1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright 2017 6WIND S.A. 3 * Copyright 2017 Mellanox Technologies, Ltd 4 */ 5 6 #ifndef RTE_PMD_MLX5_RXTX_VEC_ALTIVEC_H_ 7 #define RTE_PMD_MLX5_RXTX_VEC_ALTIVEC_H_ 8 9 #include <stdint.h> 10 #include <string.h> 11 #include <stdlib.h> 12 13 #include <rte_altivec.h> 14 15 #include <rte_mbuf.h> 16 #include <rte_mempool.h> 17 #include <rte_prefetch.h> 18 19 #include <mlx5_prm.h> 20 21 #include "mlx5_defs.h" 22 #include "mlx5.h" 23 #include "mlx5_utils.h" 24 #include "mlx5_rxtx.h" 25 #include "mlx5_rxtx_vec.h" 26 #include "mlx5_autoconf.h" 27 28 #ifndef __INTEL_COMPILER 29 #pragma GCC diagnostic ignored "-Wcast-qual" 30 #pragma GCC diagnostic ignored "-Wstrict-aliasing" 31 #endif 32 33 /** 34 * Store free buffers to RX SW ring. 35 * 36 * @param elts 37 * Pointer to SW ring to be filled. 38 * @param pkts 39 * Pointer to array of packets to be stored. 40 * @param pkts_n 41 * Number of packets to be stored. 42 */ 43 static inline void 44 rxq_copy_mbuf_v(struct rte_mbuf **elts, struct rte_mbuf **pkts, uint16_t n) 45 { 46 unsigned int pos; 47 uint16_t p = n & -2; 48 49 for (pos = 0; pos < p; pos += 2) { 50 __vector unsigned char mbp; 51 52 mbp = (__vector unsigned char)vec_vsx_ld(0, 53 (signed int const *)&elts[pos]); 54 *(__vector unsigned char *)&pkts[pos] = mbp; 55 } 56 if (n & 1) 57 pkts[pos] = elts[pos]; 58 } 59 60 /** 61 * Decompress a compressed completion and fill in mbufs in RX SW ring with data 62 * extracted from the title completion descriptor. 63 * 64 * @param rxq 65 * Pointer to RX queue structure. 66 * @param cq 67 * Pointer to completion array having a compressed completion at first. 68 * @param elts 69 * Pointer to SW ring to be filled. The first mbuf has to be pre-built from 70 * the title completion descriptor to be copied to the rest of mbufs. 71 * @param keep 72 * Keep unzipping if the next CQE is the miniCQE array. 73 * 74 * @return 75 * Number of mini-CQEs successfully decompressed. 76 */ 77 static inline uint16_t 78 rxq_cq_decompress_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq, 79 struct rte_mbuf **elts, bool keep) 80 { 81 volatile struct mlx5_mini_cqe8 *mcq = 82 (void *)&(cq + !rxq->cqe_comp_layout)->pkt_info; 83 /* Title packet is pre-built. */ 84 struct rte_mbuf *t_pkt = rxq->cqe_comp_layout ? &rxq->title_pkt : elts[0]; 85 const __vector unsigned char zero = (__vector unsigned char){0}; 86 /* Mask to shuffle from extracted mini CQE to mbuf. */ 87 const __vector unsigned char shuf_mask1 = (__vector unsigned char){ 88 -1, -1, -1, -1, /* skip packet_type */ 89 7, 6, -1, -1, /* bswap16, pkt_len */ 90 7, 6, /* bswap16, data_len */ 91 -1, -1, /* skip vlan_tci */ 92 3, 2, 1, 0}; /* bswap32, rss */ 93 const __vector unsigned char shuf_mask2 = (__vector unsigned char){ 94 -1, -1, -1, -1, /* skip packet_type */ 95 15, 14, -1, -1, /* bswap16, pkt_len */ 96 15, 14, /* data_len, bswap16 */ 97 -1, -1, /* skip vlan_tci */ 98 11, 10, 9, 8}; /* bswap32, rss */ 99 /* Restore the compressed count. Must be 16 bits. */ 100 uint16_t mcqe_n = (rxq->cqe_comp_layout) ? 101 (MLX5_CQE_NUM_MINIS(cq->op_own) + 1U) : rte_be_to_cpu_32(cq->byte_cnt); 102 uint16_t pkts_n = mcqe_n; 103 const __vector unsigned char rearm = 104 (__vector unsigned char)vec_vsx_ld(0, 105 (signed int const *)&t_pkt->rearm_data); 106 const __vector unsigned char rxdf = 107 (__vector unsigned char)vec_vsx_ld(0, 108 (signed int const *)&t_pkt->rx_descriptor_fields1); 109 const __vector unsigned char crc_adj = 110 (__vector unsigned char)(__vector unsigned short){ 111 0, 0, rxq->crc_present * RTE_ETHER_CRC_LEN, 0, 112 rxq->crc_present * RTE_ETHER_CRC_LEN, 0, 0, 0}; 113 const __vector unsigned short rxdf_sel_mask = 114 (__vector unsigned short){ 115 0xffff, 0xffff, 0, 0, 0, 0xffff, 0, 0}; 116 __vector unsigned char ol_flags = (__vector unsigned char){0}; 117 __vector unsigned char ol_flags_mask = (__vector unsigned char){0}; 118 unsigned int pos; 119 unsigned int i; 120 unsigned int inv = 0; 121 122 #ifdef MLX5_PMD_SOFT_COUNTERS 123 const __vector unsigned char ones = vec_splat_u8(-1); 124 uint32_t rcvd_byte = 0; 125 /* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */ 126 const __vector unsigned char len_shuf_mask = (__vector unsigned char){ 127 3, 2, 11, 10, 128 7, 6, 15, 14, 129 -1, -1, -1, -1, 130 -1, -1, -1, -1}; 131 #endif 132 133 /* 134 * A. load mCQEs into a 128bit register. 135 * B. store rearm data to mbuf. 136 * C. combine data from mCQEs with rx_descriptor_fields1. 137 * D. store rx_descriptor_fields1. 138 * E. store flow tag (rte_flow mark). 139 */ 140 cycle: 141 if (rxq->cqe_comp_layout) 142 rte_prefetch0((void *)(cq + mcqe_n)); 143 for (pos = 0; pos < mcqe_n; ) { 144 __vector unsigned char mcqe1, mcqe2; 145 __vector unsigned char rxdf1, rxdf2; 146 #ifdef MLX5_PMD_SOFT_COUNTERS 147 const __vector unsigned short mcqe_sel_mask = 148 (__vector unsigned short){0, 0, 0xffff, 0xffff, 149 0, 0, 0xfff, 0xffff}; 150 const __vector unsigned char lower_half = { 151 0, 1, 4, 5, 8, 9, 12, 13, 16, 152 17, 20, 21, 24, 25, 28, 29}; 153 const __vector unsigned char upper_half = { 154 2, 3, 6, 7, 10, 11, 14, 15, 155 18, 19, 22, 23, 26, 27, 30, 31}; 156 __vector unsigned short left, right; 157 __vector unsigned char byte_cnt, invalid_mask; 158 __vector unsigned long lshift; 159 __attribute__((altivec(vector__))) 160 __attribute__((altivec(bool__))) 161 unsigned long long shmask; 162 const __vector unsigned long shmax = {64, 64}; 163 #endif 164 165 if (!rxq->cqe_comp_layout) 166 for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i) 167 if (likely(pos + i < mcqe_n)) 168 rte_prefetch0((void *)(cq + pos + i)); 169 /* A.1 load mCQEs into a 128bit register. */ 170 mcqe1 = (__vector unsigned char)vec_vsx_ld(0, 171 (signed int const *)&mcq[pos % 8]); 172 mcqe2 = (__vector unsigned char)vec_vsx_ld(0, 173 (signed int const *)&mcq[pos % 8 + 2]); 174 175 /* B.1 store rearm data to mbuf. */ 176 *(__vector unsigned char *) 177 &elts[pos]->rearm_data = rearm; 178 *(__vector unsigned char *) 179 &elts[pos + 1]->rearm_data = rearm; 180 181 /* C.1 combine data from mCQEs with rx_descriptor_fields1. */ 182 rxdf1 = vec_perm(mcqe1, zero, shuf_mask1); 183 rxdf2 = vec_perm(mcqe1, zero, shuf_mask2); 184 rxdf1 = (__vector unsigned char) 185 ((__vector unsigned short)rxdf1 - 186 (__vector unsigned short)crc_adj); 187 rxdf2 = (__vector unsigned char) 188 ((__vector unsigned short)rxdf2 - 189 (__vector unsigned short)crc_adj); 190 rxdf1 = (__vector unsigned char) 191 vec_sel((__vector unsigned short)rxdf1, 192 (__vector unsigned short)rxdf, rxdf_sel_mask); 193 rxdf2 = (__vector unsigned char) 194 vec_sel((__vector unsigned short)rxdf2, 195 (__vector unsigned short)rxdf, rxdf_sel_mask); 196 197 /* D.1 store rx_descriptor_fields1. */ 198 *(__vector unsigned char *) 199 &elts[pos]->rx_descriptor_fields1 = rxdf1; 200 *(__vector unsigned char *) 201 &elts[pos + 1]->rx_descriptor_fields1 = rxdf2; 202 203 /* B.1 store rearm data to mbuf. */ 204 *(__vector unsigned char *) 205 &elts[pos + 2]->rearm_data = rearm; 206 *(__vector unsigned char *) 207 &elts[pos + 3]->rearm_data = rearm; 208 209 /* C.1 combine data from mCQEs with rx_descriptor_fields1. */ 210 rxdf1 = vec_perm(mcqe2, zero, shuf_mask1); 211 rxdf2 = vec_perm(mcqe2, zero, shuf_mask2); 212 rxdf1 = (__vector unsigned char) 213 ((__vector unsigned short)rxdf1 - 214 (__vector unsigned short)crc_adj); 215 rxdf2 = (__vector unsigned char) 216 ((__vector unsigned short)rxdf2 - 217 (__vector unsigned short)crc_adj); 218 rxdf1 = (__vector unsigned char) 219 vec_sel((__vector unsigned short)rxdf1, 220 (__vector unsigned short)rxdf, rxdf_sel_mask); 221 rxdf2 = (__vector unsigned char) 222 vec_sel((__vector unsigned short)rxdf2, 223 (__vector unsigned short)rxdf, rxdf_sel_mask); 224 225 /* D.1 store rx_descriptor_fields1. */ 226 *(__vector unsigned char *) 227 &elts[pos + 2]->rx_descriptor_fields1 = rxdf1; 228 *(__vector unsigned char *) 229 &elts[pos + 3]->rx_descriptor_fields1 = rxdf2; 230 231 #ifdef MLX5_PMD_SOFT_COUNTERS 232 invalid_mask = (__vector unsigned char)(__vector unsigned long){ 233 (mcqe_n - pos) * sizeof(uint16_t) * 8, 0}; 234 235 lshift = 236 vec_splat((__vector unsigned long)invalid_mask, 0); 237 shmask = vec_cmpgt(shmax, lshift); 238 invalid_mask = (__vector unsigned char) 239 vec_sl((__vector unsigned long)ones, lshift); 240 invalid_mask = (__vector unsigned char) 241 vec_sel((__vector unsigned long)shmask, 242 (__vector unsigned long)invalid_mask, shmask); 243 244 byte_cnt = (__vector unsigned char) 245 vec_sel((__vector unsigned short) 246 vec_sro((__vector unsigned short)mcqe1, 247 (__vector unsigned char){32}), 248 (__vector unsigned short)mcqe2, mcqe_sel_mask); 249 byte_cnt = vec_perm(byte_cnt, zero, len_shuf_mask); 250 byte_cnt = (__vector unsigned char) 251 vec_andc((__vector unsigned long)byte_cnt, 252 (__vector unsigned long)invalid_mask); 253 left = vec_perm((__vector unsigned short)byte_cnt, 254 (__vector unsigned short)zero, lower_half); 255 right = vec_perm((__vector unsigned short)byte_cnt, 256 (__vector unsigned short)zero, upper_half); 257 byte_cnt = (__vector unsigned char)vec_add(left, right); 258 left = vec_perm((__vector unsigned short)byte_cnt, 259 (__vector unsigned short)zero, lower_half); 260 right = vec_perm((__vector unsigned short)byte_cnt, 261 (__vector unsigned short)zero, upper_half); 262 byte_cnt = (__vector unsigned char)vec_add(left, right); 263 rcvd_byte += ((__vector unsigned long)byte_cnt)[0]; 264 #endif 265 266 if (rxq->mark) { 267 if (rxq->mcqe_format != 268 MLX5_CQE_RESP_FORMAT_FTAG_STRIDX) { 269 const uint32_t flow_tag = t_pkt->hash.fdir.hi; 270 271 /* E.1 store flow tag (rte_flow mark). */ 272 elts[pos]->hash.fdir.hi = flow_tag; 273 elts[pos + 1]->hash.fdir.hi = flow_tag; 274 elts[pos + 2]->hash.fdir.hi = flow_tag; 275 elts[pos + 3]->hash.fdir.hi = flow_tag; 276 } else { 277 const __vector unsigned char flow_mark_adj = 278 (__vector unsigned char) 279 (__vector unsigned int){ 280 -1, -1, -1, -1}; 281 const __vector unsigned char flow_mark_shuf = 282 (__vector unsigned char){ 283 -1, -1, -1, -1, 284 -1, -1, -1, -1, 285 12, 8, 9, -1, 286 4, 0, 1, -1}; 287 const __vector unsigned char ft_mask = 288 (__vector unsigned char) 289 (__vector unsigned int){ 290 0xffffff00, 0xffffff00, 291 0xffffff00, 0xffffff00}; 292 const __vector unsigned char fdir_flags = 293 (__vector unsigned char) 294 (__vector unsigned int){ 295 RTE_MBUF_F_RX_FDIR, RTE_MBUF_F_RX_FDIR, 296 RTE_MBUF_F_RX_FDIR, RTE_MBUF_F_RX_FDIR}; 297 const __vector unsigned char fdir_all_flags = 298 (__vector unsigned char) 299 (__vector unsigned int){ 300 RTE_MBUF_F_RX_FDIR | rxq->mark_flag, 301 RTE_MBUF_F_RX_FDIR | rxq->mark_flag, 302 RTE_MBUF_F_RX_FDIR | rxq->mark_flag, 303 RTE_MBUF_F_RX_FDIR | rxq->mark_flag}; 304 __vector unsigned char fdir_id_flags = 305 (__vector unsigned char) 306 (__vector unsigned int){ 307 rxq->mark_flag, rxq->mark_flag, 308 rxq->mark_flag, rxq->mark_flag}; 309 /* Extract flow_tag field. */ 310 __vector unsigned char ftag0 = vec_perm(mcqe1, 311 zero, flow_mark_shuf); 312 __vector unsigned char ftag1 = vec_perm(mcqe2, 313 zero, flow_mark_shuf); 314 __vector unsigned char ftag = 315 (__vector unsigned char) 316 vec_mergel((__vector unsigned int)ftag0, 317 (__vector unsigned int)ftag1); 318 __vector unsigned char invalid_mask = 319 (__vector unsigned char) 320 vec_cmpeq((__vector unsigned int)ftag, 321 (__vector unsigned int)zero); 322 323 ol_flags_mask = (__vector unsigned char) 324 vec_or((__vector unsigned long) 325 ol_flags_mask, 326 (__vector unsigned long)fdir_all_flags); 327 328 /* Set RTE_MBUF_F_RX_FDIR if flow tag is non-zero. */ 329 invalid_mask = (__vector unsigned char) 330 vec_cmpeq((__vector unsigned int)ftag, 331 (__vector unsigned int)zero); 332 ol_flags = (__vector unsigned char) 333 vec_or((__vector unsigned long)ol_flags, 334 (__vector unsigned long) 335 vec_andc((__vector unsigned long) 336 fdir_flags, 337 (__vector unsigned long)invalid_mask)); 338 ol_flags_mask = (__vector unsigned char) 339 vec_or((__vector unsigned long) 340 ol_flags_mask, 341 (__vector unsigned long)fdir_flags); 342 343 /* Mask out invalid entries. */ 344 fdir_id_flags = (__vector unsigned char) 345 vec_andc((__vector unsigned long) 346 fdir_id_flags, 347 (__vector unsigned long)invalid_mask); 348 349 /* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */ 350 ol_flags = (__vector unsigned char) 351 vec_or((__vector unsigned long)ol_flags, 352 (__vector unsigned long) 353 vec_andc((__vector unsigned long) 354 fdir_id_flags, 355 (__vector unsigned long) 356 vec_cmpeq((__vector unsigned int)ftag, 357 (__vector unsigned int)ft_mask))); 358 359 ftag = (__vector unsigned char) 360 ((__vector unsigned int)ftag + 361 (__vector unsigned int)flow_mark_adj); 362 elts[pos]->hash.fdir.hi = 363 ((__vector unsigned int)ftag)[0]; 364 elts[pos + 1]->hash.fdir.hi = 365 ((__vector unsigned int)ftag)[1]; 366 elts[pos + 2]->hash.fdir.hi = 367 ((__vector unsigned int)ftag)[2]; 368 elts[pos + 3]->hash.fdir.hi = 369 ((__vector unsigned int)ftag)[3]; 370 } 371 } 372 if (unlikely(rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_HASH)) { 373 if (rxq->mcqe_format == 374 MLX5_CQE_RESP_FORMAT_L34H_STRIDX) { 375 const uint8_t pkt_info = 376 (cq->pkt_info & 0x3) << 6; 377 const uint8_t pkt_hdr0 = 378 mcq[pos % 8].hdr_type; 379 const uint8_t pkt_hdr1 = 380 mcq[pos % 8 + 1].hdr_type; 381 const uint8_t pkt_hdr2 = 382 mcq[pos % 8 + 2].hdr_type; 383 const uint8_t pkt_hdr3 = 384 mcq[pos % 8 + 3].hdr_type; 385 const __vector unsigned char vlan_mask = 386 (__vector unsigned char) 387 (__vector unsigned int) { 388 (RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED), 389 (RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED), 390 (RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED), 391 (RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED)}; 392 const __vector unsigned char cv_mask = 393 (__vector unsigned char) 394 (__vector unsigned int) { 395 MLX5_CQE_VLAN_STRIPPED, 396 MLX5_CQE_VLAN_STRIPPED, 397 MLX5_CQE_VLAN_STRIPPED, 398 MLX5_CQE_VLAN_STRIPPED}; 399 __vector unsigned char pkt_cv = 400 (__vector unsigned char) 401 (__vector unsigned int) { 402 pkt_hdr0 & 0x1, pkt_hdr1 & 0x1, 403 pkt_hdr2 & 0x1, pkt_hdr3 & 0x1}; 404 405 ol_flags_mask = (__vector unsigned char) 406 vec_or((__vector unsigned long) 407 ol_flags_mask, 408 (__vector unsigned long)vlan_mask); 409 ol_flags = (__vector unsigned char) 410 vec_or((__vector unsigned long)ol_flags, 411 (__vector unsigned long) 412 vec_and((__vector unsigned long)vlan_mask, 413 (__vector unsigned long) 414 vec_cmpeq((__vector unsigned int)pkt_cv, 415 (__vector unsigned int)cv_mask))); 416 elts[pos]->packet_type = 417 mlx5_ptype_table[(pkt_hdr0 >> 2) | 418 pkt_info]; 419 elts[pos + 1]->packet_type = 420 mlx5_ptype_table[(pkt_hdr1 >> 2) | 421 pkt_info]; 422 elts[pos + 2]->packet_type = 423 mlx5_ptype_table[(pkt_hdr2 >> 2) | 424 pkt_info]; 425 elts[pos + 3]->packet_type = 426 mlx5_ptype_table[(pkt_hdr3 >> 2) | 427 pkt_info]; 428 if (rxq->tunnel) { 429 elts[pos]->packet_type |= 430 !!(((pkt_hdr0 >> 2) | 431 pkt_info) & (1 << 6)); 432 elts[pos + 1]->packet_type |= 433 !!(((pkt_hdr1 >> 2) | 434 pkt_info) & (1 << 6)); 435 elts[pos + 2]->packet_type |= 436 !!(((pkt_hdr2 >> 2) | 437 pkt_info) & (1 << 6)); 438 elts[pos + 3]->packet_type |= 439 !!(((pkt_hdr3 >> 2) | 440 pkt_info) & (1 << 6)); 441 } 442 } 443 const __vector unsigned char hash_mask = 444 (__vector unsigned char)(__vector unsigned int) { 445 RTE_MBUF_F_RX_RSS_HASH, 446 RTE_MBUF_F_RX_RSS_HASH, 447 RTE_MBUF_F_RX_RSS_HASH, 448 RTE_MBUF_F_RX_RSS_HASH}; 449 const __vector unsigned char rearm_flags = 450 (__vector unsigned char)(__vector unsigned int) { 451 (uint32_t)t_pkt->ol_flags, 452 (uint32_t)t_pkt->ol_flags, 453 (uint32_t)t_pkt->ol_flags, 454 (uint32_t)t_pkt->ol_flags}; 455 456 ol_flags_mask = (__vector unsigned char) 457 vec_or((__vector unsigned long)ol_flags_mask, 458 (__vector unsigned long)hash_mask); 459 ol_flags = (__vector unsigned char) 460 vec_or((__vector unsigned long)ol_flags, 461 (__vector unsigned long) 462 vec_andc((__vector unsigned long)rearm_flags, 463 (__vector unsigned long)ol_flags_mask)); 464 465 elts[pos]->ol_flags = 466 ((__vector unsigned int)ol_flags)[0]; 467 elts[pos + 1]->ol_flags = 468 ((__vector unsigned int)ol_flags)[1]; 469 elts[pos + 2]->ol_flags = 470 ((__vector unsigned int)ol_flags)[2]; 471 elts[pos + 3]->ol_flags = 472 ((__vector unsigned int)ol_flags)[3]; 473 elts[pos]->hash.rss = 0; 474 elts[pos + 1]->hash.rss = 0; 475 elts[pos + 2]->hash.rss = 0; 476 elts[pos + 3]->hash.rss = 0; 477 } 478 if (rxq->dynf_meta) { 479 int32_t offs = rxq->flow_meta_offset; 480 const uint32_t meta = 481 *RTE_MBUF_DYNFIELD(t_pkt, offs, uint32_t *); 482 483 /* Check if title packet has valid metadata. */ 484 if (meta) { 485 MLX5_ASSERT(t_pkt->ol_flags & 486 rxq->flow_meta_mask); 487 *RTE_MBUF_DYNFIELD(elts[pos], offs, 488 uint32_t *) = meta; 489 *RTE_MBUF_DYNFIELD(elts[pos + 1], offs, 490 uint32_t *) = meta; 491 *RTE_MBUF_DYNFIELD(elts[pos + 2], offs, 492 uint32_t *) = meta; 493 *RTE_MBUF_DYNFIELD(elts[pos + 3], offs, 494 uint32_t *) = meta; 495 } 496 } 497 498 pos += MLX5_VPMD_DESCS_PER_LOOP; 499 /* Move to next CQE and invalidate consumed CQEs. */ 500 if (!rxq->cqe_comp_layout) { 501 if (!(pos & 0x7) && pos < mcqe_n) { 502 if (pos + 8 < mcqe_n) 503 rte_prefetch0((void *)(cq + pos + 8)); 504 mcq = (void *)&(cq + pos)->pkt_info; 505 for (i = 0; i < 8; ++i) 506 cq[inv++].op_own = MLX5_CQE_INVALIDATE; 507 } 508 } 509 } 510 511 if (rxq->cqe_comp_layout && keep) { 512 int ret; 513 /* Keep unzipping if the next CQE is the miniCQE array. */ 514 cq = &cq[mcqe_n]; 515 ret = check_cqe_iteration(cq, rxq->cqe_n, rxq->cq_ci + pkts_n); 516 if (ret == MLX5_CQE_STATUS_SW_OWN && 517 MLX5_CQE_FORMAT(cq->op_own) == MLX5_COMPRESSED) { 518 pos = 0; 519 elts = &elts[mcqe_n]; 520 mcq = (void *)cq; 521 mcqe_n = MLX5_CQE_NUM_MINIS(cq->op_own) + 1; 522 pkts_n += mcqe_n; 523 goto cycle; 524 } 525 } else { 526 /* Invalidate the rest of CQEs. */ 527 for (; inv < pkts_n; ++inv) 528 cq[inv].op_own = MLX5_CQE_INVALIDATE; 529 } 530 531 #ifdef MLX5_PMD_SOFT_COUNTERS 532 rxq->stats.ipackets += pkts_n; 533 rxq->stats.ibytes += rcvd_byte; 534 #endif 535 536 return pkts_n; 537 } 538 539 /** 540 * Calculate packet type and offload flag for mbuf and store it. 541 * 542 * @param rxq 543 * Pointer to RX queue structure. 544 * @param cqes[4] 545 * Array of four 16bytes completions extracted from the original completion 546 * descriptor. 547 * @param op_err 548 * Opcode vector having responder error status. Each field is 4B. 549 * @param pkts 550 * Pointer to array of packets to be filled. 551 */ 552 static inline void 553 rxq_cq_to_ptype_oflags_v(struct mlx5_rxq_data *rxq, 554 __vector unsigned char cqes[4], __vector unsigned char op_err, 555 struct rte_mbuf **pkts) 556 { 557 __vector unsigned char pinfo0, pinfo1; 558 __vector unsigned char pinfo, ptype; 559 __vector unsigned char ol_flags = (__vector unsigned char) 560 (__vector unsigned int){ 561 rxq->rss_hash * RTE_MBUF_F_RX_RSS_HASH | 562 rxq->hw_timestamp * rxq->timestamp_rx_flag, 563 rxq->rss_hash * RTE_MBUF_F_RX_RSS_HASH | 564 rxq->hw_timestamp * rxq->timestamp_rx_flag, 565 rxq->rss_hash * RTE_MBUF_F_RX_RSS_HASH | 566 rxq->hw_timestamp * rxq->timestamp_rx_flag, 567 rxq->rss_hash * RTE_MBUF_F_RX_RSS_HASH | 568 rxq->hw_timestamp * rxq->timestamp_rx_flag}; 569 __vector unsigned char cv_flags; 570 const __vector unsigned char zero = (__vector unsigned char){0}; 571 const __vector unsigned char ptype_mask = 572 (__vector unsigned char)(__vector unsigned int){ 573 0x0000fd06, 0x0000fd06, 0x0000fd06, 0x0000fd06}; 574 const __vector unsigned char ptype_ol_mask = 575 (__vector unsigned char)(__vector unsigned int){ 576 0x00000106, 0x00000106, 0x00000106, 0x00000106}; 577 const __vector unsigned char pinfo_mask = 578 (__vector unsigned char)(__vector unsigned int){ 579 0x00000003, 0x00000003, 0x00000003, 0x00000003}; 580 const __vector unsigned char cv_flag_sel = (__vector unsigned char){ 581 0, (uint8_t)(RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED), 582 (uint8_t)(RTE_MBUF_F_RX_IP_CKSUM_GOOD >> 1), 0, 583 (uint8_t)(RTE_MBUF_F_RX_L4_CKSUM_GOOD >> 1), 0, 584 (uint8_t)((RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_GOOD) >> 1), 585 0, 0, 0, 0, 0, 0, 0, 0, 0}; 586 const __vector unsigned char cv_mask = 587 (__vector unsigned char)(__vector unsigned int){ 588 RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_GOOD | 589 RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED, 590 RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_GOOD | 591 RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED, 592 RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_GOOD | 593 RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED, 594 RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_GOOD | 595 RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED}; 596 const __vector unsigned char mbuf_init = 597 (__vector unsigned char)vec_vsx_ld 598 (0, (__vector unsigned char *)&rxq->mbuf_initializer); 599 const __vector unsigned short rearm_sel_mask = 600 (__vector unsigned short){0, 0, 0, 0, 0xffff, 0xffff, 0, 0}; 601 __vector unsigned char rearm0, rearm1, rearm2, rearm3; 602 uint8_t pt_idx0, pt_idx1, pt_idx2, pt_idx3; 603 604 /* Extract pkt_info field. */ 605 pinfo0 = (__vector unsigned char) 606 vec_mergeh((__vector unsigned int)cqes[0], 607 (__vector unsigned int)cqes[1]); 608 pinfo1 = (__vector unsigned char) 609 vec_mergeh((__vector unsigned int)cqes[2], 610 (__vector unsigned int)cqes[3]); 611 pinfo = (__vector unsigned char) 612 vec_mergeh((__vector unsigned long)pinfo0, 613 (__vector unsigned long)pinfo1); 614 615 /* Extract hdr_type_etc field. */ 616 pinfo0 = (__vector unsigned char) 617 vec_mergel((__vector unsigned int)cqes[0], 618 (__vector unsigned int)cqes[1]); 619 pinfo1 = (__vector unsigned char) 620 vec_mergel((__vector unsigned int)cqes[2], 621 (__vector unsigned int)cqes[3]); 622 ptype = (__vector unsigned char) 623 vec_mergeh((__vector unsigned long)pinfo0, 624 (__vector unsigned long)pinfo1); 625 626 if (rxq->mark) { 627 const __vector unsigned char pinfo_ft_mask = 628 (__vector unsigned char)(__vector unsigned int){ 629 0xffffff00, 0xffffff00, 0xffffff00, 0xffffff00}; 630 const __vector unsigned char fdir_flags = 631 (__vector unsigned char)(__vector unsigned int){ 632 RTE_MBUF_F_RX_FDIR, RTE_MBUF_F_RX_FDIR, 633 RTE_MBUF_F_RX_FDIR, RTE_MBUF_F_RX_FDIR}; 634 __vector unsigned char fdir_id_flags = 635 (__vector unsigned char)(__vector unsigned int){ 636 rxq->mark_flag, rxq->mark_flag, 637 rxq->mark_flag, rxq->mark_flag}; 638 __vector unsigned char flow_tag, invalid_mask; 639 640 flow_tag = (__vector unsigned char) 641 vec_and((__vector unsigned long)pinfo, 642 (__vector unsigned long)pinfo_ft_mask); 643 644 /* Check if flow tag is non-zero then set RTE_MBUF_F_RX_FDIR. */ 645 invalid_mask = (__vector unsigned char) 646 vec_cmpeq((__vector unsigned int)flow_tag, 647 (__vector unsigned int)zero); 648 ol_flags = (__vector unsigned char) 649 vec_or((__vector unsigned long)ol_flags, 650 (__vector unsigned long) 651 vec_andc((__vector unsigned long)fdir_flags, 652 (__vector unsigned long)invalid_mask)); 653 654 /* Mask out invalid entries. */ 655 fdir_id_flags = (__vector unsigned char) 656 vec_andc((__vector unsigned long)fdir_id_flags, 657 (__vector unsigned long)invalid_mask); 658 659 /* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */ 660 ol_flags = (__vector unsigned char) 661 vec_or((__vector unsigned long)ol_flags, 662 (__vector unsigned long) 663 vec_andc((__vector unsigned long)fdir_id_flags, 664 (__vector unsigned long) 665 vec_cmpeq((__vector unsigned int)flow_tag, 666 (__vector unsigned int)pinfo_ft_mask))); 667 } 668 /* 669 * Merge the two fields to generate the following: 670 * bit[1] = l3_ok 671 * bit[2] = l4_ok 672 * bit[8] = cv 673 * bit[11:10] = l3_hdr_type 674 * bit[14:12] = l4_hdr_type 675 * bit[15] = ip_frag 676 * bit[16] = tunneled 677 * bit[17] = outer_l3_type 678 */ 679 ptype = (__vector unsigned char) 680 vec_and((__vector unsigned long)ptype, 681 (__vector unsigned long)ptype_mask); 682 pinfo = (__vector unsigned char) 683 vec_and((__vector unsigned long)pinfo, 684 (__vector unsigned long)pinfo_mask); 685 pinfo = (__vector unsigned char) 686 vec_sl((__vector unsigned int)pinfo, 687 (__vector unsigned int){16, 16, 16, 16}); 688 689 /* Make pinfo has merged fields for ol_flags calculation. */ 690 pinfo = (__vector unsigned char) 691 vec_or((__vector unsigned long)ptype, 692 (__vector unsigned long)pinfo); 693 ptype = (__vector unsigned char) 694 vec_sr((__vector unsigned int)pinfo, 695 (__vector unsigned int){10, 10, 10, 10}); 696 ptype = (__vector unsigned char) 697 vec_packs((__vector unsigned int)ptype, 698 (__vector unsigned int)zero); 699 700 /* Errored packets will have RTE_PTYPE_ALL_MASK. */ 701 op_err = (__vector unsigned char) 702 vec_sr((__vector unsigned short)op_err, 703 (__vector unsigned short){8, 8, 8, 8, 8, 8, 8, 8}); 704 ptype = (__vector unsigned char) 705 vec_or((__vector unsigned long)ptype, 706 (__vector unsigned long)op_err); 707 708 pt_idx0 = (uint8_t)((__vector unsigned char)ptype)[0]; 709 pt_idx1 = (uint8_t)((__vector unsigned char)ptype)[2]; 710 pt_idx2 = (uint8_t)((__vector unsigned char)ptype)[4]; 711 pt_idx3 = (uint8_t)((__vector unsigned char)ptype)[6]; 712 713 pkts[0]->packet_type = mlx5_ptype_table[pt_idx0] | 714 !!(pt_idx0 & (1 << 6)) * rxq->tunnel; 715 pkts[1]->packet_type = mlx5_ptype_table[pt_idx1] | 716 !!(pt_idx1 & (1 << 6)) * rxq->tunnel; 717 pkts[2]->packet_type = mlx5_ptype_table[pt_idx2] | 718 !!(pt_idx2 & (1 << 6)) * rxq->tunnel; 719 pkts[3]->packet_type = mlx5_ptype_table[pt_idx3] | 720 !!(pt_idx3 & (1 << 6)) * rxq->tunnel; 721 722 /* Fill flags for checksum and VLAN. */ 723 pinfo = (__vector unsigned char) 724 vec_and((__vector unsigned long)pinfo, 725 (__vector unsigned long)ptype_ol_mask); 726 pinfo = vec_perm(cv_flag_sel, zero, pinfo); 727 728 /* Locate checksum flags at byte[2:1] and merge with VLAN flags. */ 729 cv_flags = (__vector unsigned char) 730 vec_sl((__vector unsigned int)pinfo, 731 (__vector unsigned int){9, 9, 9, 9}); 732 cv_flags = (__vector unsigned char) 733 vec_or((__vector unsigned long)pinfo, 734 (__vector unsigned long)cv_flags); 735 736 /* Move back flags to start from byte[0]. */ 737 cv_flags = (__vector unsigned char) 738 vec_sr((__vector unsigned int)cv_flags, 739 (__vector unsigned int){8, 8, 8, 8}); 740 741 /* Mask out garbage bits. */ 742 cv_flags = (__vector unsigned char) 743 vec_and((__vector unsigned long)cv_flags, 744 (__vector unsigned long)cv_mask); 745 746 /* Merge to ol_flags. */ 747 ol_flags = (__vector unsigned char) 748 vec_or((__vector unsigned long)ol_flags, 749 (__vector unsigned long)cv_flags); 750 751 /* Merge mbuf_init and ol_flags. */ 752 rearm0 = (__vector unsigned char) 753 vec_sel((__vector unsigned short)mbuf_init, 754 (__vector unsigned short) 755 vec_slo((__vector unsigned short)ol_flags, 756 (__vector unsigned char){64}), rearm_sel_mask); 757 rearm1 = (__vector unsigned char) 758 vec_sel((__vector unsigned short)mbuf_init, 759 (__vector unsigned short) 760 vec_slo((__vector unsigned short)ol_flags, 761 (__vector unsigned char){32}), rearm_sel_mask); 762 rearm2 = (__vector unsigned char) 763 vec_sel((__vector unsigned short)mbuf_init, 764 (__vector unsigned short)ol_flags, rearm_sel_mask); 765 rearm3 = (__vector unsigned char) 766 vec_sel((__vector unsigned short)mbuf_init, 767 (__vector unsigned short) 768 vec_sro((__vector unsigned short)ol_flags, 769 (__vector unsigned char){32}), rearm_sel_mask); 770 771 /* Write 8B rearm_data and 8B ol_flags. */ 772 vec_vsx_st(rearm0, 0, 773 (__vector unsigned char *)&pkts[0]->rearm_data); 774 vec_vsx_st(rearm1, 0, 775 (__vector unsigned char *)&pkts[1]->rearm_data); 776 vec_vsx_st(rearm2, 0, 777 (__vector unsigned char *)&pkts[2]->rearm_data); 778 vec_vsx_st(rearm3, 0, 779 (__vector unsigned char *)&pkts[3]->rearm_data); 780 } 781 782 /** 783 * Process a non-compressed completion and fill in mbufs in RX SW ring 784 * with data extracted from the title completion descriptor. 785 * 786 * @param rxq 787 * Pointer to RX queue structure. 788 * @param cq 789 * Pointer to completion array having a non-compressed completion at first. 790 * @param elts 791 * Pointer to SW ring to be filled. The first mbuf has to be pre-built from 792 * the title completion descriptor to be copied to the rest of mbufs. 793 * @param[out] pkts 794 * Array to store received packets. 795 * @param pkts_n 796 * Maximum number of packets in array. 797 * @param[out] err 798 * Pointer to a flag. Set non-zero value if pkts array has at least one error 799 * packet to handle. 800 * @param[out] comp 801 * Pointer to a index. Set it to the first compressed completion if any. 802 * 803 * @return 804 * Number of CQEs successfully processed. 805 */ 806 static inline uint16_t 807 rxq_cq_process_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq, 808 struct rte_mbuf **elts, struct rte_mbuf **pkts, 809 uint16_t pkts_n, uint64_t *err, uint64_t *comp) 810 { 811 const uint16_t q_n = 1 << rxq->cqe_n; 812 const uint16_t q_mask = q_n - 1; 813 unsigned int pos, adj; 814 uint64_t n = 0; 815 uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP; 816 uint16_t nocmp_n = 0; 817 const uint8_t vic = rxq->cq_ci >> rxq->cqe_n; 818 unsigned int own = !(rxq->cq_ci & (q_mask + 1)); 819 const __vector unsigned char zero = (__vector unsigned char){0}; 820 const __vector unsigned char ones = vec_splat_u8(-1); 821 const __vector unsigned char vic_check = 822 (__vector unsigned char)(__vector unsigned long){ 823 0x00ff000000ff0000LL, 0x00ff000000ff0000LL}; 824 const __vector unsigned char owner_check = 825 (__vector unsigned char)(__vector unsigned long){ 826 0x0100000001000000LL, 0x0100000001000000LL}; 827 const __vector unsigned char opcode_check = 828 (__vector unsigned char)(__vector unsigned long){ 829 0xf0000000f0000000LL, 0xf0000000f0000000LL}; 830 const __vector unsigned char format_check = 831 (__vector unsigned char)(__vector unsigned long){ 832 0x0c0000000c000000LL, 0x0c0000000c000000LL}; 833 const __vector unsigned char resp_err_check = 834 (__vector unsigned char)(__vector unsigned long){ 835 0xe0000000e0000000LL, 0xe0000000e0000000LL}; 836 #ifdef MLX5_PMD_SOFT_COUNTERS 837 uint32_t rcvd_byte = 0; 838 /* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */ 839 const __vector unsigned char len_shuf_mask = (__vector unsigned char){ 840 1, 0, 5, 4, 841 9, 8, 13, 12, 842 -1, -1, -1, -1, 843 -1, -1, -1, -1}; 844 #endif 845 /* Mask to shuffle from extracted CQE to mbuf. */ 846 const __vector unsigned char shuf_mask = (__vector unsigned char){ 847 5, 4, /* bswap16, pkt_len */ 848 -1, -1, /* zero out 2nd half of pkt_len */ 849 5, 4, /* bswap16, data_len */ 850 11, 10, /* bswap16, vlan+tci */ 851 15, 14, 13, 12, /* bswap32, rss */ 852 1, 2, 3, -1}; /* fdir.hi */ 853 /* Mask to blend from the last Qword to the first DQword. */ 854 /* Mask to blend from the last Qword to the first DQword. */ 855 const __vector unsigned char blend_mask = (__vector unsigned char){ 856 -1, 0, 0, 0, 857 0, 0, 0, 0, 858 -1, -1, -1, -1, 859 -1, -1, -1, -1}; 860 const __vector unsigned char crc_adj = 861 (__vector unsigned char)(__vector unsigned short){ 862 rxq->crc_present * RTE_ETHER_CRC_LEN, 0, 863 rxq->crc_present * RTE_ETHER_CRC_LEN, 0, 0, 0, 0, 0}; 864 const __vector unsigned char flow_mark_adj = 865 (__vector unsigned char)(__vector unsigned int){ 866 0, 0, 0, rxq->mark * (-1)}; 867 const __vector unsigned short cqe_sel_mask1 = 868 (__vector unsigned short){0, 0, 0, 0, 0xffff, 0xffff, 0, 0}; 869 const __vector unsigned short cqe_sel_mask2 = 870 (__vector unsigned short){0, 0, 0xffff, 0, 0, 0, 0, 0}; 871 const __vector unsigned char validity = (__vector unsigned char){ 872 0, 0, vic, 0, 873 0, 0, vic, 0, 874 0, 0, vic, 0, 875 0, 0, vic, 0}; 876 const __vector unsigned char ownership = (__vector unsigned char){ 877 0, 0, 0, own, 878 0, 0, 0, own, 879 0, 0, 0, own, 880 0, 0, 0, own}; 881 /* 882 * A. load first Qword (8bytes) in one loop. 883 * B. copy 4 mbuf pointers from elts ring to returning pkts. 884 * C. load remaining CQE data and extract necessary fields. 885 * Final 16bytes cqes[] extracted from original 64bytes CQE has the 886 * following structure: 887 * struct { 888 * uint8_t pkt_info; 889 * uint8_t flow_tag[3]; 890 * uint16_t byte_cnt; 891 * uint8_t validity_iteration_count; 892 * uint8_t op_own; 893 * uint16_t hdr_type_etc; 894 * uint16_t vlan_info; 895 * uint32_t rx_has_res; 896 * } c; 897 * D. fill in mbuf. 898 * E. get valid CQEs. 899 * F. find compressed CQE. 900 */ 901 for (pos = 0; 902 pos < pkts_n; 903 pos += MLX5_VPMD_DESCS_PER_LOOP) { 904 __vector unsigned char cqes[MLX5_VPMD_DESCS_PER_LOOP]; 905 __vector unsigned char cqe_tmp1, cqe_tmp2; 906 __vector unsigned char pkt_mb0, pkt_mb1, pkt_mb2, pkt_mb3; 907 __vector unsigned char op_own, op_own_tmp1, op_own_tmp2; 908 __vector unsigned char opcode, owner_mask, invalid_mask; 909 __vector unsigned char comp_mask, mini_mask; 910 __vector unsigned char mask; 911 #ifdef MLX5_PMD_SOFT_COUNTERS 912 const __vector unsigned char lower_half = { 913 0, 1, 4, 5, 8, 9, 12, 13, 914 16, 17, 20, 21, 24, 25, 28, 29}; 915 const __vector unsigned char upper_half = { 916 2, 3, 6, 7, 10, 11, 14, 15, 917 18, 19, 22, 23, 26, 27, 30, 31}; 918 const __vector unsigned long shmax = {64, 64}; 919 __vector unsigned char byte_cnt; 920 __vector unsigned short left, right; 921 __vector unsigned long lshift; 922 __vector __attribute__((altivec(bool__))) 923 unsigned long shmask; 924 #endif 925 __vector unsigned char mbp1, mbp2; 926 __vector unsigned char p = 927 (__vector unsigned char)(__vector unsigned short){ 928 0, 1, 2, 3, 0, 0, 0, 0}; 929 unsigned int p1, p2, p3; 930 931 /* Prefetch next 4 CQEs. */ 932 if (pkts_n - pos >= 2 * MLX5_VPMD_DESCS_PER_LOOP) { 933 rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP]); 934 rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 1]); 935 rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 2]); 936 rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 3]); 937 } 938 939 /* A.0 do not cross the end of CQ. */ 940 mask = (__vector unsigned char)(__vector unsigned long){ 941 (pkts_n - pos) * sizeof(uint16_t) * 8, 0}; 942 943 { 944 __vector unsigned long lshift; 945 __vector __attribute__((altivec(bool__))) 946 unsigned long shmask; 947 const __vector unsigned long shmax = {64, 64}; 948 949 lshift = vec_splat((__vector unsigned long)mask, 0); 950 shmask = vec_cmpgt(shmax, lshift); 951 mask = (__vector unsigned char) 952 vec_sl((__vector unsigned long)ones, lshift); 953 mask = (__vector unsigned char) 954 vec_sel((__vector unsigned long)shmask, 955 (__vector unsigned long)mask, shmask); 956 } 957 958 p = (__vector unsigned char) 959 vec_andc((__vector unsigned long)p, 960 (__vector unsigned long)mask); 961 962 /* A.1 load cqes. */ 963 p3 = (unsigned int)((__vector unsigned short)p)[3]; 964 cqes[3] = (__vector unsigned char)(__vector unsigned long){ 965 *(__rte_aligned(8) unsigned long *) 966 &cq[pos + p3].sop_drop_qpn, 0LL}; 967 rte_compiler_barrier(); 968 969 p2 = (unsigned int)((__vector unsigned short)p)[2]; 970 cqes[2] = (__vector unsigned char)(__vector unsigned long){ 971 *(__rte_aligned(8) unsigned long *) 972 &cq[pos + p2].sop_drop_qpn, 0LL}; 973 rte_compiler_barrier(); 974 975 /* B.1 load mbuf pointers. */ 976 mbp1 = (__vector unsigned char)vec_vsx_ld(0, 977 (signed int const *)&elts[pos]); 978 mbp2 = (__vector unsigned char)vec_vsx_ld(0, 979 (signed int const *)&elts[pos + 2]); 980 981 /* A.1 load a block having op_own. */ 982 p1 = (unsigned int)((__vector unsigned short)p)[1]; 983 cqes[1] = (__vector unsigned char)(__vector unsigned long){ 984 *(__rte_aligned(8) unsigned long *) 985 &cq[pos + p1].sop_drop_qpn, 0LL}; 986 rte_compiler_barrier(); 987 988 cqes[0] = (__vector unsigned char)(__vector unsigned long){ 989 *(__rte_aligned(8) unsigned long *) 990 &cq[pos].sop_drop_qpn, 0LL}; 991 rte_compiler_barrier(); 992 993 /* B.2 copy mbuf pointers. */ 994 *(__vector unsigned char *)&pkts[pos] = mbp1; 995 *(__vector unsigned char *)&pkts[pos + 2] = mbp2; 996 rte_io_rmb(); 997 998 /* C.1 load remaining CQE data and extract necessary fields. */ 999 cqe_tmp2 = *(__vector unsigned char *) 1000 &cq[pos + p3].pkt_info; 1001 cqe_tmp1 = *(__vector unsigned char *) 1002 &cq[pos + p2].pkt_info; 1003 cqes[3] = vec_sel(cqes[3], cqe_tmp2, blend_mask); 1004 cqes[2] = vec_sel(cqes[2], cqe_tmp1, blend_mask); 1005 cqe_tmp2 = (__vector unsigned char)vec_vsx_ld(0, 1006 (signed int const *)&cq[pos + p3].csum); 1007 cqe_tmp1 = (__vector unsigned char)vec_vsx_ld(0, 1008 (signed int const *)&cq[pos + p2].csum); 1009 cqes[3] = (__vector unsigned char) 1010 vec_sel((__vector unsigned short)cqes[3], 1011 (__vector unsigned short)cqe_tmp2, cqe_sel_mask1); 1012 cqes[2] = (__vector unsigned char) 1013 vec_sel((__vector unsigned short)cqes[2], 1014 (__vector unsigned short)cqe_tmp1, cqe_sel_mask1); 1015 cqe_tmp2 = (__vector unsigned char)(__vector unsigned long){ 1016 *(__rte_aligned(8) unsigned long *) 1017 &cq[pos + p3].rsvd4[2], 0LL}; 1018 cqe_tmp1 = (__vector unsigned char)(__vector unsigned long){ 1019 *(__rte_aligned(8) unsigned long *) 1020 &cq[pos + p2].rsvd4[2], 0LL}; 1021 cqes[3] = (__vector unsigned char) 1022 vec_sel((__vector unsigned short)cqes[3], 1023 (__vector unsigned short)cqe_tmp2, 1024 (__vector unsigned short)cqe_sel_mask2); 1025 cqes[2] = (__vector unsigned char) 1026 vec_sel((__vector unsigned short)cqes[2], 1027 (__vector unsigned short)cqe_tmp1, 1028 (__vector unsigned short)cqe_sel_mask2); 1029 1030 /* C.2 generate final structure for mbuf with swapping bytes. */ 1031 pkt_mb3 = vec_perm(cqes[3], zero, shuf_mask); 1032 pkt_mb2 = vec_perm(cqes[2], zero, shuf_mask); 1033 1034 /* C.3 adjust CRC length. */ 1035 pkt_mb3 = (__vector unsigned char) 1036 ((__vector unsigned short)pkt_mb3 - 1037 (__vector unsigned short)crc_adj); 1038 pkt_mb2 = (__vector unsigned char) 1039 ((__vector unsigned short)pkt_mb2 - 1040 (__vector unsigned short)crc_adj); 1041 1042 /* C.4 adjust flow mark. */ 1043 pkt_mb3 = (__vector unsigned char) 1044 ((__vector unsigned int)pkt_mb3 + 1045 (__vector unsigned int)flow_mark_adj); 1046 pkt_mb2 = (__vector unsigned char) 1047 ((__vector unsigned int)pkt_mb2 + 1048 (__vector unsigned int)flow_mark_adj); 1049 1050 /* D.1 fill in mbuf - rx_descriptor_fields1. */ 1051 *(__vector unsigned char *) 1052 &pkts[pos + 3]->pkt_len = pkt_mb3; 1053 *(__vector unsigned char *) 1054 &pkts[pos + 2]->pkt_len = pkt_mb2; 1055 1056 /* E.1 extract op_own field. */ 1057 op_own_tmp2 = (__vector unsigned char) 1058 vec_mergeh((__vector unsigned int)cqes[2], 1059 (__vector unsigned int)cqes[3]); 1060 1061 /* C.1 load remaining CQE data and extract necessary fields. */ 1062 cqe_tmp2 = *(__vector unsigned char *) 1063 &cq[pos + p1].pkt_info; 1064 cqe_tmp1 = *(__vector unsigned char *) 1065 &cq[pos].pkt_info; 1066 cqes[1] = vec_sel(cqes[1], cqe_tmp2, blend_mask); 1067 cqes[0] = vec_sel(cqes[0], cqe_tmp2, blend_mask); 1068 cqe_tmp2 = (__vector unsigned char)vec_vsx_ld(0, 1069 (signed int const *)&cq[pos + p1].csum); 1070 cqe_tmp1 = (__vector unsigned char)vec_vsx_ld(0, 1071 (signed int const *)&cq[pos].csum); 1072 cqes[1] = (__vector unsigned char) 1073 vec_sel((__vector unsigned short)cqes[1], 1074 (__vector unsigned short)cqe_tmp2, cqe_sel_mask1); 1075 cqes[0] = (__vector unsigned char) 1076 vec_sel((__vector unsigned short)cqes[0], 1077 (__vector unsigned short)cqe_tmp1, cqe_sel_mask1); 1078 cqe_tmp2 = (__vector unsigned char)(__vector unsigned long){ 1079 *(__rte_aligned(8) unsigned long *) 1080 &cq[pos + p1].rsvd4[2], 0LL}; 1081 cqe_tmp1 = (__vector unsigned char)(__vector unsigned long){ 1082 *(__rte_aligned(8) unsigned long *) 1083 &cq[pos].rsvd4[2], 0LL}; 1084 cqes[1] = (__vector unsigned char) 1085 vec_sel((__vector unsigned short)cqes[1], 1086 (__vector unsigned short)cqe_tmp2, cqe_sel_mask2); 1087 cqes[0] = (__vector unsigned char) 1088 vec_sel((__vector unsigned short)cqes[0], 1089 (__vector unsigned short)cqe_tmp1, cqe_sel_mask2); 1090 1091 /* C.2 generate final structure for mbuf with swapping bytes. */ 1092 pkt_mb1 = vec_perm(cqes[1], zero, shuf_mask); 1093 pkt_mb0 = vec_perm(cqes[0], zero, shuf_mask); 1094 1095 /* C.3 adjust CRC length. */ 1096 pkt_mb1 = (__vector unsigned char) 1097 ((__vector unsigned short)pkt_mb1 - 1098 (__vector unsigned short)crc_adj); 1099 pkt_mb0 = (__vector unsigned char) 1100 ((__vector unsigned short)pkt_mb0 - 1101 (__vector unsigned short)crc_adj); 1102 1103 /* C.4 adjust flow mark. */ 1104 pkt_mb1 = (__vector unsigned char) 1105 ((__vector unsigned int)pkt_mb1 + 1106 (__vector unsigned int)flow_mark_adj); 1107 pkt_mb0 = (__vector unsigned char) 1108 ((__vector unsigned int)pkt_mb0 + 1109 (__vector unsigned int)flow_mark_adj); 1110 1111 /* E.1 extract op_own byte. */ 1112 op_own_tmp1 = (__vector unsigned char) 1113 vec_mergeh((__vector unsigned int)cqes[0], 1114 (__vector unsigned int)cqes[1]); 1115 op_own = (__vector unsigned char) 1116 vec_mergel((__vector unsigned long)op_own_tmp1, 1117 (__vector unsigned long)op_own_tmp2); 1118 1119 /* D.1 fill in mbuf - rx_descriptor_fields1. */ 1120 *(__vector unsigned char *) 1121 &pkts[pos + 1]->pkt_len = pkt_mb1; 1122 *(__vector unsigned char *) 1123 &pkts[pos]->pkt_len = pkt_mb0; 1124 1125 /* E.2 mask out CQEs belonging to HW. */ 1126 if (rxq->cqe_comp_layout) { 1127 owner_mask = (__vector unsigned char) 1128 vec_and((__vector unsigned long)op_own, 1129 (__vector unsigned long)vic_check); 1130 owner_mask = (__vector unsigned char) 1131 vec_cmpeq((__vector unsigned int)owner_mask, 1132 (__vector unsigned int)validity); 1133 owner_mask = (__vector unsigned char) 1134 vec_xor((__vector unsigned long)owner_mask, 1135 (__vector unsigned long)ones); 1136 } else { 1137 owner_mask = (__vector unsigned char) 1138 vec_and((__vector unsigned long)op_own, 1139 (__vector unsigned long)owner_check); 1140 owner_mask = (__vector unsigned char) 1141 vec_cmpeq((__vector unsigned int)owner_mask, 1142 (__vector unsigned int)ownership); 1143 } 1144 owner_mask = (__vector unsigned char) 1145 vec_packs((__vector unsigned int)owner_mask, 1146 (__vector unsigned int)zero); 1147 1148 /* E.3 get mask for invalidated CQEs. */ 1149 opcode = (__vector unsigned char) 1150 vec_and((__vector unsigned long)op_own, 1151 (__vector unsigned long)opcode_check); 1152 invalid_mask = (__vector unsigned char) 1153 vec_cmpeq((__vector unsigned int)opcode_check, 1154 (__vector unsigned int)opcode); 1155 invalid_mask = (__vector unsigned char) 1156 vec_packs((__vector unsigned int)invalid_mask, 1157 (__vector unsigned int)zero); 1158 1159 /* E.4 mask out beyond boundary. */ 1160 invalid_mask = (__vector unsigned char) 1161 vec_or((__vector unsigned long)invalid_mask, 1162 (__vector unsigned long)mask); 1163 1164 /* E.5 merge invalid_mask with invalid owner. */ 1165 invalid_mask = (__vector unsigned char) 1166 vec_or((__vector unsigned long)invalid_mask, 1167 (__vector unsigned long)owner_mask); 1168 1169 /* F.1 find compressed CQE format. */ 1170 comp_mask = (__vector unsigned char) 1171 vec_and((__vector unsigned long)op_own, 1172 (__vector unsigned long)format_check); 1173 comp_mask = (__vector unsigned char) 1174 vec_cmpeq((__vector unsigned int)comp_mask, 1175 (__vector unsigned int)format_check); 1176 comp_mask = (__vector unsigned char) 1177 vec_packs((__vector unsigned int)comp_mask, 1178 (__vector unsigned int)zero); 1179 1180 /* F.2 mask out invalid entries. */ 1181 comp_mask = (__vector unsigned char) 1182 vec_andc((__vector unsigned long)comp_mask, 1183 (__vector unsigned long)invalid_mask); 1184 comp_idx = ((__vector unsigned long)comp_mask)[0]; 1185 1186 /* F.3 get the first compressed CQE. */ 1187 comp_idx = comp_idx ? rte_ctz64(comp_idx) / 1188 (sizeof(uint16_t) * 8) : MLX5_VPMD_DESCS_PER_LOOP; 1189 1190 /* E.6 mask out entries after the compressed CQE. */ 1191 mask = (__vector unsigned char)(__vector unsigned long){ 1192 (comp_idx * sizeof(uint16_t) * 8), 0}; 1193 lshift = vec_splat((__vector unsigned long)mask, 0); 1194 shmask = vec_cmpgt(shmax, lshift); 1195 mask = (__vector unsigned char) 1196 vec_sl((__vector unsigned long)ones, lshift); 1197 mask = (__vector unsigned char) 1198 vec_sel((__vector unsigned long)shmask, 1199 (__vector unsigned long)mask, shmask); 1200 invalid_mask = (__vector unsigned char) 1201 vec_or((__vector unsigned long)invalid_mask, 1202 (__vector unsigned long)mask); 1203 1204 /* E.7 count non-compressed valid CQEs. */ 1205 n = ((__vector unsigned long)invalid_mask)[0]; 1206 n = n ? rte_ctz64(n) / (sizeof(uint16_t) * 8) : 1207 MLX5_VPMD_DESCS_PER_LOOP; 1208 nocmp_n += n; 1209 1210 /* D.2 get the final invalid mask. */ 1211 mask = (__vector unsigned char)(__vector unsigned long){ 1212 (n * sizeof(uint16_t) * 8), 0}; 1213 lshift = vec_splat((__vector unsigned long)mask, 0); 1214 shmask = vec_cmpgt(shmax, lshift); 1215 mask = (__vector unsigned char) 1216 vec_sl((__vector unsigned long)ones, lshift); 1217 mask = (__vector unsigned char) 1218 vec_sel((__vector unsigned long)shmask, 1219 (__vector unsigned long)mask, shmask); 1220 invalid_mask = (__vector unsigned char) 1221 vec_or((__vector unsigned long)invalid_mask, 1222 (__vector unsigned long)mask); 1223 1224 /* D.3 check error in opcode. */ 1225 adj = (!rxq->cqe_comp_layout && 1226 comp_idx != MLX5_VPMD_DESCS_PER_LOOP && comp_idx == n); 1227 mask = (__vector unsigned char)(__vector unsigned long){ 1228 (adj * sizeof(uint16_t) * 8), 0}; 1229 lshift = vec_splat((__vector unsigned long)mask, 0); 1230 shmask = vec_cmpgt(shmax, lshift); 1231 mini_mask = (__vector unsigned char) 1232 vec_sl((__vector unsigned long)invalid_mask, lshift); 1233 mini_mask = (__vector unsigned char) 1234 vec_sel((__vector unsigned long)shmask, 1235 (__vector unsigned long)mini_mask, shmask); 1236 opcode = (__vector unsigned char) 1237 vec_cmpeq((__vector unsigned int)resp_err_check, 1238 (__vector unsigned int)opcode); 1239 opcode = (__vector unsigned char) 1240 vec_packs((__vector unsigned int)opcode, 1241 (__vector unsigned int)zero); 1242 opcode = (__vector unsigned char) 1243 vec_andc((__vector unsigned long)opcode, 1244 (__vector unsigned long)mini_mask); 1245 1246 /* D.4 mark if any error is set */ 1247 *err |= ((__vector unsigned long)opcode)[0]; 1248 1249 /* D.5 fill in mbuf - rearm_data and packet_type. */ 1250 rxq_cq_to_ptype_oflags_v(rxq, cqes, opcode, &pkts[pos]); 1251 if (unlikely(rxq->shared)) { 1252 pkts[pos]->port = cq[pos].user_index_low; 1253 pkts[pos + 1]->port = cq[pos + p1].user_index_low; 1254 pkts[pos + 2]->port = cq[pos + p2].user_index_low; 1255 pkts[pos + 3]->port = cq[pos + p3].user_index_low; 1256 } 1257 if (rxq->hw_timestamp) { 1258 int offset = rxq->timestamp_offset; 1259 if (rxq->rt_timestamp) { 1260 struct mlx5_dev_ctx_shared *sh = rxq->sh; 1261 uint64_t ts; 1262 1263 ts = rte_be_to_cpu_64(cq[pos].timestamp); 1264 mlx5_timestamp_set(pkts[pos], offset, 1265 mlx5_txpp_convert_rx_ts(sh, ts)); 1266 ts = rte_be_to_cpu_64(cq[pos + p1].timestamp); 1267 mlx5_timestamp_set(pkts[pos + 1], offset, 1268 mlx5_txpp_convert_rx_ts(sh, ts)); 1269 ts = rte_be_to_cpu_64(cq[pos + p2].timestamp); 1270 mlx5_timestamp_set(pkts[pos + 2], offset, 1271 mlx5_txpp_convert_rx_ts(sh, ts)); 1272 ts = rte_be_to_cpu_64(cq[pos + p3].timestamp); 1273 mlx5_timestamp_set(pkts[pos + 3], offset, 1274 mlx5_txpp_convert_rx_ts(sh, ts)); 1275 } else { 1276 mlx5_timestamp_set(pkts[pos], offset, 1277 rte_be_to_cpu_64(cq[pos].timestamp)); 1278 mlx5_timestamp_set(pkts[pos + 1], offset, 1279 rte_be_to_cpu_64(cq[pos + p1].timestamp)); 1280 mlx5_timestamp_set(pkts[pos + 2], offset, 1281 rte_be_to_cpu_64(cq[pos + p2].timestamp)); 1282 mlx5_timestamp_set(pkts[pos + 3], offset, 1283 rte_be_to_cpu_64(cq[pos + p3].timestamp)); 1284 } 1285 } 1286 if (rxq->dynf_meta) { 1287 uint64_t flag = rxq->flow_meta_mask; 1288 int32_t offs = rxq->flow_meta_offset; 1289 uint32_t mask = rxq->flow_meta_port_mask; 1290 uint32_t metadata; 1291 1292 /* This code is subject for further optimization. */ 1293 metadata = rte_be_to_cpu_32 1294 (cq[pos].flow_table_metadata) & mask; 1295 *RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *) = 1296 metadata; 1297 pkts[pos]->ol_flags |= metadata ? flag : 0ULL; 1298 metadata = rte_be_to_cpu_32 1299 (cq[pos + p1].flow_table_metadata) & mask; 1300 *RTE_MBUF_DYNFIELD(pkts[pos + 1], offs, uint32_t *) = 1301 metadata; 1302 pkts[pos + 1]->ol_flags |= metadata ? flag : 0ULL; 1303 metadata = rte_be_to_cpu_32 1304 (cq[pos + p2].flow_table_metadata) & mask; 1305 *RTE_MBUF_DYNFIELD(pkts[pos + 2], offs, uint32_t *) = 1306 metadata; 1307 pkts[pos + 2]->ol_flags |= metadata ? flag : 0ULL; 1308 metadata = rte_be_to_cpu_32 1309 (cq[pos + p3].flow_table_metadata) & mask; 1310 *RTE_MBUF_DYNFIELD(pkts[pos + 3], offs, uint32_t *) = 1311 metadata; 1312 pkts[pos + 3]->ol_flags |= metadata ? flag : 0ULL; 1313 } 1314 #ifdef MLX5_PMD_SOFT_COUNTERS 1315 /* Add up received bytes count. */ 1316 byte_cnt = vec_perm(op_own, zero, len_shuf_mask); 1317 byte_cnt = (__vector unsigned char) 1318 vec_andc((__vector unsigned long)byte_cnt, 1319 (__vector unsigned long)invalid_mask); 1320 left = vec_perm((__vector unsigned short)byte_cnt, 1321 (__vector unsigned short)zero, lower_half); 1322 right = vec_perm((__vector unsigned short)byte_cnt, 1323 (__vector unsigned short)zero, upper_half); 1324 byte_cnt = (__vector unsigned char)vec_add(left, right); 1325 left = vec_perm((__vector unsigned short)byte_cnt, 1326 (__vector unsigned short)zero, lower_half); 1327 right = vec_perm((__vector unsigned short)byte_cnt, 1328 (__vector unsigned short)zero, upper_half); 1329 byte_cnt = (__vector unsigned char)vec_add(left, right); 1330 rcvd_byte += ((__vector unsigned long)byte_cnt)[0]; 1331 #endif 1332 1333 /* 1334 * Break the loop unless more valid CQE is expected, or if 1335 * there's a compressed CQE. 1336 */ 1337 if (n != MLX5_VPMD_DESCS_PER_LOOP) 1338 break; 1339 } 1340 #ifdef MLX5_PMD_SOFT_COUNTERS 1341 rxq->stats.ipackets += nocmp_n; 1342 rxq->stats.ibytes += rcvd_byte; 1343 #endif 1344 if (comp_idx == n) 1345 *comp = comp_idx; 1346 return nocmp_n; 1347 } 1348 1349 #endif /* RTE_PMD_MLX5_RXTX_VEC_ALTIVEC_H_ */ 1350