1 /* 2 * Copyright (C) 2016-2018 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/bug.h> 37 #include <linux/bpf.h> 38 #include <linux/filter.h> 39 #include <linux/kernel.h> 40 #include <linux/pkt_cls.h> 41 #include <linux/reciprocal_div.h> 42 #include <linux/unistd.h> 43 44 #include "main.h" 45 #include "../nfp_asm.h" 46 #include "../nfp_net_ctrl.h" 47 48 /* --- NFP prog --- */ 49 /* Foreach "multiple" entries macros provide pos and next<n> pointers. 50 * It's safe to modify the next pointers (but not pos). 51 */ 52 #define nfp_for_each_insn_walk2(nfp_prog, pos, next) \ 53 for (pos = list_first_entry(&(nfp_prog)->insns, typeof(*pos), l), \ 54 next = list_next_entry(pos, l); \ 55 &(nfp_prog)->insns != &pos->l && \ 56 &(nfp_prog)->insns != &next->l; \ 57 pos = nfp_meta_next(pos), \ 58 next = nfp_meta_next(pos)) 59 60 #define nfp_for_each_insn_walk3(nfp_prog, pos, next, next2) \ 61 for (pos = list_first_entry(&(nfp_prog)->insns, typeof(*pos), l), \ 62 next = list_next_entry(pos, l), \ 63 next2 = list_next_entry(next, l); \ 64 &(nfp_prog)->insns != &pos->l && \ 65 &(nfp_prog)->insns != &next->l && \ 66 &(nfp_prog)->insns != &next2->l; \ 67 pos = nfp_meta_next(pos), \ 68 next = nfp_meta_next(pos), \ 69 next2 = nfp_meta_next(next)) 70 71 static bool 72 nfp_meta_has_prev(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 73 { 74 return meta->l.prev != &nfp_prog->insns; 75 } 76 77 static void nfp_prog_push(struct nfp_prog *nfp_prog, u64 insn) 78 { 79 if (nfp_prog->__prog_alloc_len / sizeof(u64) == nfp_prog->prog_len) { 80 pr_warn("instruction limit reached (%u NFP instructions)\n", 81 nfp_prog->prog_len); 82 nfp_prog->error = -ENOSPC; 83 return; 84 } 85 86 nfp_prog->prog[nfp_prog->prog_len] = insn; 87 nfp_prog->prog_len++; 88 } 89 90 static unsigned int nfp_prog_current_offset(struct nfp_prog *nfp_prog) 91 { 92 return nfp_prog->prog_len; 93 } 94 95 static bool 96 nfp_prog_confirm_current_offset(struct nfp_prog *nfp_prog, unsigned int off) 97 { 98 /* If there is a recorded error we may have dropped instructions; 99 * that doesn't have to be due to translator bug, and the translation 100 * will fail anyway, so just return OK. 101 */ 102 if (nfp_prog->error) 103 return true; 104 return !WARN_ON_ONCE(nfp_prog_current_offset(nfp_prog) != off); 105 } 106 107 /* --- Emitters --- */ 108 static void 109 __emit_cmd(struct nfp_prog *nfp_prog, enum cmd_tgt_map op, 110 u8 mode, u8 xfer, u8 areg, u8 breg, u8 size, enum cmd_ctx_swap ctx, 111 bool indir) 112 { 113 u64 insn; 114 115 insn = FIELD_PREP(OP_CMD_A_SRC, areg) | 116 FIELD_PREP(OP_CMD_CTX, ctx) | 117 FIELD_PREP(OP_CMD_B_SRC, breg) | 118 FIELD_PREP(OP_CMD_TOKEN, cmd_tgt_act[op].token) | 119 FIELD_PREP(OP_CMD_XFER, xfer) | 120 FIELD_PREP(OP_CMD_CNT, size) | 121 FIELD_PREP(OP_CMD_SIG, ctx != CMD_CTX_NO_SWAP) | 122 FIELD_PREP(OP_CMD_TGT_CMD, cmd_tgt_act[op].tgt_cmd) | 123 FIELD_PREP(OP_CMD_INDIR, indir) | 124 FIELD_PREP(OP_CMD_MODE, mode); 125 126 nfp_prog_push(nfp_prog, insn); 127 } 128 129 static void 130 emit_cmd_any(struct nfp_prog *nfp_prog, enum cmd_tgt_map op, u8 mode, u8 xfer, 131 swreg lreg, swreg rreg, u8 size, enum cmd_ctx_swap ctx, bool indir) 132 { 133 struct nfp_insn_re_regs reg; 134 int err; 135 136 err = swreg_to_restricted(reg_none(), lreg, rreg, ®, false); 137 if (err) { 138 nfp_prog->error = err; 139 return; 140 } 141 if (reg.swap) { 142 pr_err("cmd can't swap arguments\n"); 143 nfp_prog->error = -EFAULT; 144 return; 145 } 146 if (reg.dst_lmextn || reg.src_lmextn) { 147 pr_err("cmd can't use LMextn\n"); 148 nfp_prog->error = -EFAULT; 149 return; 150 } 151 152 __emit_cmd(nfp_prog, op, mode, xfer, reg.areg, reg.breg, size, ctx, 153 indir); 154 } 155 156 static void 157 emit_cmd(struct nfp_prog *nfp_prog, enum cmd_tgt_map op, u8 mode, u8 xfer, 158 swreg lreg, swreg rreg, u8 size, enum cmd_ctx_swap ctx) 159 { 160 emit_cmd_any(nfp_prog, op, mode, xfer, lreg, rreg, size, ctx, false); 161 } 162 163 static void 164 emit_cmd_indir(struct nfp_prog *nfp_prog, enum cmd_tgt_map op, u8 mode, u8 xfer, 165 swreg lreg, swreg rreg, u8 size, enum cmd_ctx_swap ctx) 166 { 167 emit_cmd_any(nfp_prog, op, mode, xfer, lreg, rreg, size, ctx, true); 168 } 169 170 static void 171 __emit_br(struct nfp_prog *nfp_prog, enum br_mask mask, enum br_ev_pip ev_pip, 172 enum br_ctx_signal_state css, u16 addr, u8 defer) 173 { 174 u16 addr_lo, addr_hi; 175 u64 insn; 176 177 addr_lo = addr & (OP_BR_ADDR_LO >> __bf_shf(OP_BR_ADDR_LO)); 178 addr_hi = addr != addr_lo; 179 180 insn = OP_BR_BASE | 181 FIELD_PREP(OP_BR_MASK, mask) | 182 FIELD_PREP(OP_BR_EV_PIP, ev_pip) | 183 FIELD_PREP(OP_BR_CSS, css) | 184 FIELD_PREP(OP_BR_DEFBR, defer) | 185 FIELD_PREP(OP_BR_ADDR_LO, addr_lo) | 186 FIELD_PREP(OP_BR_ADDR_HI, addr_hi); 187 188 nfp_prog_push(nfp_prog, insn); 189 } 190 191 static void 192 emit_br_relo(struct nfp_prog *nfp_prog, enum br_mask mask, u16 addr, u8 defer, 193 enum nfp_relo_type relo) 194 { 195 if (mask == BR_UNC && defer > 2) { 196 pr_err("BUG: branch defer out of bounds %d\n", defer); 197 nfp_prog->error = -EFAULT; 198 return; 199 } 200 201 __emit_br(nfp_prog, mask, 202 mask != BR_UNC ? BR_EV_PIP_COND : BR_EV_PIP_UNCOND, 203 BR_CSS_NONE, addr, defer); 204 205 nfp_prog->prog[nfp_prog->prog_len - 1] |= 206 FIELD_PREP(OP_RELO_TYPE, relo); 207 } 208 209 static void 210 emit_br(struct nfp_prog *nfp_prog, enum br_mask mask, u16 addr, u8 defer) 211 { 212 emit_br_relo(nfp_prog, mask, addr, defer, RELO_BR_REL); 213 } 214 215 static void 216 __emit_br_bit(struct nfp_prog *nfp_prog, u16 areg, u16 breg, u16 addr, u8 defer, 217 bool set, bool src_lmextn) 218 { 219 u16 addr_lo, addr_hi; 220 u64 insn; 221 222 addr_lo = addr & (OP_BR_BIT_ADDR_LO >> __bf_shf(OP_BR_BIT_ADDR_LO)); 223 addr_hi = addr != addr_lo; 224 225 insn = OP_BR_BIT_BASE | 226 FIELD_PREP(OP_BR_BIT_A_SRC, areg) | 227 FIELD_PREP(OP_BR_BIT_B_SRC, breg) | 228 FIELD_PREP(OP_BR_BIT_BV, set) | 229 FIELD_PREP(OP_BR_BIT_DEFBR, defer) | 230 FIELD_PREP(OP_BR_BIT_ADDR_LO, addr_lo) | 231 FIELD_PREP(OP_BR_BIT_ADDR_HI, addr_hi) | 232 FIELD_PREP(OP_BR_BIT_SRC_LMEXTN, src_lmextn); 233 234 nfp_prog_push(nfp_prog, insn); 235 } 236 237 static void 238 emit_br_bit_relo(struct nfp_prog *nfp_prog, swreg src, u8 bit, u16 addr, 239 u8 defer, bool set, enum nfp_relo_type relo) 240 { 241 struct nfp_insn_re_regs reg; 242 int err; 243 244 /* NOTE: The bit to test is specified as an rotation amount, such that 245 * the bit to test will be placed on the MSB of the result when 246 * doing a rotate right. For bit X, we need right rotate X + 1. 247 */ 248 bit += 1; 249 250 err = swreg_to_restricted(reg_none(), src, reg_imm(bit), ®, false); 251 if (err) { 252 nfp_prog->error = err; 253 return; 254 } 255 256 __emit_br_bit(nfp_prog, reg.areg, reg.breg, addr, defer, set, 257 reg.src_lmextn); 258 259 nfp_prog->prog[nfp_prog->prog_len - 1] |= 260 FIELD_PREP(OP_RELO_TYPE, relo); 261 } 262 263 static void 264 emit_br_bset(struct nfp_prog *nfp_prog, swreg src, u8 bit, u16 addr, u8 defer) 265 { 266 emit_br_bit_relo(nfp_prog, src, bit, addr, defer, true, RELO_BR_REL); 267 } 268 269 static void 270 __emit_immed(struct nfp_prog *nfp_prog, u16 areg, u16 breg, u16 imm_hi, 271 enum immed_width width, bool invert, 272 enum immed_shift shift, bool wr_both, 273 bool dst_lmextn, bool src_lmextn) 274 { 275 u64 insn; 276 277 insn = OP_IMMED_BASE | 278 FIELD_PREP(OP_IMMED_A_SRC, areg) | 279 FIELD_PREP(OP_IMMED_B_SRC, breg) | 280 FIELD_PREP(OP_IMMED_IMM, imm_hi) | 281 FIELD_PREP(OP_IMMED_WIDTH, width) | 282 FIELD_PREP(OP_IMMED_INV, invert) | 283 FIELD_PREP(OP_IMMED_SHIFT, shift) | 284 FIELD_PREP(OP_IMMED_WR_AB, wr_both) | 285 FIELD_PREP(OP_IMMED_SRC_LMEXTN, src_lmextn) | 286 FIELD_PREP(OP_IMMED_DST_LMEXTN, dst_lmextn); 287 288 nfp_prog_push(nfp_prog, insn); 289 } 290 291 static void 292 emit_immed(struct nfp_prog *nfp_prog, swreg dst, u16 imm, 293 enum immed_width width, bool invert, enum immed_shift shift) 294 { 295 struct nfp_insn_ur_regs reg; 296 int err; 297 298 if (swreg_type(dst) == NN_REG_IMM) { 299 nfp_prog->error = -EFAULT; 300 return; 301 } 302 303 err = swreg_to_unrestricted(dst, dst, reg_imm(imm & 0xff), ®); 304 if (err) { 305 nfp_prog->error = err; 306 return; 307 } 308 309 /* Use reg.dst when destination is No-Dest. */ 310 __emit_immed(nfp_prog, 311 swreg_type(dst) == NN_REG_NONE ? reg.dst : reg.areg, 312 reg.breg, imm >> 8, width, invert, shift, 313 reg.wr_both, reg.dst_lmextn, reg.src_lmextn); 314 } 315 316 static void 317 __emit_shf(struct nfp_prog *nfp_prog, u16 dst, enum alu_dst_ab dst_ab, 318 enum shf_sc sc, u8 shift, 319 u16 areg, enum shf_op op, u16 breg, bool i8, bool sw, bool wr_both, 320 bool dst_lmextn, bool src_lmextn) 321 { 322 u64 insn; 323 324 if (!FIELD_FIT(OP_SHF_SHIFT, shift)) { 325 nfp_prog->error = -EFAULT; 326 return; 327 } 328 329 if (sc == SHF_SC_L_SHF) 330 shift = 32 - shift; 331 332 insn = OP_SHF_BASE | 333 FIELD_PREP(OP_SHF_A_SRC, areg) | 334 FIELD_PREP(OP_SHF_SC, sc) | 335 FIELD_PREP(OP_SHF_B_SRC, breg) | 336 FIELD_PREP(OP_SHF_I8, i8) | 337 FIELD_PREP(OP_SHF_SW, sw) | 338 FIELD_PREP(OP_SHF_DST, dst) | 339 FIELD_PREP(OP_SHF_SHIFT, shift) | 340 FIELD_PREP(OP_SHF_OP, op) | 341 FIELD_PREP(OP_SHF_DST_AB, dst_ab) | 342 FIELD_PREP(OP_SHF_WR_AB, wr_both) | 343 FIELD_PREP(OP_SHF_SRC_LMEXTN, src_lmextn) | 344 FIELD_PREP(OP_SHF_DST_LMEXTN, dst_lmextn); 345 346 nfp_prog_push(nfp_prog, insn); 347 } 348 349 static void 350 emit_shf(struct nfp_prog *nfp_prog, swreg dst, 351 swreg lreg, enum shf_op op, swreg rreg, enum shf_sc sc, u8 shift) 352 { 353 struct nfp_insn_re_regs reg; 354 int err; 355 356 err = swreg_to_restricted(dst, lreg, rreg, ®, true); 357 if (err) { 358 nfp_prog->error = err; 359 return; 360 } 361 362 __emit_shf(nfp_prog, reg.dst, reg.dst_ab, sc, shift, 363 reg.areg, op, reg.breg, reg.i8, reg.swap, reg.wr_both, 364 reg.dst_lmextn, reg.src_lmextn); 365 } 366 367 static void 368 emit_shf_indir(struct nfp_prog *nfp_prog, swreg dst, 369 swreg lreg, enum shf_op op, swreg rreg, enum shf_sc sc) 370 { 371 if (sc == SHF_SC_R_ROT) { 372 pr_err("indirect shift is not allowed on rotation\n"); 373 nfp_prog->error = -EFAULT; 374 return; 375 } 376 377 emit_shf(nfp_prog, dst, lreg, op, rreg, sc, 0); 378 } 379 380 static void 381 __emit_alu(struct nfp_prog *nfp_prog, u16 dst, enum alu_dst_ab dst_ab, 382 u16 areg, enum alu_op op, u16 breg, bool swap, bool wr_both, 383 bool dst_lmextn, bool src_lmextn) 384 { 385 u64 insn; 386 387 insn = OP_ALU_BASE | 388 FIELD_PREP(OP_ALU_A_SRC, areg) | 389 FIELD_PREP(OP_ALU_B_SRC, breg) | 390 FIELD_PREP(OP_ALU_DST, dst) | 391 FIELD_PREP(OP_ALU_SW, swap) | 392 FIELD_PREP(OP_ALU_OP, op) | 393 FIELD_PREP(OP_ALU_DST_AB, dst_ab) | 394 FIELD_PREP(OP_ALU_WR_AB, wr_both) | 395 FIELD_PREP(OP_ALU_SRC_LMEXTN, src_lmextn) | 396 FIELD_PREP(OP_ALU_DST_LMEXTN, dst_lmextn); 397 398 nfp_prog_push(nfp_prog, insn); 399 } 400 401 static void 402 emit_alu(struct nfp_prog *nfp_prog, swreg dst, 403 swreg lreg, enum alu_op op, swreg rreg) 404 { 405 struct nfp_insn_ur_regs reg; 406 int err; 407 408 err = swreg_to_unrestricted(dst, lreg, rreg, ®); 409 if (err) { 410 nfp_prog->error = err; 411 return; 412 } 413 414 __emit_alu(nfp_prog, reg.dst, reg.dst_ab, 415 reg.areg, op, reg.breg, reg.swap, reg.wr_both, 416 reg.dst_lmextn, reg.src_lmextn); 417 } 418 419 static void 420 __emit_mul(struct nfp_prog *nfp_prog, enum alu_dst_ab dst_ab, u16 areg, 421 enum mul_type type, enum mul_step step, u16 breg, bool swap, 422 bool wr_both, bool dst_lmextn, bool src_lmextn) 423 { 424 u64 insn; 425 426 insn = OP_MUL_BASE | 427 FIELD_PREP(OP_MUL_A_SRC, areg) | 428 FIELD_PREP(OP_MUL_B_SRC, breg) | 429 FIELD_PREP(OP_MUL_STEP, step) | 430 FIELD_PREP(OP_MUL_DST_AB, dst_ab) | 431 FIELD_PREP(OP_MUL_SW, swap) | 432 FIELD_PREP(OP_MUL_TYPE, type) | 433 FIELD_PREP(OP_MUL_WR_AB, wr_both) | 434 FIELD_PREP(OP_MUL_SRC_LMEXTN, src_lmextn) | 435 FIELD_PREP(OP_MUL_DST_LMEXTN, dst_lmextn); 436 437 nfp_prog_push(nfp_prog, insn); 438 } 439 440 static void 441 emit_mul(struct nfp_prog *nfp_prog, swreg lreg, enum mul_type type, 442 enum mul_step step, swreg rreg) 443 { 444 struct nfp_insn_ur_regs reg; 445 u16 areg; 446 int err; 447 448 if (type == MUL_TYPE_START && step != MUL_STEP_NONE) { 449 nfp_prog->error = -EINVAL; 450 return; 451 } 452 453 if (step == MUL_LAST || step == MUL_LAST_2) { 454 /* When type is step and step Number is LAST or LAST2, left 455 * source is used as destination. 456 */ 457 err = swreg_to_unrestricted(lreg, reg_none(), rreg, ®); 458 areg = reg.dst; 459 } else { 460 err = swreg_to_unrestricted(reg_none(), lreg, rreg, ®); 461 areg = reg.areg; 462 } 463 464 if (err) { 465 nfp_prog->error = err; 466 return; 467 } 468 469 __emit_mul(nfp_prog, reg.dst_ab, areg, type, step, reg.breg, reg.swap, 470 reg.wr_both, reg.dst_lmextn, reg.src_lmextn); 471 } 472 473 static void 474 __emit_ld_field(struct nfp_prog *nfp_prog, enum shf_sc sc, 475 u8 areg, u8 bmask, u8 breg, u8 shift, bool imm8, 476 bool zero, bool swap, bool wr_both, 477 bool dst_lmextn, bool src_lmextn) 478 { 479 u64 insn; 480 481 insn = OP_LDF_BASE | 482 FIELD_PREP(OP_LDF_A_SRC, areg) | 483 FIELD_PREP(OP_LDF_SC, sc) | 484 FIELD_PREP(OP_LDF_B_SRC, breg) | 485 FIELD_PREP(OP_LDF_I8, imm8) | 486 FIELD_PREP(OP_LDF_SW, swap) | 487 FIELD_PREP(OP_LDF_ZF, zero) | 488 FIELD_PREP(OP_LDF_BMASK, bmask) | 489 FIELD_PREP(OP_LDF_SHF, shift) | 490 FIELD_PREP(OP_LDF_WR_AB, wr_both) | 491 FIELD_PREP(OP_LDF_SRC_LMEXTN, src_lmextn) | 492 FIELD_PREP(OP_LDF_DST_LMEXTN, dst_lmextn); 493 494 nfp_prog_push(nfp_prog, insn); 495 } 496 497 static void 498 emit_ld_field_any(struct nfp_prog *nfp_prog, swreg dst, u8 bmask, swreg src, 499 enum shf_sc sc, u8 shift, bool zero) 500 { 501 struct nfp_insn_re_regs reg; 502 int err; 503 504 /* Note: ld_field is special as it uses one of the src regs as dst */ 505 err = swreg_to_restricted(dst, dst, src, ®, true); 506 if (err) { 507 nfp_prog->error = err; 508 return; 509 } 510 511 __emit_ld_field(nfp_prog, sc, reg.areg, bmask, reg.breg, shift, 512 reg.i8, zero, reg.swap, reg.wr_both, 513 reg.dst_lmextn, reg.src_lmextn); 514 } 515 516 static void 517 emit_ld_field(struct nfp_prog *nfp_prog, swreg dst, u8 bmask, swreg src, 518 enum shf_sc sc, u8 shift) 519 { 520 emit_ld_field_any(nfp_prog, dst, bmask, src, sc, shift, false); 521 } 522 523 static void 524 __emit_lcsr(struct nfp_prog *nfp_prog, u16 areg, u16 breg, bool wr, u16 addr, 525 bool dst_lmextn, bool src_lmextn) 526 { 527 u64 insn; 528 529 insn = OP_LCSR_BASE | 530 FIELD_PREP(OP_LCSR_A_SRC, areg) | 531 FIELD_PREP(OP_LCSR_B_SRC, breg) | 532 FIELD_PREP(OP_LCSR_WRITE, wr) | 533 FIELD_PREP(OP_LCSR_ADDR, addr / 4) | 534 FIELD_PREP(OP_LCSR_SRC_LMEXTN, src_lmextn) | 535 FIELD_PREP(OP_LCSR_DST_LMEXTN, dst_lmextn); 536 537 nfp_prog_push(nfp_prog, insn); 538 } 539 540 static void emit_csr_wr(struct nfp_prog *nfp_prog, swreg src, u16 addr) 541 { 542 struct nfp_insn_ur_regs reg; 543 int err; 544 545 /* This instruction takes immeds instead of reg_none() for the ignored 546 * operand, but we can't encode 2 immeds in one instr with our normal 547 * swreg infra so if param is an immed, we encode as reg_none() and 548 * copy the immed to both operands. 549 */ 550 if (swreg_type(src) == NN_REG_IMM) { 551 err = swreg_to_unrestricted(reg_none(), src, reg_none(), ®); 552 reg.breg = reg.areg; 553 } else { 554 err = swreg_to_unrestricted(reg_none(), src, reg_imm(0), ®); 555 } 556 if (err) { 557 nfp_prog->error = err; 558 return; 559 } 560 561 __emit_lcsr(nfp_prog, reg.areg, reg.breg, true, addr, 562 false, reg.src_lmextn); 563 } 564 565 /* CSR value is read in following immed[gpr, 0] */ 566 static void __emit_csr_rd(struct nfp_prog *nfp_prog, u16 addr) 567 { 568 __emit_lcsr(nfp_prog, 0, 0, false, addr, false, false); 569 } 570 571 static void emit_nop(struct nfp_prog *nfp_prog) 572 { 573 __emit_immed(nfp_prog, UR_REG_IMM, UR_REG_IMM, 0, 0, 0, 0, 0, 0, 0); 574 } 575 576 /* --- Wrappers --- */ 577 static bool pack_immed(u32 imm, u16 *val, enum immed_shift *shift) 578 { 579 if (!(imm & 0xffff0000)) { 580 *val = imm; 581 *shift = IMMED_SHIFT_0B; 582 } else if (!(imm & 0xff0000ff)) { 583 *val = imm >> 8; 584 *shift = IMMED_SHIFT_1B; 585 } else if (!(imm & 0x0000ffff)) { 586 *val = imm >> 16; 587 *shift = IMMED_SHIFT_2B; 588 } else { 589 return false; 590 } 591 592 return true; 593 } 594 595 static void wrp_immed(struct nfp_prog *nfp_prog, swreg dst, u32 imm) 596 { 597 enum immed_shift shift; 598 u16 val; 599 600 if (pack_immed(imm, &val, &shift)) { 601 emit_immed(nfp_prog, dst, val, IMMED_WIDTH_ALL, false, shift); 602 } else if (pack_immed(~imm, &val, &shift)) { 603 emit_immed(nfp_prog, dst, val, IMMED_WIDTH_ALL, true, shift); 604 } else { 605 emit_immed(nfp_prog, dst, imm & 0xffff, IMMED_WIDTH_ALL, 606 false, IMMED_SHIFT_0B); 607 emit_immed(nfp_prog, dst, imm >> 16, IMMED_WIDTH_WORD, 608 false, IMMED_SHIFT_2B); 609 } 610 } 611 612 static void 613 wrp_immed_relo(struct nfp_prog *nfp_prog, swreg dst, u32 imm, 614 enum nfp_relo_type relo) 615 { 616 if (imm > 0xffff) { 617 pr_err("relocation of a large immediate!\n"); 618 nfp_prog->error = -EFAULT; 619 return; 620 } 621 emit_immed(nfp_prog, dst, imm, IMMED_WIDTH_ALL, false, IMMED_SHIFT_0B); 622 623 nfp_prog->prog[nfp_prog->prog_len - 1] |= 624 FIELD_PREP(OP_RELO_TYPE, relo); 625 } 626 627 /* ur_load_imm_any() - encode immediate or use tmp register (unrestricted) 628 * If the @imm is small enough encode it directly in operand and return 629 * otherwise load @imm to a spare register and return its encoding. 630 */ 631 static swreg ur_load_imm_any(struct nfp_prog *nfp_prog, u32 imm, swreg tmp_reg) 632 { 633 if (FIELD_FIT(UR_REG_IMM_MAX, imm)) 634 return reg_imm(imm); 635 636 wrp_immed(nfp_prog, tmp_reg, imm); 637 return tmp_reg; 638 } 639 640 /* re_load_imm_any() - encode immediate or use tmp register (restricted) 641 * If the @imm is small enough encode it directly in operand and return 642 * otherwise load @imm to a spare register and return its encoding. 643 */ 644 static swreg re_load_imm_any(struct nfp_prog *nfp_prog, u32 imm, swreg tmp_reg) 645 { 646 if (FIELD_FIT(RE_REG_IMM_MAX, imm)) 647 return reg_imm(imm); 648 649 wrp_immed(nfp_prog, tmp_reg, imm); 650 return tmp_reg; 651 } 652 653 static void wrp_nops(struct nfp_prog *nfp_prog, unsigned int count) 654 { 655 while (count--) 656 emit_nop(nfp_prog); 657 } 658 659 static void wrp_mov(struct nfp_prog *nfp_prog, swreg dst, swreg src) 660 { 661 emit_alu(nfp_prog, dst, reg_none(), ALU_OP_NONE, src); 662 } 663 664 static void wrp_reg_mov(struct nfp_prog *nfp_prog, u16 dst, u16 src) 665 { 666 wrp_mov(nfp_prog, reg_both(dst), reg_b(src)); 667 } 668 669 /* wrp_reg_subpart() - load @field_len bytes from @offset of @src, write the 670 * result to @dst from low end. 671 */ 672 static void 673 wrp_reg_subpart(struct nfp_prog *nfp_prog, swreg dst, swreg src, u8 field_len, 674 u8 offset) 675 { 676 enum shf_sc sc = offset ? SHF_SC_R_SHF : SHF_SC_NONE; 677 u8 mask = (1 << field_len) - 1; 678 679 emit_ld_field_any(nfp_prog, dst, mask, src, sc, offset * 8, true); 680 } 681 682 /* wrp_reg_or_subpart() - load @field_len bytes from low end of @src, or the 683 * result to @dst from offset, there is no change on the other bits of @dst. 684 */ 685 static void 686 wrp_reg_or_subpart(struct nfp_prog *nfp_prog, swreg dst, swreg src, 687 u8 field_len, u8 offset) 688 { 689 enum shf_sc sc = offset ? SHF_SC_L_SHF : SHF_SC_NONE; 690 u8 mask = ((1 << field_len) - 1) << offset; 691 692 emit_ld_field(nfp_prog, dst, mask, src, sc, 32 - offset * 8); 693 } 694 695 static void 696 addr40_offset(struct nfp_prog *nfp_prog, u8 src_gpr, swreg offset, 697 swreg *rega, swreg *regb) 698 { 699 if (offset == reg_imm(0)) { 700 *rega = reg_a(src_gpr); 701 *regb = reg_b(src_gpr + 1); 702 return; 703 } 704 705 emit_alu(nfp_prog, imm_a(nfp_prog), reg_a(src_gpr), ALU_OP_ADD, offset); 706 emit_alu(nfp_prog, imm_b(nfp_prog), reg_b(src_gpr + 1), ALU_OP_ADD_C, 707 reg_imm(0)); 708 *rega = imm_a(nfp_prog); 709 *regb = imm_b(nfp_prog); 710 } 711 712 /* NFP has Command Push Pull bus which supports bluk memory operations. */ 713 static int nfp_cpp_memcpy(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 714 { 715 bool descending_seq = meta->ldst_gather_len < 0; 716 s16 len = abs(meta->ldst_gather_len); 717 swreg src_base, off; 718 bool src_40bit_addr; 719 unsigned int i; 720 u8 xfer_num; 721 722 off = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog)); 723 src_40bit_addr = meta->ptr.type == PTR_TO_MAP_VALUE; 724 src_base = reg_a(meta->insn.src_reg * 2); 725 xfer_num = round_up(len, 4) / 4; 726 727 if (src_40bit_addr) 728 addr40_offset(nfp_prog, meta->insn.src_reg * 2, off, &src_base, 729 &off); 730 731 /* Setup PREV_ALU fields to override memory read length. */ 732 if (len > 32) 733 wrp_immed(nfp_prog, reg_none(), 734 CMD_OVE_LEN | FIELD_PREP(CMD_OV_LEN, xfer_num - 1)); 735 736 /* Memory read from source addr into transfer-in registers. */ 737 emit_cmd_any(nfp_prog, CMD_TGT_READ32_SWAP, 738 src_40bit_addr ? CMD_MODE_40b_BA : CMD_MODE_32b, 0, 739 src_base, off, xfer_num - 1, CMD_CTX_SWAP, len > 32); 740 741 /* Move from transfer-in to transfer-out. */ 742 for (i = 0; i < xfer_num; i++) 743 wrp_mov(nfp_prog, reg_xfer(i), reg_xfer(i)); 744 745 off = re_load_imm_any(nfp_prog, meta->paired_st->off, imm_b(nfp_prog)); 746 747 if (len <= 8) { 748 /* Use single direct_ref write8. */ 749 emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 0, 750 reg_a(meta->paired_st->dst_reg * 2), off, len - 1, 751 CMD_CTX_SWAP); 752 } else if (len <= 32 && IS_ALIGNED(len, 4)) { 753 /* Use single direct_ref write32. */ 754 emit_cmd(nfp_prog, CMD_TGT_WRITE32_SWAP, CMD_MODE_32b, 0, 755 reg_a(meta->paired_st->dst_reg * 2), off, xfer_num - 1, 756 CMD_CTX_SWAP); 757 } else if (len <= 32) { 758 /* Use single indirect_ref write8. */ 759 wrp_immed(nfp_prog, reg_none(), 760 CMD_OVE_LEN | FIELD_PREP(CMD_OV_LEN, len - 1)); 761 emit_cmd_indir(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 0, 762 reg_a(meta->paired_st->dst_reg * 2), off, 763 len - 1, CMD_CTX_SWAP); 764 } else if (IS_ALIGNED(len, 4)) { 765 /* Use single indirect_ref write32. */ 766 wrp_immed(nfp_prog, reg_none(), 767 CMD_OVE_LEN | FIELD_PREP(CMD_OV_LEN, xfer_num - 1)); 768 emit_cmd_indir(nfp_prog, CMD_TGT_WRITE32_SWAP, CMD_MODE_32b, 0, 769 reg_a(meta->paired_st->dst_reg * 2), off, 770 xfer_num - 1, CMD_CTX_SWAP); 771 } else if (len <= 40) { 772 /* Use one direct_ref write32 to write the first 32-bytes, then 773 * another direct_ref write8 to write the remaining bytes. 774 */ 775 emit_cmd(nfp_prog, CMD_TGT_WRITE32_SWAP, CMD_MODE_32b, 0, 776 reg_a(meta->paired_st->dst_reg * 2), off, 7, 777 CMD_CTX_SWAP); 778 779 off = re_load_imm_any(nfp_prog, meta->paired_st->off + 32, 780 imm_b(nfp_prog)); 781 emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 8, 782 reg_a(meta->paired_st->dst_reg * 2), off, len - 33, 783 CMD_CTX_SWAP); 784 } else { 785 /* Use one indirect_ref write32 to write 4-bytes aligned length, 786 * then another direct_ref write8 to write the remaining bytes. 787 */ 788 u8 new_off; 789 790 wrp_immed(nfp_prog, reg_none(), 791 CMD_OVE_LEN | FIELD_PREP(CMD_OV_LEN, xfer_num - 2)); 792 emit_cmd_indir(nfp_prog, CMD_TGT_WRITE32_SWAP, CMD_MODE_32b, 0, 793 reg_a(meta->paired_st->dst_reg * 2), off, 794 xfer_num - 2, CMD_CTX_SWAP); 795 new_off = meta->paired_st->off + (xfer_num - 1) * 4; 796 off = re_load_imm_any(nfp_prog, new_off, imm_b(nfp_prog)); 797 emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 798 xfer_num - 1, reg_a(meta->paired_st->dst_reg * 2), off, 799 (len & 0x3) - 1, CMD_CTX_SWAP); 800 } 801 802 /* TODO: The following extra load is to make sure data flow be identical 803 * before and after we do memory copy optimization. 804 * 805 * The load destination register is not guaranteed to be dead, so we 806 * need to make sure it is loaded with the value the same as before 807 * this transformation. 808 * 809 * These extra loads could be removed once we have accurate register 810 * usage information. 811 */ 812 if (descending_seq) 813 xfer_num = 0; 814 else if (BPF_SIZE(meta->insn.code) != BPF_DW) 815 xfer_num = xfer_num - 1; 816 else 817 xfer_num = xfer_num - 2; 818 819 switch (BPF_SIZE(meta->insn.code)) { 820 case BPF_B: 821 wrp_reg_subpart(nfp_prog, reg_both(meta->insn.dst_reg * 2), 822 reg_xfer(xfer_num), 1, 823 IS_ALIGNED(len, 4) ? 3 : (len & 3) - 1); 824 break; 825 case BPF_H: 826 wrp_reg_subpart(nfp_prog, reg_both(meta->insn.dst_reg * 2), 827 reg_xfer(xfer_num), 2, (len & 3) ^ 2); 828 break; 829 case BPF_W: 830 wrp_mov(nfp_prog, reg_both(meta->insn.dst_reg * 2), 831 reg_xfer(0)); 832 break; 833 case BPF_DW: 834 wrp_mov(nfp_prog, reg_both(meta->insn.dst_reg * 2), 835 reg_xfer(xfer_num)); 836 wrp_mov(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 837 reg_xfer(xfer_num + 1)); 838 break; 839 } 840 841 if (BPF_SIZE(meta->insn.code) != BPF_DW) 842 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0); 843 844 return 0; 845 } 846 847 static int 848 data_ld(struct nfp_prog *nfp_prog, swreg offset, u8 dst_gpr, int size) 849 { 850 unsigned int i; 851 u16 shift, sz; 852 853 /* We load the value from the address indicated in @offset and then 854 * shift out the data we don't need. Note: this is big endian! 855 */ 856 sz = max(size, 4); 857 shift = size < 4 ? 4 - size : 0; 858 859 emit_cmd(nfp_prog, CMD_TGT_READ8, CMD_MODE_32b, 0, 860 pptr_reg(nfp_prog), offset, sz - 1, CMD_CTX_SWAP); 861 862 i = 0; 863 if (shift) 864 emit_shf(nfp_prog, reg_both(dst_gpr), reg_none(), SHF_OP_NONE, 865 reg_xfer(0), SHF_SC_R_SHF, shift * 8); 866 else 867 for (; i * 4 < size; i++) 868 wrp_mov(nfp_prog, reg_both(dst_gpr + i), reg_xfer(i)); 869 870 if (i < 2) 871 wrp_immed(nfp_prog, reg_both(dst_gpr + 1), 0); 872 873 return 0; 874 } 875 876 static int 877 data_ld_host_order(struct nfp_prog *nfp_prog, u8 dst_gpr, 878 swreg lreg, swreg rreg, int size, enum cmd_mode mode) 879 { 880 unsigned int i; 881 u8 mask, sz; 882 883 /* We load the value from the address indicated in rreg + lreg and then 884 * mask out the data we don't need. Note: this is little endian! 885 */ 886 sz = max(size, 4); 887 mask = size < 4 ? GENMASK(size - 1, 0) : 0; 888 889 emit_cmd(nfp_prog, CMD_TGT_READ32_SWAP, mode, 0, 890 lreg, rreg, sz / 4 - 1, CMD_CTX_SWAP); 891 892 i = 0; 893 if (mask) 894 emit_ld_field_any(nfp_prog, reg_both(dst_gpr), mask, 895 reg_xfer(0), SHF_SC_NONE, 0, true); 896 else 897 for (; i * 4 < size; i++) 898 wrp_mov(nfp_prog, reg_both(dst_gpr + i), reg_xfer(i)); 899 900 if (i < 2) 901 wrp_immed(nfp_prog, reg_both(dst_gpr + 1), 0); 902 903 return 0; 904 } 905 906 static int 907 data_ld_host_order_addr32(struct nfp_prog *nfp_prog, u8 src_gpr, swreg offset, 908 u8 dst_gpr, u8 size) 909 { 910 return data_ld_host_order(nfp_prog, dst_gpr, reg_a(src_gpr), offset, 911 size, CMD_MODE_32b); 912 } 913 914 static int 915 data_ld_host_order_addr40(struct nfp_prog *nfp_prog, u8 src_gpr, swreg offset, 916 u8 dst_gpr, u8 size) 917 { 918 swreg rega, regb; 919 920 addr40_offset(nfp_prog, src_gpr, offset, ®a, ®b); 921 922 return data_ld_host_order(nfp_prog, dst_gpr, rega, regb, 923 size, CMD_MODE_40b_BA); 924 } 925 926 static int 927 construct_data_ind_ld(struct nfp_prog *nfp_prog, u16 offset, u16 src, u8 size) 928 { 929 swreg tmp_reg; 930 931 /* Calculate the true offset (src_reg + imm) */ 932 tmp_reg = ur_load_imm_any(nfp_prog, offset, imm_b(nfp_prog)); 933 emit_alu(nfp_prog, imm_both(nfp_prog), reg_a(src), ALU_OP_ADD, tmp_reg); 934 935 /* Check packet length (size guaranteed to fit b/c it's u8) */ 936 emit_alu(nfp_prog, imm_a(nfp_prog), 937 imm_a(nfp_prog), ALU_OP_ADD, reg_imm(size)); 938 emit_alu(nfp_prog, reg_none(), 939 plen_reg(nfp_prog), ALU_OP_SUB, imm_a(nfp_prog)); 940 emit_br_relo(nfp_prog, BR_BLO, BR_OFF_RELO, 0, RELO_BR_GO_ABORT); 941 942 /* Load data */ 943 return data_ld(nfp_prog, imm_b(nfp_prog), 0, size); 944 } 945 946 static int construct_data_ld(struct nfp_prog *nfp_prog, u16 offset, u8 size) 947 { 948 swreg tmp_reg; 949 950 /* Check packet length */ 951 tmp_reg = ur_load_imm_any(nfp_prog, offset + size, imm_a(nfp_prog)); 952 emit_alu(nfp_prog, reg_none(), plen_reg(nfp_prog), ALU_OP_SUB, tmp_reg); 953 emit_br_relo(nfp_prog, BR_BLO, BR_OFF_RELO, 0, RELO_BR_GO_ABORT); 954 955 /* Load data */ 956 tmp_reg = re_load_imm_any(nfp_prog, offset, imm_b(nfp_prog)); 957 return data_ld(nfp_prog, tmp_reg, 0, size); 958 } 959 960 static int 961 data_stx_host_order(struct nfp_prog *nfp_prog, u8 dst_gpr, swreg offset, 962 u8 src_gpr, u8 size) 963 { 964 unsigned int i; 965 966 for (i = 0; i * 4 < size; i++) 967 wrp_mov(nfp_prog, reg_xfer(i), reg_a(src_gpr + i)); 968 969 emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 0, 970 reg_a(dst_gpr), offset, size - 1, CMD_CTX_SWAP); 971 972 return 0; 973 } 974 975 static int 976 data_st_host_order(struct nfp_prog *nfp_prog, u8 dst_gpr, swreg offset, 977 u64 imm, u8 size) 978 { 979 wrp_immed(nfp_prog, reg_xfer(0), imm); 980 if (size == 8) 981 wrp_immed(nfp_prog, reg_xfer(1), imm >> 32); 982 983 emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 0, 984 reg_a(dst_gpr), offset, size - 1, CMD_CTX_SWAP); 985 986 return 0; 987 } 988 989 typedef int 990 (*lmem_step)(struct nfp_prog *nfp_prog, u8 gpr, u8 gpr_byte, s32 off, 991 unsigned int size, bool first, bool new_gpr, bool last, bool lm3, 992 bool needs_inc); 993 994 static int 995 wrp_lmem_load(struct nfp_prog *nfp_prog, u8 dst, u8 dst_byte, s32 off, 996 unsigned int size, bool first, bool new_gpr, bool last, bool lm3, 997 bool needs_inc) 998 { 999 bool should_inc = needs_inc && new_gpr && !last; 1000 u32 idx, src_byte; 1001 enum shf_sc sc; 1002 swreg reg; 1003 int shf; 1004 u8 mask; 1005 1006 if (WARN_ON_ONCE(dst_byte + size > 4 || off % 4 + size > 4)) 1007 return -EOPNOTSUPP; 1008 1009 idx = off / 4; 1010 1011 /* Move the entire word */ 1012 if (size == 4) { 1013 wrp_mov(nfp_prog, reg_both(dst), 1014 should_inc ? reg_lm_inc(3) : reg_lm(lm3 ? 3 : 0, idx)); 1015 return 0; 1016 } 1017 1018 if (WARN_ON_ONCE(lm3 && idx > RE_REG_LM_IDX_MAX)) 1019 return -EOPNOTSUPP; 1020 1021 src_byte = off % 4; 1022 1023 mask = (1 << size) - 1; 1024 mask <<= dst_byte; 1025 1026 if (WARN_ON_ONCE(mask > 0xf)) 1027 return -EOPNOTSUPP; 1028 1029 shf = abs(src_byte - dst_byte) * 8; 1030 if (src_byte == dst_byte) { 1031 sc = SHF_SC_NONE; 1032 } else if (src_byte < dst_byte) { 1033 shf = 32 - shf; 1034 sc = SHF_SC_L_SHF; 1035 } else { 1036 sc = SHF_SC_R_SHF; 1037 } 1038 1039 /* ld_field can address fewer indexes, if offset too large do RMW. 1040 * Because we RMV twice we waste 2 cycles on unaligned 8 byte writes. 1041 */ 1042 if (idx <= RE_REG_LM_IDX_MAX) { 1043 reg = reg_lm(lm3 ? 3 : 0, idx); 1044 } else { 1045 reg = imm_a(nfp_prog); 1046 /* If it's not the first part of the load and we start a new GPR 1047 * that means we are loading a second part of the LMEM word into 1048 * a new GPR. IOW we've already looked that LMEM word and 1049 * therefore it has been loaded into imm_a(). 1050 */ 1051 if (first || !new_gpr) 1052 wrp_mov(nfp_prog, reg, reg_lm(0, idx)); 1053 } 1054 1055 emit_ld_field_any(nfp_prog, reg_both(dst), mask, reg, sc, shf, new_gpr); 1056 1057 if (should_inc) 1058 wrp_mov(nfp_prog, reg_none(), reg_lm_inc(3)); 1059 1060 return 0; 1061 } 1062 1063 static int 1064 wrp_lmem_store(struct nfp_prog *nfp_prog, u8 src, u8 src_byte, s32 off, 1065 unsigned int size, bool first, bool new_gpr, bool last, bool lm3, 1066 bool needs_inc) 1067 { 1068 bool should_inc = needs_inc && new_gpr && !last; 1069 u32 idx, dst_byte; 1070 enum shf_sc sc; 1071 swreg reg; 1072 int shf; 1073 u8 mask; 1074 1075 if (WARN_ON_ONCE(src_byte + size > 4 || off % 4 + size > 4)) 1076 return -EOPNOTSUPP; 1077 1078 idx = off / 4; 1079 1080 /* Move the entire word */ 1081 if (size == 4) { 1082 wrp_mov(nfp_prog, 1083 should_inc ? reg_lm_inc(3) : reg_lm(lm3 ? 3 : 0, idx), 1084 reg_b(src)); 1085 return 0; 1086 } 1087 1088 if (WARN_ON_ONCE(lm3 && idx > RE_REG_LM_IDX_MAX)) 1089 return -EOPNOTSUPP; 1090 1091 dst_byte = off % 4; 1092 1093 mask = (1 << size) - 1; 1094 mask <<= dst_byte; 1095 1096 if (WARN_ON_ONCE(mask > 0xf)) 1097 return -EOPNOTSUPP; 1098 1099 shf = abs(src_byte - dst_byte) * 8; 1100 if (src_byte == dst_byte) { 1101 sc = SHF_SC_NONE; 1102 } else if (src_byte < dst_byte) { 1103 shf = 32 - shf; 1104 sc = SHF_SC_L_SHF; 1105 } else { 1106 sc = SHF_SC_R_SHF; 1107 } 1108 1109 /* ld_field can address fewer indexes, if offset too large do RMW. 1110 * Because we RMV twice we waste 2 cycles on unaligned 8 byte writes. 1111 */ 1112 if (idx <= RE_REG_LM_IDX_MAX) { 1113 reg = reg_lm(lm3 ? 3 : 0, idx); 1114 } else { 1115 reg = imm_a(nfp_prog); 1116 /* Only first and last LMEM locations are going to need RMW, 1117 * the middle location will be overwritten fully. 1118 */ 1119 if (first || last) 1120 wrp_mov(nfp_prog, reg, reg_lm(0, idx)); 1121 } 1122 1123 emit_ld_field(nfp_prog, reg, mask, reg_b(src), sc, shf); 1124 1125 if (new_gpr || last) { 1126 if (idx > RE_REG_LM_IDX_MAX) 1127 wrp_mov(nfp_prog, reg_lm(0, idx), reg); 1128 if (should_inc) 1129 wrp_mov(nfp_prog, reg_none(), reg_lm_inc(3)); 1130 } 1131 1132 return 0; 1133 } 1134 1135 static int 1136 mem_op_stack(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1137 unsigned int size, unsigned int ptr_off, u8 gpr, u8 ptr_gpr, 1138 bool clr_gpr, lmem_step step) 1139 { 1140 s32 off = nfp_prog->stack_depth + meta->insn.off + ptr_off; 1141 bool first = true, last; 1142 bool needs_inc = false; 1143 swreg stack_off_reg; 1144 u8 prev_gpr = 255; 1145 u32 gpr_byte = 0; 1146 bool lm3 = true; 1147 int ret; 1148 1149 if (meta->ptr_not_const) { 1150 /* Use of the last encountered ptr_off is OK, they all have 1151 * the same alignment. Depend on low bits of value being 1152 * discarded when written to LMaddr register. 1153 */ 1154 stack_off_reg = ur_load_imm_any(nfp_prog, meta->insn.off, 1155 stack_imm(nfp_prog)); 1156 1157 emit_alu(nfp_prog, imm_b(nfp_prog), 1158 reg_a(ptr_gpr), ALU_OP_ADD, stack_off_reg); 1159 1160 needs_inc = true; 1161 } else if (off + size <= 64) { 1162 /* We can reach bottom 64B with LMaddr0 */ 1163 lm3 = false; 1164 } else if (round_down(off, 32) == round_down(off + size - 1, 32)) { 1165 /* We have to set up a new pointer. If we know the offset 1166 * and the entire access falls into a single 32 byte aligned 1167 * window we won't have to increment the LM pointer. 1168 * The 32 byte alignment is imporant because offset is ORed in 1169 * not added when doing *l$indexN[off]. 1170 */ 1171 stack_off_reg = ur_load_imm_any(nfp_prog, round_down(off, 32), 1172 stack_imm(nfp_prog)); 1173 emit_alu(nfp_prog, imm_b(nfp_prog), 1174 stack_reg(nfp_prog), ALU_OP_ADD, stack_off_reg); 1175 1176 off %= 32; 1177 } else { 1178 stack_off_reg = ur_load_imm_any(nfp_prog, round_down(off, 4), 1179 stack_imm(nfp_prog)); 1180 1181 emit_alu(nfp_prog, imm_b(nfp_prog), 1182 stack_reg(nfp_prog), ALU_OP_ADD, stack_off_reg); 1183 1184 needs_inc = true; 1185 } 1186 if (lm3) { 1187 emit_csr_wr(nfp_prog, imm_b(nfp_prog), NFP_CSR_ACT_LM_ADDR3); 1188 /* For size < 4 one slot will be filled by zeroing of upper. */ 1189 wrp_nops(nfp_prog, clr_gpr && size < 8 ? 2 : 3); 1190 } 1191 1192 if (clr_gpr && size < 8) 1193 wrp_immed(nfp_prog, reg_both(gpr + 1), 0); 1194 1195 while (size) { 1196 u32 slice_end; 1197 u8 slice_size; 1198 1199 slice_size = min(size, 4 - gpr_byte); 1200 slice_end = min(off + slice_size, round_up(off + 1, 4)); 1201 slice_size = slice_end - off; 1202 1203 last = slice_size == size; 1204 1205 if (needs_inc) 1206 off %= 4; 1207 1208 ret = step(nfp_prog, gpr, gpr_byte, off, slice_size, 1209 first, gpr != prev_gpr, last, lm3, needs_inc); 1210 if (ret) 1211 return ret; 1212 1213 prev_gpr = gpr; 1214 first = false; 1215 1216 gpr_byte += slice_size; 1217 if (gpr_byte >= 4) { 1218 gpr_byte -= 4; 1219 gpr++; 1220 } 1221 1222 size -= slice_size; 1223 off += slice_size; 1224 } 1225 1226 return 0; 1227 } 1228 1229 static void 1230 wrp_alu_imm(struct nfp_prog *nfp_prog, u8 dst, enum alu_op alu_op, u32 imm) 1231 { 1232 swreg tmp_reg; 1233 1234 if (alu_op == ALU_OP_AND) { 1235 if (!imm) 1236 wrp_immed(nfp_prog, reg_both(dst), 0); 1237 if (!imm || !~imm) 1238 return; 1239 } 1240 if (alu_op == ALU_OP_OR) { 1241 if (!~imm) 1242 wrp_immed(nfp_prog, reg_both(dst), ~0U); 1243 if (!imm || !~imm) 1244 return; 1245 } 1246 if (alu_op == ALU_OP_XOR) { 1247 if (!~imm) 1248 emit_alu(nfp_prog, reg_both(dst), reg_none(), 1249 ALU_OP_NOT, reg_b(dst)); 1250 if (!imm || !~imm) 1251 return; 1252 } 1253 1254 tmp_reg = ur_load_imm_any(nfp_prog, imm, imm_b(nfp_prog)); 1255 emit_alu(nfp_prog, reg_both(dst), reg_a(dst), alu_op, tmp_reg); 1256 } 1257 1258 static int 1259 wrp_alu64_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1260 enum alu_op alu_op, bool skip) 1261 { 1262 const struct bpf_insn *insn = &meta->insn; 1263 u64 imm = insn->imm; /* sign extend */ 1264 1265 if (skip) { 1266 meta->skip = true; 1267 return 0; 1268 } 1269 1270 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, alu_op, imm & ~0U); 1271 wrp_alu_imm(nfp_prog, insn->dst_reg * 2 + 1, alu_op, imm >> 32); 1272 1273 return 0; 1274 } 1275 1276 static int 1277 wrp_alu64_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1278 enum alu_op alu_op) 1279 { 1280 u8 dst = meta->insn.dst_reg * 2, src = meta->insn.src_reg * 2; 1281 1282 emit_alu(nfp_prog, reg_both(dst), reg_a(dst), alu_op, reg_b(src)); 1283 emit_alu(nfp_prog, reg_both(dst + 1), 1284 reg_a(dst + 1), alu_op, reg_b(src + 1)); 1285 1286 return 0; 1287 } 1288 1289 static int 1290 wrp_alu32_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1291 enum alu_op alu_op, bool skip) 1292 { 1293 const struct bpf_insn *insn = &meta->insn; 1294 1295 if (skip) { 1296 meta->skip = true; 1297 return 0; 1298 } 1299 1300 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, alu_op, insn->imm); 1301 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0); 1302 1303 return 0; 1304 } 1305 1306 static int 1307 wrp_alu32_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1308 enum alu_op alu_op) 1309 { 1310 u8 dst = meta->insn.dst_reg * 2, src = meta->insn.src_reg * 2; 1311 1312 emit_alu(nfp_prog, reg_both(dst), reg_a(dst), alu_op, reg_b(src)); 1313 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0); 1314 1315 return 0; 1316 } 1317 1318 static void 1319 wrp_test_reg_one(struct nfp_prog *nfp_prog, u8 dst, enum alu_op alu_op, u8 src, 1320 enum br_mask br_mask, u16 off) 1321 { 1322 emit_alu(nfp_prog, reg_none(), reg_a(dst), alu_op, reg_b(src)); 1323 emit_br(nfp_prog, br_mask, off, 0); 1324 } 1325 1326 static int 1327 wrp_test_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1328 enum alu_op alu_op, enum br_mask br_mask) 1329 { 1330 const struct bpf_insn *insn = &meta->insn; 1331 1332 wrp_test_reg_one(nfp_prog, insn->dst_reg * 2, alu_op, 1333 insn->src_reg * 2, br_mask, insn->off); 1334 wrp_test_reg_one(nfp_prog, insn->dst_reg * 2 + 1, alu_op, 1335 insn->src_reg * 2 + 1, br_mask, insn->off); 1336 1337 return 0; 1338 } 1339 1340 static const struct jmp_code_map { 1341 enum br_mask br_mask; 1342 bool swap; 1343 } jmp_code_map[] = { 1344 [BPF_JGT >> 4] = { BR_BLO, true }, 1345 [BPF_JGE >> 4] = { BR_BHS, false }, 1346 [BPF_JLT >> 4] = { BR_BLO, false }, 1347 [BPF_JLE >> 4] = { BR_BHS, true }, 1348 [BPF_JSGT >> 4] = { BR_BLT, true }, 1349 [BPF_JSGE >> 4] = { BR_BGE, false }, 1350 [BPF_JSLT >> 4] = { BR_BLT, false }, 1351 [BPF_JSLE >> 4] = { BR_BGE, true }, 1352 }; 1353 1354 static const struct jmp_code_map *nfp_jmp_code_get(struct nfp_insn_meta *meta) 1355 { 1356 unsigned int op; 1357 1358 op = BPF_OP(meta->insn.code) >> 4; 1359 /* br_mask of 0 is BR_BEQ which we don't use in jump code table */ 1360 if (WARN_ONCE(op >= ARRAY_SIZE(jmp_code_map) || 1361 !jmp_code_map[op].br_mask, 1362 "no code found for jump instruction")) 1363 return NULL; 1364 1365 return &jmp_code_map[op]; 1366 } 1367 1368 static int cmp_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1369 { 1370 const struct bpf_insn *insn = &meta->insn; 1371 u64 imm = insn->imm; /* sign extend */ 1372 const struct jmp_code_map *code; 1373 enum alu_op alu_op, carry_op; 1374 u8 reg = insn->dst_reg * 2; 1375 swreg tmp_reg; 1376 1377 code = nfp_jmp_code_get(meta); 1378 if (!code) 1379 return -EINVAL; 1380 1381 alu_op = meta->jump_neg_op ? ALU_OP_ADD : ALU_OP_SUB; 1382 carry_op = meta->jump_neg_op ? ALU_OP_ADD_C : ALU_OP_SUB_C; 1383 1384 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog)); 1385 if (!code->swap) 1386 emit_alu(nfp_prog, reg_none(), reg_a(reg), alu_op, tmp_reg); 1387 else 1388 emit_alu(nfp_prog, reg_none(), tmp_reg, alu_op, reg_a(reg)); 1389 1390 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog)); 1391 if (!code->swap) 1392 emit_alu(nfp_prog, reg_none(), 1393 reg_a(reg + 1), carry_op, tmp_reg); 1394 else 1395 emit_alu(nfp_prog, reg_none(), 1396 tmp_reg, carry_op, reg_a(reg + 1)); 1397 1398 emit_br(nfp_prog, code->br_mask, insn->off, 0); 1399 1400 return 0; 1401 } 1402 1403 static int cmp_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1404 { 1405 const struct bpf_insn *insn = &meta->insn; 1406 const struct jmp_code_map *code; 1407 u8 areg, breg; 1408 1409 code = nfp_jmp_code_get(meta); 1410 if (!code) 1411 return -EINVAL; 1412 1413 areg = insn->dst_reg * 2; 1414 breg = insn->src_reg * 2; 1415 1416 if (code->swap) { 1417 areg ^= breg; 1418 breg ^= areg; 1419 areg ^= breg; 1420 } 1421 1422 emit_alu(nfp_prog, reg_none(), reg_a(areg), ALU_OP_SUB, reg_b(breg)); 1423 emit_alu(nfp_prog, reg_none(), 1424 reg_a(areg + 1), ALU_OP_SUB_C, reg_b(breg + 1)); 1425 emit_br(nfp_prog, code->br_mask, insn->off, 0); 1426 1427 return 0; 1428 } 1429 1430 static void wrp_end32(struct nfp_prog *nfp_prog, swreg reg_in, u8 gpr_out) 1431 { 1432 emit_ld_field(nfp_prog, reg_both(gpr_out), 0xf, reg_in, 1433 SHF_SC_R_ROT, 8); 1434 emit_ld_field(nfp_prog, reg_both(gpr_out), 0x5, reg_a(gpr_out), 1435 SHF_SC_R_ROT, 16); 1436 } 1437 1438 static void 1439 wrp_mul_u32(struct nfp_prog *nfp_prog, swreg dst_hi, swreg dst_lo, swreg lreg, 1440 swreg rreg, bool gen_high_half) 1441 { 1442 emit_mul(nfp_prog, lreg, MUL_TYPE_START, MUL_STEP_NONE, rreg); 1443 emit_mul(nfp_prog, lreg, MUL_TYPE_STEP_32x32, MUL_STEP_1, rreg); 1444 emit_mul(nfp_prog, lreg, MUL_TYPE_STEP_32x32, MUL_STEP_2, rreg); 1445 emit_mul(nfp_prog, lreg, MUL_TYPE_STEP_32x32, MUL_STEP_3, rreg); 1446 emit_mul(nfp_prog, lreg, MUL_TYPE_STEP_32x32, MUL_STEP_4, rreg); 1447 emit_mul(nfp_prog, dst_lo, MUL_TYPE_STEP_32x32, MUL_LAST, reg_none()); 1448 if (gen_high_half) 1449 emit_mul(nfp_prog, dst_hi, MUL_TYPE_STEP_32x32, MUL_LAST_2, 1450 reg_none()); 1451 else 1452 wrp_immed(nfp_prog, dst_hi, 0); 1453 } 1454 1455 static void 1456 wrp_mul_u16(struct nfp_prog *nfp_prog, swreg dst_hi, swreg dst_lo, swreg lreg, 1457 swreg rreg) 1458 { 1459 emit_mul(nfp_prog, lreg, MUL_TYPE_START, MUL_STEP_NONE, rreg); 1460 emit_mul(nfp_prog, lreg, MUL_TYPE_STEP_16x16, MUL_STEP_1, rreg); 1461 emit_mul(nfp_prog, lreg, MUL_TYPE_STEP_16x16, MUL_STEP_2, rreg); 1462 emit_mul(nfp_prog, dst_lo, MUL_TYPE_STEP_16x16, MUL_LAST, reg_none()); 1463 } 1464 1465 static int 1466 wrp_mul(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1467 bool gen_high_half, bool ropnd_from_reg) 1468 { 1469 swreg multiplier, multiplicand, dst_hi, dst_lo; 1470 const struct bpf_insn *insn = &meta->insn; 1471 u32 lopnd_max, ropnd_max; 1472 u8 dst_reg; 1473 1474 dst_reg = insn->dst_reg; 1475 multiplicand = reg_a(dst_reg * 2); 1476 dst_hi = reg_both(dst_reg * 2 + 1); 1477 dst_lo = reg_both(dst_reg * 2); 1478 lopnd_max = meta->umax_dst; 1479 if (ropnd_from_reg) { 1480 multiplier = reg_b(insn->src_reg * 2); 1481 ropnd_max = meta->umax_src; 1482 } else { 1483 u32 imm = insn->imm; 1484 1485 multiplier = ur_load_imm_any(nfp_prog, imm, imm_b(nfp_prog)); 1486 ropnd_max = imm; 1487 } 1488 if (lopnd_max > U16_MAX || ropnd_max > U16_MAX) 1489 wrp_mul_u32(nfp_prog, dst_hi, dst_lo, multiplicand, multiplier, 1490 gen_high_half); 1491 else 1492 wrp_mul_u16(nfp_prog, dst_hi, dst_lo, multiplicand, multiplier); 1493 1494 return 0; 1495 } 1496 1497 static int wrp_div_imm(struct nfp_prog *nfp_prog, u8 dst, u64 imm) 1498 { 1499 swreg dst_both = reg_both(dst), dst_a = reg_a(dst), dst_b = reg_a(dst); 1500 struct reciprocal_value_adv rvalue; 1501 u8 pre_shift, exp; 1502 swreg magic; 1503 1504 if (imm > U32_MAX) { 1505 wrp_immed(nfp_prog, dst_both, 0); 1506 return 0; 1507 } 1508 1509 /* NOTE: because we are using "reciprocal_value_adv" which doesn't 1510 * support "divisor > (1u << 31)", we need to JIT separate NFP sequence 1511 * to handle such case which actually equals to the result of unsigned 1512 * comparison "dst >= imm" which could be calculated using the following 1513 * NFP sequence: 1514 * 1515 * alu[--, dst, -, imm] 1516 * immed[imm, 0] 1517 * alu[dst, imm, +carry, 0] 1518 * 1519 */ 1520 if (imm > 1U << 31) { 1521 swreg tmp_b = ur_load_imm_any(nfp_prog, imm, imm_b(nfp_prog)); 1522 1523 emit_alu(nfp_prog, reg_none(), dst_a, ALU_OP_SUB, tmp_b); 1524 wrp_immed(nfp_prog, imm_a(nfp_prog), 0); 1525 emit_alu(nfp_prog, dst_both, imm_a(nfp_prog), ALU_OP_ADD_C, 1526 reg_imm(0)); 1527 return 0; 1528 } 1529 1530 rvalue = reciprocal_value_adv(imm, 32); 1531 exp = rvalue.exp; 1532 if (rvalue.is_wide_m && !(imm & 1)) { 1533 pre_shift = fls(imm & -imm) - 1; 1534 rvalue = reciprocal_value_adv(imm >> pre_shift, 32 - pre_shift); 1535 } else { 1536 pre_shift = 0; 1537 } 1538 magic = ur_load_imm_any(nfp_prog, rvalue.m, imm_b(nfp_prog)); 1539 if (imm == 1U << exp) { 1540 emit_shf(nfp_prog, dst_both, reg_none(), SHF_OP_NONE, dst_b, 1541 SHF_SC_R_SHF, exp); 1542 } else if (rvalue.is_wide_m) { 1543 wrp_mul_u32(nfp_prog, imm_both(nfp_prog), reg_none(), dst_a, 1544 magic, true); 1545 emit_alu(nfp_prog, dst_both, dst_a, ALU_OP_SUB, 1546 imm_b(nfp_prog)); 1547 emit_shf(nfp_prog, dst_both, reg_none(), SHF_OP_NONE, dst_b, 1548 SHF_SC_R_SHF, 1); 1549 emit_alu(nfp_prog, dst_both, dst_a, ALU_OP_ADD, 1550 imm_b(nfp_prog)); 1551 emit_shf(nfp_prog, dst_both, reg_none(), SHF_OP_NONE, dst_b, 1552 SHF_SC_R_SHF, rvalue.sh - 1); 1553 } else { 1554 if (pre_shift) 1555 emit_shf(nfp_prog, dst_both, reg_none(), SHF_OP_NONE, 1556 dst_b, SHF_SC_R_SHF, pre_shift); 1557 wrp_mul_u32(nfp_prog, dst_both, reg_none(), dst_a, magic, true); 1558 emit_shf(nfp_prog, dst_both, reg_none(), SHF_OP_NONE, 1559 dst_b, SHF_SC_R_SHF, rvalue.sh); 1560 } 1561 1562 return 0; 1563 } 1564 1565 static int adjust_head(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1566 { 1567 swreg tmp = imm_a(nfp_prog), tmp_len = imm_b(nfp_prog); 1568 struct nfp_bpf_cap_adjust_head *adjust_head; 1569 u32 ret_einval, end; 1570 1571 adjust_head = &nfp_prog->bpf->adjust_head; 1572 1573 /* Optimized version - 5 vs 14 cycles */ 1574 if (nfp_prog->adjust_head_location != UINT_MAX) { 1575 if (WARN_ON_ONCE(nfp_prog->adjust_head_location != meta->n)) 1576 return -EINVAL; 1577 1578 emit_alu(nfp_prog, pptr_reg(nfp_prog), 1579 reg_a(2 * 2), ALU_OP_ADD, pptr_reg(nfp_prog)); 1580 emit_alu(nfp_prog, plen_reg(nfp_prog), 1581 plen_reg(nfp_prog), ALU_OP_SUB, reg_a(2 * 2)); 1582 emit_alu(nfp_prog, pv_len(nfp_prog), 1583 pv_len(nfp_prog), ALU_OP_SUB, reg_a(2 * 2)); 1584 1585 wrp_immed(nfp_prog, reg_both(0), 0); 1586 wrp_immed(nfp_prog, reg_both(1), 0); 1587 1588 /* TODO: when adjust head is guaranteed to succeed we can 1589 * also eliminate the following if (r0 == 0) branch. 1590 */ 1591 1592 return 0; 1593 } 1594 1595 ret_einval = nfp_prog_current_offset(nfp_prog) + 14; 1596 end = ret_einval + 2; 1597 1598 /* We need to use a temp because offset is just a part of the pkt ptr */ 1599 emit_alu(nfp_prog, tmp, 1600 reg_a(2 * 2), ALU_OP_ADD_2B, pptr_reg(nfp_prog)); 1601 1602 /* Validate result will fit within FW datapath constraints */ 1603 emit_alu(nfp_prog, reg_none(), 1604 tmp, ALU_OP_SUB, reg_imm(adjust_head->off_min)); 1605 emit_br(nfp_prog, BR_BLO, ret_einval, 0); 1606 emit_alu(nfp_prog, reg_none(), 1607 reg_imm(adjust_head->off_max), ALU_OP_SUB, tmp); 1608 emit_br(nfp_prog, BR_BLO, ret_einval, 0); 1609 1610 /* Validate the length is at least ETH_HLEN */ 1611 emit_alu(nfp_prog, tmp_len, 1612 plen_reg(nfp_prog), ALU_OP_SUB, reg_a(2 * 2)); 1613 emit_alu(nfp_prog, reg_none(), 1614 tmp_len, ALU_OP_SUB, reg_imm(ETH_HLEN)); 1615 emit_br(nfp_prog, BR_BMI, ret_einval, 0); 1616 1617 /* Load the ret code */ 1618 wrp_immed(nfp_prog, reg_both(0), 0); 1619 wrp_immed(nfp_prog, reg_both(1), 0); 1620 1621 /* Modify the packet metadata */ 1622 emit_ld_field(nfp_prog, pptr_reg(nfp_prog), 0x3, tmp, SHF_SC_NONE, 0); 1623 1624 /* Skip over the -EINVAL ret code (defer 2) */ 1625 emit_br(nfp_prog, BR_UNC, end, 2); 1626 1627 emit_alu(nfp_prog, plen_reg(nfp_prog), 1628 plen_reg(nfp_prog), ALU_OP_SUB, reg_a(2 * 2)); 1629 emit_alu(nfp_prog, pv_len(nfp_prog), 1630 pv_len(nfp_prog), ALU_OP_SUB, reg_a(2 * 2)); 1631 1632 /* return -EINVAL target */ 1633 if (!nfp_prog_confirm_current_offset(nfp_prog, ret_einval)) 1634 return -EINVAL; 1635 1636 wrp_immed(nfp_prog, reg_both(0), -22); 1637 wrp_immed(nfp_prog, reg_both(1), ~0); 1638 1639 if (!nfp_prog_confirm_current_offset(nfp_prog, end)) 1640 return -EINVAL; 1641 1642 return 0; 1643 } 1644 1645 static int 1646 map_call_stack_common(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1647 { 1648 bool load_lm_ptr; 1649 u32 ret_tgt; 1650 s64 lm_off; 1651 1652 /* We only have to reload LM0 if the key is not at start of stack */ 1653 lm_off = nfp_prog->stack_depth; 1654 lm_off += meta->arg2.reg.var_off.value + meta->arg2.reg.off; 1655 load_lm_ptr = meta->arg2.var_off || lm_off; 1656 1657 /* Set LM0 to start of key */ 1658 if (load_lm_ptr) 1659 emit_csr_wr(nfp_prog, reg_b(2 * 2), NFP_CSR_ACT_LM_ADDR0); 1660 if (meta->func_id == BPF_FUNC_map_update_elem) 1661 emit_csr_wr(nfp_prog, reg_b(3 * 2), NFP_CSR_ACT_LM_ADDR2); 1662 1663 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO + meta->func_id, 1664 2, RELO_BR_HELPER); 1665 ret_tgt = nfp_prog_current_offset(nfp_prog) + 2; 1666 1667 /* Load map ID into A0 */ 1668 wrp_mov(nfp_prog, reg_a(0), reg_a(2)); 1669 1670 /* Load the return address into B0 */ 1671 wrp_immed_relo(nfp_prog, reg_b(0), ret_tgt, RELO_IMMED_REL); 1672 1673 if (!nfp_prog_confirm_current_offset(nfp_prog, ret_tgt)) 1674 return -EINVAL; 1675 1676 /* Reset the LM0 pointer */ 1677 if (!load_lm_ptr) 1678 return 0; 1679 1680 emit_csr_wr(nfp_prog, stack_reg(nfp_prog), NFP_CSR_ACT_LM_ADDR0); 1681 wrp_nops(nfp_prog, 3); 1682 1683 return 0; 1684 } 1685 1686 static int 1687 nfp_get_prandom_u32(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1688 { 1689 __emit_csr_rd(nfp_prog, NFP_CSR_PSEUDO_RND_NUM); 1690 /* CSR value is read in following immed[gpr, 0] */ 1691 emit_immed(nfp_prog, reg_both(0), 0, 1692 IMMED_WIDTH_ALL, false, IMMED_SHIFT_0B); 1693 emit_immed(nfp_prog, reg_both(1), 0, 1694 IMMED_WIDTH_ALL, false, IMMED_SHIFT_0B); 1695 return 0; 1696 } 1697 1698 static int 1699 nfp_perf_event_output(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1700 { 1701 swreg ptr_type; 1702 u32 ret_tgt; 1703 1704 ptr_type = ur_load_imm_any(nfp_prog, meta->arg1.type, imm_a(nfp_prog)); 1705 1706 ret_tgt = nfp_prog_current_offset(nfp_prog) + 3; 1707 1708 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO + meta->func_id, 1709 2, RELO_BR_HELPER); 1710 1711 /* Load ptr type into A1 */ 1712 wrp_mov(nfp_prog, reg_a(1), ptr_type); 1713 1714 /* Load the return address into B0 */ 1715 wrp_immed_relo(nfp_prog, reg_b(0), ret_tgt, RELO_IMMED_REL); 1716 1717 if (!nfp_prog_confirm_current_offset(nfp_prog, ret_tgt)) 1718 return -EINVAL; 1719 1720 return 0; 1721 } 1722 1723 static int 1724 nfp_queue_select(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1725 { 1726 u32 jmp_tgt; 1727 1728 jmp_tgt = nfp_prog_current_offset(nfp_prog) + 5; 1729 1730 /* Make sure the queue id fits into FW field */ 1731 emit_alu(nfp_prog, reg_none(), reg_a(meta->insn.src_reg * 2), 1732 ALU_OP_AND_NOT_B, reg_imm(0xff)); 1733 emit_br(nfp_prog, BR_BEQ, jmp_tgt, 2); 1734 1735 /* Set the 'queue selected' bit and the queue value */ 1736 emit_shf(nfp_prog, pv_qsel_set(nfp_prog), 1737 pv_qsel_set(nfp_prog), SHF_OP_OR, reg_imm(1), 1738 SHF_SC_L_SHF, PKT_VEL_QSEL_SET_BIT); 1739 emit_ld_field(nfp_prog, 1740 pv_qsel_val(nfp_prog), 0x1, reg_b(meta->insn.src_reg * 2), 1741 SHF_SC_NONE, 0); 1742 /* Delay slots end here, we will jump over next instruction if queue 1743 * value fits into the field. 1744 */ 1745 emit_ld_field(nfp_prog, 1746 pv_qsel_val(nfp_prog), 0x1, reg_imm(NFP_NET_RXR_MAX), 1747 SHF_SC_NONE, 0); 1748 1749 if (!nfp_prog_confirm_current_offset(nfp_prog, jmp_tgt)) 1750 return -EINVAL; 1751 1752 return 0; 1753 } 1754 1755 /* --- Callbacks --- */ 1756 static int mov_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1757 { 1758 const struct bpf_insn *insn = &meta->insn; 1759 u8 dst = insn->dst_reg * 2; 1760 u8 src = insn->src_reg * 2; 1761 1762 if (insn->src_reg == BPF_REG_10) { 1763 swreg stack_depth_reg; 1764 1765 stack_depth_reg = ur_load_imm_any(nfp_prog, 1766 nfp_prog->stack_depth, 1767 stack_imm(nfp_prog)); 1768 emit_alu(nfp_prog, reg_both(dst), 1769 stack_reg(nfp_prog), ALU_OP_ADD, stack_depth_reg); 1770 wrp_immed(nfp_prog, reg_both(dst + 1), 0); 1771 } else { 1772 wrp_reg_mov(nfp_prog, dst, src); 1773 wrp_reg_mov(nfp_prog, dst + 1, src + 1); 1774 } 1775 1776 return 0; 1777 } 1778 1779 static int mov_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1780 { 1781 u64 imm = meta->insn.imm; /* sign extend */ 1782 1783 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2), imm & ~0U); 1784 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), imm >> 32); 1785 1786 return 0; 1787 } 1788 1789 static int xor_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1790 { 1791 return wrp_alu64_reg(nfp_prog, meta, ALU_OP_XOR); 1792 } 1793 1794 static int xor_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1795 { 1796 return wrp_alu64_imm(nfp_prog, meta, ALU_OP_XOR, !meta->insn.imm); 1797 } 1798 1799 static int and_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1800 { 1801 return wrp_alu64_reg(nfp_prog, meta, ALU_OP_AND); 1802 } 1803 1804 static int and_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1805 { 1806 return wrp_alu64_imm(nfp_prog, meta, ALU_OP_AND, !~meta->insn.imm); 1807 } 1808 1809 static int or_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1810 { 1811 return wrp_alu64_reg(nfp_prog, meta, ALU_OP_OR); 1812 } 1813 1814 static int or_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1815 { 1816 return wrp_alu64_imm(nfp_prog, meta, ALU_OP_OR, !meta->insn.imm); 1817 } 1818 1819 static int add_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1820 { 1821 const struct bpf_insn *insn = &meta->insn; 1822 1823 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2), 1824 reg_a(insn->dst_reg * 2), ALU_OP_ADD, 1825 reg_b(insn->src_reg * 2)); 1826 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 1827 reg_a(insn->dst_reg * 2 + 1), ALU_OP_ADD_C, 1828 reg_b(insn->src_reg * 2 + 1)); 1829 1830 return 0; 1831 } 1832 1833 static int add_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1834 { 1835 const struct bpf_insn *insn = &meta->insn; 1836 u64 imm = insn->imm; /* sign extend */ 1837 1838 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, ALU_OP_ADD, imm & ~0U); 1839 wrp_alu_imm(nfp_prog, insn->dst_reg * 2 + 1, ALU_OP_ADD_C, imm >> 32); 1840 1841 return 0; 1842 } 1843 1844 static int sub_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1845 { 1846 const struct bpf_insn *insn = &meta->insn; 1847 1848 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2), 1849 reg_a(insn->dst_reg * 2), ALU_OP_SUB, 1850 reg_b(insn->src_reg * 2)); 1851 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 1852 reg_a(insn->dst_reg * 2 + 1), ALU_OP_SUB_C, 1853 reg_b(insn->src_reg * 2 + 1)); 1854 1855 return 0; 1856 } 1857 1858 static int sub_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1859 { 1860 const struct bpf_insn *insn = &meta->insn; 1861 u64 imm = insn->imm; /* sign extend */ 1862 1863 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, ALU_OP_SUB, imm & ~0U); 1864 wrp_alu_imm(nfp_prog, insn->dst_reg * 2 + 1, ALU_OP_SUB_C, imm >> 32); 1865 1866 return 0; 1867 } 1868 1869 static int mul_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1870 { 1871 return wrp_mul(nfp_prog, meta, true, true); 1872 } 1873 1874 static int mul_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1875 { 1876 return wrp_mul(nfp_prog, meta, true, false); 1877 } 1878 1879 static int div_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1880 { 1881 const struct bpf_insn *insn = &meta->insn; 1882 1883 return wrp_div_imm(nfp_prog, insn->dst_reg * 2, insn->imm); 1884 } 1885 1886 static int div_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1887 { 1888 /* NOTE: verifier hook has rejected cases for which verifier doesn't 1889 * know whether the source operand is constant or not. 1890 */ 1891 return wrp_div_imm(nfp_prog, meta->insn.dst_reg * 2, meta->umin_src); 1892 } 1893 1894 static int neg_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1895 { 1896 const struct bpf_insn *insn = &meta->insn; 1897 1898 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2), reg_imm(0), 1899 ALU_OP_SUB, reg_b(insn->dst_reg * 2)); 1900 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2 + 1), reg_imm(0), 1901 ALU_OP_SUB_C, reg_b(insn->dst_reg * 2 + 1)); 1902 1903 return 0; 1904 } 1905 1906 /* Pseudo code: 1907 * if shift_amt >= 32 1908 * dst_high = dst_low << shift_amt[4:0] 1909 * dst_low = 0; 1910 * else 1911 * dst_high = (dst_high, dst_low) >> (32 - shift_amt) 1912 * dst_low = dst_low << shift_amt 1913 * 1914 * The indirect shift will use the same logic at runtime. 1915 */ 1916 static int __shl_imm64(struct nfp_prog *nfp_prog, u8 dst, u8 shift_amt) 1917 { 1918 if (shift_amt < 32) { 1919 emit_shf(nfp_prog, reg_both(dst + 1), reg_a(dst + 1), 1920 SHF_OP_NONE, reg_b(dst), SHF_SC_R_DSHF, 1921 32 - shift_amt); 1922 emit_shf(nfp_prog, reg_both(dst), reg_none(), SHF_OP_NONE, 1923 reg_b(dst), SHF_SC_L_SHF, shift_amt); 1924 } else if (shift_amt == 32) { 1925 wrp_reg_mov(nfp_prog, dst + 1, dst); 1926 wrp_immed(nfp_prog, reg_both(dst), 0); 1927 } else if (shift_amt > 32) { 1928 emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_NONE, 1929 reg_b(dst), SHF_SC_L_SHF, shift_amt - 32); 1930 wrp_immed(nfp_prog, reg_both(dst), 0); 1931 } 1932 1933 return 0; 1934 } 1935 1936 static int shl_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1937 { 1938 const struct bpf_insn *insn = &meta->insn; 1939 u8 dst = insn->dst_reg * 2; 1940 1941 return __shl_imm64(nfp_prog, dst, insn->imm); 1942 } 1943 1944 static void shl_reg64_lt32_high(struct nfp_prog *nfp_prog, u8 dst, u8 src) 1945 { 1946 emit_alu(nfp_prog, imm_both(nfp_prog), reg_imm(32), ALU_OP_SUB, 1947 reg_b(src)); 1948 emit_alu(nfp_prog, reg_none(), imm_a(nfp_prog), ALU_OP_OR, reg_imm(0)); 1949 emit_shf_indir(nfp_prog, reg_both(dst + 1), reg_a(dst + 1), SHF_OP_NONE, 1950 reg_b(dst), SHF_SC_R_DSHF); 1951 } 1952 1953 /* NOTE: for indirect left shift, HIGH part should be calculated first. */ 1954 static void shl_reg64_lt32_low(struct nfp_prog *nfp_prog, u8 dst, u8 src) 1955 { 1956 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0)); 1957 emit_shf_indir(nfp_prog, reg_both(dst), reg_none(), SHF_OP_NONE, 1958 reg_b(dst), SHF_SC_L_SHF); 1959 } 1960 1961 static void shl_reg64_lt32(struct nfp_prog *nfp_prog, u8 dst, u8 src) 1962 { 1963 shl_reg64_lt32_high(nfp_prog, dst, src); 1964 shl_reg64_lt32_low(nfp_prog, dst, src); 1965 } 1966 1967 static void shl_reg64_ge32(struct nfp_prog *nfp_prog, u8 dst, u8 src) 1968 { 1969 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0)); 1970 emit_shf_indir(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_NONE, 1971 reg_b(dst), SHF_SC_L_SHF); 1972 wrp_immed(nfp_prog, reg_both(dst), 0); 1973 } 1974 1975 static int shl_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1976 { 1977 const struct bpf_insn *insn = &meta->insn; 1978 u64 umin, umax; 1979 u8 dst, src; 1980 1981 dst = insn->dst_reg * 2; 1982 umin = meta->umin_src; 1983 umax = meta->umax_src; 1984 if (umin == umax) 1985 return __shl_imm64(nfp_prog, dst, umin); 1986 1987 src = insn->src_reg * 2; 1988 if (umax < 32) { 1989 shl_reg64_lt32(nfp_prog, dst, src); 1990 } else if (umin >= 32) { 1991 shl_reg64_ge32(nfp_prog, dst, src); 1992 } else { 1993 /* Generate different instruction sequences depending on runtime 1994 * value of shift amount. 1995 */ 1996 u16 label_ge32, label_end; 1997 1998 label_ge32 = nfp_prog_current_offset(nfp_prog) + 7; 1999 emit_br_bset(nfp_prog, reg_a(src), 5, label_ge32, 0); 2000 2001 shl_reg64_lt32_high(nfp_prog, dst, src); 2002 label_end = nfp_prog_current_offset(nfp_prog) + 6; 2003 emit_br(nfp_prog, BR_UNC, label_end, 2); 2004 /* shl_reg64_lt32_low packed in delay slot. */ 2005 shl_reg64_lt32_low(nfp_prog, dst, src); 2006 2007 if (!nfp_prog_confirm_current_offset(nfp_prog, label_ge32)) 2008 return -EINVAL; 2009 shl_reg64_ge32(nfp_prog, dst, src); 2010 2011 if (!nfp_prog_confirm_current_offset(nfp_prog, label_end)) 2012 return -EINVAL; 2013 } 2014 2015 return 0; 2016 } 2017 2018 /* Pseudo code: 2019 * if shift_amt >= 32 2020 * dst_high = 0; 2021 * dst_low = dst_high >> shift_amt[4:0] 2022 * else 2023 * dst_high = dst_high >> shift_amt 2024 * dst_low = (dst_high, dst_low) >> shift_amt 2025 * 2026 * The indirect shift will use the same logic at runtime. 2027 */ 2028 static int __shr_imm64(struct nfp_prog *nfp_prog, u8 dst, u8 shift_amt) 2029 { 2030 if (shift_amt < 32) { 2031 emit_shf(nfp_prog, reg_both(dst), reg_a(dst + 1), SHF_OP_NONE, 2032 reg_b(dst), SHF_SC_R_DSHF, shift_amt); 2033 emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_NONE, 2034 reg_b(dst + 1), SHF_SC_R_SHF, shift_amt); 2035 } else if (shift_amt == 32) { 2036 wrp_reg_mov(nfp_prog, dst, dst + 1); 2037 wrp_immed(nfp_prog, reg_both(dst + 1), 0); 2038 } else if (shift_amt > 32) { 2039 emit_shf(nfp_prog, reg_both(dst), reg_none(), SHF_OP_NONE, 2040 reg_b(dst + 1), SHF_SC_R_SHF, shift_amt - 32); 2041 wrp_immed(nfp_prog, reg_both(dst + 1), 0); 2042 } 2043 2044 return 0; 2045 } 2046 2047 static int shr_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2048 { 2049 const struct bpf_insn *insn = &meta->insn; 2050 u8 dst = insn->dst_reg * 2; 2051 2052 return __shr_imm64(nfp_prog, dst, insn->imm); 2053 } 2054 2055 /* NOTE: for indirect right shift, LOW part should be calculated first. */ 2056 static void shr_reg64_lt32_high(struct nfp_prog *nfp_prog, u8 dst, u8 src) 2057 { 2058 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0)); 2059 emit_shf_indir(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_NONE, 2060 reg_b(dst + 1), SHF_SC_R_SHF); 2061 } 2062 2063 static void shr_reg64_lt32_low(struct nfp_prog *nfp_prog, u8 dst, u8 src) 2064 { 2065 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0)); 2066 emit_shf_indir(nfp_prog, reg_both(dst), reg_a(dst + 1), SHF_OP_NONE, 2067 reg_b(dst), SHF_SC_R_DSHF); 2068 } 2069 2070 static void shr_reg64_lt32(struct nfp_prog *nfp_prog, u8 dst, u8 src) 2071 { 2072 shr_reg64_lt32_low(nfp_prog, dst, src); 2073 shr_reg64_lt32_high(nfp_prog, dst, src); 2074 } 2075 2076 static void shr_reg64_ge32(struct nfp_prog *nfp_prog, u8 dst, u8 src) 2077 { 2078 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0)); 2079 emit_shf_indir(nfp_prog, reg_both(dst), reg_none(), SHF_OP_NONE, 2080 reg_b(dst + 1), SHF_SC_R_SHF); 2081 wrp_immed(nfp_prog, reg_both(dst + 1), 0); 2082 } 2083 2084 static int shr_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2085 { 2086 const struct bpf_insn *insn = &meta->insn; 2087 u64 umin, umax; 2088 u8 dst, src; 2089 2090 dst = insn->dst_reg * 2; 2091 umin = meta->umin_src; 2092 umax = meta->umax_src; 2093 if (umin == umax) 2094 return __shr_imm64(nfp_prog, dst, umin); 2095 2096 src = insn->src_reg * 2; 2097 if (umax < 32) { 2098 shr_reg64_lt32(nfp_prog, dst, src); 2099 } else if (umin >= 32) { 2100 shr_reg64_ge32(nfp_prog, dst, src); 2101 } else { 2102 /* Generate different instruction sequences depending on runtime 2103 * value of shift amount. 2104 */ 2105 u16 label_ge32, label_end; 2106 2107 label_ge32 = nfp_prog_current_offset(nfp_prog) + 6; 2108 emit_br_bset(nfp_prog, reg_a(src), 5, label_ge32, 0); 2109 shr_reg64_lt32_low(nfp_prog, dst, src); 2110 label_end = nfp_prog_current_offset(nfp_prog) + 6; 2111 emit_br(nfp_prog, BR_UNC, label_end, 2); 2112 /* shr_reg64_lt32_high packed in delay slot. */ 2113 shr_reg64_lt32_high(nfp_prog, dst, src); 2114 2115 if (!nfp_prog_confirm_current_offset(nfp_prog, label_ge32)) 2116 return -EINVAL; 2117 shr_reg64_ge32(nfp_prog, dst, src); 2118 2119 if (!nfp_prog_confirm_current_offset(nfp_prog, label_end)) 2120 return -EINVAL; 2121 } 2122 2123 return 0; 2124 } 2125 2126 /* Code logic is the same as __shr_imm64 except ashr requires signedness bit 2127 * told through PREV_ALU result. 2128 */ 2129 static int __ashr_imm64(struct nfp_prog *nfp_prog, u8 dst, u8 shift_amt) 2130 { 2131 if (shift_amt < 32) { 2132 emit_shf(nfp_prog, reg_both(dst), reg_a(dst + 1), SHF_OP_NONE, 2133 reg_b(dst), SHF_SC_R_DSHF, shift_amt); 2134 /* Set signedness bit. */ 2135 emit_alu(nfp_prog, reg_none(), reg_a(dst + 1), ALU_OP_OR, 2136 reg_imm(0)); 2137 emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_ASHR, 2138 reg_b(dst + 1), SHF_SC_R_SHF, shift_amt); 2139 } else if (shift_amt == 32) { 2140 /* NOTE: this also helps setting signedness bit. */ 2141 wrp_reg_mov(nfp_prog, dst, dst + 1); 2142 emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_ASHR, 2143 reg_b(dst + 1), SHF_SC_R_SHF, 31); 2144 } else if (shift_amt > 32) { 2145 emit_alu(nfp_prog, reg_none(), reg_a(dst + 1), ALU_OP_OR, 2146 reg_imm(0)); 2147 emit_shf(nfp_prog, reg_both(dst), reg_none(), SHF_OP_ASHR, 2148 reg_b(dst + 1), SHF_SC_R_SHF, shift_amt - 32); 2149 emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_ASHR, 2150 reg_b(dst + 1), SHF_SC_R_SHF, 31); 2151 } 2152 2153 return 0; 2154 } 2155 2156 static int ashr_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2157 { 2158 const struct bpf_insn *insn = &meta->insn; 2159 u8 dst = insn->dst_reg * 2; 2160 2161 return __ashr_imm64(nfp_prog, dst, insn->imm); 2162 } 2163 2164 static void ashr_reg64_lt32_high(struct nfp_prog *nfp_prog, u8 dst, u8 src) 2165 { 2166 /* NOTE: the first insn will set both indirect shift amount (source A) 2167 * and signedness bit (MSB of result). 2168 */ 2169 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_b(dst + 1)); 2170 emit_shf_indir(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_ASHR, 2171 reg_b(dst + 1), SHF_SC_R_SHF); 2172 } 2173 2174 static void ashr_reg64_lt32_low(struct nfp_prog *nfp_prog, u8 dst, u8 src) 2175 { 2176 /* NOTE: it is the same as logic shift because we don't need to shift in 2177 * signedness bit when the shift amount is less than 32. 2178 */ 2179 return shr_reg64_lt32_low(nfp_prog, dst, src); 2180 } 2181 2182 static void ashr_reg64_lt32(struct nfp_prog *nfp_prog, u8 dst, u8 src) 2183 { 2184 ashr_reg64_lt32_low(nfp_prog, dst, src); 2185 ashr_reg64_lt32_high(nfp_prog, dst, src); 2186 } 2187 2188 static void ashr_reg64_ge32(struct nfp_prog *nfp_prog, u8 dst, u8 src) 2189 { 2190 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_b(dst + 1)); 2191 emit_shf_indir(nfp_prog, reg_both(dst), reg_none(), SHF_OP_ASHR, 2192 reg_b(dst + 1), SHF_SC_R_SHF); 2193 emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_ASHR, 2194 reg_b(dst + 1), SHF_SC_R_SHF, 31); 2195 } 2196 2197 /* Like ashr_imm64, but need to use indirect shift. */ 2198 static int ashr_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2199 { 2200 const struct bpf_insn *insn = &meta->insn; 2201 u64 umin, umax; 2202 u8 dst, src; 2203 2204 dst = insn->dst_reg * 2; 2205 umin = meta->umin_src; 2206 umax = meta->umax_src; 2207 if (umin == umax) 2208 return __ashr_imm64(nfp_prog, dst, umin); 2209 2210 src = insn->src_reg * 2; 2211 if (umax < 32) { 2212 ashr_reg64_lt32(nfp_prog, dst, src); 2213 } else if (umin >= 32) { 2214 ashr_reg64_ge32(nfp_prog, dst, src); 2215 } else { 2216 u16 label_ge32, label_end; 2217 2218 label_ge32 = nfp_prog_current_offset(nfp_prog) + 6; 2219 emit_br_bset(nfp_prog, reg_a(src), 5, label_ge32, 0); 2220 ashr_reg64_lt32_low(nfp_prog, dst, src); 2221 label_end = nfp_prog_current_offset(nfp_prog) + 6; 2222 emit_br(nfp_prog, BR_UNC, label_end, 2); 2223 /* ashr_reg64_lt32_high packed in delay slot. */ 2224 ashr_reg64_lt32_high(nfp_prog, dst, src); 2225 2226 if (!nfp_prog_confirm_current_offset(nfp_prog, label_ge32)) 2227 return -EINVAL; 2228 ashr_reg64_ge32(nfp_prog, dst, src); 2229 2230 if (!nfp_prog_confirm_current_offset(nfp_prog, label_end)) 2231 return -EINVAL; 2232 } 2233 2234 return 0; 2235 } 2236 2237 static int mov_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2238 { 2239 const struct bpf_insn *insn = &meta->insn; 2240 2241 wrp_reg_mov(nfp_prog, insn->dst_reg * 2, insn->src_reg * 2); 2242 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0); 2243 2244 return 0; 2245 } 2246 2247 static int mov_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2248 { 2249 const struct bpf_insn *insn = &meta->insn; 2250 2251 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2), insn->imm); 2252 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0); 2253 2254 return 0; 2255 } 2256 2257 static int xor_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2258 { 2259 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_XOR); 2260 } 2261 2262 static int xor_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2263 { 2264 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_XOR, !~meta->insn.imm); 2265 } 2266 2267 static int and_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2268 { 2269 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_AND); 2270 } 2271 2272 static int and_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2273 { 2274 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_AND, !~meta->insn.imm); 2275 } 2276 2277 static int or_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2278 { 2279 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_OR); 2280 } 2281 2282 static int or_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2283 { 2284 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_OR, !meta->insn.imm); 2285 } 2286 2287 static int add_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2288 { 2289 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_ADD); 2290 } 2291 2292 static int add_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2293 { 2294 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_ADD, !meta->insn.imm); 2295 } 2296 2297 static int sub_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2298 { 2299 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_SUB); 2300 } 2301 2302 static int sub_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2303 { 2304 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_SUB, !meta->insn.imm); 2305 } 2306 2307 static int mul_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2308 { 2309 return wrp_mul(nfp_prog, meta, false, true); 2310 } 2311 2312 static int mul_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2313 { 2314 return wrp_mul(nfp_prog, meta, false, false); 2315 } 2316 2317 static int div_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2318 { 2319 return div_reg64(nfp_prog, meta); 2320 } 2321 2322 static int div_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2323 { 2324 return div_imm64(nfp_prog, meta); 2325 } 2326 2327 static int neg_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2328 { 2329 u8 dst = meta->insn.dst_reg * 2; 2330 2331 emit_alu(nfp_prog, reg_both(dst), reg_imm(0), ALU_OP_SUB, reg_b(dst)); 2332 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0); 2333 2334 return 0; 2335 } 2336 2337 static int shl_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2338 { 2339 const struct bpf_insn *insn = &meta->insn; 2340 2341 if (!insn->imm) 2342 return 1; /* TODO: zero shift means indirect */ 2343 2344 emit_shf(nfp_prog, reg_both(insn->dst_reg * 2), 2345 reg_none(), SHF_OP_NONE, reg_b(insn->dst_reg * 2), 2346 SHF_SC_L_SHF, insn->imm); 2347 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0); 2348 2349 return 0; 2350 } 2351 2352 static int end_reg32(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2353 { 2354 const struct bpf_insn *insn = &meta->insn; 2355 u8 gpr = insn->dst_reg * 2; 2356 2357 switch (insn->imm) { 2358 case 16: 2359 emit_ld_field(nfp_prog, reg_both(gpr), 0x9, reg_b(gpr), 2360 SHF_SC_R_ROT, 8); 2361 emit_ld_field(nfp_prog, reg_both(gpr), 0xe, reg_a(gpr), 2362 SHF_SC_R_SHF, 16); 2363 2364 wrp_immed(nfp_prog, reg_both(gpr + 1), 0); 2365 break; 2366 case 32: 2367 wrp_end32(nfp_prog, reg_a(gpr), gpr); 2368 wrp_immed(nfp_prog, reg_both(gpr + 1), 0); 2369 break; 2370 case 64: 2371 wrp_mov(nfp_prog, imm_a(nfp_prog), reg_b(gpr + 1)); 2372 2373 wrp_end32(nfp_prog, reg_a(gpr), gpr + 1); 2374 wrp_end32(nfp_prog, imm_a(nfp_prog), gpr); 2375 break; 2376 } 2377 2378 return 0; 2379 } 2380 2381 static int imm_ld8_part2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2382 { 2383 struct nfp_insn_meta *prev = nfp_meta_prev(meta); 2384 u32 imm_lo, imm_hi; 2385 u8 dst; 2386 2387 dst = prev->insn.dst_reg * 2; 2388 imm_lo = prev->insn.imm; 2389 imm_hi = meta->insn.imm; 2390 2391 wrp_immed(nfp_prog, reg_both(dst), imm_lo); 2392 2393 /* mov is always 1 insn, load imm may be two, so try to use mov */ 2394 if (imm_hi == imm_lo) 2395 wrp_mov(nfp_prog, reg_both(dst + 1), reg_a(dst)); 2396 else 2397 wrp_immed(nfp_prog, reg_both(dst + 1), imm_hi); 2398 2399 return 0; 2400 } 2401 2402 static int imm_ld8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2403 { 2404 meta->double_cb = imm_ld8_part2; 2405 return 0; 2406 } 2407 2408 static int data_ld1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2409 { 2410 return construct_data_ld(nfp_prog, meta->insn.imm, 1); 2411 } 2412 2413 static int data_ld2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2414 { 2415 return construct_data_ld(nfp_prog, meta->insn.imm, 2); 2416 } 2417 2418 static int data_ld4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2419 { 2420 return construct_data_ld(nfp_prog, meta->insn.imm, 4); 2421 } 2422 2423 static int data_ind_ld1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2424 { 2425 return construct_data_ind_ld(nfp_prog, meta->insn.imm, 2426 meta->insn.src_reg * 2, 1); 2427 } 2428 2429 static int data_ind_ld2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2430 { 2431 return construct_data_ind_ld(nfp_prog, meta->insn.imm, 2432 meta->insn.src_reg * 2, 2); 2433 } 2434 2435 static int data_ind_ld4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2436 { 2437 return construct_data_ind_ld(nfp_prog, meta->insn.imm, 2438 meta->insn.src_reg * 2, 4); 2439 } 2440 2441 static int 2442 mem_ldx_stack(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2443 unsigned int size, unsigned int ptr_off) 2444 { 2445 return mem_op_stack(nfp_prog, meta, size, ptr_off, 2446 meta->insn.dst_reg * 2, meta->insn.src_reg * 2, 2447 true, wrp_lmem_load); 2448 } 2449 2450 static int mem_ldx_skb(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2451 u8 size) 2452 { 2453 swreg dst = reg_both(meta->insn.dst_reg * 2); 2454 2455 switch (meta->insn.off) { 2456 case offsetof(struct __sk_buff, len): 2457 if (size != FIELD_SIZEOF(struct __sk_buff, len)) 2458 return -EOPNOTSUPP; 2459 wrp_mov(nfp_prog, dst, plen_reg(nfp_prog)); 2460 break; 2461 case offsetof(struct __sk_buff, data): 2462 if (size != FIELD_SIZEOF(struct __sk_buff, data)) 2463 return -EOPNOTSUPP; 2464 wrp_mov(nfp_prog, dst, pptr_reg(nfp_prog)); 2465 break; 2466 case offsetof(struct __sk_buff, data_end): 2467 if (size != FIELD_SIZEOF(struct __sk_buff, data_end)) 2468 return -EOPNOTSUPP; 2469 emit_alu(nfp_prog, dst, 2470 plen_reg(nfp_prog), ALU_OP_ADD, pptr_reg(nfp_prog)); 2471 break; 2472 default: 2473 return -EOPNOTSUPP; 2474 } 2475 2476 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0); 2477 2478 return 0; 2479 } 2480 2481 static int mem_ldx_xdp(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2482 u8 size) 2483 { 2484 swreg dst = reg_both(meta->insn.dst_reg * 2); 2485 2486 switch (meta->insn.off) { 2487 case offsetof(struct xdp_md, data): 2488 if (size != FIELD_SIZEOF(struct xdp_md, data)) 2489 return -EOPNOTSUPP; 2490 wrp_mov(nfp_prog, dst, pptr_reg(nfp_prog)); 2491 break; 2492 case offsetof(struct xdp_md, data_end): 2493 if (size != FIELD_SIZEOF(struct xdp_md, data_end)) 2494 return -EOPNOTSUPP; 2495 emit_alu(nfp_prog, dst, 2496 plen_reg(nfp_prog), ALU_OP_ADD, pptr_reg(nfp_prog)); 2497 break; 2498 default: 2499 return -EOPNOTSUPP; 2500 } 2501 2502 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0); 2503 2504 return 0; 2505 } 2506 2507 static int 2508 mem_ldx_data(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2509 unsigned int size) 2510 { 2511 swreg tmp_reg; 2512 2513 tmp_reg = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog)); 2514 2515 return data_ld_host_order_addr32(nfp_prog, meta->insn.src_reg * 2, 2516 tmp_reg, meta->insn.dst_reg * 2, size); 2517 } 2518 2519 static int 2520 mem_ldx_emem(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2521 unsigned int size) 2522 { 2523 swreg tmp_reg; 2524 2525 tmp_reg = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog)); 2526 2527 return data_ld_host_order_addr40(nfp_prog, meta->insn.src_reg * 2, 2528 tmp_reg, meta->insn.dst_reg * 2, size); 2529 } 2530 2531 static void 2532 mem_ldx_data_init_pktcache(struct nfp_prog *nfp_prog, 2533 struct nfp_insn_meta *meta) 2534 { 2535 s16 range_start = meta->pkt_cache.range_start; 2536 s16 range_end = meta->pkt_cache.range_end; 2537 swreg src_base, off; 2538 u8 xfer_num, len; 2539 bool indir; 2540 2541 off = re_load_imm_any(nfp_prog, range_start, imm_b(nfp_prog)); 2542 src_base = reg_a(meta->insn.src_reg * 2); 2543 len = range_end - range_start; 2544 xfer_num = round_up(len, REG_WIDTH) / REG_WIDTH; 2545 2546 indir = len > 8 * REG_WIDTH; 2547 /* Setup PREV_ALU for indirect mode. */ 2548 if (indir) 2549 wrp_immed(nfp_prog, reg_none(), 2550 CMD_OVE_LEN | FIELD_PREP(CMD_OV_LEN, xfer_num - 1)); 2551 2552 /* Cache memory into transfer-in registers. */ 2553 emit_cmd_any(nfp_prog, CMD_TGT_READ32_SWAP, CMD_MODE_32b, 0, src_base, 2554 off, xfer_num - 1, CMD_CTX_SWAP, indir); 2555 } 2556 2557 static int 2558 mem_ldx_data_from_pktcache_unaligned(struct nfp_prog *nfp_prog, 2559 struct nfp_insn_meta *meta, 2560 unsigned int size) 2561 { 2562 s16 range_start = meta->pkt_cache.range_start; 2563 s16 insn_off = meta->insn.off - range_start; 2564 swreg dst_lo, dst_hi, src_lo, src_mid; 2565 u8 dst_gpr = meta->insn.dst_reg * 2; 2566 u8 len_lo = size, len_mid = 0; 2567 u8 idx = insn_off / REG_WIDTH; 2568 u8 off = insn_off % REG_WIDTH; 2569 2570 dst_hi = reg_both(dst_gpr + 1); 2571 dst_lo = reg_both(dst_gpr); 2572 src_lo = reg_xfer(idx); 2573 2574 /* The read length could involve as many as three registers. */ 2575 if (size > REG_WIDTH - off) { 2576 /* Calculate the part in the second register. */ 2577 len_lo = REG_WIDTH - off; 2578 len_mid = size - len_lo; 2579 2580 /* Calculate the part in the third register. */ 2581 if (size > 2 * REG_WIDTH - off) 2582 len_mid = REG_WIDTH; 2583 } 2584 2585 wrp_reg_subpart(nfp_prog, dst_lo, src_lo, len_lo, off); 2586 2587 if (!len_mid) { 2588 wrp_immed(nfp_prog, dst_hi, 0); 2589 return 0; 2590 } 2591 2592 src_mid = reg_xfer(idx + 1); 2593 2594 if (size <= REG_WIDTH) { 2595 wrp_reg_or_subpart(nfp_prog, dst_lo, src_mid, len_mid, len_lo); 2596 wrp_immed(nfp_prog, dst_hi, 0); 2597 } else { 2598 swreg src_hi = reg_xfer(idx + 2); 2599 2600 wrp_reg_or_subpart(nfp_prog, dst_lo, src_mid, 2601 REG_WIDTH - len_lo, len_lo); 2602 wrp_reg_subpart(nfp_prog, dst_hi, src_mid, len_lo, 2603 REG_WIDTH - len_lo); 2604 wrp_reg_or_subpart(nfp_prog, dst_hi, src_hi, REG_WIDTH - len_lo, 2605 len_lo); 2606 } 2607 2608 return 0; 2609 } 2610 2611 static int 2612 mem_ldx_data_from_pktcache_aligned(struct nfp_prog *nfp_prog, 2613 struct nfp_insn_meta *meta, 2614 unsigned int size) 2615 { 2616 swreg dst_lo, dst_hi, src_lo; 2617 u8 dst_gpr, idx; 2618 2619 idx = (meta->insn.off - meta->pkt_cache.range_start) / REG_WIDTH; 2620 dst_gpr = meta->insn.dst_reg * 2; 2621 dst_hi = reg_both(dst_gpr + 1); 2622 dst_lo = reg_both(dst_gpr); 2623 src_lo = reg_xfer(idx); 2624 2625 if (size < REG_WIDTH) { 2626 wrp_reg_subpart(nfp_prog, dst_lo, src_lo, size, 0); 2627 wrp_immed(nfp_prog, dst_hi, 0); 2628 } else if (size == REG_WIDTH) { 2629 wrp_mov(nfp_prog, dst_lo, src_lo); 2630 wrp_immed(nfp_prog, dst_hi, 0); 2631 } else { 2632 swreg src_hi = reg_xfer(idx + 1); 2633 2634 wrp_mov(nfp_prog, dst_lo, src_lo); 2635 wrp_mov(nfp_prog, dst_hi, src_hi); 2636 } 2637 2638 return 0; 2639 } 2640 2641 static int 2642 mem_ldx_data_from_pktcache(struct nfp_prog *nfp_prog, 2643 struct nfp_insn_meta *meta, unsigned int size) 2644 { 2645 u8 off = meta->insn.off - meta->pkt_cache.range_start; 2646 2647 if (IS_ALIGNED(off, REG_WIDTH)) 2648 return mem_ldx_data_from_pktcache_aligned(nfp_prog, meta, size); 2649 2650 return mem_ldx_data_from_pktcache_unaligned(nfp_prog, meta, size); 2651 } 2652 2653 static int 2654 mem_ldx(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2655 unsigned int size) 2656 { 2657 if (meta->ldst_gather_len) 2658 return nfp_cpp_memcpy(nfp_prog, meta); 2659 2660 if (meta->ptr.type == PTR_TO_CTX) { 2661 if (nfp_prog->type == BPF_PROG_TYPE_XDP) 2662 return mem_ldx_xdp(nfp_prog, meta, size); 2663 else 2664 return mem_ldx_skb(nfp_prog, meta, size); 2665 } 2666 2667 if (meta->ptr.type == PTR_TO_PACKET) { 2668 if (meta->pkt_cache.range_end) { 2669 if (meta->pkt_cache.do_init) 2670 mem_ldx_data_init_pktcache(nfp_prog, meta); 2671 2672 return mem_ldx_data_from_pktcache(nfp_prog, meta, size); 2673 } else { 2674 return mem_ldx_data(nfp_prog, meta, size); 2675 } 2676 } 2677 2678 if (meta->ptr.type == PTR_TO_STACK) 2679 return mem_ldx_stack(nfp_prog, meta, size, 2680 meta->ptr.off + meta->ptr.var_off.value); 2681 2682 if (meta->ptr.type == PTR_TO_MAP_VALUE) 2683 return mem_ldx_emem(nfp_prog, meta, size); 2684 2685 return -EOPNOTSUPP; 2686 } 2687 2688 static int mem_ldx1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2689 { 2690 return mem_ldx(nfp_prog, meta, 1); 2691 } 2692 2693 static int mem_ldx2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2694 { 2695 return mem_ldx(nfp_prog, meta, 2); 2696 } 2697 2698 static int mem_ldx4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2699 { 2700 return mem_ldx(nfp_prog, meta, 4); 2701 } 2702 2703 static int mem_ldx8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2704 { 2705 return mem_ldx(nfp_prog, meta, 8); 2706 } 2707 2708 static int 2709 mem_st_data(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2710 unsigned int size) 2711 { 2712 u64 imm = meta->insn.imm; /* sign extend */ 2713 swreg off_reg; 2714 2715 off_reg = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog)); 2716 2717 return data_st_host_order(nfp_prog, meta->insn.dst_reg * 2, off_reg, 2718 imm, size); 2719 } 2720 2721 static int mem_st(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2722 unsigned int size) 2723 { 2724 if (meta->ptr.type == PTR_TO_PACKET) 2725 return mem_st_data(nfp_prog, meta, size); 2726 2727 return -EOPNOTSUPP; 2728 } 2729 2730 static int mem_st1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2731 { 2732 return mem_st(nfp_prog, meta, 1); 2733 } 2734 2735 static int mem_st2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2736 { 2737 return mem_st(nfp_prog, meta, 2); 2738 } 2739 2740 static int mem_st4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2741 { 2742 return mem_st(nfp_prog, meta, 4); 2743 } 2744 2745 static int mem_st8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2746 { 2747 return mem_st(nfp_prog, meta, 8); 2748 } 2749 2750 static int 2751 mem_stx_data(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2752 unsigned int size) 2753 { 2754 swreg off_reg; 2755 2756 off_reg = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog)); 2757 2758 return data_stx_host_order(nfp_prog, meta->insn.dst_reg * 2, off_reg, 2759 meta->insn.src_reg * 2, size); 2760 } 2761 2762 static int 2763 mem_stx_stack(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2764 unsigned int size, unsigned int ptr_off) 2765 { 2766 return mem_op_stack(nfp_prog, meta, size, ptr_off, 2767 meta->insn.src_reg * 2, meta->insn.dst_reg * 2, 2768 false, wrp_lmem_store); 2769 } 2770 2771 static int mem_stx_xdp(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2772 { 2773 switch (meta->insn.off) { 2774 case offsetof(struct xdp_md, rx_queue_index): 2775 return nfp_queue_select(nfp_prog, meta); 2776 } 2777 2778 WARN_ON_ONCE(1); /* verifier should have rejected bad accesses */ 2779 return -EOPNOTSUPP; 2780 } 2781 2782 static int 2783 mem_stx(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2784 unsigned int size) 2785 { 2786 if (meta->ptr.type == PTR_TO_PACKET) 2787 return mem_stx_data(nfp_prog, meta, size); 2788 2789 if (meta->ptr.type == PTR_TO_STACK) 2790 return mem_stx_stack(nfp_prog, meta, size, 2791 meta->ptr.off + meta->ptr.var_off.value); 2792 2793 return -EOPNOTSUPP; 2794 } 2795 2796 static int mem_stx1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2797 { 2798 return mem_stx(nfp_prog, meta, 1); 2799 } 2800 2801 static int mem_stx2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2802 { 2803 return mem_stx(nfp_prog, meta, 2); 2804 } 2805 2806 static int mem_stx4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2807 { 2808 if (meta->ptr.type == PTR_TO_CTX) 2809 if (nfp_prog->type == BPF_PROG_TYPE_XDP) 2810 return mem_stx_xdp(nfp_prog, meta); 2811 return mem_stx(nfp_prog, meta, 4); 2812 } 2813 2814 static int mem_stx8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2815 { 2816 return mem_stx(nfp_prog, meta, 8); 2817 } 2818 2819 static int 2820 mem_xadd(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, bool is64) 2821 { 2822 u8 dst_gpr = meta->insn.dst_reg * 2; 2823 u8 src_gpr = meta->insn.src_reg * 2; 2824 unsigned int full_add, out; 2825 swreg addra, addrb, off; 2826 2827 off = ur_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog)); 2828 2829 /* We can fit 16 bits into command immediate, if we know the immediate 2830 * is guaranteed to either always or never fit into 16 bit we only 2831 * generate code to handle that particular case, otherwise generate 2832 * code for both. 2833 */ 2834 out = nfp_prog_current_offset(nfp_prog); 2835 full_add = nfp_prog_current_offset(nfp_prog); 2836 2837 if (meta->insn.off) { 2838 out += 2; 2839 full_add += 2; 2840 } 2841 if (meta->xadd_maybe_16bit) { 2842 out += 3; 2843 full_add += 3; 2844 } 2845 if (meta->xadd_over_16bit) 2846 out += 2 + is64; 2847 if (meta->xadd_maybe_16bit && meta->xadd_over_16bit) { 2848 out += 5; 2849 full_add += 5; 2850 } 2851 2852 /* Generate the branch for choosing add_imm vs add */ 2853 if (meta->xadd_maybe_16bit && meta->xadd_over_16bit) { 2854 swreg max_imm = imm_a(nfp_prog); 2855 2856 wrp_immed(nfp_prog, max_imm, 0xffff); 2857 emit_alu(nfp_prog, reg_none(), 2858 max_imm, ALU_OP_SUB, reg_b(src_gpr)); 2859 emit_alu(nfp_prog, reg_none(), 2860 reg_imm(0), ALU_OP_SUB_C, reg_b(src_gpr + 1)); 2861 emit_br(nfp_prog, BR_BLO, full_add, meta->insn.off ? 2 : 0); 2862 /* defer for add */ 2863 } 2864 2865 /* If insn has an offset add to the address */ 2866 if (!meta->insn.off) { 2867 addra = reg_a(dst_gpr); 2868 addrb = reg_b(dst_gpr + 1); 2869 } else { 2870 emit_alu(nfp_prog, imma_a(nfp_prog), 2871 reg_a(dst_gpr), ALU_OP_ADD, off); 2872 emit_alu(nfp_prog, imma_b(nfp_prog), 2873 reg_a(dst_gpr + 1), ALU_OP_ADD_C, reg_imm(0)); 2874 addra = imma_a(nfp_prog); 2875 addrb = imma_b(nfp_prog); 2876 } 2877 2878 /* Generate the add_imm if 16 bits are possible */ 2879 if (meta->xadd_maybe_16bit) { 2880 swreg prev_alu = imm_a(nfp_prog); 2881 2882 wrp_immed(nfp_prog, prev_alu, 2883 FIELD_PREP(CMD_OVE_DATA, 2) | 2884 CMD_OVE_LEN | 2885 FIELD_PREP(CMD_OV_LEN, 0x8 | is64 << 2)); 2886 wrp_reg_or_subpart(nfp_prog, prev_alu, reg_b(src_gpr), 2, 2); 2887 emit_cmd_indir(nfp_prog, CMD_TGT_ADD_IMM, CMD_MODE_40b_BA, 0, 2888 addra, addrb, 0, CMD_CTX_NO_SWAP); 2889 2890 if (meta->xadd_over_16bit) 2891 emit_br(nfp_prog, BR_UNC, out, 0); 2892 } 2893 2894 if (!nfp_prog_confirm_current_offset(nfp_prog, full_add)) 2895 return -EINVAL; 2896 2897 /* Generate the add if 16 bits are not guaranteed */ 2898 if (meta->xadd_over_16bit) { 2899 emit_cmd(nfp_prog, CMD_TGT_ADD, CMD_MODE_40b_BA, 0, 2900 addra, addrb, is64 << 2, 2901 is64 ? CMD_CTX_SWAP_DEFER2 : CMD_CTX_SWAP_DEFER1); 2902 2903 wrp_mov(nfp_prog, reg_xfer(0), reg_a(src_gpr)); 2904 if (is64) 2905 wrp_mov(nfp_prog, reg_xfer(1), reg_a(src_gpr + 1)); 2906 } 2907 2908 if (!nfp_prog_confirm_current_offset(nfp_prog, out)) 2909 return -EINVAL; 2910 2911 return 0; 2912 } 2913 2914 static int mem_xadd4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2915 { 2916 return mem_xadd(nfp_prog, meta, false); 2917 } 2918 2919 static int mem_xadd8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2920 { 2921 return mem_xadd(nfp_prog, meta, true); 2922 } 2923 2924 static int jump(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2925 { 2926 emit_br(nfp_prog, BR_UNC, meta->insn.off, 0); 2927 2928 return 0; 2929 } 2930 2931 static int jeq_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2932 { 2933 const struct bpf_insn *insn = &meta->insn; 2934 u64 imm = insn->imm; /* sign extend */ 2935 swreg or1, or2, tmp_reg; 2936 2937 or1 = reg_a(insn->dst_reg * 2); 2938 or2 = reg_b(insn->dst_reg * 2 + 1); 2939 2940 if (imm & ~0U) { 2941 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog)); 2942 emit_alu(nfp_prog, imm_a(nfp_prog), 2943 reg_a(insn->dst_reg * 2), ALU_OP_XOR, tmp_reg); 2944 or1 = imm_a(nfp_prog); 2945 } 2946 2947 if (imm >> 32) { 2948 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog)); 2949 emit_alu(nfp_prog, imm_b(nfp_prog), 2950 reg_a(insn->dst_reg * 2 + 1), ALU_OP_XOR, tmp_reg); 2951 or2 = imm_b(nfp_prog); 2952 } 2953 2954 emit_alu(nfp_prog, reg_none(), or1, ALU_OP_OR, or2); 2955 emit_br(nfp_prog, BR_BEQ, insn->off, 0); 2956 2957 return 0; 2958 } 2959 2960 static int jset_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2961 { 2962 const struct bpf_insn *insn = &meta->insn; 2963 u64 imm = insn->imm; /* sign extend */ 2964 swreg tmp_reg; 2965 2966 if (!imm) { 2967 meta->skip = true; 2968 return 0; 2969 } 2970 2971 if (imm & ~0U) { 2972 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog)); 2973 emit_alu(nfp_prog, reg_none(), 2974 reg_a(insn->dst_reg * 2), ALU_OP_AND, tmp_reg); 2975 emit_br(nfp_prog, BR_BNE, insn->off, 0); 2976 } 2977 2978 if (imm >> 32) { 2979 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog)); 2980 emit_alu(nfp_prog, reg_none(), 2981 reg_a(insn->dst_reg * 2 + 1), ALU_OP_AND, tmp_reg); 2982 emit_br(nfp_prog, BR_BNE, insn->off, 0); 2983 } 2984 2985 return 0; 2986 } 2987 2988 static int jne_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2989 { 2990 const struct bpf_insn *insn = &meta->insn; 2991 u64 imm = insn->imm; /* sign extend */ 2992 swreg tmp_reg; 2993 2994 if (!imm) { 2995 emit_alu(nfp_prog, reg_none(), reg_a(insn->dst_reg * 2), 2996 ALU_OP_OR, reg_b(insn->dst_reg * 2 + 1)); 2997 emit_br(nfp_prog, BR_BNE, insn->off, 0); 2998 return 0; 2999 } 3000 3001 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog)); 3002 emit_alu(nfp_prog, reg_none(), 3003 reg_a(insn->dst_reg * 2), ALU_OP_XOR, tmp_reg); 3004 emit_br(nfp_prog, BR_BNE, insn->off, 0); 3005 3006 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog)); 3007 emit_alu(nfp_prog, reg_none(), 3008 reg_a(insn->dst_reg * 2 + 1), ALU_OP_XOR, tmp_reg); 3009 emit_br(nfp_prog, BR_BNE, insn->off, 0); 3010 3011 return 0; 3012 } 3013 3014 static int jeq_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 3015 { 3016 const struct bpf_insn *insn = &meta->insn; 3017 3018 emit_alu(nfp_prog, imm_a(nfp_prog), reg_a(insn->dst_reg * 2), 3019 ALU_OP_XOR, reg_b(insn->src_reg * 2)); 3020 emit_alu(nfp_prog, imm_b(nfp_prog), reg_a(insn->dst_reg * 2 + 1), 3021 ALU_OP_XOR, reg_b(insn->src_reg * 2 + 1)); 3022 emit_alu(nfp_prog, reg_none(), 3023 imm_a(nfp_prog), ALU_OP_OR, imm_b(nfp_prog)); 3024 emit_br(nfp_prog, BR_BEQ, insn->off, 0); 3025 3026 return 0; 3027 } 3028 3029 static int jset_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 3030 { 3031 return wrp_test_reg(nfp_prog, meta, ALU_OP_AND, BR_BNE); 3032 } 3033 3034 static int jne_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 3035 { 3036 return wrp_test_reg(nfp_prog, meta, ALU_OP_XOR, BR_BNE); 3037 } 3038 3039 static int call(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 3040 { 3041 switch (meta->insn.imm) { 3042 case BPF_FUNC_xdp_adjust_head: 3043 return adjust_head(nfp_prog, meta); 3044 case BPF_FUNC_map_lookup_elem: 3045 case BPF_FUNC_map_update_elem: 3046 case BPF_FUNC_map_delete_elem: 3047 return map_call_stack_common(nfp_prog, meta); 3048 case BPF_FUNC_get_prandom_u32: 3049 return nfp_get_prandom_u32(nfp_prog, meta); 3050 case BPF_FUNC_perf_event_output: 3051 return nfp_perf_event_output(nfp_prog, meta); 3052 default: 3053 WARN_ONCE(1, "verifier allowed unsupported function\n"); 3054 return -EOPNOTSUPP; 3055 } 3056 } 3057 3058 static int goto_out(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 3059 { 3060 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 0, RELO_BR_GO_OUT); 3061 3062 return 0; 3063 } 3064 3065 static const instr_cb_t instr_cb[256] = { 3066 [BPF_ALU64 | BPF_MOV | BPF_X] = mov_reg64, 3067 [BPF_ALU64 | BPF_MOV | BPF_K] = mov_imm64, 3068 [BPF_ALU64 | BPF_XOR | BPF_X] = xor_reg64, 3069 [BPF_ALU64 | BPF_XOR | BPF_K] = xor_imm64, 3070 [BPF_ALU64 | BPF_AND | BPF_X] = and_reg64, 3071 [BPF_ALU64 | BPF_AND | BPF_K] = and_imm64, 3072 [BPF_ALU64 | BPF_OR | BPF_X] = or_reg64, 3073 [BPF_ALU64 | BPF_OR | BPF_K] = or_imm64, 3074 [BPF_ALU64 | BPF_ADD | BPF_X] = add_reg64, 3075 [BPF_ALU64 | BPF_ADD | BPF_K] = add_imm64, 3076 [BPF_ALU64 | BPF_SUB | BPF_X] = sub_reg64, 3077 [BPF_ALU64 | BPF_SUB | BPF_K] = sub_imm64, 3078 [BPF_ALU64 | BPF_MUL | BPF_X] = mul_reg64, 3079 [BPF_ALU64 | BPF_MUL | BPF_K] = mul_imm64, 3080 [BPF_ALU64 | BPF_DIV | BPF_X] = div_reg64, 3081 [BPF_ALU64 | BPF_DIV | BPF_K] = div_imm64, 3082 [BPF_ALU64 | BPF_NEG] = neg_reg64, 3083 [BPF_ALU64 | BPF_LSH | BPF_X] = shl_reg64, 3084 [BPF_ALU64 | BPF_LSH | BPF_K] = shl_imm64, 3085 [BPF_ALU64 | BPF_RSH | BPF_X] = shr_reg64, 3086 [BPF_ALU64 | BPF_RSH | BPF_K] = shr_imm64, 3087 [BPF_ALU64 | BPF_ARSH | BPF_X] = ashr_reg64, 3088 [BPF_ALU64 | BPF_ARSH | BPF_K] = ashr_imm64, 3089 [BPF_ALU | BPF_MOV | BPF_X] = mov_reg, 3090 [BPF_ALU | BPF_MOV | BPF_K] = mov_imm, 3091 [BPF_ALU | BPF_XOR | BPF_X] = xor_reg, 3092 [BPF_ALU | BPF_XOR | BPF_K] = xor_imm, 3093 [BPF_ALU | BPF_AND | BPF_X] = and_reg, 3094 [BPF_ALU | BPF_AND | BPF_K] = and_imm, 3095 [BPF_ALU | BPF_OR | BPF_X] = or_reg, 3096 [BPF_ALU | BPF_OR | BPF_K] = or_imm, 3097 [BPF_ALU | BPF_ADD | BPF_X] = add_reg, 3098 [BPF_ALU | BPF_ADD | BPF_K] = add_imm, 3099 [BPF_ALU | BPF_SUB | BPF_X] = sub_reg, 3100 [BPF_ALU | BPF_SUB | BPF_K] = sub_imm, 3101 [BPF_ALU | BPF_MUL | BPF_X] = mul_reg, 3102 [BPF_ALU | BPF_MUL | BPF_K] = mul_imm, 3103 [BPF_ALU | BPF_DIV | BPF_X] = div_reg, 3104 [BPF_ALU | BPF_DIV | BPF_K] = div_imm, 3105 [BPF_ALU | BPF_NEG] = neg_reg, 3106 [BPF_ALU | BPF_LSH | BPF_K] = shl_imm, 3107 [BPF_ALU | BPF_END | BPF_X] = end_reg32, 3108 [BPF_LD | BPF_IMM | BPF_DW] = imm_ld8, 3109 [BPF_LD | BPF_ABS | BPF_B] = data_ld1, 3110 [BPF_LD | BPF_ABS | BPF_H] = data_ld2, 3111 [BPF_LD | BPF_ABS | BPF_W] = data_ld4, 3112 [BPF_LD | BPF_IND | BPF_B] = data_ind_ld1, 3113 [BPF_LD | BPF_IND | BPF_H] = data_ind_ld2, 3114 [BPF_LD | BPF_IND | BPF_W] = data_ind_ld4, 3115 [BPF_LDX | BPF_MEM | BPF_B] = mem_ldx1, 3116 [BPF_LDX | BPF_MEM | BPF_H] = mem_ldx2, 3117 [BPF_LDX | BPF_MEM | BPF_W] = mem_ldx4, 3118 [BPF_LDX | BPF_MEM | BPF_DW] = mem_ldx8, 3119 [BPF_STX | BPF_MEM | BPF_B] = mem_stx1, 3120 [BPF_STX | BPF_MEM | BPF_H] = mem_stx2, 3121 [BPF_STX | BPF_MEM | BPF_W] = mem_stx4, 3122 [BPF_STX | BPF_MEM | BPF_DW] = mem_stx8, 3123 [BPF_STX | BPF_XADD | BPF_W] = mem_xadd4, 3124 [BPF_STX | BPF_XADD | BPF_DW] = mem_xadd8, 3125 [BPF_ST | BPF_MEM | BPF_B] = mem_st1, 3126 [BPF_ST | BPF_MEM | BPF_H] = mem_st2, 3127 [BPF_ST | BPF_MEM | BPF_W] = mem_st4, 3128 [BPF_ST | BPF_MEM | BPF_DW] = mem_st8, 3129 [BPF_JMP | BPF_JA | BPF_K] = jump, 3130 [BPF_JMP | BPF_JEQ | BPF_K] = jeq_imm, 3131 [BPF_JMP | BPF_JGT | BPF_K] = cmp_imm, 3132 [BPF_JMP | BPF_JGE | BPF_K] = cmp_imm, 3133 [BPF_JMP | BPF_JLT | BPF_K] = cmp_imm, 3134 [BPF_JMP | BPF_JLE | BPF_K] = cmp_imm, 3135 [BPF_JMP | BPF_JSGT | BPF_K] = cmp_imm, 3136 [BPF_JMP | BPF_JSGE | BPF_K] = cmp_imm, 3137 [BPF_JMP | BPF_JSLT | BPF_K] = cmp_imm, 3138 [BPF_JMP | BPF_JSLE | BPF_K] = cmp_imm, 3139 [BPF_JMP | BPF_JSET | BPF_K] = jset_imm, 3140 [BPF_JMP | BPF_JNE | BPF_K] = jne_imm, 3141 [BPF_JMP | BPF_JEQ | BPF_X] = jeq_reg, 3142 [BPF_JMP | BPF_JGT | BPF_X] = cmp_reg, 3143 [BPF_JMP | BPF_JGE | BPF_X] = cmp_reg, 3144 [BPF_JMP | BPF_JLT | BPF_X] = cmp_reg, 3145 [BPF_JMP | BPF_JLE | BPF_X] = cmp_reg, 3146 [BPF_JMP | BPF_JSGT | BPF_X] = cmp_reg, 3147 [BPF_JMP | BPF_JSGE | BPF_X] = cmp_reg, 3148 [BPF_JMP | BPF_JSLT | BPF_X] = cmp_reg, 3149 [BPF_JMP | BPF_JSLE | BPF_X] = cmp_reg, 3150 [BPF_JMP | BPF_JSET | BPF_X] = jset_reg, 3151 [BPF_JMP | BPF_JNE | BPF_X] = jne_reg, 3152 [BPF_JMP | BPF_CALL] = call, 3153 [BPF_JMP | BPF_EXIT] = goto_out, 3154 }; 3155 3156 /* --- Assembler logic --- */ 3157 static int nfp_fixup_branches(struct nfp_prog *nfp_prog) 3158 { 3159 struct nfp_insn_meta *meta, *jmp_dst; 3160 u32 idx, br_idx; 3161 3162 list_for_each_entry(meta, &nfp_prog->insns, l) { 3163 if (meta->skip) 3164 continue; 3165 if (meta->insn.code == (BPF_JMP | BPF_CALL)) 3166 continue; 3167 if (BPF_CLASS(meta->insn.code) != BPF_JMP) 3168 continue; 3169 3170 if (list_is_last(&meta->l, &nfp_prog->insns)) 3171 br_idx = nfp_prog->last_bpf_off; 3172 else 3173 br_idx = list_next_entry(meta, l)->off - 1; 3174 3175 if (!nfp_is_br(nfp_prog->prog[br_idx])) { 3176 pr_err("Fixup found block not ending in branch %d %02x %016llx!!\n", 3177 br_idx, meta->insn.code, nfp_prog->prog[br_idx]); 3178 return -ELOOP; 3179 } 3180 /* Leave special branches for later */ 3181 if (FIELD_GET(OP_RELO_TYPE, nfp_prog->prog[br_idx]) != 3182 RELO_BR_REL) 3183 continue; 3184 3185 if (!meta->jmp_dst) { 3186 pr_err("Non-exit jump doesn't have destination info recorded!!\n"); 3187 return -ELOOP; 3188 } 3189 3190 jmp_dst = meta->jmp_dst; 3191 3192 if (jmp_dst->skip) { 3193 pr_err("Branch landing on removed instruction!!\n"); 3194 return -ELOOP; 3195 } 3196 3197 for (idx = meta->off; idx <= br_idx; idx++) { 3198 if (!nfp_is_br(nfp_prog->prog[idx])) 3199 continue; 3200 br_set_offset(&nfp_prog->prog[idx], jmp_dst->off); 3201 } 3202 } 3203 3204 return 0; 3205 } 3206 3207 static void nfp_intro(struct nfp_prog *nfp_prog) 3208 { 3209 wrp_immed(nfp_prog, plen_reg(nfp_prog), GENMASK(13, 0)); 3210 emit_alu(nfp_prog, plen_reg(nfp_prog), 3211 plen_reg(nfp_prog), ALU_OP_AND, pv_len(nfp_prog)); 3212 } 3213 3214 static void nfp_outro_tc_da(struct nfp_prog *nfp_prog) 3215 { 3216 /* TC direct-action mode: 3217 * 0,1 ok NOT SUPPORTED[1] 3218 * 2 drop 0x22 -> drop, count as stat1 3219 * 4,5 nuke 0x02 -> drop 3220 * 7 redir 0x44 -> redir, count as stat2 3221 * * unspec 0x11 -> pass, count as stat0 3222 * 3223 * [1] We can't support OK and RECLASSIFY because we can't tell TC 3224 * the exact decision made. We are forced to support UNSPEC 3225 * to handle aborts so that's the only one we handle for passing 3226 * packets up the stack. 3227 */ 3228 /* Target for aborts */ 3229 nfp_prog->tgt_abort = nfp_prog_current_offset(nfp_prog); 3230 3231 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 2, RELO_BR_NEXT_PKT); 3232 3233 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS); 3234 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_imm(0x11), SHF_SC_L_SHF, 16); 3235 3236 /* Target for normal exits */ 3237 nfp_prog->tgt_out = nfp_prog_current_offset(nfp_prog); 3238 3239 /* if R0 > 7 jump to abort */ 3240 emit_alu(nfp_prog, reg_none(), reg_imm(7), ALU_OP_SUB, reg_b(0)); 3241 emit_br(nfp_prog, BR_BLO, nfp_prog->tgt_abort, 0); 3242 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS); 3243 3244 wrp_immed(nfp_prog, reg_b(2), 0x41221211); 3245 wrp_immed(nfp_prog, reg_b(3), 0x41001211); 3246 3247 emit_shf(nfp_prog, reg_a(1), 3248 reg_none(), SHF_OP_NONE, reg_b(0), SHF_SC_L_SHF, 2); 3249 3250 emit_alu(nfp_prog, reg_none(), reg_a(1), ALU_OP_OR, reg_imm(0)); 3251 emit_shf(nfp_prog, reg_a(2), 3252 reg_imm(0xf), SHF_OP_AND, reg_b(2), SHF_SC_R_SHF, 0); 3253 3254 emit_alu(nfp_prog, reg_none(), reg_a(1), ALU_OP_OR, reg_imm(0)); 3255 emit_shf(nfp_prog, reg_b(2), 3256 reg_imm(0xf), SHF_OP_AND, reg_b(3), SHF_SC_R_SHF, 0); 3257 3258 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 2, RELO_BR_NEXT_PKT); 3259 3260 emit_shf(nfp_prog, reg_b(2), 3261 reg_a(2), SHF_OP_OR, reg_b(2), SHF_SC_L_SHF, 4); 3262 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_b(2), SHF_SC_L_SHF, 16); 3263 } 3264 3265 static void nfp_outro_xdp(struct nfp_prog *nfp_prog) 3266 { 3267 /* XDP return codes: 3268 * 0 aborted 0x82 -> drop, count as stat3 3269 * 1 drop 0x22 -> drop, count as stat1 3270 * 2 pass 0x11 -> pass, count as stat0 3271 * 3 tx 0x44 -> redir, count as stat2 3272 * * unknown 0x82 -> drop, count as stat3 3273 */ 3274 /* Target for aborts */ 3275 nfp_prog->tgt_abort = nfp_prog_current_offset(nfp_prog); 3276 3277 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 2, RELO_BR_NEXT_PKT); 3278 3279 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS); 3280 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_imm(0x82), SHF_SC_L_SHF, 16); 3281 3282 /* Target for normal exits */ 3283 nfp_prog->tgt_out = nfp_prog_current_offset(nfp_prog); 3284 3285 /* if R0 > 3 jump to abort */ 3286 emit_alu(nfp_prog, reg_none(), reg_imm(3), ALU_OP_SUB, reg_b(0)); 3287 emit_br(nfp_prog, BR_BLO, nfp_prog->tgt_abort, 0); 3288 3289 wrp_immed(nfp_prog, reg_b(2), 0x44112282); 3290 3291 emit_shf(nfp_prog, reg_a(1), 3292 reg_none(), SHF_OP_NONE, reg_b(0), SHF_SC_L_SHF, 3); 3293 3294 emit_alu(nfp_prog, reg_none(), reg_a(1), ALU_OP_OR, reg_imm(0)); 3295 emit_shf(nfp_prog, reg_b(2), 3296 reg_imm(0xff), SHF_OP_AND, reg_b(2), SHF_SC_R_SHF, 0); 3297 3298 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 2, RELO_BR_NEXT_PKT); 3299 3300 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS); 3301 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_b(2), SHF_SC_L_SHF, 16); 3302 } 3303 3304 static void nfp_outro(struct nfp_prog *nfp_prog) 3305 { 3306 switch (nfp_prog->type) { 3307 case BPF_PROG_TYPE_SCHED_CLS: 3308 nfp_outro_tc_da(nfp_prog); 3309 break; 3310 case BPF_PROG_TYPE_XDP: 3311 nfp_outro_xdp(nfp_prog); 3312 break; 3313 default: 3314 WARN_ON(1); 3315 } 3316 } 3317 3318 static int nfp_translate(struct nfp_prog *nfp_prog) 3319 { 3320 struct nfp_insn_meta *meta; 3321 int err; 3322 3323 nfp_intro(nfp_prog); 3324 if (nfp_prog->error) 3325 return nfp_prog->error; 3326 3327 list_for_each_entry(meta, &nfp_prog->insns, l) { 3328 instr_cb_t cb = instr_cb[meta->insn.code]; 3329 3330 meta->off = nfp_prog_current_offset(nfp_prog); 3331 3332 if (meta->skip) { 3333 nfp_prog->n_translated++; 3334 continue; 3335 } 3336 3337 if (nfp_meta_has_prev(nfp_prog, meta) && 3338 nfp_meta_prev(meta)->double_cb) 3339 cb = nfp_meta_prev(meta)->double_cb; 3340 if (!cb) 3341 return -ENOENT; 3342 err = cb(nfp_prog, meta); 3343 if (err) 3344 return err; 3345 if (nfp_prog->error) 3346 return nfp_prog->error; 3347 3348 nfp_prog->n_translated++; 3349 } 3350 3351 nfp_prog->last_bpf_off = nfp_prog_current_offset(nfp_prog) - 1; 3352 3353 nfp_outro(nfp_prog); 3354 if (nfp_prog->error) 3355 return nfp_prog->error; 3356 3357 wrp_nops(nfp_prog, NFP_USTORE_PREFETCH_WINDOW); 3358 if (nfp_prog->error) 3359 return nfp_prog->error; 3360 3361 return nfp_fixup_branches(nfp_prog); 3362 } 3363 3364 /* --- Optimizations --- */ 3365 static void nfp_bpf_opt_reg_init(struct nfp_prog *nfp_prog) 3366 { 3367 struct nfp_insn_meta *meta; 3368 3369 list_for_each_entry(meta, &nfp_prog->insns, l) { 3370 struct bpf_insn insn = meta->insn; 3371 3372 /* Programs converted from cBPF start with register xoring */ 3373 if (insn.code == (BPF_ALU64 | BPF_XOR | BPF_X) && 3374 insn.src_reg == insn.dst_reg) 3375 continue; 3376 3377 /* Programs start with R6 = R1 but we ignore the skb pointer */ 3378 if (insn.code == (BPF_ALU64 | BPF_MOV | BPF_X) && 3379 insn.src_reg == 1 && insn.dst_reg == 6) 3380 meta->skip = true; 3381 3382 /* Return as soon as something doesn't match */ 3383 if (!meta->skip) 3384 return; 3385 } 3386 } 3387 3388 /* abs(insn.imm) will fit better into unrestricted reg immediate - 3389 * convert add/sub of a negative number into a sub/add of a positive one. 3390 */ 3391 static void nfp_bpf_opt_neg_add_sub(struct nfp_prog *nfp_prog) 3392 { 3393 struct nfp_insn_meta *meta; 3394 3395 list_for_each_entry(meta, &nfp_prog->insns, l) { 3396 struct bpf_insn insn = meta->insn; 3397 3398 if (meta->skip) 3399 continue; 3400 3401 if (BPF_CLASS(insn.code) != BPF_ALU && 3402 BPF_CLASS(insn.code) != BPF_ALU64 && 3403 BPF_CLASS(insn.code) != BPF_JMP) 3404 continue; 3405 if (BPF_SRC(insn.code) != BPF_K) 3406 continue; 3407 if (insn.imm >= 0) 3408 continue; 3409 3410 if (BPF_CLASS(insn.code) == BPF_JMP) { 3411 switch (BPF_OP(insn.code)) { 3412 case BPF_JGE: 3413 case BPF_JSGE: 3414 case BPF_JLT: 3415 case BPF_JSLT: 3416 meta->jump_neg_op = true; 3417 break; 3418 default: 3419 continue; 3420 } 3421 } else { 3422 if (BPF_OP(insn.code) == BPF_ADD) 3423 insn.code = BPF_CLASS(insn.code) | BPF_SUB; 3424 else if (BPF_OP(insn.code) == BPF_SUB) 3425 insn.code = BPF_CLASS(insn.code) | BPF_ADD; 3426 else 3427 continue; 3428 3429 meta->insn.code = insn.code | BPF_K; 3430 } 3431 3432 meta->insn.imm = -insn.imm; 3433 } 3434 } 3435 3436 /* Remove masking after load since our load guarantees this is not needed */ 3437 static void nfp_bpf_opt_ld_mask(struct nfp_prog *nfp_prog) 3438 { 3439 struct nfp_insn_meta *meta1, *meta2; 3440 const s32 exp_mask[] = { 3441 [BPF_B] = 0x000000ffU, 3442 [BPF_H] = 0x0000ffffU, 3443 [BPF_W] = 0xffffffffU, 3444 }; 3445 3446 nfp_for_each_insn_walk2(nfp_prog, meta1, meta2) { 3447 struct bpf_insn insn, next; 3448 3449 insn = meta1->insn; 3450 next = meta2->insn; 3451 3452 if (BPF_CLASS(insn.code) != BPF_LD) 3453 continue; 3454 if (BPF_MODE(insn.code) != BPF_ABS && 3455 BPF_MODE(insn.code) != BPF_IND) 3456 continue; 3457 3458 if (next.code != (BPF_ALU64 | BPF_AND | BPF_K)) 3459 continue; 3460 3461 if (!exp_mask[BPF_SIZE(insn.code)]) 3462 continue; 3463 if (exp_mask[BPF_SIZE(insn.code)] != next.imm) 3464 continue; 3465 3466 if (next.src_reg || next.dst_reg) 3467 continue; 3468 3469 if (meta2->flags & FLAG_INSN_IS_JUMP_DST) 3470 continue; 3471 3472 meta2->skip = true; 3473 } 3474 } 3475 3476 static void nfp_bpf_opt_ld_shift(struct nfp_prog *nfp_prog) 3477 { 3478 struct nfp_insn_meta *meta1, *meta2, *meta3; 3479 3480 nfp_for_each_insn_walk3(nfp_prog, meta1, meta2, meta3) { 3481 struct bpf_insn insn, next1, next2; 3482 3483 insn = meta1->insn; 3484 next1 = meta2->insn; 3485 next2 = meta3->insn; 3486 3487 if (BPF_CLASS(insn.code) != BPF_LD) 3488 continue; 3489 if (BPF_MODE(insn.code) != BPF_ABS && 3490 BPF_MODE(insn.code) != BPF_IND) 3491 continue; 3492 if (BPF_SIZE(insn.code) != BPF_W) 3493 continue; 3494 3495 if (!(next1.code == (BPF_LSH | BPF_K | BPF_ALU64) && 3496 next2.code == (BPF_RSH | BPF_K | BPF_ALU64)) && 3497 !(next1.code == (BPF_RSH | BPF_K | BPF_ALU64) && 3498 next2.code == (BPF_LSH | BPF_K | BPF_ALU64))) 3499 continue; 3500 3501 if (next1.src_reg || next1.dst_reg || 3502 next2.src_reg || next2.dst_reg) 3503 continue; 3504 3505 if (next1.imm != 0x20 || next2.imm != 0x20) 3506 continue; 3507 3508 if (meta2->flags & FLAG_INSN_IS_JUMP_DST || 3509 meta3->flags & FLAG_INSN_IS_JUMP_DST) 3510 continue; 3511 3512 meta2->skip = true; 3513 meta3->skip = true; 3514 } 3515 } 3516 3517 /* load/store pair that forms memory copy sould look like the following: 3518 * 3519 * ld_width R, [addr_src + offset_src] 3520 * st_width [addr_dest + offset_dest], R 3521 * 3522 * The destination register of load and source register of store should 3523 * be the same, load and store should also perform at the same width. 3524 * If either of addr_src or addr_dest is stack pointer, we don't do the 3525 * CPP optimization as stack is modelled by registers on NFP. 3526 */ 3527 static bool 3528 curr_pair_is_memcpy(struct nfp_insn_meta *ld_meta, 3529 struct nfp_insn_meta *st_meta) 3530 { 3531 struct bpf_insn *ld = &ld_meta->insn; 3532 struct bpf_insn *st = &st_meta->insn; 3533 3534 if (!is_mbpf_load(ld_meta) || !is_mbpf_store(st_meta)) 3535 return false; 3536 3537 if (ld_meta->ptr.type != PTR_TO_PACKET && 3538 ld_meta->ptr.type != PTR_TO_MAP_VALUE) 3539 return false; 3540 3541 if (st_meta->ptr.type != PTR_TO_PACKET) 3542 return false; 3543 3544 if (BPF_SIZE(ld->code) != BPF_SIZE(st->code)) 3545 return false; 3546 3547 if (ld->dst_reg != st->src_reg) 3548 return false; 3549 3550 /* There is jump to the store insn in this pair. */ 3551 if (st_meta->flags & FLAG_INSN_IS_JUMP_DST) 3552 return false; 3553 3554 return true; 3555 } 3556 3557 /* Currently, we only support chaining load/store pairs if: 3558 * 3559 * - Their address base registers are the same. 3560 * - Their address offsets are in the same order. 3561 * - They operate at the same memory width. 3562 * - There is no jump into the middle of them. 3563 */ 3564 static bool 3565 curr_pair_chain_with_previous(struct nfp_insn_meta *ld_meta, 3566 struct nfp_insn_meta *st_meta, 3567 struct bpf_insn *prev_ld, 3568 struct bpf_insn *prev_st) 3569 { 3570 u8 prev_size, curr_size, prev_ld_base, prev_st_base, prev_ld_dst; 3571 struct bpf_insn *ld = &ld_meta->insn; 3572 struct bpf_insn *st = &st_meta->insn; 3573 s16 prev_ld_off, prev_st_off; 3574 3575 /* This pair is the start pair. */ 3576 if (!prev_ld) 3577 return true; 3578 3579 prev_size = BPF_LDST_BYTES(prev_ld); 3580 curr_size = BPF_LDST_BYTES(ld); 3581 prev_ld_base = prev_ld->src_reg; 3582 prev_st_base = prev_st->dst_reg; 3583 prev_ld_dst = prev_ld->dst_reg; 3584 prev_ld_off = prev_ld->off; 3585 prev_st_off = prev_st->off; 3586 3587 if (ld->dst_reg != prev_ld_dst) 3588 return false; 3589 3590 if (ld->src_reg != prev_ld_base || st->dst_reg != prev_st_base) 3591 return false; 3592 3593 if (curr_size != prev_size) 3594 return false; 3595 3596 /* There is jump to the head of this pair. */ 3597 if (ld_meta->flags & FLAG_INSN_IS_JUMP_DST) 3598 return false; 3599 3600 /* Both in ascending order. */ 3601 if (prev_ld_off + prev_size == ld->off && 3602 prev_st_off + prev_size == st->off) 3603 return true; 3604 3605 /* Both in descending order. */ 3606 if (ld->off + curr_size == prev_ld_off && 3607 st->off + curr_size == prev_st_off) 3608 return true; 3609 3610 return false; 3611 } 3612 3613 /* Return TRUE if cross memory access happens. Cross memory access means 3614 * store area is overlapping with load area that a later load might load 3615 * the value from previous store, for this case we can't treat the sequence 3616 * as an memory copy. 3617 */ 3618 static bool 3619 cross_mem_access(struct bpf_insn *ld, struct nfp_insn_meta *head_ld_meta, 3620 struct nfp_insn_meta *head_st_meta) 3621 { 3622 s16 head_ld_off, head_st_off, ld_off; 3623 3624 /* Different pointer types does not overlap. */ 3625 if (head_ld_meta->ptr.type != head_st_meta->ptr.type) 3626 return false; 3627 3628 /* load and store are both PTR_TO_PACKET, check ID info. */ 3629 if (head_ld_meta->ptr.id != head_st_meta->ptr.id) 3630 return true; 3631 3632 /* Canonicalize the offsets. Turn all of them against the original 3633 * base register. 3634 */ 3635 head_ld_off = head_ld_meta->insn.off + head_ld_meta->ptr.off; 3636 head_st_off = head_st_meta->insn.off + head_st_meta->ptr.off; 3637 ld_off = ld->off + head_ld_meta->ptr.off; 3638 3639 /* Ascending order cross. */ 3640 if (ld_off > head_ld_off && 3641 head_ld_off < head_st_off && ld_off >= head_st_off) 3642 return true; 3643 3644 /* Descending order cross. */ 3645 if (ld_off < head_ld_off && 3646 head_ld_off > head_st_off && ld_off <= head_st_off) 3647 return true; 3648 3649 return false; 3650 } 3651 3652 /* This pass try to identify the following instructoin sequences. 3653 * 3654 * load R, [regA + offA] 3655 * store [regB + offB], R 3656 * load R, [regA + offA + const_imm_A] 3657 * store [regB + offB + const_imm_A], R 3658 * load R, [regA + offA + 2 * const_imm_A] 3659 * store [regB + offB + 2 * const_imm_A], R 3660 * ... 3661 * 3662 * Above sequence is typically generated by compiler when lowering 3663 * memcpy. NFP prefer using CPP instructions to accelerate it. 3664 */ 3665 static void nfp_bpf_opt_ldst_gather(struct nfp_prog *nfp_prog) 3666 { 3667 struct nfp_insn_meta *head_ld_meta = NULL; 3668 struct nfp_insn_meta *head_st_meta = NULL; 3669 struct nfp_insn_meta *meta1, *meta2; 3670 struct bpf_insn *prev_ld = NULL; 3671 struct bpf_insn *prev_st = NULL; 3672 u8 count = 0; 3673 3674 nfp_for_each_insn_walk2(nfp_prog, meta1, meta2) { 3675 struct bpf_insn *ld = &meta1->insn; 3676 struct bpf_insn *st = &meta2->insn; 3677 3678 /* Reset record status if any of the following if true: 3679 * - The current insn pair is not load/store. 3680 * - The load/store pair doesn't chain with previous one. 3681 * - The chained load/store pair crossed with previous pair. 3682 * - The chained load/store pair has a total size of memory 3683 * copy beyond 128 bytes which is the maximum length a 3684 * single NFP CPP command can transfer. 3685 */ 3686 if (!curr_pair_is_memcpy(meta1, meta2) || 3687 !curr_pair_chain_with_previous(meta1, meta2, prev_ld, 3688 prev_st) || 3689 (head_ld_meta && (cross_mem_access(ld, head_ld_meta, 3690 head_st_meta) || 3691 head_ld_meta->ldst_gather_len >= 128))) { 3692 if (!count) 3693 continue; 3694 3695 if (count > 1) { 3696 s16 prev_ld_off = prev_ld->off; 3697 s16 prev_st_off = prev_st->off; 3698 s16 head_ld_off = head_ld_meta->insn.off; 3699 3700 if (prev_ld_off < head_ld_off) { 3701 head_ld_meta->insn.off = prev_ld_off; 3702 head_st_meta->insn.off = prev_st_off; 3703 head_ld_meta->ldst_gather_len = 3704 -head_ld_meta->ldst_gather_len; 3705 } 3706 3707 head_ld_meta->paired_st = &head_st_meta->insn; 3708 head_st_meta->skip = true; 3709 } else { 3710 head_ld_meta->ldst_gather_len = 0; 3711 } 3712 3713 /* If the chain is ended by an load/store pair then this 3714 * could serve as the new head of the the next chain. 3715 */ 3716 if (curr_pair_is_memcpy(meta1, meta2)) { 3717 head_ld_meta = meta1; 3718 head_st_meta = meta2; 3719 head_ld_meta->ldst_gather_len = 3720 BPF_LDST_BYTES(ld); 3721 meta1 = nfp_meta_next(meta1); 3722 meta2 = nfp_meta_next(meta2); 3723 prev_ld = ld; 3724 prev_st = st; 3725 count = 1; 3726 } else { 3727 head_ld_meta = NULL; 3728 head_st_meta = NULL; 3729 prev_ld = NULL; 3730 prev_st = NULL; 3731 count = 0; 3732 } 3733 3734 continue; 3735 } 3736 3737 if (!head_ld_meta) { 3738 head_ld_meta = meta1; 3739 head_st_meta = meta2; 3740 } else { 3741 meta1->skip = true; 3742 meta2->skip = true; 3743 } 3744 3745 head_ld_meta->ldst_gather_len += BPF_LDST_BYTES(ld); 3746 meta1 = nfp_meta_next(meta1); 3747 meta2 = nfp_meta_next(meta2); 3748 prev_ld = ld; 3749 prev_st = st; 3750 count++; 3751 } 3752 } 3753 3754 static void nfp_bpf_opt_pkt_cache(struct nfp_prog *nfp_prog) 3755 { 3756 struct nfp_insn_meta *meta, *range_node = NULL; 3757 s16 range_start = 0, range_end = 0; 3758 bool cache_avail = false; 3759 struct bpf_insn *insn; 3760 s32 range_ptr_off = 0; 3761 u32 range_ptr_id = 0; 3762 3763 list_for_each_entry(meta, &nfp_prog->insns, l) { 3764 if (meta->flags & FLAG_INSN_IS_JUMP_DST) 3765 cache_avail = false; 3766 3767 if (meta->skip) 3768 continue; 3769 3770 insn = &meta->insn; 3771 3772 if (is_mbpf_store_pkt(meta) || 3773 insn->code == (BPF_JMP | BPF_CALL) || 3774 is_mbpf_classic_store_pkt(meta) || 3775 is_mbpf_classic_load(meta)) { 3776 cache_avail = false; 3777 continue; 3778 } 3779 3780 if (!is_mbpf_load(meta)) 3781 continue; 3782 3783 if (meta->ptr.type != PTR_TO_PACKET || meta->ldst_gather_len) { 3784 cache_avail = false; 3785 continue; 3786 } 3787 3788 if (!cache_avail) { 3789 cache_avail = true; 3790 if (range_node) 3791 goto end_current_then_start_new; 3792 goto start_new; 3793 } 3794 3795 /* Check ID to make sure two reads share the same 3796 * variable offset against PTR_TO_PACKET, and check OFF 3797 * to make sure they also share the same constant 3798 * offset. 3799 * 3800 * OFFs don't really need to be the same, because they 3801 * are the constant offsets against PTR_TO_PACKET, so 3802 * for different OFFs, we could canonicalize them to 3803 * offsets against original packet pointer. We don't 3804 * support this. 3805 */ 3806 if (meta->ptr.id == range_ptr_id && 3807 meta->ptr.off == range_ptr_off) { 3808 s16 new_start = range_start; 3809 s16 end, off = insn->off; 3810 s16 new_end = range_end; 3811 bool changed = false; 3812 3813 if (off < range_start) { 3814 new_start = off; 3815 changed = true; 3816 } 3817 3818 end = off + BPF_LDST_BYTES(insn); 3819 if (end > range_end) { 3820 new_end = end; 3821 changed = true; 3822 } 3823 3824 if (!changed) 3825 continue; 3826 3827 if (new_end - new_start <= 64) { 3828 /* Install new range. */ 3829 range_start = new_start; 3830 range_end = new_end; 3831 continue; 3832 } 3833 } 3834 3835 end_current_then_start_new: 3836 range_node->pkt_cache.range_start = range_start; 3837 range_node->pkt_cache.range_end = range_end; 3838 start_new: 3839 range_node = meta; 3840 range_node->pkt_cache.do_init = true; 3841 range_ptr_id = range_node->ptr.id; 3842 range_ptr_off = range_node->ptr.off; 3843 range_start = insn->off; 3844 range_end = insn->off + BPF_LDST_BYTES(insn); 3845 } 3846 3847 if (range_node) { 3848 range_node->pkt_cache.range_start = range_start; 3849 range_node->pkt_cache.range_end = range_end; 3850 } 3851 3852 list_for_each_entry(meta, &nfp_prog->insns, l) { 3853 if (meta->skip) 3854 continue; 3855 3856 if (is_mbpf_load_pkt(meta) && !meta->ldst_gather_len) { 3857 if (meta->pkt_cache.do_init) { 3858 range_start = meta->pkt_cache.range_start; 3859 range_end = meta->pkt_cache.range_end; 3860 } else { 3861 meta->pkt_cache.range_start = range_start; 3862 meta->pkt_cache.range_end = range_end; 3863 } 3864 } 3865 } 3866 } 3867 3868 static int nfp_bpf_optimize(struct nfp_prog *nfp_prog) 3869 { 3870 nfp_bpf_opt_reg_init(nfp_prog); 3871 3872 nfp_bpf_opt_neg_add_sub(nfp_prog); 3873 nfp_bpf_opt_ld_mask(nfp_prog); 3874 nfp_bpf_opt_ld_shift(nfp_prog); 3875 nfp_bpf_opt_ldst_gather(nfp_prog); 3876 nfp_bpf_opt_pkt_cache(nfp_prog); 3877 3878 return 0; 3879 } 3880 3881 static int nfp_bpf_replace_map_ptrs(struct nfp_prog *nfp_prog) 3882 { 3883 struct nfp_insn_meta *meta1, *meta2; 3884 struct nfp_bpf_map *nfp_map; 3885 struct bpf_map *map; 3886 3887 nfp_for_each_insn_walk2(nfp_prog, meta1, meta2) { 3888 if (meta1->skip || meta2->skip) 3889 continue; 3890 3891 if (meta1->insn.code != (BPF_LD | BPF_IMM | BPF_DW) || 3892 meta1->insn.src_reg != BPF_PSEUDO_MAP_FD) 3893 continue; 3894 3895 map = (void *)(unsigned long)((u32)meta1->insn.imm | 3896 (u64)meta2->insn.imm << 32); 3897 if (bpf_map_offload_neutral(map)) 3898 continue; 3899 nfp_map = map_to_offmap(map)->dev_priv; 3900 3901 meta1->insn.imm = nfp_map->tid; 3902 meta2->insn.imm = 0; 3903 } 3904 3905 return 0; 3906 } 3907 3908 static int nfp_bpf_ustore_calc(u64 *prog, unsigned int len) 3909 { 3910 __le64 *ustore = (__force __le64 *)prog; 3911 int i; 3912 3913 for (i = 0; i < len; i++) { 3914 int err; 3915 3916 err = nfp_ustore_check_valid_no_ecc(prog[i]); 3917 if (err) 3918 return err; 3919 3920 ustore[i] = cpu_to_le64(nfp_ustore_calc_ecc_insn(prog[i])); 3921 } 3922 3923 return 0; 3924 } 3925 3926 static void nfp_bpf_prog_trim(struct nfp_prog *nfp_prog) 3927 { 3928 void *prog; 3929 3930 prog = kvmalloc_array(nfp_prog->prog_len, sizeof(u64), GFP_KERNEL); 3931 if (!prog) 3932 return; 3933 3934 nfp_prog->__prog_alloc_len = nfp_prog->prog_len * sizeof(u64); 3935 memcpy(prog, nfp_prog->prog, nfp_prog->__prog_alloc_len); 3936 kvfree(nfp_prog->prog); 3937 nfp_prog->prog = prog; 3938 } 3939 3940 int nfp_bpf_jit(struct nfp_prog *nfp_prog) 3941 { 3942 int ret; 3943 3944 ret = nfp_bpf_replace_map_ptrs(nfp_prog); 3945 if (ret) 3946 return ret; 3947 3948 ret = nfp_bpf_optimize(nfp_prog); 3949 if (ret) 3950 return ret; 3951 3952 ret = nfp_translate(nfp_prog); 3953 if (ret) { 3954 pr_err("Translation failed with error %d (translated: %u)\n", 3955 ret, nfp_prog->n_translated); 3956 return -EINVAL; 3957 } 3958 3959 nfp_bpf_prog_trim(nfp_prog); 3960 3961 return ret; 3962 } 3963 3964 void nfp_bpf_jit_prepare(struct nfp_prog *nfp_prog, unsigned int cnt) 3965 { 3966 struct nfp_insn_meta *meta; 3967 3968 /* Another pass to record jump information. */ 3969 list_for_each_entry(meta, &nfp_prog->insns, l) { 3970 u64 code = meta->insn.code; 3971 3972 if (BPF_CLASS(code) == BPF_JMP && BPF_OP(code) != BPF_EXIT && 3973 BPF_OP(code) != BPF_CALL) { 3974 struct nfp_insn_meta *dst_meta; 3975 unsigned short dst_indx; 3976 3977 dst_indx = meta->n + 1 + meta->insn.off; 3978 dst_meta = nfp_bpf_goto_meta(nfp_prog, meta, dst_indx, 3979 cnt); 3980 3981 meta->jmp_dst = dst_meta; 3982 dst_meta->flags |= FLAG_INSN_IS_JUMP_DST; 3983 } 3984 } 3985 } 3986 3987 bool nfp_bpf_supported_opcode(u8 code) 3988 { 3989 return !!instr_cb[code]; 3990 } 3991 3992 void *nfp_bpf_relo_for_vnic(struct nfp_prog *nfp_prog, struct nfp_bpf_vnic *bv) 3993 { 3994 unsigned int i; 3995 u64 *prog; 3996 int err; 3997 3998 prog = kmemdup(nfp_prog->prog, nfp_prog->prog_len * sizeof(u64), 3999 GFP_KERNEL); 4000 if (!prog) 4001 return ERR_PTR(-ENOMEM); 4002 4003 for (i = 0; i < nfp_prog->prog_len; i++) { 4004 enum nfp_relo_type special; 4005 u32 val; 4006 4007 special = FIELD_GET(OP_RELO_TYPE, prog[i]); 4008 switch (special) { 4009 case RELO_NONE: 4010 continue; 4011 case RELO_BR_REL: 4012 br_add_offset(&prog[i], bv->start_off); 4013 break; 4014 case RELO_BR_GO_OUT: 4015 br_set_offset(&prog[i], 4016 nfp_prog->tgt_out + bv->start_off); 4017 break; 4018 case RELO_BR_GO_ABORT: 4019 br_set_offset(&prog[i], 4020 nfp_prog->tgt_abort + bv->start_off); 4021 break; 4022 case RELO_BR_NEXT_PKT: 4023 br_set_offset(&prog[i], bv->tgt_done); 4024 break; 4025 case RELO_BR_HELPER: 4026 val = br_get_offset(prog[i]); 4027 val -= BR_OFF_RELO; 4028 switch (val) { 4029 case BPF_FUNC_map_lookup_elem: 4030 val = nfp_prog->bpf->helpers.map_lookup; 4031 break; 4032 case BPF_FUNC_map_update_elem: 4033 val = nfp_prog->bpf->helpers.map_update; 4034 break; 4035 case BPF_FUNC_map_delete_elem: 4036 val = nfp_prog->bpf->helpers.map_delete; 4037 break; 4038 case BPF_FUNC_perf_event_output: 4039 val = nfp_prog->bpf->helpers.perf_event_output; 4040 break; 4041 default: 4042 pr_err("relocation of unknown helper %d\n", 4043 val); 4044 err = -EINVAL; 4045 goto err_free_prog; 4046 } 4047 br_set_offset(&prog[i], val); 4048 break; 4049 case RELO_IMMED_REL: 4050 immed_add_value(&prog[i], bv->start_off); 4051 break; 4052 } 4053 4054 prog[i] &= ~OP_RELO_TYPE; 4055 } 4056 4057 err = nfp_bpf_ustore_calc(prog, nfp_prog->prog_len); 4058 if (err) 4059 goto err_free_prog; 4060 4061 return prog; 4062 4063 err_free_prog: 4064 kfree(prog); 4065 return ERR_PTR(err); 4066 } 4067