drivers/net/i40e/i40e_rxtx_vec_sse.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2015 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <stdint.h>
  35 #include <rte_ethdev.h>
  36 #include <rte_malloc.h>
  37
  38 #include "base/i40e_prototype.h"
  39 #include "base/i40e_type.h"
  40 #include "i40e_ethdev.h"
  41 #include "i40e_rxtx.h"
  42 #include "i40e_rxtx_vec_common.h"
  43
  44 #include <tmmintrin.h>
  45
  46 #ifndef __INTEL_COMPILER
  47 #pragma GCC diagnostic ignored "-Wcast-qual"
  48 #endif
  49
  50 static inline void
  51 i40e_rxq_rearm(struct i40e_rx_queue *rxq)
  52 {
  53         int i;
  54         uint16_t rx_id;
  55         volatile union i40e_rx_desc *rxdp;
  56         struct i40e_rx_entry *rxep = &rxq->sw_ring[rxq->rxrearm_start];
  57         struct rte_mbuf *mb0, *mb1;
  58         __m128i hdr_room = _mm_set_epi64x(RTE_PKTMBUF_HEADROOM,
  59                         RTE_PKTMBUF_HEADROOM);
  60         __m128i dma_addr0, dma_addr1;
  61
  62         rxdp = rxq->rx_ring + rxq->rxrearm_start;
  63
  64         /* Pull 'n' more MBUFs into the software ring */
  65         if (rte_mempool_get_bulk(rxq->mp,
  66                                  (void *)rxep,
  67                                  RTE_I40E_RXQ_REARM_THRESH) < 0) {
  68                 if (rxq->rxrearm_nb + RTE_I40E_RXQ_REARM_THRESH >=
  69                     rxq->nb_rx_desc) {
  70                         dma_addr0 = _mm_setzero_si128();
  71                         for (i = 0; i < RTE_I40E_DESCS_PER_LOOP; i++) {
  72                                 rxep[i].mbuf = &rxq->fake_mbuf;
  73                                 _mm_store_si128((__m128i *)&rxdp[i].read,
  74                                                 dma_addr0);
  75                         }
  76                 }
  77                 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
  78                         RTE_I40E_RXQ_REARM_THRESH;
  79                 return;
  80         }
  81
  82         /* Initialize the mbufs in vector, process 2 mbufs in one loop */
  83         for (i = 0; i < RTE_I40E_RXQ_REARM_THRESH; i += 2, rxep += 2) {
  84                 __m128i vaddr0, vaddr1;
  85                 uintptr_t p0, p1;
  86
  87                 mb0 = rxep[0].mbuf;
  88                 mb1 = rxep[1].mbuf;
  89
  90                 /* Flush mbuf with pkt template.
  91                  * Data to be rearmed is 6 bytes long.
  92                  */
  93                 p0 = (uintptr_t)&mb0->rearm_data;
  94                 *(uint64_t *)p0 = rxq->mbuf_initializer;
  95                 p1 = (uintptr_t)&mb1->rearm_data;
  96                 *(uint64_t *)p1 = rxq->mbuf_initializer;
  97
  98                 /* load buf_addr(lo 64bit) and buf_physaddr(hi 64bit) */
  99                 vaddr0 = _mm_loadu_si128((__m128i *)&mb0->buf_addr);
 100                 vaddr1 = _mm_loadu_si128((__m128i *)&mb1->buf_addr);
 101
 102                 /* convert pa to dma_addr hdr/data */
 103                 dma_addr0 = _mm_unpackhi_epi64(vaddr0, vaddr0);
 104                 dma_addr1 = _mm_unpackhi_epi64(vaddr1, vaddr1);
 105
 106                 /* add headroom to pa values */
 107                 dma_addr0 = _mm_add_epi64(dma_addr0, hdr_room);
 108                 dma_addr1 = _mm_add_epi64(dma_addr1, hdr_room);
 109
 110                 /* flush desc with pa dma_addr */
 111                 _mm_store_si128((__m128i *)&rxdp++->read, dma_addr0);
 112                 _mm_store_si128((__m128i *)&rxdp++->read, dma_addr1);
 113         }
 114
 115         rxq->rxrearm_start += RTE_I40E_RXQ_REARM_THRESH;
 116         if (rxq->rxrearm_start >= rxq->nb_rx_desc)
 117                 rxq->rxrearm_start = 0;
 118
 119         rxq->rxrearm_nb -= RTE_I40E_RXQ_REARM_THRESH;
 120
 121         rx_id = (uint16_t)((rxq->rxrearm_start == 0) ?
 122                              (rxq->nb_rx_desc - 1) : (rxq->rxrearm_start - 1));
 123
 124         /* Update the tail pointer on the NIC */
 125         I40E_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
 126 }
 127
 128 /* Handling the offload flags (olflags) field takes computation
 129  * time when receiving packets. Therefore we provide a flag to disable
 130  * the processing of the olflags field when they are not needed. This
 131  * gives improved performance, at the cost of losing the offload info
 132  * in the received packet
 133  */
 134 #ifdef RTE_LIBRTE_I40E_RX_OLFLAGS_ENABLE
 135
 136 static inline void
 137 desc_to_olflags_v(__m128i descs[4], struct rte_mbuf **rx_pkts)
 138 {
 139         __m128i vlan0, vlan1, rss, l3_l4e;
 140
 141         /* mask everything except RSS, flow director and VLAN flags
 142          * bit2 is for VLAN tag, bit11 for flow director indication
 143          * bit13:12 for RSS indication.
 144          */
 145         const __m128i rss_vlan_msk = _mm_set_epi32(
 146                         0x1c03804, 0x1c03804, 0x1c03804, 0x1c03804);
 147
 148         const __m128i cksum_mask = _mm_set_epi32(
 149                         PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
 150                         PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
 151                         PKT_RX_EIP_CKSUM_BAD,
 152                         PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
 153                         PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
 154                         PKT_RX_EIP_CKSUM_BAD,
 155                         PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
 156                         PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
 157                         PKT_RX_EIP_CKSUM_BAD,
 158                         PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
 159                         PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
 160                         PKT_RX_EIP_CKSUM_BAD);
 161
 162         /* map rss and vlan type to rss hash and vlan flag */
 163         const __m128i vlan_flags = _mm_set_epi8(0, 0, 0, 0,
 164                         0, 0, 0, 0,
 165                         0, 0, 0, PKT_RX_VLAN_PKT | PKT_RX_VLAN_STRIPPED,
 166                         0, 0, 0, 0);
 167
 168         const __m128i rss_flags = _mm_set_epi8(0, 0, 0, 0,
 169                         0, 0, 0, 0,
 170                         PKT_RX_RSS_HASH | PKT_RX_FDIR, PKT_RX_RSS_HASH, 0, 0,
 171                         0, 0, PKT_RX_FDIR, 0);
 172
 173         const __m128i l3_l4e_flags = _mm_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
 174                         /* shift right 1 bit to make sure it not exceed 255 */
 175                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
 176                          PKT_RX_IP_CKSUM_BAD) >> 1,
 177                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD |
 178                          PKT_RX_L4_CKSUM_BAD) >> 1,
 179                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
 180                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD) >> 1,
 181                         (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
 182                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD) >> 1,
 183                         PKT_RX_IP_CKSUM_BAD >> 1,
 184                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1);
 185
 186         vlan0 = _mm_unpackhi_epi32(descs[0], descs[1]);
 187         vlan1 = _mm_unpackhi_epi32(descs[2], descs[3]);
 188         vlan0 = _mm_unpacklo_epi64(vlan0, vlan1);
 189
 190         vlan1 = _mm_and_si128(vlan0, rss_vlan_msk);
 191         vlan0 = _mm_shuffle_epi8(vlan_flags, vlan1);
 192
 193         rss = _mm_srli_epi32(vlan1, 11);
 194         rss = _mm_shuffle_epi8(rss_flags, rss);
 195
 196         l3_l4e = _mm_srli_epi32(vlan1, 22);
 197         l3_l4e = _mm_shuffle_epi8(l3_l4e_flags, l3_l4e);
 198         /* then we shift left 1 bit */
 199         l3_l4e = _mm_slli_epi32(l3_l4e, 1);
 200         /* we need to mask out the reduntant bits */
 201         l3_l4e = _mm_and_si128(l3_l4e, cksum_mask);
 202
 203         vlan0 = _mm_or_si128(vlan0, rss);
 204         vlan0 = _mm_or_si128(vlan0, l3_l4e);
 205
 206         rx_pkts[0]->ol_flags = _mm_extract_epi16(vlan0, 0);
 207         rx_pkts[1]->ol_flags = _mm_extract_epi16(vlan0, 2);
 208         rx_pkts[2]->ol_flags = _mm_extract_epi16(vlan0, 4);
 209         rx_pkts[3]->ol_flags = _mm_extract_epi16(vlan0, 6);
 210 }
 211 #else
 212 #define desc_to_olflags_v(desc, rx_pkts) do {} while (0)
 213 #endif
 214
 215 #define PKTLEN_SHIFT     10
 216
 217 static inline void
 218 desc_to_ptype_v(__m128i descs[4], struct rte_mbuf **rx_pkts)
 219 {
 220         __m128i ptype0 = _mm_unpackhi_epi64(descs[0], descs[1]);
 221         __m128i ptype1 = _mm_unpackhi_epi64(descs[2], descs[3]);
 222
 223         ptype0 = _mm_srli_epi64(ptype0, 30);
 224         ptype1 = _mm_srli_epi64(ptype1, 30);
 225
 226         rx_pkts[0]->packet_type = i40e_rxd_pkt_type_mapping(_mm_extract_epi8(ptype0, 0));
 227         rx_pkts[1]->packet_type = i40e_rxd_pkt_type_mapping(_mm_extract_epi8(ptype0, 8));
 228         rx_pkts[2]->packet_type = i40e_rxd_pkt_type_mapping(_mm_extract_epi8(ptype1, 0));
 229         rx_pkts[3]->packet_type = i40e_rxd_pkt_type_mapping(_mm_extract_epi8(ptype1, 8));
 230 }
 231
 232  /*
 233  * Notice:
 234  * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
 235  * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
 236  *   numbers of DD bits
 237  */
 238 static inline uint16_t
 239 _recv_raw_pkts_vec(struct i40e_rx_queue *rxq, struct rte_mbuf **rx_pkts,
 240                    uint16_t nb_pkts, uint8_t *split_packet)
 241 {
 242         volatile union i40e_rx_desc *rxdp;
 243         struct i40e_rx_entry *sw_ring;
 244         uint16_t nb_pkts_recd;
 245         int pos;
 246         uint64_t var;
 247         __m128i shuf_msk;
 248
 249         __m128i crc_adjust = _mm_set_epi16(
 250                                 0, 0, 0,    /* ignore non-length fields */
 251                                 -rxq->crc_len, /* sub crc on data_len */
 252                                 0,          /* ignore high-16bits of pkt_len */
 253                                 -rxq->crc_len, /* sub crc on pkt_len */
 254                                 0, 0            /* ignore pkt_type field */
 255                         );
 256         __m128i dd_check, eop_check;
 257
 258         /* nb_pkts shall be less equal than RTE_I40E_MAX_RX_BURST */
 259         nb_pkts = RTE_MIN(nb_pkts, RTE_I40E_MAX_RX_BURST);
 260
 261         /* nb_pkts has to be floor-aligned to RTE_I40E_DESCS_PER_LOOP */
 262         nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_I40E_DESCS_PER_LOOP);
 263
 264         /* Just the act of getting into the function from the application is
 265          * going to cost about 7 cycles
 266          */
 267         rxdp = rxq->rx_ring + rxq->rx_tail;
 268
 269         rte_prefetch0(rxdp);
 270
 271         /* See if we need to rearm the RX queue - gives the prefetch a bit
 272          * of time to act
 273          */
 274         if (rxq->rxrearm_nb > RTE_I40E_RXQ_REARM_THRESH)
 275                 i40e_rxq_rearm(rxq);
 276
 277         /* Before we start moving massive data around, check to see if
 278          * there is actually a packet available
 279          */
 280         if (!(rxdp->wb.qword1.status_error_len &
 281                         rte_cpu_to_le_32(1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
 282                 return 0;
 283
 284         /* 4 packets DD mask */
 285         dd_check = _mm_set_epi64x(0x0000000100000001LL, 0x0000000100000001LL);
 286
 287         /* 4 packets EOP mask */
 288         eop_check = _mm_set_epi64x(0x0000000200000002LL, 0x0000000200000002LL);
 289
 290         /* mask to shuffle from desc. to mbuf */
 291         shuf_msk = _mm_set_epi8(
 292                 7, 6, 5, 4,  /* octet 4~7, 32bits rss */
 293                 3, 2,        /* octet 2~3, low 16 bits vlan_macip */
 294                 15, 14,      /* octet 15~14, 16 bits data_len */
 295                 0xFF, 0xFF,  /* skip high 16 bits pkt_len, zero out */
 296                 15, 14,      /* octet 15~14, low 16 bits pkt_len */
 297                 0xFF, 0xFF,  /* pkt_type set as unknown */
 298                 0xFF, 0xFF  /*pkt_type set as unknown */
 299                 );
 300
 301         /* Cache is empty -> need to scan the buffer rings, but first move
 302          * the next 'n' mbufs into the cache
 303          */
 304         sw_ring = &rxq->sw_ring[rxq->rx_tail];
 305
 306         /* A. load 4 packet in one loop
 307          * [A*. mask out 4 unused dirty field in desc]
 308          * B. copy 4 mbuf point from swring to rx_pkts
 309          * C. calc the number of DD bits among the 4 packets
 310          * [C*. extract the end-of-packet bit, if requested]
 311          * D. fill info. from desc to mbuf
 312          */
 313
 314         for (pos = 0, nb_pkts_recd = 0; pos < nb_pkts;
 315                         pos += RTE_I40E_DESCS_PER_LOOP,
 316                         rxdp += RTE_I40E_DESCS_PER_LOOP) {
 317                 __m128i descs[RTE_I40E_DESCS_PER_LOOP];
 318                 __m128i pkt_mb1, pkt_mb2, pkt_mb3, pkt_mb4;
 319                 __m128i zero, staterr, sterr_tmp1, sterr_tmp2;
 320                 __m128i mbp1, mbp2; /* two mbuf pointer in one XMM reg. */
 321
 322                 /* B.1 load 1 mbuf point */
 323                 mbp1 = _mm_loadu_si128((__m128i *)&sw_ring[pos]);
 324                 /* Read desc statuses backwards to avoid race condition */
 325                 /* A.1 load 4 pkts desc */
 326                 descs[3] = _mm_loadu_si128((__m128i *)(rxdp + 3));
 327                 rte_compiler_barrier();
 328
 329                 /* B.2 copy 2 mbuf point into rx_pkts  */
 330                 _mm_storeu_si128((__m128i *)&rx_pkts[pos], mbp1);
 331
 332                 /* B.1 load 1 mbuf point */
 333                 mbp2 = _mm_loadu_si128((__m128i *)&sw_ring[pos+2]);
 334
 335                 descs[2] = _mm_loadu_si128((__m128i *)(rxdp + 2));
 336                 rte_compiler_barrier();
 337                 /* B.1 load 2 mbuf point */
 338                 descs[1] = _mm_loadu_si128((__m128i *)(rxdp + 1));
 339                 rte_compiler_barrier();
 340                 descs[0] = _mm_loadu_si128((__m128i *)(rxdp));
 341
 342                 /* B.2 copy 2 mbuf point into rx_pkts  */
 343                 _mm_storeu_si128((__m128i *)&rx_pkts[pos+2], mbp2);
 344
 345                 if (split_packet) {
 346                         rte_mbuf_prefetch_part2(rx_pkts[pos]);
 347                         rte_mbuf_prefetch_part2(rx_pkts[pos + 1]);
 348                         rte_mbuf_prefetch_part2(rx_pkts[pos + 2]);
 349                         rte_mbuf_prefetch_part2(rx_pkts[pos + 3]);
 350                 }
 351
 352                 /* avoid compiler reorder optimization */
 353                 rte_compiler_barrier();
 354
 355                 /* pkt 3,4 shift the pktlen field to be 16-bit aligned*/
 356                 const __m128i len3 = _mm_slli_epi32(descs[3], PKTLEN_SHIFT);
 357                 const __m128i len2 = _mm_slli_epi32(descs[2], PKTLEN_SHIFT);
 358
 359                 /* merge the now-aligned packet length fields back in */
 360                 descs[3] = _mm_blend_epi16(descs[3], len3, 0x80);
 361                 descs[2] = _mm_blend_epi16(descs[2], len2, 0x80);
 362
 363                 /* D.1 pkt 3,4 convert format from desc to pktmbuf */
 364                 pkt_mb4 = _mm_shuffle_epi8(descs[3], shuf_msk);
 365                 pkt_mb3 = _mm_shuffle_epi8(descs[2], shuf_msk);
 366
 367                 /* C.1 4=>2 filter staterr info only */
 368                 sterr_tmp2 = _mm_unpackhi_epi32(descs[3], descs[2]);
 369                 /* C.1 4=>2 filter staterr info only */
 370                 sterr_tmp1 = _mm_unpackhi_epi32(descs[1], descs[0]);
 371
 372                 desc_to_olflags_v(descs, &rx_pkts[pos]);
 373
 374                 /* D.2 pkt 3,4 set in_port/nb_seg and remove crc */
 375                 pkt_mb4 = _mm_add_epi16(pkt_mb4, crc_adjust);
 376                 pkt_mb3 = _mm_add_epi16(pkt_mb3, crc_adjust);
 377
 378                 /* pkt 1,2 shift the pktlen field to be 16-bit aligned*/
 379                 const __m128i len1 = _mm_slli_epi32(descs[1], PKTLEN_SHIFT);
 380                 const __m128i len0 = _mm_slli_epi32(descs[0], PKTLEN_SHIFT);
 381
 382                 /* merge the now-aligned packet length fields back in */
 383                 descs[1] = _mm_blend_epi16(descs[1], len1, 0x80);
 384                 descs[0] = _mm_blend_epi16(descs[0], len0, 0x80);
 385
 386                 /* D.1 pkt 1,2 convert format from desc to pktmbuf */
 387                 pkt_mb2 = _mm_shuffle_epi8(descs[1], shuf_msk);
 388                 pkt_mb1 = _mm_shuffle_epi8(descs[0], shuf_msk);
 389
 390                 /* C.2 get 4 pkts staterr value  */
 391                 zero = _mm_xor_si128(dd_check, dd_check);
 392                 staterr = _mm_unpacklo_epi32(sterr_tmp1, sterr_tmp2);
 393
 394                 /* D.3 copy final 3,4 data to rx_pkts */
 395                 _mm_storeu_si128((void *)&rx_pkts[pos+3]->rx_descriptor_fields1,
 396                                  pkt_mb4);
 397                 _mm_storeu_si128((void *)&rx_pkts[pos+2]->rx_descriptor_fields1,
 398                                  pkt_mb3);
 399
 400                 /* D.2 pkt 1,2 set in_port/nb_seg and remove crc */
 401                 pkt_mb2 = _mm_add_epi16(pkt_mb2, crc_adjust);
 402                 pkt_mb1 = _mm_add_epi16(pkt_mb1, crc_adjust);
 403
 404                 /* C* extract and record EOP bit */
 405                 if (split_packet) {
 406                         __m128i eop_shuf_mask = _mm_set_epi8(
 407                                         0xFF, 0xFF, 0xFF, 0xFF,
 408                                         0xFF, 0xFF, 0xFF, 0xFF,
 409                                         0xFF, 0xFF, 0xFF, 0xFF,
 410                                         0x04, 0x0C, 0x00, 0x08
 411                                         );
 412
 413                         /* and with mask to extract bits, flipping 1-0 */
 414                         __m128i eop_bits = _mm_andnot_si128(staterr, eop_check);
 415                         /* the staterr values are not in order, as the count
 416                          * count of dd bits doesn't care. However, for end of
 417                          * packet tracking, we do care, so shuffle. This also
 418                          * compresses the 32-bit values to 8-bit
 419                          */
 420                         eop_bits = _mm_shuffle_epi8(eop_bits, eop_shuf_mask);
 421                         /* store the resulting 32-bit value */
 422                         *(int *)split_packet = _mm_cvtsi128_si32(eop_bits);
 423                         split_packet += RTE_I40E_DESCS_PER_LOOP;
 424                 }
 425
 426                 /* C.3 calc available number of desc */
 427                 staterr = _mm_and_si128(staterr, dd_check);
 428                 staterr = _mm_packs_epi32(staterr, zero);
 429
 430                 /* D.3 copy final 1,2 data to rx_pkts */
 431                 _mm_storeu_si128((void *)&rx_pkts[pos+1]->rx_descriptor_fields1,
 432                                  pkt_mb2);
 433                 _mm_storeu_si128((void *)&rx_pkts[pos]->rx_descriptor_fields1,
 434                                  pkt_mb1);
 435                 desc_to_ptype_v(descs, &rx_pkts[pos]);
 436                 /* C.4 calc avaialbe number of desc */
 437                 var = __builtin_popcountll(_mm_cvtsi128_si64(staterr));
 438                 nb_pkts_recd += var;
 439                 if (likely(var != RTE_I40E_DESCS_PER_LOOP))
 440                         break;
 441         }
 442
 443         /* Update our internal tail pointer */
 444         rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_pkts_recd);
 445         rxq->rx_tail = (uint16_t)(rxq->rx_tail & (rxq->nb_rx_desc - 1));
 446         rxq->rxrearm_nb = (uint16_t)(rxq->rxrearm_nb + nb_pkts_recd);
 447
 448         return nb_pkts_recd;
 449 }
 450
 451  /*
 452  * Notice:
 453  * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
 454  * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
 455  *   numbers of DD bits
 456  */
 457 uint16_t
 458 i40e_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 459                    uint16_t nb_pkts)
 460 {
 461         return _recv_raw_pkts_vec(rx_queue, rx_pkts, nb_pkts, NULL);
 462 }
 463
 464  /* vPMD receive routine that reassembles scattered packets
 465  * Notice:
 466  * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
 467  * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
 468  *   numbers of DD bits
 469  */
 470 uint16_t
 471 i40e_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 472                              uint16_t nb_pkts)
 473 {
 474
 475         struct i40e_rx_queue *rxq = rx_queue;
 476         uint8_t split_flags[RTE_I40E_VPMD_RX_BURST] = {0};
 477
 478         /* get some new buffers */
 479         uint16_t nb_bufs = _recv_raw_pkts_vec(rxq, rx_pkts, nb_pkts,
 480                         split_flags);
 481         if (nb_bufs == 0)
 482                 return 0;
 483
 484         /* happy day case, full burst + no packets to be joined */
 485         const uint64_t *split_fl64 = (uint64_t *)split_flags;
 486
 487         if (rxq->pkt_first_seg == NULL &&
 488                         split_fl64[0] == 0 && split_fl64[1] == 0 &&
 489                         split_fl64[2] == 0 && split_fl64[3] == 0)
 490                 return nb_bufs;
 491
 492         /* reassemble any packets that need reassembly*/
 493         unsigned i = 0;
 494
 495         if (rxq->pkt_first_seg == NULL) {
 496                 /* find the first split flag, and only reassemble then*/
 497                 while (i < nb_bufs && !split_flags[i])
 498                         i++;
 499                 if (i == nb_bufs)
 500                         return nb_bufs;
 501         }
 502         return i + reassemble_packets(rxq, &rx_pkts[i], nb_bufs - i,
 503                 &split_flags[i]);
 504 }
 505
 506 static inline void
 507 vtx1(volatile struct i40e_tx_desc *txdp,
 508                 struct rte_mbuf *pkt, uint64_t flags)
 509 {
 510         uint64_t high_qw = (I40E_TX_DESC_DTYPE_DATA |
 511                         ((uint64_t)flags  << I40E_TXD_QW1_CMD_SHIFT) |
 512                         ((uint64_t)pkt->data_len << I40E_TXD_QW1_TX_BUF_SZ_SHIFT));
 513
 514         __m128i descriptor = _mm_set_epi64x(high_qw,
 515                                 pkt->buf_physaddr + pkt->data_off);
 516         _mm_store_si128((__m128i *)txdp, descriptor);
 517 }
 518
 519 static inline void
 520 vtx(volatile struct i40e_tx_desc *txdp,
 521                 struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
 522 {
 523         int i;
 524
 525         for (i = 0; i < nb_pkts; ++i, ++txdp, ++pkt)
 526                 vtx1(txdp, *pkt, flags);
 527 }
 528
 529 uint16_t
 530 i40e_xmit_fixed_burst_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
 531                           uint16_t nb_pkts)
 532 {
 533         struct i40e_tx_queue *txq = (struct i40e_tx_queue *)tx_queue;
 534         volatile struct i40e_tx_desc *txdp;
 535         struct i40e_tx_entry *txep;
 536         uint16_t n, nb_commit, tx_id;
 537         uint64_t flags = I40E_TD_CMD;
 538         uint64_t rs = I40E_TX_DESC_CMD_RS | I40E_TD_CMD;
 539         int i;
 540
 541         /* cross rx_thresh boundary is not allowed */
 542         nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
 543
 544         if (txq->nb_tx_free < txq->tx_free_thresh)
 545                 i40e_tx_free_bufs(txq);
 546
 547         nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
 548         if (unlikely(nb_pkts == 0))
 549                 return 0;
 550
 551         tx_id = txq->tx_tail;
 552         txdp = &txq->tx_ring[tx_id];
 553         txep = &txq->sw_ring[tx_id];
 554
 555         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
 556
 557         n = (uint16_t)(txq->nb_tx_desc - tx_id);
 558         if (nb_commit >= n) {
 559                 tx_backlog_entry(txep, tx_pkts, n);
 560
 561                 for (i = 0; i < n - 1; ++i, ++tx_pkts, ++txdp)
 562                         vtx1(txdp, *tx_pkts, flags);
 563
 564                 vtx1(txdp, *tx_pkts++, rs);
 565
 566                 nb_commit = (uint16_t)(nb_commit - n);
 567
 568                 tx_id = 0;
 569                 txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
 570
 571                 /* avoid reach the end of ring */
 572                 txdp = &txq->tx_ring[tx_id];
 573                 txep = &txq->sw_ring[tx_id];
 574         }
 575
 576         tx_backlog_entry(txep, tx_pkts, nb_commit);
 577
 578         vtx(txdp, tx_pkts, nb_commit, flags);
 579
 580         tx_id = (uint16_t)(tx_id + nb_commit);
 581         if (tx_id > txq->tx_next_rs) {
 582                 txq->tx_ring[txq->tx_next_rs].cmd_type_offset_bsz |=
 583                         rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
 584                                                 I40E_TXD_QW1_CMD_SHIFT);
 585                 txq->tx_next_rs =
 586                         (uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
 587         }
 588
 589         txq->tx_tail = tx_id;
 590
 591         I40E_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
 592
 593         return nb_pkts;
 594 }
 595
 596 void __attribute__((cold))
 597 i40e_rx_queue_release_mbufs_vec(struct i40e_rx_queue *rxq)
 598 {
 599         _i40e_rx_queue_release_mbufs_vec(rxq);
 600 }
 601
 602 int __attribute__((cold))
 603 i40e_rxq_vec_setup(struct i40e_rx_queue *rxq)
 604 {
 605         return i40e_rxq_vec_setup_default(rxq);
 606 }
 607
 608 int __attribute__((cold))
 609 i40e_txq_vec_setup(struct i40e_tx_queue __rte_unused *txq)
 610 {
 611         return 0;
 612 }
 613
 614 int __attribute__((cold))
 615 i40e_rx_vec_dev_conf_condition_check(struct rte_eth_dev *dev)
 616 {
 617 #ifndef RTE_LIBRTE_IEEE1588
 618         /* need SSE4.1 support */
 619         if (!rte_cpu_get_flag_enabled(RTE_CPUFLAG_SSE4_1))
 620                 return -1;
 621 #endif
 622
 623         return i40e_rx_vec_dev_conf_condition_check_default(dev);
 624 }