1 // SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause) 2 /* Copyright (C) 2016-2018 Netronome Systems, Inc. */ 3 4 #define pr_fmt(fmt) "NFP net bpf: " fmt 5 6 #include <linux/bug.h> 7 #include <linux/bpf.h> 8 #include <linux/filter.h> 9 #include <linux/kernel.h> 10 #include <linux/pkt_cls.h> 11 #include <linux/reciprocal_div.h> 12 #include <linux/unistd.h> 13 14 #include "main.h" 15 #include "../nfp_asm.h" 16 #include "../nfp_net_ctrl.h" 17 18 /* --- NFP prog --- */ 19 /* Foreach "multiple" entries macros provide pos and next<n> pointers. 20 * It's safe to modify the next pointers (but not pos). 21 */ 22 #define nfp_for_each_insn_walk2(nfp_prog, pos, next) \ 23 for (pos = list_first_entry(&(nfp_prog)->insns, typeof(*pos), l), \ 24 next = list_next_entry(pos, l); \ 25 &(nfp_prog)->insns != &pos->l && \ 26 &(nfp_prog)->insns != &next->l; \ 27 pos = nfp_meta_next(pos), \ 28 next = nfp_meta_next(pos)) 29 30 #define nfp_for_each_insn_walk3(nfp_prog, pos, next, next2) \ 31 for (pos = list_first_entry(&(nfp_prog)->insns, typeof(*pos), l), \ 32 next = list_next_entry(pos, l), \ 33 next2 = list_next_entry(next, l); \ 34 &(nfp_prog)->insns != &pos->l && \ 35 &(nfp_prog)->insns != &next->l && \ 36 &(nfp_prog)->insns != &next2->l; \ 37 pos = nfp_meta_next(pos), \ 38 next = nfp_meta_next(pos), \ 39 next2 = nfp_meta_next(next)) 40 41 static bool 42 nfp_meta_has_prev(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 43 { 44 return meta->l.prev != &nfp_prog->insns; 45 } 46 47 static void nfp_prog_push(struct nfp_prog *nfp_prog, u64 insn) 48 { 49 if (nfp_prog->__prog_alloc_len / sizeof(u64) == nfp_prog->prog_len) { 50 pr_warn("instruction limit reached (%u NFP instructions)\n", 51 nfp_prog->prog_len); 52 nfp_prog->error = -ENOSPC; 53 return; 54 } 55 56 nfp_prog->prog[nfp_prog->prog_len] = insn; 57 nfp_prog->prog_len++; 58 } 59 60 static unsigned int nfp_prog_current_offset(struct nfp_prog *nfp_prog) 61 { 62 return nfp_prog->prog_len; 63 } 64 65 static bool 66 nfp_prog_confirm_current_offset(struct nfp_prog *nfp_prog, unsigned int off) 67 { 68 /* If there is a recorded error we may have dropped instructions; 69 * that doesn't have to be due to translator bug, and the translation 70 * will fail anyway, so just return OK. 71 */ 72 if (nfp_prog->error) 73 return true; 74 return !WARN_ON_ONCE(nfp_prog_current_offset(nfp_prog) != off); 75 } 76 77 /* --- Emitters --- */ 78 static void 79 __emit_cmd(struct nfp_prog *nfp_prog, enum cmd_tgt_map op, 80 u8 mode, u8 xfer, u8 areg, u8 breg, u8 size, enum cmd_ctx_swap ctx, 81 bool indir) 82 { 83 u64 insn; 84 85 insn = FIELD_PREP(OP_CMD_A_SRC, areg) | 86 FIELD_PREP(OP_CMD_CTX, ctx) | 87 FIELD_PREP(OP_CMD_B_SRC, breg) | 88 FIELD_PREP(OP_CMD_TOKEN, cmd_tgt_act[op].token) | 89 FIELD_PREP(OP_CMD_XFER, xfer) | 90 FIELD_PREP(OP_CMD_CNT, size) | 91 FIELD_PREP(OP_CMD_SIG, ctx != CMD_CTX_NO_SWAP) | 92 FIELD_PREP(OP_CMD_TGT_CMD, cmd_tgt_act[op].tgt_cmd) | 93 FIELD_PREP(OP_CMD_INDIR, indir) | 94 FIELD_PREP(OP_CMD_MODE, mode); 95 96 nfp_prog_push(nfp_prog, insn); 97 } 98 99 static void 100 emit_cmd_any(struct nfp_prog *nfp_prog, enum cmd_tgt_map op, u8 mode, u8 xfer, 101 swreg lreg, swreg rreg, u8 size, enum cmd_ctx_swap ctx, bool indir) 102 { 103 struct nfp_insn_re_regs reg; 104 int err; 105 106 err = swreg_to_restricted(reg_none(), lreg, rreg, ®, false); 107 if (err) { 108 nfp_prog->error = err; 109 return; 110 } 111 if (reg.swap) { 112 pr_err("cmd can't swap arguments\n"); 113 nfp_prog->error = -EFAULT; 114 return; 115 } 116 if (reg.dst_lmextn || reg.src_lmextn) { 117 pr_err("cmd can't use LMextn\n"); 118 nfp_prog->error = -EFAULT; 119 return; 120 } 121 122 __emit_cmd(nfp_prog, op, mode, xfer, reg.areg, reg.breg, size, ctx, 123 indir); 124 } 125 126 static void 127 emit_cmd(struct nfp_prog *nfp_prog, enum cmd_tgt_map op, u8 mode, u8 xfer, 128 swreg lreg, swreg rreg, u8 size, enum cmd_ctx_swap ctx) 129 { 130 emit_cmd_any(nfp_prog, op, mode, xfer, lreg, rreg, size, ctx, false); 131 } 132 133 static void 134 emit_cmd_indir(struct nfp_prog *nfp_prog, enum cmd_tgt_map op, u8 mode, u8 xfer, 135 swreg lreg, swreg rreg, u8 size, enum cmd_ctx_swap ctx) 136 { 137 emit_cmd_any(nfp_prog, op, mode, xfer, lreg, rreg, size, ctx, true); 138 } 139 140 static void 141 __emit_br(struct nfp_prog *nfp_prog, enum br_mask mask, enum br_ev_pip ev_pip, 142 enum br_ctx_signal_state css, u16 addr, u8 defer) 143 { 144 u16 addr_lo, addr_hi; 145 u64 insn; 146 147 addr_lo = addr & (OP_BR_ADDR_LO >> __bf_shf(OP_BR_ADDR_LO)); 148 addr_hi = addr != addr_lo; 149 150 insn = OP_BR_BASE | 151 FIELD_PREP(OP_BR_MASK, mask) | 152 FIELD_PREP(OP_BR_EV_PIP, ev_pip) | 153 FIELD_PREP(OP_BR_CSS, css) | 154 FIELD_PREP(OP_BR_DEFBR, defer) | 155 FIELD_PREP(OP_BR_ADDR_LO, addr_lo) | 156 FIELD_PREP(OP_BR_ADDR_HI, addr_hi); 157 158 nfp_prog_push(nfp_prog, insn); 159 } 160 161 static void 162 emit_br_relo(struct nfp_prog *nfp_prog, enum br_mask mask, u16 addr, u8 defer, 163 enum nfp_relo_type relo) 164 { 165 if (mask == BR_UNC && defer > 2) { 166 pr_err("BUG: branch defer out of bounds %d\n", defer); 167 nfp_prog->error = -EFAULT; 168 return; 169 } 170 171 __emit_br(nfp_prog, mask, 172 mask != BR_UNC ? BR_EV_PIP_COND : BR_EV_PIP_UNCOND, 173 BR_CSS_NONE, addr, defer); 174 175 nfp_prog->prog[nfp_prog->prog_len - 1] |= 176 FIELD_PREP(OP_RELO_TYPE, relo); 177 } 178 179 static void 180 emit_br(struct nfp_prog *nfp_prog, enum br_mask mask, u16 addr, u8 defer) 181 { 182 emit_br_relo(nfp_prog, mask, addr, defer, RELO_BR_REL); 183 } 184 185 static void 186 __emit_br_bit(struct nfp_prog *nfp_prog, u16 areg, u16 breg, u16 addr, u8 defer, 187 bool set, bool src_lmextn) 188 { 189 u16 addr_lo, addr_hi; 190 u64 insn; 191 192 addr_lo = addr & (OP_BR_BIT_ADDR_LO >> __bf_shf(OP_BR_BIT_ADDR_LO)); 193 addr_hi = addr != addr_lo; 194 195 insn = OP_BR_BIT_BASE | 196 FIELD_PREP(OP_BR_BIT_A_SRC, areg) | 197 FIELD_PREP(OP_BR_BIT_B_SRC, breg) | 198 FIELD_PREP(OP_BR_BIT_BV, set) | 199 FIELD_PREP(OP_BR_BIT_DEFBR, defer) | 200 FIELD_PREP(OP_BR_BIT_ADDR_LO, addr_lo) | 201 FIELD_PREP(OP_BR_BIT_ADDR_HI, addr_hi) | 202 FIELD_PREP(OP_BR_BIT_SRC_LMEXTN, src_lmextn); 203 204 nfp_prog_push(nfp_prog, insn); 205 } 206 207 static void 208 emit_br_bit_relo(struct nfp_prog *nfp_prog, swreg src, u8 bit, u16 addr, 209 u8 defer, bool set, enum nfp_relo_type relo) 210 { 211 struct nfp_insn_re_regs reg; 212 int err; 213 214 /* NOTE: The bit to test is specified as an rotation amount, such that 215 * the bit to test will be placed on the MSB of the result when 216 * doing a rotate right. For bit X, we need right rotate X + 1. 217 */ 218 bit += 1; 219 220 err = swreg_to_restricted(reg_none(), src, reg_imm(bit), ®, false); 221 if (err) { 222 nfp_prog->error = err; 223 return; 224 } 225 226 __emit_br_bit(nfp_prog, reg.areg, reg.breg, addr, defer, set, 227 reg.src_lmextn); 228 229 nfp_prog->prog[nfp_prog->prog_len - 1] |= 230 FIELD_PREP(OP_RELO_TYPE, relo); 231 } 232 233 static void 234 emit_br_bset(struct nfp_prog *nfp_prog, swreg src, u8 bit, u16 addr, u8 defer) 235 { 236 emit_br_bit_relo(nfp_prog, src, bit, addr, defer, true, RELO_BR_REL); 237 } 238 239 static void 240 __emit_br_alu(struct nfp_prog *nfp_prog, u16 areg, u16 breg, u16 imm_hi, 241 u8 defer, bool dst_lmextn, bool src_lmextn) 242 { 243 u64 insn; 244 245 insn = OP_BR_ALU_BASE | 246 FIELD_PREP(OP_BR_ALU_A_SRC, areg) | 247 FIELD_PREP(OP_BR_ALU_B_SRC, breg) | 248 FIELD_PREP(OP_BR_ALU_DEFBR, defer) | 249 FIELD_PREP(OP_BR_ALU_IMM_HI, imm_hi) | 250 FIELD_PREP(OP_BR_ALU_SRC_LMEXTN, src_lmextn) | 251 FIELD_PREP(OP_BR_ALU_DST_LMEXTN, dst_lmextn); 252 253 nfp_prog_push(nfp_prog, insn); 254 } 255 256 static void emit_rtn(struct nfp_prog *nfp_prog, swreg base, u8 defer) 257 { 258 struct nfp_insn_ur_regs reg; 259 int err; 260 261 err = swreg_to_unrestricted(reg_none(), base, reg_imm(0), ®); 262 if (err) { 263 nfp_prog->error = err; 264 return; 265 } 266 267 __emit_br_alu(nfp_prog, reg.areg, reg.breg, 0, defer, reg.dst_lmextn, 268 reg.src_lmextn); 269 } 270 271 static void 272 __emit_immed(struct nfp_prog *nfp_prog, u16 areg, u16 breg, u16 imm_hi, 273 enum immed_width width, bool invert, 274 enum immed_shift shift, bool wr_both, 275 bool dst_lmextn, bool src_lmextn) 276 { 277 u64 insn; 278 279 insn = OP_IMMED_BASE | 280 FIELD_PREP(OP_IMMED_A_SRC, areg) | 281 FIELD_PREP(OP_IMMED_B_SRC, breg) | 282 FIELD_PREP(OP_IMMED_IMM, imm_hi) | 283 FIELD_PREP(OP_IMMED_WIDTH, width) | 284 FIELD_PREP(OP_IMMED_INV, invert) | 285 FIELD_PREP(OP_IMMED_SHIFT, shift) | 286 FIELD_PREP(OP_IMMED_WR_AB, wr_both) | 287 FIELD_PREP(OP_IMMED_SRC_LMEXTN, src_lmextn) | 288 FIELD_PREP(OP_IMMED_DST_LMEXTN, dst_lmextn); 289 290 nfp_prog_push(nfp_prog, insn); 291 } 292 293 static void 294 emit_immed(struct nfp_prog *nfp_prog, swreg dst, u16 imm, 295 enum immed_width width, bool invert, enum immed_shift shift) 296 { 297 struct nfp_insn_ur_regs reg; 298 int err; 299 300 if (swreg_type(dst) == NN_REG_IMM) { 301 nfp_prog->error = -EFAULT; 302 return; 303 } 304 305 err = swreg_to_unrestricted(dst, dst, reg_imm(imm & 0xff), ®); 306 if (err) { 307 nfp_prog->error = err; 308 return; 309 } 310 311 /* Use reg.dst when destination is No-Dest. */ 312 __emit_immed(nfp_prog, 313 swreg_type(dst) == NN_REG_NONE ? reg.dst : reg.areg, 314 reg.breg, imm >> 8, width, invert, shift, 315 reg.wr_both, reg.dst_lmextn, reg.src_lmextn); 316 } 317 318 static void 319 __emit_shf(struct nfp_prog *nfp_prog, u16 dst, enum alu_dst_ab dst_ab, 320 enum shf_sc sc, u8 shift, 321 u16 areg, enum shf_op op, u16 breg, bool i8, bool sw, bool wr_both, 322 bool dst_lmextn, bool src_lmextn) 323 { 324 u64 insn; 325 326 if (!FIELD_FIT(OP_SHF_SHIFT, shift)) { 327 nfp_prog->error = -EFAULT; 328 return; 329 } 330 331 if (sc == SHF_SC_L_SHF) 332 shift = 32 - shift; 333 334 insn = OP_SHF_BASE | 335 FIELD_PREP(OP_SHF_A_SRC, areg) | 336 FIELD_PREP(OP_SHF_SC, sc) | 337 FIELD_PREP(OP_SHF_B_SRC, breg) | 338 FIELD_PREP(OP_SHF_I8, i8) | 339 FIELD_PREP(OP_SHF_SW, sw) | 340 FIELD_PREP(OP_SHF_DST, dst) | 341 FIELD_PREP(OP_SHF_SHIFT, shift) | 342 FIELD_PREP(OP_SHF_OP, op) | 343 FIELD_PREP(OP_SHF_DST_AB, dst_ab) | 344 FIELD_PREP(OP_SHF_WR_AB, wr_both) | 345 FIELD_PREP(OP_SHF_SRC_LMEXTN, src_lmextn) | 346 FIELD_PREP(OP_SHF_DST_LMEXTN, dst_lmextn); 347 348 nfp_prog_push(nfp_prog, insn); 349 } 350 351 static void 352 emit_shf(struct nfp_prog *nfp_prog, swreg dst, 353 swreg lreg, enum shf_op op, swreg rreg, enum shf_sc sc, u8 shift) 354 { 355 struct nfp_insn_re_regs reg; 356 int err; 357 358 err = swreg_to_restricted(dst, lreg, rreg, ®, true); 359 if (err) { 360 nfp_prog->error = err; 361 return; 362 } 363 364 __emit_shf(nfp_prog, reg.dst, reg.dst_ab, sc, shift, 365 reg.areg, op, reg.breg, reg.i8, reg.swap, reg.wr_both, 366 reg.dst_lmextn, reg.src_lmextn); 367 } 368 369 static void 370 emit_shf_indir(struct nfp_prog *nfp_prog, swreg dst, 371 swreg lreg, enum shf_op op, swreg rreg, enum shf_sc sc) 372 { 373 if (sc == SHF_SC_R_ROT) { 374 pr_err("indirect shift is not allowed on rotation\n"); 375 nfp_prog->error = -EFAULT; 376 return; 377 } 378 379 emit_shf(nfp_prog, dst, lreg, op, rreg, sc, 0); 380 } 381 382 static void 383 __emit_alu(struct nfp_prog *nfp_prog, u16 dst, enum alu_dst_ab dst_ab, 384 u16 areg, enum alu_op op, u16 breg, bool swap, bool wr_both, 385 bool dst_lmextn, bool src_lmextn) 386 { 387 u64 insn; 388 389 insn = OP_ALU_BASE | 390 FIELD_PREP(OP_ALU_A_SRC, areg) | 391 FIELD_PREP(OP_ALU_B_SRC, breg) | 392 FIELD_PREP(OP_ALU_DST, dst) | 393 FIELD_PREP(OP_ALU_SW, swap) | 394 FIELD_PREP(OP_ALU_OP, op) | 395 FIELD_PREP(OP_ALU_DST_AB, dst_ab) | 396 FIELD_PREP(OP_ALU_WR_AB, wr_both) | 397 FIELD_PREP(OP_ALU_SRC_LMEXTN, src_lmextn) | 398 FIELD_PREP(OP_ALU_DST_LMEXTN, dst_lmextn); 399 400 nfp_prog_push(nfp_prog, insn); 401 } 402 403 static void 404 emit_alu(struct nfp_prog *nfp_prog, swreg dst, 405 swreg lreg, enum alu_op op, swreg rreg) 406 { 407 struct nfp_insn_ur_regs reg; 408 int err; 409 410 err = swreg_to_unrestricted(dst, lreg, rreg, ®); 411 if (err) { 412 nfp_prog->error = err; 413 return; 414 } 415 416 __emit_alu(nfp_prog, reg.dst, reg.dst_ab, 417 reg.areg, op, reg.breg, reg.swap, reg.wr_both, 418 reg.dst_lmextn, reg.src_lmextn); 419 } 420 421 static void 422 __emit_mul(struct nfp_prog *nfp_prog, enum alu_dst_ab dst_ab, u16 areg, 423 enum mul_type type, enum mul_step step, u16 breg, bool swap, 424 bool wr_both, bool dst_lmextn, bool src_lmextn) 425 { 426 u64 insn; 427 428 insn = OP_MUL_BASE | 429 FIELD_PREP(OP_MUL_A_SRC, areg) | 430 FIELD_PREP(OP_MUL_B_SRC, breg) | 431 FIELD_PREP(OP_MUL_STEP, step) | 432 FIELD_PREP(OP_MUL_DST_AB, dst_ab) | 433 FIELD_PREP(OP_MUL_SW, swap) | 434 FIELD_PREP(OP_MUL_TYPE, type) | 435 FIELD_PREP(OP_MUL_WR_AB, wr_both) | 436 FIELD_PREP(OP_MUL_SRC_LMEXTN, src_lmextn) | 437 FIELD_PREP(OP_MUL_DST_LMEXTN, dst_lmextn); 438 439 nfp_prog_push(nfp_prog, insn); 440 } 441 442 static void 443 emit_mul(struct nfp_prog *nfp_prog, swreg lreg, enum mul_type type, 444 enum mul_step step, swreg rreg) 445 { 446 struct nfp_insn_ur_regs reg; 447 u16 areg; 448 int err; 449 450 if (type == MUL_TYPE_START && step != MUL_STEP_NONE) { 451 nfp_prog->error = -EINVAL; 452 return; 453 } 454 455 if (step == MUL_LAST || step == MUL_LAST_2) { 456 /* When type is step and step Number is LAST or LAST2, left 457 * source is used as destination. 458 */ 459 err = swreg_to_unrestricted(lreg, reg_none(), rreg, ®); 460 areg = reg.dst; 461 } else { 462 err = swreg_to_unrestricted(reg_none(), lreg, rreg, ®); 463 areg = reg.areg; 464 } 465 466 if (err) { 467 nfp_prog->error = err; 468 return; 469 } 470 471 __emit_mul(nfp_prog, reg.dst_ab, areg, type, step, reg.breg, reg.swap, 472 reg.wr_both, reg.dst_lmextn, reg.src_lmextn); 473 } 474 475 static void 476 __emit_ld_field(struct nfp_prog *nfp_prog, enum shf_sc sc, 477 u8 areg, u8 bmask, u8 breg, u8 shift, bool imm8, 478 bool zero, bool swap, bool wr_both, 479 bool dst_lmextn, bool src_lmextn) 480 { 481 u64 insn; 482 483 insn = OP_LDF_BASE | 484 FIELD_PREP(OP_LDF_A_SRC, areg) | 485 FIELD_PREP(OP_LDF_SC, sc) | 486 FIELD_PREP(OP_LDF_B_SRC, breg) | 487 FIELD_PREP(OP_LDF_I8, imm8) | 488 FIELD_PREP(OP_LDF_SW, swap) | 489 FIELD_PREP(OP_LDF_ZF, zero) | 490 FIELD_PREP(OP_LDF_BMASK, bmask) | 491 FIELD_PREP(OP_LDF_SHF, shift) | 492 FIELD_PREP(OP_LDF_WR_AB, wr_both) | 493 FIELD_PREP(OP_LDF_SRC_LMEXTN, src_lmextn) | 494 FIELD_PREP(OP_LDF_DST_LMEXTN, dst_lmextn); 495 496 nfp_prog_push(nfp_prog, insn); 497 } 498 499 static void 500 emit_ld_field_any(struct nfp_prog *nfp_prog, swreg dst, u8 bmask, swreg src, 501 enum shf_sc sc, u8 shift, bool zero) 502 { 503 struct nfp_insn_re_regs reg; 504 int err; 505 506 /* Note: ld_field is special as it uses one of the src regs as dst */ 507 err = swreg_to_restricted(dst, dst, src, ®, true); 508 if (err) { 509 nfp_prog->error = err; 510 return; 511 } 512 513 __emit_ld_field(nfp_prog, sc, reg.areg, bmask, reg.breg, shift, 514 reg.i8, zero, reg.swap, reg.wr_both, 515 reg.dst_lmextn, reg.src_lmextn); 516 } 517 518 static void 519 emit_ld_field(struct nfp_prog *nfp_prog, swreg dst, u8 bmask, swreg src, 520 enum shf_sc sc, u8 shift) 521 { 522 emit_ld_field_any(nfp_prog, dst, bmask, src, sc, shift, false); 523 } 524 525 static void 526 __emit_lcsr(struct nfp_prog *nfp_prog, u16 areg, u16 breg, bool wr, u16 addr, 527 bool dst_lmextn, bool src_lmextn) 528 { 529 u64 insn; 530 531 insn = OP_LCSR_BASE | 532 FIELD_PREP(OP_LCSR_A_SRC, areg) | 533 FIELD_PREP(OP_LCSR_B_SRC, breg) | 534 FIELD_PREP(OP_LCSR_WRITE, wr) | 535 FIELD_PREP(OP_LCSR_ADDR, addr / 4) | 536 FIELD_PREP(OP_LCSR_SRC_LMEXTN, src_lmextn) | 537 FIELD_PREP(OP_LCSR_DST_LMEXTN, dst_lmextn); 538 539 nfp_prog_push(nfp_prog, insn); 540 } 541 542 static void emit_csr_wr(struct nfp_prog *nfp_prog, swreg src, u16 addr) 543 { 544 struct nfp_insn_ur_regs reg; 545 int err; 546 547 /* This instruction takes immeds instead of reg_none() for the ignored 548 * operand, but we can't encode 2 immeds in one instr with our normal 549 * swreg infra so if param is an immed, we encode as reg_none() and 550 * copy the immed to both operands. 551 */ 552 if (swreg_type(src) == NN_REG_IMM) { 553 err = swreg_to_unrestricted(reg_none(), src, reg_none(), ®); 554 reg.breg = reg.areg; 555 } else { 556 err = swreg_to_unrestricted(reg_none(), src, reg_imm(0), ®); 557 } 558 if (err) { 559 nfp_prog->error = err; 560 return; 561 } 562 563 __emit_lcsr(nfp_prog, reg.areg, reg.breg, true, addr, 564 false, reg.src_lmextn); 565 } 566 567 /* CSR value is read in following immed[gpr, 0] */ 568 static void __emit_csr_rd(struct nfp_prog *nfp_prog, u16 addr) 569 { 570 __emit_lcsr(nfp_prog, 0, 0, false, addr, false, false); 571 } 572 573 static void emit_nop(struct nfp_prog *nfp_prog) 574 { 575 __emit_immed(nfp_prog, UR_REG_IMM, UR_REG_IMM, 0, 0, 0, 0, 0, 0, 0); 576 } 577 578 /* --- Wrappers --- */ 579 static bool pack_immed(u32 imm, u16 *val, enum immed_shift *shift) 580 { 581 if (!(imm & 0xffff0000)) { 582 *val = imm; 583 *shift = IMMED_SHIFT_0B; 584 } else if (!(imm & 0xff0000ff)) { 585 *val = imm >> 8; 586 *shift = IMMED_SHIFT_1B; 587 } else if (!(imm & 0x0000ffff)) { 588 *val = imm >> 16; 589 *shift = IMMED_SHIFT_2B; 590 } else { 591 return false; 592 } 593 594 return true; 595 } 596 597 static void wrp_immed(struct nfp_prog *nfp_prog, swreg dst, u32 imm) 598 { 599 enum immed_shift shift; 600 u16 val; 601 602 if (pack_immed(imm, &val, &shift)) { 603 emit_immed(nfp_prog, dst, val, IMMED_WIDTH_ALL, false, shift); 604 } else if (pack_immed(~imm, &val, &shift)) { 605 emit_immed(nfp_prog, dst, val, IMMED_WIDTH_ALL, true, shift); 606 } else { 607 emit_immed(nfp_prog, dst, imm & 0xffff, IMMED_WIDTH_ALL, 608 false, IMMED_SHIFT_0B); 609 emit_immed(nfp_prog, dst, imm >> 16, IMMED_WIDTH_WORD, 610 false, IMMED_SHIFT_2B); 611 } 612 } 613 614 static void 615 wrp_immed_relo(struct nfp_prog *nfp_prog, swreg dst, u32 imm, 616 enum nfp_relo_type relo) 617 { 618 if (imm > 0xffff) { 619 pr_err("relocation of a large immediate!\n"); 620 nfp_prog->error = -EFAULT; 621 return; 622 } 623 emit_immed(nfp_prog, dst, imm, IMMED_WIDTH_ALL, false, IMMED_SHIFT_0B); 624 625 nfp_prog->prog[nfp_prog->prog_len - 1] |= 626 FIELD_PREP(OP_RELO_TYPE, relo); 627 } 628 629 /* ur_load_imm_any() - encode immediate or use tmp register (unrestricted) 630 * If the @imm is small enough encode it directly in operand and return 631 * otherwise load @imm to a spare register and return its encoding. 632 */ 633 static swreg ur_load_imm_any(struct nfp_prog *nfp_prog, u32 imm, swreg tmp_reg) 634 { 635 if (FIELD_FIT(UR_REG_IMM_MAX, imm)) 636 return reg_imm(imm); 637 638 wrp_immed(nfp_prog, tmp_reg, imm); 639 return tmp_reg; 640 } 641 642 /* re_load_imm_any() - encode immediate or use tmp register (restricted) 643 * If the @imm is small enough encode it directly in operand and return 644 * otherwise load @imm to a spare register and return its encoding. 645 */ 646 static swreg re_load_imm_any(struct nfp_prog *nfp_prog, u32 imm, swreg tmp_reg) 647 { 648 if (FIELD_FIT(RE_REG_IMM_MAX, imm)) 649 return reg_imm(imm); 650 651 wrp_immed(nfp_prog, tmp_reg, imm); 652 return tmp_reg; 653 } 654 655 static void wrp_nops(struct nfp_prog *nfp_prog, unsigned int count) 656 { 657 while (count--) 658 emit_nop(nfp_prog); 659 } 660 661 static void wrp_mov(struct nfp_prog *nfp_prog, swreg dst, swreg src) 662 { 663 emit_alu(nfp_prog, dst, reg_none(), ALU_OP_NONE, src); 664 } 665 666 static void wrp_reg_mov(struct nfp_prog *nfp_prog, u16 dst, u16 src) 667 { 668 wrp_mov(nfp_prog, reg_both(dst), reg_b(src)); 669 } 670 671 /* wrp_reg_subpart() - load @field_len bytes from @offset of @src, write the 672 * result to @dst from low end. 673 */ 674 static void 675 wrp_reg_subpart(struct nfp_prog *nfp_prog, swreg dst, swreg src, u8 field_len, 676 u8 offset) 677 { 678 enum shf_sc sc = offset ? SHF_SC_R_SHF : SHF_SC_NONE; 679 u8 mask = (1 << field_len) - 1; 680 681 emit_ld_field_any(nfp_prog, dst, mask, src, sc, offset * 8, true); 682 } 683 684 /* wrp_reg_or_subpart() - load @field_len bytes from low end of @src, or the 685 * result to @dst from offset, there is no change on the other bits of @dst. 686 */ 687 static void 688 wrp_reg_or_subpart(struct nfp_prog *nfp_prog, swreg dst, swreg src, 689 u8 field_len, u8 offset) 690 { 691 enum shf_sc sc = offset ? SHF_SC_L_SHF : SHF_SC_NONE; 692 u8 mask = ((1 << field_len) - 1) << offset; 693 694 emit_ld_field(nfp_prog, dst, mask, src, sc, 32 - offset * 8); 695 } 696 697 static void 698 addr40_offset(struct nfp_prog *nfp_prog, u8 src_gpr, swreg offset, 699 swreg *rega, swreg *regb) 700 { 701 if (offset == reg_imm(0)) { 702 *rega = reg_a(src_gpr); 703 *regb = reg_b(src_gpr + 1); 704 return; 705 } 706 707 emit_alu(nfp_prog, imm_a(nfp_prog), reg_a(src_gpr), ALU_OP_ADD, offset); 708 emit_alu(nfp_prog, imm_b(nfp_prog), reg_b(src_gpr + 1), ALU_OP_ADD_C, 709 reg_imm(0)); 710 *rega = imm_a(nfp_prog); 711 *regb = imm_b(nfp_prog); 712 } 713 714 /* NFP has Command Push Pull bus which supports bluk memory operations. */ 715 static int nfp_cpp_memcpy(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 716 { 717 bool descending_seq = meta->ldst_gather_len < 0; 718 s16 len = abs(meta->ldst_gather_len); 719 swreg src_base, off; 720 bool src_40bit_addr; 721 unsigned int i; 722 u8 xfer_num; 723 724 off = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog)); 725 src_40bit_addr = meta->ptr.type == PTR_TO_MAP_VALUE; 726 src_base = reg_a(meta->insn.src_reg * 2); 727 xfer_num = round_up(len, 4) / 4; 728 729 if (src_40bit_addr) 730 addr40_offset(nfp_prog, meta->insn.src_reg * 2, off, &src_base, 731 &off); 732 733 /* Setup PREV_ALU fields to override memory read length. */ 734 if (len > 32) 735 wrp_immed(nfp_prog, reg_none(), 736 CMD_OVE_LEN | FIELD_PREP(CMD_OV_LEN, xfer_num - 1)); 737 738 /* Memory read from source addr into transfer-in registers. */ 739 emit_cmd_any(nfp_prog, CMD_TGT_READ32_SWAP, 740 src_40bit_addr ? CMD_MODE_40b_BA : CMD_MODE_32b, 0, 741 src_base, off, xfer_num - 1, CMD_CTX_SWAP, len > 32); 742 743 /* Move from transfer-in to transfer-out. */ 744 for (i = 0; i < xfer_num; i++) 745 wrp_mov(nfp_prog, reg_xfer(i), reg_xfer(i)); 746 747 off = re_load_imm_any(nfp_prog, meta->paired_st->off, imm_b(nfp_prog)); 748 749 if (len <= 8) { 750 /* Use single direct_ref write8. */ 751 emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 0, 752 reg_a(meta->paired_st->dst_reg * 2), off, len - 1, 753 CMD_CTX_SWAP); 754 } else if (len <= 32 && IS_ALIGNED(len, 4)) { 755 /* Use single direct_ref write32. */ 756 emit_cmd(nfp_prog, CMD_TGT_WRITE32_SWAP, CMD_MODE_32b, 0, 757 reg_a(meta->paired_st->dst_reg * 2), off, xfer_num - 1, 758 CMD_CTX_SWAP); 759 } else if (len <= 32) { 760 /* Use single indirect_ref write8. */ 761 wrp_immed(nfp_prog, reg_none(), 762 CMD_OVE_LEN | FIELD_PREP(CMD_OV_LEN, len - 1)); 763 emit_cmd_indir(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 0, 764 reg_a(meta->paired_st->dst_reg * 2), off, 765 len - 1, CMD_CTX_SWAP); 766 } else if (IS_ALIGNED(len, 4)) { 767 /* Use single indirect_ref write32. */ 768 wrp_immed(nfp_prog, reg_none(), 769 CMD_OVE_LEN | FIELD_PREP(CMD_OV_LEN, xfer_num - 1)); 770 emit_cmd_indir(nfp_prog, CMD_TGT_WRITE32_SWAP, CMD_MODE_32b, 0, 771 reg_a(meta->paired_st->dst_reg * 2), off, 772 xfer_num - 1, CMD_CTX_SWAP); 773 } else if (len <= 40) { 774 /* Use one direct_ref write32 to write the first 32-bytes, then 775 * another direct_ref write8 to write the remaining bytes. 776 */ 777 emit_cmd(nfp_prog, CMD_TGT_WRITE32_SWAP, CMD_MODE_32b, 0, 778 reg_a(meta->paired_st->dst_reg * 2), off, 7, 779 CMD_CTX_SWAP); 780 781 off = re_load_imm_any(nfp_prog, meta->paired_st->off + 32, 782 imm_b(nfp_prog)); 783 emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 8, 784 reg_a(meta->paired_st->dst_reg * 2), off, len - 33, 785 CMD_CTX_SWAP); 786 } else { 787 /* Use one indirect_ref write32 to write 4-bytes aligned length, 788 * then another direct_ref write8 to write the remaining bytes. 789 */ 790 u8 new_off; 791 792 wrp_immed(nfp_prog, reg_none(), 793 CMD_OVE_LEN | FIELD_PREP(CMD_OV_LEN, xfer_num - 2)); 794 emit_cmd_indir(nfp_prog, CMD_TGT_WRITE32_SWAP, CMD_MODE_32b, 0, 795 reg_a(meta->paired_st->dst_reg * 2), off, 796 xfer_num - 2, CMD_CTX_SWAP); 797 new_off = meta->paired_st->off + (xfer_num - 1) * 4; 798 off = re_load_imm_any(nfp_prog, new_off, imm_b(nfp_prog)); 799 emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 800 xfer_num - 1, reg_a(meta->paired_st->dst_reg * 2), off, 801 (len & 0x3) - 1, CMD_CTX_SWAP); 802 } 803 804 /* TODO: The following extra load is to make sure data flow be identical 805 * before and after we do memory copy optimization. 806 * 807 * The load destination register is not guaranteed to be dead, so we 808 * need to make sure it is loaded with the value the same as before 809 * this transformation. 810 * 811 * These extra loads could be removed once we have accurate register 812 * usage information. 813 */ 814 if (descending_seq) 815 xfer_num = 0; 816 else if (BPF_SIZE(meta->insn.code) != BPF_DW) 817 xfer_num = xfer_num - 1; 818 else 819 xfer_num = xfer_num - 2; 820 821 switch (BPF_SIZE(meta->insn.code)) { 822 case BPF_B: 823 wrp_reg_subpart(nfp_prog, reg_both(meta->insn.dst_reg * 2), 824 reg_xfer(xfer_num), 1, 825 IS_ALIGNED(len, 4) ? 3 : (len & 3) - 1); 826 break; 827 case BPF_H: 828 wrp_reg_subpart(nfp_prog, reg_both(meta->insn.dst_reg * 2), 829 reg_xfer(xfer_num), 2, (len & 3) ^ 2); 830 break; 831 case BPF_W: 832 wrp_mov(nfp_prog, reg_both(meta->insn.dst_reg * 2), 833 reg_xfer(0)); 834 break; 835 case BPF_DW: 836 wrp_mov(nfp_prog, reg_both(meta->insn.dst_reg * 2), 837 reg_xfer(xfer_num)); 838 wrp_mov(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 839 reg_xfer(xfer_num + 1)); 840 break; 841 } 842 843 if (BPF_SIZE(meta->insn.code) != BPF_DW) 844 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0); 845 846 return 0; 847 } 848 849 static int 850 data_ld(struct nfp_prog *nfp_prog, swreg offset, u8 dst_gpr, int size) 851 { 852 unsigned int i; 853 u16 shift, sz; 854 855 /* We load the value from the address indicated in @offset and then 856 * shift out the data we don't need. Note: this is big endian! 857 */ 858 sz = max(size, 4); 859 shift = size < 4 ? 4 - size : 0; 860 861 emit_cmd(nfp_prog, CMD_TGT_READ8, CMD_MODE_32b, 0, 862 pptr_reg(nfp_prog), offset, sz - 1, CMD_CTX_SWAP); 863 864 i = 0; 865 if (shift) 866 emit_shf(nfp_prog, reg_both(dst_gpr), reg_none(), SHF_OP_NONE, 867 reg_xfer(0), SHF_SC_R_SHF, shift * 8); 868 else 869 for (; i * 4 < size; i++) 870 wrp_mov(nfp_prog, reg_both(dst_gpr + i), reg_xfer(i)); 871 872 if (i < 2) 873 wrp_immed(nfp_prog, reg_both(dst_gpr + 1), 0); 874 875 return 0; 876 } 877 878 static int 879 data_ld_host_order(struct nfp_prog *nfp_prog, u8 dst_gpr, 880 swreg lreg, swreg rreg, int size, enum cmd_mode mode) 881 { 882 unsigned int i; 883 u8 mask, sz; 884 885 /* We load the value from the address indicated in rreg + lreg and then 886 * mask out the data we don't need. Note: this is little endian! 887 */ 888 sz = max(size, 4); 889 mask = size < 4 ? GENMASK(size - 1, 0) : 0; 890 891 emit_cmd(nfp_prog, CMD_TGT_READ32_SWAP, mode, 0, 892 lreg, rreg, sz / 4 - 1, CMD_CTX_SWAP); 893 894 i = 0; 895 if (mask) 896 emit_ld_field_any(nfp_prog, reg_both(dst_gpr), mask, 897 reg_xfer(0), SHF_SC_NONE, 0, true); 898 else 899 for (; i * 4 < size; i++) 900 wrp_mov(nfp_prog, reg_both(dst_gpr + i), reg_xfer(i)); 901 902 if (i < 2) 903 wrp_immed(nfp_prog, reg_both(dst_gpr + 1), 0); 904 905 return 0; 906 } 907 908 static int 909 data_ld_host_order_addr32(struct nfp_prog *nfp_prog, u8 src_gpr, swreg offset, 910 u8 dst_gpr, u8 size) 911 { 912 return data_ld_host_order(nfp_prog, dst_gpr, reg_a(src_gpr), offset, 913 size, CMD_MODE_32b); 914 } 915 916 static int 917 data_ld_host_order_addr40(struct nfp_prog *nfp_prog, u8 src_gpr, swreg offset, 918 u8 dst_gpr, u8 size) 919 { 920 swreg rega, regb; 921 922 addr40_offset(nfp_prog, src_gpr, offset, ®a, ®b); 923 924 return data_ld_host_order(nfp_prog, dst_gpr, rega, regb, 925 size, CMD_MODE_40b_BA); 926 } 927 928 static int 929 construct_data_ind_ld(struct nfp_prog *nfp_prog, u16 offset, u16 src, u8 size) 930 { 931 swreg tmp_reg; 932 933 /* Calculate the true offset (src_reg + imm) */ 934 tmp_reg = ur_load_imm_any(nfp_prog, offset, imm_b(nfp_prog)); 935 emit_alu(nfp_prog, imm_both(nfp_prog), reg_a(src), ALU_OP_ADD, tmp_reg); 936 937 /* Check packet length (size guaranteed to fit b/c it's u8) */ 938 emit_alu(nfp_prog, imm_a(nfp_prog), 939 imm_a(nfp_prog), ALU_OP_ADD, reg_imm(size)); 940 emit_alu(nfp_prog, reg_none(), 941 plen_reg(nfp_prog), ALU_OP_SUB, imm_a(nfp_prog)); 942 emit_br_relo(nfp_prog, BR_BLO, BR_OFF_RELO, 0, RELO_BR_GO_ABORT); 943 944 /* Load data */ 945 return data_ld(nfp_prog, imm_b(nfp_prog), 0, size); 946 } 947 948 static int construct_data_ld(struct nfp_prog *nfp_prog, u16 offset, u8 size) 949 { 950 swreg tmp_reg; 951 952 /* Check packet length */ 953 tmp_reg = ur_load_imm_any(nfp_prog, offset + size, imm_a(nfp_prog)); 954 emit_alu(nfp_prog, reg_none(), plen_reg(nfp_prog), ALU_OP_SUB, tmp_reg); 955 emit_br_relo(nfp_prog, BR_BLO, BR_OFF_RELO, 0, RELO_BR_GO_ABORT); 956 957 /* Load data */ 958 tmp_reg = re_load_imm_any(nfp_prog, offset, imm_b(nfp_prog)); 959 return data_ld(nfp_prog, tmp_reg, 0, size); 960 } 961 962 static int 963 data_stx_host_order(struct nfp_prog *nfp_prog, u8 dst_gpr, swreg offset, 964 u8 src_gpr, u8 size) 965 { 966 unsigned int i; 967 968 for (i = 0; i * 4 < size; i++) 969 wrp_mov(nfp_prog, reg_xfer(i), reg_a(src_gpr + i)); 970 971 emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 0, 972 reg_a(dst_gpr), offset, size - 1, CMD_CTX_SWAP); 973 974 return 0; 975 } 976 977 static int 978 data_st_host_order(struct nfp_prog *nfp_prog, u8 dst_gpr, swreg offset, 979 u64 imm, u8 size) 980 { 981 wrp_immed(nfp_prog, reg_xfer(0), imm); 982 if (size == 8) 983 wrp_immed(nfp_prog, reg_xfer(1), imm >> 32); 984 985 emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 0, 986 reg_a(dst_gpr), offset, size - 1, CMD_CTX_SWAP); 987 988 return 0; 989 } 990 991 typedef int 992 (*lmem_step)(struct nfp_prog *nfp_prog, u8 gpr, u8 gpr_byte, s32 off, 993 unsigned int size, bool first, bool new_gpr, bool last, bool lm3, 994 bool needs_inc); 995 996 static int 997 wrp_lmem_load(struct nfp_prog *nfp_prog, u8 dst, u8 dst_byte, s32 off, 998 unsigned int size, bool first, bool new_gpr, bool last, bool lm3, 999 bool needs_inc) 1000 { 1001 bool should_inc = needs_inc && new_gpr && !last; 1002 u32 idx, src_byte; 1003 enum shf_sc sc; 1004 swreg reg; 1005 int shf; 1006 u8 mask; 1007 1008 if (WARN_ON_ONCE(dst_byte + size > 4 || off % 4 + size > 4)) 1009 return -EOPNOTSUPP; 1010 1011 idx = off / 4; 1012 1013 /* Move the entire word */ 1014 if (size == 4) { 1015 wrp_mov(nfp_prog, reg_both(dst), 1016 should_inc ? reg_lm_inc(3) : reg_lm(lm3 ? 3 : 0, idx)); 1017 return 0; 1018 } 1019 1020 if (WARN_ON_ONCE(lm3 && idx > RE_REG_LM_IDX_MAX)) 1021 return -EOPNOTSUPP; 1022 1023 src_byte = off % 4; 1024 1025 mask = (1 << size) - 1; 1026 mask <<= dst_byte; 1027 1028 if (WARN_ON_ONCE(mask > 0xf)) 1029 return -EOPNOTSUPP; 1030 1031 shf = abs(src_byte - dst_byte) * 8; 1032 if (src_byte == dst_byte) { 1033 sc = SHF_SC_NONE; 1034 } else if (src_byte < dst_byte) { 1035 shf = 32 - shf; 1036 sc = SHF_SC_L_SHF; 1037 } else { 1038 sc = SHF_SC_R_SHF; 1039 } 1040 1041 /* ld_field can address fewer indexes, if offset too large do RMW. 1042 * Because we RMV twice we waste 2 cycles on unaligned 8 byte writes. 1043 */ 1044 if (idx <= RE_REG_LM_IDX_MAX) { 1045 reg = reg_lm(lm3 ? 3 : 0, idx); 1046 } else { 1047 reg = imm_a(nfp_prog); 1048 /* If it's not the first part of the load and we start a new GPR 1049 * that means we are loading a second part of the LMEM word into 1050 * a new GPR. IOW we've already looked that LMEM word and 1051 * therefore it has been loaded into imm_a(). 1052 */ 1053 if (first || !new_gpr) 1054 wrp_mov(nfp_prog, reg, reg_lm(0, idx)); 1055 } 1056 1057 emit_ld_field_any(nfp_prog, reg_both(dst), mask, reg, sc, shf, new_gpr); 1058 1059 if (should_inc) 1060 wrp_mov(nfp_prog, reg_none(), reg_lm_inc(3)); 1061 1062 return 0; 1063 } 1064 1065 static int 1066 wrp_lmem_store(struct nfp_prog *nfp_prog, u8 src, u8 src_byte, s32 off, 1067 unsigned int size, bool first, bool new_gpr, bool last, bool lm3, 1068 bool needs_inc) 1069 { 1070 bool should_inc = needs_inc && new_gpr && !last; 1071 u32 idx, dst_byte; 1072 enum shf_sc sc; 1073 swreg reg; 1074 int shf; 1075 u8 mask; 1076 1077 if (WARN_ON_ONCE(src_byte + size > 4 || off % 4 + size > 4)) 1078 return -EOPNOTSUPP; 1079 1080 idx = off / 4; 1081 1082 /* Move the entire word */ 1083 if (size == 4) { 1084 wrp_mov(nfp_prog, 1085 should_inc ? reg_lm_inc(3) : reg_lm(lm3 ? 3 : 0, idx), 1086 reg_b(src)); 1087 return 0; 1088 } 1089 1090 if (WARN_ON_ONCE(lm3 && idx > RE_REG_LM_IDX_MAX)) 1091 return -EOPNOTSUPP; 1092 1093 dst_byte = off % 4; 1094 1095 mask = (1 << size) - 1; 1096 mask <<= dst_byte; 1097 1098 if (WARN_ON_ONCE(mask > 0xf)) 1099 return -EOPNOTSUPP; 1100 1101 shf = abs(src_byte - dst_byte) * 8; 1102 if (src_byte == dst_byte) { 1103 sc = SHF_SC_NONE; 1104 } else if (src_byte < dst_byte) { 1105 shf = 32 - shf; 1106 sc = SHF_SC_L_SHF; 1107 } else { 1108 sc = SHF_SC_R_SHF; 1109 } 1110 1111 /* ld_field can address fewer indexes, if offset too large do RMW. 1112 * Because we RMV twice we waste 2 cycles on unaligned 8 byte writes. 1113 */ 1114 if (idx <= RE_REG_LM_IDX_MAX) { 1115 reg = reg_lm(lm3 ? 3 : 0, idx); 1116 } else { 1117 reg = imm_a(nfp_prog); 1118 /* Only first and last LMEM locations are going to need RMW, 1119 * the middle location will be overwritten fully. 1120 */ 1121 if (first || last) 1122 wrp_mov(nfp_prog, reg, reg_lm(0, idx)); 1123 } 1124 1125 emit_ld_field(nfp_prog, reg, mask, reg_b(src), sc, shf); 1126 1127 if (new_gpr || last) { 1128 if (idx > RE_REG_LM_IDX_MAX) 1129 wrp_mov(nfp_prog, reg_lm(0, idx), reg); 1130 if (should_inc) 1131 wrp_mov(nfp_prog, reg_none(), reg_lm_inc(3)); 1132 } 1133 1134 return 0; 1135 } 1136 1137 static int 1138 mem_op_stack(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1139 unsigned int size, unsigned int ptr_off, u8 gpr, u8 ptr_gpr, 1140 bool clr_gpr, lmem_step step) 1141 { 1142 s32 off = nfp_prog->stack_frame_depth + meta->insn.off + ptr_off; 1143 bool first = true, last; 1144 bool needs_inc = false; 1145 swreg stack_off_reg; 1146 u8 prev_gpr = 255; 1147 u32 gpr_byte = 0; 1148 bool lm3 = true; 1149 int ret; 1150 1151 if (meta->ptr_not_const || 1152 meta->flags & FLAG_INSN_PTR_CALLER_STACK_FRAME) { 1153 /* Use of the last encountered ptr_off is OK, they all have 1154 * the same alignment. Depend on low bits of value being 1155 * discarded when written to LMaddr register. 1156 */ 1157 stack_off_reg = ur_load_imm_any(nfp_prog, meta->insn.off, 1158 stack_imm(nfp_prog)); 1159 1160 emit_alu(nfp_prog, imm_b(nfp_prog), 1161 reg_a(ptr_gpr), ALU_OP_ADD, stack_off_reg); 1162 1163 needs_inc = true; 1164 } else if (off + size <= 64) { 1165 /* We can reach bottom 64B with LMaddr0 */ 1166 lm3 = false; 1167 } else if (round_down(off, 32) == round_down(off + size - 1, 32)) { 1168 /* We have to set up a new pointer. If we know the offset 1169 * and the entire access falls into a single 32 byte aligned 1170 * window we won't have to increment the LM pointer. 1171 * The 32 byte alignment is imporant because offset is ORed in 1172 * not added when doing *l$indexN[off]. 1173 */ 1174 stack_off_reg = ur_load_imm_any(nfp_prog, round_down(off, 32), 1175 stack_imm(nfp_prog)); 1176 emit_alu(nfp_prog, imm_b(nfp_prog), 1177 stack_reg(nfp_prog), ALU_OP_ADD, stack_off_reg); 1178 1179 off %= 32; 1180 } else { 1181 stack_off_reg = ur_load_imm_any(nfp_prog, round_down(off, 4), 1182 stack_imm(nfp_prog)); 1183 1184 emit_alu(nfp_prog, imm_b(nfp_prog), 1185 stack_reg(nfp_prog), ALU_OP_ADD, stack_off_reg); 1186 1187 needs_inc = true; 1188 } 1189 if (lm3) { 1190 emit_csr_wr(nfp_prog, imm_b(nfp_prog), NFP_CSR_ACT_LM_ADDR3); 1191 /* For size < 4 one slot will be filled by zeroing of upper. */ 1192 wrp_nops(nfp_prog, clr_gpr && size < 8 ? 2 : 3); 1193 } 1194 1195 if (clr_gpr && size < 8) 1196 wrp_immed(nfp_prog, reg_both(gpr + 1), 0); 1197 1198 while (size) { 1199 u32 slice_end; 1200 u8 slice_size; 1201 1202 slice_size = min(size, 4 - gpr_byte); 1203 slice_end = min(off + slice_size, round_up(off + 1, 4)); 1204 slice_size = slice_end - off; 1205 1206 last = slice_size == size; 1207 1208 if (needs_inc) 1209 off %= 4; 1210 1211 ret = step(nfp_prog, gpr, gpr_byte, off, slice_size, 1212 first, gpr != prev_gpr, last, lm3, needs_inc); 1213 if (ret) 1214 return ret; 1215 1216 prev_gpr = gpr; 1217 first = false; 1218 1219 gpr_byte += slice_size; 1220 if (gpr_byte >= 4) { 1221 gpr_byte -= 4; 1222 gpr++; 1223 } 1224 1225 size -= slice_size; 1226 off += slice_size; 1227 } 1228 1229 return 0; 1230 } 1231 1232 static void 1233 wrp_alu_imm(struct nfp_prog *nfp_prog, u8 dst, enum alu_op alu_op, u32 imm) 1234 { 1235 swreg tmp_reg; 1236 1237 if (alu_op == ALU_OP_AND) { 1238 if (!imm) 1239 wrp_immed(nfp_prog, reg_both(dst), 0); 1240 if (!imm || !~imm) 1241 return; 1242 } 1243 if (alu_op == ALU_OP_OR) { 1244 if (!~imm) 1245 wrp_immed(nfp_prog, reg_both(dst), ~0U); 1246 if (!imm || !~imm) 1247 return; 1248 } 1249 if (alu_op == ALU_OP_XOR) { 1250 if (!~imm) 1251 emit_alu(nfp_prog, reg_both(dst), reg_none(), 1252 ALU_OP_NOT, reg_b(dst)); 1253 if (!imm || !~imm) 1254 return; 1255 } 1256 1257 tmp_reg = ur_load_imm_any(nfp_prog, imm, imm_b(nfp_prog)); 1258 emit_alu(nfp_prog, reg_both(dst), reg_a(dst), alu_op, tmp_reg); 1259 } 1260 1261 static int 1262 wrp_alu64_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1263 enum alu_op alu_op, bool skip) 1264 { 1265 const struct bpf_insn *insn = &meta->insn; 1266 u64 imm = insn->imm; /* sign extend */ 1267 1268 if (skip) { 1269 meta->skip = true; 1270 return 0; 1271 } 1272 1273 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, alu_op, imm & ~0U); 1274 wrp_alu_imm(nfp_prog, insn->dst_reg * 2 + 1, alu_op, imm >> 32); 1275 1276 return 0; 1277 } 1278 1279 static int 1280 wrp_alu64_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1281 enum alu_op alu_op) 1282 { 1283 u8 dst = meta->insn.dst_reg * 2, src = meta->insn.src_reg * 2; 1284 1285 emit_alu(nfp_prog, reg_both(dst), reg_a(dst), alu_op, reg_b(src)); 1286 emit_alu(nfp_prog, reg_both(dst + 1), 1287 reg_a(dst + 1), alu_op, reg_b(src + 1)); 1288 1289 return 0; 1290 } 1291 1292 static int 1293 wrp_alu32_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1294 enum alu_op alu_op, bool skip) 1295 { 1296 const struct bpf_insn *insn = &meta->insn; 1297 1298 if (skip) { 1299 meta->skip = true; 1300 return 0; 1301 } 1302 1303 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, alu_op, insn->imm); 1304 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0); 1305 1306 return 0; 1307 } 1308 1309 static int 1310 wrp_alu32_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1311 enum alu_op alu_op) 1312 { 1313 u8 dst = meta->insn.dst_reg * 2, src = meta->insn.src_reg * 2; 1314 1315 emit_alu(nfp_prog, reg_both(dst), reg_a(dst), alu_op, reg_b(src)); 1316 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0); 1317 1318 return 0; 1319 } 1320 1321 static void 1322 wrp_test_reg_one(struct nfp_prog *nfp_prog, u8 dst, enum alu_op alu_op, u8 src, 1323 enum br_mask br_mask, u16 off) 1324 { 1325 emit_alu(nfp_prog, reg_none(), reg_a(dst), alu_op, reg_b(src)); 1326 emit_br(nfp_prog, br_mask, off, 0); 1327 } 1328 1329 static int 1330 wrp_test_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1331 enum alu_op alu_op, enum br_mask br_mask) 1332 { 1333 const struct bpf_insn *insn = &meta->insn; 1334 1335 wrp_test_reg_one(nfp_prog, insn->dst_reg * 2, alu_op, 1336 insn->src_reg * 2, br_mask, insn->off); 1337 wrp_test_reg_one(nfp_prog, insn->dst_reg * 2 + 1, alu_op, 1338 insn->src_reg * 2 + 1, br_mask, insn->off); 1339 1340 return 0; 1341 } 1342 1343 static const struct jmp_code_map { 1344 enum br_mask br_mask; 1345 bool swap; 1346 } jmp_code_map[] = { 1347 [BPF_JGT >> 4] = { BR_BLO, true }, 1348 [BPF_JGE >> 4] = { BR_BHS, false }, 1349 [BPF_JLT >> 4] = { BR_BLO, false }, 1350 [BPF_JLE >> 4] = { BR_BHS, true }, 1351 [BPF_JSGT >> 4] = { BR_BLT, true }, 1352 [BPF_JSGE >> 4] = { BR_BGE, false }, 1353 [BPF_JSLT >> 4] = { BR_BLT, false }, 1354 [BPF_JSLE >> 4] = { BR_BGE, true }, 1355 }; 1356 1357 static const struct jmp_code_map *nfp_jmp_code_get(struct nfp_insn_meta *meta) 1358 { 1359 unsigned int op; 1360 1361 op = BPF_OP(meta->insn.code) >> 4; 1362 /* br_mask of 0 is BR_BEQ which we don't use in jump code table */ 1363 if (WARN_ONCE(op >= ARRAY_SIZE(jmp_code_map) || 1364 !jmp_code_map[op].br_mask, 1365 "no code found for jump instruction")) 1366 return NULL; 1367 1368 return &jmp_code_map[op]; 1369 } 1370 1371 static int cmp_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1372 { 1373 const struct bpf_insn *insn = &meta->insn; 1374 u64 imm = insn->imm; /* sign extend */ 1375 const struct jmp_code_map *code; 1376 enum alu_op alu_op, carry_op; 1377 u8 reg = insn->dst_reg * 2; 1378 swreg tmp_reg; 1379 1380 code = nfp_jmp_code_get(meta); 1381 if (!code) 1382 return -EINVAL; 1383 1384 alu_op = meta->jump_neg_op ? ALU_OP_ADD : ALU_OP_SUB; 1385 carry_op = meta->jump_neg_op ? ALU_OP_ADD_C : ALU_OP_SUB_C; 1386 1387 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog)); 1388 if (!code->swap) 1389 emit_alu(nfp_prog, reg_none(), reg_a(reg), alu_op, tmp_reg); 1390 else 1391 emit_alu(nfp_prog, reg_none(), tmp_reg, alu_op, reg_a(reg)); 1392 1393 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog)); 1394 if (!code->swap) 1395 emit_alu(nfp_prog, reg_none(), 1396 reg_a(reg + 1), carry_op, tmp_reg); 1397 else 1398 emit_alu(nfp_prog, reg_none(), 1399 tmp_reg, carry_op, reg_a(reg + 1)); 1400 1401 emit_br(nfp_prog, code->br_mask, insn->off, 0); 1402 1403 return 0; 1404 } 1405 1406 static int cmp_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1407 { 1408 const struct bpf_insn *insn = &meta->insn; 1409 const struct jmp_code_map *code; 1410 u8 areg, breg; 1411 1412 code = nfp_jmp_code_get(meta); 1413 if (!code) 1414 return -EINVAL; 1415 1416 areg = insn->dst_reg * 2; 1417 breg = insn->src_reg * 2; 1418 1419 if (code->swap) { 1420 areg ^= breg; 1421 breg ^= areg; 1422 areg ^= breg; 1423 } 1424 1425 emit_alu(nfp_prog, reg_none(), reg_a(areg), ALU_OP_SUB, reg_b(breg)); 1426 emit_alu(nfp_prog, reg_none(), 1427 reg_a(areg + 1), ALU_OP_SUB_C, reg_b(breg + 1)); 1428 emit_br(nfp_prog, code->br_mask, insn->off, 0); 1429 1430 return 0; 1431 } 1432 1433 static void wrp_end32(struct nfp_prog *nfp_prog, swreg reg_in, u8 gpr_out) 1434 { 1435 emit_ld_field(nfp_prog, reg_both(gpr_out), 0xf, reg_in, 1436 SHF_SC_R_ROT, 8); 1437 emit_ld_field(nfp_prog, reg_both(gpr_out), 0x5, reg_a(gpr_out), 1438 SHF_SC_R_ROT, 16); 1439 } 1440 1441 static void 1442 wrp_mul_u32(struct nfp_prog *nfp_prog, swreg dst_hi, swreg dst_lo, swreg lreg, 1443 swreg rreg, bool gen_high_half) 1444 { 1445 emit_mul(nfp_prog, lreg, MUL_TYPE_START, MUL_STEP_NONE, rreg); 1446 emit_mul(nfp_prog, lreg, MUL_TYPE_STEP_32x32, MUL_STEP_1, rreg); 1447 emit_mul(nfp_prog, lreg, MUL_TYPE_STEP_32x32, MUL_STEP_2, rreg); 1448 emit_mul(nfp_prog, lreg, MUL_TYPE_STEP_32x32, MUL_STEP_3, rreg); 1449 emit_mul(nfp_prog, lreg, MUL_TYPE_STEP_32x32, MUL_STEP_4, rreg); 1450 emit_mul(nfp_prog, dst_lo, MUL_TYPE_STEP_32x32, MUL_LAST, reg_none()); 1451 if (gen_high_half) 1452 emit_mul(nfp_prog, dst_hi, MUL_TYPE_STEP_32x32, MUL_LAST_2, 1453 reg_none()); 1454 else 1455 wrp_immed(nfp_prog, dst_hi, 0); 1456 } 1457 1458 static void 1459 wrp_mul_u16(struct nfp_prog *nfp_prog, swreg dst_hi, swreg dst_lo, swreg lreg, 1460 swreg rreg) 1461 { 1462 emit_mul(nfp_prog, lreg, MUL_TYPE_START, MUL_STEP_NONE, rreg); 1463 emit_mul(nfp_prog, lreg, MUL_TYPE_STEP_16x16, MUL_STEP_1, rreg); 1464 emit_mul(nfp_prog, lreg, MUL_TYPE_STEP_16x16, MUL_STEP_2, rreg); 1465 emit_mul(nfp_prog, dst_lo, MUL_TYPE_STEP_16x16, MUL_LAST, reg_none()); 1466 } 1467 1468 static int 1469 wrp_mul(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 1470 bool gen_high_half, bool ropnd_from_reg) 1471 { 1472 swreg multiplier, multiplicand, dst_hi, dst_lo; 1473 const struct bpf_insn *insn = &meta->insn; 1474 u32 lopnd_max, ropnd_max; 1475 u8 dst_reg; 1476 1477 dst_reg = insn->dst_reg; 1478 multiplicand = reg_a(dst_reg * 2); 1479 dst_hi = reg_both(dst_reg * 2 + 1); 1480 dst_lo = reg_both(dst_reg * 2); 1481 lopnd_max = meta->umax_dst; 1482 if (ropnd_from_reg) { 1483 multiplier = reg_b(insn->src_reg * 2); 1484 ropnd_max = meta->umax_src; 1485 } else { 1486 u32 imm = insn->imm; 1487 1488 multiplier = ur_load_imm_any(nfp_prog, imm, imm_b(nfp_prog)); 1489 ropnd_max = imm; 1490 } 1491 if (lopnd_max > U16_MAX || ropnd_max > U16_MAX) 1492 wrp_mul_u32(nfp_prog, dst_hi, dst_lo, multiplicand, multiplier, 1493 gen_high_half); 1494 else 1495 wrp_mul_u16(nfp_prog, dst_hi, dst_lo, multiplicand, multiplier); 1496 1497 return 0; 1498 } 1499 1500 static int wrp_div_imm(struct nfp_prog *nfp_prog, u8 dst, u64 imm) 1501 { 1502 swreg dst_both = reg_both(dst), dst_a = reg_a(dst), dst_b = reg_a(dst); 1503 struct reciprocal_value_adv rvalue; 1504 u8 pre_shift, exp; 1505 swreg magic; 1506 1507 if (imm > U32_MAX) { 1508 wrp_immed(nfp_prog, dst_both, 0); 1509 return 0; 1510 } 1511 1512 /* NOTE: because we are using "reciprocal_value_adv" which doesn't 1513 * support "divisor > (1u << 31)", we need to JIT separate NFP sequence 1514 * to handle such case which actually equals to the result of unsigned 1515 * comparison "dst >= imm" which could be calculated using the following 1516 * NFP sequence: 1517 * 1518 * alu[--, dst, -, imm] 1519 * immed[imm, 0] 1520 * alu[dst, imm, +carry, 0] 1521 * 1522 */ 1523 if (imm > 1U << 31) { 1524 swreg tmp_b = ur_load_imm_any(nfp_prog, imm, imm_b(nfp_prog)); 1525 1526 emit_alu(nfp_prog, reg_none(), dst_a, ALU_OP_SUB, tmp_b); 1527 wrp_immed(nfp_prog, imm_a(nfp_prog), 0); 1528 emit_alu(nfp_prog, dst_both, imm_a(nfp_prog), ALU_OP_ADD_C, 1529 reg_imm(0)); 1530 return 0; 1531 } 1532 1533 rvalue = reciprocal_value_adv(imm, 32); 1534 exp = rvalue.exp; 1535 if (rvalue.is_wide_m && !(imm & 1)) { 1536 pre_shift = fls(imm & -imm) - 1; 1537 rvalue = reciprocal_value_adv(imm >> pre_shift, 32 - pre_shift); 1538 } else { 1539 pre_shift = 0; 1540 } 1541 magic = ur_load_imm_any(nfp_prog, rvalue.m, imm_b(nfp_prog)); 1542 if (imm == 1U << exp) { 1543 emit_shf(nfp_prog, dst_both, reg_none(), SHF_OP_NONE, dst_b, 1544 SHF_SC_R_SHF, exp); 1545 } else if (rvalue.is_wide_m) { 1546 wrp_mul_u32(nfp_prog, imm_both(nfp_prog), reg_none(), dst_a, 1547 magic, true); 1548 emit_alu(nfp_prog, dst_both, dst_a, ALU_OP_SUB, 1549 imm_b(nfp_prog)); 1550 emit_shf(nfp_prog, dst_both, reg_none(), SHF_OP_NONE, dst_b, 1551 SHF_SC_R_SHF, 1); 1552 emit_alu(nfp_prog, dst_both, dst_a, ALU_OP_ADD, 1553 imm_b(nfp_prog)); 1554 emit_shf(nfp_prog, dst_both, reg_none(), SHF_OP_NONE, dst_b, 1555 SHF_SC_R_SHF, rvalue.sh - 1); 1556 } else { 1557 if (pre_shift) 1558 emit_shf(nfp_prog, dst_both, reg_none(), SHF_OP_NONE, 1559 dst_b, SHF_SC_R_SHF, pre_shift); 1560 wrp_mul_u32(nfp_prog, dst_both, reg_none(), dst_a, magic, true); 1561 emit_shf(nfp_prog, dst_both, reg_none(), SHF_OP_NONE, 1562 dst_b, SHF_SC_R_SHF, rvalue.sh); 1563 } 1564 1565 return 0; 1566 } 1567 1568 static int adjust_head(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1569 { 1570 swreg tmp = imm_a(nfp_prog), tmp_len = imm_b(nfp_prog); 1571 struct nfp_bpf_cap_adjust_head *adjust_head; 1572 u32 ret_einval, end; 1573 1574 adjust_head = &nfp_prog->bpf->adjust_head; 1575 1576 /* Optimized version - 5 vs 14 cycles */ 1577 if (nfp_prog->adjust_head_location != UINT_MAX) { 1578 if (WARN_ON_ONCE(nfp_prog->adjust_head_location != meta->n)) 1579 return -EINVAL; 1580 1581 emit_alu(nfp_prog, pptr_reg(nfp_prog), 1582 reg_a(2 * 2), ALU_OP_ADD, pptr_reg(nfp_prog)); 1583 emit_alu(nfp_prog, plen_reg(nfp_prog), 1584 plen_reg(nfp_prog), ALU_OP_SUB, reg_a(2 * 2)); 1585 emit_alu(nfp_prog, pv_len(nfp_prog), 1586 pv_len(nfp_prog), ALU_OP_SUB, reg_a(2 * 2)); 1587 1588 wrp_immed(nfp_prog, reg_both(0), 0); 1589 wrp_immed(nfp_prog, reg_both(1), 0); 1590 1591 /* TODO: when adjust head is guaranteed to succeed we can 1592 * also eliminate the following if (r0 == 0) branch. 1593 */ 1594 1595 return 0; 1596 } 1597 1598 ret_einval = nfp_prog_current_offset(nfp_prog) + 14; 1599 end = ret_einval + 2; 1600 1601 /* We need to use a temp because offset is just a part of the pkt ptr */ 1602 emit_alu(nfp_prog, tmp, 1603 reg_a(2 * 2), ALU_OP_ADD_2B, pptr_reg(nfp_prog)); 1604 1605 /* Validate result will fit within FW datapath constraints */ 1606 emit_alu(nfp_prog, reg_none(), 1607 tmp, ALU_OP_SUB, reg_imm(adjust_head->off_min)); 1608 emit_br(nfp_prog, BR_BLO, ret_einval, 0); 1609 emit_alu(nfp_prog, reg_none(), 1610 reg_imm(adjust_head->off_max), ALU_OP_SUB, tmp); 1611 emit_br(nfp_prog, BR_BLO, ret_einval, 0); 1612 1613 /* Validate the length is at least ETH_HLEN */ 1614 emit_alu(nfp_prog, tmp_len, 1615 plen_reg(nfp_prog), ALU_OP_SUB, reg_a(2 * 2)); 1616 emit_alu(nfp_prog, reg_none(), 1617 tmp_len, ALU_OP_SUB, reg_imm(ETH_HLEN)); 1618 emit_br(nfp_prog, BR_BMI, ret_einval, 0); 1619 1620 /* Load the ret code */ 1621 wrp_immed(nfp_prog, reg_both(0), 0); 1622 wrp_immed(nfp_prog, reg_both(1), 0); 1623 1624 /* Modify the packet metadata */ 1625 emit_ld_field(nfp_prog, pptr_reg(nfp_prog), 0x3, tmp, SHF_SC_NONE, 0); 1626 1627 /* Skip over the -EINVAL ret code (defer 2) */ 1628 emit_br(nfp_prog, BR_UNC, end, 2); 1629 1630 emit_alu(nfp_prog, plen_reg(nfp_prog), 1631 plen_reg(nfp_prog), ALU_OP_SUB, reg_a(2 * 2)); 1632 emit_alu(nfp_prog, pv_len(nfp_prog), 1633 pv_len(nfp_prog), ALU_OP_SUB, reg_a(2 * 2)); 1634 1635 /* return -EINVAL target */ 1636 if (!nfp_prog_confirm_current_offset(nfp_prog, ret_einval)) 1637 return -EINVAL; 1638 1639 wrp_immed(nfp_prog, reg_both(0), -22); 1640 wrp_immed(nfp_prog, reg_both(1), ~0); 1641 1642 if (!nfp_prog_confirm_current_offset(nfp_prog, end)) 1643 return -EINVAL; 1644 1645 return 0; 1646 } 1647 1648 static int adjust_tail(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1649 { 1650 u32 ret_einval, end; 1651 swreg plen, delta; 1652 1653 BUILD_BUG_ON(plen_reg(nfp_prog) != reg_b(STATIC_REG_PKT_LEN)); 1654 1655 plen = imm_a(nfp_prog); 1656 delta = reg_a(2 * 2); 1657 1658 ret_einval = nfp_prog_current_offset(nfp_prog) + 9; 1659 end = nfp_prog_current_offset(nfp_prog) + 11; 1660 1661 /* Calculate resulting length */ 1662 emit_alu(nfp_prog, plen, plen_reg(nfp_prog), ALU_OP_ADD, delta); 1663 /* delta == 0 is not allowed by the kernel, add must overflow to make 1664 * length smaller. 1665 */ 1666 emit_br(nfp_prog, BR_BCC, ret_einval, 0); 1667 1668 /* if (new_len < 14) then -EINVAL */ 1669 emit_alu(nfp_prog, reg_none(), plen, ALU_OP_SUB, reg_imm(ETH_HLEN)); 1670 emit_br(nfp_prog, BR_BMI, ret_einval, 0); 1671 1672 emit_alu(nfp_prog, plen_reg(nfp_prog), 1673 plen_reg(nfp_prog), ALU_OP_ADD, delta); 1674 emit_alu(nfp_prog, pv_len(nfp_prog), 1675 pv_len(nfp_prog), ALU_OP_ADD, delta); 1676 1677 emit_br(nfp_prog, BR_UNC, end, 2); 1678 wrp_immed(nfp_prog, reg_both(0), 0); 1679 wrp_immed(nfp_prog, reg_both(1), 0); 1680 1681 if (!nfp_prog_confirm_current_offset(nfp_prog, ret_einval)) 1682 return -EINVAL; 1683 1684 wrp_immed(nfp_prog, reg_both(0), -22); 1685 wrp_immed(nfp_prog, reg_both(1), ~0); 1686 1687 if (!nfp_prog_confirm_current_offset(nfp_prog, end)) 1688 return -EINVAL; 1689 1690 return 0; 1691 } 1692 1693 static int 1694 map_call_stack_common(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1695 { 1696 bool load_lm_ptr; 1697 u32 ret_tgt; 1698 s64 lm_off; 1699 1700 /* We only have to reload LM0 if the key is not at start of stack */ 1701 lm_off = nfp_prog->stack_frame_depth; 1702 lm_off += meta->arg2.reg.var_off.value + meta->arg2.reg.off; 1703 load_lm_ptr = meta->arg2.var_off || lm_off; 1704 1705 /* Set LM0 to start of key */ 1706 if (load_lm_ptr) 1707 emit_csr_wr(nfp_prog, reg_b(2 * 2), NFP_CSR_ACT_LM_ADDR0); 1708 if (meta->func_id == BPF_FUNC_map_update_elem) 1709 emit_csr_wr(nfp_prog, reg_b(3 * 2), NFP_CSR_ACT_LM_ADDR2); 1710 1711 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO + meta->func_id, 1712 2, RELO_BR_HELPER); 1713 ret_tgt = nfp_prog_current_offset(nfp_prog) + 2; 1714 1715 /* Load map ID into A0 */ 1716 wrp_mov(nfp_prog, reg_a(0), reg_a(2)); 1717 1718 /* Load the return address into B0 */ 1719 wrp_immed_relo(nfp_prog, reg_b(0), ret_tgt, RELO_IMMED_REL); 1720 1721 if (!nfp_prog_confirm_current_offset(nfp_prog, ret_tgt)) 1722 return -EINVAL; 1723 1724 /* Reset the LM0 pointer */ 1725 if (!load_lm_ptr) 1726 return 0; 1727 1728 emit_csr_wr(nfp_prog, stack_reg(nfp_prog), NFP_CSR_ACT_LM_ADDR0); 1729 wrp_nops(nfp_prog, 3); 1730 1731 return 0; 1732 } 1733 1734 static int 1735 nfp_get_prandom_u32(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1736 { 1737 __emit_csr_rd(nfp_prog, NFP_CSR_PSEUDO_RND_NUM); 1738 /* CSR value is read in following immed[gpr, 0] */ 1739 emit_immed(nfp_prog, reg_both(0), 0, 1740 IMMED_WIDTH_ALL, false, IMMED_SHIFT_0B); 1741 emit_immed(nfp_prog, reg_both(1), 0, 1742 IMMED_WIDTH_ALL, false, IMMED_SHIFT_0B); 1743 return 0; 1744 } 1745 1746 static int 1747 nfp_perf_event_output(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1748 { 1749 swreg ptr_type; 1750 u32 ret_tgt; 1751 1752 ptr_type = ur_load_imm_any(nfp_prog, meta->arg1.type, imm_a(nfp_prog)); 1753 1754 ret_tgt = nfp_prog_current_offset(nfp_prog) + 3; 1755 1756 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO + meta->func_id, 1757 2, RELO_BR_HELPER); 1758 1759 /* Load ptr type into A1 */ 1760 wrp_mov(nfp_prog, reg_a(1), ptr_type); 1761 1762 /* Load the return address into B0 */ 1763 wrp_immed_relo(nfp_prog, reg_b(0), ret_tgt, RELO_IMMED_REL); 1764 1765 if (!nfp_prog_confirm_current_offset(nfp_prog, ret_tgt)) 1766 return -EINVAL; 1767 1768 return 0; 1769 } 1770 1771 static int 1772 nfp_queue_select(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1773 { 1774 u32 jmp_tgt; 1775 1776 jmp_tgt = nfp_prog_current_offset(nfp_prog) + 5; 1777 1778 /* Make sure the queue id fits into FW field */ 1779 emit_alu(nfp_prog, reg_none(), reg_a(meta->insn.src_reg * 2), 1780 ALU_OP_AND_NOT_B, reg_imm(0xff)); 1781 emit_br(nfp_prog, BR_BEQ, jmp_tgt, 2); 1782 1783 /* Set the 'queue selected' bit and the queue value */ 1784 emit_shf(nfp_prog, pv_qsel_set(nfp_prog), 1785 pv_qsel_set(nfp_prog), SHF_OP_OR, reg_imm(1), 1786 SHF_SC_L_SHF, PKT_VEL_QSEL_SET_BIT); 1787 emit_ld_field(nfp_prog, 1788 pv_qsel_val(nfp_prog), 0x1, reg_b(meta->insn.src_reg * 2), 1789 SHF_SC_NONE, 0); 1790 /* Delay slots end here, we will jump over next instruction if queue 1791 * value fits into the field. 1792 */ 1793 emit_ld_field(nfp_prog, 1794 pv_qsel_val(nfp_prog), 0x1, reg_imm(NFP_NET_RXR_MAX), 1795 SHF_SC_NONE, 0); 1796 1797 if (!nfp_prog_confirm_current_offset(nfp_prog, jmp_tgt)) 1798 return -EINVAL; 1799 1800 return 0; 1801 } 1802 1803 /* --- Callbacks --- */ 1804 static int mov_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1805 { 1806 const struct bpf_insn *insn = &meta->insn; 1807 u8 dst = insn->dst_reg * 2; 1808 u8 src = insn->src_reg * 2; 1809 1810 if (insn->src_reg == BPF_REG_10) { 1811 swreg stack_depth_reg; 1812 1813 stack_depth_reg = ur_load_imm_any(nfp_prog, 1814 nfp_prog->stack_frame_depth, 1815 stack_imm(nfp_prog)); 1816 emit_alu(nfp_prog, reg_both(dst), stack_reg(nfp_prog), 1817 ALU_OP_ADD, stack_depth_reg); 1818 wrp_immed(nfp_prog, reg_both(dst + 1), 0); 1819 } else { 1820 wrp_reg_mov(nfp_prog, dst, src); 1821 wrp_reg_mov(nfp_prog, dst + 1, src + 1); 1822 } 1823 1824 return 0; 1825 } 1826 1827 static int mov_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1828 { 1829 u64 imm = meta->insn.imm; /* sign extend */ 1830 1831 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2), imm & ~0U); 1832 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), imm >> 32); 1833 1834 return 0; 1835 } 1836 1837 static int xor_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1838 { 1839 return wrp_alu64_reg(nfp_prog, meta, ALU_OP_XOR); 1840 } 1841 1842 static int xor_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1843 { 1844 return wrp_alu64_imm(nfp_prog, meta, ALU_OP_XOR, !meta->insn.imm); 1845 } 1846 1847 static int and_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1848 { 1849 return wrp_alu64_reg(nfp_prog, meta, ALU_OP_AND); 1850 } 1851 1852 static int and_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1853 { 1854 return wrp_alu64_imm(nfp_prog, meta, ALU_OP_AND, !~meta->insn.imm); 1855 } 1856 1857 static int or_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1858 { 1859 return wrp_alu64_reg(nfp_prog, meta, ALU_OP_OR); 1860 } 1861 1862 static int or_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1863 { 1864 return wrp_alu64_imm(nfp_prog, meta, ALU_OP_OR, !meta->insn.imm); 1865 } 1866 1867 static int add_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1868 { 1869 const struct bpf_insn *insn = &meta->insn; 1870 1871 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2), 1872 reg_a(insn->dst_reg * 2), ALU_OP_ADD, 1873 reg_b(insn->src_reg * 2)); 1874 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 1875 reg_a(insn->dst_reg * 2 + 1), ALU_OP_ADD_C, 1876 reg_b(insn->src_reg * 2 + 1)); 1877 1878 return 0; 1879 } 1880 1881 static int add_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1882 { 1883 const struct bpf_insn *insn = &meta->insn; 1884 u64 imm = insn->imm; /* sign extend */ 1885 1886 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, ALU_OP_ADD, imm & ~0U); 1887 wrp_alu_imm(nfp_prog, insn->dst_reg * 2 + 1, ALU_OP_ADD_C, imm >> 32); 1888 1889 return 0; 1890 } 1891 1892 static int sub_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1893 { 1894 const struct bpf_insn *insn = &meta->insn; 1895 1896 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2), 1897 reg_a(insn->dst_reg * 2), ALU_OP_SUB, 1898 reg_b(insn->src_reg * 2)); 1899 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 1900 reg_a(insn->dst_reg * 2 + 1), ALU_OP_SUB_C, 1901 reg_b(insn->src_reg * 2 + 1)); 1902 1903 return 0; 1904 } 1905 1906 static int sub_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1907 { 1908 const struct bpf_insn *insn = &meta->insn; 1909 u64 imm = insn->imm; /* sign extend */ 1910 1911 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, ALU_OP_SUB, imm & ~0U); 1912 wrp_alu_imm(nfp_prog, insn->dst_reg * 2 + 1, ALU_OP_SUB_C, imm >> 32); 1913 1914 return 0; 1915 } 1916 1917 static int mul_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1918 { 1919 return wrp_mul(nfp_prog, meta, true, true); 1920 } 1921 1922 static int mul_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1923 { 1924 return wrp_mul(nfp_prog, meta, true, false); 1925 } 1926 1927 static int div_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1928 { 1929 const struct bpf_insn *insn = &meta->insn; 1930 1931 return wrp_div_imm(nfp_prog, insn->dst_reg * 2, insn->imm); 1932 } 1933 1934 static int div_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1935 { 1936 /* NOTE: verifier hook has rejected cases for which verifier doesn't 1937 * know whether the source operand is constant or not. 1938 */ 1939 return wrp_div_imm(nfp_prog, meta->insn.dst_reg * 2, meta->umin_src); 1940 } 1941 1942 static int neg_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1943 { 1944 const struct bpf_insn *insn = &meta->insn; 1945 1946 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2), reg_imm(0), 1947 ALU_OP_SUB, reg_b(insn->dst_reg * 2)); 1948 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2 + 1), reg_imm(0), 1949 ALU_OP_SUB_C, reg_b(insn->dst_reg * 2 + 1)); 1950 1951 return 0; 1952 } 1953 1954 /* Pseudo code: 1955 * if shift_amt >= 32 1956 * dst_high = dst_low << shift_amt[4:0] 1957 * dst_low = 0; 1958 * else 1959 * dst_high = (dst_high, dst_low) >> (32 - shift_amt) 1960 * dst_low = dst_low << shift_amt 1961 * 1962 * The indirect shift will use the same logic at runtime. 1963 */ 1964 static int __shl_imm64(struct nfp_prog *nfp_prog, u8 dst, u8 shift_amt) 1965 { 1966 if (shift_amt < 32) { 1967 emit_shf(nfp_prog, reg_both(dst + 1), reg_a(dst + 1), 1968 SHF_OP_NONE, reg_b(dst), SHF_SC_R_DSHF, 1969 32 - shift_amt); 1970 emit_shf(nfp_prog, reg_both(dst), reg_none(), SHF_OP_NONE, 1971 reg_b(dst), SHF_SC_L_SHF, shift_amt); 1972 } else if (shift_amt == 32) { 1973 wrp_reg_mov(nfp_prog, dst + 1, dst); 1974 wrp_immed(nfp_prog, reg_both(dst), 0); 1975 } else if (shift_amt > 32) { 1976 emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_NONE, 1977 reg_b(dst), SHF_SC_L_SHF, shift_amt - 32); 1978 wrp_immed(nfp_prog, reg_both(dst), 0); 1979 } 1980 1981 return 0; 1982 } 1983 1984 static int shl_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 1985 { 1986 const struct bpf_insn *insn = &meta->insn; 1987 u8 dst = insn->dst_reg * 2; 1988 1989 return __shl_imm64(nfp_prog, dst, insn->imm); 1990 } 1991 1992 static void shl_reg64_lt32_high(struct nfp_prog *nfp_prog, u8 dst, u8 src) 1993 { 1994 emit_alu(nfp_prog, imm_both(nfp_prog), reg_imm(32), ALU_OP_SUB, 1995 reg_b(src)); 1996 emit_alu(nfp_prog, reg_none(), imm_a(nfp_prog), ALU_OP_OR, reg_imm(0)); 1997 emit_shf_indir(nfp_prog, reg_both(dst + 1), reg_a(dst + 1), SHF_OP_NONE, 1998 reg_b(dst), SHF_SC_R_DSHF); 1999 } 2000 2001 /* NOTE: for indirect left shift, HIGH part should be calculated first. */ 2002 static void shl_reg64_lt32_low(struct nfp_prog *nfp_prog, u8 dst, u8 src) 2003 { 2004 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0)); 2005 emit_shf_indir(nfp_prog, reg_both(dst), reg_none(), SHF_OP_NONE, 2006 reg_b(dst), SHF_SC_L_SHF); 2007 } 2008 2009 static void shl_reg64_lt32(struct nfp_prog *nfp_prog, u8 dst, u8 src) 2010 { 2011 shl_reg64_lt32_high(nfp_prog, dst, src); 2012 shl_reg64_lt32_low(nfp_prog, dst, src); 2013 } 2014 2015 static void shl_reg64_ge32(struct nfp_prog *nfp_prog, u8 dst, u8 src) 2016 { 2017 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0)); 2018 emit_shf_indir(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_NONE, 2019 reg_b(dst), SHF_SC_L_SHF); 2020 wrp_immed(nfp_prog, reg_both(dst), 0); 2021 } 2022 2023 static int shl_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2024 { 2025 const struct bpf_insn *insn = &meta->insn; 2026 u64 umin, umax; 2027 u8 dst, src; 2028 2029 dst = insn->dst_reg * 2; 2030 umin = meta->umin_src; 2031 umax = meta->umax_src; 2032 if (umin == umax) 2033 return __shl_imm64(nfp_prog, dst, umin); 2034 2035 src = insn->src_reg * 2; 2036 if (umax < 32) { 2037 shl_reg64_lt32(nfp_prog, dst, src); 2038 } else if (umin >= 32) { 2039 shl_reg64_ge32(nfp_prog, dst, src); 2040 } else { 2041 /* Generate different instruction sequences depending on runtime 2042 * value of shift amount. 2043 */ 2044 u16 label_ge32, label_end; 2045 2046 label_ge32 = nfp_prog_current_offset(nfp_prog) + 7; 2047 emit_br_bset(nfp_prog, reg_a(src), 5, label_ge32, 0); 2048 2049 shl_reg64_lt32_high(nfp_prog, dst, src); 2050 label_end = nfp_prog_current_offset(nfp_prog) + 6; 2051 emit_br(nfp_prog, BR_UNC, label_end, 2); 2052 /* shl_reg64_lt32_low packed in delay slot. */ 2053 shl_reg64_lt32_low(nfp_prog, dst, src); 2054 2055 if (!nfp_prog_confirm_current_offset(nfp_prog, label_ge32)) 2056 return -EINVAL; 2057 shl_reg64_ge32(nfp_prog, dst, src); 2058 2059 if (!nfp_prog_confirm_current_offset(nfp_prog, label_end)) 2060 return -EINVAL; 2061 } 2062 2063 return 0; 2064 } 2065 2066 /* Pseudo code: 2067 * if shift_amt >= 32 2068 * dst_high = 0; 2069 * dst_low = dst_high >> shift_amt[4:0] 2070 * else 2071 * dst_high = dst_high >> shift_amt 2072 * dst_low = (dst_high, dst_low) >> shift_amt 2073 * 2074 * The indirect shift will use the same logic at runtime. 2075 */ 2076 static int __shr_imm64(struct nfp_prog *nfp_prog, u8 dst, u8 shift_amt) 2077 { 2078 if (shift_amt < 32) { 2079 emit_shf(nfp_prog, reg_both(dst), reg_a(dst + 1), SHF_OP_NONE, 2080 reg_b(dst), SHF_SC_R_DSHF, shift_amt); 2081 emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_NONE, 2082 reg_b(dst + 1), SHF_SC_R_SHF, shift_amt); 2083 } else if (shift_amt == 32) { 2084 wrp_reg_mov(nfp_prog, dst, dst + 1); 2085 wrp_immed(nfp_prog, reg_both(dst + 1), 0); 2086 } else if (shift_amt > 32) { 2087 emit_shf(nfp_prog, reg_both(dst), reg_none(), SHF_OP_NONE, 2088 reg_b(dst + 1), SHF_SC_R_SHF, shift_amt - 32); 2089 wrp_immed(nfp_prog, reg_both(dst + 1), 0); 2090 } 2091 2092 return 0; 2093 } 2094 2095 static int shr_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2096 { 2097 const struct bpf_insn *insn = &meta->insn; 2098 u8 dst = insn->dst_reg * 2; 2099 2100 return __shr_imm64(nfp_prog, dst, insn->imm); 2101 } 2102 2103 /* NOTE: for indirect right shift, LOW part should be calculated first. */ 2104 static void shr_reg64_lt32_high(struct nfp_prog *nfp_prog, u8 dst, u8 src) 2105 { 2106 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0)); 2107 emit_shf_indir(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_NONE, 2108 reg_b(dst + 1), SHF_SC_R_SHF); 2109 } 2110 2111 static void shr_reg64_lt32_low(struct nfp_prog *nfp_prog, u8 dst, u8 src) 2112 { 2113 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0)); 2114 emit_shf_indir(nfp_prog, reg_both(dst), reg_a(dst + 1), SHF_OP_NONE, 2115 reg_b(dst), SHF_SC_R_DSHF); 2116 } 2117 2118 static void shr_reg64_lt32(struct nfp_prog *nfp_prog, u8 dst, u8 src) 2119 { 2120 shr_reg64_lt32_low(nfp_prog, dst, src); 2121 shr_reg64_lt32_high(nfp_prog, dst, src); 2122 } 2123 2124 static void shr_reg64_ge32(struct nfp_prog *nfp_prog, u8 dst, u8 src) 2125 { 2126 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0)); 2127 emit_shf_indir(nfp_prog, reg_both(dst), reg_none(), SHF_OP_NONE, 2128 reg_b(dst + 1), SHF_SC_R_SHF); 2129 wrp_immed(nfp_prog, reg_both(dst + 1), 0); 2130 } 2131 2132 static int shr_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2133 { 2134 const struct bpf_insn *insn = &meta->insn; 2135 u64 umin, umax; 2136 u8 dst, src; 2137 2138 dst = insn->dst_reg * 2; 2139 umin = meta->umin_src; 2140 umax = meta->umax_src; 2141 if (umin == umax) 2142 return __shr_imm64(nfp_prog, dst, umin); 2143 2144 src = insn->src_reg * 2; 2145 if (umax < 32) { 2146 shr_reg64_lt32(nfp_prog, dst, src); 2147 } else if (umin >= 32) { 2148 shr_reg64_ge32(nfp_prog, dst, src); 2149 } else { 2150 /* Generate different instruction sequences depending on runtime 2151 * value of shift amount. 2152 */ 2153 u16 label_ge32, label_end; 2154 2155 label_ge32 = nfp_prog_current_offset(nfp_prog) + 6; 2156 emit_br_bset(nfp_prog, reg_a(src), 5, label_ge32, 0); 2157 shr_reg64_lt32_low(nfp_prog, dst, src); 2158 label_end = nfp_prog_current_offset(nfp_prog) + 6; 2159 emit_br(nfp_prog, BR_UNC, label_end, 2); 2160 /* shr_reg64_lt32_high packed in delay slot. */ 2161 shr_reg64_lt32_high(nfp_prog, dst, src); 2162 2163 if (!nfp_prog_confirm_current_offset(nfp_prog, label_ge32)) 2164 return -EINVAL; 2165 shr_reg64_ge32(nfp_prog, dst, src); 2166 2167 if (!nfp_prog_confirm_current_offset(nfp_prog, label_end)) 2168 return -EINVAL; 2169 } 2170 2171 return 0; 2172 } 2173 2174 /* Code logic is the same as __shr_imm64 except ashr requires signedness bit 2175 * told through PREV_ALU result. 2176 */ 2177 static int __ashr_imm64(struct nfp_prog *nfp_prog, u8 dst, u8 shift_amt) 2178 { 2179 if (shift_amt < 32) { 2180 emit_shf(nfp_prog, reg_both(dst), reg_a(dst + 1), SHF_OP_NONE, 2181 reg_b(dst), SHF_SC_R_DSHF, shift_amt); 2182 /* Set signedness bit. */ 2183 emit_alu(nfp_prog, reg_none(), reg_a(dst + 1), ALU_OP_OR, 2184 reg_imm(0)); 2185 emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_ASHR, 2186 reg_b(dst + 1), SHF_SC_R_SHF, shift_amt); 2187 } else if (shift_amt == 32) { 2188 /* NOTE: this also helps setting signedness bit. */ 2189 wrp_reg_mov(nfp_prog, dst, dst + 1); 2190 emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_ASHR, 2191 reg_b(dst + 1), SHF_SC_R_SHF, 31); 2192 } else if (shift_amt > 32) { 2193 emit_alu(nfp_prog, reg_none(), reg_a(dst + 1), ALU_OP_OR, 2194 reg_imm(0)); 2195 emit_shf(nfp_prog, reg_both(dst), reg_none(), SHF_OP_ASHR, 2196 reg_b(dst + 1), SHF_SC_R_SHF, shift_amt - 32); 2197 emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_ASHR, 2198 reg_b(dst + 1), SHF_SC_R_SHF, 31); 2199 } 2200 2201 return 0; 2202 } 2203 2204 static int ashr_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2205 { 2206 const struct bpf_insn *insn = &meta->insn; 2207 u8 dst = insn->dst_reg * 2; 2208 2209 return __ashr_imm64(nfp_prog, dst, insn->imm); 2210 } 2211 2212 static void ashr_reg64_lt32_high(struct nfp_prog *nfp_prog, u8 dst, u8 src) 2213 { 2214 /* NOTE: the first insn will set both indirect shift amount (source A) 2215 * and signedness bit (MSB of result). 2216 */ 2217 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_b(dst + 1)); 2218 emit_shf_indir(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_ASHR, 2219 reg_b(dst + 1), SHF_SC_R_SHF); 2220 } 2221 2222 static void ashr_reg64_lt32_low(struct nfp_prog *nfp_prog, u8 dst, u8 src) 2223 { 2224 /* NOTE: it is the same as logic shift because we don't need to shift in 2225 * signedness bit when the shift amount is less than 32. 2226 */ 2227 return shr_reg64_lt32_low(nfp_prog, dst, src); 2228 } 2229 2230 static void ashr_reg64_lt32(struct nfp_prog *nfp_prog, u8 dst, u8 src) 2231 { 2232 ashr_reg64_lt32_low(nfp_prog, dst, src); 2233 ashr_reg64_lt32_high(nfp_prog, dst, src); 2234 } 2235 2236 static void ashr_reg64_ge32(struct nfp_prog *nfp_prog, u8 dst, u8 src) 2237 { 2238 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_b(dst + 1)); 2239 emit_shf_indir(nfp_prog, reg_both(dst), reg_none(), SHF_OP_ASHR, 2240 reg_b(dst + 1), SHF_SC_R_SHF); 2241 emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_ASHR, 2242 reg_b(dst + 1), SHF_SC_R_SHF, 31); 2243 } 2244 2245 /* Like ashr_imm64, but need to use indirect shift. */ 2246 static int ashr_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2247 { 2248 const struct bpf_insn *insn = &meta->insn; 2249 u64 umin, umax; 2250 u8 dst, src; 2251 2252 dst = insn->dst_reg * 2; 2253 umin = meta->umin_src; 2254 umax = meta->umax_src; 2255 if (umin == umax) 2256 return __ashr_imm64(nfp_prog, dst, umin); 2257 2258 src = insn->src_reg * 2; 2259 if (umax < 32) { 2260 ashr_reg64_lt32(nfp_prog, dst, src); 2261 } else if (umin >= 32) { 2262 ashr_reg64_ge32(nfp_prog, dst, src); 2263 } else { 2264 u16 label_ge32, label_end; 2265 2266 label_ge32 = nfp_prog_current_offset(nfp_prog) + 6; 2267 emit_br_bset(nfp_prog, reg_a(src), 5, label_ge32, 0); 2268 ashr_reg64_lt32_low(nfp_prog, dst, src); 2269 label_end = nfp_prog_current_offset(nfp_prog) + 6; 2270 emit_br(nfp_prog, BR_UNC, label_end, 2); 2271 /* ashr_reg64_lt32_high packed in delay slot. */ 2272 ashr_reg64_lt32_high(nfp_prog, dst, src); 2273 2274 if (!nfp_prog_confirm_current_offset(nfp_prog, label_ge32)) 2275 return -EINVAL; 2276 ashr_reg64_ge32(nfp_prog, dst, src); 2277 2278 if (!nfp_prog_confirm_current_offset(nfp_prog, label_end)) 2279 return -EINVAL; 2280 } 2281 2282 return 0; 2283 } 2284 2285 static int mov_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2286 { 2287 const struct bpf_insn *insn = &meta->insn; 2288 2289 wrp_reg_mov(nfp_prog, insn->dst_reg * 2, insn->src_reg * 2); 2290 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0); 2291 2292 return 0; 2293 } 2294 2295 static int mov_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2296 { 2297 const struct bpf_insn *insn = &meta->insn; 2298 2299 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2), insn->imm); 2300 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0); 2301 2302 return 0; 2303 } 2304 2305 static int xor_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2306 { 2307 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_XOR); 2308 } 2309 2310 static int xor_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2311 { 2312 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_XOR, !~meta->insn.imm); 2313 } 2314 2315 static int and_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2316 { 2317 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_AND); 2318 } 2319 2320 static int and_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2321 { 2322 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_AND, !~meta->insn.imm); 2323 } 2324 2325 static int or_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2326 { 2327 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_OR); 2328 } 2329 2330 static int or_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2331 { 2332 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_OR, !meta->insn.imm); 2333 } 2334 2335 static int add_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2336 { 2337 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_ADD); 2338 } 2339 2340 static int add_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2341 { 2342 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_ADD, !meta->insn.imm); 2343 } 2344 2345 static int sub_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2346 { 2347 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_SUB); 2348 } 2349 2350 static int sub_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2351 { 2352 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_SUB, !meta->insn.imm); 2353 } 2354 2355 static int mul_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2356 { 2357 return wrp_mul(nfp_prog, meta, false, true); 2358 } 2359 2360 static int mul_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2361 { 2362 return wrp_mul(nfp_prog, meta, false, false); 2363 } 2364 2365 static int div_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2366 { 2367 return div_reg64(nfp_prog, meta); 2368 } 2369 2370 static int div_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2371 { 2372 return div_imm64(nfp_prog, meta); 2373 } 2374 2375 static int neg_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2376 { 2377 u8 dst = meta->insn.dst_reg * 2; 2378 2379 emit_alu(nfp_prog, reg_both(dst), reg_imm(0), ALU_OP_SUB, reg_b(dst)); 2380 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0); 2381 2382 return 0; 2383 } 2384 2385 static int shl_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2386 { 2387 const struct bpf_insn *insn = &meta->insn; 2388 2389 if (!insn->imm) 2390 return 1; /* TODO: zero shift means indirect */ 2391 2392 emit_shf(nfp_prog, reg_both(insn->dst_reg * 2), 2393 reg_none(), SHF_OP_NONE, reg_b(insn->dst_reg * 2), 2394 SHF_SC_L_SHF, insn->imm); 2395 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0); 2396 2397 return 0; 2398 } 2399 2400 static int end_reg32(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2401 { 2402 const struct bpf_insn *insn = &meta->insn; 2403 u8 gpr = insn->dst_reg * 2; 2404 2405 switch (insn->imm) { 2406 case 16: 2407 emit_ld_field(nfp_prog, reg_both(gpr), 0x9, reg_b(gpr), 2408 SHF_SC_R_ROT, 8); 2409 emit_ld_field(nfp_prog, reg_both(gpr), 0xe, reg_a(gpr), 2410 SHF_SC_R_SHF, 16); 2411 2412 wrp_immed(nfp_prog, reg_both(gpr + 1), 0); 2413 break; 2414 case 32: 2415 wrp_end32(nfp_prog, reg_a(gpr), gpr); 2416 wrp_immed(nfp_prog, reg_both(gpr + 1), 0); 2417 break; 2418 case 64: 2419 wrp_mov(nfp_prog, imm_a(nfp_prog), reg_b(gpr + 1)); 2420 2421 wrp_end32(nfp_prog, reg_a(gpr), gpr + 1); 2422 wrp_end32(nfp_prog, imm_a(nfp_prog), gpr); 2423 break; 2424 } 2425 2426 return 0; 2427 } 2428 2429 static int imm_ld8_part2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2430 { 2431 struct nfp_insn_meta *prev = nfp_meta_prev(meta); 2432 u32 imm_lo, imm_hi; 2433 u8 dst; 2434 2435 dst = prev->insn.dst_reg * 2; 2436 imm_lo = prev->insn.imm; 2437 imm_hi = meta->insn.imm; 2438 2439 wrp_immed(nfp_prog, reg_both(dst), imm_lo); 2440 2441 /* mov is always 1 insn, load imm may be two, so try to use mov */ 2442 if (imm_hi == imm_lo) 2443 wrp_mov(nfp_prog, reg_both(dst + 1), reg_a(dst)); 2444 else 2445 wrp_immed(nfp_prog, reg_both(dst + 1), imm_hi); 2446 2447 return 0; 2448 } 2449 2450 static int imm_ld8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2451 { 2452 meta->double_cb = imm_ld8_part2; 2453 return 0; 2454 } 2455 2456 static int data_ld1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2457 { 2458 return construct_data_ld(nfp_prog, meta->insn.imm, 1); 2459 } 2460 2461 static int data_ld2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2462 { 2463 return construct_data_ld(nfp_prog, meta->insn.imm, 2); 2464 } 2465 2466 static int data_ld4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2467 { 2468 return construct_data_ld(nfp_prog, meta->insn.imm, 4); 2469 } 2470 2471 static int data_ind_ld1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2472 { 2473 return construct_data_ind_ld(nfp_prog, meta->insn.imm, 2474 meta->insn.src_reg * 2, 1); 2475 } 2476 2477 static int data_ind_ld2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2478 { 2479 return construct_data_ind_ld(nfp_prog, meta->insn.imm, 2480 meta->insn.src_reg * 2, 2); 2481 } 2482 2483 static int data_ind_ld4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2484 { 2485 return construct_data_ind_ld(nfp_prog, meta->insn.imm, 2486 meta->insn.src_reg * 2, 4); 2487 } 2488 2489 static int 2490 mem_ldx_stack(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2491 unsigned int size, unsigned int ptr_off) 2492 { 2493 return mem_op_stack(nfp_prog, meta, size, ptr_off, 2494 meta->insn.dst_reg * 2, meta->insn.src_reg * 2, 2495 true, wrp_lmem_load); 2496 } 2497 2498 static int mem_ldx_skb(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2499 u8 size) 2500 { 2501 swreg dst = reg_both(meta->insn.dst_reg * 2); 2502 2503 switch (meta->insn.off) { 2504 case offsetof(struct __sk_buff, len): 2505 if (size != FIELD_SIZEOF(struct __sk_buff, len)) 2506 return -EOPNOTSUPP; 2507 wrp_mov(nfp_prog, dst, plen_reg(nfp_prog)); 2508 break; 2509 case offsetof(struct __sk_buff, data): 2510 if (size != FIELD_SIZEOF(struct __sk_buff, data)) 2511 return -EOPNOTSUPP; 2512 wrp_mov(nfp_prog, dst, pptr_reg(nfp_prog)); 2513 break; 2514 case offsetof(struct __sk_buff, data_end): 2515 if (size != FIELD_SIZEOF(struct __sk_buff, data_end)) 2516 return -EOPNOTSUPP; 2517 emit_alu(nfp_prog, dst, 2518 plen_reg(nfp_prog), ALU_OP_ADD, pptr_reg(nfp_prog)); 2519 break; 2520 default: 2521 return -EOPNOTSUPP; 2522 } 2523 2524 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0); 2525 2526 return 0; 2527 } 2528 2529 static int mem_ldx_xdp(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2530 u8 size) 2531 { 2532 swreg dst = reg_both(meta->insn.dst_reg * 2); 2533 2534 switch (meta->insn.off) { 2535 case offsetof(struct xdp_md, data): 2536 if (size != FIELD_SIZEOF(struct xdp_md, data)) 2537 return -EOPNOTSUPP; 2538 wrp_mov(nfp_prog, dst, pptr_reg(nfp_prog)); 2539 break; 2540 case offsetof(struct xdp_md, data_end): 2541 if (size != FIELD_SIZEOF(struct xdp_md, data_end)) 2542 return -EOPNOTSUPP; 2543 emit_alu(nfp_prog, dst, 2544 plen_reg(nfp_prog), ALU_OP_ADD, pptr_reg(nfp_prog)); 2545 break; 2546 default: 2547 return -EOPNOTSUPP; 2548 } 2549 2550 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0); 2551 2552 return 0; 2553 } 2554 2555 static int 2556 mem_ldx_data(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2557 unsigned int size) 2558 { 2559 swreg tmp_reg; 2560 2561 tmp_reg = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog)); 2562 2563 return data_ld_host_order_addr32(nfp_prog, meta->insn.src_reg * 2, 2564 tmp_reg, meta->insn.dst_reg * 2, size); 2565 } 2566 2567 static int 2568 mem_ldx_emem(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2569 unsigned int size) 2570 { 2571 swreg tmp_reg; 2572 2573 tmp_reg = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog)); 2574 2575 return data_ld_host_order_addr40(nfp_prog, meta->insn.src_reg * 2, 2576 tmp_reg, meta->insn.dst_reg * 2, size); 2577 } 2578 2579 static void 2580 mem_ldx_data_init_pktcache(struct nfp_prog *nfp_prog, 2581 struct nfp_insn_meta *meta) 2582 { 2583 s16 range_start = meta->pkt_cache.range_start; 2584 s16 range_end = meta->pkt_cache.range_end; 2585 swreg src_base, off; 2586 u8 xfer_num, len; 2587 bool indir; 2588 2589 off = re_load_imm_any(nfp_prog, range_start, imm_b(nfp_prog)); 2590 src_base = reg_a(meta->insn.src_reg * 2); 2591 len = range_end - range_start; 2592 xfer_num = round_up(len, REG_WIDTH) / REG_WIDTH; 2593 2594 indir = len > 8 * REG_WIDTH; 2595 /* Setup PREV_ALU for indirect mode. */ 2596 if (indir) 2597 wrp_immed(nfp_prog, reg_none(), 2598 CMD_OVE_LEN | FIELD_PREP(CMD_OV_LEN, xfer_num - 1)); 2599 2600 /* Cache memory into transfer-in registers. */ 2601 emit_cmd_any(nfp_prog, CMD_TGT_READ32_SWAP, CMD_MODE_32b, 0, src_base, 2602 off, xfer_num - 1, CMD_CTX_SWAP, indir); 2603 } 2604 2605 static int 2606 mem_ldx_data_from_pktcache_unaligned(struct nfp_prog *nfp_prog, 2607 struct nfp_insn_meta *meta, 2608 unsigned int size) 2609 { 2610 s16 range_start = meta->pkt_cache.range_start; 2611 s16 insn_off = meta->insn.off - range_start; 2612 swreg dst_lo, dst_hi, src_lo, src_mid; 2613 u8 dst_gpr = meta->insn.dst_reg * 2; 2614 u8 len_lo = size, len_mid = 0; 2615 u8 idx = insn_off / REG_WIDTH; 2616 u8 off = insn_off % REG_WIDTH; 2617 2618 dst_hi = reg_both(dst_gpr + 1); 2619 dst_lo = reg_both(dst_gpr); 2620 src_lo = reg_xfer(idx); 2621 2622 /* The read length could involve as many as three registers. */ 2623 if (size > REG_WIDTH - off) { 2624 /* Calculate the part in the second register. */ 2625 len_lo = REG_WIDTH - off; 2626 len_mid = size - len_lo; 2627 2628 /* Calculate the part in the third register. */ 2629 if (size > 2 * REG_WIDTH - off) 2630 len_mid = REG_WIDTH; 2631 } 2632 2633 wrp_reg_subpart(nfp_prog, dst_lo, src_lo, len_lo, off); 2634 2635 if (!len_mid) { 2636 wrp_immed(nfp_prog, dst_hi, 0); 2637 return 0; 2638 } 2639 2640 src_mid = reg_xfer(idx + 1); 2641 2642 if (size <= REG_WIDTH) { 2643 wrp_reg_or_subpart(nfp_prog, dst_lo, src_mid, len_mid, len_lo); 2644 wrp_immed(nfp_prog, dst_hi, 0); 2645 } else { 2646 swreg src_hi = reg_xfer(idx + 2); 2647 2648 wrp_reg_or_subpart(nfp_prog, dst_lo, src_mid, 2649 REG_WIDTH - len_lo, len_lo); 2650 wrp_reg_subpart(nfp_prog, dst_hi, src_mid, len_lo, 2651 REG_WIDTH - len_lo); 2652 wrp_reg_or_subpart(nfp_prog, dst_hi, src_hi, REG_WIDTH - len_lo, 2653 len_lo); 2654 } 2655 2656 return 0; 2657 } 2658 2659 static int 2660 mem_ldx_data_from_pktcache_aligned(struct nfp_prog *nfp_prog, 2661 struct nfp_insn_meta *meta, 2662 unsigned int size) 2663 { 2664 swreg dst_lo, dst_hi, src_lo; 2665 u8 dst_gpr, idx; 2666 2667 idx = (meta->insn.off - meta->pkt_cache.range_start) / REG_WIDTH; 2668 dst_gpr = meta->insn.dst_reg * 2; 2669 dst_hi = reg_both(dst_gpr + 1); 2670 dst_lo = reg_both(dst_gpr); 2671 src_lo = reg_xfer(idx); 2672 2673 if (size < REG_WIDTH) { 2674 wrp_reg_subpart(nfp_prog, dst_lo, src_lo, size, 0); 2675 wrp_immed(nfp_prog, dst_hi, 0); 2676 } else if (size == REG_WIDTH) { 2677 wrp_mov(nfp_prog, dst_lo, src_lo); 2678 wrp_immed(nfp_prog, dst_hi, 0); 2679 } else { 2680 swreg src_hi = reg_xfer(idx + 1); 2681 2682 wrp_mov(nfp_prog, dst_lo, src_lo); 2683 wrp_mov(nfp_prog, dst_hi, src_hi); 2684 } 2685 2686 return 0; 2687 } 2688 2689 static int 2690 mem_ldx_data_from_pktcache(struct nfp_prog *nfp_prog, 2691 struct nfp_insn_meta *meta, unsigned int size) 2692 { 2693 u8 off = meta->insn.off - meta->pkt_cache.range_start; 2694 2695 if (IS_ALIGNED(off, REG_WIDTH)) 2696 return mem_ldx_data_from_pktcache_aligned(nfp_prog, meta, size); 2697 2698 return mem_ldx_data_from_pktcache_unaligned(nfp_prog, meta, size); 2699 } 2700 2701 static int 2702 mem_ldx(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2703 unsigned int size) 2704 { 2705 if (meta->ldst_gather_len) 2706 return nfp_cpp_memcpy(nfp_prog, meta); 2707 2708 if (meta->ptr.type == PTR_TO_CTX) { 2709 if (nfp_prog->type == BPF_PROG_TYPE_XDP) 2710 return mem_ldx_xdp(nfp_prog, meta, size); 2711 else 2712 return mem_ldx_skb(nfp_prog, meta, size); 2713 } 2714 2715 if (meta->ptr.type == PTR_TO_PACKET) { 2716 if (meta->pkt_cache.range_end) { 2717 if (meta->pkt_cache.do_init) 2718 mem_ldx_data_init_pktcache(nfp_prog, meta); 2719 2720 return mem_ldx_data_from_pktcache(nfp_prog, meta, size); 2721 } else { 2722 return mem_ldx_data(nfp_prog, meta, size); 2723 } 2724 } 2725 2726 if (meta->ptr.type == PTR_TO_STACK) 2727 return mem_ldx_stack(nfp_prog, meta, size, 2728 meta->ptr.off + meta->ptr.var_off.value); 2729 2730 if (meta->ptr.type == PTR_TO_MAP_VALUE) 2731 return mem_ldx_emem(nfp_prog, meta, size); 2732 2733 return -EOPNOTSUPP; 2734 } 2735 2736 static int mem_ldx1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2737 { 2738 return mem_ldx(nfp_prog, meta, 1); 2739 } 2740 2741 static int mem_ldx2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2742 { 2743 return mem_ldx(nfp_prog, meta, 2); 2744 } 2745 2746 static int mem_ldx4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2747 { 2748 return mem_ldx(nfp_prog, meta, 4); 2749 } 2750 2751 static int mem_ldx8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2752 { 2753 return mem_ldx(nfp_prog, meta, 8); 2754 } 2755 2756 static int 2757 mem_st_data(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2758 unsigned int size) 2759 { 2760 u64 imm = meta->insn.imm; /* sign extend */ 2761 swreg off_reg; 2762 2763 off_reg = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog)); 2764 2765 return data_st_host_order(nfp_prog, meta->insn.dst_reg * 2, off_reg, 2766 imm, size); 2767 } 2768 2769 static int mem_st(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2770 unsigned int size) 2771 { 2772 if (meta->ptr.type == PTR_TO_PACKET) 2773 return mem_st_data(nfp_prog, meta, size); 2774 2775 return -EOPNOTSUPP; 2776 } 2777 2778 static int mem_st1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2779 { 2780 return mem_st(nfp_prog, meta, 1); 2781 } 2782 2783 static int mem_st2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2784 { 2785 return mem_st(nfp_prog, meta, 2); 2786 } 2787 2788 static int mem_st4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2789 { 2790 return mem_st(nfp_prog, meta, 4); 2791 } 2792 2793 static int mem_st8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2794 { 2795 return mem_st(nfp_prog, meta, 8); 2796 } 2797 2798 static int 2799 mem_stx_data(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2800 unsigned int size) 2801 { 2802 swreg off_reg; 2803 2804 off_reg = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog)); 2805 2806 return data_stx_host_order(nfp_prog, meta->insn.dst_reg * 2, off_reg, 2807 meta->insn.src_reg * 2, size); 2808 } 2809 2810 static int 2811 mem_stx_stack(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2812 unsigned int size, unsigned int ptr_off) 2813 { 2814 return mem_op_stack(nfp_prog, meta, size, ptr_off, 2815 meta->insn.src_reg * 2, meta->insn.dst_reg * 2, 2816 false, wrp_lmem_store); 2817 } 2818 2819 static int mem_stx_xdp(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2820 { 2821 switch (meta->insn.off) { 2822 case offsetof(struct xdp_md, rx_queue_index): 2823 return nfp_queue_select(nfp_prog, meta); 2824 } 2825 2826 WARN_ON_ONCE(1); /* verifier should have rejected bad accesses */ 2827 return -EOPNOTSUPP; 2828 } 2829 2830 static int 2831 mem_stx(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 2832 unsigned int size) 2833 { 2834 if (meta->ptr.type == PTR_TO_PACKET) 2835 return mem_stx_data(nfp_prog, meta, size); 2836 2837 if (meta->ptr.type == PTR_TO_STACK) 2838 return mem_stx_stack(nfp_prog, meta, size, 2839 meta->ptr.off + meta->ptr.var_off.value); 2840 2841 return -EOPNOTSUPP; 2842 } 2843 2844 static int mem_stx1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2845 { 2846 return mem_stx(nfp_prog, meta, 1); 2847 } 2848 2849 static int mem_stx2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2850 { 2851 return mem_stx(nfp_prog, meta, 2); 2852 } 2853 2854 static int mem_stx4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2855 { 2856 if (meta->ptr.type == PTR_TO_CTX) 2857 if (nfp_prog->type == BPF_PROG_TYPE_XDP) 2858 return mem_stx_xdp(nfp_prog, meta); 2859 return mem_stx(nfp_prog, meta, 4); 2860 } 2861 2862 static int mem_stx8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2863 { 2864 return mem_stx(nfp_prog, meta, 8); 2865 } 2866 2867 static int 2868 mem_xadd(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, bool is64) 2869 { 2870 u8 dst_gpr = meta->insn.dst_reg * 2; 2871 u8 src_gpr = meta->insn.src_reg * 2; 2872 unsigned int full_add, out; 2873 swreg addra, addrb, off; 2874 2875 off = ur_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog)); 2876 2877 /* We can fit 16 bits into command immediate, if we know the immediate 2878 * is guaranteed to either always or never fit into 16 bit we only 2879 * generate code to handle that particular case, otherwise generate 2880 * code for both. 2881 */ 2882 out = nfp_prog_current_offset(nfp_prog); 2883 full_add = nfp_prog_current_offset(nfp_prog); 2884 2885 if (meta->insn.off) { 2886 out += 2; 2887 full_add += 2; 2888 } 2889 if (meta->xadd_maybe_16bit) { 2890 out += 3; 2891 full_add += 3; 2892 } 2893 if (meta->xadd_over_16bit) 2894 out += 2 + is64; 2895 if (meta->xadd_maybe_16bit && meta->xadd_over_16bit) { 2896 out += 5; 2897 full_add += 5; 2898 } 2899 2900 /* Generate the branch for choosing add_imm vs add */ 2901 if (meta->xadd_maybe_16bit && meta->xadd_over_16bit) { 2902 swreg max_imm = imm_a(nfp_prog); 2903 2904 wrp_immed(nfp_prog, max_imm, 0xffff); 2905 emit_alu(nfp_prog, reg_none(), 2906 max_imm, ALU_OP_SUB, reg_b(src_gpr)); 2907 emit_alu(nfp_prog, reg_none(), 2908 reg_imm(0), ALU_OP_SUB_C, reg_b(src_gpr + 1)); 2909 emit_br(nfp_prog, BR_BLO, full_add, meta->insn.off ? 2 : 0); 2910 /* defer for add */ 2911 } 2912 2913 /* If insn has an offset add to the address */ 2914 if (!meta->insn.off) { 2915 addra = reg_a(dst_gpr); 2916 addrb = reg_b(dst_gpr + 1); 2917 } else { 2918 emit_alu(nfp_prog, imma_a(nfp_prog), 2919 reg_a(dst_gpr), ALU_OP_ADD, off); 2920 emit_alu(nfp_prog, imma_b(nfp_prog), 2921 reg_a(dst_gpr + 1), ALU_OP_ADD_C, reg_imm(0)); 2922 addra = imma_a(nfp_prog); 2923 addrb = imma_b(nfp_prog); 2924 } 2925 2926 /* Generate the add_imm if 16 bits are possible */ 2927 if (meta->xadd_maybe_16bit) { 2928 swreg prev_alu = imm_a(nfp_prog); 2929 2930 wrp_immed(nfp_prog, prev_alu, 2931 FIELD_PREP(CMD_OVE_DATA, 2) | 2932 CMD_OVE_LEN | 2933 FIELD_PREP(CMD_OV_LEN, 0x8 | is64 << 2)); 2934 wrp_reg_or_subpart(nfp_prog, prev_alu, reg_b(src_gpr), 2, 2); 2935 emit_cmd_indir(nfp_prog, CMD_TGT_ADD_IMM, CMD_MODE_40b_BA, 0, 2936 addra, addrb, 0, CMD_CTX_NO_SWAP); 2937 2938 if (meta->xadd_over_16bit) 2939 emit_br(nfp_prog, BR_UNC, out, 0); 2940 } 2941 2942 if (!nfp_prog_confirm_current_offset(nfp_prog, full_add)) 2943 return -EINVAL; 2944 2945 /* Generate the add if 16 bits are not guaranteed */ 2946 if (meta->xadd_over_16bit) { 2947 emit_cmd(nfp_prog, CMD_TGT_ADD, CMD_MODE_40b_BA, 0, 2948 addra, addrb, is64 << 2, 2949 is64 ? CMD_CTX_SWAP_DEFER2 : CMD_CTX_SWAP_DEFER1); 2950 2951 wrp_mov(nfp_prog, reg_xfer(0), reg_a(src_gpr)); 2952 if (is64) 2953 wrp_mov(nfp_prog, reg_xfer(1), reg_a(src_gpr + 1)); 2954 } 2955 2956 if (!nfp_prog_confirm_current_offset(nfp_prog, out)) 2957 return -EINVAL; 2958 2959 return 0; 2960 } 2961 2962 static int mem_xadd4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2963 { 2964 return mem_xadd(nfp_prog, meta, false); 2965 } 2966 2967 static int mem_xadd8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2968 { 2969 return mem_xadd(nfp_prog, meta, true); 2970 } 2971 2972 static int jump(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2973 { 2974 emit_br(nfp_prog, BR_UNC, meta->insn.off, 0); 2975 2976 return 0; 2977 } 2978 2979 static int jeq_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 2980 { 2981 const struct bpf_insn *insn = &meta->insn; 2982 u64 imm = insn->imm; /* sign extend */ 2983 swreg or1, or2, tmp_reg; 2984 2985 or1 = reg_a(insn->dst_reg * 2); 2986 or2 = reg_b(insn->dst_reg * 2 + 1); 2987 2988 if (imm & ~0U) { 2989 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog)); 2990 emit_alu(nfp_prog, imm_a(nfp_prog), 2991 reg_a(insn->dst_reg * 2), ALU_OP_XOR, tmp_reg); 2992 or1 = imm_a(nfp_prog); 2993 } 2994 2995 if (imm >> 32) { 2996 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog)); 2997 emit_alu(nfp_prog, imm_b(nfp_prog), 2998 reg_a(insn->dst_reg * 2 + 1), ALU_OP_XOR, tmp_reg); 2999 or2 = imm_b(nfp_prog); 3000 } 3001 3002 emit_alu(nfp_prog, reg_none(), or1, ALU_OP_OR, or2); 3003 emit_br(nfp_prog, BR_BEQ, insn->off, 0); 3004 3005 return 0; 3006 } 3007 3008 static int jset_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 3009 { 3010 const struct bpf_insn *insn = &meta->insn; 3011 u64 imm = insn->imm; /* sign extend */ 3012 swreg tmp_reg; 3013 3014 if (!imm) { 3015 meta->skip = true; 3016 return 0; 3017 } 3018 3019 if (imm & ~0U) { 3020 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog)); 3021 emit_alu(nfp_prog, reg_none(), 3022 reg_a(insn->dst_reg * 2), ALU_OP_AND, tmp_reg); 3023 emit_br(nfp_prog, BR_BNE, insn->off, 0); 3024 } 3025 3026 if (imm >> 32) { 3027 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog)); 3028 emit_alu(nfp_prog, reg_none(), 3029 reg_a(insn->dst_reg * 2 + 1), ALU_OP_AND, tmp_reg); 3030 emit_br(nfp_prog, BR_BNE, insn->off, 0); 3031 } 3032 3033 return 0; 3034 } 3035 3036 static int jne_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 3037 { 3038 const struct bpf_insn *insn = &meta->insn; 3039 u64 imm = insn->imm; /* sign extend */ 3040 swreg tmp_reg; 3041 3042 if (!imm) { 3043 emit_alu(nfp_prog, reg_none(), reg_a(insn->dst_reg * 2), 3044 ALU_OP_OR, reg_b(insn->dst_reg * 2 + 1)); 3045 emit_br(nfp_prog, BR_BNE, insn->off, 0); 3046 return 0; 3047 } 3048 3049 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog)); 3050 emit_alu(nfp_prog, reg_none(), 3051 reg_a(insn->dst_reg * 2), ALU_OP_XOR, tmp_reg); 3052 emit_br(nfp_prog, BR_BNE, insn->off, 0); 3053 3054 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog)); 3055 emit_alu(nfp_prog, reg_none(), 3056 reg_a(insn->dst_reg * 2 + 1), ALU_OP_XOR, tmp_reg); 3057 emit_br(nfp_prog, BR_BNE, insn->off, 0); 3058 3059 return 0; 3060 } 3061 3062 static int jeq_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 3063 { 3064 const struct bpf_insn *insn = &meta->insn; 3065 3066 emit_alu(nfp_prog, imm_a(nfp_prog), reg_a(insn->dst_reg * 2), 3067 ALU_OP_XOR, reg_b(insn->src_reg * 2)); 3068 emit_alu(nfp_prog, imm_b(nfp_prog), reg_a(insn->dst_reg * 2 + 1), 3069 ALU_OP_XOR, reg_b(insn->src_reg * 2 + 1)); 3070 emit_alu(nfp_prog, reg_none(), 3071 imm_a(nfp_prog), ALU_OP_OR, imm_b(nfp_prog)); 3072 emit_br(nfp_prog, BR_BEQ, insn->off, 0); 3073 3074 return 0; 3075 } 3076 3077 static int jset_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 3078 { 3079 return wrp_test_reg(nfp_prog, meta, ALU_OP_AND, BR_BNE); 3080 } 3081 3082 static int jne_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 3083 { 3084 return wrp_test_reg(nfp_prog, meta, ALU_OP_XOR, BR_BNE); 3085 } 3086 3087 static int 3088 bpf_to_bpf_call(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 3089 { 3090 u32 ret_tgt, stack_depth, offset_br; 3091 swreg tmp_reg; 3092 3093 stack_depth = round_up(nfp_prog->stack_frame_depth, STACK_FRAME_ALIGN); 3094 /* Space for saving the return address is accounted for by the callee, 3095 * so stack_depth can be zero for the main function. 3096 */ 3097 if (stack_depth) { 3098 tmp_reg = ur_load_imm_any(nfp_prog, stack_depth, 3099 stack_imm(nfp_prog)); 3100 emit_alu(nfp_prog, stack_reg(nfp_prog), 3101 stack_reg(nfp_prog), ALU_OP_ADD, tmp_reg); 3102 emit_csr_wr(nfp_prog, stack_reg(nfp_prog), 3103 NFP_CSR_ACT_LM_ADDR0); 3104 } 3105 3106 /* Two cases for jumping to the callee: 3107 * 3108 * - If callee uses and needs to save R6~R9 then: 3109 * 1. Put the start offset of the callee into imm_b(). This will 3110 * require a fixup step, as we do not necessarily know this 3111 * address yet. 3112 * 2. Put the return address from the callee to the caller into 3113 * register ret_reg(). 3114 * 3. (After defer slots are consumed) Jump to the subroutine that 3115 * pushes the registers to the stack. 3116 * The subroutine acts as a trampoline, and returns to the address in 3117 * imm_b(), i.e. jumps to the callee. 3118 * 3119 * - If callee does not need to save R6~R9 then just load return 3120 * address to the caller in ret_reg(), and jump to the callee 3121 * directly. 3122 * 3123 * Using ret_reg() to pass the return address to the callee is set here 3124 * as a convention. The callee can then push this address onto its 3125 * stack frame in its prologue. The advantages of passing the return 3126 * address through ret_reg(), instead of pushing it to the stack right 3127 * here, are the following: 3128 * - It looks cleaner. 3129 * - If the called function is called multiple time, we get a lower 3130 * program size. 3131 * - We save two no-op instructions that should be added just before 3132 * the emit_br() when stack depth is not null otherwise. 3133 * - If we ever find a register to hold the return address during whole 3134 * execution of the callee, we will not have to push the return 3135 * address to the stack for leaf functions. 3136 */ 3137 if (!meta->jmp_dst) { 3138 pr_err("BUG: BPF-to-BPF call has no destination recorded\n"); 3139 return -ELOOP; 3140 } 3141 if (nfp_prog->subprog[meta->jmp_dst->subprog_idx].needs_reg_push) { 3142 ret_tgt = nfp_prog_current_offset(nfp_prog) + 3; 3143 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 2, 3144 RELO_BR_GO_CALL_PUSH_REGS); 3145 offset_br = nfp_prog_current_offset(nfp_prog); 3146 wrp_immed_relo(nfp_prog, imm_b(nfp_prog), 0, RELO_IMMED_REL); 3147 } else { 3148 ret_tgt = nfp_prog_current_offset(nfp_prog) + 2; 3149 emit_br(nfp_prog, BR_UNC, meta->n + 1 + meta->insn.imm, 1); 3150 offset_br = nfp_prog_current_offset(nfp_prog); 3151 } 3152 wrp_immed_relo(nfp_prog, ret_reg(nfp_prog), ret_tgt, RELO_IMMED_REL); 3153 3154 if (!nfp_prog_confirm_current_offset(nfp_prog, ret_tgt)) 3155 return -EINVAL; 3156 3157 if (stack_depth) { 3158 tmp_reg = ur_load_imm_any(nfp_prog, stack_depth, 3159 stack_imm(nfp_prog)); 3160 emit_alu(nfp_prog, stack_reg(nfp_prog), 3161 stack_reg(nfp_prog), ALU_OP_SUB, tmp_reg); 3162 emit_csr_wr(nfp_prog, stack_reg(nfp_prog), 3163 NFP_CSR_ACT_LM_ADDR0); 3164 wrp_nops(nfp_prog, 3); 3165 } 3166 3167 meta->num_insns_after_br = nfp_prog_current_offset(nfp_prog); 3168 meta->num_insns_after_br -= offset_br; 3169 3170 return 0; 3171 } 3172 3173 static int helper_call(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 3174 { 3175 switch (meta->insn.imm) { 3176 case BPF_FUNC_xdp_adjust_head: 3177 return adjust_head(nfp_prog, meta); 3178 case BPF_FUNC_xdp_adjust_tail: 3179 return adjust_tail(nfp_prog, meta); 3180 case BPF_FUNC_map_lookup_elem: 3181 case BPF_FUNC_map_update_elem: 3182 case BPF_FUNC_map_delete_elem: 3183 return map_call_stack_common(nfp_prog, meta); 3184 case BPF_FUNC_get_prandom_u32: 3185 return nfp_get_prandom_u32(nfp_prog, meta); 3186 case BPF_FUNC_perf_event_output: 3187 return nfp_perf_event_output(nfp_prog, meta); 3188 default: 3189 WARN_ONCE(1, "verifier allowed unsupported function\n"); 3190 return -EOPNOTSUPP; 3191 } 3192 } 3193 3194 static int call(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 3195 { 3196 if (is_mbpf_pseudo_call(meta)) 3197 return bpf_to_bpf_call(nfp_prog, meta); 3198 else 3199 return helper_call(nfp_prog, meta); 3200 } 3201 3202 static bool nfp_is_main_function(struct nfp_insn_meta *meta) 3203 { 3204 return meta->subprog_idx == 0; 3205 } 3206 3207 static int goto_out(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 3208 { 3209 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 0, RELO_BR_GO_OUT); 3210 3211 return 0; 3212 } 3213 3214 static int 3215 nfp_subprog_epilogue(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 3216 { 3217 if (nfp_prog->subprog[meta->subprog_idx].needs_reg_push) { 3218 /* Pop R6~R9 to the stack via related subroutine. 3219 * We loaded the return address to the caller into ret_reg(). 3220 * This means that the subroutine does not come back here, we 3221 * make it jump back to the subprogram caller directly! 3222 */ 3223 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 1, 3224 RELO_BR_GO_CALL_POP_REGS); 3225 /* Pop return address from the stack. */ 3226 wrp_mov(nfp_prog, ret_reg(nfp_prog), reg_lm(0, 0)); 3227 } else { 3228 /* Pop return address from the stack. */ 3229 wrp_mov(nfp_prog, ret_reg(nfp_prog), reg_lm(0, 0)); 3230 /* Jump back to caller if no callee-saved registers were used 3231 * by the subprogram. 3232 */ 3233 emit_rtn(nfp_prog, ret_reg(nfp_prog), 0); 3234 } 3235 3236 return 0; 3237 } 3238 3239 static int jmp_exit(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 3240 { 3241 if (nfp_is_main_function(meta)) 3242 return goto_out(nfp_prog, meta); 3243 else 3244 return nfp_subprog_epilogue(nfp_prog, meta); 3245 } 3246 3247 static const instr_cb_t instr_cb[256] = { 3248 [BPF_ALU64 | BPF_MOV | BPF_X] = mov_reg64, 3249 [BPF_ALU64 | BPF_MOV | BPF_K] = mov_imm64, 3250 [BPF_ALU64 | BPF_XOR | BPF_X] = xor_reg64, 3251 [BPF_ALU64 | BPF_XOR | BPF_K] = xor_imm64, 3252 [BPF_ALU64 | BPF_AND | BPF_X] = and_reg64, 3253 [BPF_ALU64 | BPF_AND | BPF_K] = and_imm64, 3254 [BPF_ALU64 | BPF_OR | BPF_X] = or_reg64, 3255 [BPF_ALU64 | BPF_OR | BPF_K] = or_imm64, 3256 [BPF_ALU64 | BPF_ADD | BPF_X] = add_reg64, 3257 [BPF_ALU64 | BPF_ADD | BPF_K] = add_imm64, 3258 [BPF_ALU64 | BPF_SUB | BPF_X] = sub_reg64, 3259 [BPF_ALU64 | BPF_SUB | BPF_K] = sub_imm64, 3260 [BPF_ALU64 | BPF_MUL | BPF_X] = mul_reg64, 3261 [BPF_ALU64 | BPF_MUL | BPF_K] = mul_imm64, 3262 [BPF_ALU64 | BPF_DIV | BPF_X] = div_reg64, 3263 [BPF_ALU64 | BPF_DIV | BPF_K] = div_imm64, 3264 [BPF_ALU64 | BPF_NEG] = neg_reg64, 3265 [BPF_ALU64 | BPF_LSH | BPF_X] = shl_reg64, 3266 [BPF_ALU64 | BPF_LSH | BPF_K] = shl_imm64, 3267 [BPF_ALU64 | BPF_RSH | BPF_X] = shr_reg64, 3268 [BPF_ALU64 | BPF_RSH | BPF_K] = shr_imm64, 3269 [BPF_ALU64 | BPF_ARSH | BPF_X] = ashr_reg64, 3270 [BPF_ALU64 | BPF_ARSH | BPF_K] = ashr_imm64, 3271 [BPF_ALU | BPF_MOV | BPF_X] = mov_reg, 3272 [BPF_ALU | BPF_MOV | BPF_K] = mov_imm, 3273 [BPF_ALU | BPF_XOR | BPF_X] = xor_reg, 3274 [BPF_ALU | BPF_XOR | BPF_K] = xor_imm, 3275 [BPF_ALU | BPF_AND | BPF_X] = and_reg, 3276 [BPF_ALU | BPF_AND | BPF_K] = and_imm, 3277 [BPF_ALU | BPF_OR | BPF_X] = or_reg, 3278 [BPF_ALU | BPF_OR | BPF_K] = or_imm, 3279 [BPF_ALU | BPF_ADD | BPF_X] = add_reg, 3280 [BPF_ALU | BPF_ADD | BPF_K] = add_imm, 3281 [BPF_ALU | BPF_SUB | BPF_X] = sub_reg, 3282 [BPF_ALU | BPF_SUB | BPF_K] = sub_imm, 3283 [BPF_ALU | BPF_MUL | BPF_X] = mul_reg, 3284 [BPF_ALU | BPF_MUL | BPF_K] = mul_imm, 3285 [BPF_ALU | BPF_DIV | BPF_X] = div_reg, 3286 [BPF_ALU | BPF_DIV | BPF_K] = div_imm, 3287 [BPF_ALU | BPF_NEG] = neg_reg, 3288 [BPF_ALU | BPF_LSH | BPF_K] = shl_imm, 3289 [BPF_ALU | BPF_END | BPF_X] = end_reg32, 3290 [BPF_LD | BPF_IMM | BPF_DW] = imm_ld8, 3291 [BPF_LD | BPF_ABS | BPF_B] = data_ld1, 3292 [BPF_LD | BPF_ABS | BPF_H] = data_ld2, 3293 [BPF_LD | BPF_ABS | BPF_W] = data_ld4, 3294 [BPF_LD | BPF_IND | BPF_B] = data_ind_ld1, 3295 [BPF_LD | BPF_IND | BPF_H] = data_ind_ld2, 3296 [BPF_LD | BPF_IND | BPF_W] = data_ind_ld4, 3297 [BPF_LDX | BPF_MEM | BPF_B] = mem_ldx1, 3298 [BPF_LDX | BPF_MEM | BPF_H] = mem_ldx2, 3299 [BPF_LDX | BPF_MEM | BPF_W] = mem_ldx4, 3300 [BPF_LDX | BPF_MEM | BPF_DW] = mem_ldx8, 3301 [BPF_STX | BPF_MEM | BPF_B] = mem_stx1, 3302 [BPF_STX | BPF_MEM | BPF_H] = mem_stx2, 3303 [BPF_STX | BPF_MEM | BPF_W] = mem_stx4, 3304 [BPF_STX | BPF_MEM | BPF_DW] = mem_stx8, 3305 [BPF_STX | BPF_XADD | BPF_W] = mem_xadd4, 3306 [BPF_STX | BPF_XADD | BPF_DW] = mem_xadd8, 3307 [BPF_ST | BPF_MEM | BPF_B] = mem_st1, 3308 [BPF_ST | BPF_MEM | BPF_H] = mem_st2, 3309 [BPF_ST | BPF_MEM | BPF_W] = mem_st4, 3310 [BPF_ST | BPF_MEM | BPF_DW] = mem_st8, 3311 [BPF_JMP | BPF_JA | BPF_K] = jump, 3312 [BPF_JMP | BPF_JEQ | BPF_K] = jeq_imm, 3313 [BPF_JMP | BPF_JGT | BPF_K] = cmp_imm, 3314 [BPF_JMP | BPF_JGE | BPF_K] = cmp_imm, 3315 [BPF_JMP | BPF_JLT | BPF_K] = cmp_imm, 3316 [BPF_JMP | BPF_JLE | BPF_K] = cmp_imm, 3317 [BPF_JMP | BPF_JSGT | BPF_K] = cmp_imm, 3318 [BPF_JMP | BPF_JSGE | BPF_K] = cmp_imm, 3319 [BPF_JMP | BPF_JSLT | BPF_K] = cmp_imm, 3320 [BPF_JMP | BPF_JSLE | BPF_K] = cmp_imm, 3321 [BPF_JMP | BPF_JSET | BPF_K] = jset_imm, 3322 [BPF_JMP | BPF_JNE | BPF_K] = jne_imm, 3323 [BPF_JMP | BPF_JEQ | BPF_X] = jeq_reg, 3324 [BPF_JMP | BPF_JGT | BPF_X] = cmp_reg, 3325 [BPF_JMP | BPF_JGE | BPF_X] = cmp_reg, 3326 [BPF_JMP | BPF_JLT | BPF_X] = cmp_reg, 3327 [BPF_JMP | BPF_JLE | BPF_X] = cmp_reg, 3328 [BPF_JMP | BPF_JSGT | BPF_X] = cmp_reg, 3329 [BPF_JMP | BPF_JSGE | BPF_X] = cmp_reg, 3330 [BPF_JMP | BPF_JSLT | BPF_X] = cmp_reg, 3331 [BPF_JMP | BPF_JSLE | BPF_X] = cmp_reg, 3332 [BPF_JMP | BPF_JSET | BPF_X] = jset_reg, 3333 [BPF_JMP | BPF_JNE | BPF_X] = jne_reg, 3334 [BPF_JMP | BPF_CALL] = call, 3335 [BPF_JMP | BPF_EXIT] = jmp_exit, 3336 }; 3337 3338 /* --- Assembler logic --- */ 3339 static int 3340 nfp_fixup_immed_relo(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, 3341 struct nfp_insn_meta *jmp_dst, u32 br_idx) 3342 { 3343 if (immed_get_value(nfp_prog->prog[br_idx + 1])) { 3344 pr_err("BUG: failed to fix up callee register saving\n"); 3345 return -EINVAL; 3346 } 3347 3348 immed_set_value(&nfp_prog->prog[br_idx + 1], jmp_dst->off); 3349 3350 return 0; 3351 } 3352 3353 static int nfp_fixup_branches(struct nfp_prog *nfp_prog) 3354 { 3355 struct nfp_insn_meta *meta, *jmp_dst; 3356 u32 idx, br_idx; 3357 int err; 3358 3359 list_for_each_entry(meta, &nfp_prog->insns, l) { 3360 if (meta->skip) 3361 continue; 3362 if (BPF_CLASS(meta->insn.code) != BPF_JMP) 3363 continue; 3364 if (meta->insn.code == (BPF_JMP | BPF_EXIT) && 3365 !nfp_is_main_function(meta)) 3366 continue; 3367 if (is_mbpf_helper_call(meta)) 3368 continue; 3369 3370 if (list_is_last(&meta->l, &nfp_prog->insns)) 3371 br_idx = nfp_prog->last_bpf_off; 3372 else 3373 br_idx = list_next_entry(meta, l)->off - 1; 3374 3375 /* For BPF-to-BPF function call, a stack adjustment sequence is 3376 * generated after the return instruction. Therefore, we must 3377 * withdraw the length of this sequence to have br_idx pointing 3378 * to where the "branch" NFP instruction is expected to be. 3379 */ 3380 if (is_mbpf_pseudo_call(meta)) 3381 br_idx -= meta->num_insns_after_br; 3382 3383 if (!nfp_is_br(nfp_prog->prog[br_idx])) { 3384 pr_err("Fixup found block not ending in branch %d %02x %016llx!!\n", 3385 br_idx, meta->insn.code, nfp_prog->prog[br_idx]); 3386 return -ELOOP; 3387 } 3388 3389 if (meta->insn.code == (BPF_JMP | BPF_EXIT)) 3390 continue; 3391 3392 /* Leave special branches for later */ 3393 if (FIELD_GET(OP_RELO_TYPE, nfp_prog->prog[br_idx]) != 3394 RELO_BR_REL && !is_mbpf_pseudo_call(meta)) 3395 continue; 3396 3397 if (!meta->jmp_dst) { 3398 pr_err("Non-exit jump doesn't have destination info recorded!!\n"); 3399 return -ELOOP; 3400 } 3401 3402 jmp_dst = meta->jmp_dst; 3403 3404 if (jmp_dst->skip) { 3405 pr_err("Branch landing on removed instruction!!\n"); 3406 return -ELOOP; 3407 } 3408 3409 if (is_mbpf_pseudo_call(meta) && 3410 nfp_prog->subprog[jmp_dst->subprog_idx].needs_reg_push) { 3411 err = nfp_fixup_immed_relo(nfp_prog, meta, 3412 jmp_dst, br_idx); 3413 if (err) 3414 return err; 3415 } 3416 3417 if (FIELD_GET(OP_RELO_TYPE, nfp_prog->prog[br_idx]) != 3418 RELO_BR_REL) 3419 continue; 3420 3421 for (idx = meta->off; idx <= br_idx; idx++) { 3422 if (!nfp_is_br(nfp_prog->prog[idx])) 3423 continue; 3424 br_set_offset(&nfp_prog->prog[idx], jmp_dst->off); 3425 } 3426 } 3427 3428 return 0; 3429 } 3430 3431 static void nfp_intro(struct nfp_prog *nfp_prog) 3432 { 3433 wrp_immed(nfp_prog, plen_reg(nfp_prog), GENMASK(13, 0)); 3434 emit_alu(nfp_prog, plen_reg(nfp_prog), 3435 plen_reg(nfp_prog), ALU_OP_AND, pv_len(nfp_prog)); 3436 } 3437 3438 static void 3439 nfp_subprog_prologue(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 3440 { 3441 /* Save return address into the stack. */ 3442 wrp_mov(nfp_prog, reg_lm(0, 0), ret_reg(nfp_prog)); 3443 } 3444 3445 static void 3446 nfp_start_subprog(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta) 3447 { 3448 unsigned int depth = nfp_prog->subprog[meta->subprog_idx].stack_depth; 3449 3450 nfp_prog->stack_frame_depth = round_up(depth, 4); 3451 nfp_subprog_prologue(nfp_prog, meta); 3452 } 3453 3454 bool nfp_is_subprog_start(struct nfp_insn_meta *meta) 3455 { 3456 return meta->flags & FLAG_INSN_IS_SUBPROG_START; 3457 } 3458 3459 static void nfp_outro_tc_da(struct nfp_prog *nfp_prog) 3460 { 3461 /* TC direct-action mode: 3462 * 0,1 ok NOT SUPPORTED[1] 3463 * 2 drop 0x22 -> drop, count as stat1 3464 * 4,5 nuke 0x02 -> drop 3465 * 7 redir 0x44 -> redir, count as stat2 3466 * * unspec 0x11 -> pass, count as stat0 3467 * 3468 * [1] We can't support OK and RECLASSIFY because we can't tell TC 3469 * the exact decision made. We are forced to support UNSPEC 3470 * to handle aborts so that's the only one we handle for passing 3471 * packets up the stack. 3472 */ 3473 /* Target for aborts */ 3474 nfp_prog->tgt_abort = nfp_prog_current_offset(nfp_prog); 3475 3476 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 2, RELO_BR_NEXT_PKT); 3477 3478 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS); 3479 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_imm(0x11), SHF_SC_L_SHF, 16); 3480 3481 /* Target for normal exits */ 3482 nfp_prog->tgt_out = nfp_prog_current_offset(nfp_prog); 3483 3484 /* if R0 > 7 jump to abort */ 3485 emit_alu(nfp_prog, reg_none(), reg_imm(7), ALU_OP_SUB, reg_b(0)); 3486 emit_br(nfp_prog, BR_BLO, nfp_prog->tgt_abort, 0); 3487 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS); 3488 3489 wrp_immed(nfp_prog, reg_b(2), 0x41221211); 3490 wrp_immed(nfp_prog, reg_b(3), 0x41001211); 3491 3492 emit_shf(nfp_prog, reg_a(1), 3493 reg_none(), SHF_OP_NONE, reg_b(0), SHF_SC_L_SHF, 2); 3494 3495 emit_alu(nfp_prog, reg_none(), reg_a(1), ALU_OP_OR, reg_imm(0)); 3496 emit_shf(nfp_prog, reg_a(2), 3497 reg_imm(0xf), SHF_OP_AND, reg_b(2), SHF_SC_R_SHF, 0); 3498 3499 emit_alu(nfp_prog, reg_none(), reg_a(1), ALU_OP_OR, reg_imm(0)); 3500 emit_shf(nfp_prog, reg_b(2), 3501 reg_imm(0xf), SHF_OP_AND, reg_b(3), SHF_SC_R_SHF, 0); 3502 3503 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 2, RELO_BR_NEXT_PKT); 3504 3505 emit_shf(nfp_prog, reg_b(2), 3506 reg_a(2), SHF_OP_OR, reg_b(2), SHF_SC_L_SHF, 4); 3507 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_b(2), SHF_SC_L_SHF, 16); 3508 } 3509 3510 static void nfp_outro_xdp(struct nfp_prog *nfp_prog) 3511 { 3512 /* XDP return codes: 3513 * 0 aborted 0x82 -> drop, count as stat3 3514 * 1 drop 0x22 -> drop, count as stat1 3515 * 2 pass 0x11 -> pass, count as stat0 3516 * 3 tx 0x44 -> redir, count as stat2 3517 * * unknown 0x82 -> drop, count as stat3 3518 */ 3519 /* Target for aborts */ 3520 nfp_prog->tgt_abort = nfp_prog_current_offset(nfp_prog); 3521 3522 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 2, RELO_BR_NEXT_PKT); 3523 3524 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS); 3525 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_imm(0x82), SHF_SC_L_SHF, 16); 3526 3527 /* Target for normal exits */ 3528 nfp_prog->tgt_out = nfp_prog_current_offset(nfp_prog); 3529 3530 /* if R0 > 3 jump to abort */ 3531 emit_alu(nfp_prog, reg_none(), reg_imm(3), ALU_OP_SUB, reg_b(0)); 3532 emit_br(nfp_prog, BR_BLO, nfp_prog->tgt_abort, 0); 3533 3534 wrp_immed(nfp_prog, reg_b(2), 0x44112282); 3535 3536 emit_shf(nfp_prog, reg_a(1), 3537 reg_none(), SHF_OP_NONE, reg_b(0), SHF_SC_L_SHF, 3); 3538 3539 emit_alu(nfp_prog, reg_none(), reg_a(1), ALU_OP_OR, reg_imm(0)); 3540 emit_shf(nfp_prog, reg_b(2), 3541 reg_imm(0xff), SHF_OP_AND, reg_b(2), SHF_SC_R_SHF, 0); 3542 3543 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 2, RELO_BR_NEXT_PKT); 3544 3545 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS); 3546 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_b(2), SHF_SC_L_SHF, 16); 3547 } 3548 3549 static bool nfp_prog_needs_callee_reg_save(struct nfp_prog *nfp_prog) 3550 { 3551 unsigned int idx; 3552 3553 for (idx = 1; idx < nfp_prog->subprog_cnt; idx++) 3554 if (nfp_prog->subprog[idx].needs_reg_push) 3555 return true; 3556 3557 return false; 3558 } 3559 3560 static void nfp_push_callee_registers(struct nfp_prog *nfp_prog) 3561 { 3562 u8 reg; 3563 3564 /* Subroutine: Save all callee saved registers (R6 ~ R9). 3565 * imm_b() holds the return address. 3566 */ 3567 nfp_prog->tgt_call_push_regs = nfp_prog_current_offset(nfp_prog); 3568 for (reg = BPF_REG_6; reg <= BPF_REG_9; reg++) { 3569 u8 adj = (reg - BPF_REG_0) * 2; 3570 u8 idx = (reg - BPF_REG_6) * 2; 3571 3572 /* The first slot in the stack frame is used to push the return 3573 * address in bpf_to_bpf_call(), start just after. 3574 */ 3575 wrp_mov(nfp_prog, reg_lm(0, 1 + idx), reg_b(adj)); 3576 3577 if (reg == BPF_REG_8) 3578 /* Prepare to jump back, last 3 insns use defer slots */ 3579 emit_rtn(nfp_prog, imm_b(nfp_prog), 3); 3580 3581 wrp_mov(nfp_prog, reg_lm(0, 1 + idx + 1), reg_b(adj + 1)); 3582 } 3583 } 3584 3585 static void nfp_pop_callee_registers(struct nfp_prog *nfp_prog) 3586 { 3587 u8 reg; 3588 3589 /* Subroutine: Restore all callee saved registers (R6 ~ R9). 3590 * ret_reg() holds the return address. 3591 */ 3592 nfp_prog->tgt_call_pop_regs = nfp_prog_current_offset(nfp_prog); 3593 for (reg = BPF_REG_6; reg <= BPF_REG_9; reg++) { 3594 u8 adj = (reg - BPF_REG_0) * 2; 3595 u8 idx = (reg - BPF_REG_6) * 2; 3596 3597 /* The first slot in the stack frame holds the return address, 3598 * start popping just after that. 3599 */ 3600 wrp_mov(nfp_prog, reg_both(adj), reg_lm(0, 1 + idx)); 3601 3602 if (reg == BPF_REG_8) 3603 /* Prepare to jump back, last 3 insns use defer slots */ 3604 emit_rtn(nfp_prog, ret_reg(nfp_prog), 3); 3605 3606 wrp_mov(nfp_prog, reg_both(adj + 1), reg_lm(0, 1 + idx + 1)); 3607 } 3608 } 3609 3610 static void nfp_outro(struct nfp_prog *nfp_prog) 3611 { 3612 switch (nfp_prog->type) { 3613 case BPF_PROG_TYPE_SCHED_CLS: 3614 nfp_outro_tc_da(nfp_prog); 3615 break; 3616 case BPF_PROG_TYPE_XDP: 3617 nfp_outro_xdp(nfp_prog); 3618 break; 3619 default: 3620 WARN_ON(1); 3621 } 3622 3623 if (!nfp_prog_needs_callee_reg_save(nfp_prog)) 3624 return; 3625 3626 nfp_push_callee_registers(nfp_prog); 3627 nfp_pop_callee_registers(nfp_prog); 3628 } 3629 3630 static int nfp_translate(struct nfp_prog *nfp_prog) 3631 { 3632 struct nfp_insn_meta *meta; 3633 unsigned int depth; 3634 int err; 3635 3636 depth = nfp_prog->subprog[0].stack_depth; 3637 nfp_prog->stack_frame_depth = round_up(depth, 4); 3638 3639 nfp_intro(nfp_prog); 3640 if (nfp_prog->error) 3641 return nfp_prog->error; 3642 3643 list_for_each_entry(meta, &nfp_prog->insns, l) { 3644 instr_cb_t cb = instr_cb[meta->insn.code]; 3645 3646 meta->off = nfp_prog_current_offset(nfp_prog); 3647 3648 if (nfp_is_subprog_start(meta)) { 3649 nfp_start_subprog(nfp_prog, meta); 3650 if (nfp_prog->error) 3651 return nfp_prog->error; 3652 } 3653 3654 if (meta->skip) { 3655 nfp_prog->n_translated++; 3656 continue; 3657 } 3658 3659 if (nfp_meta_has_prev(nfp_prog, meta) && 3660 nfp_meta_prev(meta)->double_cb) 3661 cb = nfp_meta_prev(meta)->double_cb; 3662 if (!cb) 3663 return -ENOENT; 3664 err = cb(nfp_prog, meta); 3665 if (err) 3666 return err; 3667 if (nfp_prog->error) 3668 return nfp_prog->error; 3669 3670 nfp_prog->n_translated++; 3671 } 3672 3673 nfp_prog->last_bpf_off = nfp_prog_current_offset(nfp_prog) - 1; 3674 3675 nfp_outro(nfp_prog); 3676 if (nfp_prog->error) 3677 return nfp_prog->error; 3678 3679 wrp_nops(nfp_prog, NFP_USTORE_PREFETCH_WINDOW); 3680 if (nfp_prog->error) 3681 return nfp_prog->error; 3682 3683 return nfp_fixup_branches(nfp_prog); 3684 } 3685 3686 /* --- Optimizations --- */ 3687 static void nfp_bpf_opt_reg_init(struct nfp_prog *nfp_prog) 3688 { 3689 struct nfp_insn_meta *meta; 3690 3691 list_for_each_entry(meta, &nfp_prog->insns, l) { 3692 struct bpf_insn insn = meta->insn; 3693 3694 /* Programs converted from cBPF start with register xoring */ 3695 if (insn.code == (BPF_ALU64 | BPF_XOR | BPF_X) && 3696 insn.src_reg == insn.dst_reg) 3697 continue; 3698 3699 /* Programs start with R6 = R1 but we ignore the skb pointer */ 3700 if (insn.code == (BPF_ALU64 | BPF_MOV | BPF_X) && 3701 insn.src_reg == 1 && insn.dst_reg == 6) 3702 meta->skip = true; 3703 3704 /* Return as soon as something doesn't match */ 3705 if (!meta->skip) 3706 return; 3707 } 3708 } 3709 3710 /* abs(insn.imm) will fit better into unrestricted reg immediate - 3711 * convert add/sub of a negative number into a sub/add of a positive one. 3712 */ 3713 static void nfp_bpf_opt_neg_add_sub(struct nfp_prog *nfp_prog) 3714 { 3715 struct nfp_insn_meta *meta; 3716 3717 list_for_each_entry(meta, &nfp_prog->insns, l) { 3718 struct bpf_insn insn = meta->insn; 3719 3720 if (meta->skip) 3721 continue; 3722 3723 if (BPF_CLASS(insn.code) != BPF_ALU && 3724 BPF_CLASS(insn.code) != BPF_ALU64 && 3725 BPF_CLASS(insn.code) != BPF_JMP) 3726 continue; 3727 if (BPF_SRC(insn.code) != BPF_K) 3728 continue; 3729 if (insn.imm >= 0) 3730 continue; 3731 3732 if (BPF_CLASS(insn.code) == BPF_JMP) { 3733 switch (BPF_OP(insn.code)) { 3734 case BPF_JGE: 3735 case BPF_JSGE: 3736 case BPF_JLT: 3737 case BPF_JSLT: 3738 meta->jump_neg_op = true; 3739 break; 3740 default: 3741 continue; 3742 } 3743 } else { 3744 if (BPF_OP(insn.code) == BPF_ADD) 3745 insn.code = BPF_CLASS(insn.code) | BPF_SUB; 3746 else if (BPF_OP(insn.code) == BPF_SUB) 3747 insn.code = BPF_CLASS(insn.code) | BPF_ADD; 3748 else 3749 continue; 3750 3751 meta->insn.code = insn.code | BPF_K; 3752 } 3753 3754 meta->insn.imm = -insn.imm; 3755 } 3756 } 3757 3758 /* Remove masking after load since our load guarantees this is not needed */ 3759 static void nfp_bpf_opt_ld_mask(struct nfp_prog *nfp_prog) 3760 { 3761 struct nfp_insn_meta *meta1, *meta2; 3762 const s32 exp_mask[] = { 3763 [BPF_B] = 0x000000ffU, 3764 [BPF_H] = 0x0000ffffU, 3765 [BPF_W] = 0xffffffffU, 3766 }; 3767 3768 nfp_for_each_insn_walk2(nfp_prog, meta1, meta2) { 3769 struct bpf_insn insn, next; 3770 3771 insn = meta1->insn; 3772 next = meta2->insn; 3773 3774 if (BPF_CLASS(insn.code) != BPF_LD) 3775 continue; 3776 if (BPF_MODE(insn.code) != BPF_ABS && 3777 BPF_MODE(insn.code) != BPF_IND) 3778 continue; 3779 3780 if (next.code != (BPF_ALU64 | BPF_AND | BPF_K)) 3781 continue; 3782 3783 if (!exp_mask[BPF_SIZE(insn.code)]) 3784 continue; 3785 if (exp_mask[BPF_SIZE(insn.code)] != next.imm) 3786 continue; 3787 3788 if (next.src_reg || next.dst_reg) 3789 continue; 3790 3791 if (meta2->flags & FLAG_INSN_IS_JUMP_DST) 3792 continue; 3793 3794 meta2->skip = true; 3795 } 3796 } 3797 3798 static void nfp_bpf_opt_ld_shift(struct nfp_prog *nfp_prog) 3799 { 3800 struct nfp_insn_meta *meta1, *meta2, *meta3; 3801 3802 nfp_for_each_insn_walk3(nfp_prog, meta1, meta2, meta3) { 3803 struct bpf_insn insn, next1, next2; 3804 3805 insn = meta1->insn; 3806 next1 = meta2->insn; 3807 next2 = meta3->insn; 3808 3809 if (BPF_CLASS(insn.code) != BPF_LD) 3810 continue; 3811 if (BPF_MODE(insn.code) != BPF_ABS && 3812 BPF_MODE(insn.code) != BPF_IND) 3813 continue; 3814 if (BPF_SIZE(insn.code) != BPF_W) 3815 continue; 3816 3817 if (!(next1.code == (BPF_LSH | BPF_K | BPF_ALU64) && 3818 next2.code == (BPF_RSH | BPF_K | BPF_ALU64)) && 3819 !(next1.code == (BPF_RSH | BPF_K | BPF_ALU64) && 3820 next2.code == (BPF_LSH | BPF_K | BPF_ALU64))) 3821 continue; 3822 3823 if (next1.src_reg || next1.dst_reg || 3824 next2.src_reg || next2.dst_reg) 3825 continue; 3826 3827 if (next1.imm != 0x20 || next2.imm != 0x20) 3828 continue; 3829 3830 if (meta2->flags & FLAG_INSN_IS_JUMP_DST || 3831 meta3->flags & FLAG_INSN_IS_JUMP_DST) 3832 continue; 3833 3834 meta2->skip = true; 3835 meta3->skip = true; 3836 } 3837 } 3838 3839 /* load/store pair that forms memory copy sould look like the following: 3840 * 3841 * ld_width R, [addr_src + offset_src] 3842 * st_width [addr_dest + offset_dest], R 3843 * 3844 * The destination register of load and source register of store should 3845 * be the same, load and store should also perform at the same width. 3846 * If either of addr_src or addr_dest is stack pointer, we don't do the 3847 * CPP optimization as stack is modelled by registers on NFP. 3848 */ 3849 static bool 3850 curr_pair_is_memcpy(struct nfp_insn_meta *ld_meta, 3851 struct nfp_insn_meta *st_meta) 3852 { 3853 struct bpf_insn *ld = &ld_meta->insn; 3854 struct bpf_insn *st = &st_meta->insn; 3855 3856 if (!is_mbpf_load(ld_meta) || !is_mbpf_store(st_meta)) 3857 return false; 3858 3859 if (ld_meta->ptr.type != PTR_TO_PACKET && 3860 ld_meta->ptr.type != PTR_TO_MAP_VALUE) 3861 return false; 3862 3863 if (st_meta->ptr.type != PTR_TO_PACKET) 3864 return false; 3865 3866 if (BPF_SIZE(ld->code) != BPF_SIZE(st->code)) 3867 return false; 3868 3869 if (ld->dst_reg != st->src_reg) 3870 return false; 3871 3872 /* There is jump to the store insn in this pair. */ 3873 if (st_meta->flags & FLAG_INSN_IS_JUMP_DST) 3874 return false; 3875 3876 return true; 3877 } 3878 3879 /* Currently, we only support chaining load/store pairs if: 3880 * 3881 * - Their address base registers are the same. 3882 * - Their address offsets are in the same order. 3883 * - They operate at the same memory width. 3884 * - There is no jump into the middle of them. 3885 */ 3886 static bool 3887 curr_pair_chain_with_previous(struct nfp_insn_meta *ld_meta, 3888 struct nfp_insn_meta *st_meta, 3889 struct bpf_insn *prev_ld, 3890 struct bpf_insn *prev_st) 3891 { 3892 u8 prev_size, curr_size, prev_ld_base, prev_st_base, prev_ld_dst; 3893 struct bpf_insn *ld = &ld_meta->insn; 3894 struct bpf_insn *st = &st_meta->insn; 3895 s16 prev_ld_off, prev_st_off; 3896 3897 /* This pair is the start pair. */ 3898 if (!prev_ld) 3899 return true; 3900 3901 prev_size = BPF_LDST_BYTES(prev_ld); 3902 curr_size = BPF_LDST_BYTES(ld); 3903 prev_ld_base = prev_ld->src_reg; 3904 prev_st_base = prev_st->dst_reg; 3905 prev_ld_dst = prev_ld->dst_reg; 3906 prev_ld_off = prev_ld->off; 3907 prev_st_off = prev_st->off; 3908 3909 if (ld->dst_reg != prev_ld_dst) 3910 return false; 3911 3912 if (ld->src_reg != prev_ld_base || st->dst_reg != prev_st_base) 3913 return false; 3914 3915 if (curr_size != prev_size) 3916 return false; 3917 3918 /* There is jump to the head of this pair. */ 3919 if (ld_meta->flags & FLAG_INSN_IS_JUMP_DST) 3920 return false; 3921 3922 /* Both in ascending order. */ 3923 if (prev_ld_off + prev_size == ld->off && 3924 prev_st_off + prev_size == st->off) 3925 return true; 3926 3927 /* Both in descending order. */ 3928 if (ld->off + curr_size == prev_ld_off && 3929 st->off + curr_size == prev_st_off) 3930 return true; 3931 3932 return false; 3933 } 3934 3935 /* Return TRUE if cross memory access happens. Cross memory access means 3936 * store area is overlapping with load area that a later load might load 3937 * the value from previous store, for this case we can't treat the sequence 3938 * as an memory copy. 3939 */ 3940 static bool 3941 cross_mem_access(struct bpf_insn *ld, struct nfp_insn_meta *head_ld_meta, 3942 struct nfp_insn_meta *head_st_meta) 3943 { 3944 s16 head_ld_off, head_st_off, ld_off; 3945 3946 /* Different pointer types does not overlap. */ 3947 if (head_ld_meta->ptr.type != head_st_meta->ptr.type) 3948 return false; 3949 3950 /* load and store are both PTR_TO_PACKET, check ID info. */ 3951 if (head_ld_meta->ptr.id != head_st_meta->ptr.id) 3952 return true; 3953 3954 /* Canonicalize the offsets. Turn all of them against the original 3955 * base register. 3956 */ 3957 head_ld_off = head_ld_meta->insn.off + head_ld_meta->ptr.off; 3958 head_st_off = head_st_meta->insn.off + head_st_meta->ptr.off; 3959 ld_off = ld->off + head_ld_meta->ptr.off; 3960 3961 /* Ascending order cross. */ 3962 if (ld_off > head_ld_off && 3963 head_ld_off < head_st_off && ld_off >= head_st_off) 3964 return true; 3965 3966 /* Descending order cross. */ 3967 if (ld_off < head_ld_off && 3968 head_ld_off > head_st_off && ld_off <= head_st_off) 3969 return true; 3970 3971 return false; 3972 } 3973 3974 /* This pass try to identify the following instructoin sequences. 3975 * 3976 * load R, [regA + offA] 3977 * store [regB + offB], R 3978 * load R, [regA + offA + const_imm_A] 3979 * store [regB + offB + const_imm_A], R 3980 * load R, [regA + offA + 2 * const_imm_A] 3981 * store [regB + offB + 2 * const_imm_A], R 3982 * ... 3983 * 3984 * Above sequence is typically generated by compiler when lowering 3985 * memcpy. NFP prefer using CPP instructions to accelerate it. 3986 */ 3987 static void nfp_bpf_opt_ldst_gather(struct nfp_prog *nfp_prog) 3988 { 3989 struct nfp_insn_meta *head_ld_meta = NULL; 3990 struct nfp_insn_meta *head_st_meta = NULL; 3991 struct nfp_insn_meta *meta1, *meta2; 3992 struct bpf_insn *prev_ld = NULL; 3993 struct bpf_insn *prev_st = NULL; 3994 u8 count = 0; 3995 3996 nfp_for_each_insn_walk2(nfp_prog, meta1, meta2) { 3997 struct bpf_insn *ld = &meta1->insn; 3998 struct bpf_insn *st = &meta2->insn; 3999 4000 /* Reset record status if any of the following if true: 4001 * - The current insn pair is not load/store. 4002 * - The load/store pair doesn't chain with previous one. 4003 * - The chained load/store pair crossed with previous pair. 4004 * - The chained load/store pair has a total size of memory 4005 * copy beyond 128 bytes which is the maximum length a 4006 * single NFP CPP command can transfer. 4007 */ 4008 if (!curr_pair_is_memcpy(meta1, meta2) || 4009 !curr_pair_chain_with_previous(meta1, meta2, prev_ld, 4010 prev_st) || 4011 (head_ld_meta && (cross_mem_access(ld, head_ld_meta, 4012 head_st_meta) || 4013 head_ld_meta->ldst_gather_len >= 128))) { 4014 if (!count) 4015 continue; 4016 4017 if (count > 1) { 4018 s16 prev_ld_off = prev_ld->off; 4019 s16 prev_st_off = prev_st->off; 4020 s16 head_ld_off = head_ld_meta->insn.off; 4021 4022 if (prev_ld_off < head_ld_off) { 4023 head_ld_meta->insn.off = prev_ld_off; 4024 head_st_meta->insn.off = prev_st_off; 4025 head_ld_meta->ldst_gather_len = 4026 -head_ld_meta->ldst_gather_len; 4027 } 4028 4029 head_ld_meta->paired_st = &head_st_meta->insn; 4030 head_st_meta->skip = true; 4031 } else { 4032 head_ld_meta->ldst_gather_len = 0; 4033 } 4034 4035 /* If the chain is ended by an load/store pair then this 4036 * could serve as the new head of the the next chain. 4037 */ 4038 if (curr_pair_is_memcpy(meta1, meta2)) { 4039 head_ld_meta = meta1; 4040 head_st_meta = meta2; 4041 head_ld_meta->ldst_gather_len = 4042 BPF_LDST_BYTES(ld); 4043 meta1 = nfp_meta_next(meta1); 4044 meta2 = nfp_meta_next(meta2); 4045 prev_ld = ld; 4046 prev_st = st; 4047 count = 1; 4048 } else { 4049 head_ld_meta = NULL; 4050 head_st_meta = NULL; 4051 prev_ld = NULL; 4052 prev_st = NULL; 4053 count = 0; 4054 } 4055 4056 continue; 4057 } 4058 4059 if (!head_ld_meta) { 4060 head_ld_meta = meta1; 4061 head_st_meta = meta2; 4062 } else { 4063 meta1->skip = true; 4064 meta2->skip = true; 4065 } 4066 4067 head_ld_meta->ldst_gather_len += BPF_LDST_BYTES(ld); 4068 meta1 = nfp_meta_next(meta1); 4069 meta2 = nfp_meta_next(meta2); 4070 prev_ld = ld; 4071 prev_st = st; 4072 count++; 4073 } 4074 } 4075 4076 static void nfp_bpf_opt_pkt_cache(struct nfp_prog *nfp_prog) 4077 { 4078 struct nfp_insn_meta *meta, *range_node = NULL; 4079 s16 range_start = 0, range_end = 0; 4080 bool cache_avail = false; 4081 struct bpf_insn *insn; 4082 s32 range_ptr_off = 0; 4083 u32 range_ptr_id = 0; 4084 4085 list_for_each_entry(meta, &nfp_prog->insns, l) { 4086 if (meta->flags & FLAG_INSN_IS_JUMP_DST) 4087 cache_avail = false; 4088 4089 if (meta->skip) 4090 continue; 4091 4092 insn = &meta->insn; 4093 4094 if (is_mbpf_store_pkt(meta) || 4095 insn->code == (BPF_JMP | BPF_CALL) || 4096 is_mbpf_classic_store_pkt(meta) || 4097 is_mbpf_classic_load(meta)) { 4098 cache_avail = false; 4099 continue; 4100 } 4101 4102 if (!is_mbpf_load(meta)) 4103 continue; 4104 4105 if (meta->ptr.type != PTR_TO_PACKET || meta->ldst_gather_len) { 4106 cache_avail = false; 4107 continue; 4108 } 4109 4110 if (!cache_avail) { 4111 cache_avail = true; 4112 if (range_node) 4113 goto end_current_then_start_new; 4114 goto start_new; 4115 } 4116 4117 /* Check ID to make sure two reads share the same 4118 * variable offset against PTR_TO_PACKET, and check OFF 4119 * to make sure they also share the same constant 4120 * offset. 4121 * 4122 * OFFs don't really need to be the same, because they 4123 * are the constant offsets against PTR_TO_PACKET, so 4124 * for different OFFs, we could canonicalize them to 4125 * offsets against original packet pointer. We don't 4126 * support this. 4127 */ 4128 if (meta->ptr.id == range_ptr_id && 4129 meta->ptr.off == range_ptr_off) { 4130 s16 new_start = range_start; 4131 s16 end, off = insn->off; 4132 s16 new_end = range_end; 4133 bool changed = false; 4134 4135 if (off < range_start) { 4136 new_start = off; 4137 changed = true; 4138 } 4139 4140 end = off + BPF_LDST_BYTES(insn); 4141 if (end > range_end) { 4142 new_end = end; 4143 changed = true; 4144 } 4145 4146 if (!changed) 4147 continue; 4148 4149 if (new_end - new_start <= 64) { 4150 /* Install new range. */ 4151 range_start = new_start; 4152 range_end = new_end; 4153 continue; 4154 } 4155 } 4156 4157 end_current_then_start_new: 4158 range_node->pkt_cache.range_start = range_start; 4159 range_node->pkt_cache.range_end = range_end; 4160 start_new: 4161 range_node = meta; 4162 range_node->pkt_cache.do_init = true; 4163 range_ptr_id = range_node->ptr.id; 4164 range_ptr_off = range_node->ptr.off; 4165 range_start = insn->off; 4166 range_end = insn->off + BPF_LDST_BYTES(insn); 4167 } 4168 4169 if (range_node) { 4170 range_node->pkt_cache.range_start = range_start; 4171 range_node->pkt_cache.range_end = range_end; 4172 } 4173 4174 list_for_each_entry(meta, &nfp_prog->insns, l) { 4175 if (meta->skip) 4176 continue; 4177 4178 if (is_mbpf_load_pkt(meta) && !meta->ldst_gather_len) { 4179 if (meta->pkt_cache.do_init) { 4180 range_start = meta->pkt_cache.range_start; 4181 range_end = meta->pkt_cache.range_end; 4182 } else { 4183 meta->pkt_cache.range_start = range_start; 4184 meta->pkt_cache.range_end = range_end; 4185 } 4186 } 4187 } 4188 } 4189 4190 static int nfp_bpf_optimize(struct nfp_prog *nfp_prog) 4191 { 4192 nfp_bpf_opt_reg_init(nfp_prog); 4193 4194 nfp_bpf_opt_neg_add_sub(nfp_prog); 4195 nfp_bpf_opt_ld_mask(nfp_prog); 4196 nfp_bpf_opt_ld_shift(nfp_prog); 4197 nfp_bpf_opt_ldst_gather(nfp_prog); 4198 nfp_bpf_opt_pkt_cache(nfp_prog); 4199 4200 return 0; 4201 } 4202 4203 static int nfp_bpf_replace_map_ptrs(struct nfp_prog *nfp_prog) 4204 { 4205 struct nfp_insn_meta *meta1, *meta2; 4206 struct nfp_bpf_map *nfp_map; 4207 struct bpf_map *map; 4208 u32 id; 4209 4210 nfp_for_each_insn_walk2(nfp_prog, meta1, meta2) { 4211 if (meta1->skip || meta2->skip) 4212 continue; 4213 4214 if (meta1->insn.code != (BPF_LD | BPF_IMM | BPF_DW) || 4215 meta1->insn.src_reg != BPF_PSEUDO_MAP_FD) 4216 continue; 4217 4218 map = (void *)(unsigned long)((u32)meta1->insn.imm | 4219 (u64)meta2->insn.imm << 32); 4220 if (bpf_map_offload_neutral(map)) { 4221 id = map->id; 4222 } else { 4223 nfp_map = map_to_offmap(map)->dev_priv; 4224 id = nfp_map->tid; 4225 } 4226 4227 meta1->insn.imm = id; 4228 meta2->insn.imm = 0; 4229 } 4230 4231 return 0; 4232 } 4233 4234 static int nfp_bpf_ustore_calc(u64 *prog, unsigned int len) 4235 { 4236 __le64 *ustore = (__force __le64 *)prog; 4237 int i; 4238 4239 for (i = 0; i < len; i++) { 4240 int err; 4241 4242 err = nfp_ustore_check_valid_no_ecc(prog[i]); 4243 if (err) 4244 return err; 4245 4246 ustore[i] = cpu_to_le64(nfp_ustore_calc_ecc_insn(prog[i])); 4247 } 4248 4249 return 0; 4250 } 4251 4252 static void nfp_bpf_prog_trim(struct nfp_prog *nfp_prog) 4253 { 4254 void *prog; 4255 4256 prog = kvmalloc_array(nfp_prog->prog_len, sizeof(u64), GFP_KERNEL); 4257 if (!prog) 4258 return; 4259 4260 nfp_prog->__prog_alloc_len = nfp_prog->prog_len * sizeof(u64); 4261 memcpy(prog, nfp_prog->prog, nfp_prog->__prog_alloc_len); 4262 kvfree(nfp_prog->prog); 4263 nfp_prog->prog = prog; 4264 } 4265 4266 int nfp_bpf_jit(struct nfp_prog *nfp_prog) 4267 { 4268 int ret; 4269 4270 ret = nfp_bpf_replace_map_ptrs(nfp_prog); 4271 if (ret) 4272 return ret; 4273 4274 ret = nfp_bpf_optimize(nfp_prog); 4275 if (ret) 4276 return ret; 4277 4278 ret = nfp_translate(nfp_prog); 4279 if (ret) { 4280 pr_err("Translation failed with error %d (translated: %u)\n", 4281 ret, nfp_prog->n_translated); 4282 return -EINVAL; 4283 } 4284 4285 nfp_bpf_prog_trim(nfp_prog); 4286 4287 return ret; 4288 } 4289 4290 void nfp_bpf_jit_prepare(struct nfp_prog *nfp_prog, unsigned int cnt) 4291 { 4292 struct nfp_insn_meta *meta; 4293 4294 /* Another pass to record jump information. */ 4295 list_for_each_entry(meta, &nfp_prog->insns, l) { 4296 struct nfp_insn_meta *dst_meta; 4297 u64 code = meta->insn.code; 4298 unsigned int dst_idx; 4299 bool pseudo_call; 4300 4301 if (BPF_CLASS(code) != BPF_JMP) 4302 continue; 4303 if (BPF_OP(code) == BPF_EXIT) 4304 continue; 4305 if (is_mbpf_helper_call(meta)) 4306 continue; 4307 4308 /* If opcode is BPF_CALL at this point, this can only be a 4309 * BPF-to-BPF call (a.k.a pseudo call). 4310 */ 4311 pseudo_call = BPF_OP(code) == BPF_CALL; 4312 4313 if (pseudo_call) 4314 dst_idx = meta->n + 1 + meta->insn.imm; 4315 else 4316 dst_idx = meta->n + 1 + meta->insn.off; 4317 4318 dst_meta = nfp_bpf_goto_meta(nfp_prog, meta, dst_idx, cnt); 4319 4320 if (pseudo_call) 4321 dst_meta->flags |= FLAG_INSN_IS_SUBPROG_START; 4322 4323 dst_meta->flags |= FLAG_INSN_IS_JUMP_DST; 4324 meta->jmp_dst = dst_meta; 4325 } 4326 } 4327 4328 bool nfp_bpf_supported_opcode(u8 code) 4329 { 4330 return !!instr_cb[code]; 4331 } 4332 4333 void *nfp_bpf_relo_for_vnic(struct nfp_prog *nfp_prog, struct nfp_bpf_vnic *bv) 4334 { 4335 unsigned int i; 4336 u64 *prog; 4337 int err; 4338 4339 prog = kmemdup(nfp_prog->prog, nfp_prog->prog_len * sizeof(u64), 4340 GFP_KERNEL); 4341 if (!prog) 4342 return ERR_PTR(-ENOMEM); 4343 4344 for (i = 0; i < nfp_prog->prog_len; i++) { 4345 enum nfp_relo_type special; 4346 u32 val; 4347 u16 off; 4348 4349 special = FIELD_GET(OP_RELO_TYPE, prog[i]); 4350 switch (special) { 4351 case RELO_NONE: 4352 continue; 4353 case RELO_BR_REL: 4354 br_add_offset(&prog[i], bv->start_off); 4355 break; 4356 case RELO_BR_GO_OUT: 4357 br_set_offset(&prog[i], 4358 nfp_prog->tgt_out + bv->start_off); 4359 break; 4360 case RELO_BR_GO_ABORT: 4361 br_set_offset(&prog[i], 4362 nfp_prog->tgt_abort + bv->start_off); 4363 break; 4364 case RELO_BR_GO_CALL_PUSH_REGS: 4365 if (!nfp_prog->tgt_call_push_regs) { 4366 pr_err("BUG: failed to detect subprogram registers needs\n"); 4367 err = -EINVAL; 4368 goto err_free_prog; 4369 } 4370 off = nfp_prog->tgt_call_push_regs + bv->start_off; 4371 br_set_offset(&prog[i], off); 4372 break; 4373 case RELO_BR_GO_CALL_POP_REGS: 4374 if (!nfp_prog->tgt_call_pop_regs) { 4375 pr_err("BUG: failed to detect subprogram registers needs\n"); 4376 err = -EINVAL; 4377 goto err_free_prog; 4378 } 4379 off = nfp_prog->tgt_call_pop_regs + bv->start_off; 4380 br_set_offset(&prog[i], off); 4381 break; 4382 case RELO_BR_NEXT_PKT: 4383 br_set_offset(&prog[i], bv->tgt_done); 4384 break; 4385 case RELO_BR_HELPER: 4386 val = br_get_offset(prog[i]); 4387 val -= BR_OFF_RELO; 4388 switch (val) { 4389 case BPF_FUNC_map_lookup_elem: 4390 val = nfp_prog->bpf->helpers.map_lookup; 4391 break; 4392 case BPF_FUNC_map_update_elem: 4393 val = nfp_prog->bpf->helpers.map_update; 4394 break; 4395 case BPF_FUNC_map_delete_elem: 4396 val = nfp_prog->bpf->helpers.map_delete; 4397 break; 4398 case BPF_FUNC_perf_event_output: 4399 val = nfp_prog->bpf->helpers.perf_event_output; 4400 break; 4401 default: 4402 pr_err("relocation of unknown helper %d\n", 4403 val); 4404 err = -EINVAL; 4405 goto err_free_prog; 4406 } 4407 br_set_offset(&prog[i], val); 4408 break; 4409 case RELO_IMMED_REL: 4410 immed_add_value(&prog[i], bv->start_off); 4411 break; 4412 } 4413 4414 prog[i] &= ~OP_RELO_TYPE; 4415 } 4416 4417 err = nfp_bpf_ustore_calc(prog, nfp_prog->prog_len); 4418 if (err) 4419 goto err_free_prog; 4420 4421 return prog; 4422 4423 err_free_prog: 4424 kfree(prog); 4425 return ERR_PTR(err); 4426 } 4427