1 /* 2 * Copyright (C) 2016-2017 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_prev(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 70 { 71 return meta->l.prev != &nfp_prog->insns; 72 } 73 74 static void nfp_prog_push(struct nfp_prog *nfp_prog, u64 insn) 75 { 76 if (nfp_prog->__prog_alloc_len == nfp_prog->prog_len) { 77 nfp_prog->error = -ENOSPC; 78 return; 79 } 80 81 nfp_prog->prog[nfp_prog->prog_len] = insn; 82 nfp_prog->prog_len++; 83 } 84 85 static unsigned int nfp_prog_current_offset(struct nfp_prog *nfp_prog) 86 { 87 return nfp_prog->start_off + nfp_prog->prog_len; 88 } 89 90 static unsigned int 91 nfp_prog_offset_to_index(struct nfp_prog *nfp_prog, unsigned int offset) 92 { 93 return offset - nfp_prog->start_off; 94 } 95 96 /* --- Emitters --- */ 97 static void 98 __emit_cmd(struct nfp_prog *nfp_prog, enum cmd_tgt_map op, 99 u8 mode, u8 xfer, u8 areg, u8 breg, u8 size, bool sync) 100 { 101 enum cmd_ctx_swap ctx; 102 u64 insn; 103 104 if (sync) 105 ctx = CMD_CTX_SWAP; 106 else 107 ctx = CMD_CTX_NO_SWAP; 108 109 insn = FIELD_PREP(OP_CMD_A_SRC, areg) | 110 FIELD_PREP(OP_CMD_CTX, ctx) | 111 FIELD_PREP(OP_CMD_B_SRC, breg) | 112 FIELD_PREP(OP_CMD_TOKEN, cmd_tgt_act[op].token) | 113 FIELD_PREP(OP_CMD_XFER, xfer) | 114 FIELD_PREP(OP_CMD_CNT, size) | 115 FIELD_PREP(OP_CMD_SIG, sync) | 116 FIELD_PREP(OP_CMD_TGT_CMD, cmd_tgt_act[op].tgt_cmd) | 117 FIELD_PREP(OP_CMD_MODE, mode); 118 119 nfp_prog_push(nfp_prog, insn); 120 } 121 122 static void 123 emit_cmd(struct nfp_prog *nfp_prog, enum cmd_tgt_map op, 124 u8 mode, u8 xfer, swreg lreg, swreg rreg, u8 size, bool sync) 125 { 126 struct nfp_insn_re_regs reg; 127 int err; 128 129 err = swreg_to_restricted(reg_none(), lreg, rreg, ®, false); 130 if (err) { 131 nfp_prog->error = err; 132 return; 133 } 134 if (reg.swap) { 135 pr_err("cmd can't swap arguments\n"); 136 nfp_prog->error = -EFAULT; 137 return; 138 } 139 if (reg.dst_lmextn || reg.src_lmextn) { 140 pr_err("cmd can't use LMextn\n"); 141 nfp_prog->error = -EFAULT; 142 return; 143 } 144 145 __emit_cmd(nfp_prog, op, mode, xfer, reg.areg, reg.breg, size, sync); 146 } 147 148 static void 149 __emit_br(struct nfp_prog *nfp_prog, enum br_mask mask, enum br_ev_pip ev_pip, 150 enum br_ctx_signal_state css, u16 addr, u8 defer) 151 { 152 u16 addr_lo, addr_hi; 153 u64 insn; 154 155 addr_lo = addr & (OP_BR_ADDR_LO >> __bf_shf(OP_BR_ADDR_LO)); 156 addr_hi = addr != addr_lo; 157 158 insn = OP_BR_BASE | 159 FIELD_PREP(OP_BR_MASK, mask) | 160 FIELD_PREP(OP_BR_EV_PIP, ev_pip) | 161 FIELD_PREP(OP_BR_CSS, css) | 162 FIELD_PREP(OP_BR_DEFBR, defer) | 163 FIELD_PREP(OP_BR_ADDR_LO, addr_lo) | 164 FIELD_PREP(OP_BR_ADDR_HI, addr_hi); 165 166 nfp_prog_push(nfp_prog, insn); 167 } 168 169 static void emit_br_def(struct nfp_prog *nfp_prog, u16 addr, u8 defer) 170 { 171 if (defer > 2) { 172 pr_err("BUG: branch defer out of bounds %d\n", defer); 173 nfp_prog->error = -EFAULT; 174 return; 175 } 176 __emit_br(nfp_prog, BR_UNC, BR_EV_PIP_UNCOND, BR_CSS_NONE, addr, defer); 177 } 178 179 static void 180 emit_br(struct nfp_prog *nfp_prog, enum br_mask mask, u16 addr, u8 defer) 181 { 182 __emit_br(nfp_prog, mask, 183 mask != BR_UNC ? BR_EV_PIP_COND : BR_EV_PIP_UNCOND, 184 BR_CSS_NONE, addr, defer); 185 } 186 187 static void 188 __emit_immed(struct nfp_prog *nfp_prog, u16 areg, u16 breg, u16 imm_hi, 189 enum immed_width width, bool invert, 190 enum immed_shift shift, bool wr_both, 191 bool dst_lmextn, bool src_lmextn) 192 { 193 u64 insn; 194 195 insn = OP_IMMED_BASE | 196 FIELD_PREP(OP_IMMED_A_SRC, areg) | 197 FIELD_PREP(OP_IMMED_B_SRC, breg) | 198 FIELD_PREP(OP_IMMED_IMM, imm_hi) | 199 FIELD_PREP(OP_IMMED_WIDTH, width) | 200 FIELD_PREP(OP_IMMED_INV, invert) | 201 FIELD_PREP(OP_IMMED_SHIFT, shift) | 202 FIELD_PREP(OP_IMMED_WR_AB, wr_both) | 203 FIELD_PREP(OP_IMMED_SRC_LMEXTN, src_lmextn) | 204 FIELD_PREP(OP_IMMED_DST_LMEXTN, dst_lmextn); 205 206 nfp_prog_push(nfp_prog, insn); 207 } 208 209 static void 210 emit_immed(struct nfp_prog *nfp_prog, swreg dst, u16 imm, 211 enum immed_width width, bool invert, enum immed_shift shift) 212 { 213 struct nfp_insn_ur_regs reg; 214 int err; 215 216 if (swreg_type(dst) == NN_REG_IMM) { 217 nfp_prog->error = -EFAULT; 218 return; 219 } 220 221 err = swreg_to_unrestricted(dst, dst, reg_imm(imm & 0xff), ®); 222 if (err) { 223 nfp_prog->error = err; 224 return; 225 } 226 227 __emit_immed(nfp_prog, reg.areg, reg.breg, imm >> 8, width, 228 invert, shift, reg.wr_both, 229 reg.dst_lmextn, reg.src_lmextn); 230 } 231 232 static void 233 __emit_shf(struct nfp_prog *nfp_prog, u16 dst, enum alu_dst_ab dst_ab, 234 enum shf_sc sc, u8 shift, 235 u16 areg, enum shf_op op, u16 breg, bool i8, bool sw, bool wr_both, 236 bool dst_lmextn, bool src_lmextn) 237 { 238 u64 insn; 239 240 if (!FIELD_FIT(OP_SHF_SHIFT, shift)) { 241 nfp_prog->error = -EFAULT; 242 return; 243 } 244 245 if (sc == SHF_SC_L_SHF) 246 shift = 32 - shift; 247 248 insn = OP_SHF_BASE | 249 FIELD_PREP(OP_SHF_A_SRC, areg) | 250 FIELD_PREP(OP_SHF_SC, sc) | 251 FIELD_PREP(OP_SHF_B_SRC, breg) | 252 FIELD_PREP(OP_SHF_I8, i8) | 253 FIELD_PREP(OP_SHF_SW, sw) | 254 FIELD_PREP(OP_SHF_DST, dst) | 255 FIELD_PREP(OP_SHF_SHIFT, shift) | 256 FIELD_PREP(OP_SHF_OP, op) | 257 FIELD_PREP(OP_SHF_DST_AB, dst_ab) | 258 FIELD_PREP(OP_SHF_WR_AB, wr_both) | 259 FIELD_PREP(OP_SHF_SRC_LMEXTN, src_lmextn) | 260 FIELD_PREP(OP_SHF_DST_LMEXTN, dst_lmextn); 261 262 nfp_prog_push(nfp_prog, insn); 263 } 264 265 static void 266 emit_shf(struct nfp_prog *nfp_prog, swreg dst, 267 swreg lreg, enum shf_op op, swreg rreg, enum shf_sc sc, u8 shift) 268 { 269 struct nfp_insn_re_regs reg; 270 int err; 271 272 err = swreg_to_restricted(dst, lreg, rreg, ®, true); 273 if (err) { 274 nfp_prog->error = err; 275 return; 276 } 277 278 __emit_shf(nfp_prog, reg.dst, reg.dst_ab, sc, shift, 279 reg.areg, op, reg.breg, reg.i8, reg.swap, reg.wr_both, 280 reg.dst_lmextn, reg.src_lmextn); 281 } 282 283 static void 284 __emit_alu(struct nfp_prog *nfp_prog, u16 dst, enum alu_dst_ab dst_ab, 285 u16 areg, enum alu_op op, u16 breg, bool swap, bool wr_both, 286 bool dst_lmextn, bool src_lmextn) 287 { 288 u64 insn; 289 290 insn = OP_ALU_BASE | 291 FIELD_PREP(OP_ALU_A_SRC, areg) | 292 FIELD_PREP(OP_ALU_B_SRC, breg) | 293 FIELD_PREP(OP_ALU_DST, dst) | 294 FIELD_PREP(OP_ALU_SW, swap) | 295 FIELD_PREP(OP_ALU_OP, op) | 296 FIELD_PREP(OP_ALU_DST_AB, dst_ab) | 297 FIELD_PREP(OP_ALU_WR_AB, wr_both) | 298 FIELD_PREP(OP_ALU_SRC_LMEXTN, src_lmextn) | 299 FIELD_PREP(OP_ALU_DST_LMEXTN, dst_lmextn); 300 301 nfp_prog_push(nfp_prog, insn); 302 } 303 304 static void 305 emit_alu(struct nfp_prog *nfp_prog, swreg dst, 306 swreg lreg, enum alu_op op, swreg rreg) 307 { 308 struct nfp_insn_ur_regs reg; 309 int err; 310 311 err = swreg_to_unrestricted(dst, lreg, rreg, ®); 312 if (err) { 313 nfp_prog->error = err; 314 return; 315 } 316 317 __emit_alu(nfp_prog, reg.dst, reg.dst_ab, 318 reg.areg, op, reg.breg, reg.swap, reg.wr_both, 319 reg.dst_lmextn, reg.src_lmextn); 320 } 321 322 static void 323 __emit_ld_field(struct nfp_prog *nfp_prog, enum shf_sc sc, 324 u8 areg, u8 bmask, u8 breg, u8 shift, bool imm8, 325 bool zero, bool swap, bool wr_both, 326 bool dst_lmextn, bool src_lmextn) 327 { 328 u64 insn; 329 330 insn = OP_LDF_BASE | 331 FIELD_PREP(OP_LDF_A_SRC, areg) | 332 FIELD_PREP(OP_LDF_SC, sc) | 333 FIELD_PREP(OP_LDF_B_SRC, breg) | 334 FIELD_PREP(OP_LDF_I8, imm8) | 335 FIELD_PREP(OP_LDF_SW, swap) | 336 FIELD_PREP(OP_LDF_ZF, zero) | 337 FIELD_PREP(OP_LDF_BMASK, bmask) | 338 FIELD_PREP(OP_LDF_SHF, shift) | 339 FIELD_PREP(OP_LDF_WR_AB, wr_both) | 340 FIELD_PREP(OP_LDF_SRC_LMEXTN, src_lmextn) | 341 FIELD_PREP(OP_LDF_DST_LMEXTN, dst_lmextn); 342 343 nfp_prog_push(nfp_prog, insn); 344 } 345 346 static void 347 emit_ld_field_any(struct nfp_prog *nfp_prog, swreg dst, u8 bmask, swreg src, 348 enum shf_sc sc, u8 shift, bool zero) 349 { 350 struct nfp_insn_re_regs reg; 351 int err; 352 353 /* Note: ld_field is special as it uses one of the src regs as dst */ 354 err = swreg_to_restricted(dst, dst, src, ®, true); 355 if (err) { 356 nfp_prog->error = err; 357 return; 358 } 359 360 __emit_ld_field(nfp_prog, sc, reg.areg, bmask, reg.breg, shift, 361 reg.i8, zero, reg.swap, reg.wr_both, 362 reg.dst_lmextn, reg.src_lmextn); 363 } 364 365 static void 366 emit_ld_field(struct nfp_prog *nfp_prog, swreg dst, u8 bmask, swreg src, 367 enum shf_sc sc, u8 shift) 368 { 369 emit_ld_field_any(nfp_prog, dst, bmask, src, sc, shift, false); 370 } 371 372 static void 373 __emit_lcsr(struct nfp_prog *nfp_prog, u16 areg, u16 breg, bool wr, u16 addr, 374 bool dst_lmextn, bool src_lmextn) 375 { 376 u64 insn; 377 378 insn = OP_LCSR_BASE | 379 FIELD_PREP(OP_LCSR_A_SRC, areg) | 380 FIELD_PREP(OP_LCSR_B_SRC, breg) | 381 FIELD_PREP(OP_LCSR_WRITE, wr) | 382 FIELD_PREP(OP_LCSR_ADDR, addr) | 383 FIELD_PREP(OP_LCSR_SRC_LMEXTN, src_lmextn) | 384 FIELD_PREP(OP_LCSR_DST_LMEXTN, dst_lmextn); 385 386 nfp_prog_push(nfp_prog, insn); 387 } 388 389 static void emit_csr_wr(struct nfp_prog *nfp_prog, swreg src, u16 addr) 390 { 391 struct nfp_insn_ur_regs reg; 392 int err; 393 394 /* This instruction takes immeds instead of reg_none() for the ignored 395 * operand, but we can't encode 2 immeds in one instr with our normal 396 * swreg infra so if param is an immed, we encode as reg_none() and 397 * copy the immed to both operands. 398 */ 399 if (swreg_type(src) == NN_REG_IMM) { 400 err = swreg_to_unrestricted(reg_none(), src, reg_none(), ®); 401 reg.breg = reg.areg; 402 } else { 403 err = swreg_to_unrestricted(reg_none(), src, reg_imm(0), ®); 404 } 405 if (err) { 406 nfp_prog->error = err; 407 return; 408 } 409 410 __emit_lcsr(nfp_prog, reg.areg, reg.breg, true, addr / 4, 411 false, reg.src_lmextn); 412 } 413 414 static void emit_nop(struct nfp_prog *nfp_prog) 415 { 416 __emit_immed(nfp_prog, UR_REG_IMM, UR_REG_IMM, 0, 0, 0, 0, 0, 0, 0); 417 } 418 419 /* --- Wrappers --- */ 420 static bool pack_immed(u32 imm, u16 *val, enum immed_shift *shift) 421 { 422 if (!(imm & 0xffff0000)) { 423 *val = imm; 424 *shift = IMMED_SHIFT_0B; 425 } else if (!(imm & 0xff0000ff)) { 426 *val = imm >> 8; 427 *shift = IMMED_SHIFT_1B; 428 } else if (!(imm & 0x0000ffff)) { 429 *val = imm >> 16; 430 *shift = IMMED_SHIFT_2B; 431 } else { 432 return false; 433 } 434 435 return true; 436 } 437 438 static void wrp_immed(struct nfp_prog *nfp_prog, swreg dst, u32 imm) 439 { 440 enum immed_shift shift; 441 u16 val; 442 443 if (pack_immed(imm, &val, &shift)) { 444 emit_immed(nfp_prog, dst, val, IMMED_WIDTH_ALL, false, shift); 445 } else if (pack_immed(~imm, &val, &shift)) { 446 emit_immed(nfp_prog, dst, val, IMMED_WIDTH_ALL, true, shift); 447 } else { 448 emit_immed(nfp_prog, dst, imm & 0xffff, IMMED_WIDTH_ALL, 449 false, IMMED_SHIFT_0B); 450 emit_immed(nfp_prog, dst, imm >> 16, IMMED_WIDTH_WORD, 451 false, IMMED_SHIFT_2B); 452 } 453 } 454 455 /* ur_load_imm_any() - encode immediate or use tmp register (unrestricted) 456 * If the @imm is small enough encode it directly in operand and return 457 * otherwise load @imm to a spare register and return its encoding. 458 */ 459 static swreg ur_load_imm_any(struct nfp_prog *nfp_prog, u32 imm, swreg tmp_reg) 460 { 461 if (FIELD_FIT(UR_REG_IMM_MAX, imm)) 462 return reg_imm(imm); 463 464 wrp_immed(nfp_prog, tmp_reg, imm); 465 return tmp_reg; 466 } 467 468 /* re_load_imm_any() - encode immediate or use tmp register (restricted) 469 * If the @imm is small enough encode it directly in operand and return 470 * otherwise load @imm to a spare register and return its encoding. 471 */ 472 static swreg re_load_imm_any(struct nfp_prog *nfp_prog, u32 imm, swreg tmp_reg) 473 { 474 if (FIELD_FIT(RE_REG_IMM_MAX, imm)) 475 return reg_imm(imm); 476 477 wrp_immed(nfp_prog, tmp_reg, imm); 478 return tmp_reg; 479 } 480 481 static void wrp_nops(struct nfp_prog *nfp_prog, unsigned int count) 482 { 483 while (count--) 484 emit_nop(nfp_prog); 485 } 486 487 static void 488 wrp_br_special(struct nfp_prog *nfp_prog, enum br_mask mask, 489 enum br_special special) 490 { 491 emit_br(nfp_prog, mask, 0, 0); 492 493 nfp_prog->prog[nfp_prog->prog_len - 1] |= 494 FIELD_PREP(OP_BR_SPECIAL, special); 495 } 496 497 static void wrp_mov(struct nfp_prog *nfp_prog, swreg dst, swreg src) 498 { 499 emit_alu(nfp_prog, dst, reg_none(), ALU_OP_NONE, src); 500 } 501 502 static void wrp_reg_mov(struct nfp_prog *nfp_prog, u16 dst, u16 src) 503 { 504 wrp_mov(nfp_prog, reg_both(dst), reg_b(src)); 505 } 506 507 static int 508 data_ld(struct nfp_prog *nfp_prog, swreg offset, u8 dst_gpr, int size) 509 { 510 unsigned int i; 511 u16 shift, sz; 512 513 /* We load the value from the address indicated in @offset and then 514 * shift out the data we don't need. Note: this is big endian! 515 */ 516 sz = max(size, 4); 517 shift = size < 4 ? 4 - size : 0; 518 519 emit_cmd(nfp_prog, CMD_TGT_READ8, CMD_MODE_32b, 0, 520 pptr_reg(nfp_prog), offset, sz - 1, true); 521 522 i = 0; 523 if (shift) 524 emit_shf(nfp_prog, reg_both(dst_gpr), reg_none(), SHF_OP_NONE, 525 reg_xfer(0), SHF_SC_R_SHF, shift * 8); 526 else 527 for (; i * 4 < size; i++) 528 wrp_mov(nfp_prog, reg_both(dst_gpr + i), reg_xfer(i)); 529 530 if (i < 2) 531 wrp_immed(nfp_prog, reg_both(dst_gpr + 1), 0); 532 533 return 0; 534 } 535 536 static int 537 data_ld_host_order(struct nfp_prog *nfp_prog, u8 src_gpr, swreg offset, 538 u8 dst_gpr, int size) 539 { 540 unsigned int i; 541 u8 mask, sz; 542 543 /* We load the value from the address indicated in @offset and then 544 * mask out the data we don't need. Note: this is little endian! 545 */ 546 sz = max(size, 4); 547 mask = size < 4 ? GENMASK(size - 1, 0) : 0; 548 549 emit_cmd(nfp_prog, CMD_TGT_READ32_SWAP, CMD_MODE_32b, 0, 550 reg_a(src_gpr), offset, sz / 4 - 1, true); 551 552 i = 0; 553 if (mask) 554 emit_ld_field_any(nfp_prog, reg_both(dst_gpr), mask, 555 reg_xfer(0), SHF_SC_NONE, 0, true); 556 else 557 for (; i * 4 < size; i++) 558 wrp_mov(nfp_prog, reg_both(dst_gpr + i), reg_xfer(i)); 559 560 if (i < 2) 561 wrp_immed(nfp_prog, reg_both(dst_gpr + 1), 0); 562 563 return 0; 564 } 565 566 static int 567 construct_data_ind_ld(struct nfp_prog *nfp_prog, u16 offset, u16 src, u8 size) 568 { 569 swreg tmp_reg; 570 571 /* Calculate the true offset (src_reg + imm) */ 572 tmp_reg = ur_load_imm_any(nfp_prog, offset, imm_b(nfp_prog)); 573 emit_alu(nfp_prog, imm_both(nfp_prog), reg_a(src), ALU_OP_ADD, tmp_reg); 574 575 /* Check packet length (size guaranteed to fit b/c it's u8) */ 576 emit_alu(nfp_prog, imm_a(nfp_prog), 577 imm_a(nfp_prog), ALU_OP_ADD, reg_imm(size)); 578 emit_alu(nfp_prog, reg_none(), 579 plen_reg(nfp_prog), ALU_OP_SUB, imm_a(nfp_prog)); 580 wrp_br_special(nfp_prog, BR_BLO, OP_BR_GO_ABORT); 581 582 /* Load data */ 583 return data_ld(nfp_prog, imm_b(nfp_prog), 0, size); 584 } 585 586 static int construct_data_ld(struct nfp_prog *nfp_prog, u16 offset, u8 size) 587 { 588 swreg tmp_reg; 589 590 /* Check packet length */ 591 tmp_reg = ur_load_imm_any(nfp_prog, offset + size, imm_a(nfp_prog)); 592 emit_alu(nfp_prog, reg_none(), plen_reg(nfp_prog), ALU_OP_SUB, tmp_reg); 593 wrp_br_special(nfp_prog, BR_BLO, OP_BR_GO_ABORT); 594 595 /* Load data */ 596 tmp_reg = re_load_imm_any(nfp_prog, offset, imm_b(nfp_prog)); 597 return data_ld(nfp_prog, tmp_reg, 0, size); 598 } 599 600 static int 601 data_stx_host_order(struct nfp_prog *nfp_prog, u8 dst_gpr, swreg offset, 602 u8 src_gpr, u8 size) 603 { 604 unsigned int i; 605 606 for (i = 0; i * 4 < size; i++) 607 wrp_mov(nfp_prog, reg_xfer(i), reg_a(src_gpr + i)); 608 609 emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 0, 610 reg_a(dst_gpr), offset, size - 1, true); 611 612 return 0; 613 } 614 615 static int 616 data_st_host_order(struct nfp_prog *nfp_prog, u8 dst_gpr, swreg offset, 617 u64 imm, u8 size) 618 { 619 wrp_immed(nfp_prog, reg_xfer(0), imm); 620 if (size == 8) 621 wrp_immed(nfp_prog, reg_xfer(1), imm >> 32); 622 623 emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 0, 624 reg_a(dst_gpr), offset, size - 1, true); 625 626 return 0; 627 } 628 629 typedef int 630 (*lmem_step)(struct nfp_prog *nfp_prog, u8 gpr, u8 gpr_byte, s32 off, 631 unsigned int size, bool first, bool new_gpr, bool last, bool lm3, 632 bool needs_inc); 633 634 static int 635 wrp_lmem_load(struct nfp_prog *nfp_prog, u8 dst, u8 dst_byte, s32 off, 636 unsigned int size, bool first, bool new_gpr, bool last, bool lm3, 637 bool needs_inc) 638 { 639 bool should_inc = needs_inc && new_gpr && !last; 640 u32 idx, src_byte; 641 enum shf_sc sc; 642 swreg reg; 643 int shf; 644 u8 mask; 645 646 if (WARN_ON_ONCE(dst_byte + size > 4 || off % 4 + size > 4)) 647 return -EOPNOTSUPP; 648 649 idx = off / 4; 650 651 /* Move the entire word */ 652 if (size == 4) { 653 wrp_mov(nfp_prog, reg_both(dst), 654 should_inc ? reg_lm_inc(3) : reg_lm(lm3 ? 3 : 0, idx)); 655 return 0; 656 } 657 658 if (WARN_ON_ONCE(lm3 && idx > RE_REG_LM_IDX_MAX)) 659 return -EOPNOTSUPP; 660 661 src_byte = off % 4; 662 663 mask = (1 << size) - 1; 664 mask <<= dst_byte; 665 666 if (WARN_ON_ONCE(mask > 0xf)) 667 return -EOPNOTSUPP; 668 669 shf = abs(src_byte - dst_byte) * 8; 670 if (src_byte == dst_byte) { 671 sc = SHF_SC_NONE; 672 } else if (src_byte < dst_byte) { 673 shf = 32 - shf; 674 sc = SHF_SC_L_SHF; 675 } else { 676 sc = SHF_SC_R_SHF; 677 } 678 679 /* ld_field can address fewer indexes, if offset too large do RMW. 680 * Because we RMV twice we waste 2 cycles on unaligned 8 byte writes. 681 */ 682 if (idx <= RE_REG_LM_IDX_MAX) { 683 reg = reg_lm(lm3 ? 3 : 0, idx); 684 } else { 685 reg = imm_a(nfp_prog); 686 /* If it's not the first part of the load and we start a new GPR 687 * that means we are loading a second part of the LMEM word into 688 * a new GPR. IOW we've already looked that LMEM word and 689 * therefore it has been loaded into imm_a(). 690 */ 691 if (first || !new_gpr) 692 wrp_mov(nfp_prog, reg, reg_lm(0, idx)); 693 } 694 695 emit_ld_field_any(nfp_prog, reg_both(dst), mask, reg, sc, shf, new_gpr); 696 697 if (should_inc) 698 wrp_mov(nfp_prog, reg_none(), reg_lm_inc(3)); 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 first, bool new_gpr, bool last, bool lm3, 706 bool needs_inc) 707 { 708 bool should_inc = needs_inc && new_gpr && !last; 709 u32 idx, dst_byte; 710 enum shf_sc sc; 711 swreg reg; 712 int shf; 713 u8 mask; 714 715 if (WARN_ON_ONCE(src_byte + size > 4 || off % 4 + size > 4)) 716 return -EOPNOTSUPP; 717 718 idx = off / 4; 719 720 /* Move the entire word */ 721 if (size == 4) { 722 wrp_mov(nfp_prog, 723 should_inc ? reg_lm_inc(3) : reg_lm(lm3 ? 3 : 0, idx), 724 reg_b(src)); 725 return 0; 726 } 727 728 if (WARN_ON_ONCE(lm3 && idx > RE_REG_LM_IDX_MAX)) 729 return -EOPNOTSUPP; 730 731 dst_byte = off % 4; 732 733 mask = (1 << size) - 1; 734 mask <<= dst_byte; 735 736 if (WARN_ON_ONCE(mask > 0xf)) 737 return -EOPNOTSUPP; 738 739 shf = abs(src_byte - dst_byte) * 8; 740 if (src_byte == dst_byte) { 741 sc = SHF_SC_NONE; 742 } else if (src_byte < dst_byte) { 743 shf = 32 - shf; 744 sc = SHF_SC_L_SHF; 745 } else { 746 sc = SHF_SC_R_SHF; 747 } 748 749 /* ld_field can address fewer indexes, if offset too large do RMW. 750 * Because we RMV twice we waste 2 cycles on unaligned 8 byte writes. 751 */ 752 if (idx <= RE_REG_LM_IDX_MAX) { 753 reg = reg_lm(lm3 ? 3 : 0, idx); 754 } else { 755 reg = imm_a(nfp_prog); 756 /* Only first and last LMEM locations are going to need RMW, 757 * the middle location will be overwritten fully. 758 */ 759 if (first || last) 760 wrp_mov(nfp_prog, reg, reg_lm(0, idx)); 761 } 762 763 emit_ld_field(nfp_prog, reg, mask, reg_b(src), sc, shf); 764 765 if (new_gpr || last) { 766 if (idx > RE_REG_LM_IDX_MAX) 767 wrp_mov(nfp_prog, reg_lm(0, idx), reg); 768 if (should_inc) 769 wrp_mov(nfp_prog, reg_none(), reg_lm_inc(3)); 770 } 771 772 return 0; 773 } 774 775 static int 776 mem_op_stack(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 777 unsigned int size, unsigned int ptr_off, u8 gpr, u8 ptr_gpr, 778 bool clr_gpr, lmem_step step) 779 { 780 s32 off = nfp_prog->stack_depth + meta->insn.off + ptr_off; 781 bool first = true, last; 782 bool needs_inc = false; 783 swreg stack_off_reg; 784 u8 prev_gpr = 255; 785 u32 gpr_byte = 0; 786 bool lm3 = true; 787 int ret; 788 789 if (meta->ptr_not_const) { 790 /* Use of the last encountered ptr_off is OK, they all have 791 * the same alignment. Depend on low bits of value being 792 * discarded when written to LMaddr register. 793 */ 794 stack_off_reg = ur_load_imm_any(nfp_prog, meta->insn.off, 795 stack_imm(nfp_prog)); 796 797 emit_alu(nfp_prog, imm_b(nfp_prog), 798 reg_a(ptr_gpr), ALU_OP_ADD, stack_off_reg); 799 800 needs_inc = true; 801 } else if (off + size <= 64) { 802 /* We can reach bottom 64B with LMaddr0 */ 803 lm3 = false; 804 } else if (round_down(off, 32) == round_down(off + size - 1, 32)) { 805 /* We have to set up a new pointer. If we know the offset 806 * and the entire access falls into a single 32 byte aligned 807 * window we won't have to increment the LM pointer. 808 * The 32 byte alignment is imporant because offset is ORed in 809 * not added when doing *l$indexN[off]. 810 */ 811 stack_off_reg = ur_load_imm_any(nfp_prog, round_down(off, 32), 812 stack_imm(nfp_prog)); 813 emit_alu(nfp_prog, imm_b(nfp_prog), 814 stack_reg(nfp_prog), ALU_OP_ADD, stack_off_reg); 815 816 off %= 32; 817 } else { 818 stack_off_reg = ur_load_imm_any(nfp_prog, round_down(off, 4), 819 stack_imm(nfp_prog)); 820 821 emit_alu(nfp_prog, imm_b(nfp_prog), 822 stack_reg(nfp_prog), ALU_OP_ADD, stack_off_reg); 823 824 needs_inc = true; 825 } 826 if (lm3) { 827 emit_csr_wr(nfp_prog, imm_b(nfp_prog), NFP_CSR_ACT_LM_ADDR3); 828 /* For size < 4 one slot will be filled by zeroing of upper. */ 829 wrp_nops(nfp_prog, clr_gpr && size < 8 ? 2 : 3); 830 } 831 832 if (clr_gpr && size < 8) 833 wrp_immed(nfp_prog, reg_both(gpr + 1), 0); 834 835 while (size) { 836 u32 slice_end; 837 u8 slice_size; 838 839 slice_size = min(size, 4 - gpr_byte); 840 slice_end = min(off + slice_size, round_up(off + 1, 4)); 841 slice_size = slice_end - off; 842 843 last = slice_size == size; 844 845 if (needs_inc) 846 off %= 4; 847 848 ret = step(nfp_prog, gpr, gpr_byte, off, slice_size, 849 first, gpr != prev_gpr, last, lm3, needs_inc); 850 if (ret) 851 return ret; 852 853 prev_gpr = gpr; 854 first = false; 855 856 gpr_byte += slice_size; 857 if (gpr_byte >= 4) { 858 gpr_byte -= 4; 859 gpr++; 860 } 861 862 size -= slice_size; 863 off += slice_size; 864 } 865 866 return 0; 867 } 868 869 static void 870 wrp_alu_imm(struct nfp_prog *nfp_prog, u8 dst, enum alu_op alu_op, u32 imm) 871 { 872 swreg tmp_reg; 873 874 if (alu_op == ALU_OP_AND) { 875 if (!imm) 876 wrp_immed(nfp_prog, reg_both(dst), 0); 877 if (!imm || !~imm) 878 return; 879 } 880 if (alu_op == ALU_OP_OR) { 881 if (!~imm) 882 wrp_immed(nfp_prog, reg_both(dst), ~0U); 883 if (!imm || !~imm) 884 return; 885 } 886 if (alu_op == ALU_OP_XOR) { 887 if (!~imm) 888 emit_alu(nfp_prog, reg_both(dst), reg_none(), 889 ALU_OP_NOT, reg_b(dst)); 890 if (!imm || !~imm) 891 return; 892 } 893 894 tmp_reg = ur_load_imm_any(nfp_prog, imm, imm_b(nfp_prog)); 895 emit_alu(nfp_prog, reg_both(dst), reg_a(dst), alu_op, tmp_reg); 896 } 897 898 static int 899 wrp_alu64_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 900 enum alu_op alu_op, bool skip) 901 { 902 const struct bpf_insn *insn = &meta->insn; 903 u64 imm = insn->imm; /* sign extend */ 904 905 if (skip) { 906 meta->skip = true; 907 return 0; 908 } 909 910 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, alu_op, imm & ~0U); 911 wrp_alu_imm(nfp_prog, insn->dst_reg * 2 + 1, alu_op, imm >> 32); 912 913 return 0; 914 } 915 916 static int 917 wrp_alu64_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 918 enum alu_op alu_op) 919 { 920 u8 dst = meta->insn.dst_reg * 2, src = meta->insn.src_reg * 2; 921 922 emit_alu(nfp_prog, reg_both(dst), reg_a(dst), alu_op, reg_b(src)); 923 emit_alu(nfp_prog, reg_both(dst + 1), 924 reg_a(dst + 1), alu_op, reg_b(src + 1)); 925 926 return 0; 927 } 928 929 static int 930 wrp_alu32_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 931 enum alu_op alu_op, bool skip) 932 { 933 const struct bpf_insn *insn = &meta->insn; 934 935 if (skip) { 936 meta->skip = true; 937 return 0; 938 } 939 940 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, alu_op, insn->imm); 941 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0); 942 943 return 0; 944 } 945 946 static int 947 wrp_alu32_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 948 enum alu_op alu_op) 949 { 950 u8 dst = meta->insn.dst_reg * 2, src = meta->insn.src_reg * 2; 951 952 emit_alu(nfp_prog, reg_both(dst), reg_a(dst), alu_op, reg_b(src)); 953 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0); 954 955 return 0; 956 } 957 958 static void 959 wrp_test_reg_one(struct nfp_prog *nfp_prog, u8 dst, enum alu_op alu_op, u8 src, 960 enum br_mask br_mask, u16 off) 961 { 962 emit_alu(nfp_prog, reg_none(), reg_a(dst), alu_op, reg_b(src)); 963 emit_br(nfp_prog, br_mask, off, 0); 964 } 965 966 static int 967 wrp_test_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 968 enum alu_op alu_op, enum br_mask br_mask) 969 { 970 const struct bpf_insn *insn = &meta->insn; 971 972 wrp_test_reg_one(nfp_prog, insn->dst_reg * 2, alu_op, 973 insn->src_reg * 2, br_mask, insn->off); 974 wrp_test_reg_one(nfp_prog, insn->dst_reg * 2 + 1, alu_op, 975 insn->src_reg * 2 + 1, br_mask, insn->off); 976 977 return 0; 978 } 979 980 static int 981 wrp_cmp_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 982 enum br_mask br_mask, bool swap) 983 { 984 const struct bpf_insn *insn = &meta->insn; 985 u64 imm = insn->imm; /* sign extend */ 986 u8 reg = insn->dst_reg * 2; 987 swreg tmp_reg; 988 989 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog)); 990 if (!swap) 991 emit_alu(nfp_prog, reg_none(), reg_a(reg), ALU_OP_SUB, tmp_reg); 992 else 993 emit_alu(nfp_prog, reg_none(), tmp_reg, ALU_OP_SUB, reg_a(reg)); 994 995 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog)); 996 if (!swap) 997 emit_alu(nfp_prog, reg_none(), 998 reg_a(reg + 1), ALU_OP_SUB_C, tmp_reg); 999 else 1000 emit_alu(nfp_prog, reg_none(), 1001 tmp_reg, ALU_OP_SUB_C, reg_a(reg + 1)); 1002 1003 emit_br(nfp_prog, br_mask, insn->off, 0); 1004 1005 return 0; 1006 } 1007 1008 static int 1009 wrp_cmp_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1010 enum br_mask br_mask, bool swap) 1011 { 1012 const struct bpf_insn *insn = &meta->insn; 1013 u8 areg, breg; 1014 1015 areg = insn->dst_reg * 2; 1016 breg = insn->src_reg * 2; 1017 1018 if (swap) { 1019 areg ^= breg; 1020 breg ^= areg; 1021 areg ^= breg; 1022 } 1023 1024 emit_alu(nfp_prog, reg_none(), reg_a(areg), ALU_OP_SUB, reg_b(breg)); 1025 emit_alu(nfp_prog, reg_none(), 1026 reg_a(areg + 1), ALU_OP_SUB_C, reg_b(breg + 1)); 1027 emit_br(nfp_prog, br_mask, insn->off, 0); 1028 1029 return 0; 1030 } 1031 1032 static void wrp_end32(struct nfp_prog *nfp_prog, swreg reg_in, u8 gpr_out) 1033 { 1034 emit_ld_field(nfp_prog, reg_both(gpr_out), 0xf, reg_in, 1035 SHF_SC_R_ROT, 8); 1036 emit_ld_field(nfp_prog, reg_both(gpr_out), 0x5, reg_a(gpr_out), 1037 SHF_SC_R_ROT, 16); 1038 } 1039 1040 /* --- Callbacks --- */ 1041 static int mov_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1042 { 1043 const struct bpf_insn *insn = &meta->insn; 1044 u8 dst = insn->dst_reg * 2; 1045 u8 src = insn->src_reg * 2; 1046 1047 if (insn->src_reg == BPF_REG_10) { 1048 swreg stack_depth_reg; 1049 1050 stack_depth_reg = ur_load_imm_any(nfp_prog, 1051 nfp_prog->stack_depth, 1052 stack_imm(nfp_prog)); 1053 emit_alu(nfp_prog, reg_both(dst), 1054 stack_reg(nfp_prog), ALU_OP_ADD, stack_depth_reg); 1055 wrp_immed(nfp_prog, reg_both(dst + 1), 0); 1056 } else { 1057 wrp_reg_mov(nfp_prog, dst, src); 1058 wrp_reg_mov(nfp_prog, dst + 1, src + 1); 1059 } 1060 1061 return 0; 1062 } 1063 1064 static int mov_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1065 { 1066 u64 imm = meta->insn.imm; /* sign extend */ 1067 1068 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2), imm & ~0U); 1069 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), imm >> 32); 1070 1071 return 0; 1072 } 1073 1074 static int xor_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1075 { 1076 return wrp_alu64_reg(nfp_prog, meta, ALU_OP_XOR); 1077 } 1078 1079 static int xor_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1080 { 1081 return wrp_alu64_imm(nfp_prog, meta, ALU_OP_XOR, !meta->insn.imm); 1082 } 1083 1084 static int and_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1085 { 1086 return wrp_alu64_reg(nfp_prog, meta, ALU_OP_AND); 1087 } 1088 1089 static int and_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1090 { 1091 return wrp_alu64_imm(nfp_prog, meta, ALU_OP_AND, !~meta->insn.imm); 1092 } 1093 1094 static int or_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1095 { 1096 return wrp_alu64_reg(nfp_prog, meta, ALU_OP_OR); 1097 } 1098 1099 static int or_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1100 { 1101 return wrp_alu64_imm(nfp_prog, meta, ALU_OP_OR, !meta->insn.imm); 1102 } 1103 1104 static int add_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1105 { 1106 const struct bpf_insn *insn = &meta->insn; 1107 1108 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2), 1109 reg_a(insn->dst_reg * 2), ALU_OP_ADD, 1110 reg_b(insn->src_reg * 2)); 1111 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 1112 reg_a(insn->dst_reg * 2 + 1), ALU_OP_ADD_C, 1113 reg_b(insn->src_reg * 2 + 1)); 1114 1115 return 0; 1116 } 1117 1118 static int add_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1119 { 1120 const struct bpf_insn *insn = &meta->insn; 1121 u64 imm = insn->imm; /* sign extend */ 1122 1123 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, ALU_OP_ADD, imm & ~0U); 1124 wrp_alu_imm(nfp_prog, insn->dst_reg * 2 + 1, ALU_OP_ADD_C, imm >> 32); 1125 1126 return 0; 1127 } 1128 1129 static int sub_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1130 { 1131 const struct bpf_insn *insn = &meta->insn; 1132 1133 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2), 1134 reg_a(insn->dst_reg * 2), ALU_OP_SUB, 1135 reg_b(insn->src_reg * 2)); 1136 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 1137 reg_a(insn->dst_reg * 2 + 1), ALU_OP_SUB_C, 1138 reg_b(insn->src_reg * 2 + 1)); 1139 1140 return 0; 1141 } 1142 1143 static int sub_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1144 { 1145 const struct bpf_insn *insn = &meta->insn; 1146 u64 imm = insn->imm; /* sign extend */ 1147 1148 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, ALU_OP_SUB, imm & ~0U); 1149 wrp_alu_imm(nfp_prog, insn->dst_reg * 2 + 1, ALU_OP_SUB_C, imm >> 32); 1150 1151 return 0; 1152 } 1153 1154 static int neg_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1155 { 1156 const struct bpf_insn *insn = &meta->insn; 1157 1158 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2), reg_imm(0), 1159 ALU_OP_SUB, reg_b(insn->dst_reg * 2)); 1160 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2 + 1), reg_imm(0), 1161 ALU_OP_SUB_C, reg_b(insn->dst_reg * 2 + 1)); 1162 1163 return 0; 1164 } 1165 1166 static int shl_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1167 { 1168 const struct bpf_insn *insn = &meta->insn; 1169 u8 dst = insn->dst_reg * 2; 1170 1171 if (insn->imm < 32) { 1172 emit_shf(nfp_prog, reg_both(dst + 1), 1173 reg_a(dst + 1), SHF_OP_NONE, reg_b(dst), 1174 SHF_SC_R_DSHF, 32 - insn->imm); 1175 emit_shf(nfp_prog, reg_both(dst), 1176 reg_none(), SHF_OP_NONE, reg_b(dst), 1177 SHF_SC_L_SHF, insn->imm); 1178 } else if (insn->imm == 32) { 1179 wrp_reg_mov(nfp_prog, dst + 1, dst); 1180 wrp_immed(nfp_prog, reg_both(dst), 0); 1181 } else if (insn->imm > 32) { 1182 emit_shf(nfp_prog, reg_both(dst + 1), 1183 reg_none(), SHF_OP_NONE, reg_b(dst), 1184 SHF_SC_L_SHF, insn->imm - 32); 1185 wrp_immed(nfp_prog, reg_both(dst), 0); 1186 } 1187 1188 return 0; 1189 } 1190 1191 static int shr_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1192 { 1193 const struct bpf_insn *insn = &meta->insn; 1194 u8 dst = insn->dst_reg * 2; 1195 1196 if (insn->imm < 32) { 1197 emit_shf(nfp_prog, reg_both(dst), 1198 reg_a(dst + 1), SHF_OP_NONE, reg_b(dst), 1199 SHF_SC_R_DSHF, insn->imm); 1200 emit_shf(nfp_prog, reg_both(dst + 1), 1201 reg_none(), SHF_OP_NONE, reg_b(dst + 1), 1202 SHF_SC_R_SHF, insn->imm); 1203 } else if (insn->imm == 32) { 1204 wrp_reg_mov(nfp_prog, dst, dst + 1); 1205 wrp_immed(nfp_prog, reg_both(dst + 1), 0); 1206 } else if (insn->imm > 32) { 1207 emit_shf(nfp_prog, reg_both(dst), 1208 reg_none(), SHF_OP_NONE, reg_b(dst + 1), 1209 SHF_SC_R_SHF, insn->imm - 32); 1210 wrp_immed(nfp_prog, reg_both(dst + 1), 0); 1211 } 1212 1213 return 0; 1214 } 1215 1216 static int mov_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1217 { 1218 const struct bpf_insn *insn = &meta->insn; 1219 1220 wrp_reg_mov(nfp_prog, insn->dst_reg * 2, insn->src_reg * 2); 1221 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0); 1222 1223 return 0; 1224 } 1225 1226 static int mov_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1227 { 1228 const struct bpf_insn *insn = &meta->insn; 1229 1230 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2), insn->imm); 1231 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0); 1232 1233 return 0; 1234 } 1235 1236 static int xor_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1237 { 1238 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_XOR); 1239 } 1240 1241 static int xor_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1242 { 1243 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_XOR, !~meta->insn.imm); 1244 } 1245 1246 static int and_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1247 { 1248 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_AND); 1249 } 1250 1251 static int and_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1252 { 1253 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_AND, !~meta->insn.imm); 1254 } 1255 1256 static int or_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1257 { 1258 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_OR); 1259 } 1260 1261 static int or_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1262 { 1263 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_OR, !meta->insn.imm); 1264 } 1265 1266 static int add_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1267 { 1268 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_ADD); 1269 } 1270 1271 static int add_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1272 { 1273 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_ADD, !meta->insn.imm); 1274 } 1275 1276 static int sub_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1277 { 1278 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_SUB); 1279 } 1280 1281 static int sub_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1282 { 1283 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_SUB, !meta->insn.imm); 1284 } 1285 1286 static int neg_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1287 { 1288 u8 dst = meta->insn.dst_reg * 2; 1289 1290 emit_alu(nfp_prog, reg_both(dst), reg_imm(0), ALU_OP_SUB, reg_b(dst)); 1291 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0); 1292 1293 return 0; 1294 } 1295 1296 static int shl_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1297 { 1298 const struct bpf_insn *insn = &meta->insn; 1299 1300 if (!insn->imm) 1301 return 1; /* TODO: zero shift means indirect */ 1302 1303 emit_shf(nfp_prog, reg_both(insn->dst_reg * 2), 1304 reg_none(), SHF_OP_NONE, reg_b(insn->dst_reg * 2), 1305 SHF_SC_L_SHF, insn->imm); 1306 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0); 1307 1308 return 0; 1309 } 1310 1311 static int end_reg32(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1312 { 1313 const struct bpf_insn *insn = &meta->insn; 1314 u8 gpr = insn->dst_reg * 2; 1315 1316 switch (insn->imm) { 1317 case 16: 1318 emit_ld_field(nfp_prog, reg_both(gpr), 0x9, reg_b(gpr), 1319 SHF_SC_R_ROT, 8); 1320 emit_ld_field(nfp_prog, reg_both(gpr), 0xe, reg_a(gpr), 1321 SHF_SC_R_SHF, 16); 1322 1323 wrp_immed(nfp_prog, reg_both(gpr + 1), 0); 1324 break; 1325 case 32: 1326 wrp_end32(nfp_prog, reg_a(gpr), gpr); 1327 wrp_immed(nfp_prog, reg_both(gpr + 1), 0); 1328 break; 1329 case 64: 1330 wrp_mov(nfp_prog, imm_a(nfp_prog), reg_b(gpr + 1)); 1331 1332 wrp_end32(nfp_prog, reg_a(gpr), gpr + 1); 1333 wrp_end32(nfp_prog, imm_a(nfp_prog), gpr); 1334 break; 1335 } 1336 1337 return 0; 1338 } 1339 1340 static int imm_ld8_part2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1341 { 1342 struct nfp_insn_meta *prev = nfp_meta_prev(meta); 1343 u32 imm_lo, imm_hi; 1344 u8 dst; 1345 1346 dst = prev->insn.dst_reg * 2; 1347 imm_lo = prev->insn.imm; 1348 imm_hi = meta->insn.imm; 1349 1350 wrp_immed(nfp_prog, reg_both(dst), imm_lo); 1351 1352 /* mov is always 1 insn, load imm may be two, so try to use mov */ 1353 if (imm_hi == imm_lo) 1354 wrp_mov(nfp_prog, reg_both(dst + 1), reg_a(dst)); 1355 else 1356 wrp_immed(nfp_prog, reg_both(dst + 1), imm_hi); 1357 1358 return 0; 1359 } 1360 1361 static int imm_ld8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1362 { 1363 meta->double_cb = imm_ld8_part2; 1364 return 0; 1365 } 1366 1367 static int data_ld1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1368 { 1369 return construct_data_ld(nfp_prog, meta->insn.imm, 1); 1370 } 1371 1372 static int data_ld2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1373 { 1374 return construct_data_ld(nfp_prog, meta->insn.imm, 2); 1375 } 1376 1377 static int data_ld4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1378 { 1379 return construct_data_ld(nfp_prog, meta->insn.imm, 4); 1380 } 1381 1382 static int data_ind_ld1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1383 { 1384 return construct_data_ind_ld(nfp_prog, meta->insn.imm, 1385 meta->insn.src_reg * 2, 1); 1386 } 1387 1388 static int data_ind_ld2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1389 { 1390 return construct_data_ind_ld(nfp_prog, meta->insn.imm, 1391 meta->insn.src_reg * 2, 2); 1392 } 1393 1394 static int data_ind_ld4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1395 { 1396 return construct_data_ind_ld(nfp_prog, meta->insn.imm, 1397 meta->insn.src_reg * 2, 4); 1398 } 1399 1400 static int 1401 mem_ldx_stack(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1402 unsigned int size, unsigned int ptr_off) 1403 { 1404 return mem_op_stack(nfp_prog, meta, size, ptr_off, 1405 meta->insn.dst_reg * 2, meta->insn.src_reg * 2, 1406 true, wrp_lmem_load); 1407 } 1408 1409 static int mem_ldx_skb(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1410 u8 size) 1411 { 1412 swreg dst = reg_both(meta->insn.dst_reg * 2); 1413 1414 switch (meta->insn.off) { 1415 case offsetof(struct __sk_buff, len): 1416 if (size != FIELD_SIZEOF(struct __sk_buff, len)) 1417 return -EOPNOTSUPP; 1418 wrp_mov(nfp_prog, dst, plen_reg(nfp_prog)); 1419 break; 1420 case offsetof(struct __sk_buff, data): 1421 if (size != FIELD_SIZEOF(struct __sk_buff, data)) 1422 return -EOPNOTSUPP; 1423 wrp_mov(nfp_prog, dst, pptr_reg(nfp_prog)); 1424 break; 1425 case offsetof(struct __sk_buff, data_end): 1426 if (size != FIELD_SIZEOF(struct __sk_buff, data_end)) 1427 return -EOPNOTSUPP; 1428 emit_alu(nfp_prog, dst, 1429 plen_reg(nfp_prog), ALU_OP_ADD, pptr_reg(nfp_prog)); 1430 break; 1431 default: 1432 return -EOPNOTSUPP; 1433 } 1434 1435 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0); 1436 1437 return 0; 1438 } 1439 1440 static int mem_ldx_xdp(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1441 u8 size) 1442 { 1443 swreg dst = reg_both(meta->insn.dst_reg * 2); 1444 1445 switch (meta->insn.off) { 1446 case offsetof(struct xdp_md, data): 1447 if (size != FIELD_SIZEOF(struct xdp_md, data)) 1448 return -EOPNOTSUPP; 1449 wrp_mov(nfp_prog, dst, pptr_reg(nfp_prog)); 1450 break; 1451 case offsetof(struct xdp_md, data_end): 1452 if (size != FIELD_SIZEOF(struct xdp_md, data_end)) 1453 return -EOPNOTSUPP; 1454 emit_alu(nfp_prog, dst, 1455 plen_reg(nfp_prog), ALU_OP_ADD, pptr_reg(nfp_prog)); 1456 break; 1457 default: 1458 return -EOPNOTSUPP; 1459 } 1460 1461 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0); 1462 1463 return 0; 1464 } 1465 1466 static int 1467 mem_ldx_data(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1468 unsigned int size) 1469 { 1470 swreg tmp_reg; 1471 1472 tmp_reg = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog)); 1473 1474 return data_ld_host_order(nfp_prog, meta->insn.src_reg * 2, tmp_reg, 1475 meta->insn.dst_reg * 2, size); 1476 } 1477 1478 static int 1479 mem_ldx(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1480 unsigned int size) 1481 { 1482 if (meta->ptr.type == PTR_TO_CTX) { 1483 if (nfp_prog->type == BPF_PROG_TYPE_XDP) 1484 return mem_ldx_xdp(nfp_prog, meta, size); 1485 else 1486 return mem_ldx_skb(nfp_prog, meta, size); 1487 } 1488 1489 if (meta->ptr.type == PTR_TO_PACKET) 1490 return mem_ldx_data(nfp_prog, meta, size); 1491 1492 if (meta->ptr.type == PTR_TO_STACK) 1493 return mem_ldx_stack(nfp_prog, meta, size, 1494 meta->ptr.off + meta->ptr.var_off.value); 1495 1496 return -EOPNOTSUPP; 1497 } 1498 1499 static int mem_ldx1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1500 { 1501 return mem_ldx(nfp_prog, meta, 1); 1502 } 1503 1504 static int mem_ldx2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1505 { 1506 return mem_ldx(nfp_prog, meta, 2); 1507 } 1508 1509 static int mem_ldx4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1510 { 1511 return mem_ldx(nfp_prog, meta, 4); 1512 } 1513 1514 static int mem_ldx8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1515 { 1516 return mem_ldx(nfp_prog, meta, 8); 1517 } 1518 1519 static int 1520 mem_st_data(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1521 unsigned int size) 1522 { 1523 u64 imm = meta->insn.imm; /* sign extend */ 1524 swreg off_reg; 1525 1526 off_reg = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog)); 1527 1528 return data_st_host_order(nfp_prog, meta->insn.dst_reg * 2, off_reg, 1529 imm, size); 1530 } 1531 1532 static int mem_st(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1533 unsigned int size) 1534 { 1535 if (meta->ptr.type == PTR_TO_PACKET) 1536 return mem_st_data(nfp_prog, meta, size); 1537 1538 return -EOPNOTSUPP; 1539 } 1540 1541 static int mem_st1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1542 { 1543 return mem_st(nfp_prog, meta, 1); 1544 } 1545 1546 static int mem_st2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1547 { 1548 return mem_st(nfp_prog, meta, 2); 1549 } 1550 1551 static int mem_st4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1552 { 1553 return mem_st(nfp_prog, meta, 4); 1554 } 1555 1556 static int mem_st8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1557 { 1558 return mem_st(nfp_prog, meta, 8); 1559 } 1560 1561 static int 1562 mem_stx_data(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1563 unsigned int size) 1564 { 1565 swreg off_reg; 1566 1567 off_reg = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog)); 1568 1569 return data_stx_host_order(nfp_prog, meta->insn.dst_reg * 2, off_reg, 1570 meta->insn.src_reg * 2, size); 1571 } 1572 1573 static int 1574 mem_stx_stack(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1575 unsigned int size, unsigned int ptr_off) 1576 { 1577 return mem_op_stack(nfp_prog, meta, size, ptr_off, 1578 meta->insn.src_reg * 2, meta->insn.dst_reg * 2, 1579 false, wrp_lmem_store); 1580 } 1581 1582 static int 1583 mem_stx(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1584 unsigned int size) 1585 { 1586 if (meta->ptr.type == PTR_TO_PACKET) 1587 return mem_stx_data(nfp_prog, meta, size); 1588 1589 if (meta->ptr.type == PTR_TO_STACK) 1590 return mem_stx_stack(nfp_prog, meta, size, 1591 meta->ptr.off + meta->ptr.var_off.value); 1592 1593 return -EOPNOTSUPP; 1594 } 1595 1596 static int mem_stx1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1597 { 1598 return mem_stx(nfp_prog, meta, 1); 1599 } 1600 1601 static int mem_stx2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1602 { 1603 return mem_stx(nfp_prog, meta, 2); 1604 } 1605 1606 static int mem_stx4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1607 { 1608 return mem_stx(nfp_prog, meta, 4); 1609 } 1610 1611 static int mem_stx8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1612 { 1613 return mem_stx(nfp_prog, meta, 8); 1614 } 1615 1616 static int jump(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1617 { 1618 emit_br(nfp_prog, BR_UNC, meta->insn.off, 0); 1619 1620 return 0; 1621 } 1622 1623 static int jeq_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1624 { 1625 const struct bpf_insn *insn = &meta->insn; 1626 u64 imm = insn->imm; /* sign extend */ 1627 swreg or1, or2, tmp_reg; 1628 1629 or1 = reg_a(insn->dst_reg * 2); 1630 or2 = reg_b(insn->dst_reg * 2 + 1); 1631 1632 if (imm & ~0U) { 1633 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog)); 1634 emit_alu(nfp_prog, imm_a(nfp_prog), 1635 reg_a(insn->dst_reg * 2), ALU_OP_XOR, tmp_reg); 1636 or1 = imm_a(nfp_prog); 1637 } 1638 1639 if (imm >> 32) { 1640 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog)); 1641 emit_alu(nfp_prog, imm_b(nfp_prog), 1642 reg_a(insn->dst_reg * 2 + 1), ALU_OP_XOR, tmp_reg); 1643 or2 = imm_b(nfp_prog); 1644 } 1645 1646 emit_alu(nfp_prog, reg_none(), or1, ALU_OP_OR, or2); 1647 emit_br(nfp_prog, BR_BEQ, insn->off, 0); 1648 1649 return 0; 1650 } 1651 1652 static int jgt_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1653 { 1654 return wrp_cmp_imm(nfp_prog, meta, BR_BLO, true); 1655 } 1656 1657 static int jge_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1658 { 1659 return wrp_cmp_imm(nfp_prog, meta, BR_BHS, false); 1660 } 1661 1662 static int jlt_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1663 { 1664 return wrp_cmp_imm(nfp_prog, meta, BR_BLO, false); 1665 } 1666 1667 static int jle_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1668 { 1669 return wrp_cmp_imm(nfp_prog, meta, BR_BHS, true); 1670 } 1671 1672 static int jset_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1673 { 1674 const struct bpf_insn *insn = &meta->insn; 1675 u64 imm = insn->imm; /* sign extend */ 1676 swreg tmp_reg; 1677 1678 if (!imm) { 1679 meta->skip = true; 1680 return 0; 1681 } 1682 1683 if (imm & ~0U) { 1684 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog)); 1685 emit_alu(nfp_prog, reg_none(), 1686 reg_a(insn->dst_reg * 2), ALU_OP_AND, tmp_reg); 1687 emit_br(nfp_prog, BR_BNE, insn->off, 0); 1688 } 1689 1690 if (imm >> 32) { 1691 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog)); 1692 emit_alu(nfp_prog, reg_none(), 1693 reg_a(insn->dst_reg * 2 + 1), ALU_OP_AND, tmp_reg); 1694 emit_br(nfp_prog, BR_BNE, insn->off, 0); 1695 } 1696 1697 return 0; 1698 } 1699 1700 static int jne_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1701 { 1702 const struct bpf_insn *insn = &meta->insn; 1703 u64 imm = insn->imm; /* sign extend */ 1704 swreg tmp_reg; 1705 1706 if (!imm) { 1707 emit_alu(nfp_prog, reg_none(), reg_a(insn->dst_reg * 2), 1708 ALU_OP_OR, reg_b(insn->dst_reg * 2 + 1)); 1709 emit_br(nfp_prog, BR_BNE, insn->off, 0); 1710 return 0; 1711 } 1712 1713 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog)); 1714 emit_alu(nfp_prog, reg_none(), 1715 reg_a(insn->dst_reg * 2), ALU_OP_XOR, tmp_reg); 1716 emit_br(nfp_prog, BR_BNE, insn->off, 0); 1717 1718 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog)); 1719 emit_alu(nfp_prog, reg_none(), 1720 reg_a(insn->dst_reg * 2 + 1), ALU_OP_XOR, tmp_reg); 1721 emit_br(nfp_prog, BR_BNE, insn->off, 0); 1722 1723 return 0; 1724 } 1725 1726 static int jeq_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1727 { 1728 const struct bpf_insn *insn = &meta->insn; 1729 1730 emit_alu(nfp_prog, imm_a(nfp_prog), reg_a(insn->dst_reg * 2), 1731 ALU_OP_XOR, reg_b(insn->src_reg * 2)); 1732 emit_alu(nfp_prog, imm_b(nfp_prog), reg_a(insn->dst_reg * 2 + 1), 1733 ALU_OP_XOR, reg_b(insn->src_reg * 2 + 1)); 1734 emit_alu(nfp_prog, reg_none(), 1735 imm_a(nfp_prog), ALU_OP_OR, imm_b(nfp_prog)); 1736 emit_br(nfp_prog, BR_BEQ, insn->off, 0); 1737 1738 return 0; 1739 } 1740 1741 static int jgt_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1742 { 1743 return wrp_cmp_reg(nfp_prog, meta, BR_BLO, true); 1744 } 1745 1746 static int jge_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1747 { 1748 return wrp_cmp_reg(nfp_prog, meta, BR_BHS, false); 1749 } 1750 1751 static int jlt_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1752 { 1753 return wrp_cmp_reg(nfp_prog, meta, BR_BLO, false); 1754 } 1755 1756 static int jle_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1757 { 1758 return wrp_cmp_reg(nfp_prog, meta, BR_BHS, true); 1759 } 1760 1761 static int jset_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1762 { 1763 return wrp_test_reg(nfp_prog, meta, ALU_OP_AND, BR_BNE); 1764 } 1765 1766 static int jne_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1767 { 1768 return wrp_test_reg(nfp_prog, meta, ALU_OP_XOR, BR_BNE); 1769 } 1770 1771 static int goto_out(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1772 { 1773 wrp_br_special(nfp_prog, BR_UNC, OP_BR_GO_OUT); 1774 1775 return 0; 1776 } 1777 1778 static const instr_cb_t instr_cb[256] = { 1779 [BPF_ALU64 | BPF_MOV | BPF_X] = mov_reg64, 1780 [BPF_ALU64 | BPF_MOV | BPF_K] = mov_imm64, 1781 [BPF_ALU64 | BPF_XOR | BPF_X] = xor_reg64, 1782 [BPF_ALU64 | BPF_XOR | BPF_K] = xor_imm64, 1783 [BPF_ALU64 | BPF_AND | BPF_X] = and_reg64, 1784 [BPF_ALU64 | BPF_AND | BPF_K] = and_imm64, 1785 [BPF_ALU64 | BPF_OR | BPF_X] = or_reg64, 1786 [BPF_ALU64 | BPF_OR | BPF_K] = or_imm64, 1787 [BPF_ALU64 | BPF_ADD | BPF_X] = add_reg64, 1788 [BPF_ALU64 | BPF_ADD | BPF_K] = add_imm64, 1789 [BPF_ALU64 | BPF_SUB | BPF_X] = sub_reg64, 1790 [BPF_ALU64 | BPF_SUB | BPF_K] = sub_imm64, 1791 [BPF_ALU64 | BPF_NEG] = neg_reg64, 1792 [BPF_ALU64 | BPF_LSH | BPF_K] = shl_imm64, 1793 [BPF_ALU64 | BPF_RSH | BPF_K] = shr_imm64, 1794 [BPF_ALU | BPF_MOV | BPF_X] = mov_reg, 1795 [BPF_ALU | BPF_MOV | BPF_K] = mov_imm, 1796 [BPF_ALU | BPF_XOR | BPF_X] = xor_reg, 1797 [BPF_ALU | BPF_XOR | BPF_K] = xor_imm, 1798 [BPF_ALU | BPF_AND | BPF_X] = and_reg, 1799 [BPF_ALU | BPF_AND | BPF_K] = and_imm, 1800 [BPF_ALU | BPF_OR | BPF_X] = or_reg, 1801 [BPF_ALU | BPF_OR | BPF_K] = or_imm, 1802 [BPF_ALU | BPF_ADD | BPF_X] = add_reg, 1803 [BPF_ALU | BPF_ADD | BPF_K] = add_imm, 1804 [BPF_ALU | BPF_SUB | BPF_X] = sub_reg, 1805 [BPF_ALU | BPF_SUB | BPF_K] = sub_imm, 1806 [BPF_ALU | BPF_NEG] = neg_reg, 1807 [BPF_ALU | BPF_LSH | BPF_K] = shl_imm, 1808 [BPF_ALU | BPF_END | BPF_X] = end_reg32, 1809 [BPF_LD | BPF_IMM | BPF_DW] = imm_ld8, 1810 [BPF_LD | BPF_ABS | BPF_B] = data_ld1, 1811 [BPF_LD | BPF_ABS | BPF_H] = data_ld2, 1812 [BPF_LD | BPF_ABS | BPF_W] = data_ld4, 1813 [BPF_LD | BPF_IND | BPF_B] = data_ind_ld1, 1814 [BPF_LD | BPF_IND | BPF_H] = data_ind_ld2, 1815 [BPF_LD | BPF_IND | BPF_W] = data_ind_ld4, 1816 [BPF_LDX | BPF_MEM | BPF_B] = mem_ldx1, 1817 [BPF_LDX | BPF_MEM | BPF_H] = mem_ldx2, 1818 [BPF_LDX | BPF_MEM | BPF_W] = mem_ldx4, 1819 [BPF_LDX | BPF_MEM | BPF_DW] = mem_ldx8, 1820 [BPF_STX | BPF_MEM | BPF_B] = mem_stx1, 1821 [BPF_STX | BPF_MEM | BPF_H] = mem_stx2, 1822 [BPF_STX | BPF_MEM | BPF_W] = mem_stx4, 1823 [BPF_STX | BPF_MEM | BPF_DW] = mem_stx8, 1824 [BPF_ST | BPF_MEM | BPF_B] = mem_st1, 1825 [BPF_ST | BPF_MEM | BPF_H] = mem_st2, 1826 [BPF_ST | BPF_MEM | BPF_W] = mem_st4, 1827 [BPF_ST | BPF_MEM | BPF_DW] = mem_st8, 1828 [BPF_JMP | BPF_JA | BPF_K] = jump, 1829 [BPF_JMP | BPF_JEQ | BPF_K] = jeq_imm, 1830 [BPF_JMP | BPF_JGT | BPF_K] = jgt_imm, 1831 [BPF_JMP | BPF_JGE | BPF_K] = jge_imm, 1832 [BPF_JMP | BPF_JLT | BPF_K] = jlt_imm, 1833 [BPF_JMP | BPF_JLE | BPF_K] = jle_imm, 1834 [BPF_JMP | BPF_JSET | BPF_K] = jset_imm, 1835 [BPF_JMP | BPF_JNE | BPF_K] = jne_imm, 1836 [BPF_JMP | BPF_JEQ | BPF_X] = jeq_reg, 1837 [BPF_JMP | BPF_JGT | BPF_X] = jgt_reg, 1838 [BPF_JMP | BPF_JGE | BPF_X] = jge_reg, 1839 [BPF_JMP | BPF_JLT | BPF_X] = jlt_reg, 1840 [BPF_JMP | BPF_JLE | BPF_X] = jle_reg, 1841 [BPF_JMP | BPF_JSET | BPF_X] = jset_reg, 1842 [BPF_JMP | BPF_JNE | BPF_X] = jne_reg, 1843 [BPF_JMP | BPF_EXIT] = goto_out, 1844 }; 1845 1846 /* --- Misc code --- */ 1847 static void br_set_offset(u64 *instr, u16 offset) 1848 { 1849 u16 addr_lo, addr_hi; 1850 1851 addr_lo = offset & (OP_BR_ADDR_LO >> __bf_shf(OP_BR_ADDR_LO)); 1852 addr_hi = offset != addr_lo; 1853 *instr &= ~(OP_BR_ADDR_HI | OP_BR_ADDR_LO); 1854 *instr |= FIELD_PREP(OP_BR_ADDR_HI, addr_hi); 1855 *instr |= FIELD_PREP(OP_BR_ADDR_LO, addr_lo); 1856 } 1857 1858 /* --- Assembler logic --- */ 1859 static int nfp_fixup_branches(struct nfp_prog *nfp_prog) 1860 { 1861 struct nfp_insn_meta *meta, *jmp_dst; 1862 u32 idx, br_idx; 1863 1864 list_for_each_entry(meta, &nfp_prog->insns, l) { 1865 if (meta->skip) 1866 continue; 1867 if (BPF_CLASS(meta->insn.code) != BPF_JMP) 1868 continue; 1869 1870 if (list_is_last(&meta->l, &nfp_prog->insns)) 1871 idx = nfp_prog->last_bpf_off; 1872 else 1873 idx = list_next_entry(meta, l)->off - 1; 1874 1875 br_idx = nfp_prog_offset_to_index(nfp_prog, idx); 1876 1877 if (!nfp_is_br(nfp_prog->prog[br_idx])) { 1878 pr_err("Fixup found block not ending in branch %d %02x %016llx!!\n", 1879 br_idx, meta->insn.code, nfp_prog->prog[br_idx]); 1880 return -ELOOP; 1881 } 1882 /* Leave special branches for later */ 1883 if (FIELD_GET(OP_BR_SPECIAL, nfp_prog->prog[br_idx])) 1884 continue; 1885 1886 if (!meta->jmp_dst) { 1887 pr_err("Non-exit jump doesn't have destination info recorded!!\n"); 1888 return -ELOOP; 1889 } 1890 1891 jmp_dst = meta->jmp_dst; 1892 1893 if (jmp_dst->skip) { 1894 pr_err("Branch landing on removed instruction!!\n"); 1895 return -ELOOP; 1896 } 1897 1898 for (idx = nfp_prog_offset_to_index(nfp_prog, meta->off); 1899 idx <= br_idx; idx++) { 1900 if (!nfp_is_br(nfp_prog->prog[idx])) 1901 continue; 1902 br_set_offset(&nfp_prog->prog[idx], jmp_dst->off); 1903 } 1904 } 1905 1906 /* Fixup 'goto out's separately, they can be scattered around */ 1907 for (br_idx = 0; br_idx < nfp_prog->prog_len; br_idx++) { 1908 enum br_special special; 1909 1910 if ((nfp_prog->prog[br_idx] & OP_BR_BASE_MASK) != OP_BR_BASE) 1911 continue; 1912 1913 special = FIELD_GET(OP_BR_SPECIAL, nfp_prog->prog[br_idx]); 1914 switch (special) { 1915 case OP_BR_NORMAL: 1916 break; 1917 case OP_BR_GO_OUT: 1918 br_set_offset(&nfp_prog->prog[br_idx], 1919 nfp_prog->tgt_out); 1920 break; 1921 case OP_BR_GO_ABORT: 1922 br_set_offset(&nfp_prog->prog[br_idx], 1923 nfp_prog->tgt_abort); 1924 break; 1925 } 1926 1927 nfp_prog->prog[br_idx] &= ~OP_BR_SPECIAL; 1928 } 1929 1930 return 0; 1931 } 1932 1933 static void nfp_intro(struct nfp_prog *nfp_prog) 1934 { 1935 wrp_immed(nfp_prog, plen_reg(nfp_prog), GENMASK(13, 0)); 1936 emit_alu(nfp_prog, plen_reg(nfp_prog), 1937 plen_reg(nfp_prog), ALU_OP_AND, pv_len(nfp_prog)); 1938 } 1939 1940 static void nfp_outro_tc_da(struct nfp_prog *nfp_prog) 1941 { 1942 /* TC direct-action mode: 1943 * 0,1 ok NOT SUPPORTED[1] 1944 * 2 drop 0x22 -> drop, count as stat1 1945 * 4,5 nuke 0x02 -> drop 1946 * 7 redir 0x44 -> redir, count as stat2 1947 * * unspec 0x11 -> pass, count as stat0 1948 * 1949 * [1] We can't support OK and RECLASSIFY because we can't tell TC 1950 * the exact decision made. We are forced to support UNSPEC 1951 * to handle aborts so that's the only one we handle for passing 1952 * packets up the stack. 1953 */ 1954 /* Target for aborts */ 1955 nfp_prog->tgt_abort = nfp_prog_current_offset(nfp_prog); 1956 1957 emit_br_def(nfp_prog, nfp_prog->tgt_done, 2); 1958 1959 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS); 1960 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_imm(0x11), SHF_SC_L_SHF, 16); 1961 1962 /* Target for normal exits */ 1963 nfp_prog->tgt_out = nfp_prog_current_offset(nfp_prog); 1964 1965 /* if R0 > 7 jump to abort */ 1966 emit_alu(nfp_prog, reg_none(), reg_imm(7), ALU_OP_SUB, reg_b(0)); 1967 emit_br(nfp_prog, BR_BLO, nfp_prog->tgt_abort, 0); 1968 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS); 1969 1970 wrp_immed(nfp_prog, reg_b(2), 0x41221211); 1971 wrp_immed(nfp_prog, reg_b(3), 0x41001211); 1972 1973 emit_shf(nfp_prog, reg_a(1), 1974 reg_none(), SHF_OP_NONE, reg_b(0), SHF_SC_L_SHF, 2); 1975 1976 emit_alu(nfp_prog, reg_none(), reg_a(1), ALU_OP_OR, reg_imm(0)); 1977 emit_shf(nfp_prog, reg_a(2), 1978 reg_imm(0xf), SHF_OP_AND, reg_b(2), SHF_SC_R_SHF, 0); 1979 1980 emit_alu(nfp_prog, reg_none(), reg_a(1), ALU_OP_OR, reg_imm(0)); 1981 emit_shf(nfp_prog, reg_b(2), 1982 reg_imm(0xf), SHF_OP_AND, reg_b(3), SHF_SC_R_SHF, 0); 1983 1984 emit_br_def(nfp_prog, nfp_prog->tgt_done, 2); 1985 1986 emit_shf(nfp_prog, reg_b(2), 1987 reg_a(2), SHF_OP_OR, reg_b(2), SHF_SC_L_SHF, 4); 1988 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_b(2), SHF_SC_L_SHF, 16); 1989 } 1990 1991 static void nfp_outro_xdp(struct nfp_prog *nfp_prog) 1992 { 1993 /* XDP return codes: 1994 * 0 aborted 0x82 -> drop, count as stat3 1995 * 1 drop 0x22 -> drop, count as stat1 1996 * 2 pass 0x11 -> pass, count as stat0 1997 * 3 tx 0x44 -> redir, count as stat2 1998 * * unknown 0x82 -> drop, count as stat3 1999 */ 2000 /* Target for aborts */ 2001 nfp_prog->tgt_abort = nfp_prog_current_offset(nfp_prog); 2002 2003 emit_br_def(nfp_prog, nfp_prog->tgt_done, 2); 2004 2005 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS); 2006 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_imm(0x82), SHF_SC_L_SHF, 16); 2007 2008 /* Target for normal exits */ 2009 nfp_prog->tgt_out = nfp_prog_current_offset(nfp_prog); 2010 2011 /* if R0 > 3 jump to abort */ 2012 emit_alu(nfp_prog, reg_none(), reg_imm(3), ALU_OP_SUB, reg_b(0)); 2013 emit_br(nfp_prog, BR_BLO, nfp_prog->tgt_abort, 0); 2014 2015 wrp_immed(nfp_prog, reg_b(2), 0x44112282); 2016 2017 emit_shf(nfp_prog, reg_a(1), 2018 reg_none(), SHF_OP_NONE, reg_b(0), SHF_SC_L_SHF, 3); 2019 2020 emit_alu(nfp_prog, reg_none(), reg_a(1), ALU_OP_OR, reg_imm(0)); 2021 emit_shf(nfp_prog, reg_b(2), 2022 reg_imm(0xff), SHF_OP_AND, reg_b(2), SHF_SC_R_SHF, 0); 2023 2024 emit_br_def(nfp_prog, nfp_prog->tgt_done, 2); 2025 2026 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS); 2027 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_b(2), SHF_SC_L_SHF, 16); 2028 } 2029 2030 static void nfp_outro(struct nfp_prog *nfp_prog) 2031 { 2032 switch (nfp_prog->type) { 2033 case BPF_PROG_TYPE_SCHED_CLS: 2034 nfp_outro_tc_da(nfp_prog); 2035 break; 2036 case BPF_PROG_TYPE_XDP: 2037 nfp_outro_xdp(nfp_prog); 2038 break; 2039 default: 2040 WARN_ON(1); 2041 } 2042 } 2043 2044 static int nfp_translate(struct nfp_prog *nfp_prog) 2045 { 2046 struct nfp_insn_meta *meta; 2047 int err; 2048 2049 nfp_intro(nfp_prog); 2050 if (nfp_prog->error) 2051 return nfp_prog->error; 2052 2053 list_for_each_entry(meta, &nfp_prog->insns, l) { 2054 instr_cb_t cb = instr_cb[meta->insn.code]; 2055 2056 meta->off = nfp_prog_current_offset(nfp_prog); 2057 2058 if (meta->skip) { 2059 nfp_prog->n_translated++; 2060 continue; 2061 } 2062 2063 if (nfp_meta_has_prev(nfp_prog, meta) && 2064 nfp_meta_prev(meta)->double_cb) 2065 cb = nfp_meta_prev(meta)->double_cb; 2066 if (!cb) 2067 return -ENOENT; 2068 err = cb(nfp_prog, meta); 2069 if (err) 2070 return err; 2071 2072 nfp_prog->n_translated++; 2073 } 2074 2075 nfp_prog->last_bpf_off = nfp_prog_current_offset(nfp_prog) - 1; 2076 2077 nfp_outro(nfp_prog); 2078 if (nfp_prog->error) 2079 return nfp_prog->error; 2080 2081 wrp_nops(nfp_prog, NFP_USTORE_PREFETCH_WINDOW); 2082 if (nfp_prog->error) 2083 return nfp_prog->error; 2084 2085 return nfp_fixup_branches(nfp_prog); 2086 } 2087 2088 /* --- Optimizations --- */ 2089 static void nfp_bpf_opt_reg_init(struct nfp_prog *nfp_prog) 2090 { 2091 struct nfp_insn_meta *meta; 2092 2093 list_for_each_entry(meta, &nfp_prog->insns, l) { 2094 struct bpf_insn insn = meta->insn; 2095 2096 /* Programs converted from cBPF start with register xoring */ 2097 if (insn.code == (BPF_ALU64 | BPF_XOR | BPF_X) && 2098 insn.src_reg == insn.dst_reg) 2099 continue; 2100 2101 /* Programs start with R6 = R1 but we ignore the skb pointer */ 2102 if (insn.code == (BPF_ALU64 | BPF_MOV | BPF_X) && 2103 insn.src_reg == 1 && insn.dst_reg == 6) 2104 meta->skip = true; 2105 2106 /* Return as soon as something doesn't match */ 2107 if (!meta->skip) 2108 return; 2109 } 2110 } 2111 2112 /* Remove masking after load since our load guarantees this is not needed */ 2113 static void nfp_bpf_opt_ld_mask(struct nfp_prog *nfp_prog) 2114 { 2115 struct nfp_insn_meta *meta1, *meta2; 2116 const s32 exp_mask[] = { 2117 [BPF_B] = 0x000000ffU, 2118 [BPF_H] = 0x0000ffffU, 2119 [BPF_W] = 0xffffffffU, 2120 }; 2121 2122 nfp_for_each_insn_walk2(nfp_prog, meta1, meta2) { 2123 struct bpf_insn insn, next; 2124 2125 insn = meta1->insn; 2126 next = meta2->insn; 2127 2128 if (BPF_CLASS(insn.code) != BPF_LD) 2129 continue; 2130 if (BPF_MODE(insn.code) != BPF_ABS && 2131 BPF_MODE(insn.code) != BPF_IND) 2132 continue; 2133 2134 if (next.code != (BPF_ALU64 | BPF_AND | BPF_K)) 2135 continue; 2136 2137 if (!exp_mask[BPF_SIZE(insn.code)]) 2138 continue; 2139 if (exp_mask[BPF_SIZE(insn.code)] != next.imm) 2140 continue; 2141 2142 if (next.src_reg || next.dst_reg) 2143 continue; 2144 2145 meta2->skip = true; 2146 } 2147 } 2148 2149 static void nfp_bpf_opt_ld_shift(struct nfp_prog *nfp_prog) 2150 { 2151 struct nfp_insn_meta *meta1, *meta2, *meta3; 2152 2153 nfp_for_each_insn_walk3(nfp_prog, meta1, meta2, meta3) { 2154 struct bpf_insn insn, next1, next2; 2155 2156 insn = meta1->insn; 2157 next1 = meta2->insn; 2158 next2 = meta3->insn; 2159 2160 if (BPF_CLASS(insn.code) != BPF_LD) 2161 continue; 2162 if (BPF_MODE(insn.code) != BPF_ABS && 2163 BPF_MODE(insn.code) != BPF_IND) 2164 continue; 2165 if (BPF_SIZE(insn.code) != BPF_W) 2166 continue; 2167 2168 if (!(next1.code == (BPF_LSH | BPF_K | BPF_ALU64) && 2169 next2.code == (BPF_RSH | BPF_K | BPF_ALU64)) && 2170 !(next1.code == (BPF_RSH | BPF_K | BPF_ALU64) && 2171 next2.code == (BPF_LSH | BPF_K | BPF_ALU64))) 2172 continue; 2173 2174 if (next1.src_reg || next1.dst_reg || 2175 next2.src_reg || next2.dst_reg) 2176 continue; 2177 2178 if (next1.imm != 0x20 || next2.imm != 0x20) 2179 continue; 2180 2181 meta2->skip = true; 2182 meta3->skip = true; 2183 } 2184 } 2185 2186 static int nfp_bpf_optimize(struct nfp_prog *nfp_prog) 2187 { 2188 nfp_bpf_opt_reg_init(nfp_prog); 2189 2190 nfp_bpf_opt_ld_mask(nfp_prog); 2191 nfp_bpf_opt_ld_shift(nfp_prog); 2192 2193 return 0; 2194 } 2195 2196 static int nfp_bpf_ustore_calc(struct nfp_prog *nfp_prog, __le64 *ustore) 2197 { 2198 int i; 2199 2200 for (i = 0; i < nfp_prog->prog_len; i++) { 2201 int err; 2202 2203 err = nfp_ustore_check_valid_no_ecc(nfp_prog->prog[i]); 2204 if (err) 2205 return err; 2206 2207 nfp_prog->prog[i] = nfp_ustore_calc_ecc_insn(nfp_prog->prog[i]); 2208 2209 ustore[i] = cpu_to_le64(nfp_prog->prog[i]); 2210 } 2211 2212 return 0; 2213 } 2214 2215 int nfp_bpf_jit(struct nfp_prog *nfp_prog) 2216 { 2217 int ret; 2218 2219 ret = nfp_bpf_optimize(nfp_prog); 2220 if (ret) 2221 return ret; 2222 2223 ret = nfp_translate(nfp_prog); 2224 if (ret) { 2225 pr_err("Translation failed with error %d (translated: %u)\n", 2226 ret, nfp_prog->n_translated); 2227 return -EINVAL; 2228 } 2229 2230 return nfp_bpf_ustore_calc(nfp_prog, (__force __le64 *)nfp_prog->prog); 2231 } 2232