drivers/net/bnxt/bnxt_rxtx_vec_sse.c

   1 // SPDX-License-Identifier: BSD-3-Clause
   2 /* Copyright(c) 2019 Broadcom All rights reserved. */
   3
   4 #include <inttypes.h>
   5 #include <stdbool.h>
   6
   7 #include <rte_bitmap.h>
   8 #include <rte_byteorder.h>
   9 #include <rte_malloc.h>
  10 #include <rte_memory.h>
  11 #if defined(RTE_ARCH_X86)
  12 #include <tmmintrin.h>
  13 #else
  14 #error "bnxt vector pmd: unsupported target."
  15 #endif
  16
  17 #include "bnxt.h"
  18 #include "bnxt_cpr.h"
  19 #include "bnxt_ring.h"
  20 #include "bnxt_rxr.h"
  21 #include "bnxt_rxq.h"
  22 #include "hsi_struct_def_dpdk.h"
  23
  24 #include "bnxt_txq.h"
  25 #include "bnxt_txr.h"
  26
  27 /*
  28  * RX Ring handling
  29  */
  30
  31 #define RTE_BNXT_MAX_RX_BURST           32
  32 #define RTE_BNXT_MAX_TX_BURST           32
  33 #define RTE_BNXT_RXQ_REARM_THRESH       32
  34 #define RTE_BNXT_DESCS_PER_LOOP         4
  35
  36 static inline void
  37 bnxt_rxq_rearm(struct bnxt_rx_queue *rxq, struct bnxt_rx_ring_info *rxr)
  38 {
  39         struct rx_prod_pkt_bd *rxbds = &rxr->rx_desc_ring[rxq->rxrearm_start];
  40         struct bnxt_sw_rx_bd *rx_bufs = &rxr->rx_buf_ring[rxq->rxrearm_start];
  41         struct rte_mbuf *mb0, *mb1;
  42         int i;
  43
  44         const __m128i hdr_room = _mm_set_epi64x(RTE_PKTMBUF_HEADROOM, 0);
  45         const __m128i addrmask = _mm_set_epi64x(UINT64_MAX, 0);
  46
  47         /* Pull RTE_BNXT_RXQ_REARM_THRESH more mbufs into the software ring */
  48         if (rte_mempool_get_bulk(rxq->mb_pool,
  49                                  (void *)rx_bufs,
  50                                  RTE_BNXT_RXQ_REARM_THRESH) < 0) {
  51                 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
  52                         RTE_BNXT_RXQ_REARM_THRESH;
  53
  54                 return;
  55         }
  56
  57         /* Initialize the mbufs in vector, process 2 mbufs in one loop */
  58         for (i = 0; i < RTE_BNXT_RXQ_REARM_THRESH; i += 2, rx_bufs += 2) {
  59                 __m128i buf_addr0, buf_addr1;
  60                 __m128i rxbd0, rxbd1;
  61
  62                 mb0 = rx_bufs[0].mbuf;
  63                 mb1 = rx_bufs[1].mbuf;
  64
  65                 /* Load address fields from both mbufs */
  66                 buf_addr0 = _mm_loadu_si128((__m128i *)&mb0->buf_addr);
  67                 buf_addr1 = _mm_loadu_si128((__m128i *)&mb1->buf_addr);
  68
  69                 /* Load both rx descriptors (preserving some existing fields) */
  70                 rxbd0 = _mm_loadu_si128((__m128i *)(rxbds + 0));
  71                 rxbd1 = _mm_loadu_si128((__m128i *)(rxbds + 1));
  72
  73                 /* Add default offset to buffer address. */
  74                 buf_addr0 = _mm_add_epi64(buf_addr0, hdr_room);
  75                 buf_addr1 = _mm_add_epi64(buf_addr1, hdr_room);
  76
  77                 /* Clear all fields except address. */
  78                 buf_addr0 =  _mm_and_si128(buf_addr0, addrmask);
  79                 buf_addr1 =  _mm_and_si128(buf_addr1, addrmask);
  80
  81                 /* Clear address field in descriptor. */
  82                 rxbd0 = _mm_andnot_si128(addrmask, rxbd0);
  83                 rxbd1 = _mm_andnot_si128(addrmask, rxbd1);
  84
  85                 /* Set address field in descriptor. */
  86                 rxbd0 = _mm_add_epi64(rxbd0, buf_addr0);
  87                 rxbd1 = _mm_add_epi64(rxbd1, buf_addr1);
  88
  89                 /* Store descriptors to memory. */
  90                 _mm_store_si128((__m128i *)(rxbds++), rxbd0);
  91                 _mm_store_si128((__m128i *)(rxbds++), rxbd1);
  92         }
  93
  94         rxq->rxrearm_start += RTE_BNXT_RXQ_REARM_THRESH;
  95         bnxt_db_write(&rxr->rx_db, rxq->rxrearm_start - 1);
  96         if (rxq->rxrearm_start >= rxq->nb_rx_desc)
  97                 rxq->rxrearm_start = 0;
  98
  99         rxq->rxrearm_nb -= RTE_BNXT_RXQ_REARM_THRESH;
 100 }
 101
 102 static uint32_t
 103 bnxt_parse_pkt_type(struct rx_pkt_cmpl *rxcmp, struct rx_pkt_cmpl_hi *rxcmp1)
 104 {
 105         uint32_t l3, pkt_type = 0;
 106         uint32_t t_ipcs = 0, ip6 = 0, vlan = 0;
 107         uint32_t flags_type;
 108
 109         vlan = !!(rxcmp1->flags2 &
 110                 rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN));
 111         pkt_type |= vlan ? RTE_PTYPE_L2_ETHER_VLAN : RTE_PTYPE_L2_ETHER;
 112
 113         t_ipcs = !!(rxcmp1->flags2 &
 114                 rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS2_T_IP_CS_CALC));
 115         ip6 = !!(rxcmp1->flags2 &
 116                  rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS2_IP_TYPE));
 117
 118         flags_type = rxcmp->flags_type &
 119                 rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS_ITYPE_MASK);
 120
 121         if (!t_ipcs && !ip6)
 122                 l3 = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN;
 123         else if (!t_ipcs && ip6)
 124                 l3 = RTE_PTYPE_L3_IPV6_EXT_UNKNOWN;
 125         else if (t_ipcs && !ip6)
 126                 l3 = RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN;
 127         else
 128                 l3 = RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN;
 129
 130         switch (flags_type) {
 131         case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_ICMP):
 132                 if (!t_ipcs)
 133                         pkt_type |= l3 | RTE_PTYPE_L4_ICMP;
 134                 else
 135                         pkt_type |= l3 | RTE_PTYPE_INNER_L4_ICMP;
 136                 break;
 137
 138         case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_TCP):
 139                 if (!t_ipcs)
 140                         pkt_type |= l3 | RTE_PTYPE_L4_TCP;
 141                 else
 142                         pkt_type |= l3 | RTE_PTYPE_INNER_L4_TCP;
 143                 break;
 144
 145         case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_UDP):
 146                 if (!t_ipcs)
 147                         pkt_type |= l3 | RTE_PTYPE_L4_UDP;
 148                 else
 149                         pkt_type |= l3 | RTE_PTYPE_INNER_L4_UDP;
 150                 break;
 151
 152         case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_IP):
 153                 pkt_type |= l3;
 154                 break;
 155         }
 156
 157         return pkt_type;
 158 }
 159
 160 static void
 161 bnxt_parse_csum(struct rte_mbuf *mbuf, struct rx_pkt_cmpl_hi *rxcmp1)
 162 {
 163         uint32_t flags;
 164
 165         flags = flags2_0xf(rxcmp1);
 166         /* IP Checksum */
 167         if (likely(IS_IP_NONTUNNEL_PKT(flags))) {
 168                 if (unlikely(RX_CMP_IP_CS_ERROR(rxcmp1)))
 169                         mbuf->ol_flags |= PKT_RX_IP_CKSUM_BAD;
 170                 else
 171                         mbuf->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
 172         } else if (IS_IP_TUNNEL_PKT(flags)) {
 173                 if (unlikely(RX_CMP_IP_OUTER_CS_ERROR(rxcmp1) ||
 174                              RX_CMP_IP_CS_ERROR(rxcmp1)))
 175                         mbuf->ol_flags |= PKT_RX_IP_CKSUM_BAD;
 176                 else
 177                         mbuf->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
 178         } else if (unlikely(RX_CMP_IP_CS_UNKNOWN(rxcmp1))) {
 179                 mbuf->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
 180         }
 181
 182         /* L4 Checksum */
 183         if (likely(IS_L4_NONTUNNEL_PKT(flags))) {
 184                 if (unlikely(RX_CMP_L4_INNER_CS_ERR2(rxcmp1)))
 185                         mbuf->ol_flags |= PKT_RX_L4_CKSUM_BAD;
 186                 else
 187                         mbuf->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
 188         } else if (IS_L4_TUNNEL_PKT(flags)) {
 189                 if (unlikely(RX_CMP_L4_INNER_CS_ERR2(rxcmp1)))
 190                         mbuf->ol_flags |= PKT_RX_L4_CKSUM_BAD;
 191                 else
 192                         mbuf->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
 193                 if (unlikely(RX_CMP_L4_OUTER_CS_ERR2(rxcmp1))) {
 194                         mbuf->ol_flags |= PKT_RX_OUTER_L4_CKSUM_BAD;
 195                 } else if (unlikely(IS_L4_TUNNEL_PKT_ONLY_INNER_L4_CS
 196                                     (flags))) {
 197                         mbuf->ol_flags |= PKT_RX_OUTER_L4_CKSUM_UNKNOWN;
 198                 } else {
 199                         mbuf->ol_flags |= PKT_RX_OUTER_L4_CKSUM_GOOD;
 200                 }
 201         } else if (unlikely(RX_CMP_L4_CS_UNKNOWN(rxcmp1))) {
 202                 mbuf->ol_flags |= PKT_RX_L4_CKSUM_UNKNOWN;
 203         }
 204 }
 205
 206 uint16_t
 207 bnxt_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 208                    uint16_t nb_pkts)
 209 {
 210         struct bnxt_rx_queue *rxq = rx_queue;
 211         struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
 212         struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
 213         uint32_t raw_cons = cpr->cp_raw_cons;
 214         uint32_t cons;
 215         int nb_rx_pkts = 0;
 216         struct rx_pkt_cmpl *rxcmp;
 217         bool evt = false;
 218         const __m128i mbuf_init = _mm_set_epi64x(0, rxq->mbuf_initializer);
 219         const __m128i shuf_msk =
 220                 _mm_set_epi8(15, 14, 13, 12,          /* rss */
 221                              0xFF, 0xFF,              /* vlan_tci (zeroes) */
 222                              3, 2,                    /* data_len */
 223                              0xFF, 0xFF, 3, 2,        /* pkt_len */
 224                              0xFF, 0xFF, 0xFF, 0xFF); /* pkt_type (zeroes) */
 225
 226         /* If Rx Q was stopped return */
 227         if (unlikely(!rxq->rx_started))
 228                 return 0;
 229
 230         if (rxq->rxrearm_nb >= RTE_BNXT_RXQ_REARM_THRESH)
 231                 bnxt_rxq_rearm(rxq, rxr);
 232
 233         /* Return no more than RTE_BNXT_MAX_RX_BURST per call. */
 234         nb_pkts = RTE_MIN(nb_pkts, RTE_BNXT_MAX_RX_BURST);
 235
 236         /* Make nb_pkts an integer multiple of RTE_BNXT_DESCS_PER_LOOP */
 237         nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_BNXT_DESCS_PER_LOOP);
 238
 239         /* Handle RX burst request */
 240         while (1) {
 241                 cons = RING_CMP(cpr->cp_ring_struct, raw_cons);
 242
 243                 rxcmp = (struct rx_pkt_cmpl *)&cpr->cp_desc_ring[cons];
 244
 245                 if (!CMP_VALID(rxcmp, raw_cons, cpr->cp_ring_struct))
 246                         break;
 247
 248                 cpr->valid = FLIP_VALID(cons,
 249                                         cpr->cp_ring_struct->ring_mask,
 250                                         cpr->valid);
 251
 252                 if (likely(CMP_TYPE(rxcmp) == RX_PKT_CMPL_TYPE_RX_L2)) {
 253                         struct rx_pkt_cmpl_hi *rxcmp1;
 254                         uint32_t tmp_raw_cons;
 255                         uint16_t cp_cons;
 256                         struct rte_mbuf *mbuf;
 257                         __m128i mm_rxcmp, pkt_mb;
 258
 259                         tmp_raw_cons = NEXT_RAW_CMP(raw_cons);
 260                         cp_cons = RING_CMP(cpr->cp_ring_struct, tmp_raw_cons);
 261                         rxcmp1 = (struct rx_pkt_cmpl_hi *)
 262                                                 &cpr->cp_desc_ring[cp_cons];
 263
 264                         if (!CMP_VALID(rxcmp1, tmp_raw_cons,
 265                                        cpr->cp_ring_struct))
 266                                 break;
 267
 268                         raw_cons = tmp_raw_cons;
 269                         cons = rxcmp->opaque;
 270
 271                         mbuf = rxr->rx_buf_ring[cons].mbuf;
 272                         rte_prefetch0(mbuf);
 273                         rxr->rx_buf_ring[cons].mbuf = NULL;
 274
 275                         cpr->valid = FLIP_VALID(cp_cons,
 276                                                 cpr->cp_ring_struct->ring_mask,
 277                                                 cpr->valid);
 278
 279                         /* Set constant fields from mbuf initializer. */
 280                         _mm_store_si128((__m128i *)&mbuf->rearm_data,
 281                                         mbuf_init);
 282
 283                         /* Set mbuf pkt_len, data_len, and rss_hash fields. */
 284                         mm_rxcmp = _mm_load_si128((__m128i *)rxcmp);
 285                         pkt_mb = _mm_shuffle_epi8(mm_rxcmp, shuf_msk);
 286                         _mm_storeu_si128((void *)&mbuf->rx_descriptor_fields1,
 287                                          pkt_mb);
 288
 289                         rte_compiler_barrier();
 290
 291                         if (rxcmp->flags_type & RX_PKT_CMPL_FLAGS_RSS_VALID)
 292                                 mbuf->ol_flags |= PKT_RX_RSS_HASH;
 293
 294                         if (rxcmp1->flags2 &
 295                             RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN) {
 296                                 mbuf->vlan_tci = rxcmp1->metadata &
 297                                         (RX_PKT_CMPL_METADATA_VID_MASK |
 298                                         RX_PKT_CMPL_METADATA_DE |
 299                                         RX_PKT_CMPL_METADATA_PRI_MASK);
 300                                 mbuf->ol_flags |= PKT_RX_VLAN;
 301                         }
 302
 303                         bnxt_parse_csum(mbuf, rxcmp1);
 304                         mbuf->packet_type = bnxt_parse_pkt_type(rxcmp, rxcmp1);
 305
 306                         rx_pkts[nb_rx_pkts++] = mbuf;
 307                 } else if (!BNXT_NUM_ASYNC_CPR(rxq->bp)) {
 308                         evt =
 309                         bnxt_event_hwrm_resp_handler(rxq->bp,
 310                                                      (struct cmpl_base *)rxcmp);
 311                 }
 312
 313                 raw_cons = NEXT_RAW_CMP(raw_cons);
 314                 if (nb_rx_pkts == nb_pkts || evt)
 315                         break;
 316         }
 317         rxr->rx_prod = RING_ADV(rxr->rx_ring_struct, rxr->rx_prod, nb_rx_pkts);
 318
 319         rxq->rxrearm_nb += nb_rx_pkts;
 320         cpr->cp_raw_cons = raw_cons;
 321         if (nb_rx_pkts || evt)
 322                 bnxt_db_cq(cpr);
 323
 324         return nb_rx_pkts;
 325 }
 326
 327 static inline void bnxt_next_cmpl(struct bnxt_cp_ring_info *cpr, uint32_t *idx,
 328                                   bool *v, uint32_t inc)
 329 {
 330         *idx += inc;
 331         if (unlikely(*idx == cpr->cp_ring_struct->ring_size)) {
 332                 *v = !*v;
 333                 *idx = 0;
 334         }
 335 }
 336
 337 static void
 338 bnxt_tx_cmp_vec(struct bnxt_tx_queue *txq, int nr_pkts)
 339 {
 340         struct bnxt_tx_ring_info *txr = txq->tx_ring;
 341         struct rte_mbuf **free = txq->free;
 342         uint16_t cons = txr->tx_cons;
 343         unsigned int blk = 0;
 344
 345         while (nr_pkts--) {
 346                 struct bnxt_sw_tx_bd *tx_buf;
 347                 struct rte_mbuf *mbuf;
 348
 349                 tx_buf = &txr->tx_buf_ring[cons];
 350                 cons = RING_NEXT(txr->tx_ring_struct, cons);
 351                 mbuf = rte_pktmbuf_prefree_seg(tx_buf->mbuf);
 352                 tx_buf->mbuf = NULL;
 353
 354                 if (blk && mbuf->pool != free[0]->pool) {
 355                         rte_mempool_put_bulk(free[0]->pool, (void **)free, blk);
 356                         blk = 0;
 357                 }
 358                 free[blk++] = mbuf;
 359         }
 360         if (blk)
 361                 rte_mempool_put_bulk(free[0]->pool, (void **)free, blk);
 362
 363         txr->tx_cons = cons;
 364 }
 365
 366 static void
 367 bnxt_handle_tx_cp_vec(struct bnxt_tx_queue *txq)
 368 {
 369         struct bnxt_cp_ring_info *cpr = txq->cp_ring;
 370         uint32_t raw_cons = cpr->cp_raw_cons;
 371         uint32_t cons;
 372         uint32_t nb_tx_pkts = 0;
 373         struct tx_cmpl *txcmp;
 374         struct cmpl_base *cp_desc_ring = cpr->cp_desc_ring;
 375         struct bnxt_ring *cp_ring_struct = cpr->cp_ring_struct;
 376         uint32_t ring_mask = cp_ring_struct->ring_mask;
 377
 378         do {
 379                 cons = RING_CMPL(ring_mask, raw_cons);
 380                 txcmp = (struct tx_cmpl *)&cp_desc_ring[cons];
 381
 382                 if (!CMPL_VALID(txcmp, cpr->valid))
 383                         break;
 384                 bnxt_next_cmpl(cpr, &cons, &cpr->valid, 1);
 385                 rte_prefetch0(&cp_desc_ring[cons]);
 386
 387                 if (likely(CMP_TYPE(txcmp) == TX_CMPL_TYPE_TX_L2))
 388                         nb_tx_pkts += txcmp->opaque;
 389                 else
 390                         RTE_LOG_DP(ERR, PMD,
 391                                    "Unhandled CMP type %02x\n",
 392                                    CMP_TYPE(txcmp));
 393                 raw_cons = cons;
 394         } while (nb_tx_pkts < ring_mask);
 395
 396         if (nb_tx_pkts) {
 397                 bnxt_tx_cmp_vec(txq, nb_tx_pkts);
 398                 cpr->cp_raw_cons = raw_cons;
 399                 bnxt_db_cq(cpr);
 400         }
 401 }
 402
 403 #define TX_BD_FLAGS_CMPL ((1 << TX_BD_LONG_FLAGS_BD_CNT_SFT) | \
 404                           TX_BD_SHORT_FLAGS_COAL_NOW | \
 405                           TX_BD_SHORT_TYPE_TX_BD_SHORT | \
 406                           TX_BD_LONG_FLAGS_PACKET_END)
 407
 408 #define TX_BD_FLAGS_NOCMPL (TX_BD_FLAGS_CMPL | TX_BD_LONG_FLAGS_NO_CMPL)
 409
 410 static inline uint32_t
 411 bnxt_xmit_flags_len(uint16_t len, uint16_t flags)
 412 {
 413         switch (len >> 9) {
 414         case 0:
 415                 return flags | TX_BD_LONG_FLAGS_LHINT_LT512;
 416         case 1:
 417                 return flags | TX_BD_LONG_FLAGS_LHINT_LT1K;
 418         case 2:
 419                 return flags | TX_BD_LONG_FLAGS_LHINT_LT2K;
 420         case 3:
 421                 return flags | TX_BD_LONG_FLAGS_LHINT_LT2K;
 422         default:
 423                 return flags | TX_BD_LONG_FLAGS_LHINT_GTE2K;
 424         }
 425 }
 426
 427 static uint16_t
 428 bnxt_xmit_fixed_burst_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
 429                           uint16_t nb_pkts)
 430 {
 431         struct bnxt_tx_queue *txq = tx_queue;
 432         struct bnxt_tx_ring_info *txr = txq->tx_ring;
 433         uint16_t prod = txr->tx_prod;
 434         struct rte_mbuf *tx_mbuf;
 435         struct tx_bd_long *txbd = NULL;
 436         struct bnxt_sw_tx_bd *tx_buf;
 437         uint16_t to_send;
 438
 439         nb_pkts = RTE_MIN(nb_pkts, bnxt_tx_avail(txq));
 440
 441         if (unlikely(nb_pkts == 0))
 442                 return 0;
 443
 444         /* Handle TX burst request */
 445         to_send = nb_pkts;
 446         while (to_send) {
 447                 tx_mbuf = *tx_pkts++;
 448                 rte_prefetch0(tx_mbuf);
 449
 450                 tx_buf = &txr->tx_buf_ring[prod];
 451                 tx_buf->mbuf = tx_mbuf;
 452                 tx_buf->nr_bds = 1;
 453
 454                 txbd = &txr->tx_desc_ring[prod];
 455                 txbd->address = tx_mbuf->buf_iova + tx_mbuf->data_off;
 456                 txbd->len = tx_mbuf->data_len;
 457                 txbd->flags_type = bnxt_xmit_flags_len(tx_mbuf->data_len,
 458                                                        TX_BD_FLAGS_NOCMPL);
 459                 prod = RING_NEXT(txr->tx_ring_struct, prod);
 460                 to_send--;
 461         }
 462
 463         /* Request a completion for last packet in burst */
 464         if (txbd) {
 465                 txbd->opaque = nb_pkts;
 466                 txbd->flags_type &= ~TX_BD_LONG_FLAGS_NO_CMPL;
 467         }
 468
 469         rte_compiler_barrier();
 470         bnxt_db_write(&txr->tx_db, prod);
 471
 472         txr->tx_prod = prod;
 473
 474         return nb_pkts;
 475 }
 476
 477 uint16_t
 478 bnxt_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
 479                    uint16_t nb_pkts)
 480 {
 481         int nb_sent = 0;
 482         struct bnxt_tx_queue *txq = tx_queue;
 483
 484         /* Tx queue was stopped; wait for it to be restarted */
 485         if (unlikely(!txq->tx_started)) {
 486                 PMD_DRV_LOG(DEBUG, "Tx q stopped;return\n");
 487                 return 0;
 488         }
 489
 490         /* Handle TX completions */
 491         if (bnxt_tx_bds_in_hw(txq) >= txq->tx_free_thresh)
 492                 bnxt_handle_tx_cp_vec(txq);
 493
 494         while (nb_pkts) {
 495                 uint16_t ret, num;
 496
 497                 num = RTE_MIN(nb_pkts, RTE_BNXT_MAX_TX_BURST);
 498                 ret = bnxt_xmit_fixed_burst_vec(tx_queue,
 499                                                 &tx_pkts[nb_sent],
 500                                                 num);
 501                 nb_sent += ret;
 502                 nb_pkts -= ret;
 503                 if (ret < num)
 504                         break;
 505         }
 506
 507         return nb_sent;
 508 }
 509
 510 int __attribute__((cold))
 511 bnxt_rxq_vec_setup(struct bnxt_rx_queue *rxq)
 512 {
 513         uintptr_t p;
 514         struct rte_mbuf mb_def = { .buf_addr = 0 }; /* zeroed mbuf */
 515
 516         mb_def.nb_segs = 1;
 517         mb_def.data_off = RTE_PKTMBUF_HEADROOM;
 518         mb_def.port = rxq->port_id;
 519         rte_mbuf_refcnt_set(&mb_def, 1);
 520
 521         /* prevent compiler reordering: rearm_data covers previous fields */
 522         rte_compiler_barrier();
 523         p = (uintptr_t)&mb_def.rearm_data;
 524         rxq->mbuf_initializer = *(uint64_t *)p;
 525         rxq->rxrearm_nb = 0;
 526         rxq->rxrearm_start = 0;
 527         return 0;
 528 }