drivers/net/iavf/iavf_rxtx_vec_sse.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2017 Intel Corporation
   3  */
   4
   5 #include <stdint.h>
   6 #include <rte_ethdev_driver.h>
   7 #include <rte_malloc.h>
   8
   9 #include "iavf.h"
  10 #include "iavf_rxtx.h"
  11 #include "iavf_rxtx_vec_common.h"
  12
  13 #include <tmmintrin.h>
  14
  15 #ifndef __INTEL_COMPILER
  16 #pragma GCC diagnostic ignored "-Wcast-qual"
  17 #endif
  18
  19 static inline void
  20 iavf_rxq_rearm(struct iavf_rx_queue *rxq)
  21 {
  22         int i;
  23         uint16_t rx_id;
  24
  25         volatile union iavf_rx_desc *rxdp;
  26         struct rte_mbuf **rxp = &rxq->sw_ring[rxq->rxrearm_start];
  27         struct rte_mbuf *mb0, *mb1;
  28         __m128i hdr_room = _mm_set_epi64x(RTE_PKTMBUF_HEADROOM,
  29                         RTE_PKTMBUF_HEADROOM);
  30         __m128i dma_addr0, dma_addr1;
  31
  32         rxdp = rxq->rx_ring + rxq->rxrearm_start;
  33
  34         /* Pull 'n' more MBUFs into the software ring */
  35         if (rte_mempool_get_bulk(rxq->mp, (void *)rxp,
  36                                  rxq->rx_free_thresh) < 0) {
  37                 if (rxq->rxrearm_nb + rxq->rx_free_thresh >= rxq->nb_rx_desc) {
  38                         dma_addr0 = _mm_setzero_si128();
  39                         for (i = 0; i < IAVF_VPMD_DESCS_PER_LOOP; i++) {
  40                                 rxp[i] = &rxq->fake_mbuf;
  41                                 _mm_store_si128((__m128i *)&rxdp[i].read,
  42                                                 dma_addr0);
  43                         }
  44                 }
  45                 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
  46                         rxq->rx_free_thresh;
  47                 return;
  48         }
  49
  50         /* Initialize the mbufs in vector, process 2 mbufs in one loop */
  51         for (i = 0; i < rxq->rx_free_thresh; i += 2, rxp += 2) {
  52                 __m128i vaddr0, vaddr1;
  53
  54                 mb0 = rxp[0];
  55                 mb1 = rxp[1];
  56
  57                 /* load buf_addr(lo 64bit) and buf_iova(hi 64bit) */
  58                 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, buf_iova) !=
  59                                 offsetof(struct rte_mbuf, buf_addr) + 8);
  60                 vaddr0 = _mm_loadu_si128((__m128i *)&mb0->buf_addr);
  61                 vaddr1 = _mm_loadu_si128((__m128i *)&mb1->buf_addr);
  62
  63                 /* convert pa to dma_addr hdr/data */
  64                 dma_addr0 = _mm_unpackhi_epi64(vaddr0, vaddr0);
  65                 dma_addr1 = _mm_unpackhi_epi64(vaddr1, vaddr1);
  66
  67                 /* add headroom to pa values */
  68                 dma_addr0 = _mm_add_epi64(dma_addr0, hdr_room);
  69                 dma_addr1 = _mm_add_epi64(dma_addr1, hdr_room);
  70
  71                 /* flush desc with pa dma_addr */
  72                 _mm_store_si128((__m128i *)&rxdp++->read, dma_addr0);
  73                 _mm_store_si128((__m128i *)&rxdp++->read, dma_addr1);
  74         }
  75
  76         rxq->rxrearm_start += rxq->rx_free_thresh;
  77         if (rxq->rxrearm_start >= rxq->nb_rx_desc)
  78                 rxq->rxrearm_start = 0;
  79
  80         rxq->rxrearm_nb -= rxq->rx_free_thresh;
  81
  82         rx_id = (uint16_t)((rxq->rxrearm_start == 0) ?
  83                            (rxq->nb_rx_desc - 1) : (rxq->rxrearm_start - 1));
  84
  85         PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_tail=%u "
  86                    "rearm_start=%u rearm_nb=%u",
  87                    rxq->port_id, rxq->queue_id,
  88                    rx_id, rxq->rxrearm_start, rxq->rxrearm_nb);
  89
  90         /* Update the tail pointer on the NIC */
  91         IAVF_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
  92 }
  93
  94 static inline void
  95 desc_to_olflags_v(struct iavf_rx_queue *rxq, __m128i descs[4],
  96                   struct rte_mbuf **rx_pkts)
  97 {
  98         const __m128i mbuf_init = _mm_set_epi64x(0, rxq->mbuf_initializer);
  99         __m128i rearm0, rearm1, rearm2, rearm3;
 100
 101         __m128i vlan0, vlan1, rss, l3_l4e;
 102
 103         /* mask everything except RSS, flow director and VLAN flags
 104          * bit2 is for VLAN tag, bit11 for flow director indication
 105          * bit13:12 for RSS indication.
 106          */
 107         const __m128i rss_vlan_msk = _mm_set_epi32(
 108                         0x1c03804, 0x1c03804, 0x1c03804, 0x1c03804);
 109
 110         const __m128i cksum_mask = _mm_set_epi32(
 111                         PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
 112                         PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
 113                         PKT_RX_EIP_CKSUM_BAD,
 114                         PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
 115                         PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
 116                         PKT_RX_EIP_CKSUM_BAD,
 117                         PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
 118                         PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
 119                         PKT_RX_EIP_CKSUM_BAD,
 120                         PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
 121                         PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
 122                         PKT_RX_EIP_CKSUM_BAD);
 123
 124         /* map rss and vlan type to rss hash and vlan flag */
 125         const __m128i vlan_flags = _mm_set_epi8(0, 0, 0, 0,
 126                         0, 0, 0, 0,
 127                         0, 0, 0, PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
 128                         0, 0, 0, 0);
 129
 130         const __m128i rss_flags = _mm_set_epi8(0, 0, 0, 0,
 131                         0, 0, 0, 0,
 132                         PKT_RX_RSS_HASH | PKT_RX_FDIR, PKT_RX_RSS_HASH, 0, 0,
 133                         0, 0, PKT_RX_FDIR, 0);
 134
 135         const __m128i l3_l4e_flags = _mm_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
 136                         /* shift right 1 bit to make sure it not exceed 255 */
 137                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
 138                          PKT_RX_IP_CKSUM_BAD) >> 1,
 139                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD |
 140                          PKT_RX_L4_CKSUM_BAD) >> 1,
 141                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
 142                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD) >> 1,
 143                         (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
 144                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD) >> 1,
 145                         PKT_RX_IP_CKSUM_BAD >> 1,
 146                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1);
 147
 148         vlan0 = _mm_unpackhi_epi32(descs[0], descs[1]);
 149         vlan1 = _mm_unpackhi_epi32(descs[2], descs[3]);
 150         vlan0 = _mm_unpacklo_epi64(vlan0, vlan1);
 151
 152         vlan1 = _mm_and_si128(vlan0, rss_vlan_msk);
 153         vlan0 = _mm_shuffle_epi8(vlan_flags, vlan1);
 154
 155         rss = _mm_srli_epi32(vlan1, 11);
 156         rss = _mm_shuffle_epi8(rss_flags, rss);
 157
 158         l3_l4e = _mm_srli_epi32(vlan1, 22);
 159         l3_l4e = _mm_shuffle_epi8(l3_l4e_flags, l3_l4e);
 160         /* then we shift left 1 bit */
 161         l3_l4e = _mm_slli_epi32(l3_l4e, 1);
 162         /* we need to mask out the reduntant bits */
 163         l3_l4e = _mm_and_si128(l3_l4e, cksum_mask);
 164
 165         vlan0 = _mm_or_si128(vlan0, rss);
 166         vlan0 = _mm_or_si128(vlan0, l3_l4e);
 167
 168         /* At this point, we have the 4 sets of flags in the low 16-bits
 169          * of each 32-bit value in vlan0.
 170          * We want to extract these, and merge them with the mbuf init data
 171          * so we can do a single 16-byte write to the mbuf to set the flags
 172          * and all the other initialization fields. Extracting the
 173          * appropriate flags means that we have to do a shift and blend for
 174          * each mbuf before we do the write.
 175          */
 176         rearm0 = _mm_blend_epi16(mbuf_init, _mm_slli_si128(vlan0, 8), 0x10);
 177         rearm1 = _mm_blend_epi16(mbuf_init, _mm_slli_si128(vlan0, 4), 0x10);
 178         rearm2 = _mm_blend_epi16(mbuf_init, vlan0, 0x10);
 179         rearm3 = _mm_blend_epi16(mbuf_init, _mm_srli_si128(vlan0, 4), 0x10);
 180
 181         /* write the rearm data and the olflags in one write */
 182         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, ol_flags) !=
 183                         offsetof(struct rte_mbuf, rearm_data) + 8);
 184         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, rearm_data) !=
 185                         RTE_ALIGN(offsetof(struct rte_mbuf, rearm_data), 16));
 186         _mm_store_si128((__m128i *)&rx_pkts[0]->rearm_data, rearm0);
 187         _mm_store_si128((__m128i *)&rx_pkts[1]->rearm_data, rearm1);
 188         _mm_store_si128((__m128i *)&rx_pkts[2]->rearm_data, rearm2);
 189         _mm_store_si128((__m128i *)&rx_pkts[3]->rearm_data, rearm3);
 190 }
 191
 192 static inline __m128i
 193 flex_rxd_to_fdir_flags_vec(const __m128i fdir_id0_3)
 194 {
 195 #define FDID_MIS_MAGIC 0xFFFFFFFF
 196         RTE_BUILD_BUG_ON(PKT_RX_FDIR != (1 << 2));
 197         RTE_BUILD_BUG_ON(PKT_RX_FDIR_ID != (1 << 13));
 198         const __m128i pkt_fdir_bit = _mm_set1_epi32(PKT_RX_FDIR |
 199                         PKT_RX_FDIR_ID);
 200         /* desc->flow_id field == 0xFFFFFFFF means fdir mismatch */
 201         const __m128i fdir_mis_mask = _mm_set1_epi32(FDID_MIS_MAGIC);
 202         __m128i fdir_mask = _mm_cmpeq_epi32(fdir_id0_3,
 203                         fdir_mis_mask);
 204         /* this XOR op results to bit-reverse the fdir_mask */
 205         fdir_mask = _mm_xor_si128(fdir_mask, fdir_mis_mask);
 206         const __m128i fdir_flags = _mm_and_si128(fdir_mask, pkt_fdir_bit);
 207
 208         return fdir_flags;
 209 }
 210
 211 static inline void
 212 flex_desc_to_olflags_v(struct iavf_rx_queue *rxq, __m128i descs[4],
 213                        struct rte_mbuf **rx_pkts)
 214 {
 215         const __m128i mbuf_init = _mm_set_epi64x(0, rxq->mbuf_initializer);
 216         __m128i rearm0, rearm1, rearm2, rearm3;
 217
 218         __m128i tmp_desc, flags, rss_vlan;
 219
 220         /* mask everything except checksum, RSS and VLAN flags.
 221          * bit6:4 for checksum.
 222          * bit12 for RSS indication.
 223          * bit13 for VLAN indication.
 224          */
 225         const __m128i desc_mask = _mm_set_epi32(0x3070, 0x3070,
 226                                                 0x3070, 0x3070);
 227
 228         const __m128i cksum_mask = _mm_set_epi32(PKT_RX_IP_CKSUM_MASK |
 229                                                  PKT_RX_L4_CKSUM_MASK |
 230                                                  PKT_RX_EIP_CKSUM_BAD,
 231                                                  PKT_RX_IP_CKSUM_MASK |
 232                                                  PKT_RX_L4_CKSUM_MASK |
 233                                                  PKT_RX_EIP_CKSUM_BAD,
 234                                                  PKT_RX_IP_CKSUM_MASK |
 235                                                  PKT_RX_L4_CKSUM_MASK |
 236                                                  PKT_RX_EIP_CKSUM_BAD,
 237                                                  PKT_RX_IP_CKSUM_MASK |
 238                                                  PKT_RX_L4_CKSUM_MASK |
 239                                                  PKT_RX_EIP_CKSUM_BAD);
 240
 241         /* map the checksum, rss and vlan fields to the checksum, rss
 242          * and vlan flag
 243          */
 244         const __m128i cksum_flags = _mm_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
 245                         /* shift right 1 bit to make sure it not exceed 255 */
 246                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
 247                          PKT_RX_IP_CKSUM_BAD) >> 1,
 248                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
 249                          PKT_RX_IP_CKSUM_GOOD) >> 1,
 250                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
 251                          PKT_RX_IP_CKSUM_BAD) >> 1,
 252                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
 253                          PKT_RX_IP_CKSUM_GOOD) >> 1,
 254                         (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
 255                         (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
 256                         (PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
 257                         (PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_GOOD) >> 1);
 258
 259         const __m128i rss_vlan_flags = _mm_set_epi8(0, 0, 0, 0,
 260                         0, 0, 0, 0,
 261                         0, 0, 0, 0,
 262                         PKT_RX_RSS_HASH | PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
 263                         PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
 264                         PKT_RX_RSS_HASH, 0);
 265
 266         /* merge 4 descriptors */
 267         flags = _mm_unpackhi_epi32(descs[0], descs[1]);
 268         tmp_desc = _mm_unpackhi_epi32(descs[2], descs[3]);
 269         tmp_desc = _mm_unpacklo_epi64(flags, tmp_desc);
 270         tmp_desc = _mm_and_si128(flags, desc_mask);
 271
 272         /* checksum flags */
 273         tmp_desc = _mm_srli_epi32(tmp_desc, 4);
 274         flags = _mm_shuffle_epi8(cksum_flags, tmp_desc);
 275         /* then we shift left 1 bit */
 276         flags = _mm_slli_epi32(flags, 1);
 277         /* we need to mask out the redundant bits introduced by RSS or
 278          * VLAN fields.
 279          */
 280         flags = _mm_and_si128(flags, cksum_mask);
 281
 282         /* RSS, VLAN flag */
 283         tmp_desc = _mm_srli_epi32(tmp_desc, 8);
 284         rss_vlan = _mm_shuffle_epi8(rss_vlan_flags, tmp_desc);
 285
 286         /* merge the flags */
 287         flags = _mm_or_si128(flags, rss_vlan);
 288
 289         if (rxq->fdir_enabled) {
 290                 const __m128i fdir_id0_1 =
 291                         _mm_unpackhi_epi32(descs[0], descs[1]);
 292
 293                 const __m128i fdir_id2_3 =
 294                         _mm_unpackhi_epi32(descs[2], descs[3]);
 295
 296                 const __m128i fdir_id0_3 =
 297                         _mm_unpackhi_epi64(fdir_id0_1, fdir_id2_3);
 298
 299                 const __m128i fdir_flags =
 300                         flex_rxd_to_fdir_flags_vec(fdir_id0_3);
 301
 302                 /* merge with fdir_flags */
 303                 flags = _mm_or_si128(flags, fdir_flags);
 304
 305                 /* write fdir_id to mbuf */
 306                 rx_pkts[0]->hash.fdir.hi =
 307                         _mm_extract_epi32(fdir_id0_3, 0);
 308
 309                 rx_pkts[1]->hash.fdir.hi =
 310                         _mm_extract_epi32(fdir_id0_3, 1);
 311
 312                 rx_pkts[2]->hash.fdir.hi =
 313                         _mm_extract_epi32(fdir_id0_3, 2);
 314
 315                 rx_pkts[3]->hash.fdir.hi =
 316                         _mm_extract_epi32(fdir_id0_3, 3);
 317         } /* if() on fdir_enabled */
 318
 319         /**
 320          * At this point, we have the 4 sets of flags in the low 16-bits
 321          * of each 32-bit value in flags.
 322          * We want to extract these, and merge them with the mbuf init data
 323          * so we can do a single 16-byte write to the mbuf to set the flags
 324          * and all the other initialization fields. Extracting the
 325          * appropriate flags means that we have to do a shift and blend for
 326          * each mbuf before we do the write.
 327          */
 328         rearm0 = _mm_blend_epi16(mbuf_init, _mm_slli_si128(flags, 8), 0x10);
 329         rearm1 = _mm_blend_epi16(mbuf_init, _mm_slli_si128(flags, 4), 0x10);
 330         rearm2 = _mm_blend_epi16(mbuf_init, flags, 0x10);
 331         rearm3 = _mm_blend_epi16(mbuf_init, _mm_srli_si128(flags, 4), 0x10);
 332
 333         /* write the rearm data and the olflags in one write */
 334         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, ol_flags) !=
 335                          offsetof(struct rte_mbuf, rearm_data) + 8);
 336         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, rearm_data) !=
 337                          RTE_ALIGN(offsetof(struct rte_mbuf, rearm_data), 16));
 338         _mm_store_si128((__m128i *)&rx_pkts[0]->rearm_data, rearm0);
 339         _mm_store_si128((__m128i *)&rx_pkts[1]->rearm_data, rearm1);
 340         _mm_store_si128((__m128i *)&rx_pkts[2]->rearm_data, rearm2);
 341         _mm_store_si128((__m128i *)&rx_pkts[3]->rearm_data, rearm3);
 342 }
 343
 344 #define PKTLEN_SHIFT     10
 345
 346 static inline void
 347 desc_to_ptype_v(__m128i descs[4], struct rte_mbuf **rx_pkts,
 348                 const uint32_t *type_table)
 349 {
 350         __m128i ptype0 = _mm_unpackhi_epi64(descs[0], descs[1]);
 351         __m128i ptype1 = _mm_unpackhi_epi64(descs[2], descs[3]);
 352
 353         ptype0 = _mm_srli_epi64(ptype0, 30);
 354         ptype1 = _mm_srli_epi64(ptype1, 30);
 355
 356         rx_pkts[0]->packet_type = type_table[_mm_extract_epi8(ptype0, 0)];
 357         rx_pkts[1]->packet_type = type_table[_mm_extract_epi8(ptype0, 8)];
 358         rx_pkts[2]->packet_type = type_table[_mm_extract_epi8(ptype1, 0)];
 359         rx_pkts[3]->packet_type = type_table[_mm_extract_epi8(ptype1, 8)];
 360 }
 361
 362 static inline void
 363 flex_desc_to_ptype_v(__m128i descs[4], struct rte_mbuf **rx_pkts,
 364                      const uint32_t *type_table)
 365 {
 366         const __m128i ptype_mask = _mm_set_epi16(0, IAVF_RX_FLEX_DESC_PTYPE_M,
 367                                                  0, IAVF_RX_FLEX_DESC_PTYPE_M,
 368                                                  0, IAVF_RX_FLEX_DESC_PTYPE_M,
 369                                                  0, IAVF_RX_FLEX_DESC_PTYPE_M);
 370         __m128i ptype_01 = _mm_unpacklo_epi32(descs[0], descs[1]);
 371         __m128i ptype_23 = _mm_unpacklo_epi32(descs[2], descs[3]);
 372         __m128i ptype_all = _mm_unpacklo_epi64(ptype_01, ptype_23);
 373
 374         ptype_all = _mm_and_si128(ptype_all, ptype_mask);
 375
 376         rx_pkts[0]->packet_type = type_table[_mm_extract_epi16(ptype_all, 1)];
 377         rx_pkts[1]->packet_type = type_table[_mm_extract_epi16(ptype_all, 3)];
 378         rx_pkts[2]->packet_type = type_table[_mm_extract_epi16(ptype_all, 5)];
 379         rx_pkts[3]->packet_type = type_table[_mm_extract_epi16(ptype_all, 7)];
 380 }
 381
 382 /* Notice:
 383  * - nb_pkts < IAVF_VPMD_DESCS_PER_LOOP, just return no packet
 384  * - nb_pkts > IAVF_VPMD_RX_MAX_BURST, only scan IAVF_VPMD_RX_MAX_BURST
 385  *   numbers of DD bits
 386  */
 387 static inline uint16_t
 388 _recv_raw_pkts_vec(struct iavf_rx_queue *rxq, struct rte_mbuf **rx_pkts,
 389                    uint16_t nb_pkts, uint8_t *split_packet)
 390 {
 391         volatile union iavf_rx_desc *rxdp;
 392         struct rte_mbuf **sw_ring;
 393         uint16_t nb_pkts_recd;
 394         int pos;
 395         uint64_t var;
 396         __m128i shuf_msk;
 397         const uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
 398
 399         __m128i crc_adjust = _mm_set_epi16(
 400                                 0, 0, 0,    /* ignore non-length fields */
 401                                 -rxq->crc_len, /* sub crc on data_len */
 402                                 0,          /* ignore high-16bits of pkt_len */
 403                                 -rxq->crc_len, /* sub crc on pkt_len */
 404                                 0, 0            /* ignore pkt_type field */
 405                         );
 406         /* compile-time check the above crc_adjust layout is correct.
 407          * NOTE: the first field (lowest address) is given last in set_epi16
 408          * call above.
 409          */
 410         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
 411                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
 412         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
 413                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
 414         __m128i dd_check, eop_check;
 415
 416         /* nb_pkts shall be less equal than IAVF_VPMD_RX_MAX_BURST */
 417         nb_pkts = RTE_MIN(nb_pkts, IAVF_VPMD_RX_MAX_BURST);
 418
 419         /* nb_pkts has to be floor-aligned to IAVF_VPMD_DESCS_PER_LOOP */
 420         nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_VPMD_DESCS_PER_LOOP);
 421
 422         /* Just the act of getting into the function from the application is
 423          * going to cost about 7 cycles
 424          */
 425         rxdp = rxq->rx_ring + rxq->rx_tail;
 426
 427         rte_prefetch0(rxdp);
 428
 429         /* See if we need to rearm the RX queue - gives the prefetch a bit
 430          * of time to act
 431          */
 432         if (rxq->rxrearm_nb > rxq->rx_free_thresh)
 433                 iavf_rxq_rearm(rxq);
 434
 435         /* Before we start moving massive data around, check to see if
 436          * there is actually a packet available
 437          */
 438         if (!(rxdp->wb.qword1.status_error_len &
 439               rte_cpu_to_le_32(1 << IAVF_RX_DESC_STATUS_DD_SHIFT)))
 440                 return 0;
 441
 442         /* 4 packets DD mask */
 443         dd_check = _mm_set_epi64x(0x0000000100000001LL, 0x0000000100000001LL);
 444
 445         /* 4 packets EOP mask */
 446         eop_check = _mm_set_epi64x(0x0000000200000002LL, 0x0000000200000002LL);
 447
 448         /* mask to shuffle from desc. to mbuf */
 449         shuf_msk = _mm_set_epi8(
 450                 7, 6, 5, 4,  /* octet 4~7, 32bits rss */
 451                 3, 2,        /* octet 2~3, low 16 bits vlan_macip */
 452                 15, 14,      /* octet 15~14, 16 bits data_len */
 453                 0xFF, 0xFF,  /* skip high 16 bits pkt_len, zero out */
 454                 15, 14,      /* octet 15~14, low 16 bits pkt_len */
 455                 0xFF, 0xFF, 0xFF, 0xFF /* pkt_type set as unknown */
 456                 );
 457         /* Compile-time verify the shuffle mask
 458          * NOTE: some field positions already verified above, but duplicated
 459          * here for completeness in case of future modifications.
 460          */
 461         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
 462                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
 463         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
 464                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
 465         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, vlan_tci) !=
 466                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 10);
 467         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) !=
 468                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
 469
 470         /* Cache is empty -> need to scan the buffer rings, but first move
 471          * the next 'n' mbufs into the cache
 472          */
 473         sw_ring = &rxq->sw_ring[rxq->rx_tail];
 474
 475         /* A. load 4 packet in one loop
 476          * [A*. mask out 4 unused dirty field in desc]
 477          * B. copy 4 mbuf point from swring to rx_pkts
 478          * C. calc the number of DD bits among the 4 packets
 479          * [C*. extract the end-of-packet bit, if requested]
 480          * D. fill info. from desc to mbuf
 481          */
 482
 483         for (pos = 0, nb_pkts_recd = 0; pos < nb_pkts;
 484              pos += IAVF_VPMD_DESCS_PER_LOOP,
 485              rxdp += IAVF_VPMD_DESCS_PER_LOOP) {
 486                 __m128i descs[IAVF_VPMD_DESCS_PER_LOOP];
 487                 __m128i pkt_mb1, pkt_mb2, pkt_mb3, pkt_mb4;
 488                 __m128i zero, staterr, sterr_tmp1, sterr_tmp2;
 489                 /* 2 64 bit or 4 32 bit mbuf pointers in one XMM reg. */
 490                 __m128i mbp1;
 491 #if defined(RTE_ARCH_X86_64)
 492                 __m128i mbp2;
 493 #endif
 494
 495                 /* B.1 load 2 (64 bit) or 4 (32 bit) mbuf points */
 496                 mbp1 = _mm_loadu_si128((__m128i *)&sw_ring[pos]);
 497                 /* Read desc statuses backwards to avoid race condition */
 498                 /* A.1 load 4 pkts desc */
 499                 descs[3] = _mm_loadu_si128((__m128i *)(rxdp + 3));
 500                 rte_compiler_barrier();
 501
 502                 /* B.2 copy 2 64 bit or 4 32 bit mbuf point into rx_pkts */
 503                 _mm_storeu_si128((__m128i *)&rx_pkts[pos], mbp1);
 504
 505 #if defined(RTE_ARCH_X86_64)
 506                 /* B.1 load 2 64 bit mbuf points */
 507                 mbp2 = _mm_loadu_si128((__m128i *)&sw_ring[pos + 2]);
 508 #endif
 509
 510                 descs[2] = _mm_loadu_si128((__m128i *)(rxdp + 2));
 511                 rte_compiler_barrier();
 512                 /* B.1 load 2 mbuf point */
 513                 descs[1] = _mm_loadu_si128((__m128i *)(rxdp + 1));
 514                 rte_compiler_barrier();
 515                 descs[0] = _mm_loadu_si128((__m128i *)(rxdp));
 516
 517 #if defined(RTE_ARCH_X86_64)
 518                 /* B.2 copy 2 mbuf point into rx_pkts  */
 519                 _mm_storeu_si128((__m128i *)&rx_pkts[pos + 2], mbp2);
 520 #endif
 521
 522                 if (split_packet) {
 523                         rte_mbuf_prefetch_part2(rx_pkts[pos]);
 524                         rte_mbuf_prefetch_part2(rx_pkts[pos + 1]);
 525                         rte_mbuf_prefetch_part2(rx_pkts[pos + 2]);
 526                         rte_mbuf_prefetch_part2(rx_pkts[pos + 3]);
 527                 }
 528
 529                 /* avoid compiler reorder optimization */
 530                 rte_compiler_barrier();
 531
 532                 /* pkt 3,4 shift the pktlen field to be 16-bit aligned*/
 533                 const __m128i len3 = _mm_slli_epi32(descs[3], PKTLEN_SHIFT);
 534                 const __m128i len2 = _mm_slli_epi32(descs[2], PKTLEN_SHIFT);
 535
 536                 /* merge the now-aligned packet length fields back in */
 537                 descs[3] = _mm_blend_epi16(descs[3], len3, 0x80);
 538                 descs[2] = _mm_blend_epi16(descs[2], len2, 0x80);
 539
 540                 /* D.1 pkt 3,4 convert format from desc to pktmbuf */
 541                 pkt_mb4 = _mm_shuffle_epi8(descs[3], shuf_msk);
 542                 pkt_mb3 = _mm_shuffle_epi8(descs[2], shuf_msk);
 543
 544                 /* C.1 4=>2 status err info only */
 545                 sterr_tmp2 = _mm_unpackhi_epi32(descs[3], descs[2]);
 546                 sterr_tmp1 = _mm_unpackhi_epi32(descs[1], descs[0]);
 547
 548                 desc_to_olflags_v(rxq, descs, &rx_pkts[pos]);
 549
 550                 /* D.2 pkt 3,4 set in_port/nb_seg and remove crc */
 551                 pkt_mb4 = _mm_add_epi16(pkt_mb4, crc_adjust);
 552                 pkt_mb3 = _mm_add_epi16(pkt_mb3, crc_adjust);
 553
 554                 /* pkt 1,2 shift the pktlen field to be 16-bit aligned*/
 555                 const __m128i len1 = _mm_slli_epi32(descs[1], PKTLEN_SHIFT);
 556                 const __m128i len0 = _mm_slli_epi32(descs[0], PKTLEN_SHIFT);
 557
 558                 /* merge the now-aligned packet length fields back in */
 559                 descs[1] = _mm_blend_epi16(descs[1], len1, 0x80);
 560                 descs[0] = _mm_blend_epi16(descs[0], len0, 0x80);
 561
 562                 /* D.1 pkt 1,2 convert format from desc to pktmbuf */
 563                 pkt_mb2 = _mm_shuffle_epi8(descs[1], shuf_msk);
 564                 pkt_mb1 = _mm_shuffle_epi8(descs[0], shuf_msk);
 565
 566                 /* C.2 get 4 pkts status err value  */
 567                 zero = _mm_xor_si128(dd_check, dd_check);
 568                 staterr = _mm_unpacklo_epi32(sterr_tmp1, sterr_tmp2);
 569
 570                 /* D.3 copy final 3,4 data to rx_pkts */
 571                 _mm_storeu_si128(
 572                         (void *)&rx_pkts[pos + 3]->rx_descriptor_fields1,
 573                         pkt_mb4);
 574                 _mm_storeu_si128(
 575                         (void *)&rx_pkts[pos + 2]->rx_descriptor_fields1,
 576                         pkt_mb3);
 577
 578                 /* D.2 pkt 1,2 remove crc */
 579                 pkt_mb2 = _mm_add_epi16(pkt_mb2, crc_adjust);
 580                 pkt_mb1 = _mm_add_epi16(pkt_mb1, crc_adjust);
 581
 582                 /* C* extract and record EOP bit */
 583                 if (split_packet) {
 584                         __m128i eop_shuf_mask = _mm_set_epi8(
 585                                         0xFF, 0xFF, 0xFF, 0xFF,
 586                                         0xFF, 0xFF, 0xFF, 0xFF,
 587                                         0xFF, 0xFF, 0xFF, 0xFF,
 588                                         0x04, 0x0C, 0x00, 0x08
 589                                         );
 590
 591                         /* and with mask to extract bits, flipping 1-0 */
 592                         __m128i eop_bits = _mm_andnot_si128(staterr, eop_check);
 593                         /* the staterr values are not in order, as the count
 594                          * count of dd bits doesn't care. However, for end of
 595                          * packet tracking, we do care, so shuffle. This also
 596                          * compresses the 32-bit values to 8-bit
 597                          */
 598                         eop_bits = _mm_shuffle_epi8(eop_bits, eop_shuf_mask);
 599                         /* store the resulting 32-bit value */
 600                         *(int *)split_packet = _mm_cvtsi128_si32(eop_bits);
 601                         split_packet += IAVF_VPMD_DESCS_PER_LOOP;
 602                 }
 603
 604                 /* C.3 calc available number of desc */
 605                 staterr = _mm_and_si128(staterr, dd_check);
 606                 staterr = _mm_packs_epi32(staterr, zero);
 607
 608                 /* D.3 copy final 1,2 data to rx_pkts */
 609                 _mm_storeu_si128(
 610                         (void *)&rx_pkts[pos + 1]->rx_descriptor_fields1,
 611                         pkt_mb2);
 612                 _mm_storeu_si128((void *)&rx_pkts[pos]->rx_descriptor_fields1,
 613                                  pkt_mb1);
 614                 desc_to_ptype_v(descs, &rx_pkts[pos], ptype_tbl);
 615                 /* C.4 calc avaialbe number of desc */
 616                 var = __builtin_popcountll(_mm_cvtsi128_si64(staterr));
 617                 nb_pkts_recd += var;
 618                 if (likely(var != IAVF_VPMD_DESCS_PER_LOOP))
 619                         break;
 620         }
 621
 622         /* Update our internal tail pointer */
 623         rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_pkts_recd);
 624         rxq->rx_tail = (uint16_t)(rxq->rx_tail & (rxq->nb_rx_desc - 1));
 625         rxq->rxrearm_nb = (uint16_t)(rxq->rxrearm_nb + nb_pkts_recd);
 626
 627         return nb_pkts_recd;
 628 }
 629
 630 /* Notice:
 631  * - nb_pkts < IAVF_VPMD_DESCS_PER_LOOP, just return no packet
 632  * - nb_pkts > IAVF_VPMD_RX_MAX_BURST, only scan IAVF_VPMD_RX_MAX_BURST
 633  *   numbers of DD bits
 634  */
 635 static inline uint16_t
 636 _recv_raw_pkts_vec_flex_rxd(struct iavf_rx_queue *rxq,
 637                             struct rte_mbuf **rx_pkts,
 638                             uint16_t nb_pkts, uint8_t *split_packet)
 639 {
 640         volatile union iavf_rx_flex_desc *rxdp;
 641         struct rte_mbuf **sw_ring;
 642         uint16_t nb_pkts_recd;
 643         int pos;
 644         uint64_t var;
 645         const uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
 646         __m128i crc_adjust = _mm_set_epi16
 647                                 (0, 0, 0,       /* ignore non-length fields */
 648                                  -rxq->crc_len, /* sub crc on data_len */
 649                                  0,          /* ignore high-16bits of pkt_len */
 650                                  -rxq->crc_len, /* sub crc on pkt_len */
 651                                  0, 0           /* ignore pkt_type field */
 652                                 );
 653         const __m128i zero = _mm_setzero_si128();
 654         /* mask to shuffle from desc. to mbuf */
 655         const __m128i shuf_msk = _mm_set_epi8
 656                         (0xFF, 0xFF,
 657                          0xFF, 0xFF,  /* rss hash parsed separately */
 658                          11, 10,      /* octet 10~11, 16 bits vlan_macip */
 659                          5, 4,        /* octet 4~5, 16 bits data_len */
 660                          0xFF, 0xFF,  /* skip high 16 bits pkt_len, zero out */
 661                          5, 4,        /* octet 4~5, low 16 bits pkt_len */
 662                          0xFF, 0xFF,  /* pkt_type set as unknown */
 663                          0xFF, 0xFF   /* pkt_type set as unknown */
 664                         );
 665         const __m128i eop_shuf_mask = _mm_set_epi8(0xFF, 0xFF,
 666                                                    0xFF, 0xFF,
 667                                                    0xFF, 0xFF,
 668                                                    0xFF, 0xFF,
 669                                                    0xFF, 0xFF,
 670                                                    0xFF, 0xFF,
 671                                                    0x04, 0x0C,
 672                                                    0x00, 0x08);
 673
 674         /**
 675          * compile-time check the above crc_adjust layout is correct.
 676          * NOTE: the first field (lowest address) is given last in set_epi16
 677          * call above.
 678          */
 679         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
 680                          offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
 681         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
 682                          offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
 683
 684         /* 4 packets DD mask */
 685         const __m128i dd_check = _mm_set_epi64x(0x0000000100000001LL,
 686                                                 0x0000000100000001LL);
 687         /* 4 packets EOP mask */
 688         const __m128i eop_check = _mm_set_epi64x(0x0000000200000002LL,
 689                                                  0x0000000200000002LL);
 690
 691         /* nb_pkts shall be less equal than IAVF_VPMD_RX_MAX_BURST */
 692         nb_pkts = RTE_MIN(nb_pkts, IAVF_VPMD_RX_MAX_BURST);
 693
 694         /* nb_pkts has to be floor-aligned to IAVF_VPMD_DESCS_PER_LOOP */
 695         nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_VPMD_DESCS_PER_LOOP);
 696
 697         /* Just the act of getting into the function from the application is
 698          * going to cost about 7 cycles
 699          */
 700         rxdp = (union iavf_rx_flex_desc *)rxq->rx_ring + rxq->rx_tail;
 701
 702         rte_prefetch0(rxdp);
 703
 704         /* See if we need to rearm the RX queue - gives the prefetch a bit
 705          * of time to act
 706          */
 707         if (rxq->rxrearm_nb > rxq->rx_free_thresh)
 708                 iavf_rxq_rearm(rxq);
 709
 710         /* Before we start moving massive data around, check to see if
 711          * there is actually a packet available
 712          */
 713         if (!(rxdp->wb.status_error0 &
 714               rte_cpu_to_le_32(1 << IAVF_RX_FLEX_DESC_STATUS0_DD_S)))
 715                 return 0;
 716
 717         /**
 718          * Compile-time verify the shuffle mask
 719          * NOTE: some field positions already verified above, but duplicated
 720          * here for completeness in case of future modifications.
 721          */
 722         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
 723                          offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
 724         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
 725                          offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
 726         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, vlan_tci) !=
 727                          offsetof(struct rte_mbuf, rx_descriptor_fields1) + 10);
 728         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) !=
 729                          offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
 730
 731         /* Cache is empty -> need to scan the buffer rings, but first move
 732          * the next 'n' mbufs into the cache
 733          */
 734         sw_ring = &rxq->sw_ring[rxq->rx_tail];
 735
 736         /* A. load 4 packet in one loop
 737          * [A*. mask out 4 unused dirty field in desc]
 738          * B. copy 4 mbuf point from swring to rx_pkts
 739          * C. calc the number of DD bits among the 4 packets
 740          * [C*. extract the end-of-packet bit, if requested]
 741          * D. fill info. from desc to mbuf
 742          */
 743
 744         for (pos = 0, nb_pkts_recd = 0; pos < nb_pkts;
 745              pos += IAVF_VPMD_DESCS_PER_LOOP,
 746              rxdp += IAVF_VPMD_DESCS_PER_LOOP) {
 747                 __m128i descs[IAVF_VPMD_DESCS_PER_LOOP];
 748                 __m128i pkt_mb0, pkt_mb1, pkt_mb2, pkt_mb3;
 749                 __m128i staterr, sterr_tmp1, sterr_tmp2;
 750                 /* 2 64 bit or 4 32 bit mbuf pointers in one XMM reg. */
 751                 __m128i mbp1;
 752 #if defined(RTE_ARCH_X86_64)
 753                 __m128i mbp2;
 754 #endif
 755
 756                 /* B.1 load 2 (64 bit) or 4 (32 bit) mbuf points */
 757                 mbp1 = _mm_loadu_si128((__m128i *)&sw_ring[pos]);
 758                 /* Read desc statuses backwards to avoid race condition */
 759                 /* A.1 load 4 pkts desc */
 760                 descs[3] = _mm_loadu_si128((__m128i *)(rxdp + 3));
 761                 rte_compiler_barrier();
 762
 763                 /* B.2 copy 2 64 bit or 4 32 bit mbuf point into rx_pkts */
 764                 _mm_storeu_si128((__m128i *)&rx_pkts[pos], mbp1);
 765
 766 #if defined(RTE_ARCH_X86_64)
 767                 /* B.1 load 2 64 bit mbuf points */
 768                 mbp2 = _mm_loadu_si128((__m128i *)&sw_ring[pos + 2]);
 769 #endif
 770
 771                 descs[2] = _mm_loadu_si128((__m128i *)(rxdp + 2));
 772                 rte_compiler_barrier();
 773                 /* B.1 load 2 mbuf point */
 774                 descs[1] = _mm_loadu_si128((__m128i *)(rxdp + 1));
 775                 rte_compiler_barrier();
 776                 descs[0] = _mm_loadu_si128((__m128i *)(rxdp));
 777
 778 #if defined(RTE_ARCH_X86_64)
 779                 /* B.2 copy 2 mbuf point into rx_pkts  */
 780                 _mm_storeu_si128((__m128i *)&rx_pkts[pos + 2], mbp2);
 781 #endif
 782
 783                 if (split_packet) {
 784                         rte_mbuf_prefetch_part2(rx_pkts[pos]);
 785                         rte_mbuf_prefetch_part2(rx_pkts[pos + 1]);
 786                         rte_mbuf_prefetch_part2(rx_pkts[pos + 2]);
 787                         rte_mbuf_prefetch_part2(rx_pkts[pos + 3]);
 788                 }
 789
 790                 /* avoid compiler reorder optimization */
 791                 rte_compiler_barrier();
 792
 793                 /* D.1 pkt 3,4 convert format from desc to pktmbuf */
 794                 pkt_mb3 = _mm_shuffle_epi8(descs[3], shuf_msk);
 795                 pkt_mb2 = _mm_shuffle_epi8(descs[2], shuf_msk);
 796
 797                 /* D.1 pkt 1,2 convert format from desc to pktmbuf */
 798                 pkt_mb1 = _mm_shuffle_epi8(descs[1], shuf_msk);
 799                 pkt_mb0 = _mm_shuffle_epi8(descs[0], shuf_msk);
 800
 801                 /* C.1 4=>2 filter staterr info only */
 802                 sterr_tmp2 = _mm_unpackhi_epi32(descs[3], descs[2]);
 803                 /* C.1 4=>2 filter staterr info only */
 804                 sterr_tmp1 = _mm_unpackhi_epi32(descs[1], descs[0]);
 805
 806                 flex_desc_to_olflags_v(rxq, descs, &rx_pkts[pos]);
 807
 808                 /* D.2 pkt 3,4 set in_port/nb_seg and remove crc */
 809                 pkt_mb3 = _mm_add_epi16(pkt_mb3, crc_adjust);
 810                 pkt_mb2 = _mm_add_epi16(pkt_mb2, crc_adjust);
 811
 812                 /* D.2 pkt 1,2 set in_port/nb_seg and remove crc */
 813                 pkt_mb1 = _mm_add_epi16(pkt_mb1, crc_adjust);
 814                 pkt_mb0 = _mm_add_epi16(pkt_mb0, crc_adjust);
 815
 816 #ifndef RTE_LIBRTE_IAVF_16BYTE_RX_DESC
 817                 /**
 818                  * needs to load 2nd 16B of each desc for RSS hash parsing,
 819                  * will cause performance drop to get into this context.
 820                  */
 821                 if (rxq->vsi->adapter->eth_dev->data->dev_conf.rxmode.offloads &
 822                                 DEV_RX_OFFLOAD_RSS_HASH) {
 823                         /* load bottom half of every 32B desc */
 824                         const __m128i raw_desc_bh3 =
 825                                 _mm_load_si128
 826                                         ((void *)(&rxdp[3].wb.status_error1));
 827                         rte_compiler_barrier();
 828                         const __m128i raw_desc_bh2 =
 829                                 _mm_load_si128
 830                                         ((void *)(&rxdp[2].wb.status_error1));
 831                         rte_compiler_barrier();
 832                         const __m128i raw_desc_bh1 =
 833                                 _mm_load_si128
 834                                         ((void *)(&rxdp[1].wb.status_error1));
 835                         rte_compiler_barrier();
 836                         const __m128i raw_desc_bh0 =
 837                                 _mm_load_si128
 838                                         ((void *)(&rxdp[0].wb.status_error1));
 839
 840                         /**
 841                          * to shift the 32b RSS hash value to the
 842                          * highest 32b of each 128b before mask
 843                          */
 844                         __m128i rss_hash3 =
 845                                 _mm_slli_epi64(raw_desc_bh3, 32);
 846                         __m128i rss_hash2 =
 847                                 _mm_slli_epi64(raw_desc_bh2, 32);
 848                         __m128i rss_hash1 =
 849                                 _mm_slli_epi64(raw_desc_bh1, 32);
 850                         __m128i rss_hash0 =
 851                                 _mm_slli_epi64(raw_desc_bh0, 32);
 852
 853                         __m128i rss_hash_msk =
 854                                 _mm_set_epi32(0xFFFFFFFF, 0, 0, 0);
 855
 856                         rss_hash3 = _mm_and_si128
 857                                         (rss_hash3, rss_hash_msk);
 858                         rss_hash2 = _mm_and_si128
 859                                         (rss_hash2, rss_hash_msk);
 860                         rss_hash1 = _mm_and_si128
 861                                         (rss_hash1, rss_hash_msk);
 862                         rss_hash0 = _mm_and_si128
 863                                         (rss_hash0, rss_hash_msk);
 864
 865                         pkt_mb3 = _mm_or_si128(pkt_mb3, rss_hash3);
 866                         pkt_mb2 = _mm_or_si128(pkt_mb2, rss_hash2);
 867                         pkt_mb1 = _mm_or_si128(pkt_mb1, rss_hash1);
 868                         pkt_mb0 = _mm_or_si128(pkt_mb0, rss_hash0);
 869                 } /* if() on RSS hash parsing */
 870 #endif
 871
 872                 /* C.2 get 4 pkts staterr value  */
 873                 staterr = _mm_unpacklo_epi32(sterr_tmp1, sterr_tmp2);
 874
 875                 /* D.3 copy final 3,4 data to rx_pkts */
 876                 _mm_storeu_si128
 877                         ((void *)&rx_pkts[pos + 3]->rx_descriptor_fields1,
 878                          pkt_mb3);
 879                 _mm_storeu_si128
 880                         ((void *)&rx_pkts[pos + 2]->rx_descriptor_fields1,
 881                          pkt_mb2);
 882
 883                 /* C* extract and record EOP bit */
 884                 if (split_packet) {
 885                         /* and with mask to extract bits, flipping 1-0 */
 886                         __m128i eop_bits = _mm_andnot_si128(staterr, eop_check);
 887                         /* the staterr values are not in order, as the count
 888                          * count of dd bits doesn't care. However, for end of
 889                          * packet tracking, we do care, so shuffle. This also
 890                          * compresses the 32-bit values to 8-bit
 891                          */
 892                         eop_bits = _mm_shuffle_epi8(eop_bits, eop_shuf_mask);
 893                         /* store the resulting 32-bit value */
 894                         *(int *)split_packet = _mm_cvtsi128_si32(eop_bits);
 895                         split_packet += IAVF_VPMD_DESCS_PER_LOOP;
 896                 }
 897
 898                 /* C.3 calc available number of desc */
 899                 staterr = _mm_and_si128(staterr, dd_check);
 900                 staterr = _mm_packs_epi32(staterr, zero);
 901
 902                 /* D.3 copy final 1,2 data to rx_pkts */
 903                 _mm_storeu_si128
 904                         ((void *)&rx_pkts[pos + 1]->rx_descriptor_fields1,
 905                          pkt_mb1);
 906                 _mm_storeu_si128((void *)&rx_pkts[pos]->rx_descriptor_fields1,
 907                                  pkt_mb0);
 908                 flex_desc_to_ptype_v(descs, &rx_pkts[pos], ptype_tbl);
 909                 /* C.4 calc available number of desc */
 910                 var = __builtin_popcountll(_mm_cvtsi128_si64(staterr));
 911                 nb_pkts_recd += var;
 912                 if (likely(var != IAVF_VPMD_DESCS_PER_LOOP))
 913                         break;
 914         }
 915
 916         /* Update our internal tail pointer */
 917         rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_pkts_recd);
 918         rxq->rx_tail = (uint16_t)(rxq->rx_tail & (rxq->nb_rx_desc - 1));
 919         rxq->rxrearm_nb = (uint16_t)(rxq->rxrearm_nb + nb_pkts_recd);
 920
 921         return nb_pkts_recd;
 922 }
 923
 924 /* Notice:
 925  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
 926  * - nb_pkts > IAVF_VPMD_RX_MAX_BURST, only scan IAVF_VPMD_RX_MAX_BURST
 927  *   numbers of DD bits
 928  */
 929 uint16_t
 930 iavf_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 931                   uint16_t nb_pkts)
 932 {
 933         return _recv_raw_pkts_vec(rx_queue, rx_pkts, nb_pkts, NULL);
 934 }
 935
 936 /* Notice:
 937  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
 938  * - nb_pkts > IAVF_VPMD_RX_MAX_BURST, only scan IAVF_VPMD_RX_MAX_BURST
 939  *   numbers of DD bits
 940  */
 941 uint16_t
 942 iavf_recv_pkts_vec_flex_rxd(void *rx_queue, struct rte_mbuf **rx_pkts,
 943                             uint16_t nb_pkts)
 944 {
 945         return _recv_raw_pkts_vec_flex_rxd(rx_queue, rx_pkts, nb_pkts, NULL);
 946 }
 947
 948 /* vPMD receive routine that reassembles scattered packets
 949  * Notice:
 950  * - nb_pkts < IAVF_VPMD_DESCS_PER_LOOP, just return no packet
 951  * - nb_pkts > VPMD_RX_MAX_BURST, only scan IAVF_VPMD_RX_MAX_BURST
 952  *   numbers of DD bits
 953  */
 954 uint16_t
 955 iavf_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 956                             uint16_t nb_pkts)
 957 {
 958         struct iavf_rx_queue *rxq = rx_queue;
 959         uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
 960         unsigned int i = 0;
 961
 962         /* get some new buffers */
 963         uint16_t nb_bufs = _recv_raw_pkts_vec(rxq, rx_pkts, nb_pkts,
 964                                               split_flags);
 965         if (nb_bufs == 0)
 966                 return 0;
 967
 968         /* happy day case, full burst + no packets to be joined */
 969         const uint64_t *split_fl64 = (uint64_t *)split_flags;
 970
 971         if (!rxq->pkt_first_seg &&
 972             split_fl64[0] == 0 && split_fl64[1] == 0 &&
 973             split_fl64[2] == 0 && split_fl64[3] == 0)
 974                 return nb_bufs;
 975
 976         /* reassemble any packets that need reassembly*/
 977         if (!rxq->pkt_first_seg) {
 978                 /* find the first split flag, and only reassemble then*/
 979                 while (i < nb_bufs && !split_flags[i])
 980                         i++;
 981                 if (i == nb_bufs)
 982                         return nb_bufs;
 983                 rxq->pkt_first_seg = rx_pkts[i];
 984         }
 985         return i + reassemble_packets(rxq, &rx_pkts[i], nb_bufs - i,
 986                 &split_flags[i]);
 987 }
 988
 989 /* vPMD receive routine that reassembles scattered packets for flex RxD
 990  * Notice:
 991  * - nb_pkts < IAVF_VPMD_DESCS_PER_LOOP, just return no packet
 992  * - nb_pkts > VPMD_RX_MAX_BURST, only scan IAVF_VPMD_RX_MAX_BURST
 993  *   numbers of DD bits
 994  */
 995 uint16_t
 996 iavf_recv_scattered_pkts_vec_flex_rxd(void *rx_queue,
 997                                       struct rte_mbuf **rx_pkts,
 998                                       uint16_t nb_pkts)
 999 {
1000         struct iavf_rx_queue *rxq = rx_queue;
1001         uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
1002         unsigned int i = 0;
1003
1004         /* get some new buffers */
1005         uint16_t nb_bufs = _recv_raw_pkts_vec_flex_rxd(rxq, rx_pkts, nb_pkts,
1006                                               split_flags);
1007         if (nb_bufs == 0)
1008                 return 0;
1009
1010         /* happy day case, full burst + no packets to be joined */
1011         const uint64_t *split_fl64 = (uint64_t *)split_flags;
1012
1013         if (!rxq->pkt_first_seg &&
1014             split_fl64[0] == 0 && split_fl64[1] == 0 &&
1015             split_fl64[2] == 0 && split_fl64[3] == 0)
1016                 return nb_bufs;
1017
1018         /* reassemble any packets that need reassembly*/
1019         if (!rxq->pkt_first_seg) {
1020                 /* find the first split flag, and only reassemble then*/
1021                 while (i < nb_bufs && !split_flags[i])
1022                         i++;
1023                 if (i == nb_bufs)
1024                         return nb_bufs;
1025                 rxq->pkt_first_seg = rx_pkts[i];
1026         }
1027         return i + reassemble_packets(rxq, &rx_pkts[i], nb_bufs - i,
1028                 &split_flags[i]);
1029 }
1030
1031 static inline void
1032 vtx1(volatile struct iavf_tx_desc *txdp, struct rte_mbuf *pkt, uint64_t flags)
1033 {
1034         uint64_t high_qw =
1035                         (IAVF_TX_DESC_DTYPE_DATA |
1036                          ((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT) |
1037                          ((uint64_t)pkt->data_len <<
1038                           IAVF_TXD_QW1_TX_BUF_SZ_SHIFT));
1039
1040         __m128i descriptor = _mm_set_epi64x(high_qw,
1041                                             pkt->buf_iova + pkt->data_off);
1042         _mm_store_si128((__m128i *)txdp, descriptor);
1043 }
1044
1045 static inline void
1046 iavf_vtx(volatile struct iavf_tx_desc *txdp, struct rte_mbuf **pkt,
1047         uint16_t nb_pkts,  uint64_t flags)
1048 {
1049         int i;
1050
1051         for (i = 0; i < nb_pkts; ++i, ++txdp, ++pkt)
1052                 vtx1(txdp, *pkt, flags);
1053 }
1054
1055 uint16_t
1056 iavf_xmit_fixed_burst_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
1057                          uint16_t nb_pkts)
1058 {
1059         struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
1060         volatile struct iavf_tx_desc *txdp;
1061         struct iavf_tx_entry *txep;
1062         uint16_t n, nb_commit, tx_id;
1063         uint64_t flags = IAVF_TX_DESC_CMD_EOP | 0x04;  /* bit 2 must be set */
1064         uint64_t rs = IAVF_TX_DESC_CMD_RS | flags;
1065         int i;
1066
1067         /* cross rx_thresh boundary is not allowed */
1068         nb_pkts = RTE_MIN(nb_pkts, txq->rs_thresh);
1069
1070         if (txq->nb_free < txq->free_thresh)
1071                 iavf_tx_free_bufs(txq);
1072
1073         nb_pkts = (uint16_t)RTE_MIN(txq->nb_free, nb_pkts);
1074         if (unlikely(nb_pkts == 0))
1075                 return 0;
1076         nb_commit = nb_pkts;
1077
1078         tx_id = txq->tx_tail;
1079         txdp = &txq->tx_ring[tx_id];
1080         txep = &txq->sw_ring[tx_id];
1081
1082         txq->nb_free = (uint16_t)(txq->nb_free - nb_pkts);
1083
1084         n = (uint16_t)(txq->nb_tx_desc - tx_id);
1085         if (nb_commit >= n) {
1086                 tx_backlog_entry(txep, tx_pkts, n);
1087
1088                 for (i = 0; i < n - 1; ++i, ++tx_pkts, ++txdp)
1089                         vtx1(txdp, *tx_pkts, flags);
1090
1091                 vtx1(txdp, *tx_pkts++, rs);
1092
1093                 nb_commit = (uint16_t)(nb_commit - n);
1094
1095                 tx_id = 0;
1096                 txq->next_rs = (uint16_t)(txq->rs_thresh - 1);
1097
1098                 /* avoid reach the end of ring */
1099                 txdp = &txq->tx_ring[tx_id];
1100                 txep = &txq->sw_ring[tx_id];
1101         }
1102
1103         tx_backlog_entry(txep, tx_pkts, nb_commit);
1104
1105         iavf_vtx(txdp, tx_pkts, nb_commit, flags);
1106
1107         tx_id = (uint16_t)(tx_id + nb_commit);
1108         if (tx_id > txq->next_rs) {
1109                 txq->tx_ring[txq->next_rs].cmd_type_offset_bsz |=
1110                         rte_cpu_to_le_64(((uint64_t)IAVF_TX_DESC_CMD_RS) <<
1111                                          IAVF_TXD_QW1_CMD_SHIFT);
1112                 txq->next_rs =
1113                         (uint16_t)(txq->next_rs + txq->rs_thresh);
1114         }
1115
1116         txq->tx_tail = tx_id;
1117
1118         PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u tx_tail=%u nb_pkts=%u",
1119                    txq->port_id, txq->queue_id, tx_id, nb_pkts);
1120
1121         IAVF_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
1122
1123         return nb_pkts;
1124 }
1125
1126 uint16_t
1127 iavf_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
1128                    uint16_t nb_pkts)
1129 {
1130         uint16_t nb_tx = 0;
1131         struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
1132
1133         while (nb_pkts) {
1134                 uint16_t ret, num;
1135
1136                 num = (uint16_t)RTE_MIN(nb_pkts, txq->rs_thresh);
1137                 ret = iavf_xmit_fixed_burst_vec(tx_queue, &tx_pkts[nb_tx], num);
1138                 nb_tx += ret;
1139                 nb_pkts -= ret;
1140                 if (ret < num)
1141                         break;
1142         }
1143
1144         return nb_tx;
1145 }
1146
1147 static void __rte_cold
1148 iavf_rx_queue_release_mbufs_sse(struct iavf_rx_queue *rxq)
1149 {
1150         _iavf_rx_queue_release_mbufs_vec(rxq);
1151 }
1152
1153 static void __rte_cold
1154 iavf_tx_queue_release_mbufs_sse(struct iavf_tx_queue *txq)
1155 {
1156         _iavf_tx_queue_release_mbufs_vec(txq);
1157 }
1158
1159 static const struct iavf_rxq_ops sse_vec_rxq_ops = {
1160         .release_mbufs = iavf_rx_queue_release_mbufs_sse,
1161 };
1162
1163 static const struct iavf_txq_ops sse_vec_txq_ops = {
1164         .release_mbufs = iavf_tx_queue_release_mbufs_sse,
1165 };
1166
1167 int __rte_cold
1168 iavf_txq_vec_setup(struct iavf_tx_queue *txq)
1169 {
1170         txq->ops = &sse_vec_txq_ops;
1171         return 0;
1172 }
1173
1174 int __rte_cold
1175 iavf_rxq_vec_setup(struct iavf_rx_queue *rxq)
1176 {
1177         rxq->ops = &sse_vec_rxq_ops;
1178         return iavf_rxq_vec_setup_default(rxq);
1179 }
1180
1181 int __rte_cold
1182 iavf_rx_vec_dev_check(struct rte_eth_dev *dev)
1183 {
1184         return iavf_rx_vec_dev_check_default(dev);
1185 }
1186
1187 int __rte_cold
1188 iavf_tx_vec_dev_check(struct rte_eth_dev *dev)
1189 {
1190         return iavf_tx_vec_dev_check_default(dev);
1191 }