copyright notice, this list of conditions and the following 17 * disclaimer. 18 * 19 * 2. Redistributions in binary form must reproduce the above 20 * copyright notice, this list of conditions and the following 21 * disclaimer in the documentation and/or other materials 22 * provided with the distribution. 23 * 24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 31 * SOFTWARE. 32 */ 33 34 #define pr_fmt(fmt) "NFP net bpf: " fmt 35 36 #include <linux/kernel.h> 37 #include <linux/bpf.h> 38 #include <linux/filter.h> 39 #include <linux/pkt_cls.h> 40 #include <linux/unistd.h> 41 42 #include "main.h" 43 #include "../nfp_asm.h" 44 45 /* --- NFP prog --- */ 46 /* Foreach "multiple" entries macros provide pos and next<n> pointers. 47 * It's safe to modify the next pointers (but not pos). 48 */ 49 #define nfp_for_each_insn_walk2(nfp_prog, pos, next) \ 50 for (pos = list_first_entry(&(nfp_prog)->insns, typeof(*pos), l), \ 51 next = list_next_entry(pos, l); \ 52 &(nfp_prog)->insns != &pos->l && \ 53 &(nfp_prog)->insns != &next->l; \ 54 pos = nfp_meta_next(pos), \ 55 next = nfp_meta_next(pos)) 56 57 #define nfp_for_each_insn_walk3(nfp_prog, pos, next, next2) \ 58 for (pos = list_first_entry(&(nfp_prog)->insns, typeof(*pos), l), \ 59 next = list_next_entry(pos, l), \ 60 next2 = list_next_entry(next, l); \ 61 &(nfp_prog)->insns != &pos->l && \ 62 &(nfp_prog)->insns != &next->l && \ 63 &(nfp_prog)->insns != &next2->l; \ 64 pos = nfp_meta_next(pos), \ 65 next = nfp_meta_next(pos), \ 66 next2 = nfp_meta_next(next)) 67 68 static bool 69 nfp_meta_has_next(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 70 { 71 return meta->l.next != &nfp_prog->insns; 72 } 73 74 static bool 75 nfp_meta_has_prev(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 76 { 77 return meta->l.prev != &nfp_prog->insns; 78 } 79 80 static void nfp_prog_free(struct nfp_prog *nfp_prog) 81 { 82 struct nfp_insn_meta *meta, *tmp; 83 84 list_for_each_entry_safe(meta, tmp, &nfp_prog->insns, l) { 85 list_del(&meta->l); 86 kfree(meta); 87 } 88 kfree(nfp_prog); 89 } 90 91 static void nfp_prog_push(struct nfp_prog *nfp_prog, u64 insn) 92 { 93 if (nfp_prog->__prog_alloc_len == nfp_prog->prog_len) { 94 nfp_prog->error = -ENOSPC; 95 return; 96 } 97 98 nfp_prog->prog[nfp_prog->prog_len] = insn; 99 nfp_prog->prog_len++; 100 } 101 102 static unsigned int nfp_prog_current_offset(struct nfp_prog *nfp_prog) 103 { 104 return nfp_prog->start_off + nfp_prog->prog_len; 105 } 106 107 static unsigned int 108 nfp_prog_offset_to_index(struct nfp_prog *nfp_prog, unsigned int offset) 109 { 110 return offset - nfp_prog->start_off; 111 } 112 113 /* --- Emitters --- */ 114 static void 115 __emit_cmd(struct nfp_prog *nfp_prog, enum cmd_tgt_map op, 116 u8 mode, u8 xfer, u8 areg, u8 breg, u8 size, bool sync) 117 { 118 enum cmd_ctx_swap ctx; 119 u64 insn; 120 121 if (sync) 122 ctx = CMD_CTX_SWAP; 123 else 124 ctx = CMD_CTX_NO_SWAP; 125 126 insn = FIELD_PREP(OP_CMD_A_SRC, areg) | 127 FIELD_PREP(OP_CMD_CTX, ctx) | 128 FIELD_PREP(OP_CMD_B_SRC, breg) | 129 FIELD_PREP(OP_CMD_TOKEN, cmd_tgt_act[op].token) | 130 FIELD_PREP(OP_CMD_XFER, xfer) | 131 FIELD_PREP(OP_CMD_CNT, size) | 132 FIELD_PREP(OP_CMD_SIG, sync) | 133 FIELD_PREP(OP_CMD_TGT_CMD, cmd_tgt_act[op].tgt_cmd) | 134 FIELD_PREP(OP_CMD_MODE, mode); 135 136 nfp_prog_push(nfp_prog, insn); 137 } 138 139 static void 140 emit_cmd(struct nfp_prog *nfp_prog, enum cmd_tgt_map op, 141 u8 mode, u8 xfer, swreg lreg, swreg rreg, u8 size, bool sync) 142 { 143 struct nfp_insn_re_regs reg; 144 int err; 145 146 err = swreg_to_restricted(reg_none(), lreg, rreg, ®, false); 147 if (err) { 148 nfp_prog->error = err; 149 return; 150 } 151 if (reg.swap) { 152 pr_err("cmd can't swap arguments\n"); 153 nfp_prog->error = -EFAULT; 154 return; 155 } 156 if (reg.dst_lmextn || reg.src_lmextn) { 157 pr_err("cmd can't use LMextn\n"); 158 nfp_prog->error = -EFAULT; 159 return; 160 } 161 162 __emit_cmd(nfp_prog, op, mode, xfer, reg.areg, reg.breg, size, sync); 163 } 164 165 static void 166 __emit_br(struct nfp_prog *nfp_prog, enum br_mask mask, enum br_ev_pip ev_pip, 167 enum br_ctx_signal_state css, u16 addr, u8 defer) 168 { 169 u16 addr_lo, addr_hi; 170 u64 insn; 171 172 addr_lo = addr & (OP_BR_ADDR_LO >> __bf_shf(OP_BR_ADDR_LO)); 173 addr_hi = addr != addr_lo; 174 175 insn = OP_BR_BASE | 176 FIELD_PREP(OP_BR_MASK, mask) | 177 FIELD_PREP(OP_BR_EV_PIP, ev_pip) | 178 FIELD_PREP(OP_BR_CSS, css) | 179 FIELD_PREP(OP_BR_DEFBR, defer) | 180 FIELD_PREP(OP_BR_ADDR_LO, addr_lo) | 181 FIELD_PREP(OP_BR_ADDR_HI, addr_hi); 182 183 nfp_prog_push(nfp_prog, insn); 184 } 185 186 static void emit_br_def(struct nfp_prog *nfp_prog, u16 addr, u8 defer) 187 { 188 if (defer > 2) { 189 pr_err("BUG: branch defer out of bounds %d\n", defer); 190 nfp_prog->error = -EFAULT; 191 return; 192 } 193 __emit_br(nfp_prog, BR_UNC, BR_EV_PIP_UNCOND, BR_CSS_NONE, addr, defer); 194 } 195 196 static void 197 emit_br(struct nfp_prog *nfp_prog, enum br_mask mask, u16 addr, u8 defer) 198 { 199 __emit_br(nfp_prog, mask, 200 mask != BR_UNC ? BR_EV_PIP_COND : BR_EV_PIP_UNCOND, 201 BR_CSS_NONE, addr, defer); 202 } 203 204 static void 205 __emit_br_byte(struct nfp_prog *nfp_prog, u8 areg, u8 breg, bool imm8, 206 u8 byte, bool equal, u16 addr, u8 defer, bool src_lmextn) 207 { 208 u16 addr_lo, addr_hi; 209 u64 insn; 210 211 addr_lo = addr & (OP_BB_ADDR_LO >> __bf_shf(OP_BB_ADDR_LO)); 212 addr_hi = addr != addr_lo; 213 214 insn = OP_BBYTE_BASE | 215 FIELD_PREP(OP_BB_A_SRC, areg) | 216 FIELD_PREP(OP_BB_BYTE, byte) | 217 FIELD_PREP(OP_BB_B_SRC, breg) | 218 FIELD_PREP(OP_BB_I8, imm8) | 219 FIELD_PREP(OP_BB_EQ, equal) | 220 FIELD_PREP(OP_BB_DEFBR, defer) | 221 FIELD_PREP(OP_BB_ADDR_LO, addr_lo) | 222 FIELD_PREP(OP_BB_ADDR_HI, addr_hi) | 223 FIELD_PREP(OP_BB_SRC_LMEXTN, src_lmextn); 224 225 nfp_prog_push(nfp_prog, insn); 226 } 227 228 static void 229 emit_br_byte_neq(struct nfp_prog *nfp_prog, 230 swreg src, u8 imm, u8 byte, u16 addr, u8 defer) 231 { 232 struct nfp_insn_re_regs reg; 233 int err; 234 235 err = swreg_to_restricted(reg_none(), src, reg_imm(imm), ®, true); 236 if (err) { 237 nfp_prog->error = err; 238 return; 239 } 240 241 __emit_br_byte(nfp_prog, reg.areg, reg.breg, reg.i8, byte, false, addr, 242 defer, reg.src_lmextn); 243 } 244 245 static void 246 __emit_immed(struct nfp_prog *nfp_prog, u16 areg, u16 breg, u16 imm_hi, 247 enum immed_width width, bool invert, 248 enum immed_shift shift, bool wr_both, 249 bool dst_lmextn, bool src_lmextn) 250 { 251 u64 insn; 252 253 insn = OP_IMMED_BASE | 254 FIELD_PREP(OP_IMMED_A_SRC, areg) | 255 FIELD_PREP(OP_IMMED_B_SRC, breg) | 256 FIELD_PREP(OP_IMMED_IMM, imm_hi) | 257 FIELD_PREP(OP_IMMED_WIDTH, width) | 258 FIELD_PREP(OP_IMMED_INV, invert) | 259 FIELD_PREP(OP_IMMED_SHIFT, shift) | 260 FIELD_PREP(OP_IMMED_WR_AB, wr_both) | 261 FIELD_PREP(OP_IMMED_SRC_LMEXTN, src_lmextn) | 262 FIELD_PREP(OP_IMMED_DST_LMEXTN, dst_lmextn); 263 264 nfp_prog_push(nfp_prog, insn); 265 } 266 267 static void 268 emit_immed(struct nfp_prog *nfp_prog, swreg dst, u16 imm, 269 enum immed_width width, bool invert, enum immed_shift shift) 270 { 271 struct nfp_insn_ur_regs reg; 272 int err; 273 274 if (swreg_type(dst) == NN_REG_IMM) { 275 nfp_prog->error = -EFAULT; 276 return; 277 } 278 279 err = swreg_to_unrestricted(dst, dst, reg_imm(imm & 0xff), ®); 280 if (err) { 281 nfp_prog->error = err; 282 return; 283 } 284 285 __emit_immed(nfp_prog, reg.areg, reg.breg, imm >> 8, width, 286 invert, shift, reg.wr_both, 287 reg.dst_lmextn, reg.src_lmextn); 288 } 289 290 static void 291 __emit_shf(struct nfp_prog *nfp_prog, u16 dst, enum alu_dst_ab dst_ab, 292 enum shf_sc sc, u8 shift, 293 u16 areg, enum shf_op op, u16 breg, bool i8, bool sw, bool wr_both, 294 bool dst_lmextn, bool src_lmextn) 295 { 296 u64 insn; 297 298 if (!FIELD_FIT(OP_SHF_SHIFT, shift)) { 299 nfp_prog->error = -EFAULT; 300 return; 301 } 302 303 if (sc == SHF_SC_L_SHF) 304 shift = 32 - shift; 305 306 insn = OP_SHF_BASE | 307 FIELD_PREP(OP_SHF_A_SRC, areg) | 308 FIELD_PREP(OP_SHF_SC, sc) | 309 FIELD_PREP(OP_SHF_B_SRC, breg) | 310 FIELD_PREP(OP_SHF_I8, i8) | 311 FIELD_PREP(OP_SHF_SW, sw) | 312 FIELD_PREP(OP_SHF_DST, dst) | 313 FIELD_PREP(OP_SHF_SHIFT, shift) | 314 FIELD_PREP(OP_SHF_OP, op) | 315 FIELD_PREP(OP_SHF_DST_AB, dst_ab) | 316 FIELD_PREP(OP_SHF_WR_AB, wr_both) | 317 FIELD_PREP(OP_SHF_SRC_LMEXTN, src_lmextn) | 318 FIELD_PREP(OP_SHF_DST_LMEXTN, dst_lmextn); 319 320 nfp_prog_push(nfp_prog, insn); 321 } 322 323 static void 324 emit_shf(struct nfp_prog *nfp_prog, swreg dst, 325 swreg lreg, enum shf_op op, swreg rreg, enum shf_sc sc, u8 shift) 326 { 327 struct nfp_insn_re_regs reg; 328 int err; 329 330 err = swreg_to_restricted(dst, lreg, rreg, ®, true); 331 if (err) { 332 nfp_prog->error = err; 333 return; 334 } 335 336 __emit_shf(nfp_prog, reg.dst, reg.dst_ab, sc, shift, 337 reg.areg, op, reg.breg, reg.i8, reg.swap, reg.wr_both, 338 reg.dst_lmextn, reg.src_lmextn); 339 } 340 341 static void 342 __emit_alu(struct nfp_prog *nfp_prog, u16 dst, enum alu_dst_ab dst_ab, 343 u16 areg, enum alu_op op, u16 breg, bool swap, bool wr_both, 344 bool dst_lmextn, bool src_lmextn) 345 { 346 u64 insn; 347 348 insn = OP_ALU_BASE | 349 FIELD_PREP(OP_ALU_A_SRC, areg) | 350 FIELD_PREP(OP_ALU_B_SRC, breg) | 351 FIELD_PREP(OP_ALU_DST, dst) | 352 FIELD_PREP(OP_ALU_SW, swap) | 353 FIELD_PREP(OP_ALU_OP, op) | 354 FIELD_PREP(OP_ALU_DST_AB, dst_ab) | 355 FIELD_PREP(OP_ALU_WR_AB, wr_both) | 356 FIELD_PREP(OP_ALU_SRC_LMEXTN, src_lmextn) | 357 FIELD_PREP(OP_ALU_DST_LMEXTN, dst_lmextn); 358 359 nfp_prog_push(nfp_prog, insn); 360 } 361 362 static void 363 emit_alu(struct nfp_prog *nfp_prog, swreg dst, 364 swreg lreg, enum alu_op op, swreg rreg) 365 { 366 struct nfp_insn_ur_regs reg; 367 int err; 368 369 err = swreg_to_unrestricted(dst, lreg, rreg, ®); 370 if (err) { 371 nfp_prog->error = err; 372 return; 373 } 374 375 __emit_alu(nfp_prog, reg.dst, reg.dst_ab, 376 reg.areg, op, reg.breg, reg.swap, reg.wr_both, 377 reg.dst_lmextn, reg.src_lmextn); 378 } 379 380 static void 381 __emit_ld_field(struct nfp_prog *nfp_prog, enum shf_sc sc, 382 u8 areg, u8 bmask, u8 breg, u8 shift, bool imm8, 383 bool zero, bool swap, bool wr_both, 384 bool dst_lmextn, bool src_lmextn) 385 { 386 u64 insn; 387 388 insn = OP_LDF_BASE | 389 FIELD_PREP(OP_LDF_A_SRC, areg) | 390 FIELD_PREP(OP_LDF_SC, sc) | 391 FIELD_PREP(OP_LDF_B_SRC, breg) | 392 FIELD_PREP(OP_LDF_I8, imm8) | 393 FIELD_PREP(OP_LDF_SW, swap) | 394 FIELD_PREP(OP_LDF_ZF, zero) | 395 FIELD_PREP(OP_LDF_BMASK, bmask) | 396 FIELD_PREP(OP_LDF_SHF, shift) | 397 FIELD_PREP(OP_LDF_WR_AB, wr_both) | 398 FIELD_PREP(OP_LDF_SRC_LMEXTN, src_lmextn) | 399 FIELD_PREP(OP_LDF_DST_LMEXTN, dst_lmextn); 400 401 nfp_prog_push(nfp_prog, insn); 402 } 403 404 static void 405 emit_ld_field_any(struct nfp_prog *nfp_prog, swreg dst, u8 bmask, swreg src, 406 enum shf_sc sc, u8 shift, bool zero) 407 { 408 struct nfp_insn_re_regs reg; 409 int err; 410 411 /* Note: ld_field is special as it uses one of the src regs as dst */ 412 err = swreg_to_restricted(dst, dst, src, ®, true); 413 if (err) { 414 nfp_prog->error = err; 415 return; 416 } 417 418 __emit_ld_field(nfp_prog, sc, reg.areg, bmask, reg.breg, shift, 419 reg.i8, zero, reg.swap, reg.wr_both, 420 reg.dst_lmextn, reg.src_lmextn); 421 } 422 423 static void 424 emit_ld_field(struct nfp_prog *nfp_prog, swreg dst, u8 bmask, swreg src, 425 enum shf_sc sc, u8 shift) 426 { 427 emit_ld_field_any(nfp_prog, dst, bmask, src, sc, shift, false); 428 } 429 430 static void emit_nop(struct nfp_prog *nfp_prog) 431 { 432 __emit_immed(nfp_prog, UR_REG_IMM, UR_REG_IMM, 0, 0, 0, 0, 0, 0, 0); 433 } 434 435 /* --- Wrappers --- */ 436 static bool pack_immed(u32 imm, u16 *val, enum immed_shift *shift) 437 { 438 if (!(imm & 0xffff0000)) { 439 *val = imm; 440 *shift = IMMED_SHIFT_0B; 441 } else if (!(imm & 0xff0000ff)) { 442 *val = imm >> 8; 443 *shift = IMMED_SHIFT_1B; 444 } else if (!(imm & 0x0000ffff)) { 445 *val = imm >> 16; 446 *shift = IMMED_SHIFT_2B; 447 } else { 448 return false; 449 } 450 451 return true; 452 } 453 454 static void wrp_immed(struct nfp_prog *nfp_prog, swreg dst, u32 imm) 455 { 456 enum immed_shift shift; 457 u16 val; 458 459 if (pack_immed(imm, &val, &shift)) { 460 emit_immed(nfp_prog, dst, val, IMMED_WIDTH_ALL, false, shift); 461 } else if (pack_immed(~imm, &val, &shift)) { 462 emit_immed(nfp_prog, dst, val, IMMED_WIDTH_ALL, true, shift); 463 } else { 464 emit_immed(nfp_prog, dst, imm & 0xffff, IMMED_WIDTH_ALL, 465 false, IMMED_SHIFT_0B); 466 emit_immed(nfp_prog, dst, imm >> 16, IMMED_WIDTH_WORD, 467 false, IMMED_SHIFT_2B); 468 } 469 } 470 471 /* ur_load_imm_any() - encode immediate or use tmp register (unrestricted) 472 * If the @imm is small enough encode it directly in operand and return 473 * otherwise load @imm to a spare register and return its encoding. 474 */ 475 static swreg ur_load_imm_any(struct nfp_prog *nfp_prog, u32 imm, swreg tmp_reg) 476 { 477 if (FIELD_FIT(UR_REG_IMM_MAX, imm)) 478 return reg_imm(imm); 479 480 wrp_immed(nfp_prog, tmp_reg, imm); 481 return tmp_reg; 482 } 483 484 /* re_load_imm_any() - encode immediate or use tmp register (restricted) 485 * If the @imm is small enough encode it directly in operand and return 486 * otherwise load @imm to a spare register and return its encoding. 487 */ 488 static swreg re_load_imm_any(struct nfp_prog *nfp_prog, u32 imm, swreg tmp_reg) 489 { 490 if (FIELD_FIT(RE_REG_IMM_MAX, imm)) 491 return reg_imm(imm); 492 493 wrp_immed(nfp_prog, tmp_reg, imm); 494 return tmp_reg; 495 } 496 497 static void wrp_nops(struct nfp_prog *nfp_prog, unsigned int count) 498 { 499 while (count--) 500 emit_nop(nfp_prog); 501 } 502 503 static void 504 wrp_br_special(struct nfp_prog *nfp_prog, enum br_mask mask, 505 enum br_special special) 506 { 507 emit_br(nfp_prog, mask, 0, 0); 508 509 nfp_prog->prog[nfp_prog->prog_len - 1] |= 510 FIELD_PREP(OP_BR_SPECIAL, special); 511 } 512 513 static void wrp_mov(struct nfp_prog *nfp_prog, swreg dst, swreg src) 514 { 515 emit_alu(nfp_prog, dst, reg_none(), ALU_OP_NONE, src); 516 } 517 518 static void wrp_reg_mov(struct nfp_prog *nfp_prog, u16 dst, u16 src) 519 { 520 wrp_mov(nfp_prog, reg_both(dst), reg_b(src)); 521 } 522 523 static int 524 data_ld(struct nfp_prog *nfp_prog, swreg offset, u8 dst_gpr, int size) 525 { 526 unsigned int i; 527 u16 shift, sz; 528 529 /* We load the value from the address indicated in @offset and then 530 * shift out the data we don't need. Note: this is big endian! 531 */ 532 sz = max(size, 4); 533 shift = size < 4 ? 4 - size : 0; 534 535 emit_cmd(nfp_prog, CMD_TGT_READ8, CMD_MODE_32b, 0, 536 pptr_reg(nfp_prog), offset, sz - 1, true); 537 538 i = 0; 539 if (shift) 540 emit_shf(nfp_prog, reg_both(dst_gpr), reg_none(), SHF_OP_NONE, 541 reg_xfer(0), SHF_SC_R_SHF, shift * 8); 542 else 543 for (; i * 4 < size; i++) 544 wrp_mov(nfp_prog, reg_both(dst_gpr + i), reg_xfer(i)); 545 546 if (i < 2) 547 wrp_immed(nfp_prog, reg_both(dst_gpr + 1), 0); 548 549 return 0; 550 } 551 552 static int 553 data_ld_host_order(struct nfp_prog *nfp_prog, u8 src_gpr, swreg offset, 554 u8 dst_gpr, int size) 555 { 556 unsigned int i; 557 u8 mask, sz; 558 559 /* We load the value from the address indicated in @offset and then 560 * mask out the data we don't need. Note: this is little endian! 561 */ 562 sz = max(size, 4); 563 mask = size < 4 ? GENMASK(size - 1, 0) : 0; 564 565 emit_cmd(nfp_prog, CMD_TGT_READ32_SWAP, CMD_MODE_32b, 0, 566 reg_a(src_gpr), offset, sz / 4 - 1, true); 567 568 i = 0; 569 if (mask) 570 emit_ld_field_any(nfp_prog, reg_both(dst_gpr), mask, 571 reg_xfer(0), SHF_SC_NONE, 0, true); 572 else 573 for (; i * 4 < size; i++) 574 wrp_mov(nfp_prog, reg_both(dst_gpr + i), reg_xfer(i)); 575 576 if (i < 2) 577 wrp_immed(nfp_prog, reg_both(dst_gpr + 1), 0); 578 579 return 0; 580 } 581 582 static int 583 construct_data_ind_ld(struct nfp_prog *nfp_prog, u16 offset, u16 src, u8 size) 584 { 585 swreg tmp_reg; 586 587 /* Calculate the true offset (src_reg + imm) */ 588 tmp_reg = ur_load_imm_any(nfp_prog, offset, imm_b(nfp_prog)); 589 emit_alu(nfp_prog, imm_both(nfp_prog), reg_a(src), ALU_OP_ADD, tmp_reg); 590 591 /* Check packet length (size guaranteed to fit b/c it's u8) */ 592 emit_alu(nfp_prog, imm_a(nfp_prog), 593 imm_a(nfp_prog), ALU_OP_ADD, reg_imm(size)); 594 emit_alu(nfp_prog, reg_none(), 595 plen_reg(nfp_prog), ALU_OP_SUB, imm_a(nfp_prog)); 596 wrp_br_special(nfp_prog, BR_BLO, OP_BR_GO_ABORT); 597 598 /* Load data */ 599 return data_ld(nfp_prog, imm_b(nfp_prog), 0, size); 600 } 601 602 static int construct_data_ld(struct nfp_prog *nfp_prog, u16 offset, u8 size) 603 { 604 swreg tmp_reg; 605 606 /* Check packet length */ 607 tmp_reg = ur_load_imm_any(nfp_prog, offset + size, imm_a(nfp_prog)); 608 emit_alu(nfp_prog, reg_none(), plen_reg(nfp_prog), ALU_OP_SUB, tmp_reg); 609 wrp_br_special(nfp_prog, BR_BLO, OP_BR_GO_ABORT); 610 611 /* Load data */ 612 tmp_reg = re_load_imm_any(nfp_prog, offset, imm_b(nfp_prog)); 613 return data_ld(nfp_prog, tmp_reg, 0, size); 614 } 615 616 static int 617 data_stx_host_order(struct nfp_prog *nfp_prog, u8 dst_gpr, swreg offset, 618 u8 src_gpr, u8 size) 619 { 620 unsigned int i; 621 622 for (i = 0; i * 4 < size; i++) 623 wrp_mov(nfp_prog, reg_xfer(i), reg_a(src_gpr + i)); 624 625 emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 0, 626 reg_a(dst_gpr), offset, size - 1, true); 627 628 return 0; 629 } 630 631 static int 632 data_st_host_order(struct nfp_prog *nfp_prog, u8 dst_gpr, swreg offset, 633 u64 imm, u8 size) 634 { 635 wrp_immed(nfp_prog, reg_xfer(0), imm); 636 if (size == 8) 637 wrp_immed(nfp_prog, reg_xfer(1), imm >> 32); 638 639 emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 0, 640 reg_a(dst_gpr), offset, size - 1, true); 641 642 return 0; 643 } 644 645 typedef int 646 (*lmem_step)(struct nfp_prog *nfp_prog, u8 gpr, u8 gpr_byte, s32 off, 647 unsigned int size, bool new_gpr); 648 649 static int 650 wrp_lmem_load(struct nfp_prog *nfp_prog, u8 dst, u8 dst_byte, s32 off, 651 unsigned int size, bool new_gpr) 652 { 653 u32 idx, src_byte; 654 enum shf_sc sc; 655 swreg reg; 656 int shf; 657 u8 mask; 658 659 if (WARN_ON_ONCE(dst_byte + size > 4 || off % 4 + size > 4)) 660 return -EOPNOTSUPP; 661 662 idx = off / 4; 663 664 /* Move the entire word */ 665 if (size == 4) { 666 wrp_mov(nfp_prog, reg_both(dst), reg_lm(0, idx)); 667 return 0; 668 } 669 670 src_byte = off % 4; 671 672 mask = (1 << size) - 1; 673 mask <<= dst_byte; 674 675 if (WARN_ON_ONCE(mask > 0xf)) 676 return -EOPNOTSUPP; 677 678 shf = abs(src_byte - dst_byte) * 8; 679 if (src_byte == dst_byte) { 680 sc = SHF_SC_NONE; 681 } else if (src_byte < dst_byte) { 682 shf = 32 - shf; 683 sc = SHF_SC_L_SHF; 684 } else { 685 sc = SHF_SC_R_SHF; 686 } 687 688 /* ld_field can address fewer indexes, if offset too large do RMW. 689 * Because we RMV twice we waste 2 cycles on unaligned 8 byte writes. 690 */ 691 if (idx <= RE_REG_LM_IDX_MAX) { 692 reg = reg_lm(0, idx); 693 } else { 694 reg = imm_a(nfp_prog); 695 wrp_mov(nfp_prog, reg, reg_lm(0, idx)); 696 } 697 698 emit_ld_field_any(nfp_prog, reg_both(dst), mask, reg, sc, shf, new_gpr); 699 700 return 0; 701 } 702 703 static int 704 wrp_lmem_store(struct nfp_prog *nfp_prog, u8 src, u8 src_byte, s32 off, 705 unsigned int size, bool new_gpr) 706 { 707 u32 idx, dst_byte; 708 enum shf_sc sc; 709 swreg reg; 710 int shf; 711 u8 mask; 712 713 if (WARN_ON_ONCE(src_byte + size > 4 || off % 4 + size > 4)) 714 return -EOPNOTSUPP; 715 716 idx = off / 4; 717 718 /* Move the entire word */ 719 if (size == 4) { 720 wrp_mov(nfp_prog, reg_lm(0, idx), reg_b(src)); 721 return 0; 722 } 723 724 dst_byte = off % 4; 725 726 mask = (1 << size) - 1; 727 mask <<= dst_byte; 728 729 if (WARN_ON_ONCE(mask > 0xf)) 730 return -EOPNOTSUPP; 731 732 shf = abs(src_byte - dst_byte) * 8; 733 if (src_byte == dst_byte) { 734 sc = SHF_SC_NONE; 735 } else if (src_byte < dst_byte) { 736 shf = 32 - shf; 737 sc = SHF_SC_L_SHF; 738 } else { 739 sc = SHF_SC_R_SHF; 740 } 741 742 /* ld_field can address fewer indexes, if offset too large do RMW. 743 * Because we RMV twice we waste 2 cycles on unaligned 8 byte writes. 744 */ 745 if (idx <= RE_REG_LM_IDX_MAX) { 746 reg = reg_lm(0, idx); 747 } else { 748 reg = imm_a(nfp_prog); 749 wrp_mov(nfp_prog, reg, reg_lm(0, idx)); 750 } 751 752 emit_ld_field(nfp_prog, reg, mask, reg_b(src), sc, shf); 753 754 if (idx > RE_REG_LM_IDX_MAX) 755 wrp_mov(nfp_prog, reg_lm(0, idx), reg); 756 757 return 0; 758 } 759 760 static int 761 mem_op_stack(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 762 unsigned int size, u8 gpr, bool clr_gpr, lmem_step step) 763 { 764 s32 off = nfp_prog->stack_depth + meta->insn.off; 765 u8 prev_gpr = 255; 766 u32 gpr_byte = 0; 767 int ret; 768 769 if (clr_gpr && size < 8) 770 wrp_immed(nfp_prog, reg_both(gpr + 1), 0); 771 772 while (size) { 773 u32 slice_end; 774 u8 slice_size; 775 776 slice_size = min(size, 4 - gpr_byte); 777 slice_end = min(off + slice_size, round_up(off + 1, 4)); 778 slice_size = slice_end - off; 779 780 ret = step(nfp_prog, gpr, gpr_byte, off, slice_size, 781 gpr != prev_gpr); 782 if (ret) 783 return ret; 784 785 prev_gpr = gpr; 786 gpr_byte += slice_size; 787 if (gpr_byte >= 4) { 788 gpr_byte -= 4; 789 gpr++; 790 } 791 792 size -= slice_size; 793 off += slice_size; 794 } 795 796 return 0; 797 } 798 799 static void 800 wrp_alu_imm(struct nfp_prog *nfp_prog, u8 dst, enum alu_op alu_op, u32 imm) 801 { 802 swreg tmp_reg; 803 804 if (alu_op == ALU_OP_AND) { 805 if (!imm) 806 wrp_immed(nfp_prog, reg_both(dst), 0); 807 if (!imm || !~imm) 808 return; 809 } 810 if (alu_op == ALU_OP_OR) { 811 if (!~imm) 812 wrp_immed(nfp_prog, reg_both(dst), ~0U); 813 if (!imm || !~imm) 814 return; 815 } 816 if (alu_op == ALU_OP_XOR) { 817 if (!~imm) 818 emit_alu(nfp_prog, reg_both(dst), reg_none(), 819 ALU_OP_NEG, reg_b(dst)); 820 if (!imm || !~imm) 821 return; 822 } 823 824 tmp_reg = ur_load_imm_any(nfp_prog, imm, imm_b(nfp_prog)); 825 emit_alu(nfp_prog, reg_both(dst), reg_a(dst), alu_op, tmp_reg); 826 } 827 828 static int 829 wrp_alu64_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 830 enum alu_op alu_op, bool skip) 831 { 832 const struct bpf_insn *insn = &meta->insn; 833 u64 imm = insn->imm; /* sign extend */ 834 835 if (skip) { 836 meta->skip = true; 837 return 0; 838 } 839 840 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, alu_op, imm & ~0U); 841 wrp_alu_imm(nfp_prog, insn->dst_reg * 2 + 1, alu_op, imm >> 32); 842 843 return 0; 844 } 845 846 static int 847 wrp_alu64_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 848 enum alu_op alu_op) 849 { 850 u8 dst = meta->insn.dst_reg * 2, src = meta->insn.src_reg * 2; 851 852 emit_alu(nfp_prog, reg_both(dst), reg_a(dst), alu_op, reg_b(src)); 853 emit_alu(nfp_prog, reg_both(dst + 1), 854 reg_a(dst + 1), alu_op, reg_b(src + 1)); 855 856 return 0; 857 } 858 859 static int 860 wrp_alu32_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 861 enum alu_op alu_op, bool skip) 862 { 863 const struct bpf_insn *insn = &meta->insn; 864 865 if (skip) { 866 meta->skip = true; 867 return 0; 868 } 869 870 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, alu_op, insn->imm); 871 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0); 872 873 return 0; 874 } 875 876 static int 877 wrp_alu32_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 878 enum alu_op alu_op) 879 { 880 u8 dst = meta->insn.dst_reg * 2, src = meta->insn.src_reg * 2; 881 882 emit_alu(nfp_prog, reg_both(dst), reg_a(dst), alu_op, reg_b(src)); 883 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0); 884 885 return 0; 886 } 887 888 static void 889 wrp_test_reg_one(struct nfp_prog *nfp_prog, u8 dst, enum alu_op alu_op, u8 src, 890 enum br_mask br_mask, u16 off) 891 { 892 emit_alu(nfp_prog, reg_none(), reg_a(dst), alu_op, reg_b(src)); 893 emit_br(nfp_prog, br_mask, off, 0); 894 } 895 896 static int 897 wrp_test_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 898 enum alu_op alu_op, enum br_mask br_mask) 899 { 900 const struct bpf_insn *insn = &meta->insn; 901 902 if (insn->off < 0) /* TODO */ 903 return -EOPNOTSUPP; 904 905 wrp_test_reg_one(nfp_prog, insn->dst_reg * 2, alu_op, 906 insn->src_reg * 2, br_mask, insn->off); 907 wrp_test_reg_one(nfp_prog, insn->dst_reg * 2 + 1, alu_op, 908 insn->src_reg * 2 + 1, br_mask, insn->off); 909 910 return 0; 911 } 912 913 static int 914 wrp_cmp_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 915 enum br_mask br_mask, bool swap) 916 { 917 const struct bpf_insn *insn = &meta->insn; 918 u64 imm = insn->imm; /* sign extend */ 919 u8 reg = insn->dst_reg * 2; 920 swreg tmp_reg; 921 922 if (insn->off < 0) /* TODO */ 923 return -EOPNOTSUPP; 924 925 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog)); 926 if (!swap) 927 emit_alu(nfp_prog, reg_none(), reg_a(reg), ALU_OP_SUB, tmp_reg); 928 else 929 emit_alu(nfp_prog, reg_none(), tmp_reg, ALU_OP_SUB, reg_a(reg)); 930 931 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog)); 932 if (!swap) 933 emit_alu(nfp_prog, reg_none(), 934 reg_a(reg + 1), ALU_OP_SUB_C, tmp_reg); 935 else 936 emit_alu(nfp_prog, reg_none(), 937 tmp_reg, ALU_OP_SUB_C, reg_a(reg + 1)); 938 939 emit_br(nfp_prog, br_mask, insn->off, 0); 940 941 return 0; 942 } 943 944 static int 945 wrp_cmp_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 946 enum br_mask br_mask, bool swap) 947 { 948 const struct bpf_insn *insn = &meta->insn; 949 u8 areg, breg; 950 951 areg = insn->dst_reg * 2; 952 breg = insn->src_reg * 2; 953 954 if (insn->off < 0) /* TODO */ 955 return -EOPNOTSUPP; 956 957 if (swap) { 958 areg ^= breg; 959 breg ^= areg; 960 areg ^= breg; 961 } 962 963 emit_alu(nfp_prog, reg_none(), reg_a(areg), ALU_OP_SUB, reg_b(breg)); 964 emit_alu(nfp_prog, reg_none(), 965 reg_a(areg + 1), ALU_OP_SUB_C, reg_b(breg + 1)); 966 emit_br(nfp_prog, br_mask, insn->off, 0); 967 968 return 0; 969 } 970 971 static void wrp_end32(struct nfp_prog *nfp_prog, swreg reg_in, u8 gpr_out) 972 { 973 emit_ld_field(nfp_prog, reg_both(gpr_out), 0xf, reg_in, 974 SHF_SC_R_ROT, 8); 975 emit_ld_field(nfp_prog, reg_both(gpr_out), 0x5, reg_a(gpr_out), 976 SHF_SC_R_ROT, 16); 977 } 978 979 /* --- Callbacks --- */ 980 static int mov_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 981 { 982 const struct bpf_insn *insn = &meta->insn; 983 984 wrp_reg_mov(nfp_prog, insn->dst_reg * 2, insn->src_reg * 2); 985 wrp_reg_mov(nfp_prog, insn->dst_reg * 2 + 1, insn->src_reg * 2 + 1); 986 987 return 0; 988 } 989 990 static int mov_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 991 { 992 u64 imm = meta->insn.imm; /* sign extend */ 993 994 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2), imm & ~0U); 995 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), imm >> 32); 996 997 return 0; 998 } 999 1000 static int xor_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1001 { 1002 return wrp_alu64_reg(nfp_prog, meta, ALU_OP_XOR); 1003 } 1004 1005 static int xor_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1006 { 1007 return wrp_alu64_imm(nfp_prog, meta, ALU_OP_XOR, !meta->insn.imm); 1008 } 1009 1010 static int and_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1011 { 1012 return wrp_alu64_reg(nfp_prog, meta, ALU_OP_AND); 1013 } 1014 1015 static int and_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1016 { 1017 return wrp_alu64_imm(nfp_prog, meta, ALU_OP_AND, !~meta->insn.imm); 1018 } 1019 1020 static int or_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1021 { 1022 return wrp_alu64_reg(nfp_prog, meta, ALU_OP_OR); 1023 } 1024 1025 static int or_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1026 { 1027 return wrp_alu64_imm(nfp_prog, meta, ALU_OP_OR, !meta->insn.imm); 1028 } 1029 1030 static int add_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1031 { 1032 const struct bpf_insn *insn = &meta->insn; 1033 1034 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2), 1035 reg_a(insn->dst_reg * 2), ALU_OP_ADD, 1036 reg_b(insn->src_reg * 2)); 1037 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 1038 reg_a(insn->dst_reg * 2 + 1), ALU_OP_ADD_C, 1039 reg_b(insn->src_reg * 2 + 1)); 1040 1041 return 0; 1042 } 1043 1044 static int add_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1045 { 1046 const struct bpf_insn *insn = &meta->insn; 1047 u64 imm = insn->imm; /* sign extend */ 1048 1049 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, ALU_OP_ADD, imm & ~0U); 1050 wrp_alu_imm(nfp_prog, insn->dst_reg * 2 + 1, ALU_OP_ADD_C, imm >> 32); 1051 1052 return 0; 1053 } 1054 1055 static int sub_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1056 { 1057 const struct bpf_insn *insn = &meta->insn; 1058 1059 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2), 1060 reg_a(insn->dst_reg * 2), ALU_OP_SUB, 1061 reg_b(insn->src_reg * 2)); 1062 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 1063 reg_a(insn->dst_reg * 2 + 1), ALU_OP_SUB_C, 1064 reg_b(insn->src_reg * 2 + 1)); 1065 1066 return 0; 1067 } 1068 1069 static int sub_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1070 { 1071 const struct bpf_insn *insn = &meta->insn; 1072 u64 imm = insn->imm; /* sign extend */ 1073 1074 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, ALU_OP_SUB, imm & ~0U); 1075 wrp_alu_imm(nfp_prog, insn->dst_reg * 2 + 1, ALU_OP_SUB_C, imm >> 32); 1076 1077 return 0; 1078 } 1079 1080 static int shl_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1081 { 1082 const struct bpf_insn *insn = &meta->insn; 1083 u8 dst = insn->dst_reg * 2; 1084 1085 if (insn->imm < 32) { 1086 emit_shf(nfp_prog, reg_both(dst + 1), 1087 reg_a(dst + 1), SHF_OP_NONE, reg_b(dst), 1088 SHF_SC_R_DSHF, 32 - insn->imm); 1089 emit_shf(nfp_prog, reg_both(dst), 1090 reg_none(), SHF_OP_NONE, reg_b(dst), 1091 SHF_SC_L_SHF, insn->imm); 1092 } else if (insn->imm == 32) { 1093 wrp_reg_mov(nfp_prog, dst + 1, dst); 1094 wrp_immed(nfp_prog, reg_both(dst), 0); 1095 } else if (insn->imm > 32) { 1096 emit_shf(nfp_prog, reg_both(dst + 1), 1097 reg_none(), SHF_OP_NONE, reg_b(dst), 1098 SHF_SC_L_SHF, insn->imm - 32); 1099 wrp_immed(nfp_prog, reg_both(dst), 0); 1100 } 1101 1102 return 0; 1103 } 1104 1105 static int shr_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1106 { 1107 const struct bpf_insn *insn = &meta->insn; 1108 u8 dst = insn->dst_reg * 2; 1109 1110 if (insn->imm < 32) { 1111 emit_shf(nfp_prog, reg_both(dst), 1112 reg_a(dst + 1), SHF_OP_NONE, reg_b(dst), 1113 SHF_SC_R_DSHF, insn->imm); 1114 emit_shf(nfp_prog, reg_both(dst + 1), 1115 reg_none(), SHF_OP_NONE, reg_b(dst + 1), 1116 SHF_SC_R_SHF, insn->imm); 1117 } else if (insn->imm == 32) { 1118 wrp_reg_mov(nfp_prog, dst, dst + 1); 1119 wrp_immed(nfp_prog, reg_both(dst + 1), 0); 1120 } else if (insn->imm > 32) { 1121 emit_shf(nfp_prog, reg_both(dst), 1122 reg_none(), SHF_OP_NONE, reg_b(dst + 1), 1123 SHF_SC_R_SHF, insn->imm - 32); 1124 wrp_immed(nfp_prog, reg_both(dst + 1), 0); 1125 } 1126 1127 return 0; 1128 } 1129 1130 static int mov_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1131 { 1132 const struct bpf_insn *insn = &meta->insn; 1133 1134 wrp_reg_mov(nfp_prog, insn->dst_reg * 2, insn->src_reg * 2); 1135 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0); 1136 1137 return 0; 1138 } 1139 1140 static int mov_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1141 { 1142 const struct bpf_insn *insn = &meta->insn; 1143 1144 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2), insn->imm); 1145 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0); 1146 1147 return 0; 1148 } 1149 1150 static int xor_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1151 { 1152 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_XOR); 1153 } 1154 1155 static int xor_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1156 { 1157 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_XOR, !~meta->insn.imm); 1158 } 1159 1160 static int and_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1161 { 1162 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_AND); 1163 } 1164 1165 static int and_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1166 { 1167 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_AND, !~meta->insn.imm); 1168 } 1169 1170 static int or_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1171 { 1172 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_OR); 1173 } 1174 1175 static int or_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1176 { 1177 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_OR, !meta->insn.imm); 1178 } 1179 1180 static int add_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1181 { 1182 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_ADD); 1183 } 1184 1185 static int add_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1186 { 1187 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_ADD, !meta->insn.imm); 1188 } 1189 1190 static int sub_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1191 { 1192 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_SUB); 1193 } 1194 1195 static int sub_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1196 { 1197 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_SUB, !meta->insn.imm); 1198 } 1199 1200 static int shl_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1201 { 1202 const struct bpf_insn *insn = &meta->insn; 1203 1204 if (!insn->imm) 1205 return 1; /* TODO: zero shift means indirect */ 1206 1207 emit_shf(nfp_prog, reg_both(insn->dst_reg * 2), 1208 reg_none(), SHF_OP_NONE, reg_b(insn->dst_reg * 2), 1209 SHF_SC_L_SHF, insn->imm); 1210 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0); 1211 1212 return 0; 1213 } 1214 1215 static int end_reg32(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1216 { 1217 const struct bpf_insn *insn = &meta->insn; 1218 u8 gpr = insn->dst_reg * 2; 1219 1220 switch (insn->imm) { 1221 case 16: 1222 emit_ld_field(nfp_prog, reg_both(gpr), 0x9, reg_b(gpr), 1223 SHF_SC_R_ROT, 8); 1224 emit_ld_field(nfp_prog, reg_both(gpr), 0xe, reg_a(gpr), 1225 SHF_SC_R_SHF, 16); 1226 1227 wrp_immed(nfp_prog, reg_both(gpr + 1), 0); 1228 break; 1229 case 32: 1230 wrp_end32(nfp_prog, reg_a(gpr), gpr); 1231 wrp_immed(nfp_prog, reg_both(gpr + 1), 0); 1232 break; 1233 case 64: 1234 wrp_mov(nfp_prog, imm_a(nfp_prog), reg_b(gpr + 1)); 1235 1236 wrp_end32(nfp_prog, reg_a(gpr), gpr + 1); 1237 wrp_end32(nfp_prog, imm_a(nfp_prog), gpr); 1238 break; 1239 } 1240 1241 return 0; 1242 } 1243 1244 static int imm_ld8_part2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1245 { 1246 wrp_immed(nfp_prog, reg_both(nfp_meta_prev(meta)->insn.dst_reg * 2 + 1), 1247 meta->insn.imm); 1248 1249 return 0; 1250 } 1251 1252 static int imm_ld8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1253 { 1254 const struct bpf_insn *insn = &meta->insn; 1255 1256 meta->double_cb = imm_ld8_part2; 1257 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2), insn->imm); 1258 1259 return 0; 1260 } 1261 1262 static int data_ld1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1263 { 1264 return construct_data_ld(nfp_prog, meta->insn.imm, 1); 1265 } 1266 1267 static int data_ld2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1268 { 1269 return construct_data_ld(nfp_prog, meta->insn.imm, 2); 1270 } 1271 1272 static int data_ld4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1273 { 1274 return construct_data_ld(nfp_prog, meta->insn.imm, 4); 1275 } 1276 1277 static int data_ind_ld1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1278 { 1279 return construct_data_ind_ld(nfp_prog, meta->insn.imm, 1280 meta->insn.src_reg * 2, 1); 1281 } 1282 1283 static int data_ind_ld2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1284 { 1285 return construct_data_ind_ld(nfp_prog, meta->insn.imm, 1286 meta->insn.src_reg * 2, 2); 1287 } 1288 1289 static int data_ind_ld4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1290 { 1291 return construct_data_ind_ld(nfp_prog, meta->insn.imm, 1292 meta->insn.src_reg * 2, 4); 1293 } 1294 1295 static int 1296 mem_ldx_stack(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1297 unsigned int size) 1298 { 1299 return mem_op_stack(nfp_prog, meta, size, meta->insn.dst_reg * 2, true, 1300 wrp_lmem_load); 1301 } 1302 1303 static int mem_ldx_skb(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1304 u8 size) 1305 { 1306 swreg dst = reg_both(meta->insn.dst_reg * 2); 1307 1308 switch (meta->insn.off) { 1309 case offsetof(struct sk_buff, len): 1310 if (size != FIELD_SIZEOF(struct sk_buff, len)) 1311 return -EOPNOTSUPP; 1312 wrp_mov(nfp_prog, dst, plen_reg(nfp_prog)); 1313 break; 1314 case offsetof(struct sk_buff, data): 1315 if (size != sizeof(void *)) 1316 return -EOPNOTSUPP; 1317 wrp_mov(nfp_prog, dst, pptr_reg(nfp_prog)); 1318 break; 1319 case offsetof(struct sk_buff, cb) + 1320 offsetof(struct bpf_skb_data_end, data_end): 1321 if (size != sizeof(void *)) 1322 return -EOPNOTSUPP; 1323 emit_alu(nfp_prog, dst, 1324 plen_reg(nfp_prog), ALU_OP_ADD, pptr_reg(nfp_prog)); 1325 break; 1326 default: 1327 return -EOPNOTSUPP; 1328 } 1329 1330 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0); 1331 1332 return 0; 1333 } 1334 1335 static int mem_ldx_xdp(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1336 u8 size) 1337 { 1338 swreg dst = reg_both(meta->insn.dst_reg * 2); 1339 1340 if (size != sizeof(void *)) 1341 return -EINVAL; 1342 1343 switch (meta->insn.off) { 1344 case offsetof(struct xdp_buff, data): 1345 wrp_mov(nfp_prog, dst, pptr_reg(nfp_prog)); 1346 break; 1347 case offsetof(struct xdp_buff, data_end): 1348 emit_alu(nfp_prog, dst, 1349 plen_reg(nfp_prog), ALU_OP_ADD, pptr_reg(nfp_prog)); 1350 break; 1351 default: 1352 return -EOPNOTSUPP; 1353 } 1354 1355 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0); 1356 1357 return 0; 1358 } 1359 1360 static int 1361 mem_ldx_data(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1362 unsigned int size) 1363 { 1364 swreg tmp_reg; 1365 1366 tmp_reg = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog)); 1367 1368 return data_ld_host_order(nfp_prog, meta->insn.src_reg * 2, tmp_reg, 1369 meta->insn.dst_reg * 2, size); 1370 } 1371 1372 static int 1373 mem_ldx(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1374 unsigned int size) 1375 { 1376 if (meta->ptr.type == PTR_TO_CTX) { 1377 if (nfp_prog->act == NN_ACT_XDP) 1378 return mem_ldx_xdp(nfp_prog, meta, size); 1379 else 1380 return mem_ldx_skb(nfp_prog, meta, size); 1381 } 1382 1383 if (meta->ptr.type == PTR_TO_PACKET) 1384 return mem_ldx_data(nfp_prog, meta, size); 1385 1386 if (meta->ptr.type == PTR_TO_STACK) 1387 return mem_ldx_stack(nfp_prog, meta, size); 1388 1389 return -EOPNOTSUPP; 1390 } 1391 1392 static int mem_ldx1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1393 { 1394 return mem_ldx(nfp_prog, meta, 1); 1395 } 1396 1397 static int mem_ldx2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1398 { 1399 return mem_ldx(nfp_prog, meta, 2); 1400 } 1401 1402 static int mem_ldx4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1403 { 1404 return mem_ldx(nfp_prog, meta, 4); 1405 } 1406 1407 static int mem_ldx8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1408 { 1409 return mem_ldx(nfp_prog, meta, 8); 1410 } 1411 1412 static int 1413 mem_st_data(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1414 unsigned int size) 1415 { 1416 u64 imm = meta->insn.imm; /* sign extend */ 1417 swreg off_reg; 1418 1419 off_reg = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog)); 1420 1421 return data_st_host_order(nfp_prog, meta->insn.dst_reg * 2, off_reg, 1422 imm, size); 1423 } 1424 1425 static int mem_st(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1426 unsigned int size) 1427 { 1428 if (meta->ptr.type == PTR_TO_PACKET) 1429 return mem_st_data(nfp_prog, meta, size); 1430 1431 return -EOPNOTSUPP; 1432 } 1433 1434 static int mem_st1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1435 { 1436 return mem_st(nfp_prog, meta, 1); 1437 } 1438 1439 static int mem_st2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1440 { 1441 return mem_st(nfp_prog, meta, 2); 1442 } 1443 1444 static int mem_st4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1445 { 1446 return mem_st(nfp_prog, meta, 4); 1447 } 1448 1449 static int mem_st8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1450 { 1451 return mem_st(nfp_prog, meta, 8); 1452 } 1453 1454 static int 1455 mem_stx_data(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1456 unsigned int size) 1457 { 1458 swreg off_reg; 1459 1460 off_reg = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog)); 1461 1462 return data_stx_host_order(nfp_prog, meta->insn.dst_reg * 2, off_reg, 1463 meta->insn.src_reg * 2, size); 1464 } 1465 1466 static int 1467 mem_stx_stack(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1468 unsigned int size) 1469 { 1470 return mem_op_stack(nfp_prog, meta, size, meta->insn.src_reg * 2, false, 1471 wrp_lmem_store); 1472 } 1473 1474 static int 1475 mem_stx(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1476 unsigned int size) 1477 { 1478 if (meta->ptr.type == PTR_TO_PACKET) 1479 return mem_stx_data(nfp_prog, meta, size); 1480 1481 if (meta->ptr.type == PTR_TO_STACK) 1482 return mem_stx_stack(nfp_prog, meta, size); 1483 1484 return -EOPNOTSUPP; 1485 } 1486 1487 static int mem_stx1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1488 { 1489 return mem_stx(nfp_prog, meta, 1); 1490 } 1491 1492 static int mem_stx2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1493 { 1494 return mem_stx(nfp_prog, meta, 2); 1495 } 1496 1497 static int mem_stx4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1498 { 1499 return mem_stx(nfp_prog, meta, 4); 1500 } 1501 1502 static int mem_stx8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1503 { 1504 return mem_stx(nfp_prog, meta, 8); 1505 } 1506 1507 static int jump(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1508 { 1509 if (meta->insn.off < 0) /* TODO */ 1510 return -EOPNOTSUPP; 1511 emit_br(nfp_prog, BR_UNC, meta->insn.off, 0); 1512 1513 return 0; 1514 } 1515 1516 static int jeq_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1517 { 1518 const struct bpf_insn *insn = &meta->insn; 1519 u64 imm = insn->imm; /* sign extend */ 1520 swreg or1, or2, tmp_reg; 1521 1522 or1 = reg_a(insn->dst_reg * 2); 1523 or2 = reg_b(insn->dst_reg * 2 + 1); 1524 1525 if (insn->off < 0) /* TODO */ 1526 return -EOPNOTSUPP; 1527 1528 if (imm & ~0U) { 1529 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog)); 1530 emit_alu(nfp_prog, imm_a(nfp_prog), 1531 reg_a(insn->dst_reg * 2), ALU_OP_XOR, tmp_reg); 1532 or1 = imm_a(nfp_prog); 1533 } 1534 1535 if (imm >> 32) { 1536 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog)); 1537 emit_alu(nfp_prog, imm_b(nfp_prog), 1538 reg_a(insn->dst_reg * 2 + 1), ALU_OP_XOR, tmp_reg); 1539 or2 = imm_b(nfp_prog); 1540 } 1541 1542 emit_alu(nfp_prog, reg_none(), or1, ALU_OP_OR, or2); 1543 emit_br(nfp_prog, BR_BEQ, insn->off, 0); 1544 1545 return 0; 1546 } 1547 1548 static int jgt_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1549 { 1550 return wrp_cmp_imm(nfp_prog, meta, BR_BLO, true); 1551 } 1552 1553 static int jge_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1554 { 1555 return wrp_cmp_imm(nfp_prog, meta, BR_BHS, false); 1556 } 1557 1558 static int jlt_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1559 { 1560 return wrp_cmp_imm(nfp_prog, meta, BR_BLO, false); 1561 } 1562 1563 static int jle_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1564 { 1565 return wrp_cmp_imm(nfp_prog, meta, BR_BHS, true); 1566 } 1567 1568 static int jset_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1569 { 1570 const struct bpf_insn *insn = &meta->insn; 1571 u64 imm = insn->imm; /* sign extend */ 1572 swreg tmp_reg; 1573 1574 if (insn->off < 0) /* TODO */ 1575 return -EOPNOTSUPP; 1576 1577 if (!imm) { 1578 meta->skip = true; 1579 return 0; 1580 } 1581 1582 if (imm & ~0U) { 1583 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog)); 1584 emit_alu(nfp_prog, reg_none(), 1585 reg_a(insn->dst_reg * 2), ALU_OP_AND, tmp_reg); 1586 emit_br(nfp_prog, BR_BNE, insn->off, 0); 1587 } 1588 1589 if (imm >> 32) { 1590 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog)); 1591 emit_alu(nfp_prog, reg_none(), 1592 reg_a(insn->dst_reg * 2 + 1), ALU_OP_AND, tmp_reg); 1593 emit_br(nfp_prog, BR_BNE, insn->off, 0); 1594 } 1595 1596 return 0; 1597 } 1598 1599 static int jne_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1600 { 1601 const struct bpf_insn *insn = &meta->insn; 1602 u64 imm = insn->imm; /* sign extend */ 1603 swreg tmp_reg; 1604 1605 if (insn->off < 0) /* TODO */ 1606 return -EOPNOTSUPP; 1607 1608 if (!imm) { 1609 emit_alu(nfp_prog, reg_none(), reg_a(insn->dst_reg * 2), 1610 ALU_OP_OR, reg_b(insn->dst_reg * 2 + 1)); 1611 emit_br(nfp_prog, BR_BNE, insn->off, 0); 1612 return 0; 1613 } 1614 1615 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog)); 1616 emit_alu(nfp_prog, reg_none(), 1617 reg_a(insn->dst_reg * 2), ALU_OP_XOR, tmp_reg); 1618 emit_br(nfp_prog, BR_BNE, insn->off, 0); 1619 1620 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog)); 1621 emit_alu(nfp_prog, reg_none(), 1622 reg_a(insn->dst_reg * 2 + 1), ALU_OP_XOR, tmp_reg); 1623 emit_br(nfp_prog, BR_BNE, insn->off, 0); 1624 1625 return 0; 1626 } 1627 1628 static int jeq_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1629 { 1630 const struct bpf_insn *insn = &meta->insn; 1631 1632 if (insn->off < 0) /* TODO */ 1633 return -EOPNOTSUPP; 1634 1635 emit_alu(nfp_prog, imm_a(nfp_prog), reg_a(insn->dst_reg * 2), 1636 ALU_OP_XOR, reg_b(insn->src_reg * 2)); 1637 emit_alu(nfp_prog, imm_b(nfp_prog), reg_a(insn->dst_reg * 2 + 1), 1638 ALU_OP_XOR, reg_b(insn->src_reg * 2 + 1)); 1639 emit_alu(nfp_prog, reg_none(), 1640 imm_a(nfp_prog), ALU_OP_OR, imm_b(nfp_prog)); 1641 emit_br(nfp_prog, BR_BEQ, insn->off, 0); 1642 1643 return 0; 1644 } 1645 1646 static int jgt_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1647 { 1648 return wrp_cmp_reg(nfp_prog, meta, BR_BLO, true); 1649 } 1650 1651 static int jge_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1652 { 1653 return wrp_cmp_reg(nfp_prog, meta, BR_BHS, false); 1654 } 1655 1656 static int jlt_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1657 { 1658 return wrp_cmp_reg(nfp_prog, meta, BR_BLO, false); 1659 } 1660 1661 static int jle_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1662 { 1663 return wrp_cmp_reg(nfp_prog, meta, BR_BHS, true); 1664 } 1665 1666 static int jset_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1667 { 1668 return wrp_test_reg(nfp_prog, meta, ALU_OP_AND, BR_BNE); 1669 } 1670 1671 static int jne_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1672 { 1673 return wrp_test_reg(nfp_prog, meta, ALU_OP_XOR, BR_BNE); 1674 } 1675 1676 static int goto_out(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1677 { 1678 wrp_br_special(nfp_prog, BR_UNC, OP_BR_GO_OUT); 1679 1680 return 0; 1681 } 1682 1683 static const instr_cb_t instr_cb[256] = { 1684 [BPF_ALU64 | BPF_MOV | BPF_X] = mov_reg64, 1685 [BPF_ALU64 | BPF_MOV | BPF_K] = mov_imm64, 1686 [BPF_ALU64 | BPF_XOR | BPF_X] = xor_reg64, 1687 [BPF_ALU64 | BPF_XOR | BPF_K] = xor_imm64, 1688 [BPF_ALU64 | BPF_AND | BPF_X] = and_reg64, 1689 [BPF_ALU64 | BPF_AND | BPF_K] = and_imm64, 1690 [BPF_ALU64 | BPF_OR | BPF_X] = or_reg64, 1691 [BPF_ALU64 | BPF_OR | BPF_K] = or_imm64, 1692 [BPF_ALU64 | BPF_ADD | BPF_X] = add_reg64, 1693 [BPF_ALU64 | BPF_ADD | BPF_K] = add_imm64, 1694 [BPF_ALU64 | BPF_SUB | BPF_X] = sub_reg64, 1695 [BPF_ALU64 | BPF_SUB | BPF_K] = sub_imm64, 1696 [BPF_ALU64 | BPF_LSH | BPF_K] = shl_imm64, 1697 [BPF_ALU64 | BPF_RSH | BPF_K] = shr_imm64, 1698 [BPF_ALU | BPF_MOV | BPF_X] = mov_reg, 1699 [BPF_ALU | BPF_MOV | BPF_K] = mov_imm, 1700 [BPF_ALU | BPF_XOR | BPF_X] = xor_reg, 1701 [BPF_ALU | BPF_XOR | BPF_K] = xor_imm, 1702 [BPF_ALU | BPF_AND | BPF_X] = and_reg, 1703 [BPF_ALU | BPF_AND | BPF_K] = and_imm, 1704 [BPF_ALU | BPF_OR | BPF_X] = or_reg, 1705 [BPF_ALU | BPF_OR | BPF_K] = or_imm, 1706 [BPF_ALU | BPF_ADD | BPF_X] = add_reg, 1707 [BPF_ALU | BPF_ADD | BPF_K] = add_imm, 1708 [BPF_ALU | BPF_SUB | BPF_X] = sub_reg, 1709 [BPF_ALU | BPF_SUB | BPF_K] = sub_imm, 1710 [BPF_ALU | BPF_LSH | BPF_K] = shl_imm, 1711 [BPF_ALU | BPF_END | BPF_X] = end_reg32, 1712 [BPF_LD | BPF_IMM | BPF_DW] = imm_ld8, 1713 [BPF_LD | BPF_ABS | BPF_B] = data_ld1, 1714 [BPF_LD | BPF_ABS | BPF_H] = data_ld2, 1715 [BPF_LD | BPF_ABS | BPF_W] = data_ld4, 1716 [BPF_LD | BPF_IND | BPF_B] = data_ind_ld1, 1717 [BPF_LD | BPF_IND | BPF_H] = data_ind_ld2, 1718 [BPF_LD | BPF_IND | BPF_W] = data_ind_ld4, 1719 [BPF_LDX | BPF_MEM | BPF_B] = mem_ldx1, 1720 [BPF_LDX | BPF_MEM | BPF_H] = mem_ldx2, 1721 [BPF_LDX | BPF_MEM | BPF_W] = mem_ldx4, 1722 [BPF_LDX | BPF_MEM | BPF_DW] = mem_ldx8, 1723 [BPF_STX | BPF_MEM | BPF_B] = mem_stx1, 1724 [BPF_STX | BPF_MEM | BPF_H] = mem_stx2, 1725 [BPF_STX | BPF_MEM | BPF_W] = mem_stx4, 1726 [BPF_STX | BPF_MEM | BPF_DW] = mem_stx8, 1727 [BPF_ST | BPF_MEM | BPF_B] = mem_st1, 1728 [BPF_ST | BPF_MEM | BPF_H] = mem_st2, 1729 [BPF_ST | BPF_MEM | BPF_W] = mem_st4, 1730 [BPF_ST | BPF_MEM | BPF_DW] = mem_st8, 1731 [BPF_JMP | BPF_JA | BPF_K] = jump, 1732 [BPF_JMP | BPF_JEQ | BPF_K] = jeq_imm, 1733 [BPF_JMP | BPF_JGT | BPF_K] = jgt_imm, 1734 [BPF_JMP | BPF_JGE | BPF_K] = jge_imm, 1735 [BPF_JMP | BPF_JLT | BPF_K] = jlt_imm, 1736 [BPF_JMP | BPF_JLE | BPF_K] = jle_imm, 1737 [BPF_JMP | BPF_JSET | BPF_K] = jset_imm, 1738 [BPF_JMP | BPF_JNE | BPF_K] = jne_imm, 1739 [BPF_JMP | BPF_JEQ | BPF_X] = jeq_reg, 1740 [BPF_JMP | BPF_JGT | BPF_X] = jgt_reg, 1741 [BPF_JMP | BPF_JGE | BPF_X] = jge_reg, 1742 [BPF_JMP | BPF_JLT | BPF_X] = jlt_reg, 1743 [BPF_JMP | BPF_JLE | BPF_X] = jle_reg, 1744 [BPF_JMP | BPF_JSET | BPF_X] = jset_reg, 1745 [BPF_JMP | BPF_JNE | BPF_X] = jne_reg, 1746 [BPF_JMP | BPF_EXIT] = goto_out, 1747 }; 1748 1749 /* --- Misc code --- */ 1750 static void br_set_offset(u64 *instr, u16 offset) 1751 { 1752 u16 addr_lo, addr_hi; 1753 1754 addr_lo = offset & (OP_BR_ADDR_LO >> __bf_shf(OP_BR_ADDR_LO)); 1755 addr_hi = offset != addr_lo; 1756 *instr &= ~(OP_BR_ADDR_HI | OP_BR_ADDR_LO); 1757 *instr |= FIELD_PREP(OP_BR_ADDR_HI, addr_hi); 1758 *instr |= FIELD_PREP(OP_BR_ADDR_LO, addr_lo); 1759 } 1760 1761 /* --- Assembler logic --- */ 1762 static int nfp_fixup_branches(struct nfp_prog *nfp_prog) 1763 { 1764 struct nfp_insn_meta *meta, *next; 1765 u32 off, br_idx; 1766 u32 idx; 1767 1768 nfp_for_each_insn_walk2(nfp_prog, meta, next) { 1769 if (meta->skip) 1770 continue; 1771 if (BPF_CLASS(meta->insn.code) != BPF_JMP) 1772 continue; 1773 1774 br_idx = nfp_prog_offset_to_index(nfp_prog, next->off) - 1; 1775 if (!nfp_is_br(nfp_prog->prog[br_idx])) { 1776 pr_err("Fixup found block not ending in branch %d %02x %016llx!!\n", 1777 br_idx, meta->insn.code, nfp_prog->prog[br_idx]); 1778 return -ELOOP; 1779 } 1780 /* Leave special branches for later */ 1781 if (FIELD_GET(OP_BR_SPECIAL, nfp_prog->prog[br_idx])) 1782 continue; 1783 1784 /* Find the target offset in assembler realm */ 1785 off = meta->insn.off; 1786 if (!off) { 1787 pr_err("Fixup found zero offset!!\n"); 1788 return -ELOOP; 1789 } 1790 1791 while (off && nfp_meta_has_next(nfp_prog, next)) { 1792 next = nfp_meta_next(next); 1793 off--; 1794 } 1795 if (off) { 1796 pr_err("Fixup found too large jump!! %d\n", off); 1797 return -ELOOP; 1798 } 1799 1800 if (next->skip) { 1801 pr_err("Branch landing on removed instruction!!\n"); 1802 return -ELOOP; 1803 } 1804 1805 for (idx = nfp_prog_offset_to_index(nfp_prog, meta->off); 1806 idx <= br_idx; idx++) { 1807 if (!nfp_is_br(nfp_prog->prog[idx])) 1808 continue; 1809 br_set_offset(&nfp_prog->prog[idx], next->off); 1810 } 1811 } 1812 1813 /* Fixup 'goto out's separately, they can be scattered around */ 1814 for (br_idx = 0; br_idx < nfp_prog->prog_len; br_idx++) { 1815 enum br_special special; 1816 1817 if ((nfp_prog->prog[br_idx] & OP_BR_BASE_MASK) != OP_BR_BASE) 1818 continue; 1819 1820 special = FIELD_GET(OP_BR_SPECIAL, nfp_prog->prog[br_idx]); 1821 switch (special) { 1822 case OP_BR_NORMAL: 1823 break; 1824 case OP_BR_GO_OUT: 1825 br_set_offset(&nfp_prog->prog[br_idx], 1826 nfp_prog->tgt_out); 1827 break; 1828 case OP_BR_GO_ABORT: 1829 br_set_offset(&nfp_prog->prog[br_idx], 1830 nfp_prog->tgt_abort); 1831 break; 1832 } 1833 1834 nfp_prog->prog[br_idx] &= ~OP_BR_SPECIAL; 1835 } 1836 1837 return 0; 1838 } 1839 1840 static void nfp_intro(struct nfp_prog *nfp_prog) 1841 { 1842 wrp_immed(nfp_prog, plen_reg(nfp_prog), GENMASK(13, 0)); 1843 emit_alu(nfp_prog, plen_reg(nfp_prog), 1844 plen_reg(nfp_prog), ALU_OP_AND, pv_len(nfp_prog)); 1845 } 1846 1847 static void nfp_outro_tc_legacy(struct nfp_prog *nfp_prog) 1848 { 1849 const u8 act2code[] = { 1850 [NN_ACT_TC_DROP] = 0x22, 1851 [NN_ACT_TC_REDIR] = 0x24 1852 }; 1853 /* Target for aborts */ 1854 nfp_prog->tgt_abort = nfp_prog_current_offset(nfp_prog); 1855 wrp_immed(nfp_prog, reg_both(0), 0); 1856 1857 /* Target for normal exits */ 1858 nfp_prog->tgt_out = nfp_prog_current_offset(nfp_prog); 1859 /* Legacy TC mode: 1860 * 0 0x11 -> pass, count as stat0 1861 * -1 drop 0x22 -> drop, count as stat1 1862 * redir 0x24 -> redir, count as stat1 1863 * ife mark 0x21 -> pass, count as stat1 1864 * ife + tx 0x24 -> redir, count as stat1 1865 */ 1866 emit_br_byte_neq(nfp_prog, reg_b(0), 0xff, 0, nfp_prog->tgt_done, 2); 1867 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS); 1868 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_imm(0x11), SHF_SC_L_SHF, 16); 1869 1870 emit_br(nfp_prog, BR_UNC, nfp_prog->tgt_done, 1); 1871 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_imm(act2code[nfp_prog->act]), 1872 SHF_SC_L_SHF, 16); 1873 } 1874 1875 static void nfp_outro_tc_da(struct nfp_prog *nfp_prog) 1876 { 1877 /* TC direct-action mode: 1878 * 0,1 ok NOT SUPPORTED[1] 1879 * 2 drop 0x22 -> drop, count as stat1 1880 * 4,5 nuke 0x02 -> drop 1881 * 7 redir 0x44 -> redir, count as stat2 1882 * * unspec 0x11 -> pass, count as stat0 1883 * 1884 * [1] We can't support OK and RECLASSIFY because we can't tell TC 1885 * the exact decision made. We are forced to support UNSPEC 1886 * to handle aborts so that's the only one we handle for passing 1887 * packets up the stack. 1888 */ 1889 /* Target for aborts */ 1890 nfp_prog->tgt_abort = nfp_prog_current_offset(nfp_prog); 1891 1892 emit_br_def(nfp_prog, nfp_prog->tgt_done, 2); 1893 1894 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS); 1895 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_imm(0x11), SHF_SC_L_SHF, 16); 1896 1897 /* Target for normal exits */ 1898 nfp_prog->tgt_out = nfp_prog_current_offset(nfp_prog); 1899 1900 /* if R0 > 7 jump to abort */ 1901 emit_alu(nfp_prog, reg_none(), reg_imm(7), ALU_OP_SUB, reg_b(0)); 1902 emit_br(nfp_prog, BR_BLO, nfp_prog->tgt_abort, 0); 1903 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS); 1904 1905 wrp_immed(nfp_prog, reg_b(2), 0x41221211); 1906 wrp_immed(nfp_prog, reg_b(3), 0x41001211); 1907 1908 emit_shf(nfp_prog, reg_a(1), 1909 reg_none(), SHF_OP_NONE, reg_b(0), SHF_SC_L_SHF, 2); 1910 1911 emit_alu(nfp_prog, reg_none(), reg_a(1), ALU_OP_OR, reg_imm(0)); 1912 emit_shf(nfp_prog, reg_a(2), 1913 reg_imm(0xf), SHF_OP_AND, reg_b(2), SHF_SC_R_SHF, 0); 1914 1915 emit_alu(nfp_prog, reg_none(), reg_a(1), ALU_OP_OR, reg_imm(0)); 1916 emit_shf(nfp_prog, reg_b(2), 1917 reg_imm(0xf), SHF_OP_AND, reg_b(3), SHF_SC_R_SHF, 0); 1918 1919 emit_br_def(nfp_prog, nfp_prog->tgt_done, 2); 1920 1921 emit_shf(nfp_prog, reg_b(2), 1922 reg_a(2), SHF_OP_OR, reg_b(2), SHF_SC_L_SHF, 4); 1923 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_b(2), SHF_SC_L_SHF, 16); 1924 } 1925 1926 static void nfp_outro_xdp(struct nfp_prog *nfp_prog) 1927 { 1928 /* XDP return codes: 1929 * 0 aborted 0x82 -> drop, count as stat3 1930 * 1 drop 0x22 -> drop, count as stat1 1931 * 2 pass 0x11 -> pass, count as stat0 1932 * 3 tx 0x44 -> redir, count as stat2 1933 * * unknown 0x82 -> drop, count as stat3 1934 */ 1935 /* Target for aborts */ 1936 nfp_prog->tgt_abort = nfp_prog_current_offset(nfp_prog); 1937 1938 emit_br_def(nfp_prog, nfp_prog->tgt_done, 2); 1939 1940 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS); 1941 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_imm(0x82), SHF_SC_L_SHF, 16); 1942 1943 /* Target for normal exits */ 1944 nfp_prog->tgt_out = nfp_prog_current_offset(nfp_prog); 1945 1946 /* if R0 > 3 jump to abort */ 1947 emit_alu(nfp_prog, reg_none(), reg_imm(3), ALU_OP_SUB, reg_b(0)); 1948 emit_br(nfp_prog, BR_BLO, nfp_prog->tgt_abort, 0); 1949 1950 wrp_immed(nfp_prog, reg_b(2), 0x44112282); 1951 1952 emit_shf(nfp_prog, reg_a(1), 1953 reg_none(), SHF_OP_NONE, reg_b(0), SHF_SC_L_SHF, 3); 1954 1955 emit_alu(nfp_prog, reg_none(), reg_a(1), ALU_OP_OR, reg_imm(0)); 1956 emit_shf(nfp_prog, reg_b(2), 1957 reg_imm(0xff), SHF_OP_AND, reg_b(2), SHF_SC_R_SHF, 0); 1958 1959 emit_br_def(nfp_prog, nfp_prog->tgt_done, 2); 1960 1961 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS); 1962 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_b(2), SHF_SC_L_SHF, 16); 1963 } 1964 1965 static void nfp_outro(struct nfp_prog *nfp_prog) 1966 { 1967 switch (nfp_prog->act) { 1968 case NN_ACT_DIRECT: 1969 nfp_outro_tc_da(nfp_prog); 1970 break; 1971 case NN_ACT_TC_DROP: 1972 case NN_ACT_TC_REDIR: 1973 nfp_outro_tc_legacy(nfp_prog); 1974 break; 1975 case NN_ACT_XDP: 1976 nfp_outro_xdp(nfp_prog); 1977 break; 1978 } 1979 } 1980 1981 static int nfp_translate(struct nfp_prog *nfp_prog) 1982 { 1983 struct nfp_insn_meta *meta; 1984 int err; 1985 1986 nfp_intro(nfp_prog); 1987 if (nfp_prog->error) 1988 return nfp_prog->error; 1989 1990 list_for_each_entry(meta, &nfp_prog->insns, l) { 1991 instr_cb_t cb = instr_cb[meta->insn.code]; 1992 1993 meta->off = nfp_prog_current_offset(nfp_prog); 1994 1995 if (meta->skip) { 1996 nfp_prog->n_translated++; 1997 continue; 1998 } 1999 2000 if (nfp_meta_has_prev(nfp_prog, meta) && 2001 nfp_meta_prev(meta)->double_cb) 2002 cb = nfp_meta_prev(meta)->double_cb; 2003 if (!cb) 2004 return -ENOENT; 2005 err = cb(nfp_prog, meta); 2006 if (err) 2007 return err; 2008 2009 nfp_prog->n_translated++; 2010 } 2011 2012 nfp_outro(nfp_prog); 2013 if (nfp_prog->error) 2014 return nfp_prog->error; 2015 2016 wrp_nops(nfp_prog, NFP_USTORE_PREFETCH_WINDOW); 2017 if (nfp_prog->error) 2018 return nfp_prog->error; 2019 2020 return nfp_fixup_branches(nfp_prog); 2021 } 2022 2023 static int 2024 nfp_prog_prepare(struct nfp_prog *nfp_prog, const struct bpf_insn *prog, 2025 unsigned int cnt) 2026 { 2027 unsigned int i; 2028 2029 for (i = 0; i < cnt; i++) { 2030 struct nfp_insn_meta *meta; 2031 2032 meta = kzalloc(sizeof(*meta), GFP_KERNEL); 2033 if (!meta) 2034 return -ENOMEM; 2035 2036 meta->insn = prog[i]; 2037 meta->n = i; 2038 2039 list_add_tail(&meta->l, &nfp_prog->insns); 2040 } 2041 2042 return 0; 2043 } 2044 2045 /* --- Optimizations --- */ 2046 static void nfp_bpf_opt_reg_init(struct nfp_prog *nfp_prog) 2047 { 2048 struct nfp_insn_meta *meta; 2049 2050 list_for_each_entry(meta, &nfp_prog->insns, l) { 2051 struct bpf_insn insn = meta->insn; 2052 2053 /* Programs converted from cBPF start with register xoring */ 2054 if (insn.code == (BPF_ALU64 | BPF_XOR | BPF_X) && 2055 insn.src_reg == insn.dst_reg) 2056 continue; 2057 2058 /* Programs start with R6 = R1 but we ignore the skb pointer */ 2059 if (insn.code == (BPF_ALU64 | BPF_MOV | BPF_X) && 2060 insn.src_reg == 1 && insn.dst_reg == 6) 2061 meta->skip = true; 2062 2063 /* Return as soon as something doesn't match */ 2064 if (!meta->skip) 2065 return; 2066 } 2067 } 2068 2069 /* Remove masking after load since our load guarantees this is not needed */ 2070 static void nfp_bpf_opt_ld_mask(struct nfp_prog *nfp_prog) 2071 { 2072 struct nfp_insn_meta *meta1, *meta2; 2073 const s32 exp_mask[] = { 2074 [BPF_B] = 0x000000ffU, 2075 [BPF_H] = 0x0000ffffU, 2076 [BPF_W] = 0xffffffffU, 2077 }; 2078 2079 nfp_for_each_insn_walk2(nfp_prog, meta1, meta2) { 2080 struct bpf_insn insn, next; 2081 2082 insn = meta1->insn; 2083 next = meta2->insn; 2084 2085 if (BPF_CLASS(insn.code) != BPF_LD) 2086 continue; 2087 if (BPF_MODE(insn.code) != BPF_ABS && 2088 BPF_MODE(insn.code) != BPF_IND) 2089 continue; 2090 2091 if (next.code != (BPF_ALU64 | BPF_AND | BPF_K)) 2092 continue; 2093 2094 if (!exp_mask[BPF_SIZE(insn.code)]) 2095 continue; 2096 if (exp_mask[BPF_SIZE(insn.code)] != next.imm) 2097 continue; 2098 2099 if (next.src_reg || next.dst_reg) 2100 continue; 2101 2102 meta2->skip = true; 2103 } 2104 } 2105 2106 static void nfp_bpf_opt_ld_shift(struct nfp_prog *nfp_prog) 2107 { 2108 struct nfp_insn_meta *meta1, *meta2, *meta3; 2109 2110 nfp_for_each_insn_walk3(nfp_prog, meta1, meta2, meta3) { 2111 struct bpf_insn insn, next1, next2; 2112 2113 insn = meta1->insn; 2114 next1 = meta2->insn; 2115 next2 = meta3->insn; 2116 2117 if (BPF_CLASS(insn.code) != BPF_LD) 2118 continue; 2119 if (BPF_MODE(insn.code) != BPF_ABS && 2120 BPF_MODE(insn.code) != BPF_IND) 2121 continue; 2122 if (BPF_SIZE(insn.code) != BPF_W) 2123 continue; 2124 2125 if (!(next1.code == (BPF_LSH | BPF_K | BPF_ALU64) && 2126 next2.code == (BPF_RSH | BPF_K | BPF_ALU64)) && 2127 !(next1.code == (BPF_RSH | BPF_K | BPF_ALU64) && 2128 next2.code == (BPF_LSH | BPF_K | BPF_ALU64))) 2129 continue; 2130 2131 if (next1.src_reg || next1.dst_reg || 2132 next2.src_reg || next2.dst_reg) 2133 continue; 2134 2135 if (next1.imm != 0x20 || next2.imm != 0x20) 2136 continue; 2137 2138 meta2->skip = true; 2139 meta3->skip = true; 2140 } 2141 } 2142 2143 static int nfp_bpf_optimize(struct nfp_prog *nfp_prog) 2144 { 2145 nfp_bpf_opt_reg_init(nfp_prog); 2146 2147 nfp_bpf_opt_ld_mask(nfp_prog); 2148 nfp_bpf_opt_ld_shift(nfp_prog); 2149 2150 return 0; 2151 } 2152 2153 static int nfp_bpf_ustore_calc(struct nfp_prog *nfp_prog, __le64 *ustore) 2154 { 2155 int i; 2156 2157 for (i = 0; i < nfp_prog->prog_len; i++) { 2158 int err; 2159 2160 err = nfp_ustore_check_valid_no_ecc(nfp_prog->prog[i]); 2161 if (err) 2162 return err; 2163 2164 nfp_prog->prog[i] = nfp_ustore_calc_ecc_insn(nfp_prog->prog[i]); 2165 2166 ustore[i] = cpu_to_le64(nfp_prog->prog[i]); 2167 } 2168 2169 return 0; 2170 } 2171 2172 /** 2173 * nfp_bpf_jit() - translate BPF code into NFP assembly 2174 * @filter: kernel BPF filter struct 2175 * @prog_mem: memory to store assembler instructions 2176 * @act: action attached to this eBPF program 2177 * @prog_start: offset of the first instruction when loaded 2178 * @prog_done: where to jump on exit 2179 * @prog_sz: size of @prog_mem in instructions 2180 * @res: achieved parameters of translation results 2181 */ 2182 int 2183 nfp_bpf_jit(struct bpf_prog *filter, void *prog_mem, 2184 enum nfp_bpf_action_type act, 2185 unsigned int prog_start, unsigned int prog_done, 2186 unsigned int prog_sz, struct nfp_bpf_result *res) 2187 { 2188 struct nfp_prog *nfp_prog; 2189 int ret; 2190 2191 nfp_prog = kzalloc(sizeof(*nfp_prog), GFP_KERNEL); 2192 if (!nfp_prog) 2193 return -ENOMEM; 2194 2195 INIT_LIST_HEAD(&nfp_prog->insns); 2196 nfp_prog->act = act; 2197 nfp_prog->start_off = prog_start; 2198 nfp_prog->tgt_done = prog_done; 2199 2200 ret = nfp_prog_prepare(nfp_prog, filter->insnsi, filter->len); 2201 if (ret) 2202 goto out; 2203 2204 ret = nfp_prog_verify(nfp_prog, filter); 2205 if (ret) 2206 goto out; 2207 2208 ret = nfp_bpf_optimize(nfp_prog); 2209 if (ret) 2210 goto out; 2211 2212 nfp_prog->num_regs = MAX_BPF_REG; 2213 nfp_prog->regs_per_thread = 32; 2214 2215 nfp_prog->prog = prog_mem; 2216 nfp_prog->__prog_alloc_len = prog_sz; 2217 2218 ret = nfp_translate(nfp_prog); 2219 if (ret) { 2220 pr_err("Translation failed with error %d (translated: %u)\n", 2221 ret, nfp_prog->n_translated); 2222 ret = -EINVAL; 2223 goto out; 2224 } 2225 2226 ret = nfp_bpf_ustore_calc(nfp_prog, (__force __le64 *)prog_mem); 2227 2228 res->n_instr = nfp_prog->prog_len; 2229 res->dense_mode = false; 2230 out: 2231 nfp_prog_free(nfp_prog); 2232 2233 return ret; 2234 } 2235