drivers/net/mlx5/mlx5_rxtx_vec_sse.h

   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_SSE_H_
   7 #define RTE_PMD_MLX5_RXTX_VEC_SSE_H_
   8
   9 #include <stdint.h>
  10 #include <string.h>
  11 #include <stdlib.h>
  12 #include <smmintrin.h>
  13
  14 #include <rte_mbuf.h>
  15 #include <rte_mempool.h>
  16 #include <rte_prefetch.h>
  17
  18 #include <mlx5_prm.h>
  19
  20 #include "mlx5_defs.h"
  21 #include "mlx5.h"
  22 #include "mlx5_utils.h"
  23 #include "mlx5_rxtx.h"
  24 #include "mlx5_rxtx_vec.h"
  25 #include "mlx5_autoconf.h"
  26
  27 #ifndef __INTEL_COMPILER
  28 #pragma GCC diagnostic ignored "-Wcast-qual"
  29 #endif
  30
  31 /**
  32  * Store free buffers to RX SW ring.
  33  *
  34  * @param rxq
  35  *   Pointer to RX queue structure.
  36  * @param pkts
  37  *   Pointer to array of packets to be stored.
  38  * @param pkts_n
  39  *   Number of packets to be stored.
  40  */
  41 static inline void
  42 rxq_copy_mbuf_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t n)
  43 {
  44         const uint16_t q_mask = (1 << rxq->elts_n) - 1;
  45         struct rte_mbuf **elts = &(*rxq->elts)[rxq->rq_pi & q_mask];
  46         unsigned int pos;
  47         uint16_t p = n & -2;
  48
  49         for (pos = 0; pos < p; pos += 2) {
  50                 __m128i mbp;
  51
  52                 mbp = _mm_loadu_si128((__m128i *)&elts[pos]);
  53                 _mm_storeu_si128((__m128i *)&pkts[pos], mbp);
  54         }
  55         if (n & 1)
  56                 pkts[pos] = elts[pos];
  57 }
  58
  59 /**
  60  * Decompress a compressed completion and fill in mbufs in RX SW ring with data
  61  * extracted from the title completion descriptor.
  62  *
  63  * @param rxq
  64  *   Pointer to RX queue structure.
  65  * @param cq
  66  *   Pointer to completion array having a compressed completion at first.
  67  * @param elts
  68  *   Pointer to SW ring to be filled. The first mbuf has to be pre-built from
  69  *   the title completion descriptor to be copied to the rest of mbufs.
  70  *
  71  * @return
  72  *   Number of mini-CQEs successfully decompressed.
  73  */
  74 static inline uint16_t
  75 rxq_cq_decompress_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
  76                     struct rte_mbuf **elts)
  77 {
  78         volatile struct mlx5_mini_cqe8 *mcq = (void *)(cq + 1);
  79         struct rte_mbuf *t_pkt = elts[0]; /* Title packet is pre-built. */
  80         unsigned int pos;
  81         unsigned int i;
  82         unsigned int inv = 0;
  83         /* Mask to shuffle from extracted mini CQE to mbuf. */
  84         const __m128i shuf_mask1 =
  85                 _mm_set_epi8(0,  1,  2,  3, /* rss, bswap32 */
  86                             -1, -1,         /* skip vlan_tci */
  87                              6,  7,         /* data_len, bswap16 */
  88                             -1, -1,  6,  7, /* pkt_len, bswap16 */
  89                             -1, -1, -1, -1  /* skip packet_type */);
  90         const __m128i shuf_mask2 =
  91                 _mm_set_epi8(8,  9, 10, 11, /* rss, bswap32 */
  92                             -1, -1,         /* skip vlan_tci */
  93                             14, 15,         /* data_len, bswap16 */
  94                             -1, -1, 14, 15, /* pkt_len, bswap16 */
  95                             -1, -1, -1, -1  /* skip packet_type */);
  96         /* Restore the compressed count. Must be 16 bits. */
  97         const uint16_t mcqe_n = t_pkt->data_len +
  98                                 (rxq->crc_present * RTE_ETHER_CRC_LEN);
  99         const __m128i rearm =
 100                 _mm_loadu_si128((__m128i *)&t_pkt->rearm_data);
 101         const __m128i rxdf =
 102                 _mm_loadu_si128((__m128i *)&t_pkt->rx_descriptor_fields1);
 103         const __m128i crc_adj =
 104                 _mm_set_epi16(0, 0, 0,
 105                               rxq->crc_present * RTE_ETHER_CRC_LEN,
 106                               0,
 107                               rxq->crc_present * RTE_ETHER_CRC_LEN,
 108                               0, 0);
 109         const uint32_t flow_tag = t_pkt->hash.fdir.hi;
 110 #ifdef MLX5_PMD_SOFT_COUNTERS
 111         const __m128i zero = _mm_setzero_si128();
 112         const __m128i ones = _mm_cmpeq_epi32(zero, zero);
 113         uint32_t rcvd_byte = 0;
 114         /* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
 115         const __m128i len_shuf_mask =
 116                 _mm_set_epi8(-1, -1, -1, -1,
 117                              -1, -1, -1, -1,
 118                              14, 15,  6,  7,
 119                              10, 11,  2,  3);
 120 #endif
 121         /*
 122          * A. load mCQEs into a 128bit register.
 123          * B. store rearm data to mbuf.
 124          * C. combine data from mCQEs with rx_descriptor_fields1.
 125          * D. store rx_descriptor_fields1.
 126          * E. store flow tag (rte_flow mark).
 127          */
 128         for (pos = 0; pos < mcqe_n; ) {
 129                 __m128i mcqe1, mcqe2;
 130                 __m128i rxdf1, rxdf2;
 131 #ifdef MLX5_PMD_SOFT_COUNTERS
 132                 __m128i byte_cnt, invalid_mask;
 133 #endif
 134
 135                 for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
 136                         if (likely(pos + i < mcqe_n))
 137                                 rte_prefetch0((void *)(cq + pos + i));
 138
 139                 /* A.1 load mCQEs into a 128bit register. */
 140                 mcqe1 = _mm_loadu_si128((__m128i *)&mcq[pos % 8]);
 141                 mcqe2 = _mm_loadu_si128((__m128i *)&mcq[pos % 8 + 2]);
 142                 /* B.1 store rearm data to mbuf. */
 143                 _mm_storeu_si128((__m128i *)&elts[pos]->rearm_data, rearm);
 144                 _mm_storeu_si128((__m128i *)&elts[pos + 1]->rearm_data, rearm);
 145                 /* C.1 combine data from mCQEs with rx_descriptor_fields1. */
 146                 rxdf1 = _mm_shuffle_epi8(mcqe1, shuf_mask1);
 147                 rxdf2 = _mm_shuffle_epi8(mcqe1, shuf_mask2);
 148                 rxdf1 = _mm_sub_epi16(rxdf1, crc_adj);
 149                 rxdf2 = _mm_sub_epi16(rxdf2, crc_adj);
 150                 rxdf1 = _mm_blend_epi16(rxdf1, rxdf, 0x23);
 151                 rxdf2 = _mm_blend_epi16(rxdf2, rxdf, 0x23);
 152                 /* D.1 store rx_descriptor_fields1. */
 153                 _mm_storeu_si128((__m128i *)
 154                                   &elts[pos]->rx_descriptor_fields1,
 155                                  rxdf1);
 156                 _mm_storeu_si128((__m128i *)
 157                                   &elts[pos + 1]->rx_descriptor_fields1,
 158                                  rxdf2);
 159                 /* B.1 store rearm data to mbuf. */
 160                 _mm_storeu_si128((__m128i *)&elts[pos + 2]->rearm_data, rearm);
 161                 _mm_storeu_si128((__m128i *)&elts[pos + 3]->rearm_data, rearm);
 162                 /* C.1 combine data from mCQEs with rx_descriptor_fields1. */
 163                 rxdf1 = _mm_shuffle_epi8(mcqe2, shuf_mask1);
 164                 rxdf2 = _mm_shuffle_epi8(mcqe2, shuf_mask2);
 165                 rxdf1 = _mm_sub_epi16(rxdf1, crc_adj);
 166                 rxdf2 = _mm_sub_epi16(rxdf2, crc_adj);
 167                 rxdf1 = _mm_blend_epi16(rxdf1, rxdf, 0x23);
 168                 rxdf2 = _mm_blend_epi16(rxdf2, rxdf, 0x23);
 169                 /* D.1 store rx_descriptor_fields1. */
 170                 _mm_storeu_si128((__m128i *)
 171                                   &elts[pos + 2]->rx_descriptor_fields1,
 172                                  rxdf1);
 173                 _mm_storeu_si128((__m128i *)
 174                                   &elts[pos + 3]->rx_descriptor_fields1,
 175                                  rxdf2);
 176 #ifdef MLX5_PMD_SOFT_COUNTERS
 177                 invalid_mask = _mm_set_epi64x(0,
 178                                               (mcqe_n - pos) *
 179                                               sizeof(uint16_t) * 8);
 180                 invalid_mask = _mm_sll_epi64(ones, invalid_mask);
 181                 mcqe1 = _mm_srli_si128(mcqe1, 4);
 182                 byte_cnt = _mm_blend_epi16(mcqe1, mcqe2, 0xcc);
 183                 byte_cnt = _mm_shuffle_epi8(byte_cnt, len_shuf_mask);
 184                 byte_cnt = _mm_andnot_si128(invalid_mask, byte_cnt);
 185                 byte_cnt = _mm_hadd_epi16(byte_cnt, zero);
 186                 rcvd_byte += _mm_cvtsi128_si64(_mm_hadd_epi16(byte_cnt, zero));
 187 #endif
 188                 if (rxq->mark) {
 189                         /* E.1 store flow tag (rte_flow mark). */
 190                         elts[pos]->hash.fdir.hi = flow_tag;
 191                         elts[pos + 1]->hash.fdir.hi = flow_tag;
 192                         elts[pos + 2]->hash.fdir.hi = flow_tag;
 193                         elts[pos + 3]->hash.fdir.hi = flow_tag;
 194                 }
 195                 if (rxq->dynf_meta) {
 196                         int32_t offs = rxq->flow_meta_offset;
 197                         const uint32_t meta =
 198                                 *RTE_MBUF_DYNFIELD(t_pkt, offs, uint32_t *);
 199
 200                         /* Check if title packet has valid metadata. */
 201                         if (meta) {
 202                                 MLX5_ASSERT(t_pkt->ol_flags & offs);
 203                                 *RTE_MBUF_DYNFIELD(elts[pos], offs,
 204                                                         uint32_t *) = meta;
 205                                 *RTE_MBUF_DYNFIELD(elts[pos + 1], offs,
 206                                                         uint32_t *) = meta;
 207                                 *RTE_MBUF_DYNFIELD(elts[pos + 2], offs,
 208                                                         uint32_t *) = meta;
 209                                 *RTE_MBUF_DYNFIELD(elts[pos + 3], offs,
 210                                                         uint32_t *) = meta;
 211                         }
 212                 }
 213                 pos += MLX5_VPMD_DESCS_PER_LOOP;
 214                 /* Move to next CQE and invalidate consumed CQEs. */
 215                 if (!(pos & 0x7) && pos < mcqe_n) {
 216                         mcq = (void *)(cq + pos);
 217                         for (i = 0; i < 8; ++i)
 218                                 cq[inv++].op_own = MLX5_CQE_INVALIDATE;
 219                 }
 220         }
 221         /* Invalidate the rest of CQEs. */
 222         for (; inv < mcqe_n; ++inv)
 223                 cq[inv].op_own = MLX5_CQE_INVALIDATE;
 224 #ifdef MLX5_PMD_SOFT_COUNTERS
 225         rxq->stats.ipackets += mcqe_n;
 226         rxq->stats.ibytes += rcvd_byte;
 227 #endif
 228         rxq->cq_ci += mcqe_n;
 229         return mcqe_n;
 230 }
 231
 232 /**
 233  * Calculate packet type and offload flag for mbuf and store it.
 234  *
 235  * @param rxq
 236  *   Pointer to RX queue structure.
 237  * @param cqes[4]
 238  *   Array of four 16bytes completions extracted from the original completion
 239  *   descriptor.
 240  * @param op_err
 241  *   Opcode vector having responder error status. Each field is 4B.
 242  * @param pkts
 243  *   Pointer to array of packets to be filled.
 244  */
 245 static inline void
 246 rxq_cq_to_ptype_oflags_v(struct mlx5_rxq_data *rxq, __m128i cqes[4],
 247                          __m128i op_err, struct rte_mbuf **pkts)
 248 {
 249         __m128i pinfo0, pinfo1;
 250         __m128i pinfo, ptype;
 251         __m128i ol_flags = _mm_set1_epi32(rxq->rss_hash * PKT_RX_RSS_HASH |
 252                                           rxq->hw_timestamp * PKT_RX_TIMESTAMP);
 253         __m128i cv_flags;
 254         const __m128i zero = _mm_setzero_si128();
 255         const __m128i ptype_mask =
 256                 _mm_set_epi32(0xfd06, 0xfd06, 0xfd06, 0xfd06);
 257         const __m128i ptype_ol_mask =
 258                 _mm_set_epi32(0x106, 0x106, 0x106, 0x106);
 259         const __m128i pinfo_mask =
 260                 _mm_set_epi32(0x3, 0x3, 0x3, 0x3);
 261         const __m128i cv_flag_sel =
 262                 _mm_set_epi8(0, 0, 0, 0, 0, 0, 0, 0, 0,
 263                              (uint8_t)((PKT_RX_IP_CKSUM_GOOD |
 264                                         PKT_RX_L4_CKSUM_GOOD) >> 1),
 265                              0,
 266                              (uint8_t)(PKT_RX_L4_CKSUM_GOOD >> 1),
 267                              0,
 268                              (uint8_t)(PKT_RX_IP_CKSUM_GOOD >> 1),
 269                              (uint8_t)(PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED),
 270                              0);
 271         const __m128i cv_mask =
 272                 _mm_set_epi32(PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
 273                               PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
 274                               PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
 275                               PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
 276                               PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
 277                               PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
 278                               PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
 279                               PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED);
 280         const __m128i mbuf_init =
 281                 _mm_load_si128((__m128i *)&rxq->mbuf_initializer);
 282         __m128i rearm0, rearm1, rearm2, rearm3;
 283         uint8_t pt_idx0, pt_idx1, pt_idx2, pt_idx3;
 284
 285         /* Extract pkt_info field. */
 286         pinfo0 = _mm_unpacklo_epi32(cqes[0], cqes[1]);
 287         pinfo1 = _mm_unpacklo_epi32(cqes[2], cqes[3]);
 288         pinfo = _mm_unpacklo_epi64(pinfo0, pinfo1);
 289         /* Extract hdr_type_etc field. */
 290         pinfo0 = _mm_unpackhi_epi32(cqes[0], cqes[1]);
 291         pinfo1 = _mm_unpackhi_epi32(cqes[2], cqes[3]);
 292         ptype = _mm_unpacklo_epi64(pinfo0, pinfo1);
 293         if (rxq->mark) {
 294                 const __m128i pinfo_ft_mask =
 295                         _mm_set_epi32(0xffffff00, 0xffffff00,
 296                                       0xffffff00, 0xffffff00);
 297                 const __m128i fdir_flags = _mm_set1_epi32(PKT_RX_FDIR);
 298                 __m128i fdir_id_flags = _mm_set1_epi32(PKT_RX_FDIR_ID);
 299                 __m128i flow_tag, invalid_mask;
 300
 301                 flow_tag = _mm_and_si128(pinfo, pinfo_ft_mask);
 302                 /* Check if flow tag is non-zero then set PKT_RX_FDIR. */
 303                 invalid_mask = _mm_cmpeq_epi32(flow_tag, zero);
 304                 ol_flags = _mm_or_si128(ol_flags,
 305                                         _mm_andnot_si128(invalid_mask,
 306                                                          fdir_flags));
 307                 /* Mask out invalid entries. */
 308                 fdir_id_flags = _mm_andnot_si128(invalid_mask, fdir_id_flags);
 309                 /* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */
 310                 ol_flags = _mm_or_si128(ol_flags,
 311                                         _mm_andnot_si128(
 312                                                 _mm_cmpeq_epi32(flow_tag,
 313                                                                 pinfo_ft_mask),
 314                                                 fdir_id_flags));
 315         }
 316         /*
 317          * Merge the two fields to generate the following:
 318          * bit[1]     = l3_ok
 319          * bit[2]     = l4_ok
 320          * bit[8]     = cv
 321          * bit[11:10] = l3_hdr_type
 322          * bit[14:12] = l4_hdr_type
 323          * bit[15]    = ip_frag
 324          * bit[16]    = tunneled
 325          * bit[17]    = outer_l3_type
 326          */
 327         ptype = _mm_and_si128(ptype, ptype_mask);
 328         pinfo = _mm_and_si128(pinfo, pinfo_mask);
 329         pinfo = _mm_slli_epi32(pinfo, 16);
 330         /* Make pinfo has merged fields for ol_flags calculation. */
 331         pinfo = _mm_or_si128(ptype, pinfo);
 332         ptype = _mm_srli_epi32(pinfo, 10);
 333         ptype = _mm_packs_epi32(ptype, zero);
 334         /* Errored packets will have RTE_PTYPE_ALL_MASK. */
 335         op_err = _mm_srli_epi16(op_err, 8);
 336         ptype = _mm_or_si128(ptype, op_err);
 337         pt_idx0 = _mm_extract_epi8(ptype, 0);
 338         pt_idx1 = _mm_extract_epi8(ptype, 2);
 339         pt_idx2 = _mm_extract_epi8(ptype, 4);
 340         pt_idx3 = _mm_extract_epi8(ptype, 6);
 341         pkts[0]->packet_type = mlx5_ptype_table[pt_idx0] |
 342                                !!(pt_idx0 & (1 << 6)) * rxq->tunnel;
 343         pkts[1]->packet_type = mlx5_ptype_table[pt_idx1] |
 344                                !!(pt_idx1 & (1 << 6)) * rxq->tunnel;
 345         pkts[2]->packet_type = mlx5_ptype_table[pt_idx2] |
 346                                !!(pt_idx2 & (1 << 6)) * rxq->tunnel;
 347         pkts[3]->packet_type = mlx5_ptype_table[pt_idx3] |
 348                                !!(pt_idx3 & (1 << 6)) * rxq->tunnel;
 349         /* Fill flags for checksum and VLAN. */
 350         pinfo = _mm_and_si128(pinfo, ptype_ol_mask);
 351         pinfo = _mm_shuffle_epi8(cv_flag_sel, pinfo);
 352         /* Locate checksum flags at byte[2:1] and merge with VLAN flags. */
 353         cv_flags = _mm_slli_epi32(pinfo, 9);
 354         cv_flags = _mm_or_si128(pinfo, cv_flags);
 355         /* Move back flags to start from byte[0]. */
 356         cv_flags = _mm_srli_epi32(cv_flags, 8);
 357         /* Mask out garbage bits. */
 358         cv_flags = _mm_and_si128(cv_flags, cv_mask);
 359         /* Merge to ol_flags. */
 360         ol_flags = _mm_or_si128(ol_flags, cv_flags);
 361         /* Merge mbuf_init and ol_flags. */
 362         rearm0 = _mm_blend_epi16(mbuf_init, _mm_slli_si128(ol_flags, 8), 0x30);
 363         rearm1 = _mm_blend_epi16(mbuf_init, _mm_slli_si128(ol_flags, 4), 0x30);
 364         rearm2 = _mm_blend_epi16(mbuf_init, ol_flags, 0x30);
 365         rearm3 = _mm_blend_epi16(mbuf_init, _mm_srli_si128(ol_flags, 4), 0x30);
 366         /* Write 8B rearm_data and 8B ol_flags. */
 367         _mm_store_si128((__m128i *)&pkts[0]->rearm_data, rearm0);
 368         _mm_store_si128((__m128i *)&pkts[1]->rearm_data, rearm1);
 369         _mm_store_si128((__m128i *)&pkts[2]->rearm_data, rearm2);
 370         _mm_store_si128((__m128i *)&pkts[3]->rearm_data, rearm3);
 371 }
 372
 373 /**
 374  * Receive burst of packets. An errored completion also consumes a mbuf, but the
 375  * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
 376  * before returning to application.
 377  *
 378  * @param rxq
 379  *   Pointer to RX queue structure.
 380  * @param[out] pkts
 381  *   Array to store received packets.
 382  * @param pkts_n
 383  *   Maximum number of packets in array.
 384  * @param[out] err
 385  *   Pointer to a flag. Set non-zero value if pkts array has at least one error
 386  *   packet to handle.
 387  *
 388  * @return
 389  *   Number of packets received including errors (<= pkts_n).
 390  */
 391 static inline uint16_t
 392 rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
 393             uint64_t *err)
 394 {
 395         const uint16_t q_n = 1 << rxq->cqe_n;
 396         const uint16_t q_mask = q_n - 1;
 397         volatile struct mlx5_cqe *cq;
 398         struct rte_mbuf **elts;
 399         unsigned int pos;
 400         uint64_t n;
 401         uint16_t repl_n;
 402         uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
 403         uint16_t nocmp_n = 0;
 404         uint16_t rcvd_pkt = 0;
 405         unsigned int cq_idx = rxq->cq_ci & q_mask;
 406         unsigned int elts_idx;
 407         unsigned int ownership = !!(rxq->cq_ci & (q_mask + 1));
 408         const __m128i owner_check =
 409                 _mm_set_epi64x(0x0100000001000000LL, 0x0100000001000000LL);
 410         const __m128i opcode_check =
 411                 _mm_set_epi64x(0xf0000000f0000000LL, 0xf0000000f0000000LL);
 412         const __m128i format_check =
 413                 _mm_set_epi64x(0x0c0000000c000000LL, 0x0c0000000c000000LL);
 414         const __m128i resp_err_check =
 415                 _mm_set_epi64x(0xe0000000e0000000LL, 0xe0000000e0000000LL);
 416 #ifdef MLX5_PMD_SOFT_COUNTERS
 417         uint32_t rcvd_byte = 0;
 418         /* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
 419         const __m128i len_shuf_mask =
 420                 _mm_set_epi8(-1, -1, -1, -1,
 421                              -1, -1, -1, -1,
 422                              12, 13,  8,  9,
 423                               4,  5,  0,  1);
 424 #endif
 425         /* Mask to shuffle from extracted CQE to mbuf. */
 426         const __m128i shuf_mask =
 427                 _mm_set_epi8(-1,  3,  2,  1, /* fdir.hi */
 428                              12, 13, 14, 15, /* rss, bswap32 */
 429                              10, 11,         /* vlan_tci, bswap16 */
 430                               4,  5,         /* data_len, bswap16 */
 431                              -1, -1,         /* zero out 2nd half of pkt_len */
 432                               4,  5          /* pkt_len, bswap16 */);
 433         /* Mask to blend from the last Qword to the first DQword. */
 434         const __m128i blend_mask =
 435                 _mm_set_epi8(-1, -1, -1, -1,
 436                              -1, -1, -1, -1,
 437                               0,  0,  0,  0,
 438                               0,  0,  0, -1);
 439         const __m128i zero = _mm_setzero_si128();
 440         const __m128i ones = _mm_cmpeq_epi32(zero, zero);
 441         const __m128i crc_adj =
 442                 _mm_set_epi16(0, 0, 0, 0, 0,
 443                               rxq->crc_present * RTE_ETHER_CRC_LEN,
 444                               0,
 445                               rxq->crc_present * RTE_ETHER_CRC_LEN);
 446         const __m128i flow_mark_adj = _mm_set_epi32(rxq->mark * (-1), 0, 0, 0);
 447
 448         MLX5_ASSERT(rxq->sges_n == 0);
 449         MLX5_ASSERT(rxq->cqe_n == rxq->elts_n);
 450         cq = &(*rxq->cqes)[cq_idx];
 451         rte_prefetch0(cq);
 452         rte_prefetch0(cq + 1);
 453         rte_prefetch0(cq + 2);
 454         rte_prefetch0(cq + 3);
 455         pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
 456         repl_n = q_n - (rxq->rq_ci - rxq->rq_pi);
 457         if (repl_n >= rxq->rq_repl_thresh)
 458                 mlx5_rx_replenish_bulk_mbuf(rxq, repl_n);
 459         /* See if there're unreturned mbufs from compressed CQE. */
 460         rcvd_pkt = rxq->decompressed;
 461         if (rcvd_pkt > 0) {
 462                 rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
 463                 rxq_copy_mbuf_v(rxq, pkts, rcvd_pkt);
 464                 rxq->rq_pi += rcvd_pkt;
 465                 rxq->decompressed -= rcvd_pkt;
 466                 pkts += rcvd_pkt;
 467         }
 468         elts_idx = rxq->rq_pi & q_mask;
 469         elts = &(*rxq->elts)[elts_idx];
 470         /* Not to overflow pkts array. */
 471         pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
 472         /* Not to cross queue end. */
 473         pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
 474         pkts_n = RTE_MIN(pkts_n, q_n - cq_idx);
 475         if (!pkts_n)
 476                 return rcvd_pkt;
 477         /* At this point, there shouldn't be any remained packets. */
 478         MLX5_ASSERT(rxq->decompressed == 0);
 479         /*
 480          * A. load first Qword (8bytes) in one loop.
 481          * B. copy 4 mbuf pointers from elts ring to returing pkts.
 482          * C. load remained CQE data and extract necessary fields.
 483          *    Final 16bytes cqes[] extracted from original 64bytes CQE has the
 484          *    following structure:
 485          *        struct {
 486          *          uint8_t  pkt_info;
 487          *          uint8_t  flow_tag[3];
 488          *          uint16_t byte_cnt;
 489          *          uint8_t  rsvd4;
 490          *          uint8_t  op_own;
 491          *          uint16_t hdr_type_etc;
 492          *          uint16_t vlan_info;
 493          *          uint32_t rx_has_res;
 494          *        } c;
 495          * D. fill in mbuf.
 496          * E. get valid CQEs.
 497          * F. find compressed CQE.
 498          */
 499         for (pos = 0;
 500              pos < pkts_n;
 501              pos += MLX5_VPMD_DESCS_PER_LOOP) {
 502                 __m128i cqes[MLX5_VPMD_DESCS_PER_LOOP];
 503                 __m128i cqe_tmp1, cqe_tmp2;
 504                 __m128i pkt_mb0, pkt_mb1, pkt_mb2, pkt_mb3;
 505                 __m128i op_own, op_own_tmp1, op_own_tmp2;
 506                 __m128i opcode, owner_mask, invalid_mask;
 507                 __m128i comp_mask;
 508                 __m128i mask;
 509 #ifdef MLX5_PMD_SOFT_COUNTERS
 510                 __m128i byte_cnt;
 511 #endif
 512                 __m128i mbp1, mbp2;
 513                 __m128i p = _mm_set_epi16(0, 0, 0, 0, 3, 2, 1, 0);
 514                 unsigned int p1, p2, p3;
 515
 516                 /* Prefetch next 4 CQEs. */
 517                 if (pkts_n - pos >= 2 * MLX5_VPMD_DESCS_PER_LOOP) {
 518                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP]);
 519                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 1]);
 520                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 2]);
 521                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 3]);
 522                 }
 523                 /* A.0 do not cross the end of CQ. */
 524                 mask = _mm_set_epi64x(0, (pkts_n - pos) * sizeof(uint16_t) * 8);
 525                 mask = _mm_sll_epi64(ones, mask);
 526                 p = _mm_andnot_si128(mask, p);
 527                 /* A.1 load cqes. */
 528                 p3 = _mm_extract_epi16(p, 3);
 529                 cqes[3] = _mm_loadl_epi64((__m128i *)
 530                                            &cq[pos + p3].sop_drop_qpn);
 531                 rte_compiler_barrier();
 532                 p2 = _mm_extract_epi16(p, 2);
 533                 cqes[2] = _mm_loadl_epi64((__m128i *)
 534                                            &cq[pos + p2].sop_drop_qpn);
 535                 rte_compiler_barrier();
 536                 /* B.1 load mbuf pointers. */
 537                 mbp1 = _mm_loadu_si128((__m128i *)&elts[pos]);
 538                 mbp2 = _mm_loadu_si128((__m128i *)&elts[pos + 2]);
 539                 /* A.1 load a block having op_own. */
 540                 p1 = _mm_extract_epi16(p, 1);
 541                 cqes[1] = _mm_loadl_epi64((__m128i *)
 542                                            &cq[pos + p1].sop_drop_qpn);
 543                 rte_compiler_barrier();
 544                 cqes[0] = _mm_loadl_epi64((__m128i *)
 545                                            &cq[pos].sop_drop_qpn);
 546                 /* B.2 copy mbuf pointers. */
 547                 _mm_storeu_si128((__m128i *)&pkts[pos], mbp1);
 548                 _mm_storeu_si128((__m128i *)&pkts[pos + 2], mbp2);
 549                 rte_cio_rmb();
 550                 /* C.1 load remained CQE data and extract necessary fields. */
 551                 cqe_tmp2 = _mm_load_si128((__m128i *)&cq[pos + p3]);
 552                 cqe_tmp1 = _mm_load_si128((__m128i *)&cq[pos + p2]);
 553                 cqes[3] = _mm_blendv_epi8(cqes[3], cqe_tmp2, blend_mask);
 554                 cqes[2] = _mm_blendv_epi8(cqes[2], cqe_tmp1, blend_mask);
 555                 cqe_tmp2 = _mm_loadu_si128((__m128i *)&cq[pos + p3].csum);
 556                 cqe_tmp1 = _mm_loadu_si128((__m128i *)&cq[pos + p2].csum);
 557                 cqes[3] = _mm_blend_epi16(cqes[3], cqe_tmp2, 0x30);
 558                 cqes[2] = _mm_blend_epi16(cqes[2], cqe_tmp1, 0x30);
 559                 cqe_tmp2 = _mm_loadl_epi64((__m128i *)&cq[pos + p3].rsvd4[2]);
 560                 cqe_tmp1 = _mm_loadl_epi64((__m128i *)&cq[pos + p2].rsvd4[2]);
 561                 cqes[3] = _mm_blend_epi16(cqes[3], cqe_tmp2, 0x04);
 562                 cqes[2] = _mm_blend_epi16(cqes[2], cqe_tmp1, 0x04);
 563                 /* C.2 generate final structure for mbuf with swapping bytes. */
 564                 pkt_mb3 = _mm_shuffle_epi8(cqes[3], shuf_mask);
 565                 pkt_mb2 = _mm_shuffle_epi8(cqes[2], shuf_mask);
 566                 /* C.3 adjust CRC length. */
 567                 pkt_mb3 = _mm_sub_epi16(pkt_mb3, crc_adj);
 568                 pkt_mb2 = _mm_sub_epi16(pkt_mb2, crc_adj);
 569                 /* C.4 adjust flow mark. */
 570                 pkt_mb3 = _mm_add_epi32(pkt_mb3, flow_mark_adj);
 571                 pkt_mb2 = _mm_add_epi32(pkt_mb2, flow_mark_adj);
 572                 /* D.1 fill in mbuf - rx_descriptor_fields1. */
 573                 _mm_storeu_si128((void *)&pkts[pos + 3]->pkt_len, pkt_mb3);
 574                 _mm_storeu_si128((void *)&pkts[pos + 2]->pkt_len, pkt_mb2);
 575                 /* E.1 extract op_own field. */
 576                 op_own_tmp2 = _mm_unpacklo_epi32(cqes[2], cqes[3]);
 577                 /* C.1 load remained CQE data and extract necessary fields. */
 578                 cqe_tmp2 = _mm_load_si128((__m128i *)&cq[pos + p1]);
 579                 cqe_tmp1 = _mm_load_si128((__m128i *)&cq[pos]);
 580                 cqes[1] = _mm_blendv_epi8(cqes[1], cqe_tmp2, blend_mask);
 581                 cqes[0] = _mm_blendv_epi8(cqes[0], cqe_tmp1, blend_mask);
 582                 cqe_tmp2 = _mm_loadu_si128((__m128i *)&cq[pos + p1].csum);
 583                 cqe_tmp1 = _mm_loadu_si128((__m128i *)&cq[pos].csum);
 584                 cqes[1] = _mm_blend_epi16(cqes[1], cqe_tmp2, 0x30);
 585                 cqes[0] = _mm_blend_epi16(cqes[0], cqe_tmp1, 0x30);
 586                 cqe_tmp2 = _mm_loadl_epi64((__m128i *)&cq[pos + p1].rsvd4[2]);
 587                 cqe_tmp1 = _mm_loadl_epi64((__m128i *)&cq[pos].rsvd4[2]);
 588                 cqes[1] = _mm_blend_epi16(cqes[1], cqe_tmp2, 0x04);
 589                 cqes[0] = _mm_blend_epi16(cqes[0], cqe_tmp1, 0x04);
 590                 /* C.2 generate final structure for mbuf with swapping bytes. */
 591                 pkt_mb1 = _mm_shuffle_epi8(cqes[1], shuf_mask);
 592                 pkt_mb0 = _mm_shuffle_epi8(cqes[0], shuf_mask);
 593                 /* C.3 adjust CRC length. */
 594                 pkt_mb1 = _mm_sub_epi16(pkt_mb1, crc_adj);
 595                 pkt_mb0 = _mm_sub_epi16(pkt_mb0, crc_adj);
 596                 /* C.4 adjust flow mark. */
 597                 pkt_mb1 = _mm_add_epi32(pkt_mb1, flow_mark_adj);
 598                 pkt_mb0 = _mm_add_epi32(pkt_mb0, flow_mark_adj);
 599                 /* E.1 extract op_own byte. */
 600                 op_own_tmp1 = _mm_unpacklo_epi32(cqes[0], cqes[1]);
 601                 op_own = _mm_unpackhi_epi64(op_own_tmp1, op_own_tmp2);
 602                 /* D.1 fill in mbuf - rx_descriptor_fields1. */
 603                 _mm_storeu_si128((void *)&pkts[pos + 1]->pkt_len, pkt_mb1);
 604                 _mm_storeu_si128((void *)&pkts[pos]->pkt_len, pkt_mb0);
 605                 /* E.2 flip owner bit to mark CQEs from last round. */
 606                 owner_mask = _mm_and_si128(op_own, owner_check);
 607                 if (ownership)
 608                         owner_mask = _mm_xor_si128(owner_mask, owner_check);
 609                 owner_mask = _mm_cmpeq_epi32(owner_mask, owner_check);
 610                 owner_mask = _mm_packs_epi32(owner_mask, zero);
 611                 /* E.3 get mask for invalidated CQEs. */
 612                 opcode = _mm_and_si128(op_own, opcode_check);
 613                 invalid_mask = _mm_cmpeq_epi32(opcode_check, opcode);
 614                 invalid_mask = _mm_packs_epi32(invalid_mask, zero);
 615                 /* E.4 mask out beyond boundary. */
 616                 invalid_mask = _mm_or_si128(invalid_mask, mask);
 617                 /* E.5 merge invalid_mask with invalid owner. */
 618                 invalid_mask = _mm_or_si128(invalid_mask, owner_mask);
 619                 /* F.1 find compressed CQE format. */
 620                 comp_mask = _mm_and_si128(op_own, format_check);
 621                 comp_mask = _mm_cmpeq_epi32(comp_mask, format_check);
 622                 comp_mask = _mm_packs_epi32(comp_mask, zero);
 623                 /* F.2 mask out invalid entries. */
 624                 comp_mask = _mm_andnot_si128(invalid_mask, comp_mask);
 625                 comp_idx = _mm_cvtsi128_si64(comp_mask);
 626                 /* F.3 get the first compressed CQE. */
 627                 comp_idx = comp_idx ?
 628                                 __builtin_ctzll(comp_idx) /
 629                                         (sizeof(uint16_t) * 8) :
 630                                 MLX5_VPMD_DESCS_PER_LOOP;
 631                 /* E.6 mask out entries after the compressed CQE. */
 632                 mask = _mm_set_epi64x(0, comp_idx * sizeof(uint16_t) * 8);
 633                 mask = _mm_sll_epi64(ones, mask);
 634                 invalid_mask = _mm_or_si128(invalid_mask, mask);
 635                 /* E.7 count non-compressed valid CQEs. */
 636                 n = _mm_cvtsi128_si64(invalid_mask);
 637                 n = n ? __builtin_ctzll(n) / (sizeof(uint16_t) * 8) :
 638                         MLX5_VPMD_DESCS_PER_LOOP;
 639                 nocmp_n += n;
 640                 /* D.2 get the final invalid mask. */
 641                 mask = _mm_set_epi64x(0, n * sizeof(uint16_t) * 8);
 642                 mask = _mm_sll_epi64(ones, mask);
 643                 invalid_mask = _mm_or_si128(invalid_mask, mask);
 644                 /* D.3 check error in opcode. */
 645                 opcode = _mm_cmpeq_epi32(resp_err_check, opcode);
 646                 opcode = _mm_packs_epi32(opcode, zero);
 647                 opcode = _mm_andnot_si128(invalid_mask, opcode);
 648                 /* D.4 mark if any error is set */
 649                 *err |= _mm_cvtsi128_si64(opcode);
 650                 /* D.5 fill in mbuf - rearm_data and packet_type. */
 651                 rxq_cq_to_ptype_oflags_v(rxq, cqes, opcode, &pkts[pos]);
 652                 if (rxq->hw_timestamp) {
 653                         pkts[pos]->timestamp =
 654                                 rte_be_to_cpu_64(cq[pos].timestamp);
 655                         pkts[pos + 1]->timestamp =
 656                                 rte_be_to_cpu_64(cq[pos + p1].timestamp);
 657                         pkts[pos + 2]->timestamp =
 658                                 rte_be_to_cpu_64(cq[pos + p2].timestamp);
 659                         pkts[pos + 3]->timestamp =
 660                                 rte_be_to_cpu_64(cq[pos + p3].timestamp);
 661                 }
 662                 if (rxq->dynf_meta) {
 663                         /* This code is subject for futher optimization. */
 664                         int32_t offs = rxq->flow_meta_offset;
 665
 666                         *RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *) =
 667                                 cq[pos].flow_table_metadata;
 668                         *RTE_MBUF_DYNFIELD(pkts[pos + 1], offs, uint32_t *) =
 669                                 cq[pos + p1].flow_table_metadata;
 670                         *RTE_MBUF_DYNFIELD(pkts[pos + 2], offs, uint32_t *) =
 671                                 cq[pos + p2].flow_table_metadata;
 672                         *RTE_MBUF_DYNFIELD(pkts[pos + 3], offs, uint32_t *) =
 673                                 cq[pos + p3].flow_table_metadata;
 674                         if (*RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *))
 675                                 pkts[pos]->ol_flags |= rxq->flow_meta_mask;
 676                         if (*RTE_MBUF_DYNFIELD(pkts[pos + 1], offs, uint32_t *))
 677                                 pkts[pos + 1]->ol_flags |= rxq->flow_meta_mask;
 678                         if (*RTE_MBUF_DYNFIELD(pkts[pos + 2], offs, uint32_t *))
 679                                 pkts[pos + 2]->ol_flags |= rxq->flow_meta_mask;
 680                         if (*RTE_MBUF_DYNFIELD(pkts[pos + 3], offs, uint32_t *))
 681                                 pkts[pos + 3]->ol_flags |= rxq->flow_meta_mask;
 682                 }
 683 #ifdef MLX5_PMD_SOFT_COUNTERS
 684                 /* Add up received bytes count. */
 685                 byte_cnt = _mm_shuffle_epi8(op_own, len_shuf_mask);
 686                 byte_cnt = _mm_andnot_si128(invalid_mask, byte_cnt);
 687                 byte_cnt = _mm_hadd_epi16(byte_cnt, zero);
 688                 rcvd_byte += _mm_cvtsi128_si64(_mm_hadd_epi16(byte_cnt, zero));
 689 #endif
 690                 /*
 691                  * Break the loop unless more valid CQE is expected, or if
 692                  * there's a compressed CQE.
 693                  */
 694                 if (n != MLX5_VPMD_DESCS_PER_LOOP)
 695                         break;
 696         }
 697         /* If no new CQE seen, return without updating cq_db. */
 698         if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP))
 699                 return rcvd_pkt;
 700         /* Update the consumer indexes for non-compressed CQEs. */
 701         MLX5_ASSERT(nocmp_n <= pkts_n);
 702         rxq->cq_ci += nocmp_n;
 703         rxq->rq_pi += nocmp_n;
 704         rcvd_pkt += nocmp_n;
 705 #ifdef MLX5_PMD_SOFT_COUNTERS
 706         rxq->stats.ipackets += nocmp_n;
 707         rxq->stats.ibytes += rcvd_byte;
 708 #endif
 709         /* Decompress the last CQE if compressed. */
 710         if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP && comp_idx == n) {
 711                 MLX5_ASSERT(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
 712                 rxq->decompressed = rxq_cq_decompress_v(rxq, &cq[nocmp_n],
 713                                                         &elts[nocmp_n]);
 714                 /* Return more packets if needed. */
 715                 if (nocmp_n < pkts_n) {
 716                         uint16_t n = rxq->decompressed;
 717
 718                         n = RTE_MIN(n, pkts_n - nocmp_n);
 719                         rxq_copy_mbuf_v(rxq, &pkts[nocmp_n], n);
 720                         rxq->rq_pi += n;
 721                         rcvd_pkt += n;
 722                         rxq->decompressed -= n;
 723                 }
 724         }
 725         rte_compiler_barrier();
 726         *rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
 727         return rcvd_pkt;
 728 }
 729
 730 #endif /* RTE_PMD_MLX5_RXTX_VEC_SSE_H_ */