drivers/net/i40e/i40e_rxtx_vec_neon.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2015 Intel Corporation.
   3  * Copyright(c) 2016-2018, Linaro Limited.
   4  */
   5
   6 #include <stdint.h>
   7 #include <ethdev_driver.h>
   8 #include <rte_malloc.h>
   9 #include <rte_vect.h>
  10
  11 #include "base/i40e_prototype.h"
  12 #include "base/i40e_type.h"
  13 #include "i40e_ethdev.h"
  14 #include "i40e_rxtx.h"
  15 #include "i40e_rxtx_vec_common.h"
  16
  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         struct i40e_rx_entry *rxep = &rxq->sw_ring[rxq->rxrearm_start];
  27         struct rte_mbuf *mb0, *mb1;
  28         uint64x2_t dma_addr0, dma_addr1;
  29         uint64x2_t zero = vdupq_n_u64(0);
  30         uint64_t paddr;
  31
  32         rxdp = rxq->rx_ring + rxq->rxrearm_start;
  33
  34         /* Pull 'n' more MBUFs into the software ring */
  35         if (unlikely(rte_mempool_get_bulk(rxq->mp,
  36                                           (void *)rxep,
  37                                           RTE_I40E_RXQ_REARM_THRESH) < 0)) {
  38                 if (rxq->rxrearm_nb + RTE_I40E_RXQ_REARM_THRESH >=
  39                     rxq->nb_rx_desc) {
  40                         for (i = 0; i < RTE_I40E_DESCS_PER_LOOP; i++) {
  41                                 rxep[i].mbuf = &rxq->fake_mbuf;
  42                                 vst1q_u64((uint64_t *)&rxdp[i].read, zero);
  43                         }
  44                 }
  45                 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
  46                         RTE_I40E_RXQ_REARM_THRESH;
  47                 return;
  48         }
  49
  50         /* Initialize the mbufs in vector, process 2 mbufs in one loop */
  51         for (i = 0; i < RTE_I40E_RXQ_REARM_THRESH; i += 2, rxep += 2) {
  52                 mb0 = rxep[0].mbuf;
  53                 mb1 = rxep[1].mbuf;
  54
  55                 paddr = mb0->buf_iova + RTE_PKTMBUF_HEADROOM;
  56                 dma_addr0 = vdupq_n_u64(paddr);
  57
  58                 /* flush desc with pa dma_addr */
  59                 vst1q_u64((uint64_t *)&rxdp++->read, dma_addr0);
  60
  61                 paddr = mb1->buf_iova + RTE_PKTMBUF_HEADROOM;
  62                 dma_addr1 = vdupq_n_u64(paddr);
  63                 vst1q_u64((uint64_t *)&rxdp++->read, dma_addr1);
  64         }
  65
  66         rxq->rxrearm_start += RTE_I40E_RXQ_REARM_THRESH;
  67         if (rxq->rxrearm_start >= rxq->nb_rx_desc)
  68                 rxq->rxrearm_start = 0;
  69
  70         rxq->rxrearm_nb -= RTE_I40E_RXQ_REARM_THRESH;
  71
  72         rx_id = (uint16_t)((rxq->rxrearm_start == 0) ?
  73                              (rxq->nb_rx_desc - 1) : (rxq->rxrearm_start - 1));
  74
  75         rte_io_wmb();
  76         /* Update the tail pointer on the NIC */
  77         I40E_PCI_REG_WRITE_RELAXED(rxq->qrx_tail, rx_id);
  78 }
  79
  80 static inline void
  81 desc_to_olflags_v(struct i40e_rx_queue *rxq, uint64x2_t descs[4],
  82                   struct rte_mbuf **rx_pkts)
  83 {
  84         uint32x4_t vlan0, vlan1, rss, l3_l4e;
  85         const uint64x2_t mbuf_init = {rxq->mbuf_initializer, 0};
  86         uint64x2_t rearm0, rearm1, rearm2, rearm3;
  87
  88         /* mask everything except RSS, flow director and VLAN flags
  89          * bit2 is for VLAN tag, bit11 for flow director indication
  90          * bit13:12 for RSS indication.
  91          */
  92         const uint32x4_t rss_vlan_msk = {
  93                         0x1c03804, 0x1c03804, 0x1c03804, 0x1c03804};
  94
  95         const uint32x4_t cksum_mask = {
  96                         RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD |
  97                         RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
  98                         RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD,
  99                         RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD |
 100                         RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
 101                         RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD,
 102                         RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD |
 103                         RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
 104                         RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD,
 105                         RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD |
 106                         RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
 107                         RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD};
 108
 109         /* map rss and vlan type to rss hash and vlan flag */
 110         const uint8x16_t vlan_flags = {
 111                         0, 0, 0, 0,
 112                         RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED, 0,
 113                         0, 0,
 114                         0, 0, 0, 0,
 115                         0, 0, 0, 0};
 116
 117         const uint8x16_t rss_flags = {
 118                         0, RTE_MBUF_F_RX_FDIR, 0, 0,
 119                         0, 0, RTE_MBUF_F_RX_RSS_HASH,
 120                         RTE_MBUF_F_RX_RSS_HASH | RTE_MBUF_F_RX_FDIR,
 121                         0, 0, 0, 0,
 122                         0, 0, 0, 0};
 123
 124         const uint8x16_t l3_l4e_flags = {
 125                         (RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_GOOD) >> 1,
 126                         RTE_MBUF_F_RX_IP_CKSUM_BAD >> 1,
 127                         (RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_BAD) >> 1,
 128                         (RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
 129                         (RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD) >> 1,
 130                         (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
 131                         (RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
 132                          RTE_MBUF_F_RX_L4_CKSUM_BAD) >> 1,
 133                         (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
 134                          RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
 135                         0, 0, 0, 0, 0, 0, 0, 0};
 136
 137         vlan0 = vzipq_u32(vreinterpretq_u32_u64(descs[0]),
 138                           vreinterpretq_u32_u64(descs[2])).val[1];
 139         vlan1 = vzipq_u32(vreinterpretq_u32_u64(descs[1]),
 140                           vreinterpretq_u32_u64(descs[3])).val[1];
 141         vlan0 = vzipq_u32(vlan0, vlan1).val[0];
 142
 143         vlan1 = vandq_u32(vlan0, rss_vlan_msk);
 144         vlan0 = vreinterpretq_u32_u8(vqtbl1q_u8(vlan_flags,
 145                                                 vreinterpretq_u8_u32(vlan1)));
 146
 147         rss = vshrq_n_u32(vlan1, 11);
 148         rss = vreinterpretq_u32_u8(vqtbl1q_u8(rss_flags,
 149                                               vreinterpretq_u8_u32(rss)));
 150
 151         l3_l4e = vshrq_n_u32(vlan1, 22);
 152         l3_l4e = vreinterpretq_u32_u8(vqtbl1q_u8(l3_l4e_flags,
 153                                               vreinterpretq_u8_u32(l3_l4e)));
 154         /* then we shift left 1 bit */
 155         l3_l4e = vshlq_n_u32(l3_l4e, 1);
 156         /* we need to mask out the reduntant bits */
 157         l3_l4e = vandq_u32(l3_l4e, cksum_mask);
 158
 159         vlan0 = vorrq_u32(vlan0, rss);
 160         vlan0 = vorrq_u32(vlan0, l3_l4e);
 161
 162         rearm0 = vsetq_lane_u64(vgetq_lane_u32(vlan0, 0), mbuf_init, 1);
 163         rearm1 = vsetq_lane_u64(vgetq_lane_u32(vlan0, 1), mbuf_init, 1);
 164         rearm2 = vsetq_lane_u64(vgetq_lane_u32(vlan0, 2), mbuf_init, 1);
 165         rearm3 = vsetq_lane_u64(vgetq_lane_u32(vlan0, 3), mbuf_init, 1);
 166
 167         vst1q_u64((uint64_t *)&rx_pkts[0]->rearm_data, rearm0);
 168         vst1q_u64((uint64_t *)&rx_pkts[1]->rearm_data, rearm1);
 169         vst1q_u64((uint64_t *)&rx_pkts[2]->rearm_data, rearm2);
 170         vst1q_u64((uint64_t *)&rx_pkts[3]->rearm_data, rearm3);
 171 }
 172
 173 #define PKTLEN_SHIFT     10
 174 #define I40E_UINT16_BIT (CHAR_BIT * sizeof(uint16_t))
 175
 176 static inline void
 177 desc_to_ptype_v(uint64x2_t descs[4], struct rte_mbuf **__rte_restrict rx_pkts,
 178                 uint32_t *__rte_restrict ptype_tbl)
 179 {
 180         int i;
 181         uint8_t ptype;
 182         uint8x16_t tmp;
 183
 184         for (i = 0; i < 4; i++) {
 185                 tmp = vreinterpretq_u8_u64(vshrq_n_u64(descs[i], 30));
 186                 ptype = vgetq_lane_u8(tmp, 8);
 187                 rx_pkts[i]->packet_type = ptype_tbl[ptype];
 188         }
 189
 190 }
 191
 192 /**
 193  * vPMD raw receive routine, only accept(nb_pkts >= RTE_I40E_DESCS_PER_LOOP)
 194  *
 195  * Notice:
 196  * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
 197  * - floor align nb_pkts to a RTE_I40E_DESCS_PER_LOOP power-of-two
 198  */
 199 static inline uint16_t
 200 _recv_raw_pkts_vec(struct i40e_rx_queue *__rte_restrict rxq,
 201                    struct rte_mbuf **__rte_restrict rx_pkts,
 202                    uint16_t nb_pkts, uint8_t *split_packet)
 203 {
 204         volatile union i40e_rx_desc *rxdp;
 205         struct i40e_rx_entry *sw_ring;
 206         uint16_t nb_pkts_recd;
 207         int pos;
 208         uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
 209
 210         /* mask to shuffle from desc. to mbuf */
 211         uint8x16_t shuf_msk = {
 212                 0xFF, 0xFF,   /* pkt_type set as unknown */
 213                 0xFF, 0xFF,   /* pkt_type set as unknown */
 214                 14, 15,       /* octet 15~14, low 16 bits pkt_len */
 215                 0xFF, 0xFF,   /* skip high 16 bits pkt_len, zero out */
 216                 14, 15,       /* octet 15~14, 16 bits data_len */
 217                 2, 3,         /* octet 2~3, low 16 bits vlan_macip */
 218                 4, 5, 6, 7    /* octet 4~7, 32bits rss */
 219                 };
 220
 221         uint8x16_t eop_check = {
 222                 0x02, 0x00, 0x02, 0x00,
 223                 0x02, 0x00, 0x02, 0x00,
 224                 0x00, 0x00, 0x00, 0x00,
 225                 0x00, 0x00, 0x00, 0x00
 226                 };
 227
 228         uint16x8_t crc_adjust = {
 229                 0, 0,         /* ignore pkt_type field */
 230                 rxq->crc_len, /* sub crc on pkt_len */
 231                 0,            /* ignore high-16bits of pkt_len */
 232                 rxq->crc_len, /* sub crc on data_len */
 233                 0, 0, 0       /* ignore non-length fields */
 234                 };
 235
 236         /* nb_pkts has to be floor-aligned to RTE_I40E_DESCS_PER_LOOP */
 237         nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_I40E_DESCS_PER_LOOP);
 238
 239         /* Just the act of getting into the function from the application is
 240          * going to cost about 7 cycles
 241          */
 242         rxdp = rxq->rx_ring + rxq->rx_tail;
 243
 244         rte_prefetch_non_temporal(rxdp);
 245
 246         /* See if we need to rearm the RX queue - gives the prefetch a bit
 247          * of time to act
 248          */
 249         if (rxq->rxrearm_nb > RTE_I40E_RXQ_REARM_THRESH)
 250                 i40e_rxq_rearm(rxq);
 251
 252         /* Before we start moving massive data around, check to see if
 253          * there is actually a packet available
 254          */
 255         if (!(rxdp->wb.qword1.status_error_len &
 256                         rte_cpu_to_le_32(1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
 257                 return 0;
 258
 259         /* Cache is empty -> need to scan the buffer rings, but first move
 260          * the next 'n' mbufs into the cache
 261          */
 262         sw_ring = &rxq->sw_ring[rxq->rx_tail];
 263
 264         /* A. load 4 packet in one loop
 265          * [A*. mask out 4 unused dirty field in desc]
 266          * B. copy 4 mbuf point from swring to rx_pkts
 267          * C. calc the number of DD bits among the 4 packets
 268          * [C*. extract the end-of-packet bit, if requested]
 269          * D. fill info. from desc to mbuf
 270          */
 271
 272         for (pos = 0, nb_pkts_recd = 0; pos < nb_pkts;
 273                         pos += RTE_I40E_DESCS_PER_LOOP,
 274                         rxdp += RTE_I40E_DESCS_PER_LOOP) {
 275                 uint64x2_t descs[RTE_I40E_DESCS_PER_LOOP];
 276                 uint8x16_t pkt_mb1, pkt_mb2, pkt_mb3, pkt_mb4;
 277                 uint16x8x2_t sterr_tmp1, sterr_tmp2;
 278                 uint64x2_t mbp1, mbp2;
 279                 uint16x8_t staterr;
 280                 uint16x8_t tmp;
 281                 uint64_t stat;
 282
 283                 int32x4_t len_shl = {0, 0, 0, PKTLEN_SHIFT};
 284
 285                 /* B.1 load 1 mbuf point */
 286                 mbp1 = vld1q_u64((uint64_t *)&sw_ring[pos]);
 287                 /* Read desc statuses backwards to avoid race condition */
 288                 /* A.1 load 4 pkts desc */
 289                 descs[3] =  vld1q_u64((uint64_t *)(rxdp + 3));
 290
 291                 /* B.2 copy 2 mbuf point into rx_pkts  */
 292                 vst1q_u64((uint64_t *)&rx_pkts[pos], mbp1);
 293
 294                 /* B.1 load 1 mbuf point */
 295                 mbp2 = vld1q_u64((uint64_t *)&sw_ring[pos + 2]);
 296
 297                 descs[2] =  vld1q_u64((uint64_t *)(rxdp + 2));
 298                 /* B.1 load 2 mbuf point */
 299                 descs[1] =  vld1q_u64((uint64_t *)(rxdp + 1));
 300                 descs[0] =  vld1q_u64((uint64_t *)(rxdp));
 301
 302                 /* B.2 copy 2 mbuf point into rx_pkts  */
 303                 vst1q_u64((uint64_t *)&rx_pkts[pos + 2], mbp2);
 304
 305                 if (split_packet) {
 306                         rte_mbuf_prefetch_part2(rx_pkts[pos]);
 307                         rte_mbuf_prefetch_part2(rx_pkts[pos + 1]);
 308                         rte_mbuf_prefetch_part2(rx_pkts[pos + 2]);
 309                         rte_mbuf_prefetch_part2(rx_pkts[pos + 3]);
 310                 }
 311
 312                 /* pkt 3,4 shift the pktlen field to be 16-bit aligned*/
 313                 uint32x4_t len3 = vshlq_u32(vreinterpretq_u32_u64(descs[3]),
 314                                             len_shl);
 315                 descs[3] = vreinterpretq_u64_u16(vsetq_lane_u16
 316                                 (vgetq_lane_u16(vreinterpretq_u16_u32(len3), 7),
 317                                  vreinterpretq_u16_u64(descs[3]),
 318                                  7));
 319                 uint32x4_t len2 = vshlq_u32(vreinterpretq_u32_u64(descs[2]),
 320                                             len_shl);
 321                 descs[2] = vreinterpretq_u64_u16(vsetq_lane_u16
 322                                 (vgetq_lane_u16(vreinterpretq_u16_u32(len2), 7),
 323                                  vreinterpretq_u16_u64(descs[2]),
 324                                  7));
 325
 326                 /* D.1 pkt 3,4 convert format from desc to pktmbuf */
 327                 pkt_mb4 = vqtbl1q_u8(vreinterpretq_u8_u64(descs[3]), shuf_msk);
 328                 pkt_mb3 = vqtbl1q_u8(vreinterpretq_u8_u64(descs[2]), shuf_msk);
 329
 330                 /* C.1 4=>2 filter staterr info only */
 331                 sterr_tmp2 = vzipq_u16(vreinterpretq_u16_u64(descs[1]),
 332                                        vreinterpretq_u16_u64(descs[3]));
 333                 /* C.1 4=>2 filter staterr info only */
 334                 sterr_tmp1 = vzipq_u16(vreinterpretq_u16_u64(descs[0]),
 335                                        vreinterpretq_u16_u64(descs[2]));
 336
 337                 /* C.2 get 4 pkts staterr value  */
 338                 staterr = vzipq_u16(sterr_tmp1.val[1],
 339                                     sterr_tmp2.val[1]).val[0];
 340
 341                 desc_to_olflags_v(rxq, descs, &rx_pkts[pos]);
 342
 343                 /* D.2 pkt 3,4 set in_port/nb_seg and remove crc */
 344                 tmp = vsubq_u16(vreinterpretq_u16_u8(pkt_mb4), crc_adjust);
 345                 pkt_mb4 = vreinterpretq_u8_u16(tmp);
 346                 tmp = vsubq_u16(vreinterpretq_u16_u8(pkt_mb3), crc_adjust);
 347                 pkt_mb3 = vreinterpretq_u8_u16(tmp);
 348
 349                 /* pkt 1,2 shift the pktlen field to be 16-bit aligned*/
 350                 uint32x4_t len1 = vshlq_u32(vreinterpretq_u32_u64(descs[1]),
 351                                             len_shl);
 352                 descs[1] = vreinterpretq_u64_u16(vsetq_lane_u16
 353                                 (vgetq_lane_u16(vreinterpretq_u16_u32(len1), 7),
 354                                  vreinterpretq_u16_u64(descs[1]),
 355                                  7));
 356                 uint32x4_t len0 = vshlq_u32(vreinterpretq_u32_u64(descs[0]),
 357                                             len_shl);
 358                 descs[0] = vreinterpretq_u64_u16(vsetq_lane_u16
 359                                 (vgetq_lane_u16(vreinterpretq_u16_u32(len0), 7),
 360                                  vreinterpretq_u16_u64(descs[0]),
 361                                  7));
 362
 363                 /* D.1 pkt 1,2 convert format from desc to pktmbuf */
 364                 pkt_mb2 = vqtbl1q_u8(vreinterpretq_u8_u64(descs[1]), shuf_msk);
 365                 pkt_mb1 = vqtbl1q_u8(vreinterpretq_u8_u64(descs[0]), shuf_msk);
 366
 367                 /* D.3 copy final 3,4 data to rx_pkts */
 368                 vst1q_u8((void *)&rx_pkts[pos + 3]->rx_descriptor_fields1,
 369                                  pkt_mb4);
 370                 vst1q_u8((void *)&rx_pkts[pos + 2]->rx_descriptor_fields1,
 371                                  pkt_mb3);
 372
 373                 /* D.2 pkt 1,2 set in_port/nb_seg and remove crc */
 374                 tmp = vsubq_u16(vreinterpretq_u16_u8(pkt_mb2), crc_adjust);
 375                 pkt_mb2 = vreinterpretq_u8_u16(tmp);
 376                 tmp = vsubq_u16(vreinterpretq_u16_u8(pkt_mb1), crc_adjust);
 377                 pkt_mb1 = vreinterpretq_u8_u16(tmp);
 378
 379                 /* C* extract and record EOP bit */
 380                 if (split_packet) {
 381                         uint8x16_t eop_shuf_mask = {
 382                                         0x00, 0x02, 0x04, 0x06,
 383                                         0xFF, 0xFF, 0xFF, 0xFF,
 384                                         0xFF, 0xFF, 0xFF, 0xFF,
 385                                         0xFF, 0xFF, 0xFF, 0xFF};
 386                         uint8x16_t eop_bits;
 387
 388                         /* and with mask to extract bits, flipping 1-0 */
 389                         eop_bits = vmvnq_u8(vreinterpretq_u8_u16(staterr));
 390                         eop_bits = vandq_u8(eop_bits, eop_check);
 391                         /* the staterr values are not in order, as the count
 392                          * count of dd bits doesn't care. However, for end of
 393                          * packet tracking, we do care, so shuffle. This also
 394                          * compresses the 32-bit values to 8-bit
 395                          */
 396                         eop_bits = vqtbl1q_u8(eop_bits, eop_shuf_mask);
 397
 398                         /* store the resulting 32-bit value */
 399                         vst1q_lane_u32((uint32_t *)split_packet,
 400                                        vreinterpretq_u32_u8(eop_bits), 0);
 401                         split_packet += RTE_I40E_DESCS_PER_LOOP;
 402
 403                         /* zero-out next pointers */
 404                         rx_pkts[pos]->next = NULL;
 405                         rx_pkts[pos + 1]->next = NULL;
 406                         rx_pkts[pos + 2]->next = NULL;
 407                         rx_pkts[pos + 3]->next = NULL;
 408                 }
 409
 410                 staterr = vshlq_n_u16(staterr, I40E_UINT16_BIT - 1);
 411                 staterr = vreinterpretq_u16_s16(
 412                                 vshrq_n_s16(vreinterpretq_s16_u16(staterr),
 413                                             I40E_UINT16_BIT - 1));
 414                 stat = ~vgetq_lane_u64(vreinterpretq_u64_u16(staterr), 0);
 415
 416                 rte_prefetch_non_temporal(rxdp + RTE_I40E_DESCS_PER_LOOP);
 417
 418                 /* D.3 copy final 1,2 data to rx_pkts */
 419                 vst1q_u8((void *)&rx_pkts[pos + 1]->rx_descriptor_fields1,
 420                          pkt_mb2);
 421                 vst1q_u8((void *)&rx_pkts[pos]->rx_descriptor_fields1,
 422                          pkt_mb1);
 423                 desc_to_ptype_v(descs, &rx_pkts[pos], ptype_tbl);
 424                 /* C.4 calc avaialbe number of desc */
 425                 if (unlikely(stat == 0)) {
 426                         nb_pkts_recd += RTE_I40E_DESCS_PER_LOOP;
 427                 } else {
 428                         nb_pkts_recd += __builtin_ctzl(stat) / I40E_UINT16_BIT;
 429                         break;
 430                 }
 431         }
 432
 433         /* Update our internal tail pointer */
 434         rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_pkts_recd);
 435         rxq->rx_tail = (uint16_t)(rxq->rx_tail & (rxq->nb_rx_desc - 1));
 436         rxq->rxrearm_nb = (uint16_t)(rxq->rxrearm_nb + nb_pkts_recd);
 437
 438         return nb_pkts_recd;
 439 }
 440
 441  /*
 442  * Notice:
 443  * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
 444  * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
 445  *   numbers of DD bits
 446  */
 447 uint16_t
 448 i40e_recv_pkts_vec(void *__rte_restrict rx_queue,
 449                 struct rte_mbuf **__rte_restrict rx_pkts, uint16_t nb_pkts)
 450 {
 451         return _recv_raw_pkts_vec(rx_queue, rx_pkts, nb_pkts, NULL);
 452 }
 453
 454 /**
 455  * vPMD receive routine that reassembles single burst of 32 scattered packets
 456  *
 457  * Notice:
 458  * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
 459  */
 460 static uint16_t
 461 i40e_recv_scattered_burst_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 462                               uint16_t nb_pkts)
 463 {
 464
 465         struct i40e_rx_queue *rxq = rx_queue;
 466         uint8_t split_flags[RTE_I40E_VPMD_RX_BURST] = {0};
 467
 468         /* get some new buffers */
 469         uint16_t nb_bufs = _recv_raw_pkts_vec(rxq, rx_pkts, nb_pkts,
 470                         split_flags);
 471         if (nb_bufs == 0)
 472                 return 0;
 473
 474         /* happy day case, full burst + no packets to be joined */
 475         const uint64_t *split_fl64 = (uint64_t *)split_flags;
 476
 477         if (rxq->pkt_first_seg == NULL &&
 478                         split_fl64[0] == 0 && split_fl64[1] == 0 &&
 479                         split_fl64[2] == 0 && split_fl64[3] == 0)
 480                 return nb_bufs;
 481
 482         /* reassemble any packets that need reassembly*/
 483         unsigned i = 0;
 484
 485         if (rxq->pkt_first_seg == NULL) {
 486                 /* find the first split flag, and only reassemble then*/
 487                 while (i < nb_bufs && !split_flags[i])
 488                         i++;
 489                 if (i == nb_bufs)
 490                         return nb_bufs;
 491                 rxq->pkt_first_seg = rx_pkts[i];
 492         }
 493         return i + reassemble_packets(rxq, &rx_pkts[i], nb_bufs - i,
 494                 &split_flags[i]);
 495 }
 496
 497 /**
 498  * vPMD receive routine that reassembles scattered packets.
 499  */
 500 uint16_t
 501 i40e_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 502                              uint16_t nb_pkts)
 503 {
 504         uint16_t retval = 0;
 505
 506         while (nb_pkts > RTE_I40E_VPMD_RX_BURST) {
 507                 uint16_t burst;
 508
 509                 burst = i40e_recv_scattered_burst_vec(rx_queue,
 510                                                       rx_pkts + retval,
 511                                                       RTE_I40E_VPMD_RX_BURST);
 512                 retval += burst;
 513                 nb_pkts -= burst;
 514                 if (burst < RTE_I40E_VPMD_RX_BURST)
 515                         return retval;
 516         }
 517
 518         return retval + i40e_recv_scattered_burst_vec(rx_queue,
 519                                                       rx_pkts + retval,
 520                                                       nb_pkts);
 521 }
 522
 523 static inline void
 524 vtx1(volatile struct i40e_tx_desc *txdp,
 525                 struct rte_mbuf *pkt, uint64_t flags)
 526 {
 527         uint64_t high_qw = (I40E_TX_DESC_DTYPE_DATA |
 528                         ((uint64_t)flags  << I40E_TXD_QW1_CMD_SHIFT) |
 529                         ((uint64_t)pkt->data_len << I40E_TXD_QW1_TX_BUF_SZ_SHIFT));
 530
 531         uint64x2_t descriptor = {pkt->buf_iova + pkt->data_off, high_qw};
 532         vst1q_u64((uint64_t *)txdp, descriptor);
 533 }
 534
 535 static inline void
 536 vtx(volatile struct i40e_tx_desc *txdp, struct rte_mbuf **pkt,
 537                 uint16_t nb_pkts,  uint64_t flags)
 538 {
 539         int i;
 540
 541         for (i = 0; i < nb_pkts; ++i, ++txdp, ++pkt)
 542                 vtx1(txdp, *pkt, flags);
 543 }
 544
 545 uint16_t
 546 i40e_xmit_fixed_burst_vec(void *__rte_restrict tx_queue,
 547         struct rte_mbuf **__rte_restrict tx_pkts, uint16_t nb_pkts)
 548 {
 549         struct i40e_tx_queue *txq = (struct i40e_tx_queue *)tx_queue;
 550         volatile struct i40e_tx_desc *txdp;
 551         struct i40e_tx_entry *txep;
 552         uint16_t n, nb_commit, tx_id;
 553         uint64_t flags = I40E_TD_CMD;
 554         uint64_t rs = I40E_TX_DESC_CMD_RS | I40E_TD_CMD;
 555         int i;
 556
 557         /* cross rx_thresh boundary is not allowed */
 558         nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
 559
 560         if (txq->nb_tx_free < txq->tx_free_thresh)
 561                 i40e_tx_free_bufs(txq);
 562
 563         nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
 564         if (unlikely(nb_pkts == 0))
 565                 return 0;
 566
 567         tx_id = txq->tx_tail;
 568         txdp = &txq->tx_ring[tx_id];
 569         txep = &txq->sw_ring[tx_id];
 570
 571         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
 572
 573         n = (uint16_t)(txq->nb_tx_desc - tx_id);
 574         if (nb_commit >= n) {
 575                 tx_backlog_entry(txep, tx_pkts, n);
 576
 577                 for (i = 0; i < n - 1; ++i, ++tx_pkts, ++txdp)
 578                         vtx1(txdp, *tx_pkts, flags);
 579
 580                 vtx1(txdp, *tx_pkts++, rs);
 581
 582                 nb_commit = (uint16_t)(nb_commit - n);
 583
 584                 tx_id = 0;
 585                 txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
 586
 587                 /* avoid reach the end of ring */
 588                 txdp = &txq->tx_ring[tx_id];
 589                 txep = &txq->sw_ring[tx_id];
 590         }
 591
 592         tx_backlog_entry(txep, tx_pkts, nb_commit);
 593
 594         vtx(txdp, tx_pkts, nb_commit, flags);
 595
 596         tx_id = (uint16_t)(tx_id + nb_commit);
 597         if (tx_id > txq->tx_next_rs) {
 598                 txq->tx_ring[txq->tx_next_rs].cmd_type_offset_bsz |=
 599                         rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
 600                                                 I40E_TXD_QW1_CMD_SHIFT);
 601                 txq->tx_next_rs =
 602                         (uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
 603         }
 604
 605         txq->tx_tail = tx_id;
 606
 607         rte_io_wmb();
 608         I40E_PCI_REG_WRITE_RELAXED(txq->qtx_tail, tx_id);
 609
 610         return nb_pkts;
 611 }
 612
 613 void __rte_cold
 614 i40e_rx_queue_release_mbufs_vec(struct i40e_rx_queue *rxq)
 615 {
 616         _i40e_rx_queue_release_mbufs_vec(rxq);
 617 }
 618
 619 int __rte_cold
 620 i40e_rxq_vec_setup(struct i40e_rx_queue *rxq)
 621 {
 622         return i40e_rxq_vec_setup_default(rxq);
 623 }
 624
 625 int __rte_cold
 626 i40e_txq_vec_setup(struct i40e_tx_queue __rte_unused *txq)
 627 {
 628         return 0;
 629 }
 630
 631 int __rte_cold
 632 i40e_rx_vec_dev_conf_condition_check(struct rte_eth_dev *dev)
 633 {
 634         return i40e_rx_vec_dev_conf_condition_check_default(dev);
 635 }