drivers/net/i40e/i40e_rxtx_vec_altivec.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010 - 2015 Intel Corporation
   3  * Copyright(c) 2017 IBM Corporation.
   4  */
   5
   6 #include <stdint.h>
   7 #include <rte_ethdev_driver.h>
   8 #include <rte_malloc.h>
   9
  10 #include "base/i40e_prototype.h"
  11 #include "base/i40e_type.h"
  12 #include "i40e_ethdev.h"
  13 #include "i40e_rxtx.h"
  14 #include "i40e_rxtx_vec_common.h"
  15
  16 #include <rte_altivec.h>
  17
  18 #pragma GCC diagnostic ignored "-Wcast-qual"
  19
  20 static inline void
  21 i40e_rxq_rearm(struct i40e_rx_queue *rxq)
  22 {
  23         int i;
  24         uint16_t rx_id;
  25         volatile union i40e_rx_desc *rxdp;
  26
  27         struct i40e_rx_entry *rxep = &rxq->sw_ring[rxq->rxrearm_start];
  28         struct rte_mbuf *mb0, *mb1;
  29
  30         vector unsigned long hdr_room = (vector unsigned long){
  31                                                 RTE_PKTMBUF_HEADROOM,
  32                                                 RTE_PKTMBUF_HEADROOM};
  33         vector unsigned long dma_addr0, dma_addr1;
  34
  35         rxdp = rxq->rx_ring + rxq->rxrearm_start;
  36
  37         /* Pull 'n' more MBUFs into the software ring */
  38         if (rte_mempool_get_bulk(rxq->mp,
  39                                  (void *)rxep,
  40                                  RTE_I40E_RXQ_REARM_THRESH) < 0) {
  41                 if (rxq->rxrearm_nb + RTE_I40E_RXQ_REARM_THRESH >=
  42                     rxq->nb_rx_desc) {
  43                         dma_addr0 = (vector unsigned long){};
  44                         for (i = 0; i < RTE_I40E_DESCS_PER_LOOP; i++) {
  45                                 rxep[i].mbuf = &rxq->fake_mbuf;
  46                                 vec_st(dma_addr0, 0,
  47                                        (vector unsigned long *)&rxdp[i].read);
  48                         }
  49                 }
  50                 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
  51                         RTE_I40E_RXQ_REARM_THRESH;
  52                 return;
  53         }
  54
  55         /* Initialize the mbufs in vector, process 2 mbufs in one loop */
  56         for (i = 0; i < RTE_I40E_RXQ_REARM_THRESH; i += 2, rxep += 2) {
  57                 vector unsigned long vaddr0, vaddr1;
  58                 uintptr_t p0, p1;
  59
  60                 mb0 = rxep[0].mbuf;
  61                 mb1 = rxep[1].mbuf;
  62
  63                  /* Flush mbuf with pkt template.
  64                   * Data to be rearmed is 6 bytes long.
  65                   * Though, RX will overwrite ol_flags that are coming next
  66                   * anyway. So overwrite whole 8 bytes with one load:
  67                   * 6 bytes of rearm_data plus first 2 bytes of ol_flags.
  68                   */
  69                 p0 = (uintptr_t)&mb0->rearm_data;
  70                 *(uint64_t *)p0 = rxq->mbuf_initializer;
  71                 p1 = (uintptr_t)&mb1->rearm_data;
  72                 *(uint64_t *)p1 = rxq->mbuf_initializer;
  73
  74                 /* load buf_addr(lo 64bit) and buf_iova(hi 64bit) */
  75                 vaddr0 = vec_ld(0, (vector unsigned long *)&mb0->buf_addr);
  76                 vaddr1 = vec_ld(0, (vector unsigned long *)&mb1->buf_addr);
  77
  78                 /* convert pa to dma_addr hdr/data */
  79                 dma_addr0 = vec_mergel(vaddr0, vaddr0);
  80                 dma_addr1 = vec_mergel(vaddr1, vaddr1);
  81
  82                 /* add headroom to pa values */
  83                 dma_addr0 = vec_add(dma_addr0, hdr_room);
  84                 dma_addr1 = vec_add(dma_addr1, hdr_room);
  85
  86                 /* flush desc with pa dma_addr */
  87                 vec_st(dma_addr0, 0, (vector unsigned long *)&rxdp++->read);
  88                 vec_st(dma_addr1, 0, (vector unsigned long *)&rxdp++->read);
  89         }
  90
  91         rxq->rxrearm_start += RTE_I40E_RXQ_REARM_THRESH;
  92         if (rxq->rxrearm_start >= rxq->nb_rx_desc)
  93                 rxq->rxrearm_start = 0;
  94
  95         rxq->rxrearm_nb -= RTE_I40E_RXQ_REARM_THRESH;
  96
  97         rx_id = (uint16_t)((rxq->rxrearm_start == 0) ?
  98                              (rxq->nb_rx_desc - 1) : (rxq->rxrearm_start - 1));
  99
 100         /* Update the tail pointer on the NIC */
 101         I40E_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
 102 }
 103
 104 static inline void
 105 desc_to_olflags_v(vector unsigned long descs[4], struct rte_mbuf **rx_pkts)
 106 {
 107         vector unsigned int vlan0, vlan1, rss, l3_l4e;
 108
 109         /* mask everything except RSS, flow director and VLAN flags
 110          * bit2 is for VLAN tag, bit11 for flow director indication
 111          * bit13:12 for RSS indication.
 112          */
 113         const vector unsigned int rss_vlan_msk = (vector unsigned int){
 114                         (int32_t)0x1c03804, (int32_t)0x1c03804,
 115                         (int32_t)0x1c03804, (int32_t)0x1c03804};
 116
 117         /* map rss and vlan type to rss hash and vlan flag */
 118         const vector unsigned char vlan_flags = (vector unsigned char){
 119                         0, 0, 0, 0,
 120                         PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED, 0, 0, 0,
 121                         0, 0, 0, 0,
 122                         0, 0, 0, 0};
 123
 124         const vector unsigned char rss_flags = (vector unsigned char){
 125                         0, PKT_RX_FDIR, 0, 0,
 126                         0, 0, PKT_RX_RSS_HASH, PKT_RX_RSS_HASH | PKT_RX_FDIR,
 127                         0, 0, 0, 0,
 128                         0, 0, 0, 0};
 129
 130         const vector unsigned char l3_l4e_flags = (vector unsigned char){
 131                         0,
 132                         PKT_RX_IP_CKSUM_BAD,
 133                         PKT_RX_L4_CKSUM_BAD,
 134                         PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD,
 135                         PKT_RX_EIP_CKSUM_BAD,
 136                         PKT_RX_EIP_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD,
 137                         PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD,
 138                         PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD
 139                                              | PKT_RX_IP_CKSUM_BAD,
 140                         0, 0, 0, 0, 0, 0, 0, 0};
 141
 142         vlan0 = (vector unsigned int)vec_mergel(descs[0], descs[1]);
 143         vlan1 = (vector unsigned int)vec_mergel(descs[2], descs[3]);
 144         vlan0 = (vector unsigned int)vec_mergeh(vlan0, vlan1);
 145
 146         vlan1 = vec_and(vlan0, rss_vlan_msk);
 147         vlan0 = (vector unsigned int)vec_perm(vlan_flags,
 148                                         (vector unsigned char){},
 149                                         *(vector unsigned char *)&vlan1);
 150
 151         rss = vec_sr(vlan1, (vector unsigned int){11, 11, 11, 11});
 152         rss = (vector unsigned int)vec_perm(rss_flags, (vector unsigned char){},
 153                                         *(vector unsigned char *)&rss);
 154
 155         l3_l4e = vec_sr(vlan1, (vector unsigned int){22, 22, 22, 22});
 156         l3_l4e = (vector unsigned int)vec_perm(l3_l4e_flags,
 157                                         (vector unsigned char){},
 158                                         *(vector unsigned char *)&l3_l4e);
 159
 160         vlan0 = vec_or(vlan0, rss);
 161         vlan0 = vec_or(vlan0, l3_l4e);
 162
 163         rx_pkts[0]->ol_flags = (uint64_t)vlan0[2];
 164         rx_pkts[1]->ol_flags = (uint64_t)vlan0[3];
 165         rx_pkts[2]->ol_flags = (uint64_t)vlan0[0];
 166         rx_pkts[3]->ol_flags = (uint64_t)vlan0[1];
 167 }
 168
 169 #define PKTLEN_SHIFT     10
 170
 171 static inline void
 172 desc_to_ptype_v(vector unsigned long descs[4], struct rte_mbuf **rx_pkts,
 173                 uint32_t *ptype_tbl)
 174 {
 175         vector unsigned long ptype0 = vec_mergel(descs[0], descs[1]);
 176         vector unsigned long ptype1 = vec_mergel(descs[2], descs[3]);
 177
 178         ptype0 = vec_sr(ptype0, (vector unsigned long){30, 30});
 179         ptype1 = vec_sr(ptype1, (vector unsigned long){30, 30});
 180
 181         rx_pkts[0]->packet_type =
 182                 ptype_tbl[(*(vector unsigned char *)&ptype0)[0]];
 183         rx_pkts[1]->packet_type =
 184                 ptype_tbl[(*(vector unsigned char *)&ptype0)[8]];
 185         rx_pkts[2]->packet_type =
 186                 ptype_tbl[(*(vector unsigned char *)&ptype1)[0]];
 187         rx_pkts[3]->packet_type =
 188                 ptype_tbl[(*(vector unsigned char *)&ptype1)[8]];
 189 }
 190
 191  /* Notice:
 192   * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
 193   * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
 194   *   numbers of DD bits
 195   */
 196 static inline uint16_t
 197 _recv_raw_pkts_vec(struct i40e_rx_queue *rxq, struct rte_mbuf **rx_pkts,
 198                    uint16_t nb_pkts, uint8_t *split_packet)
 199 {
 200         volatile union i40e_rx_desc *rxdp;
 201         struct i40e_rx_entry *sw_ring;
 202         uint16_t nb_pkts_recd;
 203         int pos;
 204         uint64_t var;
 205         vector unsigned char shuf_msk;
 206         uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
 207
 208         vector unsigned short crc_adjust = (vector unsigned short){
 209                 0, 0,         /* ignore pkt_type field */
 210                 rxq->crc_len, /* sub crc on pkt_len */
 211                 0,            /* ignore high-16bits of pkt_len */
 212                 rxq->crc_len, /* sub crc on data_len */
 213                 0, 0, 0       /* ignore non-length fields */
 214                 };
 215         vector unsigned long dd_check, eop_check;
 216
 217         /* nb_pkts shall be less equal than RTE_I40E_MAX_RX_BURST */
 218         nb_pkts = RTE_MIN(nb_pkts, RTE_I40E_MAX_RX_BURST);
 219
 220         /* nb_pkts has to be floor-aligned to RTE_I40E_DESCS_PER_LOOP */
 221         nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_I40E_DESCS_PER_LOOP);
 222
 223         /* Just the act of getting into the function from the application is
 224          * going to cost about 7 cycles
 225          */
 226         rxdp = rxq->rx_ring + rxq->rx_tail;
 227
 228         rte_prefetch0(rxdp);
 229
 230         /* See if we need to rearm the RX queue - gives the prefetch a bit
 231          * of time to act
 232          */
 233         if (rxq->rxrearm_nb > RTE_I40E_RXQ_REARM_THRESH)
 234                 i40e_rxq_rearm(rxq);
 235
 236         /* Before we start moving massive data around, check to see if
 237          * there is actually a packet available
 238          */
 239         if (!(rxdp->wb.qword1.status_error_len &
 240                         rte_cpu_to_le_32(1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
 241                 return 0;
 242
 243         /* 4 packets DD mask */
 244         dd_check = (vector unsigned long){0x0000000100000001ULL,
 245                                           0x0000000100000001ULL};
 246
 247         /* 4 packets EOP mask */
 248         eop_check = (vector unsigned long){0x0000000200000002ULL,
 249                                            0x0000000200000002ULL};
 250
 251         /* mask to shuffle from desc. to mbuf */
 252         shuf_msk = (vector unsigned char){
 253                 0xFF, 0xFF,   /* pkt_type set as unknown */
 254                 0xFF, 0xFF,   /* pkt_type set as unknown */
 255                 14, 15,       /* octet 15~14, low 16 bits pkt_len */
 256                 0xFF, 0xFF,   /* skip high 16 bits pkt_len, zero out */
 257                 14, 15,       /* octet 15~14, 16 bits data_len */
 258                 2, 3,         /* octet 2~3, low 16 bits vlan_macip */
 259                 4, 5, 6, 7    /* octet 4~7, 32bits rss */
 260                 };
 261
 262         /* Cache is empty -> need to scan the buffer rings, but first move
 263          * the next 'n' mbufs into the cache
 264          */
 265         sw_ring = &rxq->sw_ring[rxq->rx_tail];
 266
 267         /* A. load 4 packet in one loop
 268          * [A*. mask out 4 unused dirty field in desc]
 269          * B. copy 4 mbuf point from swring to rx_pkts
 270          * C. calc the number of DD bits among the 4 packets
 271          * [C*. extract the end-of-packet bit, if requested]
 272          * D. fill info. from desc to mbuf
 273          */
 274
 275         for (pos = 0, nb_pkts_recd = 0; pos < nb_pkts;
 276                         pos += RTE_I40E_DESCS_PER_LOOP,
 277                         rxdp += RTE_I40E_DESCS_PER_LOOP) {
 278                 vector unsigned long descs[RTE_I40E_DESCS_PER_LOOP];
 279                 vector unsigned char pkt_mb1, pkt_mb2, pkt_mb3, pkt_mb4;
 280                 vector unsigned short staterr, sterr_tmp1, sterr_tmp2;
 281                 vector unsigned long mbp1, mbp2; /* two mbuf pointer
 282                                                   * in one XMM reg.
 283                                                   */
 284
 285                 /* B.1 load 1 mbuf point */
 286                 mbp1 = *(vector unsigned long *)&sw_ring[pos];
 287                 /* Read desc statuses backwards to avoid race condition */
 288                 /* A.1 load 4 pkts desc */
 289                 descs[3] = *(vector unsigned long *)(rxdp + 3);
 290                 rte_compiler_barrier();
 291
 292                 /* B.2 copy 2 mbuf point into rx_pkts  */
 293                 *(vector unsigned long *)&rx_pkts[pos] = mbp1;
 294
 295                 /* B.1 load 1 mbuf point */
 296                 mbp2 = *(vector unsigned long *)&sw_ring[pos + 2];
 297
 298                 descs[2] = *(vector unsigned long *)(rxdp + 2);
 299                 rte_compiler_barrier();
 300                 /* B.1 load 2 mbuf point */
 301                 descs[1] = *(vector unsigned long *)(rxdp + 1);
 302                 rte_compiler_barrier();
 303                 descs[0] = *(vector unsigned long *)(rxdp);
 304
 305                 /* B.2 copy 2 mbuf point into rx_pkts  */
 306                 *(vector unsigned long *)&rx_pkts[pos + 2] =  mbp2;
 307
 308                 if (split_packet) {
 309                         rte_mbuf_prefetch_part2(rx_pkts[pos]);
 310                         rte_mbuf_prefetch_part2(rx_pkts[pos + 1]);
 311                         rte_mbuf_prefetch_part2(rx_pkts[pos + 2]);
 312                         rte_mbuf_prefetch_part2(rx_pkts[pos + 3]);
 313                 }
 314
 315                 /* avoid compiler reorder optimization */
 316                 rte_compiler_barrier();
 317
 318                 /* pkt 3,4 shift the pktlen field to be 16-bit aligned*/
 319                 const vector unsigned int len3 = vec_sl(
 320                         vec_ld(0, (vector unsigned int *)&descs[3]),
 321                         (vector unsigned int){0, 0, 0, PKTLEN_SHIFT});
 322
 323                 const vector unsigned int len2 = vec_sl(
 324                         vec_ld(0, (vector unsigned int *)&descs[2]),
 325                         (vector unsigned int){0, 0, 0, PKTLEN_SHIFT});
 326
 327                 /* merge the now-aligned packet length fields back in */
 328                 descs[3] = (vector unsigned long)len3;
 329                 descs[2] = (vector unsigned long)len2;
 330
 331                 /* D.1 pkt 3,4 convert format from desc to pktmbuf */
 332                 pkt_mb4 = vec_perm((vector unsigned char)descs[3],
 333                                   (vector unsigned char){}, shuf_msk);
 334                 pkt_mb3 = vec_perm((vector unsigned char)descs[2],
 335                                   (vector unsigned char){}, shuf_msk);
 336
 337                 /* C.1 4=>2 filter staterr info only */
 338                 sterr_tmp2 = vec_mergel((vector unsigned short)descs[3],
 339                                         (vector unsigned short)descs[2]);
 340                 /* C.1 4=>2 filter staterr info only */
 341                 sterr_tmp1 = vec_mergel((vector unsigned short)descs[1],
 342                                         (vector unsigned short)descs[0]);
 343                 /* D.2 pkt 3,4 set in_port/nb_seg and remove crc */
 344                 pkt_mb4 = (vector unsigned char)vec_sub(
 345                                 (vector unsigned short)pkt_mb4, crc_adjust);
 346                 pkt_mb3 = (vector unsigned char)vec_sub(
 347                                 (vector unsigned short)pkt_mb3, crc_adjust);
 348
 349                 /* pkt 1,2 shift the pktlen field to be 16-bit aligned*/
 350                 const vector unsigned int len1 = vec_sl(
 351                         vec_ld(0, (vector unsigned int *)&descs[1]),
 352                         (vector unsigned int){0, 0, 0, PKTLEN_SHIFT});
 353                 const vector unsigned int len0 = vec_sl(
 354                         vec_ld(0, (vector unsigned int *)&descs[0]),
 355                         (vector unsigned int){0, 0, 0, PKTLEN_SHIFT});
 356
 357                 /* merge the now-aligned packet length fields back in */
 358                 descs[1] = (vector unsigned long)len1;
 359                 descs[0] = (vector unsigned long)len0;
 360
 361                 /* D.1 pkt 1,2 convert format from desc to pktmbuf */
 362                 pkt_mb2 = vec_perm((vector unsigned char)descs[1],
 363                                    (vector unsigned char){}, shuf_msk);
 364                 pkt_mb1 = vec_perm((vector unsigned char)descs[0],
 365                                    (vector unsigned char){}, shuf_msk);
 366
 367                 /* C.2 get 4 pkts staterr value  */
 368                 staterr = (vector unsigned short)vec_mergeh(
 369                                 sterr_tmp1, sterr_tmp2);
 370
 371                 /* D.3 copy final 3,4 data to rx_pkts */
 372                 vec_st(pkt_mb4, 0,
 373                  (vector unsigned char *)&rx_pkts[pos + 3]
 374                         ->rx_descriptor_fields1
 375                 );
 376                 vec_st(pkt_mb3, 0,
 377                  (vector unsigned char *)&rx_pkts[pos + 2]
 378                         ->rx_descriptor_fields1
 379                 );
 380
 381                 /* D.2 pkt 1,2 set in_port/nb_seg and remove crc */
 382                 pkt_mb2 = (vector unsigned char)vec_sub(
 383                                 (vector unsigned short)pkt_mb2, crc_adjust);
 384                 pkt_mb1 = (vector unsigned char)vec_sub(
 385                                 (vector unsigned short)pkt_mb1, crc_adjust);
 386
 387                 /* C* extract and record EOP bit */
 388                 if (split_packet) {
 389                         vector unsigned char eop_shuf_mask =
 390                                 (vector unsigned char){
 391                                         0xFF, 0xFF, 0xFF, 0xFF,
 392                                         0xFF, 0xFF, 0xFF, 0xFF,
 393                                         0xFF, 0xFF, 0xFF, 0xFF,
 394                                         0x04, 0x0C, 0x00, 0x08
 395                                 };
 396
 397                         /* and with mask to extract bits, flipping 1-0 */
 398                         vector unsigned char eop_bits = vec_and(
 399                                 (vector unsigned char)vec_nor(staterr, staterr),
 400                                 (vector unsigned char)eop_check);
 401                         /* the staterr values are not in order, as the count
 402                          * count of dd bits doesn't care. However, for end of
 403                          * packet tracking, we do care, so shuffle. This also
 404                          * compresses the 32-bit values to 8-bit
 405                          */
 406                         eop_bits = vec_perm(eop_bits, (vector unsigned char){},
 407                                             eop_shuf_mask);
 408                         /* store the resulting 32-bit value */
 409                         *split_packet = (vec_ld(0,
 410                                          (vector unsigned int *)&eop_bits))[0];
 411                         split_packet += RTE_I40E_DESCS_PER_LOOP;
 412
 413                         /* zero-out next pointers */
 414                         rx_pkts[pos]->next = NULL;
 415                         rx_pkts[pos + 1]->next = NULL;
 416                         rx_pkts[pos + 2]->next = NULL;
 417                         rx_pkts[pos + 3]->next = NULL;
 418                 }
 419
 420                 /* C.3 calc available number of desc */
 421                 staterr = vec_and(staterr, (vector unsigned short)dd_check);
 422
 423                 /* D.3 copy final 1,2 data to rx_pkts */
 424                 vec_st(pkt_mb2, 0,
 425                  (vector unsigned char *)&rx_pkts[pos + 1]
 426                         ->rx_descriptor_fields1
 427                 );
 428                 vec_st(pkt_mb1, 0,
 429                  (vector unsigned char *)&rx_pkts[pos]->rx_descriptor_fields1
 430                 );
 431                 desc_to_ptype_v(descs, &rx_pkts[pos], ptype_tbl);
 432                 desc_to_olflags_v(descs, &rx_pkts[pos]);
 433
 434                 /* C.4 calc avaialbe number of desc */
 435                 var = __builtin_popcountll((vec_ld(0,
 436                         (vector unsigned long *)&staterr)[0]));
 437                 nb_pkts_recd += var;
 438                 if (likely(var != RTE_I40E_DESCS_PER_LOOP))
 439                         break;
 440         }
 441
 442         /* Update our internal tail pointer */
 443         rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_pkts_recd);
 444         rxq->rx_tail = (uint16_t)(rxq->rx_tail & (rxq->nb_rx_desc - 1));
 445         rxq->rxrearm_nb = (uint16_t)(rxq->rxrearm_nb + nb_pkts_recd);
 446
 447         return nb_pkts_recd;
 448 }
 449
 450  /* Notice:
 451   * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
 452   * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
 453   *   numbers of DD bits
 454   */
 455 uint16_t
 456 i40e_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 457                    uint16_t nb_pkts)
 458 {
 459         return _recv_raw_pkts_vec(rx_queue, rx_pkts, nb_pkts, NULL);
 460 }
 461
 462  /* vPMD receive routine that reassembles scattered packets
 463   * Notice:
 464   * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
 465   * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
 466   *   numbers of DD bits
 467   */
 468 uint16_t
 469 i40e_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 470                              uint16_t nb_pkts)
 471 {
 472         struct i40e_rx_queue *rxq = rx_queue;
 473         uint8_t split_flags[RTE_I40E_VPMD_RX_BURST] = {0};
 474
 475         /* get some new buffers */
 476         uint16_t nb_bufs = _recv_raw_pkts_vec(rxq, rx_pkts, nb_pkts,
 477                         split_flags);
 478         if (nb_bufs == 0)
 479                 return 0;
 480
 481         /* happy day case, full burst + no packets to be joined */
 482         const uint64_t *split_fl64 = (uint64_t *)split_flags;
 483
 484         if (rxq->pkt_first_seg == NULL &&
 485             split_fl64[0] == 0 && split_fl64[1] == 0 &&
 486             split_fl64[2] == 0 && split_fl64[3] == 0)
 487                 return nb_bufs;
 488
 489         /* reassemble any packets that need reassembly*/
 490         unsigned int i = 0;
 491
 492         if (!rxq->pkt_first_seg) {
 493                 /* find the first split flag, and only reassemble then*/
 494                 while (i < nb_bufs && !split_flags[i])
 495                         i++;
 496                 if (i == nb_bufs)
 497                         return nb_bufs;
 498         }
 499         return i + reassemble_packets(rxq, &rx_pkts[i], nb_bufs - i,
 500                 &split_flags[i]);
 501 }
 502
 503 static inline void
 504 vtx1(volatile struct i40e_tx_desc *txdp,
 505         struct rte_mbuf *pkt, uint64_t flags)
 506 {
 507         uint64_t high_qw = (I40E_TX_DESC_DTYPE_DATA |
 508                 ((uint64_t)flags  << I40E_TXD_QW1_CMD_SHIFT) |
 509                 ((uint64_t)pkt->data_len << I40E_TXD_QW1_TX_BUF_SZ_SHIFT));
 510
 511         vector unsigned long descriptor = (vector unsigned long){
 512                 pkt->buf_iova + pkt->data_off, high_qw};
 513         *(vector unsigned long *)txdp = descriptor;
 514 }
 515
 516 static inline void
 517 vtx(volatile struct i40e_tx_desc *txdp,
 518         struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
 519 {
 520         int i;
 521
 522         for (i = 0; i < nb_pkts; ++i, ++txdp, ++pkt)
 523                 vtx1(txdp, *pkt, flags);
 524 }
 525
 526 uint16_t
 527 i40e_xmit_fixed_burst_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
 528                           uint16_t nb_pkts)
 529 {
 530         struct i40e_tx_queue *txq = (struct i40e_tx_queue *)tx_queue;
 531         volatile struct i40e_tx_desc *txdp;
 532         struct i40e_tx_entry *txep;
 533         uint16_t n, nb_commit, tx_id;
 534         uint64_t flags = I40E_TD_CMD;
 535         uint64_t rs = I40E_TX_DESC_CMD_RS | I40E_TD_CMD;
 536         int i;
 537
 538         /* cross rx_thresh boundary is not allowed */
 539         nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
 540
 541         if (txq->nb_tx_free < txq->tx_free_thresh)
 542                 i40e_tx_free_bufs(txq);
 543
 544         nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
 545         nb_commit = nb_pkts;
 546         if (unlikely(nb_pkts == 0))
 547                 return 0;
 548
 549         tx_id = txq->tx_tail;
 550         txdp = &txq->tx_ring[tx_id];
 551         txep = &txq->sw_ring[tx_id];
 552
 553         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
 554
 555         n = (uint16_t)(txq->nb_tx_desc - tx_id);
 556         if (nb_commit >= n) {
 557                 tx_backlog_entry(txep, tx_pkts, n);
 558
 559                 for (i = 0; i < n - 1; ++i, ++tx_pkts, ++txdp)
 560                         vtx1(txdp, *tx_pkts, flags);
 561
 562                 vtx1(txdp, *tx_pkts++, rs);
 563
 564                 nb_commit = (uint16_t)(nb_commit - n);
 565
 566                 tx_id = 0;
 567                 txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
 568
 569                 /* avoid reach the end of ring */
 570                 txdp = &txq->tx_ring[tx_id];
 571                 txep = &txq->sw_ring[tx_id];
 572         }
 573
 574         tx_backlog_entry(txep, tx_pkts, nb_commit);
 575
 576         vtx(txdp, tx_pkts, nb_commit, flags);
 577
 578         tx_id = (uint16_t)(tx_id + nb_commit);
 579         if (tx_id > txq->tx_next_rs) {
 580                 txq->tx_ring[txq->tx_next_rs].cmd_type_offset_bsz |=
 581                         rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
 582                                                 I40E_TXD_QW1_CMD_SHIFT);
 583                 txq->tx_next_rs =
 584                         (uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
 585         }
 586
 587         txq->tx_tail = tx_id;
 588
 589         I40E_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
 590
 591         return nb_pkts;
 592 }
 593
 594 void __rte_cold
 595 i40e_rx_queue_release_mbufs_vec(struct i40e_rx_queue *rxq)
 596 {
 597         _i40e_rx_queue_release_mbufs_vec(rxq);
 598 }
 599
 600 int __rte_cold
 601 i40e_rxq_vec_setup(struct i40e_rx_queue *rxq)
 602 {
 603         return i40e_rxq_vec_setup_default(rxq);
 604 }
 605
 606 int __rte_cold
 607 i40e_txq_vec_setup(struct i40e_tx_queue __rte_unused * txq)
 608 {
 609         return 0;
 610 }
 611
 612 int __rte_cold
 613 i40e_rx_vec_dev_conf_condition_check(struct rte_eth_dev *dev)
 614 {
 615         return i40e_rx_vec_dev_conf_condition_check_default(dev);
 616 }