drivers/net/iavf/iavf_rxtx_vec_avx2.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2019 Intel Corporation
   3  */
   4
   5 #include "iavf_rxtx_vec_common.h"
   6
   7 #include <rte_vect.h>
   8
   9 #ifndef __INTEL_COMPILER
  10 #pragma GCC diagnostic ignored "-Wcast-qual"
  11 #endif
  12
  13 static __rte_always_inline void
  14 iavf_rxq_rearm(struct iavf_rx_queue *rxq)
  15 {
  16         return iavf_rxq_rearm_common(rxq, false);
  17 }
  18
  19 #define PKTLEN_SHIFT     10
  20
  21 static inline uint16_t
  22 _iavf_recv_raw_pkts_vec_avx2(struct iavf_rx_queue *rxq,
  23                              struct rte_mbuf **rx_pkts,
  24                              uint16_t nb_pkts, uint8_t *split_packet)
  25 {
  26 #define IAVF_DESCS_PER_LOOP_AVX 8
  27
  28         /* const uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl; */
  29         const uint32_t *type_table = rxq->vsi->adapter->ptype_tbl;
  30
  31         const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
  32                         0, rxq->mbuf_initializer);
  33         /* struct iavf_rx_entry *sw_ring = &rxq->sw_ring[rxq->rx_tail]; */
  34         struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
  35         volatile union iavf_rx_desc *rxdp = rxq->rx_ring + rxq->rx_tail;
  36         const int avx_aligned = ((rxq->rx_tail & 1) == 0);
  37
  38         rte_prefetch0(rxdp);
  39
  40         /* nb_pkts has to be floor-aligned to IAVF_DESCS_PER_LOOP_AVX */
  41         nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_DESCS_PER_LOOP_AVX);
  42
  43         /* See if we need to rearm the RX queue - gives the prefetch a bit
  44          * of time to act
  45          */
  46         if (rxq->rxrearm_nb > IAVF_RXQ_REARM_THRESH)
  47                 iavf_rxq_rearm(rxq);
  48
  49         /* Before we start moving massive data around, check to see if
  50          * there is actually a packet available
  51          */
  52         if (!(rxdp->wb.qword1.status_error_len &
  53                         rte_cpu_to_le_32(1 << IAVF_RX_DESC_STATUS_DD_SHIFT)))
  54                 return 0;
  55
  56         /* constants used in processing loop */
  57         const __m256i crc_adjust =
  58                 _mm256_set_epi16
  59                         (/* first descriptor */
  60                          0, 0, 0,       /* ignore non-length fields */
  61                          -rxq->crc_len, /* sub crc on data_len */
  62                          0,             /* ignore high-16bits of pkt_len */
  63                          -rxq->crc_len, /* sub crc on pkt_len */
  64                          0, 0,          /* ignore pkt_type field */
  65                          /* second descriptor */
  66                          0, 0, 0,       /* ignore non-length fields */
  67                          -rxq->crc_len, /* sub crc on data_len */
  68                          0,             /* ignore high-16bits of pkt_len */
  69                          -rxq->crc_len, /* sub crc on pkt_len */
  70                          0, 0           /* ignore pkt_type field */
  71                         );
  72
  73         /* 8 packets DD mask, LSB in each 32-bit value */
  74         const __m256i dd_check = _mm256_set1_epi32(1);
  75
  76         /* 8 packets EOP mask, second-LSB in each 32-bit value */
  77         const __m256i eop_check = _mm256_slli_epi32(dd_check,
  78                         IAVF_RX_DESC_STATUS_EOF_SHIFT);
  79
  80         /* mask to shuffle from desc. to mbuf (2 descriptors)*/
  81         const __m256i shuf_msk =
  82                 _mm256_set_epi8
  83                         (/* first descriptor */
  84                          7, 6, 5, 4,  /* octet 4~7, 32bits rss */
  85                          3, 2,        /* octet 2~3, low 16 bits vlan_macip */
  86                          15, 14,      /* octet 15~14, 16 bits data_len */
  87                          0xFF, 0xFF,  /* skip high 16 bits pkt_len, zero out */
  88                          15, 14,      /* octet 15~14, low 16 bits pkt_len */
  89                          0xFF, 0xFF,  /* pkt_type set as unknown */
  90                          0xFF, 0xFF,  /*pkt_type set as unknown */
  91                          /* second descriptor */
  92                          7, 6, 5, 4,  /* octet 4~7, 32bits rss */
  93                          3, 2,        /* octet 2~3, low 16 bits vlan_macip */
  94                          15, 14,      /* octet 15~14, 16 bits data_len */
  95                          0xFF, 0xFF,  /* skip high 16 bits pkt_len, zero out */
  96                          15, 14,      /* octet 15~14, low 16 bits pkt_len */
  97                          0xFF, 0xFF,  /* pkt_type set as unknown */
  98                          0xFF, 0xFF   /*pkt_type set as unknown */
  99                         );
 100         /**
 101          * compile-time check the above crc and shuffle layout is correct.
 102          * NOTE: the first field (lowest address) is given last in set_epi
 103          * calls above.
 104          */
 105         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
 106                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
 107         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
 108                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
 109         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, vlan_tci) !=
 110                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 10);
 111         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) !=
 112                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
 113
 114         /* Status/Error flag masks */
 115         /**
 116          * mask everything except RSS, flow director and VLAN flags
 117          * bit2 is for VLAN tag, bit11 for flow director indication
 118          * bit13:12 for RSS indication. Bits 3-5 of error
 119          * field (bits 22-24) are for IP/L4 checksum errors
 120          */
 121         const __m256i flags_mask =
 122                  _mm256_set1_epi32((1 << 2) | (1 << 11) |
 123                                    (3 << 12) | (7 << 22));
 124         /**
 125          * data to be shuffled by result of flag mask. If VLAN bit is set,
 126          * (bit 2), then position 4 in this array will be used in the
 127          * destination
 128          */
 129         const __m256i vlan_flags_shuf =
 130                 _mm256_set_epi32(0, 0, PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED, 0,
 131                                  0, 0, PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED, 0);
 132         /**
 133          * data to be shuffled by result of flag mask, shifted down 11.
 134          * If RSS/FDIR bits are set, shuffle moves appropriate flags in
 135          * place.
 136          */
 137         const __m256i rss_flags_shuf =
 138                 _mm256_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
 139                                 PKT_RX_RSS_HASH | PKT_RX_FDIR, PKT_RX_RSS_HASH,
 140                                 0, 0, 0, 0, PKT_RX_FDIR, 0,/* end up 128-bits */
 141                                 0, 0, 0, 0, 0, 0, 0, 0,
 142                                 PKT_RX_RSS_HASH | PKT_RX_FDIR, PKT_RX_RSS_HASH,
 143                                 0, 0, 0, 0, PKT_RX_FDIR, 0);
 144
 145         /**
 146          * data to be shuffled by the result of the flags mask shifted by 22
 147          * bits.  This gives use the l3_l4 flags.
 148          */
 149         const __m256i l3_l4_flags_shuf = _mm256_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
 150                         /* shift right 1 bit to make sure it not exceed 255 */
 151                         (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
 152                          PKT_RX_IP_CKSUM_BAD) >> 1,
 153                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_OUTER_IP_CKSUM_BAD |
 154                          PKT_RX_L4_CKSUM_BAD) >> 1,
 155                         (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
 156                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_OUTER_IP_CKSUM_BAD) >> 1,
 157                         (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
 158                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD) >> 1,
 159                         PKT_RX_IP_CKSUM_BAD >> 1,
 160                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1,
 161                         /* second 128-bits */
 162                         0, 0, 0, 0, 0, 0, 0, 0,
 163                         (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
 164                          PKT_RX_IP_CKSUM_BAD) >> 1,
 165                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_OUTER_IP_CKSUM_BAD |
 166                          PKT_RX_L4_CKSUM_BAD) >> 1,
 167                         (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
 168                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_OUTER_IP_CKSUM_BAD) >> 1,
 169                         (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
 170                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD) >> 1,
 171                         PKT_RX_IP_CKSUM_BAD >> 1,
 172                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1);
 173
 174         const __m256i cksum_mask =
 175                  _mm256_set1_epi32(PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
 176                                    PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
 177                                    PKT_RX_OUTER_IP_CKSUM_BAD);
 178
 179         RTE_SET_USED(avx_aligned); /* for 32B descriptors we don't use this */
 180
 181         uint16_t i, received;
 182
 183         for (i = 0, received = 0; i < nb_pkts;
 184              i += IAVF_DESCS_PER_LOOP_AVX,
 185              rxdp += IAVF_DESCS_PER_LOOP_AVX) {
 186                 /* step 1, copy over 8 mbuf pointers to rx_pkts array */
 187                 _mm256_storeu_si256((void *)&rx_pkts[i],
 188                                     _mm256_loadu_si256((void *)&sw_ring[i]));
 189 #ifdef RTE_ARCH_X86_64
 190                 _mm256_storeu_si256
 191                         ((void *)&rx_pkts[i + 4],
 192                          _mm256_loadu_si256((void *)&sw_ring[i + 4]));
 193 #endif
 194
 195                 __m256i raw_desc0_1, raw_desc2_3, raw_desc4_5, raw_desc6_7;
 196 #ifdef RTE_LIBRTE_IAVF_16BYTE_RX_DESC
 197                 /* for AVX we need alignment otherwise loads are not atomic */
 198                 if (avx_aligned) {
 199                         /* load in descriptors, 2 at a time, in reverse order */
 200                         raw_desc6_7 = _mm256_load_si256((void *)(rxdp + 6));
 201                         rte_compiler_barrier();
 202                         raw_desc4_5 = _mm256_load_si256((void *)(rxdp + 4));
 203                         rte_compiler_barrier();
 204                         raw_desc2_3 = _mm256_load_si256((void *)(rxdp + 2));
 205                         rte_compiler_barrier();
 206                         raw_desc0_1 = _mm256_load_si256((void *)(rxdp + 0));
 207                 } else
 208 #endif
 209                 {
 210                         const __m128i raw_desc7 =
 211                                 _mm_load_si128((void *)(rxdp + 7));
 212                         rte_compiler_barrier();
 213                         const __m128i raw_desc6 =
 214                                 _mm_load_si128((void *)(rxdp + 6));
 215                         rte_compiler_barrier();
 216                         const __m128i raw_desc5 =
 217                                 _mm_load_si128((void *)(rxdp + 5));
 218                         rte_compiler_barrier();
 219                         const __m128i raw_desc4 =
 220                                 _mm_load_si128((void *)(rxdp + 4));
 221                         rte_compiler_barrier();
 222                         const __m128i raw_desc3 =
 223                                 _mm_load_si128((void *)(rxdp + 3));
 224                         rte_compiler_barrier();
 225                         const __m128i raw_desc2 =
 226                                 _mm_load_si128((void *)(rxdp + 2));
 227                         rte_compiler_barrier();
 228                         const __m128i raw_desc1 =
 229                                 _mm_load_si128((void *)(rxdp + 1));
 230                         rte_compiler_barrier();
 231                         const __m128i raw_desc0 =
 232                                 _mm_load_si128((void *)(rxdp + 0));
 233
 234                         raw_desc6_7 =
 235                                 _mm256_inserti128_si256
 236                                         (_mm256_castsi128_si256(raw_desc6),
 237                                          raw_desc7, 1);
 238                         raw_desc4_5 =
 239                                 _mm256_inserti128_si256
 240                                         (_mm256_castsi128_si256(raw_desc4),
 241                                          raw_desc5, 1);
 242                         raw_desc2_3 =
 243                                 _mm256_inserti128_si256
 244                                         (_mm256_castsi128_si256(raw_desc2),
 245                                          raw_desc3, 1);
 246                         raw_desc0_1 =
 247                                 _mm256_inserti128_si256
 248                                         (_mm256_castsi128_si256(raw_desc0),
 249                                          raw_desc1, 1);
 250                 }
 251
 252                 if (split_packet) {
 253                         int j;
 254
 255                         for (j = 0; j < IAVF_DESCS_PER_LOOP_AVX; j++)
 256                                 rte_mbuf_prefetch_part2(rx_pkts[i + j]);
 257                 }
 258
 259                 /**
 260                  * convert descriptors 4-7 into mbufs, adjusting length and
 261                  * re-arranging fields. Then write into the mbuf
 262                  */
 263                 const __m256i len6_7 = _mm256_slli_epi32(raw_desc6_7,
 264                                                          PKTLEN_SHIFT);
 265                 const __m256i len4_5 = _mm256_slli_epi32(raw_desc4_5,
 266                                                          PKTLEN_SHIFT);
 267                 const __m256i desc6_7 = _mm256_blend_epi16(raw_desc6_7,
 268                                                            len6_7, 0x80);
 269                 const __m256i desc4_5 = _mm256_blend_epi16(raw_desc4_5,
 270                                                            len4_5, 0x80);
 271                 __m256i mb6_7 = _mm256_shuffle_epi8(desc6_7, shuf_msk);
 272                 __m256i mb4_5 = _mm256_shuffle_epi8(desc4_5, shuf_msk);
 273
 274                 mb6_7 = _mm256_add_epi16(mb6_7, crc_adjust);
 275                 mb4_5 = _mm256_add_epi16(mb4_5, crc_adjust);
 276                 /**
 277                  * to get packet types, shift 64-bit values down 30 bits
 278                  * and so ptype is in lower 8-bits in each
 279                  */
 280                 const __m256i ptypes6_7 = _mm256_srli_epi64(desc6_7, 30);
 281                 const __m256i ptypes4_5 = _mm256_srli_epi64(desc4_5, 30);
 282                 const uint8_t ptype7 = _mm256_extract_epi8(ptypes6_7, 24);
 283                 const uint8_t ptype6 = _mm256_extract_epi8(ptypes6_7, 8);
 284                 const uint8_t ptype5 = _mm256_extract_epi8(ptypes4_5, 24);
 285                 const uint8_t ptype4 = _mm256_extract_epi8(ptypes4_5, 8);
 286
 287                 mb6_7 = _mm256_insert_epi32(mb6_7, type_table[ptype7], 4);
 288                 mb6_7 = _mm256_insert_epi32(mb6_7, type_table[ptype6], 0);
 289                 mb4_5 = _mm256_insert_epi32(mb4_5, type_table[ptype5], 4);
 290                 mb4_5 = _mm256_insert_epi32(mb4_5, type_table[ptype4], 0);
 291                 /* merge the status bits into one register */
 292                 const __m256i status4_7 = _mm256_unpackhi_epi32(desc6_7,
 293                                 desc4_5);
 294
 295                 /**
 296                  * convert descriptors 0-3 into mbufs, adjusting length and
 297                  * re-arranging fields. Then write into the mbuf
 298                  */
 299                 const __m256i len2_3 = _mm256_slli_epi32(raw_desc2_3,
 300                                                          PKTLEN_SHIFT);
 301                 const __m256i len0_1 = _mm256_slli_epi32(raw_desc0_1,
 302                                                          PKTLEN_SHIFT);
 303                 const __m256i desc2_3 = _mm256_blend_epi16(raw_desc2_3,
 304                                                            len2_3, 0x80);
 305                 const __m256i desc0_1 = _mm256_blend_epi16(raw_desc0_1,
 306                                                            len0_1, 0x80);
 307                 __m256i mb2_3 = _mm256_shuffle_epi8(desc2_3, shuf_msk);
 308                 __m256i mb0_1 = _mm256_shuffle_epi8(desc0_1, shuf_msk);
 309
 310                 mb2_3 = _mm256_add_epi16(mb2_3, crc_adjust);
 311                 mb0_1 = _mm256_add_epi16(mb0_1, crc_adjust);
 312                 /* get the packet types */
 313                 const __m256i ptypes2_3 = _mm256_srli_epi64(desc2_3, 30);
 314                 const __m256i ptypes0_1 = _mm256_srli_epi64(desc0_1, 30);
 315                 const uint8_t ptype3 = _mm256_extract_epi8(ptypes2_3, 24);
 316                 const uint8_t ptype2 = _mm256_extract_epi8(ptypes2_3, 8);
 317                 const uint8_t ptype1 = _mm256_extract_epi8(ptypes0_1, 24);
 318                 const uint8_t ptype0 = _mm256_extract_epi8(ptypes0_1, 8);
 319
 320                 mb2_3 = _mm256_insert_epi32(mb2_3, type_table[ptype3], 4);
 321                 mb2_3 = _mm256_insert_epi32(mb2_3, type_table[ptype2], 0);
 322                 mb0_1 = _mm256_insert_epi32(mb0_1, type_table[ptype1], 4);
 323                 mb0_1 = _mm256_insert_epi32(mb0_1, type_table[ptype0], 0);
 324                 /* merge the status bits into one register */
 325                 const __m256i status0_3 = _mm256_unpackhi_epi32(desc2_3,
 326                                                                 desc0_1);
 327
 328                 /**
 329                  * take the two sets of status bits and merge to one
 330                  * After merge, the packets status flags are in the
 331                  * order (hi->lo): [1, 3, 5, 7, 0, 2, 4, 6]
 332                  */
 333                 __m256i status0_7 = _mm256_unpacklo_epi64(status4_7,
 334                                                           status0_3);
 335
 336                 /* now do flag manipulation */
 337
 338                 /* get only flag/error bits we want */
 339                 const __m256i flag_bits =
 340                         _mm256_and_si256(status0_7, flags_mask);
 341                 /* set vlan and rss flags */
 342                 const __m256i vlan_flags =
 343                         _mm256_shuffle_epi8(vlan_flags_shuf, flag_bits);
 344                 const __m256i rss_flags =
 345                         _mm256_shuffle_epi8(rss_flags_shuf,
 346                                             _mm256_srli_epi32(flag_bits, 11));
 347                 /**
 348                  * l3_l4_error flags, shuffle, then shift to correct adjustment
 349                  * of flags in flags_shuf, and finally mask out extra bits
 350                  */
 351                 __m256i l3_l4_flags = _mm256_shuffle_epi8(l3_l4_flags_shuf,
 352                                 _mm256_srli_epi32(flag_bits, 22));
 353                 l3_l4_flags = _mm256_slli_epi32(l3_l4_flags, 1);
 354                 l3_l4_flags = _mm256_and_si256(l3_l4_flags, cksum_mask);
 355
 356                 /* merge flags */
 357                 const __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
 358                                 _mm256_or_si256(rss_flags, vlan_flags));
 359                 /**
 360                  * At this point, we have the 8 sets of flags in the low 16-bits
 361                  * of each 32-bit value in vlan0.
 362                  * We want to extract these, and merge them with the mbuf init
 363                  * data so we can do a single write to the mbuf to set the flags
 364                  * and all the other initialization fields. Extracting the
 365                  * appropriate flags means that we have to do a shift and blend
 366                  * for each mbuf before we do the write. However, we can also
 367                  * add in the previously computed rx_descriptor fields to
 368                  * make a single 256-bit write per mbuf
 369                  */
 370                 /* check the structure matches expectations */
 371                 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, ol_flags) !=
 372                                  offsetof(struct rte_mbuf, rearm_data) + 8);
 373                 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, rearm_data) !=
 374                                  RTE_ALIGN(offsetof(struct rte_mbuf,
 375                                                     rearm_data),
 376                                            16));
 377                 /* build up data and do writes */
 378                 __m256i rearm0, rearm1, rearm2, rearm3, rearm4, rearm5,
 379                         rearm6, rearm7;
 380                 rearm6 = _mm256_blend_epi32(mbuf_init,
 381                                             _mm256_slli_si256(mbuf_flags, 8),
 382                                             0x04);
 383                 rearm4 = _mm256_blend_epi32(mbuf_init,
 384                                             _mm256_slli_si256(mbuf_flags, 4),
 385                                             0x04);
 386                 rearm2 = _mm256_blend_epi32(mbuf_init, mbuf_flags, 0x04);
 387                 rearm0 = _mm256_blend_epi32(mbuf_init,
 388                                             _mm256_srli_si256(mbuf_flags, 4),
 389                                             0x04);
 390                 /* permute to add in the rx_descriptor e.g. rss fields */
 391                 rearm6 = _mm256_permute2f128_si256(rearm6, mb6_7, 0x20);
 392                 rearm4 = _mm256_permute2f128_si256(rearm4, mb4_5, 0x20);
 393                 rearm2 = _mm256_permute2f128_si256(rearm2, mb2_3, 0x20);
 394                 rearm0 = _mm256_permute2f128_si256(rearm0, mb0_1, 0x20);
 395                 /* write to mbuf */
 396                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 6]->rearm_data,
 397                                     rearm6);
 398                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 4]->rearm_data,
 399                                     rearm4);
 400                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 2]->rearm_data,
 401                                     rearm2);
 402                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 0]->rearm_data,
 403                                     rearm0);
 404
 405                 /* repeat for the odd mbufs */
 406                 const __m256i odd_flags =
 407                         _mm256_castsi128_si256
 408                                 (_mm256_extracti128_si256(mbuf_flags, 1));
 409                 rearm7 = _mm256_blend_epi32(mbuf_init,
 410                                             _mm256_slli_si256(odd_flags, 8),
 411                                             0x04);
 412                 rearm5 = _mm256_blend_epi32(mbuf_init,
 413                                             _mm256_slli_si256(odd_flags, 4),
 414                                             0x04);
 415                 rearm3 = _mm256_blend_epi32(mbuf_init, odd_flags, 0x04);
 416                 rearm1 = _mm256_blend_epi32(mbuf_init,
 417                                             _mm256_srli_si256(odd_flags, 4),
 418                                             0x04);
 419                 /* since odd mbufs are already in hi 128-bits use blend */
 420                 rearm7 = _mm256_blend_epi32(rearm7, mb6_7, 0xF0);
 421                 rearm5 = _mm256_blend_epi32(rearm5, mb4_5, 0xF0);
 422                 rearm3 = _mm256_blend_epi32(rearm3, mb2_3, 0xF0);
 423                 rearm1 = _mm256_blend_epi32(rearm1, mb0_1, 0xF0);
 424                 /* again write to mbufs */
 425                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 7]->rearm_data,
 426                                     rearm7);
 427                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 5]->rearm_data,
 428                                     rearm5);
 429                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 3]->rearm_data,
 430                                     rearm3);
 431                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 1]->rearm_data,
 432                                     rearm1);
 433
 434                 /* extract and record EOP bit */
 435                 if (split_packet) {
 436                         const __m128i eop_mask =
 437                                 _mm_set1_epi16(1 << IAVF_RX_DESC_STATUS_EOF_SHIFT);
 438                         const __m256i eop_bits256 = _mm256_and_si256(status0_7,
 439                                                                      eop_check);
 440                         /* pack status bits into a single 128-bit register */
 441                         const __m128i eop_bits =
 442                                 _mm_packus_epi32
 443                                         (_mm256_castsi256_si128(eop_bits256),
 444                                          _mm256_extractf128_si256(eop_bits256,
 445                                                                   1));
 446                         /**
 447                          * flip bits, and mask out the EOP bit, which is now
 448                          * a split-packet bit i.e. !EOP, rather than EOP one.
 449                          */
 450                         __m128i split_bits = _mm_andnot_si128(eop_bits,
 451                                         eop_mask);
 452                         /**
 453                          * eop bits are out of order, so we need to shuffle them
 454                          * back into order again. In doing so, only use low 8
 455                          * bits, which acts like another pack instruction
 456                          * The original order is (hi->lo): 1,3,5,7,0,2,4,6
 457                          * [Since we use epi8, the 16-bit positions are
 458                          * multiplied by 2 in the eop_shuffle value.]
 459                          */
 460                         __m128i eop_shuffle =
 461                                 _mm_set_epi8(/* zero hi 64b */
 462                                              0xFF, 0xFF, 0xFF, 0xFF,
 463                                              0xFF, 0xFF, 0xFF, 0xFF,
 464                                              /* move values to lo 64b */
 465                                              8, 0, 10, 2,
 466                                              12, 4, 14, 6);
 467                         split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
 468                         *(uint64_t *)split_packet =
 469                                 _mm_cvtsi128_si64(split_bits);
 470                         split_packet += IAVF_DESCS_PER_LOOP_AVX;
 471                 }
 472
 473                 /* perform dd_check */
 474                 status0_7 = _mm256_and_si256(status0_7, dd_check);
 475                 status0_7 = _mm256_packs_epi32(status0_7,
 476                                                _mm256_setzero_si256());
 477
 478                 uint64_t burst = __builtin_popcountll
 479                                         (_mm_cvtsi128_si64
 480                                                 (_mm256_extracti128_si256
 481                                                         (status0_7, 1)));
 482                 burst += __builtin_popcountll
 483                                 (_mm_cvtsi128_si64
 484                                         (_mm256_castsi256_si128(status0_7)));
 485                 received += burst;
 486                 if (burst != IAVF_DESCS_PER_LOOP_AVX)
 487                         break;
 488         }
 489
 490         /* update tail pointers */
 491         rxq->rx_tail += received;
 492         rxq->rx_tail &= (rxq->nb_rx_desc - 1);
 493         if ((rxq->rx_tail & 1) == 1 && received > 1) { /* keep avx2 aligned */
 494                 rxq->rx_tail--;
 495                 received--;
 496         }
 497         rxq->rxrearm_nb += received;
 498         return received;
 499 }
 500
 501 static inline __m256i
 502 flex_rxd_to_fdir_flags_vec_avx2(const __m256i fdir_id0_7)
 503 {
 504 #define FDID_MIS_MAGIC 0xFFFFFFFF
 505         RTE_BUILD_BUG_ON(PKT_RX_FDIR != (1 << 2));
 506         RTE_BUILD_BUG_ON(PKT_RX_FDIR_ID != (1 << 13));
 507         const __m256i pkt_fdir_bit = _mm256_set1_epi32(PKT_RX_FDIR |
 508                         PKT_RX_FDIR_ID);
 509         /* desc->flow_id field == 0xFFFFFFFF means fdir mismatch */
 510         const __m256i fdir_mis_mask = _mm256_set1_epi32(FDID_MIS_MAGIC);
 511         __m256i fdir_mask = _mm256_cmpeq_epi32(fdir_id0_7,
 512                         fdir_mis_mask);
 513         /* this XOR op results to bit-reverse the fdir_mask */
 514         fdir_mask = _mm256_xor_si256(fdir_mask, fdir_mis_mask);
 515         const __m256i fdir_flags = _mm256_and_si256(fdir_mask, pkt_fdir_bit);
 516
 517         return fdir_flags;
 518 }
 519
 520 static inline uint16_t
 521 _iavf_recv_raw_pkts_vec_avx2_flex_rxd(struct iavf_rx_queue *rxq,
 522                                       struct rte_mbuf **rx_pkts,
 523                                       uint16_t nb_pkts, uint8_t *split_packet)
 524 {
 525 #define IAVF_DESCS_PER_LOOP_AVX 8
 526
 527         const uint32_t *type_table = rxq->vsi->adapter->ptype_tbl;
 528
 529         const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
 530                         0, rxq->mbuf_initializer);
 531         struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
 532         volatile union iavf_rx_flex_desc *rxdp =
 533                 (union iavf_rx_flex_desc *)rxq->rx_ring + rxq->rx_tail;
 534
 535         rte_prefetch0(rxdp);
 536
 537         /* nb_pkts has to be floor-aligned to IAVF_DESCS_PER_LOOP_AVX */
 538         nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_DESCS_PER_LOOP_AVX);
 539
 540         /* See if we need to rearm the RX queue - gives the prefetch a bit
 541          * of time to act
 542          */
 543         if (rxq->rxrearm_nb > IAVF_RXQ_REARM_THRESH)
 544                 iavf_rxq_rearm(rxq);
 545
 546         /* Before we start moving massive data around, check to see if
 547          * there is actually a packet available
 548          */
 549         if (!(rxdp->wb.status_error0 &
 550                         rte_cpu_to_le_32(1 << IAVF_RX_FLEX_DESC_STATUS0_DD_S)))
 551                 return 0;
 552
 553         /* constants used in processing loop */
 554         const __m256i crc_adjust =
 555                 _mm256_set_epi16
 556                         (/* first descriptor */
 557                          0, 0, 0,       /* ignore non-length fields */
 558                          -rxq->crc_len, /* sub crc on data_len */
 559                          0,             /* ignore high-16bits of pkt_len */
 560                          -rxq->crc_len, /* sub crc on pkt_len */
 561                          0, 0,          /* ignore pkt_type field */
 562                          /* second descriptor */
 563                          0, 0, 0,       /* ignore non-length fields */
 564                          -rxq->crc_len, /* sub crc on data_len */
 565                          0,             /* ignore high-16bits of pkt_len */
 566                          -rxq->crc_len, /* sub crc on pkt_len */
 567                          0, 0           /* ignore pkt_type field */
 568                         );
 569
 570         /* 8 packets DD mask, LSB in each 32-bit value */
 571         const __m256i dd_check = _mm256_set1_epi32(1);
 572
 573         /* 8 packets EOP mask, second-LSB in each 32-bit value */
 574         const __m256i eop_check = _mm256_slli_epi32(dd_check,
 575                         IAVF_RX_FLEX_DESC_STATUS0_EOF_S);
 576
 577         /* mask to shuffle from desc. to mbuf (2 descriptors)*/
 578         const __m256i shuf_msk =
 579                 _mm256_set_epi8
 580                         (/* first descriptor */
 581                          0xFF, 0xFF,
 582                          0xFF, 0xFF,    /* rss hash parsed separately */
 583                          11, 10,        /* octet 10~11, 16 bits vlan_macip */
 584                          5, 4,          /* octet 4~5, 16 bits data_len */
 585                          0xFF, 0xFF,    /* skip hi 16 bits pkt_len, zero out */
 586                          5, 4,          /* octet 4~5, 16 bits pkt_len */
 587                          0xFF, 0xFF,    /* pkt_type set as unknown */
 588                          0xFF, 0xFF,    /*pkt_type set as unknown */
 589                          /* second descriptor */
 590                          0xFF, 0xFF,
 591                          0xFF, 0xFF,    /* rss hash parsed separately */
 592                          11, 10,        /* octet 10~11, 16 bits vlan_macip */
 593                          5, 4,          /* octet 4~5, 16 bits data_len */
 594                          0xFF, 0xFF,    /* skip hi 16 bits pkt_len, zero out */
 595                          5, 4,          /* octet 4~5, 16 bits pkt_len */
 596                          0xFF, 0xFF,    /* pkt_type set as unknown */
 597                          0xFF, 0xFF     /*pkt_type set as unknown */
 598                         );
 599         /**
 600          * compile-time check the above crc and shuffle layout is correct.
 601          * NOTE: the first field (lowest address) is given last in set_epi
 602          * calls above.
 603          */
 604         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
 605                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
 606         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
 607                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
 608         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, vlan_tci) !=
 609                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 10);
 610         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) !=
 611                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
 612
 613         /* Status/Error flag masks */
 614         /**
 615          * mask everything except Checksum Reports, RSS indication
 616          * and VLAN indication.
 617          * bit6:4 for IP/L4 checksum errors.
 618          * bit12 is for RSS indication.
 619          * bit13 is for VLAN indication.
 620          */
 621         const __m256i flags_mask =
 622                  _mm256_set1_epi32((7 << 4) | (1 << 12) | (1 << 13));
 623         /**
 624          * data to be shuffled by the result of the flags mask shifted by 4
 625          * bits.  This gives use the l3_l4 flags.
 626          */
 627         const __m256i l3_l4_flags_shuf = _mm256_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
 628                         /* shift right 1 bit to make sure it not exceed 255 */
 629                         (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
 630                          PKT_RX_IP_CKSUM_BAD) >> 1,
 631                         (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
 632                          PKT_RX_IP_CKSUM_GOOD) >> 1,
 633                         (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
 634                          PKT_RX_IP_CKSUM_BAD) >> 1,
 635                         (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
 636                          PKT_RX_IP_CKSUM_GOOD) >> 1,
 637                         (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
 638                         (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
 639                         (PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
 640                         (PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_GOOD) >> 1,
 641                         /* second 128-bits */
 642                         0, 0, 0, 0, 0, 0, 0, 0,
 643                         (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
 644                          PKT_RX_IP_CKSUM_BAD) >> 1,
 645                         (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
 646                          PKT_RX_IP_CKSUM_GOOD) >> 1,
 647                         (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
 648                          PKT_RX_IP_CKSUM_BAD) >> 1,
 649                         (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
 650                          PKT_RX_IP_CKSUM_GOOD) >> 1,
 651                         (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
 652                         (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
 653                         (PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
 654                         (PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_GOOD) >> 1);
 655         const __m256i cksum_mask =
 656                  _mm256_set1_epi32(PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
 657                                    PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
 658                                    PKT_RX_OUTER_IP_CKSUM_BAD);
 659         /**
 660          * data to be shuffled by result of flag mask, shifted down 12.
 661          * If RSS(bit12)/VLAN(bit13) are set,
 662          * shuffle moves appropriate flags in place.
 663          */
 664         const __m256i rss_flags_shuf = _mm256_set_epi8(0, 0, 0, 0,
 665                         0, 0, 0, 0,
 666                         0, 0, 0, 0,
 667                         PKT_RX_RSS_HASH, 0,
 668                         PKT_RX_RSS_HASH, 0,
 669                         /* end up 128-bits */
 670                         0, 0, 0, 0,
 671                         0, 0, 0, 0,
 672                         0, 0, 0, 0,
 673                         PKT_RX_RSS_HASH, 0,
 674                         PKT_RX_RSS_HASH, 0);
 675
 676         const __m256i vlan_flags_shuf = _mm256_set_epi8(0, 0, 0, 0,
 677                         0, 0, 0, 0,
 678                         0, 0, 0, 0,
 679                         PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
 680                         PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
 681                         0, 0,
 682                         /* end up 128-bits */
 683                         0, 0, 0, 0,
 684                         0, 0, 0, 0,
 685                         0, 0, 0, 0,
 686                         PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
 687                         PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
 688                         0, 0);
 689
 690         uint16_t i, received;
 691
 692         for (i = 0, received = 0; i < nb_pkts;
 693              i += IAVF_DESCS_PER_LOOP_AVX,
 694              rxdp += IAVF_DESCS_PER_LOOP_AVX) {
 695                 /* step 1, copy over 8 mbuf pointers to rx_pkts array */
 696                 _mm256_storeu_si256((void *)&rx_pkts[i],
 697                                     _mm256_loadu_si256((void *)&sw_ring[i]));
 698 #ifdef RTE_ARCH_X86_64
 699                 _mm256_storeu_si256
 700                         ((void *)&rx_pkts[i + 4],
 701                          _mm256_loadu_si256((void *)&sw_ring[i + 4]));
 702 #endif
 703
 704                 __m256i raw_desc0_1, raw_desc2_3, raw_desc4_5, raw_desc6_7;
 705
 706                 const __m128i raw_desc7 =
 707                         _mm_load_si128((void *)(rxdp + 7));
 708                 rte_compiler_barrier();
 709                 const __m128i raw_desc6 =
 710                         _mm_load_si128((void *)(rxdp + 6));
 711                 rte_compiler_barrier();
 712                 const __m128i raw_desc5 =
 713                         _mm_load_si128((void *)(rxdp + 5));
 714                 rte_compiler_barrier();
 715                 const __m128i raw_desc4 =
 716                         _mm_load_si128((void *)(rxdp + 4));
 717                 rte_compiler_barrier();
 718                 const __m128i raw_desc3 =
 719                         _mm_load_si128((void *)(rxdp + 3));
 720                 rte_compiler_barrier();
 721                 const __m128i raw_desc2 =
 722                         _mm_load_si128((void *)(rxdp + 2));
 723                 rte_compiler_barrier();
 724                 const __m128i raw_desc1 =
 725                         _mm_load_si128((void *)(rxdp + 1));
 726                 rte_compiler_barrier();
 727                 const __m128i raw_desc0 =
 728                         _mm_load_si128((void *)(rxdp + 0));
 729
 730                 raw_desc6_7 =
 731                         _mm256_inserti128_si256
 732                                 (_mm256_castsi128_si256(raw_desc6),
 733                                  raw_desc7, 1);
 734                 raw_desc4_5 =
 735                         _mm256_inserti128_si256
 736                                 (_mm256_castsi128_si256(raw_desc4),
 737                                  raw_desc5, 1);
 738                 raw_desc2_3 =
 739                         _mm256_inserti128_si256
 740                                 (_mm256_castsi128_si256(raw_desc2),
 741                                  raw_desc3, 1);
 742                 raw_desc0_1 =
 743                         _mm256_inserti128_si256
 744                                 (_mm256_castsi128_si256(raw_desc0),
 745                                  raw_desc1, 1);
 746
 747                 if (split_packet) {
 748                         int j;
 749
 750                         for (j = 0; j < IAVF_DESCS_PER_LOOP_AVX; j++)
 751                                 rte_mbuf_prefetch_part2(rx_pkts[i + j]);
 752                 }
 753
 754                 /**
 755                  * convert descriptors 4-7 into mbufs, re-arrange fields.
 756                  * Then write into the mbuf.
 757                  */
 758                 __m256i mb6_7 = _mm256_shuffle_epi8(raw_desc6_7, shuf_msk);
 759                 __m256i mb4_5 = _mm256_shuffle_epi8(raw_desc4_5, shuf_msk);
 760
 761                 mb6_7 = _mm256_add_epi16(mb6_7, crc_adjust);
 762                 mb4_5 = _mm256_add_epi16(mb4_5, crc_adjust);
 763                 /**
 764                  * to get packet types, ptype is located in bit16-25
 765                  * of each 128bits
 766                  */
 767                 const __m256i ptype_mask =
 768                         _mm256_set1_epi16(IAVF_RX_FLEX_DESC_PTYPE_M);
 769                 const __m256i ptypes6_7 =
 770                         _mm256_and_si256(raw_desc6_7, ptype_mask);
 771                 const __m256i ptypes4_5 =
 772                         _mm256_and_si256(raw_desc4_5, ptype_mask);
 773                 const uint16_t ptype7 = _mm256_extract_epi16(ptypes6_7, 9);
 774                 const uint16_t ptype6 = _mm256_extract_epi16(ptypes6_7, 1);
 775                 const uint16_t ptype5 = _mm256_extract_epi16(ptypes4_5, 9);
 776                 const uint16_t ptype4 = _mm256_extract_epi16(ptypes4_5, 1);
 777
 778                 mb6_7 = _mm256_insert_epi32(mb6_7, type_table[ptype7], 4);
 779                 mb6_7 = _mm256_insert_epi32(mb6_7, type_table[ptype6], 0);
 780                 mb4_5 = _mm256_insert_epi32(mb4_5, type_table[ptype5], 4);
 781                 mb4_5 = _mm256_insert_epi32(mb4_5, type_table[ptype4], 0);
 782                 /* merge the status bits into one register */
 783                 const __m256i status4_7 = _mm256_unpackhi_epi32(raw_desc6_7,
 784                                 raw_desc4_5);
 785
 786                 /**
 787                  * convert descriptors 0-3 into mbufs, re-arrange fields.
 788                  * Then write into the mbuf.
 789                  */
 790                 __m256i mb2_3 = _mm256_shuffle_epi8(raw_desc2_3, shuf_msk);
 791                 __m256i mb0_1 = _mm256_shuffle_epi8(raw_desc0_1, shuf_msk);
 792
 793                 mb2_3 = _mm256_add_epi16(mb2_3, crc_adjust);
 794                 mb0_1 = _mm256_add_epi16(mb0_1, crc_adjust);
 795                 /**
 796                  * to get packet types, ptype is located in bit16-25
 797                  * of each 128bits
 798                  */
 799                 const __m256i ptypes2_3 =
 800                         _mm256_and_si256(raw_desc2_3, ptype_mask);
 801                 const __m256i ptypes0_1 =
 802                         _mm256_and_si256(raw_desc0_1, ptype_mask);
 803                 const uint16_t ptype3 = _mm256_extract_epi16(ptypes2_3, 9);
 804                 const uint16_t ptype2 = _mm256_extract_epi16(ptypes2_3, 1);
 805                 const uint16_t ptype1 = _mm256_extract_epi16(ptypes0_1, 9);
 806                 const uint16_t ptype0 = _mm256_extract_epi16(ptypes0_1, 1);
 807
 808                 mb2_3 = _mm256_insert_epi32(mb2_3, type_table[ptype3], 4);
 809                 mb2_3 = _mm256_insert_epi32(mb2_3, type_table[ptype2], 0);
 810                 mb0_1 = _mm256_insert_epi32(mb0_1, type_table[ptype1], 4);
 811                 mb0_1 = _mm256_insert_epi32(mb0_1, type_table[ptype0], 0);
 812                 /* merge the status bits into one register */
 813                 const __m256i status0_3 = _mm256_unpackhi_epi32(raw_desc2_3,
 814                                                                 raw_desc0_1);
 815
 816                 /**
 817                  * take the two sets of status bits and merge to one
 818                  * After merge, the packets status flags are in the
 819                  * order (hi->lo): [1, 3, 5, 7, 0, 2, 4, 6]
 820                  */
 821                 __m256i status0_7 = _mm256_unpacklo_epi64(status4_7,
 822                                                           status0_3);
 823
 824                 /* now do flag manipulation */
 825
 826                 /* get only flag/error bits we want */
 827                 const __m256i flag_bits =
 828                         _mm256_and_si256(status0_7, flags_mask);
 829                 /**
 830                  * l3_l4_error flags, shuffle, then shift to correct adjustment
 831                  * of flags in flags_shuf, and finally mask out extra bits
 832                  */
 833                 __m256i l3_l4_flags = _mm256_shuffle_epi8(l3_l4_flags_shuf,
 834                                 _mm256_srli_epi32(flag_bits, 4));
 835                 l3_l4_flags = _mm256_slli_epi32(l3_l4_flags, 1);
 836                 l3_l4_flags = _mm256_and_si256(l3_l4_flags, cksum_mask);
 837
 838                 /* set rss and vlan flags */
 839                 const __m256i rss_vlan_flag_bits =
 840                         _mm256_srli_epi32(flag_bits, 12);
 841                 const __m256i rss_flags =
 842                         _mm256_shuffle_epi8(rss_flags_shuf,
 843                                             rss_vlan_flag_bits);
 844
 845                 __m256i vlan_flags = _mm256_setzero_si256();
 846
 847                 if (rxq->rx_flags == IAVF_RX_FLAGS_VLAN_TAG_LOC_L2TAG1)
 848                         vlan_flags =
 849                                 _mm256_shuffle_epi8(vlan_flags_shuf,
 850                                                     rss_vlan_flag_bits);
 851
 852                 const __m256i rss_vlan_flags =
 853                         _mm256_or_si256(rss_flags, vlan_flags);
 854
 855                 /* merge flags */
 856                 __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
 857                                 rss_vlan_flags);
 858
 859                 if (rxq->fdir_enabled) {
 860                         const __m256i fdir_id4_7 =
 861                                 _mm256_unpackhi_epi32(raw_desc6_7, raw_desc4_5);
 862
 863                         const __m256i fdir_id0_3 =
 864                                 _mm256_unpackhi_epi32(raw_desc2_3, raw_desc0_1);
 865
 866                         const __m256i fdir_id0_7 =
 867                                 _mm256_unpackhi_epi64(fdir_id4_7, fdir_id0_3);
 868
 869                         const __m256i fdir_flags =
 870                                 flex_rxd_to_fdir_flags_vec_avx2(fdir_id0_7);
 871
 872                         /* merge with fdir_flags */
 873                         mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_flags);
 874
 875                         /* write to mbuf: have to use scalar store here */
 876                         rx_pkts[i + 0]->hash.fdir.hi =
 877                                 _mm256_extract_epi32(fdir_id0_7, 3);
 878
 879                         rx_pkts[i + 1]->hash.fdir.hi =
 880                                 _mm256_extract_epi32(fdir_id0_7, 7);
 881
 882                         rx_pkts[i + 2]->hash.fdir.hi =
 883                                 _mm256_extract_epi32(fdir_id0_7, 2);
 884
 885                         rx_pkts[i + 3]->hash.fdir.hi =
 886                                 _mm256_extract_epi32(fdir_id0_7, 6);
 887
 888                         rx_pkts[i + 4]->hash.fdir.hi =
 889                                 _mm256_extract_epi32(fdir_id0_7, 1);
 890
 891                         rx_pkts[i + 5]->hash.fdir.hi =
 892                                 _mm256_extract_epi32(fdir_id0_7, 5);
 893
 894                         rx_pkts[i + 6]->hash.fdir.hi =
 895                                 _mm256_extract_epi32(fdir_id0_7, 0);
 896
 897                         rx_pkts[i + 7]->hash.fdir.hi =
 898                                 _mm256_extract_epi32(fdir_id0_7, 4);
 899                 } /* if() on fdir_enabled */
 900
 901 #ifndef RTE_LIBRTE_IAVF_16BYTE_RX_DESC
 902                 /**
 903                  * needs to load 2nd 16B of each desc for RSS hash parsing,
 904                  * will cause performance drop to get into this context.
 905                  */
 906                 if (rxq->vsi->adapter->eth_dev->data->dev_conf.rxmode.offloads &
 907                                 DEV_RX_OFFLOAD_RSS_HASH ||
 908                                 rxq->rx_flags & IAVF_RX_FLAGS_VLAN_TAG_LOC_L2TAG2_2) {
 909                         /* load bottom half of every 32B desc */
 910                         const __m128i raw_desc_bh7 =
 911                                 _mm_load_si128
 912                                         ((void *)(&rxdp[7].wb.status_error1));
 913                         rte_compiler_barrier();
 914                         const __m128i raw_desc_bh6 =
 915                                 _mm_load_si128
 916                                         ((void *)(&rxdp[6].wb.status_error1));
 917                         rte_compiler_barrier();
 918                         const __m128i raw_desc_bh5 =
 919                                 _mm_load_si128
 920                                         ((void *)(&rxdp[5].wb.status_error1));
 921                         rte_compiler_barrier();
 922                         const __m128i raw_desc_bh4 =
 923                                 _mm_load_si128
 924                                         ((void *)(&rxdp[4].wb.status_error1));
 925                         rte_compiler_barrier();
 926                         const __m128i raw_desc_bh3 =
 927                                 _mm_load_si128
 928                                         ((void *)(&rxdp[3].wb.status_error1));
 929                         rte_compiler_barrier();
 930                         const __m128i raw_desc_bh2 =
 931                                 _mm_load_si128
 932                                         ((void *)(&rxdp[2].wb.status_error1));
 933                         rte_compiler_barrier();
 934                         const __m128i raw_desc_bh1 =
 935                                 _mm_load_si128
 936                                         ((void *)(&rxdp[1].wb.status_error1));
 937                         rte_compiler_barrier();
 938                         const __m128i raw_desc_bh0 =
 939                                 _mm_load_si128
 940                                         ((void *)(&rxdp[0].wb.status_error1));
 941
 942                         __m256i raw_desc_bh6_7 =
 943                                 _mm256_inserti128_si256
 944                                         (_mm256_castsi128_si256(raw_desc_bh6),
 945                                         raw_desc_bh7, 1);
 946                         __m256i raw_desc_bh4_5 =
 947                                 _mm256_inserti128_si256
 948                                         (_mm256_castsi128_si256(raw_desc_bh4),
 949                                         raw_desc_bh5, 1);
 950                         __m256i raw_desc_bh2_3 =
 951                                 _mm256_inserti128_si256
 952                                         (_mm256_castsi128_si256(raw_desc_bh2),
 953                                         raw_desc_bh3, 1);
 954                         __m256i raw_desc_bh0_1 =
 955                                 _mm256_inserti128_si256
 956                                         (_mm256_castsi128_si256(raw_desc_bh0),
 957                                         raw_desc_bh1, 1);
 958
 959                         if (rxq->vsi->adapter->eth_dev->data->dev_conf.rxmode.offloads &
 960                                         DEV_RX_OFFLOAD_RSS_HASH) {
 961                                 /**
 962                                  * to shift the 32b RSS hash value to the
 963                                  * highest 32b of each 128b before mask
 964                                  */
 965                                 __m256i rss_hash6_7 =
 966                                         _mm256_slli_epi64(raw_desc_bh6_7, 32);
 967                                 __m256i rss_hash4_5 =
 968                                         _mm256_slli_epi64(raw_desc_bh4_5, 32);
 969                                 __m256i rss_hash2_3 =
 970                                         _mm256_slli_epi64(raw_desc_bh2_3, 32);
 971                                 __m256i rss_hash0_1 =
 972                                         _mm256_slli_epi64(raw_desc_bh0_1, 32);
 973
 974                                 const __m256i rss_hash_msk =
 975                                         _mm256_set_epi32(0xFFFFFFFF, 0, 0, 0,
 976                                                          0xFFFFFFFF, 0, 0, 0);
 977
 978                                 rss_hash6_7 = _mm256_and_si256
 979                                                 (rss_hash6_7, rss_hash_msk);
 980                                 rss_hash4_5 = _mm256_and_si256
 981                                                 (rss_hash4_5, rss_hash_msk);
 982                                 rss_hash2_3 = _mm256_and_si256
 983                                                 (rss_hash2_3, rss_hash_msk);
 984                                 rss_hash0_1 = _mm256_and_si256
 985                                                 (rss_hash0_1, rss_hash_msk);
 986
 987                                 mb6_7 = _mm256_or_si256(mb6_7, rss_hash6_7);
 988                                 mb4_5 = _mm256_or_si256(mb4_5, rss_hash4_5);
 989                                 mb2_3 = _mm256_or_si256(mb2_3, rss_hash2_3);
 990                                 mb0_1 = _mm256_or_si256(mb0_1, rss_hash0_1);
 991                         }
 992
 993                         if (rxq->rx_flags & IAVF_RX_FLAGS_VLAN_TAG_LOC_L2TAG2_2) {
 994                                 /* merge the status/error-1 bits into one register */
 995                                 const __m256i status1_4_7 =
 996                                         _mm256_unpacklo_epi32(raw_desc_bh6_7,
 997                                                               raw_desc_bh4_5);
 998                                 const __m256i status1_0_3 =
 999                                         _mm256_unpacklo_epi32(raw_desc_bh2_3,
1000                                                               raw_desc_bh0_1);
1001
1002                                 const __m256i status1_0_7 =
1003                                         _mm256_unpacklo_epi64(status1_4_7,
1004                                                               status1_0_3);
1005
1006                                 const __m256i l2tag2p_flag_mask =
1007                                         _mm256_set1_epi32
1008                                         (1 << IAVF_RX_FLEX_DESC_STATUS1_L2TAG2P_S);
1009
1010                                 __m256i l2tag2p_flag_bits =
1011                                         _mm256_and_si256
1012                                         (status1_0_7, l2tag2p_flag_mask);
1013
1014                                 l2tag2p_flag_bits =
1015                                         _mm256_srli_epi32(l2tag2p_flag_bits,
1016                                                 IAVF_RX_FLEX_DESC_STATUS1_L2TAG2P_S);
1017
1018                                 const __m256i l2tag2_flags_shuf =
1019                                         _mm256_set_epi8(0, 0, 0, 0,
1020                                                         0, 0, 0, 0,
1021                                                         0, 0, 0, 0,
1022                                                         0, 0, 0, 0,
1023                                                         /* end up 128-bits */
1024                                                         0, 0, 0, 0,
1025                                                         0, 0, 0, 0,
1026                                                         0, 0, 0, 0,
1027                                                         0, 0,
1028                                                         PKT_RX_VLAN |
1029                                                         PKT_RX_VLAN_STRIPPED,
1030                                                         0);
1031
1032                                 vlan_flags =
1033                                         _mm256_shuffle_epi8(l2tag2_flags_shuf,
1034                                                             l2tag2p_flag_bits);
1035
1036                                 /* merge with vlan_flags */
1037                                 mbuf_flags = _mm256_or_si256
1038                                                 (mbuf_flags, vlan_flags);
1039
1040                                 /* L2TAG2_2 */
1041                                 __m256i vlan_tci6_7 =
1042                                         _mm256_slli_si256(raw_desc_bh6_7, 4);
1043                                 __m256i vlan_tci4_5 =
1044                                         _mm256_slli_si256(raw_desc_bh4_5, 4);
1045                                 __m256i vlan_tci2_3 =
1046                                         _mm256_slli_si256(raw_desc_bh2_3, 4);
1047                                 __m256i vlan_tci0_1 =
1048                                         _mm256_slli_si256(raw_desc_bh0_1, 4);
1049
1050                                 const __m256i vlan_tci_msk =
1051                                         _mm256_set_epi32(0, 0xFFFF0000, 0, 0,
1052                                                          0, 0xFFFF0000, 0, 0);
1053
1054                                 vlan_tci6_7 = _mm256_and_si256
1055                                                 (vlan_tci6_7, vlan_tci_msk);
1056                                 vlan_tci4_5 = _mm256_and_si256
1057                                                 (vlan_tci4_5, vlan_tci_msk);
1058                                 vlan_tci2_3 = _mm256_and_si256
1059                                                 (vlan_tci2_3, vlan_tci_msk);
1060                                 vlan_tci0_1 = _mm256_and_si256
1061                                                 (vlan_tci0_1, vlan_tci_msk);
1062
1063                                 mb6_7 = _mm256_or_si256(mb6_7, vlan_tci6_7);
1064                                 mb4_5 = _mm256_or_si256(mb4_5, vlan_tci4_5);
1065                                 mb2_3 = _mm256_or_si256(mb2_3, vlan_tci2_3);
1066                                 mb0_1 = _mm256_or_si256(mb0_1, vlan_tci0_1);
1067                         }
1068                 } /* if() on RSS hash parsing */
1069 #endif
1070
1071                 /**
1072                  * At this point, we have the 8 sets of flags in the low 16-bits
1073                  * of each 32-bit value in vlan0.
1074                  * We want to extract these, and merge them with the mbuf init
1075                  * data so we can do a single write to the mbuf to set the flags
1076                  * and all the other initialization fields. Extracting the
1077                  * appropriate flags means that we have to do a shift and blend
1078                  * for each mbuf before we do the write. However, we can also
1079                  * add in the previously computed rx_descriptor fields to
1080                  * make a single 256-bit write per mbuf
1081                  */
1082                 /* check the structure matches expectations */
1083                 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, ol_flags) !=
1084                                  offsetof(struct rte_mbuf, rearm_data) + 8);
1085                 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, rearm_data) !=
1086                                  RTE_ALIGN(offsetof(struct rte_mbuf,
1087                                                     rearm_data),
1088                                            16));
1089                 /* build up data and do writes */
1090                 __m256i rearm0, rearm1, rearm2, rearm3, rearm4, rearm5,
1091                         rearm6, rearm7;
1092                 rearm6 = _mm256_blend_epi32(mbuf_init,
1093                                             _mm256_slli_si256(mbuf_flags, 8),
1094                                             0x04);
1095                 rearm4 = _mm256_blend_epi32(mbuf_init,
1096                                             _mm256_slli_si256(mbuf_flags, 4),
1097                                             0x04);
1098                 rearm2 = _mm256_blend_epi32(mbuf_init, mbuf_flags, 0x04);
1099                 rearm0 = _mm256_blend_epi32(mbuf_init,
1100                                             _mm256_srli_si256(mbuf_flags, 4),
1101                                             0x04);
1102                 /* permute to add in the rx_descriptor e.g. rss fields */
1103                 rearm6 = _mm256_permute2f128_si256(rearm6, mb6_7, 0x20);
1104                 rearm4 = _mm256_permute2f128_si256(rearm4, mb4_5, 0x20);
1105                 rearm2 = _mm256_permute2f128_si256(rearm2, mb2_3, 0x20);
1106                 rearm0 = _mm256_permute2f128_si256(rearm0, mb0_1, 0x20);
1107                 /* write to mbuf */
1108                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 6]->rearm_data,
1109                                     rearm6);
1110                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 4]->rearm_data,
1111                                     rearm4);
1112                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 2]->rearm_data,
1113                                     rearm2);
1114                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 0]->rearm_data,
1115                                     rearm0);
1116
1117                 /* repeat for the odd mbufs */
1118                 const __m256i odd_flags =
1119                         _mm256_castsi128_si256
1120                                 (_mm256_extracti128_si256(mbuf_flags, 1));
1121                 rearm7 = _mm256_blend_epi32(mbuf_init,
1122                                             _mm256_slli_si256(odd_flags, 8),
1123                                             0x04);
1124                 rearm5 = _mm256_blend_epi32(mbuf_init,
1125                                             _mm256_slli_si256(odd_flags, 4),
1126                                             0x04);
1127                 rearm3 = _mm256_blend_epi32(mbuf_init, odd_flags, 0x04);
1128                 rearm1 = _mm256_blend_epi32(mbuf_init,
1129                                             _mm256_srli_si256(odd_flags, 4),
1130                                             0x04);
1131                 /* since odd mbufs are already in hi 128-bits use blend */
1132                 rearm7 = _mm256_blend_epi32(rearm7, mb6_7, 0xF0);
1133                 rearm5 = _mm256_blend_epi32(rearm5, mb4_5, 0xF0);
1134                 rearm3 = _mm256_blend_epi32(rearm3, mb2_3, 0xF0);
1135                 rearm1 = _mm256_blend_epi32(rearm1, mb0_1, 0xF0);
1136                 /* again write to mbufs */
1137                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 7]->rearm_data,
1138                                     rearm7);
1139                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 5]->rearm_data,
1140                                     rearm5);
1141                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 3]->rearm_data,
1142                                     rearm3);
1143                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 1]->rearm_data,
1144                                     rearm1);
1145
1146                 /* extract and record EOP bit */
1147                 if (split_packet) {
1148                         const __m128i eop_mask =
1149                                 _mm_set1_epi16(1 <<
1150                                                IAVF_RX_FLEX_DESC_STATUS0_EOF_S);
1151                         const __m256i eop_bits256 = _mm256_and_si256(status0_7,
1152                                                                      eop_check);
1153                         /* pack status bits into a single 128-bit register */
1154                         const __m128i eop_bits =
1155                                 _mm_packus_epi32
1156                                         (_mm256_castsi256_si128(eop_bits256),
1157                                          _mm256_extractf128_si256(eop_bits256,
1158                                                                   1));
1159                         /**
1160                          * flip bits, and mask out the EOP bit, which is now
1161                          * a split-packet bit i.e. !EOP, rather than EOP one.
1162                          */
1163                         __m128i split_bits = _mm_andnot_si128(eop_bits,
1164                                         eop_mask);
1165                         /**
1166                          * eop bits are out of order, so we need to shuffle them
1167                          * back into order again. In doing so, only use low 8
1168                          * bits, which acts like another pack instruction
1169                          * The original order is (hi->lo): 1,3,5,7,0,2,4,6
1170                          * [Since we use epi8, the 16-bit positions are
1171                          * multiplied by 2 in the eop_shuffle value.]
1172                          */
1173                         __m128i eop_shuffle =
1174                                 _mm_set_epi8(/* zero hi 64b */
1175                                              0xFF, 0xFF, 0xFF, 0xFF,
1176                                              0xFF, 0xFF, 0xFF, 0xFF,
1177                                              /* move values to lo 64b */
1178                                              8, 0, 10, 2,
1179                                              12, 4, 14, 6);
1180                         split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
1181                         *(uint64_t *)split_packet =
1182                                 _mm_cvtsi128_si64(split_bits);
1183                         split_packet += IAVF_DESCS_PER_LOOP_AVX;
1184                 }
1185
1186                 /* perform dd_check */
1187                 status0_7 = _mm256_and_si256(status0_7, dd_check);
1188                 status0_7 = _mm256_packs_epi32(status0_7,
1189                                                _mm256_setzero_si256());
1190
1191                 uint64_t burst = __builtin_popcountll
1192                                         (_mm_cvtsi128_si64
1193                                                 (_mm256_extracti128_si256
1194                                                         (status0_7, 1)));
1195                 burst += __builtin_popcountll
1196                                 (_mm_cvtsi128_si64
1197                                         (_mm256_castsi256_si128(status0_7)));
1198                 received += burst;
1199                 if (burst != IAVF_DESCS_PER_LOOP_AVX)
1200                         break;
1201         }
1202
1203         /* update tail pointers */
1204         rxq->rx_tail += received;
1205         rxq->rx_tail &= (rxq->nb_rx_desc - 1);
1206         if ((rxq->rx_tail & 1) == 1 && received > 1) { /* keep avx2 aligned */
1207                 rxq->rx_tail--;
1208                 received--;
1209         }
1210         rxq->rxrearm_nb += received;
1211         return received;
1212 }
1213
1214 /**
1215  * Notice:
1216  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
1217  */
1218 uint16_t
1219 iavf_recv_pkts_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
1220                         uint16_t nb_pkts)
1221 {
1222         return _iavf_recv_raw_pkts_vec_avx2(rx_queue, rx_pkts, nb_pkts, NULL);
1223 }
1224
1225 /**
1226  * Notice:
1227  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
1228  */
1229 uint16_t
1230 iavf_recv_pkts_vec_avx2_flex_rxd(void *rx_queue, struct rte_mbuf **rx_pkts,
1231                                  uint16_t nb_pkts)
1232 {
1233         return _iavf_recv_raw_pkts_vec_avx2_flex_rxd(rx_queue, rx_pkts,
1234                                                      nb_pkts, NULL);
1235 }
1236
1237 /**
1238  * vPMD receive routine that reassembles single burst of 32 scattered packets
1239  * Notice:
1240  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
1241  */
1242 static uint16_t
1243 iavf_recv_scattered_burst_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
1244                                    uint16_t nb_pkts)
1245 {
1246         struct iavf_rx_queue *rxq = rx_queue;
1247         uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
1248
1249         /* get some new buffers */
1250         uint16_t nb_bufs = _iavf_recv_raw_pkts_vec_avx2(rxq, rx_pkts, nb_pkts,
1251                                                        split_flags);
1252         if (nb_bufs == 0)
1253                 return 0;
1254
1255         /* happy day case, full burst + no packets to be joined */
1256         const uint64_t *split_fl64 = (uint64_t *)split_flags;
1257
1258         if (!rxq->pkt_first_seg &&
1259             split_fl64[0] == 0 && split_fl64[1] == 0 &&
1260             split_fl64[2] == 0 && split_fl64[3] == 0)
1261                 return nb_bufs;
1262
1263         /* reassemble any packets that need reassembly*/
1264         unsigned int i = 0;
1265
1266         if (!rxq->pkt_first_seg) {
1267                 /* find the first split flag, and only reassemble then*/
1268                 while (i < nb_bufs && !split_flags[i])
1269                         i++;
1270                 if (i == nb_bufs)
1271                         return nb_bufs;
1272                 rxq->pkt_first_seg = rx_pkts[i];
1273         }
1274         return i + reassemble_packets(rxq, &rx_pkts[i], nb_bufs - i,
1275                                              &split_flags[i]);
1276 }
1277
1278 /**
1279  * vPMD receive routine that reassembles scattered packets.
1280  * Main receive routine that can handle arbitrary burst sizes
1281  * Notice:
1282  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
1283  */
1284 uint16_t
1285 iavf_recv_scattered_pkts_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
1286                                   uint16_t nb_pkts)
1287 {
1288         uint16_t retval = 0;
1289
1290         while (nb_pkts > IAVF_VPMD_RX_MAX_BURST) {
1291                 uint16_t burst = iavf_recv_scattered_burst_vec_avx2(rx_queue,
1292                                 rx_pkts + retval, IAVF_VPMD_RX_MAX_BURST);
1293                 retval += burst;
1294                 nb_pkts -= burst;
1295                 if (burst < IAVF_VPMD_RX_MAX_BURST)
1296                         return retval;
1297         }
1298         return retval + iavf_recv_scattered_burst_vec_avx2(rx_queue,
1299                                 rx_pkts + retval, nb_pkts);
1300 }
1301
1302 /**
1303  * vPMD receive routine that reassembles single burst of
1304  * 32 scattered packets for flex RxD
1305  * Notice:
1306  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
1307  */
1308 static uint16_t
1309 iavf_recv_scattered_burst_vec_avx2_flex_rxd(void *rx_queue,
1310                                             struct rte_mbuf **rx_pkts,
1311                                             uint16_t nb_pkts)
1312 {
1313         struct iavf_rx_queue *rxq = rx_queue;
1314         uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
1315
1316         /* get some new buffers */
1317         uint16_t nb_bufs = _iavf_recv_raw_pkts_vec_avx2_flex_rxd(rxq,
1318                                         rx_pkts, nb_pkts, split_flags);
1319         if (nb_bufs == 0)
1320                 return 0;
1321
1322         /* happy day case, full burst + no packets to be joined */
1323         const uint64_t *split_fl64 = (uint64_t *)split_flags;
1324
1325         if (!rxq->pkt_first_seg &&
1326             split_fl64[0] == 0 && split_fl64[1] == 0 &&
1327             split_fl64[2] == 0 && split_fl64[3] == 0)
1328                 return nb_bufs;
1329
1330         /* reassemble any packets that need reassembly*/
1331         unsigned int i = 0;
1332
1333         if (!rxq->pkt_first_seg) {
1334                 /* find the first split flag, and only reassemble then*/
1335                 while (i < nb_bufs && !split_flags[i])
1336                         i++;
1337                 if (i == nb_bufs)
1338                         return nb_bufs;
1339                 rxq->pkt_first_seg = rx_pkts[i];
1340         }
1341         return i + reassemble_packets(rxq, &rx_pkts[i], nb_bufs - i,
1342                                              &split_flags[i]);
1343 }
1344
1345 /**
1346  * vPMD receive routine that reassembles scattered packets for flex RxD.
1347  * Main receive routine that can handle arbitrary burst sizes
1348  * Notice:
1349  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
1350  */
1351 uint16_t
1352 iavf_recv_scattered_pkts_vec_avx2_flex_rxd(void *rx_queue,
1353                                            struct rte_mbuf **rx_pkts,
1354                                            uint16_t nb_pkts)
1355 {
1356         uint16_t retval = 0;
1357
1358         while (nb_pkts > IAVF_VPMD_RX_MAX_BURST) {
1359                 uint16_t burst =
1360                         iavf_recv_scattered_burst_vec_avx2_flex_rxd
1361                         (rx_queue, rx_pkts + retval, IAVF_VPMD_RX_MAX_BURST);
1362                 retval += burst;
1363                 nb_pkts -= burst;
1364                 if (burst < IAVF_VPMD_RX_MAX_BURST)
1365                         return retval;
1366         }
1367         return retval + iavf_recv_scattered_burst_vec_avx2_flex_rxd(rx_queue,
1368                                 rx_pkts + retval, nb_pkts);
1369 }
1370
1371 static inline void
1372 iavf_vtx1(volatile struct iavf_tx_desc *txdp,
1373           struct rte_mbuf *pkt, uint64_t flags)
1374 {
1375         uint64_t high_qw =
1376                 (IAVF_TX_DESC_DTYPE_DATA |
1377                  ((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT) |
1378                  ((uint64_t)pkt->data_len << IAVF_TXD_QW1_TX_BUF_SZ_SHIFT));
1379
1380         __m128i descriptor = _mm_set_epi64x(high_qw,
1381                                 pkt->buf_iova + pkt->data_off);
1382         _mm_store_si128((__m128i *)txdp, descriptor);
1383 }
1384
1385 static inline void
1386 iavf_vtx(volatile struct iavf_tx_desc *txdp,
1387          struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
1388 {
1389         const uint64_t hi_qw_tmpl = (IAVF_TX_DESC_DTYPE_DATA |
1390                         ((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT));
1391
1392         /* if unaligned on 32-bit boundary, do one to align */
1393         if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
1394                 iavf_vtx1(txdp, *pkt, flags);
1395                 nb_pkts--, txdp++, pkt++;
1396         }
1397
1398         /* do two at a time while possible, in bursts */
1399         for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
1400                 uint64_t hi_qw3 =
1401                         hi_qw_tmpl |
1402                         ((uint64_t)pkt[3]->data_len <<
1403                          IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
1404                 uint64_t hi_qw2 =
1405                         hi_qw_tmpl |
1406                         ((uint64_t)pkt[2]->data_len <<
1407                          IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
1408                 uint64_t hi_qw1 =
1409                         hi_qw_tmpl |
1410                         ((uint64_t)pkt[1]->data_len <<
1411                          IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
1412                 uint64_t hi_qw0 =
1413                         hi_qw_tmpl |
1414                         ((uint64_t)pkt[0]->data_len <<
1415                          IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
1416
1417                 __m256i desc2_3 =
1418                         _mm256_set_epi64x
1419                                 (hi_qw3,
1420                                  pkt[3]->buf_iova + pkt[3]->data_off,
1421                                  hi_qw2,
1422                                  pkt[2]->buf_iova + pkt[2]->data_off);
1423                 __m256i desc0_1 =
1424                         _mm256_set_epi64x
1425                                 (hi_qw1,
1426                                  pkt[1]->buf_iova + pkt[1]->data_off,
1427                                  hi_qw0,
1428                                  pkt[0]->buf_iova + pkt[0]->data_off);
1429                 _mm256_store_si256((void *)(txdp + 2), desc2_3);
1430                 _mm256_store_si256((void *)txdp, desc0_1);
1431         }
1432
1433         /* do any last ones */
1434         while (nb_pkts) {
1435                 iavf_vtx1(txdp, *pkt, flags);
1436                 txdp++, pkt++, nb_pkts--;
1437         }
1438 }
1439
1440 static inline uint16_t
1441 iavf_xmit_fixed_burst_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
1442                                uint16_t nb_pkts)
1443 {
1444         struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
1445         volatile struct iavf_tx_desc *txdp;
1446         struct iavf_tx_entry *txep;
1447         uint16_t n, nb_commit, tx_id;
1448         /* bit2 is reserved and must be set to 1 according to Spec */
1449         uint64_t flags = IAVF_TX_DESC_CMD_EOP | IAVF_TX_DESC_CMD_ICRC;
1450         uint64_t rs = IAVF_TX_DESC_CMD_RS | flags;
1451
1452         /* cross rx_thresh boundary is not allowed */
1453         nb_pkts = RTE_MIN(nb_pkts, txq->rs_thresh);
1454
1455         if (txq->nb_free < txq->free_thresh)
1456                 iavf_tx_free_bufs(txq);
1457
1458         nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_free, nb_pkts);
1459         if (unlikely(nb_pkts == 0))
1460                 return 0;
1461
1462         tx_id = txq->tx_tail;
1463         txdp = &txq->tx_ring[tx_id];
1464         txep = &txq->sw_ring[tx_id];
1465
1466         txq->nb_free = (uint16_t)(txq->nb_free - nb_pkts);
1467
1468         n = (uint16_t)(txq->nb_tx_desc - tx_id);
1469         if (nb_commit >= n) {
1470                 tx_backlog_entry(txep, tx_pkts, n);
1471
1472                 iavf_vtx(txdp, tx_pkts, n - 1, flags);
1473                 tx_pkts += (n - 1);
1474                 txdp += (n - 1);
1475
1476                 iavf_vtx1(txdp, *tx_pkts++, rs);
1477
1478                 nb_commit = (uint16_t)(nb_commit - n);
1479
1480                 tx_id = 0;
1481                 txq->next_rs = (uint16_t)(txq->rs_thresh - 1);
1482
1483                 /* avoid reach the end of ring */
1484                 txdp = &txq->tx_ring[tx_id];
1485                 txep = &txq->sw_ring[tx_id];
1486         }
1487
1488         tx_backlog_entry(txep, tx_pkts, nb_commit);
1489
1490         iavf_vtx(txdp, tx_pkts, nb_commit, flags);
1491
1492         tx_id = (uint16_t)(tx_id + nb_commit);
1493         if (tx_id > txq->next_rs) {
1494                 txq->tx_ring[txq->next_rs].cmd_type_offset_bsz |=
1495                         rte_cpu_to_le_64(((uint64_t)IAVF_TX_DESC_CMD_RS) <<
1496                                          IAVF_TXD_QW1_CMD_SHIFT);
1497                 txq->next_rs =
1498                         (uint16_t)(txq->next_rs + txq->rs_thresh);
1499         }
1500
1501         txq->tx_tail = tx_id;
1502
1503         IAVF_PCI_REG_WC_WRITE(txq->qtx_tail, txq->tx_tail);
1504
1505         return nb_pkts;
1506 }
1507
1508 uint16_t
1509 iavf_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
1510                         uint16_t nb_pkts)
1511 {
1512         uint16_t nb_tx = 0;
1513         struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
1514
1515         while (nb_pkts) {
1516                 uint16_t ret, num;
1517
1518                 num = (uint16_t)RTE_MIN(nb_pkts, txq->rs_thresh);
1519                 ret = iavf_xmit_fixed_burst_vec_avx2(tx_queue, &tx_pkts[nb_tx],
1520                                                      num);
1521                 nb_tx += ret;
1522                 nb_pkts -= ret;
1523                 if (ret < num)
1524                         break;
1525         }
1526
1527         return nb_tx;
1528 }