drivers/net/i40e/i40e_rxtx.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2016 Intel Corporation
   3  */
   4
   5 #include <stdio.h>
   6 #include <stdlib.h>
   7 #include <string.h>
   8 #include <errno.h>
   9 #include <stdint.h>
  10 #include <stdarg.h>
  11 #include <unistd.h>
  12 #include <inttypes.h>
  13 #include <sys/queue.h>
  14
  15 #include <rte_string_fns.h>
  16 #include <rte_memzone.h>
  17 #include <rte_mbuf.h>
  18 #include <rte_malloc.h>
  19 #include <rte_ether.h>
  20 #include <rte_ethdev_driver.h>
  21 #include <rte_tcp.h>
  22 #include <rte_sctp.h>
  23 #include <rte_udp.h>
  24 #include <rte_ip.h>
  25 #include <rte_net.h>
  26
  27 #include "i40e_logs.h"
  28 #include "base/i40e_prototype.h"
  29 #include "base/i40e_type.h"
  30 #include "i40e_ethdev.h"
  31 #include "i40e_rxtx.h"
  32
  33 #define DEFAULT_TX_RS_THRESH   32
  34 #define DEFAULT_TX_FREE_THRESH 32
  35
  36 #define I40E_TX_MAX_BURST  32
  37
  38 #define I40E_DMA_MEM_ALIGN 4096
  39
  40 /* Base address of the HW descriptor ring should be 128B aligned. */
  41 #define I40E_RING_BASE_ALIGN    128
  42
  43 #define I40E_TXD_CMD (I40E_TX_DESC_CMD_EOP | I40E_TX_DESC_CMD_RS)
  44
  45 #ifdef RTE_LIBRTE_IEEE1588
  46 #define I40E_TX_IEEE1588_TMST PKT_TX_IEEE1588_TMST
  47 #else
  48 #define I40E_TX_IEEE1588_TMST 0
  49 #endif
  50
  51 #define I40E_TX_CKSUM_OFFLOAD_MASK (             \
  52                 PKT_TX_IP_CKSUM |                \
  53                 PKT_TX_L4_MASK |                 \
  54                 PKT_TX_TCP_SEG |                 \
  55                 PKT_TX_OUTER_IP_CKSUM)
  56
  57 #define I40E_TX_OFFLOAD_MASK (  \
  58                 PKT_TX_OUTER_IPV4 |     \
  59                 PKT_TX_OUTER_IPV6 |     \
  60                 PKT_TX_IPV4 |           \
  61                 PKT_TX_IPV6 |           \
  62                 PKT_TX_IP_CKSUM |       \
  63                 PKT_TX_L4_MASK |        \
  64                 PKT_TX_OUTER_IP_CKSUM | \
  65                 PKT_TX_TCP_SEG |        \
  66                 PKT_TX_QINQ_PKT |       \
  67                 PKT_TX_VLAN_PKT |       \
  68                 PKT_TX_TUNNEL_MASK |    \
  69                 I40E_TX_IEEE1588_TMST)
  70
  71 #define I40E_TX_OFFLOAD_NOTSUP_MASK \
  72                 (PKT_TX_OFFLOAD_MASK ^ I40E_TX_OFFLOAD_MASK)
  73
  74 static inline void
  75 i40e_rxd_to_vlan_tci(struct rte_mbuf *mb, volatile union i40e_rx_desc *rxdp)
  76 {
  77         if (rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len) &
  78                 (1 << I40E_RX_DESC_STATUS_L2TAG1P_SHIFT)) {
  79                 mb->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
  80                 mb->vlan_tci =
  81                         rte_le_to_cpu_16(rxdp->wb.qword0.lo_dword.l2tag1);
  82                 PMD_RX_LOG(DEBUG, "Descriptor l2tag1: %u",
  83                            rte_le_to_cpu_16(rxdp->wb.qword0.lo_dword.l2tag1));
  84         } else {
  85                 mb->vlan_tci = 0;
  86         }
  87 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
  88         if (rte_le_to_cpu_16(rxdp->wb.qword2.ext_status) &
  89                 (1 << I40E_RX_DESC_EXT_STATUS_L2TAG2P_SHIFT)) {
  90                 mb->ol_flags |= PKT_RX_QINQ_STRIPPED | PKT_RX_QINQ |
  91                         PKT_RX_VLAN_STRIPPED | PKT_RX_VLAN;
  92                 mb->vlan_tci_outer = mb->vlan_tci;
  93                 mb->vlan_tci = rte_le_to_cpu_16(rxdp->wb.qword2.l2tag2_2);
  94                 PMD_RX_LOG(DEBUG, "Descriptor l2tag2_1: %u, l2tag2_2: %u",
  95                            rte_le_to_cpu_16(rxdp->wb.qword2.l2tag2_1),
  96                            rte_le_to_cpu_16(rxdp->wb.qword2.l2tag2_2));
  97         } else {
  98                 mb->vlan_tci_outer = 0;
  99         }
 100 #endif
 101         PMD_RX_LOG(DEBUG, "Mbuf vlan_tci: %u, vlan_tci_outer: %u",
 102                    mb->vlan_tci, mb->vlan_tci_outer);
 103 }
 104
 105 /* Translate the rx descriptor status to pkt flags */
 106 static inline uint64_t
 107 i40e_rxd_status_to_pkt_flags(uint64_t qword)
 108 {
 109         uint64_t flags;
 110
 111         /* Check if RSS_HASH */
 112         flags = (((qword >> I40E_RX_DESC_STATUS_FLTSTAT_SHIFT) &
 113                                         I40E_RX_DESC_FLTSTAT_RSS_HASH) ==
 114                         I40E_RX_DESC_FLTSTAT_RSS_HASH) ? PKT_RX_RSS_HASH : 0;
 115
 116         /* Check if FDIR Match */
 117         flags |= (qword & (1 << I40E_RX_DESC_STATUS_FLM_SHIFT) ?
 118                                                         PKT_RX_FDIR : 0);
 119
 120         return flags;
 121 }
 122
 123 static inline uint64_t
 124 i40e_rxd_error_to_pkt_flags(uint64_t qword)
 125 {
 126         uint64_t flags = 0;
 127         uint64_t error_bits = (qword >> I40E_RXD_QW1_ERROR_SHIFT);
 128
 129 #define I40E_RX_ERR_BITS 0x3f
 130         if (likely((error_bits & I40E_RX_ERR_BITS) == 0)) {
 131                 flags |= (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD);
 132                 return flags;
 133         }
 134
 135         if (unlikely(error_bits & (1 << I40E_RX_DESC_ERROR_IPE_SHIFT)))
 136                 flags |= PKT_RX_IP_CKSUM_BAD;
 137         else
 138                 flags |= PKT_RX_IP_CKSUM_GOOD;
 139
 140         if (unlikely(error_bits & (1 << I40E_RX_DESC_ERROR_L4E_SHIFT)))
 141                 flags |= PKT_RX_L4_CKSUM_BAD;
 142         else
 143                 flags |= PKT_RX_L4_CKSUM_GOOD;
 144
 145         if (unlikely(error_bits & (1 << I40E_RX_DESC_ERROR_EIPE_SHIFT)))
 146                 flags |= PKT_RX_EIP_CKSUM_BAD;
 147
 148         return flags;
 149 }
 150
 151 /* Function to check and set the ieee1588 timesync index and get the
 152  * appropriate flags.
 153  */
 154 #ifdef RTE_LIBRTE_IEEE1588
 155 static inline uint64_t
 156 i40e_get_iee15888_flags(struct rte_mbuf *mb, uint64_t qword)
 157 {
 158         uint64_t pkt_flags = 0;
 159         uint16_t tsyn = (qword & (I40E_RXD_QW1_STATUS_TSYNVALID_MASK
 160                                   | I40E_RXD_QW1_STATUS_TSYNINDX_MASK))
 161                                     >> I40E_RX_DESC_STATUS_TSYNINDX_SHIFT;
 162
 163         if ((mb->packet_type & RTE_PTYPE_L2_MASK)
 164                         == RTE_PTYPE_L2_ETHER_TIMESYNC)
 165                 pkt_flags = PKT_RX_IEEE1588_PTP;
 166         if (tsyn & 0x04) {
 167                 pkt_flags |= PKT_RX_IEEE1588_TMST;
 168                 mb->timesync = tsyn & 0x03;
 169         }
 170
 171         return pkt_flags;
 172 }
 173 #endif
 174
 175 static inline uint64_t
 176 i40e_rxd_build_fdir(volatile union i40e_rx_desc *rxdp, struct rte_mbuf *mb)
 177 {
 178         uint64_t flags = 0;
 179 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
 180         uint16_t flexbh, flexbl;
 181
 182         flexbh = (rte_le_to_cpu_32(rxdp->wb.qword2.ext_status) >>
 183                 I40E_RX_DESC_EXT_STATUS_FLEXBH_SHIFT) &
 184                 I40E_RX_DESC_EXT_STATUS_FLEXBH_MASK;
 185         flexbl = (rte_le_to_cpu_32(rxdp->wb.qword2.ext_status) >>
 186                 I40E_RX_DESC_EXT_STATUS_FLEXBL_SHIFT) &
 187                 I40E_RX_DESC_EXT_STATUS_FLEXBL_MASK;
 188
 189
 190         if (flexbh == I40E_RX_DESC_EXT_STATUS_FLEXBH_FD_ID) {
 191                 mb->hash.fdir.hi =
 192                         rte_le_to_cpu_32(rxdp->wb.qword3.hi_dword.fd_id);
 193                 flags |= PKT_RX_FDIR_ID;
 194         } else if (flexbh == I40E_RX_DESC_EXT_STATUS_FLEXBH_FLEX) {
 195                 mb->hash.fdir.hi =
 196                         rte_le_to_cpu_32(rxdp->wb.qword3.hi_dword.flex_bytes_hi);
 197                 flags |= PKT_RX_FDIR_FLX;
 198         }
 199         if (flexbl == I40E_RX_DESC_EXT_STATUS_FLEXBL_FLEX) {
 200                 mb->hash.fdir.lo =
 201                         rte_le_to_cpu_32(rxdp->wb.qword3.lo_dword.flex_bytes_lo);
 202                 flags |= PKT_RX_FDIR_FLX;
 203         }
 204 #else
 205         mb->hash.fdir.hi =
 206                 rte_le_to_cpu_32(rxdp->wb.qword0.hi_dword.fd_id);
 207         flags |= PKT_RX_FDIR_ID;
 208 #endif
 209         return flags;
 210 }
 211
 212 static inline void
 213 i40e_parse_tunneling_params(uint64_t ol_flags,
 214                             union i40e_tx_offload tx_offload,
 215                             uint32_t *cd_tunneling)
 216 {
 217         /* EIPT: External (outer) IP header type */
 218         if (ol_flags & PKT_TX_OUTER_IP_CKSUM)
 219                 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV4;
 220         else if (ol_flags & PKT_TX_OUTER_IPV4)
 221                 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV4_NO_CSUM;
 222         else if (ol_flags & PKT_TX_OUTER_IPV6)
 223                 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV6;
 224
 225         /* EIPLEN: External (outer) IP header length, in DWords */
 226         *cd_tunneling |= (tx_offload.outer_l3_len >> 2) <<
 227                 I40E_TXD_CTX_QW0_EXT_IPLEN_SHIFT;
 228
 229         /* L4TUNT: L4 Tunneling Type */
 230         switch (ol_flags & PKT_TX_TUNNEL_MASK) {
 231         case PKT_TX_TUNNEL_IPIP:
 232                 /* for non UDP / GRE tunneling, set to 00b */
 233                 break;
 234         case PKT_TX_TUNNEL_VXLAN:
 235         case PKT_TX_TUNNEL_GENEVE:
 236                 *cd_tunneling |= I40E_TXD_CTX_UDP_TUNNELING;
 237                 break;
 238         case PKT_TX_TUNNEL_GRE:
 239                 *cd_tunneling |= I40E_TXD_CTX_GRE_TUNNELING;
 240                 break;
 241         default:
 242                 PMD_TX_LOG(ERR, "Tunnel type not supported");
 243                 return;
 244         }
 245
 246         /* L4TUNLEN: L4 Tunneling Length, in Words
 247          *
 248          * We depend on app to set rte_mbuf.l2_len correctly.
 249          * For IP in GRE it should be set to the length of the GRE
 250          * header;
 251          * for MAC in GRE or MAC in UDP it should be set to the length
 252          * of the GRE or UDP headers plus the inner MAC up to including
 253          * its last Ethertype.
 254          */
 255         *cd_tunneling |= (tx_offload.l2_len >> 1) <<
 256                 I40E_TXD_CTX_QW0_NATLEN_SHIFT;
 257 }
 258
 259 static inline void
 260 i40e_txd_enable_checksum(uint64_t ol_flags,
 261                         uint32_t *td_cmd,
 262                         uint32_t *td_offset,
 263                         union i40e_tx_offload tx_offload)
 264 {
 265         /* Set MACLEN */
 266         if (ol_flags & PKT_TX_TUNNEL_MASK)
 267                 *td_offset |= (tx_offload.outer_l2_len >> 1)
 268                                 << I40E_TX_DESC_LENGTH_MACLEN_SHIFT;
 269         else
 270                 *td_offset |= (tx_offload.l2_len >> 1)
 271                         << I40E_TX_DESC_LENGTH_MACLEN_SHIFT;
 272
 273         /* Enable L3 checksum offloads */
 274         if (ol_flags & PKT_TX_IP_CKSUM) {
 275                 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV4_CSUM;
 276                 *td_offset |= (tx_offload.l3_len >> 2)
 277                                 << I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
 278         } else if (ol_flags & PKT_TX_IPV4) {
 279                 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV4;
 280                 *td_offset |= (tx_offload.l3_len >> 2)
 281                                 << I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
 282         } else if (ol_flags & PKT_TX_IPV6) {
 283                 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV6;
 284                 *td_offset |= (tx_offload.l3_len >> 2)
 285                                 << I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
 286         }
 287
 288         if (ol_flags & PKT_TX_TCP_SEG) {
 289                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_TCP;
 290                 *td_offset |= (tx_offload.l4_len >> 2)
 291                         << I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
 292                 return;
 293         }
 294
 295         /* Enable L4 checksum offloads */
 296         switch (ol_flags & PKT_TX_L4_MASK) {
 297         case PKT_TX_TCP_CKSUM:
 298                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_TCP;
 299                 *td_offset |= (sizeof(struct rte_tcp_hdr) >> 2) <<
 300                                 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
 301                 break;
 302         case PKT_TX_SCTP_CKSUM:
 303                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_SCTP;
 304                 *td_offset |= (sizeof(struct rte_sctp_hdr) >> 2) <<
 305                                 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
 306                 break;
 307         case PKT_TX_UDP_CKSUM:
 308                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_UDP;
 309                 *td_offset |= (sizeof(struct rte_udp_hdr) >> 2) <<
 310                                 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
 311                 break;
 312         default:
 313                 break;
 314         }
 315 }
 316
 317 /* Construct the tx flags */
 318 static inline uint64_t
 319 i40e_build_ctob(uint32_t td_cmd,
 320                 uint32_t td_offset,
 321                 unsigned int size,
 322                 uint32_t td_tag)
 323 {
 324         return rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DATA |
 325                         ((uint64_t)td_cmd  << I40E_TXD_QW1_CMD_SHIFT) |
 326                         ((uint64_t)td_offset << I40E_TXD_QW1_OFFSET_SHIFT) |
 327                         ((uint64_t)size  << I40E_TXD_QW1_TX_BUF_SZ_SHIFT) |
 328                         ((uint64_t)td_tag  << I40E_TXD_QW1_L2TAG1_SHIFT));
 329 }
 330
 331 static inline int
 332 i40e_xmit_cleanup(struct i40e_tx_queue *txq)
 333 {
 334         struct i40e_tx_entry *sw_ring = txq->sw_ring;
 335         volatile struct i40e_tx_desc *txd = txq->tx_ring;
 336         uint16_t last_desc_cleaned = txq->last_desc_cleaned;
 337         uint16_t nb_tx_desc = txq->nb_tx_desc;
 338         uint16_t desc_to_clean_to;
 339         uint16_t nb_tx_to_clean;
 340
 341         desc_to_clean_to = (uint16_t)(last_desc_cleaned + txq->tx_rs_thresh);
 342         if (desc_to_clean_to >= nb_tx_desc)
 343                 desc_to_clean_to = (uint16_t)(desc_to_clean_to - nb_tx_desc);
 344
 345         desc_to_clean_to = sw_ring[desc_to_clean_to].last_id;
 346         if ((txd[desc_to_clean_to].cmd_type_offset_bsz &
 347                         rte_cpu_to_le_64(I40E_TXD_QW1_DTYPE_MASK)) !=
 348                         rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE)) {
 349                 PMD_TX_FREE_LOG(DEBUG, "TX descriptor %4u is not done "
 350                         "(port=%d queue=%d)", desc_to_clean_to,
 351                                 txq->port_id, txq->queue_id);
 352                 return -1;
 353         }
 354
 355         if (last_desc_cleaned > desc_to_clean_to)
 356                 nb_tx_to_clean = (uint16_t)((nb_tx_desc - last_desc_cleaned) +
 357                                                         desc_to_clean_to);
 358         else
 359                 nb_tx_to_clean = (uint16_t)(desc_to_clean_to -
 360                                         last_desc_cleaned);
 361
 362         txd[desc_to_clean_to].cmd_type_offset_bsz = 0;
 363
 364         txq->last_desc_cleaned = desc_to_clean_to;
 365         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + nb_tx_to_clean);
 366
 367         return 0;
 368 }
 369
 370 static inline int
 371 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
 372 check_rx_burst_bulk_alloc_preconditions(struct i40e_rx_queue *rxq)
 373 #else
 374 check_rx_burst_bulk_alloc_preconditions(__rte_unused struct i40e_rx_queue *rxq)
 375 #endif
 376 {
 377         int ret = 0;
 378
 379 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
 380         if (!(rxq->rx_free_thresh >= RTE_PMD_I40E_RX_MAX_BURST)) {
 381                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
 382                              "rxq->rx_free_thresh=%d, "
 383                              "RTE_PMD_I40E_RX_MAX_BURST=%d",
 384                              rxq->rx_free_thresh, RTE_PMD_I40E_RX_MAX_BURST);
 385                 ret = -EINVAL;
 386         } else if (!(rxq->rx_free_thresh < rxq->nb_rx_desc)) {
 387                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
 388                              "rxq->rx_free_thresh=%d, "
 389                              "rxq->nb_rx_desc=%d",
 390                              rxq->rx_free_thresh, rxq->nb_rx_desc);
 391                 ret = -EINVAL;
 392         } else if (rxq->nb_rx_desc % rxq->rx_free_thresh != 0) {
 393                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
 394                              "rxq->nb_rx_desc=%d, "
 395                              "rxq->rx_free_thresh=%d",
 396                              rxq->nb_rx_desc, rxq->rx_free_thresh);
 397                 ret = -EINVAL;
 398         }
 399 #else
 400         ret = -EINVAL;
 401 #endif
 402
 403         return ret;
 404 }
 405
 406 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
 407 #define I40E_LOOK_AHEAD 8
 408 #if (I40E_LOOK_AHEAD != 8)
 409 #error "PMD I40E: I40E_LOOK_AHEAD must be 8\n"
 410 #endif
 411 static inline int
 412 i40e_rx_scan_hw_ring(struct i40e_rx_queue *rxq)
 413 {
 414         volatile union i40e_rx_desc *rxdp;
 415         struct i40e_rx_entry *rxep;
 416         struct rte_mbuf *mb;
 417         uint16_t pkt_len;
 418         uint64_t qword1;
 419         uint32_t rx_status;
 420         int32_t s[I40E_LOOK_AHEAD], nb_dd;
 421         int32_t i, j, nb_rx = 0;
 422         uint64_t pkt_flags;
 423         uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
 424
 425         rxdp = &rxq->rx_ring[rxq->rx_tail];
 426         rxep = &rxq->sw_ring[rxq->rx_tail];
 427
 428         qword1 = rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len);
 429         rx_status = (qword1 & I40E_RXD_QW1_STATUS_MASK) >>
 430                                 I40E_RXD_QW1_STATUS_SHIFT;
 431
 432         /* Make sure there is at least 1 packet to receive */
 433         if (!(rx_status & (1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
 434                 return 0;
 435
 436         /**
 437          * Scan LOOK_AHEAD descriptors at a time to determine which
 438          * descriptors reference packets that are ready to be received.
 439          */
 440         for (i = 0; i < RTE_PMD_I40E_RX_MAX_BURST; i+=I40E_LOOK_AHEAD,
 441                         rxdp += I40E_LOOK_AHEAD, rxep += I40E_LOOK_AHEAD) {
 442                 /* Read desc statuses backwards to avoid race condition */
 443                 for (j = I40E_LOOK_AHEAD - 1; j >= 0; j--) {
 444                         qword1 = rte_le_to_cpu_64(\
 445                                 rxdp[j].wb.qword1.status_error_len);
 446                         s[j] = (qword1 & I40E_RXD_QW1_STATUS_MASK) >>
 447                                         I40E_RXD_QW1_STATUS_SHIFT;
 448                 }
 449
 450                 rte_smp_rmb();
 451
 452                 /* Compute how many status bits were set */
 453                 for (j = 0, nb_dd = 0; j < I40E_LOOK_AHEAD; j++)
 454                         nb_dd += s[j] & (1 << I40E_RX_DESC_STATUS_DD_SHIFT);
 455
 456                 nb_rx += nb_dd;
 457
 458                 /* Translate descriptor info to mbuf parameters */
 459                 for (j = 0; j < nb_dd; j++) {
 460                         mb = rxep[j].mbuf;
 461                         qword1 = rte_le_to_cpu_64(\
 462                                 rxdp[j].wb.qword1.status_error_len);
 463                         pkt_len = ((qword1 & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
 464                                 I40E_RXD_QW1_LENGTH_PBUF_SHIFT) - rxq->crc_len;
 465                         mb->data_len = pkt_len;
 466                         mb->pkt_len = pkt_len;
 467                         mb->ol_flags = 0;
 468                         i40e_rxd_to_vlan_tci(mb, &rxdp[j]);
 469                         pkt_flags = i40e_rxd_status_to_pkt_flags(qword1);
 470                         pkt_flags |= i40e_rxd_error_to_pkt_flags(qword1);
 471                         mb->packet_type =
 472                                 ptype_tbl[(uint8_t)((qword1 &
 473                                 I40E_RXD_QW1_PTYPE_MASK) >>
 474                                 I40E_RXD_QW1_PTYPE_SHIFT)];
 475                         if (pkt_flags & PKT_RX_RSS_HASH)
 476                                 mb->hash.rss = rte_le_to_cpu_32(\
 477                                         rxdp[j].wb.qword0.hi_dword.rss);
 478                         if (pkt_flags & PKT_RX_FDIR)
 479                                 pkt_flags |= i40e_rxd_build_fdir(&rxdp[j], mb);
 480
 481 #ifdef RTE_LIBRTE_IEEE1588
 482                         pkt_flags |= i40e_get_iee15888_flags(mb, qword1);
 483 #endif
 484                         mb->ol_flags |= pkt_flags;
 485
 486                 }
 487
 488                 for (j = 0; j < I40E_LOOK_AHEAD; j++)
 489                         rxq->rx_stage[i + j] = rxep[j].mbuf;
 490
 491                 if (nb_dd != I40E_LOOK_AHEAD)
 492                         break;
 493         }
 494
 495         /* Clear software ring entries */
 496         for (i = 0; i < nb_rx; i++)
 497                 rxq->sw_ring[rxq->rx_tail + i].mbuf = NULL;
 498
 499         return nb_rx;
 500 }
 501
 502 static inline uint16_t
 503 i40e_rx_fill_from_stage(struct i40e_rx_queue *rxq,
 504                         struct rte_mbuf **rx_pkts,
 505                         uint16_t nb_pkts)
 506 {
 507         uint16_t i;
 508         struct rte_mbuf **stage = &rxq->rx_stage[rxq->rx_next_avail];
 509
 510         nb_pkts = (uint16_t)RTE_MIN(nb_pkts, rxq->rx_nb_avail);
 511
 512         for (i = 0; i < nb_pkts; i++)
 513                 rx_pkts[i] = stage[i];
 514
 515         rxq->rx_nb_avail = (uint16_t)(rxq->rx_nb_avail - nb_pkts);
 516         rxq->rx_next_avail = (uint16_t)(rxq->rx_next_avail + nb_pkts);
 517
 518         return nb_pkts;
 519 }
 520
 521 static inline int
 522 i40e_rx_alloc_bufs(struct i40e_rx_queue *rxq)
 523 {
 524         volatile union i40e_rx_desc *rxdp;
 525         struct i40e_rx_entry *rxep;
 526         struct rte_mbuf *mb;
 527         uint16_t alloc_idx, i;
 528         uint64_t dma_addr;
 529         int diag;
 530
 531         /* Allocate buffers in bulk */
 532         alloc_idx = (uint16_t)(rxq->rx_free_trigger -
 533                                 (rxq->rx_free_thresh - 1));
 534         rxep = &(rxq->sw_ring[alloc_idx]);
 535         diag = rte_mempool_get_bulk(rxq->mp, (void *)rxep,
 536                                         rxq->rx_free_thresh);
 537         if (unlikely(diag != 0)) {
 538                 PMD_DRV_LOG(ERR, "Failed to get mbufs in bulk");
 539                 return -ENOMEM;
 540         }
 541
 542         rxdp = &rxq->rx_ring[alloc_idx];
 543         for (i = 0; i < rxq->rx_free_thresh; i++) {
 544                 if (likely(i < (rxq->rx_free_thresh - 1)))
 545                         /* Prefetch next mbuf */
 546                         rte_prefetch0(rxep[i + 1].mbuf);
 547
 548                 mb = rxep[i].mbuf;
 549                 rte_mbuf_refcnt_set(mb, 1);
 550                 mb->next = NULL;
 551                 mb->data_off = RTE_PKTMBUF_HEADROOM;
 552                 mb->nb_segs = 1;
 553                 mb->port = rxq->port_id;
 554                 dma_addr = rte_cpu_to_le_64(\
 555                         rte_mbuf_data_iova_default(mb));
 556                 rxdp[i].read.hdr_addr = 0;
 557                 rxdp[i].read.pkt_addr = dma_addr;
 558         }
 559
 560         /* Update rx tail regsiter */
 561         I40E_PCI_REG_WRITE(rxq->qrx_tail, rxq->rx_free_trigger);
 562
 563         rxq->rx_free_trigger =
 564                 (uint16_t)(rxq->rx_free_trigger + rxq->rx_free_thresh);
 565         if (rxq->rx_free_trigger >= rxq->nb_rx_desc)
 566                 rxq->rx_free_trigger = (uint16_t)(rxq->rx_free_thresh - 1);
 567
 568         return 0;
 569 }
 570
 571 static inline uint16_t
 572 rx_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 573 {
 574         struct i40e_rx_queue *rxq = (struct i40e_rx_queue *)rx_queue;
 575         struct rte_eth_dev *dev;
 576         uint16_t nb_rx = 0;
 577
 578         if (!nb_pkts)
 579                 return 0;
 580
 581         if (rxq->rx_nb_avail)
 582                 return i40e_rx_fill_from_stage(rxq, rx_pkts, nb_pkts);
 583
 584         nb_rx = (uint16_t)i40e_rx_scan_hw_ring(rxq);
 585         rxq->rx_next_avail = 0;
 586         rxq->rx_nb_avail = nb_rx;
 587         rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_rx);
 588
 589         if (rxq->rx_tail > rxq->rx_free_trigger) {
 590                 if (i40e_rx_alloc_bufs(rxq) != 0) {
 591                         uint16_t i, j;
 592
 593                         dev = I40E_VSI_TO_ETH_DEV(rxq->vsi);
 594                         dev->data->rx_mbuf_alloc_failed +=
 595                                 rxq->rx_free_thresh;
 596
 597                         rxq->rx_nb_avail = 0;
 598                         rxq->rx_tail = (uint16_t)(rxq->rx_tail - nb_rx);
 599                         for (i = 0, j = rxq->rx_tail; i < nb_rx; i++, j++)
 600                                 rxq->sw_ring[j].mbuf = rxq->rx_stage[i];
 601
 602                         return 0;
 603                 }
 604         }
 605
 606         if (rxq->rx_tail >= rxq->nb_rx_desc)
 607                 rxq->rx_tail = 0;
 608
 609         if (rxq->rx_nb_avail)
 610                 return i40e_rx_fill_from_stage(rxq, rx_pkts, nb_pkts);
 611
 612         return 0;
 613 }
 614
 615 static uint16_t
 616 i40e_recv_pkts_bulk_alloc(void *rx_queue,
 617                           struct rte_mbuf **rx_pkts,
 618                           uint16_t nb_pkts)
 619 {
 620         uint16_t nb_rx = 0, n, count;
 621
 622         if (unlikely(nb_pkts == 0))
 623                 return 0;
 624
 625         if (likely(nb_pkts <= RTE_PMD_I40E_RX_MAX_BURST))
 626                 return rx_recv_pkts(rx_queue, rx_pkts, nb_pkts);
 627
 628         while (nb_pkts) {
 629                 n = RTE_MIN(nb_pkts, RTE_PMD_I40E_RX_MAX_BURST);
 630                 count = rx_recv_pkts(rx_queue, &rx_pkts[nb_rx], n);
 631                 nb_rx = (uint16_t)(nb_rx + count);
 632                 nb_pkts = (uint16_t)(nb_pkts - count);
 633                 if (count < n)
 634                         break;
 635         }
 636
 637         return nb_rx;
 638 }
 639 #else
 640 static uint16_t
 641 i40e_recv_pkts_bulk_alloc(void __rte_unused *rx_queue,
 642                           struct rte_mbuf __rte_unused **rx_pkts,
 643                           uint16_t __rte_unused nb_pkts)
 644 {
 645         return 0;
 646 }
 647 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
 648
 649 uint16_t
 650 i40e_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 651 {
 652         struct i40e_rx_queue *rxq;
 653         volatile union i40e_rx_desc *rx_ring;
 654         volatile union i40e_rx_desc *rxdp;
 655         union i40e_rx_desc rxd;
 656         struct i40e_rx_entry *sw_ring;
 657         struct i40e_rx_entry *rxe;
 658         struct rte_eth_dev *dev;
 659         struct rte_mbuf *rxm;
 660         struct rte_mbuf *nmb;
 661         uint16_t nb_rx;
 662         uint32_t rx_status;
 663         uint64_t qword1;
 664         uint16_t rx_packet_len;
 665         uint16_t rx_id, nb_hold;
 666         uint64_t dma_addr;
 667         uint64_t pkt_flags;
 668         uint32_t *ptype_tbl;
 669
 670         nb_rx = 0;
 671         nb_hold = 0;
 672         rxq = rx_queue;
 673         rx_id = rxq->rx_tail;
 674         rx_ring = rxq->rx_ring;
 675         sw_ring = rxq->sw_ring;
 676         ptype_tbl = rxq->vsi->adapter->ptype_tbl;
 677
 678         while (nb_rx < nb_pkts) {
 679                 rxdp = &rx_ring[rx_id];
 680                 qword1 = rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len);
 681                 rx_status = (qword1 & I40E_RXD_QW1_STATUS_MASK)
 682                                 >> I40E_RXD_QW1_STATUS_SHIFT;
 683
 684                 /* Check the DD bit first */
 685                 if (!(rx_status & (1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
 686                         break;
 687
 688                 nmb = rte_mbuf_raw_alloc(rxq->mp);
 689                 if (unlikely(!nmb)) {
 690                         dev = I40E_VSI_TO_ETH_DEV(rxq->vsi);
 691                         dev->data->rx_mbuf_alloc_failed++;
 692                         break;
 693                 }
 694
 695                 rxd = *rxdp;
 696                 nb_hold++;
 697                 rxe = &sw_ring[rx_id];
 698                 rx_id++;
 699                 if (unlikely(rx_id == rxq->nb_rx_desc))
 700                         rx_id = 0;
 701
 702                 /* Prefetch next mbuf */
 703                 rte_prefetch0(sw_ring[rx_id].mbuf);
 704
 705                 /**
 706                  * When next RX descriptor is on a cache line boundary,
 707                  * prefetch the next 4 RX descriptors and next 8 pointers
 708                  * to mbufs.
 709                  */
 710                 if ((rx_id & 0x3) == 0) {
 711                         rte_prefetch0(&rx_ring[rx_id]);
 712                         rte_prefetch0(&sw_ring[rx_id]);
 713                 }
 714                 rxm = rxe->mbuf;
 715                 rxe->mbuf = nmb;
 716                 dma_addr =
 717                         rte_cpu_to_le_64(rte_mbuf_data_iova_default(nmb));
 718                 rxdp->read.hdr_addr = 0;
 719                 rxdp->read.pkt_addr = dma_addr;
 720
 721                 rx_packet_len = ((qword1 & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
 722                                 I40E_RXD_QW1_LENGTH_PBUF_SHIFT) - rxq->crc_len;
 723
 724                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 725                 rte_prefetch0(RTE_PTR_ADD(rxm->buf_addr, RTE_PKTMBUF_HEADROOM));
 726                 rxm->nb_segs = 1;
 727                 rxm->next = NULL;
 728                 rxm->pkt_len = rx_packet_len;
 729                 rxm->data_len = rx_packet_len;
 730                 rxm->port = rxq->port_id;
 731                 rxm->ol_flags = 0;
 732                 i40e_rxd_to_vlan_tci(rxm, &rxd);
 733                 pkt_flags = i40e_rxd_status_to_pkt_flags(qword1);
 734                 pkt_flags |= i40e_rxd_error_to_pkt_flags(qword1);
 735                 rxm->packet_type =
 736                         ptype_tbl[(uint8_t)((qword1 &
 737                         I40E_RXD_QW1_PTYPE_MASK) >> I40E_RXD_QW1_PTYPE_SHIFT)];
 738                 if (pkt_flags & PKT_RX_RSS_HASH)
 739                         rxm->hash.rss =
 740                                 rte_le_to_cpu_32(rxd.wb.qword0.hi_dword.rss);
 741                 if (pkt_flags & PKT_RX_FDIR)
 742                         pkt_flags |= i40e_rxd_build_fdir(&rxd, rxm);
 743
 744 #ifdef RTE_LIBRTE_IEEE1588
 745                 pkt_flags |= i40e_get_iee15888_flags(rxm, qword1);
 746 #endif
 747                 rxm->ol_flags |= pkt_flags;
 748
 749                 rx_pkts[nb_rx++] = rxm;
 750         }
 751         rxq->rx_tail = rx_id;
 752
 753         /**
 754          * If the number of free RX descriptors is greater than the RX free
 755          * threshold of the queue, advance the receive tail register of queue.
 756          * Update that register with the value of the last processed RX
 757          * descriptor minus 1.
 758          */
 759         nb_hold = (uint16_t)(nb_hold + rxq->nb_rx_hold);
 760         if (nb_hold > rxq->rx_free_thresh) {
 761                 rx_id = (uint16_t) ((rx_id == 0) ?
 762                         (rxq->nb_rx_desc - 1) : (rx_id - 1));
 763                 I40E_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
 764                 nb_hold = 0;
 765         }
 766         rxq->nb_rx_hold = nb_hold;
 767
 768         return nb_rx;
 769 }
 770
 771 uint16_t
 772 i40e_recv_scattered_pkts(void *rx_queue,
 773                          struct rte_mbuf **rx_pkts,
 774                          uint16_t nb_pkts)
 775 {
 776         struct i40e_rx_queue *rxq = rx_queue;
 777         volatile union i40e_rx_desc *rx_ring = rxq->rx_ring;
 778         volatile union i40e_rx_desc *rxdp;
 779         union i40e_rx_desc rxd;
 780         struct i40e_rx_entry *sw_ring = rxq->sw_ring;
 781         struct i40e_rx_entry *rxe;
 782         struct rte_mbuf *first_seg = rxq->pkt_first_seg;
 783         struct rte_mbuf *last_seg = rxq->pkt_last_seg;
 784         struct rte_mbuf *nmb, *rxm;
 785         uint16_t rx_id = rxq->rx_tail;
 786         uint16_t nb_rx = 0, nb_hold = 0, rx_packet_len;
 787         struct rte_eth_dev *dev;
 788         uint32_t rx_status;
 789         uint64_t qword1;
 790         uint64_t dma_addr;
 791         uint64_t pkt_flags;
 792         uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
 793
 794         while (nb_rx < nb_pkts) {
 795                 rxdp = &rx_ring[rx_id];
 796                 qword1 = rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len);
 797                 rx_status = (qword1 & I40E_RXD_QW1_STATUS_MASK) >>
 798                                         I40E_RXD_QW1_STATUS_SHIFT;
 799
 800                 /* Check the DD bit */
 801                 if (!(rx_status & (1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
 802                         break;
 803
 804                 nmb = rte_mbuf_raw_alloc(rxq->mp);
 805                 if (unlikely(!nmb)) {
 806                         dev = I40E_VSI_TO_ETH_DEV(rxq->vsi);
 807                         dev->data->rx_mbuf_alloc_failed++;
 808                         break;
 809                 }
 810
 811                 rxd = *rxdp;
 812                 nb_hold++;
 813                 rxe = &sw_ring[rx_id];
 814                 rx_id++;
 815                 if (rx_id == rxq->nb_rx_desc)
 816                         rx_id = 0;
 817
 818                 /* Prefetch next mbuf */
 819                 rte_prefetch0(sw_ring[rx_id].mbuf);
 820
 821                 /**
 822                  * When next RX descriptor is on a cache line boundary,
 823                  * prefetch the next 4 RX descriptors and next 8 pointers
 824                  * to mbufs.
 825                  */
 826                 if ((rx_id & 0x3) == 0) {
 827                         rte_prefetch0(&rx_ring[rx_id]);
 828                         rte_prefetch0(&sw_ring[rx_id]);
 829                 }
 830
 831                 rxm = rxe->mbuf;
 832                 rxe->mbuf = nmb;
 833                 dma_addr =
 834                         rte_cpu_to_le_64(rte_mbuf_data_iova_default(nmb));
 835
 836                 /* Set data buffer address and data length of the mbuf */
 837                 rxdp->read.hdr_addr = 0;
 838                 rxdp->read.pkt_addr = dma_addr;
 839                 rx_packet_len = (qword1 & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
 840                                         I40E_RXD_QW1_LENGTH_PBUF_SHIFT;
 841                 rxm->data_len = rx_packet_len;
 842                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 843
 844                 /**
 845                  * If this is the first buffer of the received packet, set the
 846                  * pointer to the first mbuf of the packet and initialize its
 847                  * context. Otherwise, update the total length and the number
 848                  * of segments of the current scattered packet, and update the
 849                  * pointer to the last mbuf of the current packet.
 850                  */
 851                 if (!first_seg) {
 852                         first_seg = rxm;
 853                         first_seg->nb_segs = 1;
 854                         first_seg->pkt_len = rx_packet_len;
 855                 } else {
 856                         first_seg->pkt_len =
 857                                 (uint16_t)(first_seg->pkt_len +
 858                                                 rx_packet_len);
 859                         first_seg->nb_segs++;
 860                         last_seg->next = rxm;
 861                 }
 862
 863                 /**
 864                  * If this is not the last buffer of the received packet,
 865                  * update the pointer to the last mbuf of the current scattered
 866                  * packet and continue to parse the RX ring.
 867                  */
 868                 if (!(rx_status & (1 << I40E_RX_DESC_STATUS_EOF_SHIFT))) {
 869                         last_seg = rxm;
 870                         continue;
 871                 }
 872
 873                 /**
 874                  * This is the last buffer of the received packet. If the CRC
 875                  * is not stripped by the hardware:
 876                  *  - Subtract the CRC length from the total packet length.
 877                  *  - If the last buffer only contains the whole CRC or a part
 878                  *  of it, free the mbuf associated to the last buffer. If part
 879                  *  of the CRC is also contained in the previous mbuf, subtract
 880                  *  the length of that CRC part from the data length of the
 881                  *  previous mbuf.
 882                  */
 883                 rxm->next = NULL;
 884                 if (unlikely(rxq->crc_len > 0)) {
 885                         first_seg->pkt_len -= RTE_ETHER_CRC_LEN;
 886                         if (rx_packet_len <= RTE_ETHER_CRC_LEN) {
 887                                 rte_pktmbuf_free_seg(rxm);
 888                                 first_seg->nb_segs--;
 889                                 last_seg->data_len =
 890                                         (uint16_t)(last_seg->data_len -
 891                                         (RTE_ETHER_CRC_LEN - rx_packet_len));
 892                                 last_seg->next = NULL;
 893                         } else
 894                                 rxm->data_len = (uint16_t)(rx_packet_len -
 895                                                         RTE_ETHER_CRC_LEN);
 896                 }
 897
 898                 first_seg->port = rxq->port_id;
 899                 first_seg->ol_flags = 0;
 900                 i40e_rxd_to_vlan_tci(first_seg, &rxd);
 901                 pkt_flags = i40e_rxd_status_to_pkt_flags(qword1);
 902                 pkt_flags |= i40e_rxd_error_to_pkt_flags(qword1);
 903                 first_seg->packet_type =
 904                         ptype_tbl[(uint8_t)((qword1 &
 905                         I40E_RXD_QW1_PTYPE_MASK) >> I40E_RXD_QW1_PTYPE_SHIFT)];
 906                 if (pkt_flags & PKT_RX_RSS_HASH)
 907                         first_seg->hash.rss =
 908                                 rte_le_to_cpu_32(rxd.wb.qword0.hi_dword.rss);
 909                 if (pkt_flags & PKT_RX_FDIR)
 910                         pkt_flags |= i40e_rxd_build_fdir(&rxd, first_seg);
 911
 912 #ifdef RTE_LIBRTE_IEEE1588
 913                 pkt_flags |= i40e_get_iee15888_flags(first_seg, qword1);
 914 #endif
 915                 first_seg->ol_flags |= pkt_flags;
 916
 917                 /* Prefetch data of first segment, if configured to do so. */
 918                 rte_prefetch0(RTE_PTR_ADD(first_seg->buf_addr,
 919                         first_seg->data_off));
 920                 rx_pkts[nb_rx++] = first_seg;
 921                 first_seg = NULL;
 922         }
 923
 924         /* Record index of the next RX descriptor to probe. */
 925         rxq->rx_tail = rx_id;
 926         rxq->pkt_first_seg = first_seg;
 927         rxq->pkt_last_seg = last_seg;
 928
 929         /**
 930          * If the number of free RX descriptors is greater than the RX free
 931          * threshold of the queue, advance the Receive Descriptor Tail (RDT)
 932          * register. Update the RDT with the value of the last processed RX
 933          * descriptor minus 1, to guarantee that the RDT register is never
 934          * equal to the RDH register, which creates a "full" ring situtation
 935          * from the hardware point of view.
 936          */
 937         nb_hold = (uint16_t)(nb_hold + rxq->nb_rx_hold);
 938         if (nb_hold > rxq->rx_free_thresh) {
 939                 rx_id = (uint16_t)(rx_id == 0 ?
 940                         (rxq->nb_rx_desc - 1) : (rx_id - 1));
 941                 I40E_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
 942                 nb_hold = 0;
 943         }
 944         rxq->nb_rx_hold = nb_hold;
 945
 946         return nb_rx;
 947 }
 948
 949 /* Check if the context descriptor is needed for TX offloading */
 950 static inline uint16_t
 951 i40e_calc_context_desc(uint64_t flags)
 952 {
 953         static uint64_t mask = PKT_TX_OUTER_IP_CKSUM |
 954                 PKT_TX_TCP_SEG |
 955                 PKT_TX_QINQ_PKT |
 956                 PKT_TX_TUNNEL_MASK;
 957
 958 #ifdef RTE_LIBRTE_IEEE1588
 959         mask |= PKT_TX_IEEE1588_TMST;
 960 #endif
 961
 962         return (flags & mask) ? 1 : 0;
 963 }
 964
 965 /* set i40e TSO context descriptor */
 966 static inline uint64_t
 967 i40e_set_tso_ctx(struct rte_mbuf *mbuf, union i40e_tx_offload tx_offload)
 968 {
 969         uint64_t ctx_desc = 0;
 970         uint32_t cd_cmd, hdr_len, cd_tso_len;
 971
 972         if (!tx_offload.l4_len) {
 973                 PMD_DRV_LOG(DEBUG, "L4 length set to 0");
 974                 return ctx_desc;
 975         }
 976
 977         hdr_len = tx_offload.l2_len + tx_offload.l3_len + tx_offload.l4_len;
 978         hdr_len += (mbuf->ol_flags & PKT_TX_TUNNEL_MASK) ?
 979                    tx_offload.outer_l2_len + tx_offload.outer_l3_len : 0;
 980
 981         cd_cmd = I40E_TX_CTX_DESC_TSO;
 982         cd_tso_len = mbuf->pkt_len - hdr_len;
 983         ctx_desc |= ((uint64_t)cd_cmd << I40E_TXD_CTX_QW1_CMD_SHIFT) |
 984                 ((uint64_t)cd_tso_len <<
 985                  I40E_TXD_CTX_QW1_TSO_LEN_SHIFT) |
 986                 ((uint64_t)mbuf->tso_segsz <<
 987                  I40E_TXD_CTX_QW1_MSS_SHIFT);
 988
 989         return ctx_desc;
 990 }
 991
 992 /* HW requires that Tx buffer size ranges from 1B up to (16K-1)B. */
 993 #define I40E_MAX_DATA_PER_TXD \
 994         (I40E_TXD_QW1_TX_BUF_SZ_MASK >> I40E_TXD_QW1_TX_BUF_SZ_SHIFT)
 995 /* Calculate the number of TX descriptors needed for each pkt */
 996 static inline uint16_t
 997 i40e_calc_pkt_desc(struct rte_mbuf *tx_pkt)
 998 {
 999         struct rte_mbuf *txd = tx_pkt;
1000         uint16_t count = 0;
1001
1002         while (txd != NULL) {
1003                 count += DIV_ROUND_UP(txd->data_len, I40E_MAX_DATA_PER_TXD);
1004                 txd = txd->next;
1005         }
1006
1007         return count;
1008 }
1009
1010 uint16_t
1011 i40e_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1012 {
1013         struct i40e_tx_queue *txq;
1014         struct i40e_tx_entry *sw_ring;
1015         struct i40e_tx_entry *txe, *txn;
1016         volatile struct i40e_tx_desc *txd;
1017         volatile struct i40e_tx_desc *txr;
1018         struct rte_mbuf *tx_pkt;
1019         struct rte_mbuf *m_seg;
1020         uint32_t cd_tunneling_params;
1021         uint16_t tx_id;
1022         uint16_t nb_tx;
1023         uint32_t td_cmd;
1024         uint32_t td_offset;
1025         uint32_t td_tag;
1026         uint64_t ol_flags;
1027         uint16_t nb_used;
1028         uint16_t nb_ctx;
1029         uint16_t tx_last;
1030         uint16_t slen;
1031         uint64_t buf_dma_addr;
1032         union i40e_tx_offload tx_offload = {0};
1033
1034         txq = tx_queue;
1035         sw_ring = txq->sw_ring;
1036         txr = txq->tx_ring;
1037         tx_id = txq->tx_tail;
1038         txe = &sw_ring[tx_id];
1039
1040         /* Check if the descriptor ring needs to be cleaned. */
1041         if (txq->nb_tx_free < txq->tx_free_thresh)
1042                 (void)i40e_xmit_cleanup(txq);
1043
1044         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1045                 td_cmd = 0;
1046                 td_tag = 0;
1047                 td_offset = 0;
1048
1049                 tx_pkt = *tx_pkts++;
1050                 RTE_MBUF_PREFETCH_TO_FREE(txe->mbuf);
1051
1052                 ol_flags = tx_pkt->ol_flags;
1053                 tx_offload.l2_len = tx_pkt->l2_len;
1054                 tx_offload.l3_len = tx_pkt->l3_len;
1055                 tx_offload.outer_l2_len = tx_pkt->outer_l2_len;
1056                 tx_offload.outer_l3_len = tx_pkt->outer_l3_len;
1057                 tx_offload.l4_len = tx_pkt->l4_len;
1058                 tx_offload.tso_segsz = tx_pkt->tso_segsz;
1059
1060                 /* Calculate the number of context descriptors needed. */
1061                 nb_ctx = i40e_calc_context_desc(ol_flags);
1062
1063                 /**
1064                  * The number of descriptors that must be allocated for
1065                  * a packet equals to the number of the segments of that
1066                  * packet plus 1 context descriptor if needed.
1067                  * Recalculate the needed tx descs when TSO enabled in case
1068                  * the mbuf data size exceeds max data size that hw allows
1069                  * per tx desc.
1070                  */
1071                 if (ol_flags & PKT_TX_TCP_SEG)
1072                         nb_used = (uint16_t)(i40e_calc_pkt_desc(tx_pkt) +
1073                                              nb_ctx);
1074                 else
1075                         nb_used = (uint16_t)(tx_pkt->nb_segs + nb_ctx);
1076                 tx_last = (uint16_t)(tx_id + nb_used - 1);
1077
1078                 /* Circular ring */
1079                 if (tx_last >= txq->nb_tx_desc)
1080                         tx_last = (uint16_t)(tx_last - txq->nb_tx_desc);
1081
1082                 if (nb_used > txq->nb_tx_free) {
1083                         if (i40e_xmit_cleanup(txq) != 0) {
1084                                 if (nb_tx == 0)
1085                                         return 0;
1086                                 goto end_of_tx;
1087                         }
1088                         if (unlikely(nb_used > txq->tx_rs_thresh)) {
1089                                 while (nb_used > txq->nb_tx_free) {
1090                                         if (i40e_xmit_cleanup(txq) != 0) {
1091                                                 if (nb_tx == 0)
1092                                                         return 0;
1093                                                 goto end_of_tx;
1094                                         }
1095                                 }
1096                         }
1097                 }
1098
1099                 /* Descriptor based VLAN insertion */
1100                 if (ol_flags & (PKT_TX_VLAN_PKT | PKT_TX_QINQ_PKT)) {
1101                         td_cmd |= I40E_TX_DESC_CMD_IL2TAG1;
1102                         td_tag = tx_pkt->vlan_tci;
1103                 }
1104
1105                 /* Always enable CRC offload insertion */
1106                 td_cmd |= I40E_TX_DESC_CMD_ICRC;
1107
1108                 /* Fill in tunneling parameters if necessary */
1109                 cd_tunneling_params = 0;
1110                 if (ol_flags & PKT_TX_TUNNEL_MASK)
1111                         i40e_parse_tunneling_params(ol_flags, tx_offload,
1112                                                     &cd_tunneling_params);
1113                 /* Enable checksum offloading */
1114                 if (ol_flags & I40E_TX_CKSUM_OFFLOAD_MASK)
1115                         i40e_txd_enable_checksum(ol_flags, &td_cmd,
1116                                                  &td_offset, tx_offload);
1117
1118                 if (nb_ctx) {
1119                         /* Setup TX context descriptor if required */
1120                         volatile struct i40e_tx_context_desc *ctx_txd =
1121                                 (volatile struct i40e_tx_context_desc *)\
1122                                                         &txr[tx_id];
1123                         uint16_t cd_l2tag2 = 0;
1124                         uint64_t cd_type_cmd_tso_mss =
1125                                 I40E_TX_DESC_DTYPE_CONTEXT;
1126
1127                         txn = &sw_ring[txe->next_id];
1128                         RTE_MBUF_PREFETCH_TO_FREE(txn->mbuf);
1129                         if (txe->mbuf != NULL) {
1130                                 rte_pktmbuf_free_seg(txe->mbuf);
1131                                 txe->mbuf = NULL;
1132                         }
1133
1134                         /* TSO enabled means no timestamp */
1135                         if (ol_flags & PKT_TX_TCP_SEG)
1136                                 cd_type_cmd_tso_mss |=
1137                                         i40e_set_tso_ctx(tx_pkt, tx_offload);
1138                         else {
1139 #ifdef RTE_LIBRTE_IEEE1588
1140                                 if (ol_flags & PKT_TX_IEEE1588_TMST)
1141                                         cd_type_cmd_tso_mss |=
1142                                                 ((uint64_t)I40E_TX_CTX_DESC_TSYN <<
1143                                                  I40E_TXD_CTX_QW1_CMD_SHIFT);
1144 #endif
1145                         }
1146
1147                         ctx_txd->tunneling_params =
1148                                 rte_cpu_to_le_32(cd_tunneling_params);
1149                         if (ol_flags & PKT_TX_QINQ_PKT) {
1150                                 cd_l2tag2 = tx_pkt->vlan_tci_outer;
1151                                 cd_type_cmd_tso_mss |=
1152                                         ((uint64_t)I40E_TX_CTX_DESC_IL2TAG2 <<
1153                                                 I40E_TXD_CTX_QW1_CMD_SHIFT);
1154                         }
1155                         ctx_txd->l2tag2 = rte_cpu_to_le_16(cd_l2tag2);
1156                         ctx_txd->type_cmd_tso_mss =
1157                                 rte_cpu_to_le_64(cd_type_cmd_tso_mss);
1158
1159                         PMD_TX_LOG(DEBUG, "mbuf: %p, TCD[%u]:\n"
1160                                 "tunneling_params: %#x;\n"
1161                                 "l2tag2: %#hx;\n"
1162                                 "rsvd: %#hx;\n"
1163                                 "type_cmd_tso_mss: %#"PRIx64";\n",
1164                                 tx_pkt, tx_id,
1165                                 ctx_txd->tunneling_params,
1166                                 ctx_txd->l2tag2,
1167                                 ctx_txd->rsvd,
1168                                 ctx_txd->type_cmd_tso_mss);
1169
1170                         txe->last_id = tx_last;
1171                         tx_id = txe->next_id;
1172                         txe = txn;
1173                 }
1174
1175                 m_seg = tx_pkt;
1176                 do {
1177                         txd = &txr[tx_id];
1178                         txn = &sw_ring[txe->next_id];
1179
1180                         if (txe->mbuf)
1181                                 rte_pktmbuf_free_seg(txe->mbuf);
1182                         txe->mbuf = m_seg;
1183
1184                         /* Setup TX Descriptor */
1185                         slen = m_seg->data_len;
1186                         buf_dma_addr = rte_mbuf_data_iova(m_seg);
1187
1188                         while ((ol_flags & PKT_TX_TCP_SEG) &&
1189                                 unlikely(slen > I40E_MAX_DATA_PER_TXD)) {
1190                                 txd->buffer_addr =
1191                                         rte_cpu_to_le_64(buf_dma_addr);
1192                                 txd->cmd_type_offset_bsz =
1193                                         i40e_build_ctob(td_cmd,
1194                                         td_offset, I40E_MAX_DATA_PER_TXD,
1195                                         td_tag);
1196
1197                                 buf_dma_addr += I40E_MAX_DATA_PER_TXD;
1198                                 slen -= I40E_MAX_DATA_PER_TXD;
1199
1200                                 txe->last_id = tx_last;
1201                                 tx_id = txe->next_id;
1202                                 txe = txn;
1203                                 txd = &txr[tx_id];
1204                                 txn = &sw_ring[txe->next_id];
1205                         }
1206                         PMD_TX_LOG(DEBUG, "mbuf: %p, TDD[%u]:\n"
1207                                 "buf_dma_addr: %#"PRIx64";\n"
1208                                 "td_cmd: %#x;\n"
1209                                 "td_offset: %#x;\n"
1210                                 "td_len: %u;\n"
1211                                 "td_tag: %#x;\n",
1212                                 tx_pkt, tx_id, buf_dma_addr,
1213                                 td_cmd, td_offset, slen, td_tag);
1214
1215                         txd->buffer_addr = rte_cpu_to_le_64(buf_dma_addr);
1216                         txd->cmd_type_offset_bsz = i40e_build_ctob(td_cmd,
1217                                                 td_offset, slen, td_tag);
1218                         txe->last_id = tx_last;
1219                         tx_id = txe->next_id;
1220                         txe = txn;
1221                         m_seg = m_seg->next;
1222                 } while (m_seg != NULL);
1223
1224                 /* The last packet data descriptor needs End Of Packet (EOP) */
1225                 td_cmd |= I40E_TX_DESC_CMD_EOP;
1226                 txq->nb_tx_used = (uint16_t)(txq->nb_tx_used + nb_used);
1227                 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_used);
1228
1229                 if (txq->nb_tx_used >= txq->tx_rs_thresh) {
1230                         PMD_TX_FREE_LOG(DEBUG,
1231                                         "Setting RS bit on TXD id="
1232                                         "%4u (port=%d queue=%d)",
1233                                         tx_last, txq->port_id, txq->queue_id);
1234
1235                         td_cmd |= I40E_TX_DESC_CMD_RS;
1236
1237                         /* Update txq RS bit counters */
1238                         txq->nb_tx_used = 0;
1239                 }
1240
1241                 txd->cmd_type_offset_bsz |=
1242                         rte_cpu_to_le_64(((uint64_t)td_cmd) <<
1243                                         I40E_TXD_QW1_CMD_SHIFT);
1244         }
1245
1246 end_of_tx:
1247         PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u tx_tail=%u nb_tx=%u",
1248                    (unsigned) txq->port_id, (unsigned) txq->queue_id,
1249                    (unsigned) tx_id, (unsigned) nb_tx);
1250
1251         rte_cio_wmb();
1252         I40E_PCI_REG_WRITE_RELAXED(txq->qtx_tail, tx_id);
1253         txq->tx_tail = tx_id;
1254
1255         return nb_tx;
1256 }
1257
1258 static __rte_always_inline int
1259 i40e_tx_free_bufs(struct i40e_tx_queue *txq)
1260 {
1261         struct i40e_tx_entry *txep;
1262         uint16_t i;
1263
1264         if ((txq->tx_ring[txq->tx_next_dd].cmd_type_offset_bsz &
1265                         rte_cpu_to_le_64(I40E_TXD_QW1_DTYPE_MASK)) !=
1266                         rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE))
1267                 return 0;
1268
1269         txep = &(txq->sw_ring[txq->tx_next_dd - (txq->tx_rs_thresh - 1)]);
1270
1271         for (i = 0; i < txq->tx_rs_thresh; i++)
1272                 rte_prefetch0((txep + i)->mbuf);
1273
1274         if (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE) {
1275                 for (i = 0; i < txq->tx_rs_thresh; ++i, ++txep) {
1276                         rte_mempool_put(txep->mbuf->pool, txep->mbuf);
1277                         txep->mbuf = NULL;
1278                 }
1279         } else {
1280                 for (i = 0; i < txq->tx_rs_thresh; ++i, ++txep) {
1281                         rte_pktmbuf_free_seg(txep->mbuf);
1282                         txep->mbuf = NULL;
1283                 }
1284         }
1285
1286         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + txq->tx_rs_thresh);
1287         txq->tx_next_dd = (uint16_t)(txq->tx_next_dd + txq->tx_rs_thresh);
1288         if (txq->tx_next_dd >= txq->nb_tx_desc)
1289                 txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
1290
1291         return txq->tx_rs_thresh;
1292 }
1293
1294 /* Populate 4 descriptors with data from 4 mbufs */
1295 static inline void
1296 tx4(volatile struct i40e_tx_desc *txdp, struct rte_mbuf **pkts)
1297 {
1298         uint64_t dma_addr;
1299         uint32_t i;
1300
1301         for (i = 0; i < 4; i++, txdp++, pkts++) {
1302                 dma_addr = rte_mbuf_data_iova(*pkts);
1303                 txdp->buffer_addr = rte_cpu_to_le_64(dma_addr);
1304                 txdp->cmd_type_offset_bsz =
1305                         i40e_build_ctob((uint32_t)I40E_TD_CMD, 0,
1306                                         (*pkts)->data_len, 0);
1307         }
1308 }
1309
1310 /* Populate 1 descriptor with data from 1 mbuf */
1311 static inline void
1312 tx1(volatile struct i40e_tx_desc *txdp, struct rte_mbuf **pkts)
1313 {
1314         uint64_t dma_addr;
1315
1316         dma_addr = rte_mbuf_data_iova(*pkts);
1317         txdp->buffer_addr = rte_cpu_to_le_64(dma_addr);
1318         txdp->cmd_type_offset_bsz =
1319                 i40e_build_ctob((uint32_t)I40E_TD_CMD, 0,
1320                                 (*pkts)->data_len, 0);
1321 }
1322
1323 /* Fill hardware descriptor ring with mbuf data */
1324 static inline void
1325 i40e_tx_fill_hw_ring(struct i40e_tx_queue *txq,
1326                      struct rte_mbuf **pkts,
1327                      uint16_t nb_pkts)
1328 {
1329         volatile struct i40e_tx_desc *txdp = &(txq->tx_ring[txq->tx_tail]);
1330         struct i40e_tx_entry *txep = &(txq->sw_ring[txq->tx_tail]);
1331         const int N_PER_LOOP = 4;
1332         const int N_PER_LOOP_MASK = N_PER_LOOP - 1;
1333         int mainpart, leftover;
1334         int i, j;
1335
1336         mainpart = (nb_pkts & ((uint32_t) ~N_PER_LOOP_MASK));
1337         leftover = (nb_pkts & ((uint32_t)  N_PER_LOOP_MASK));
1338         for (i = 0; i < mainpart; i += N_PER_LOOP) {
1339                 for (j = 0; j < N_PER_LOOP; ++j) {
1340                         (txep + i + j)->mbuf = *(pkts + i + j);
1341                 }
1342                 tx4(txdp + i, pkts + i);
1343         }
1344         if (unlikely(leftover > 0)) {
1345                 for (i = 0; i < leftover; ++i) {
1346                         (txep + mainpart + i)->mbuf = *(pkts + mainpart + i);
1347                         tx1(txdp + mainpart + i, pkts + mainpart + i);
1348                 }
1349         }
1350 }
1351
1352 static inline uint16_t
1353 tx_xmit_pkts(struct i40e_tx_queue *txq,
1354              struct rte_mbuf **tx_pkts,
1355              uint16_t nb_pkts)
1356 {
1357         volatile struct i40e_tx_desc *txr = txq->tx_ring;
1358         uint16_t n = 0;
1359
1360         /**
1361          * Begin scanning the H/W ring for done descriptors when the number
1362          * of available descriptors drops below tx_free_thresh. For each done
1363          * descriptor, free the associated buffer.
1364          */
1365         if (txq->nb_tx_free < txq->tx_free_thresh)
1366                 i40e_tx_free_bufs(txq);
1367
1368         /* Use available descriptor only */
1369         nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
1370         if (unlikely(!nb_pkts))
1371                 return 0;
1372
1373         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
1374         if ((txq->tx_tail + nb_pkts) > txq->nb_tx_desc) {
1375                 n = (uint16_t)(txq->nb_tx_desc - txq->tx_tail);
1376                 i40e_tx_fill_hw_ring(txq, tx_pkts, n);
1377                 txr[txq->tx_next_rs].cmd_type_offset_bsz |=
1378                         rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
1379                                                 I40E_TXD_QW1_CMD_SHIFT);
1380                 txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
1381                 txq->tx_tail = 0;
1382         }
1383
1384         /* Fill hardware descriptor ring with mbuf data */
1385         i40e_tx_fill_hw_ring(txq, tx_pkts + n, (uint16_t)(nb_pkts - n));
1386         txq->tx_tail = (uint16_t)(txq->tx_tail + (nb_pkts - n));
1387
1388         /* Determin if RS bit needs to be set */
1389         if (txq->tx_tail > txq->tx_next_rs) {
1390                 txr[txq->tx_next_rs].cmd_type_offset_bsz |=
1391                         rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
1392                                                 I40E_TXD_QW1_CMD_SHIFT);
1393                 txq->tx_next_rs =
1394                         (uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
1395                 if (txq->tx_next_rs >= txq->nb_tx_desc)
1396                         txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
1397         }
1398
1399         if (txq->tx_tail >= txq->nb_tx_desc)
1400                 txq->tx_tail = 0;
1401
1402         /* Update the tx tail register */
1403         I40E_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
1404
1405         return nb_pkts;
1406 }
1407
1408 static uint16_t
1409 i40e_xmit_pkts_simple(void *tx_queue,
1410                       struct rte_mbuf **tx_pkts,
1411                       uint16_t nb_pkts)
1412 {
1413         uint16_t nb_tx = 0;
1414
1415         if (likely(nb_pkts <= I40E_TX_MAX_BURST))
1416                 return tx_xmit_pkts((struct i40e_tx_queue *)tx_queue,
1417                                                 tx_pkts, nb_pkts);
1418
1419         while (nb_pkts) {
1420                 uint16_t ret, num = (uint16_t)RTE_MIN(nb_pkts,
1421                                                 I40E_TX_MAX_BURST);
1422
1423                 ret = tx_xmit_pkts((struct i40e_tx_queue *)tx_queue,
1424                                                 &tx_pkts[nb_tx], num);
1425                 nb_tx = (uint16_t)(nb_tx + ret);
1426                 nb_pkts = (uint16_t)(nb_pkts - ret);
1427                 if (ret < num)
1428                         break;
1429         }
1430
1431         return nb_tx;
1432 }
1433
1434 static uint16_t
1435 i40e_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
1436                    uint16_t nb_pkts)
1437 {
1438         uint16_t nb_tx = 0;
1439         struct i40e_tx_queue *txq = (struct i40e_tx_queue *)tx_queue;
1440
1441         while (nb_pkts) {
1442                 uint16_t ret, num;
1443
1444                 num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh);
1445                 ret = i40e_xmit_fixed_burst_vec(tx_queue, &tx_pkts[nb_tx],
1446                                                 num);
1447                 nb_tx += ret;
1448                 nb_pkts -= ret;
1449                 if (ret < num)
1450                         break;
1451         }
1452
1453         return nb_tx;
1454 }
1455
1456 /*********************************************************************
1457  *
1458  *  TX prep functions
1459  *
1460  **********************************************************************/
1461 uint16_t
1462 i40e_prep_pkts(__rte_unused void *tx_queue, struct rte_mbuf **tx_pkts,
1463                 uint16_t nb_pkts)
1464 {
1465         int i, ret;
1466         uint64_t ol_flags;
1467         struct rte_mbuf *m;
1468
1469         for (i = 0; i < nb_pkts; i++) {
1470                 m = tx_pkts[i];
1471                 ol_flags = m->ol_flags;
1472
1473                 /* Check for m->nb_segs to not exceed the limits. */
1474                 if (!(ol_flags & PKT_TX_TCP_SEG)) {
1475                         if (m->nb_segs > I40E_TX_MAX_MTU_SEG ||
1476                             m->pkt_len > I40E_FRAME_SIZE_MAX) {
1477                                 rte_errno = EINVAL;
1478                                 return i;
1479                         }
1480                 } else if (m->nb_segs > I40E_TX_MAX_SEG ||
1481                            m->tso_segsz < I40E_MIN_TSO_MSS ||
1482                            m->tso_segsz > I40E_MAX_TSO_MSS ||
1483                            m->pkt_len > I40E_TSO_FRAME_SIZE_MAX) {
1484                         /* MSS outside the range (256B - 9674B) are considered
1485                          * malicious
1486                          */
1487                         rte_errno = EINVAL;
1488                         return i;
1489                 }
1490
1491                 if (ol_flags & I40E_TX_OFFLOAD_NOTSUP_MASK) {
1492                         rte_errno = ENOTSUP;
1493                         return i;
1494                 }
1495
1496                 /* check the size of packet */
1497                 if (m->pkt_len < I40E_TX_MIN_PKT_LEN) {
1498                         rte_errno = EINVAL;
1499                         return i;
1500                 }
1501
1502 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
1503                 ret = rte_validate_tx_offload(m);
1504                 if (ret != 0) {
1505                         rte_errno = -ret;
1506                         return i;
1507                 }
1508 #endif
1509                 ret = rte_net_intel_cksum_prepare(m);
1510                 if (ret != 0) {
1511                         rte_errno = -ret;
1512                         return i;
1513                 }
1514         }
1515         return i;
1516 }
1517
1518 /*
1519  * Find the VSI the queue belongs to. 'queue_idx' is the queue index
1520  * application used, which assume having sequential ones. But from driver's
1521  * perspective, it's different. For example, q0 belongs to FDIR VSI, q1-q64
1522  * to MAIN VSI, , q65-96 to SRIOV VSIs, q97-128 to VMDQ VSIs. For application
1523  * running on host, q1-64 and q97-128 can be used, total 96 queues. They can
1524  * use queue_idx from 0 to 95 to access queues, while real queue would be
1525  * different. This function will do a queue mapping to find VSI the queue
1526  * belongs to.
1527  */
1528 static struct i40e_vsi*
1529 i40e_pf_get_vsi_by_qindex(struct i40e_pf *pf, uint16_t queue_idx)
1530 {
1531         /* the queue in MAIN VSI range */
1532         if (queue_idx < pf->main_vsi->nb_qps)
1533                 return pf->main_vsi;
1534
1535         queue_idx -= pf->main_vsi->nb_qps;
1536
1537         /* queue_idx is greater than VMDQ VSIs range */
1538         if (queue_idx > pf->nb_cfg_vmdq_vsi * pf->vmdq_nb_qps - 1) {
1539                 PMD_INIT_LOG(ERR, "queue_idx out of range. VMDQ configured?");
1540                 return NULL;
1541         }
1542
1543         return pf->vmdq[queue_idx / pf->vmdq_nb_qps].vsi;
1544 }
1545
1546 static uint16_t
1547 i40e_get_queue_offset_by_qindex(struct i40e_pf *pf, uint16_t queue_idx)
1548 {
1549         /* the queue in MAIN VSI range */
1550         if (queue_idx < pf->main_vsi->nb_qps)
1551                 return queue_idx;
1552
1553         /* It's VMDQ queues */
1554         queue_idx -= pf->main_vsi->nb_qps;
1555
1556         if (pf->nb_cfg_vmdq_vsi)
1557                 return queue_idx % pf->vmdq_nb_qps;
1558         else {
1559                 PMD_INIT_LOG(ERR, "Fail to get queue offset");
1560                 return (uint16_t)(-1);
1561         }
1562 }
1563
1564 int
1565 i40e_dev_rx_queue_start(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1566 {
1567         struct i40e_rx_queue *rxq;
1568         int err;
1569         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1570
1571         PMD_INIT_FUNC_TRACE();
1572
1573         rxq = dev->data->rx_queues[rx_queue_id];
1574
1575         err = i40e_alloc_rx_queue_mbufs(rxq);
1576         if (err) {
1577                 PMD_DRV_LOG(ERR, "Failed to allocate RX queue mbuf");
1578                 return err;
1579         }
1580
1581         /* Init the RX tail regieter. */
1582         I40E_PCI_REG_WRITE(rxq->qrx_tail, rxq->nb_rx_desc - 1);
1583
1584         err = i40e_switch_rx_queue(hw, rxq->reg_idx, TRUE);
1585         if (err) {
1586                 PMD_DRV_LOG(ERR, "Failed to switch RX queue %u on",
1587                             rx_queue_id);
1588
1589                 i40e_rx_queue_release_mbufs(rxq);
1590                 i40e_reset_rx_queue(rxq);
1591                 return err;
1592         }
1593         dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
1594
1595         return 0;
1596 }
1597
1598 int
1599 i40e_dev_rx_queue_stop(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1600 {
1601         struct i40e_rx_queue *rxq;
1602         int err;
1603         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1604
1605         rxq = dev->data->rx_queues[rx_queue_id];
1606
1607         /*
1608          * rx_queue_id is queue id application refers to, while
1609          * rxq->reg_idx is the real queue index.
1610          */
1611         err = i40e_switch_rx_queue(hw, rxq->reg_idx, FALSE);
1612         if (err) {
1613                 PMD_DRV_LOG(ERR, "Failed to switch RX queue %u off",
1614                             rx_queue_id);
1615                 return err;
1616         }
1617         i40e_rx_queue_release_mbufs(rxq);
1618         i40e_reset_rx_queue(rxq);
1619         dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
1620
1621         return 0;
1622 }
1623
1624 int
1625 i40e_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id)
1626 {
1627         int err;
1628         struct i40e_tx_queue *txq;
1629         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1630
1631         PMD_INIT_FUNC_TRACE();
1632
1633         txq = dev->data->tx_queues[tx_queue_id];
1634
1635         /*
1636          * tx_queue_id is queue id application refers to, while
1637          * rxq->reg_idx is the real queue index.
1638          */
1639         err = i40e_switch_tx_queue(hw, txq->reg_idx, TRUE);
1640         if (err) {
1641                 PMD_DRV_LOG(ERR, "Failed to switch TX queue %u on",
1642                             tx_queue_id);
1643                 return err;
1644         }
1645         dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
1646
1647         return 0;
1648 }
1649
1650 int
1651 i40e_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id)
1652 {
1653         struct i40e_tx_queue *txq;
1654         int err;
1655         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1656
1657         txq = dev->data->tx_queues[tx_queue_id];
1658
1659         /*
1660          * tx_queue_id is queue id application refers to, while
1661          * txq->reg_idx is the real queue index.
1662          */
1663         err = i40e_switch_tx_queue(hw, txq->reg_idx, FALSE);
1664         if (err) {
1665                 PMD_DRV_LOG(ERR, "Failed to switch TX queue %u of",
1666                             tx_queue_id);
1667                 return err;
1668         }
1669
1670         i40e_tx_queue_release_mbufs(txq);
1671         i40e_reset_tx_queue(txq);
1672         dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
1673
1674         return 0;
1675 }
1676
1677 const uint32_t *
1678 i40e_dev_supported_ptypes_get(struct rte_eth_dev *dev)
1679 {
1680         static const uint32_t ptypes[] = {
1681                 /* refers to i40e_rxd_pkt_type_mapping() */
1682                 RTE_PTYPE_L2_ETHER,
1683                 RTE_PTYPE_L2_ETHER_TIMESYNC,
1684                 RTE_PTYPE_L2_ETHER_LLDP,
1685                 RTE_PTYPE_L2_ETHER_ARP,
1686                 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
1687                 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
1688                 RTE_PTYPE_L4_FRAG,
1689                 RTE_PTYPE_L4_ICMP,
1690                 RTE_PTYPE_L4_NONFRAG,
1691                 RTE_PTYPE_L4_SCTP,
1692                 RTE_PTYPE_L4_TCP,
1693                 RTE_PTYPE_L4_UDP,
1694                 RTE_PTYPE_TUNNEL_GRENAT,
1695                 RTE_PTYPE_TUNNEL_IP,
1696                 RTE_PTYPE_INNER_L2_ETHER,
1697                 RTE_PTYPE_INNER_L2_ETHER_VLAN,
1698                 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
1699                 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
1700                 RTE_PTYPE_INNER_L4_FRAG,
1701                 RTE_PTYPE_INNER_L4_ICMP,
1702                 RTE_PTYPE_INNER_L4_NONFRAG,
1703                 RTE_PTYPE_INNER_L4_SCTP,
1704                 RTE_PTYPE_INNER_L4_TCP,
1705                 RTE_PTYPE_INNER_L4_UDP,
1706                 RTE_PTYPE_UNKNOWN
1707         };
1708
1709         if (dev->rx_pkt_burst == i40e_recv_pkts ||
1710 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
1711             dev->rx_pkt_burst == i40e_recv_pkts_bulk_alloc ||
1712 #endif
1713             dev->rx_pkt_burst == i40e_recv_scattered_pkts ||
1714             dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec ||
1715             dev->rx_pkt_burst == i40e_recv_pkts_vec ||
1716             dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec_avx2 ||
1717             dev->rx_pkt_burst == i40e_recv_pkts_vec_avx2)
1718                 return ptypes;
1719         return NULL;
1720 }
1721
1722 static int
1723 i40e_dev_first_queue(uint16_t idx, void **queues, int num)
1724 {
1725         uint16_t i;
1726
1727         for (i = 0; i < num; i++) {
1728                 if (i != idx && queues[i])
1729                         return 0;
1730         }
1731
1732         return 1;
1733 }
1734
1735 static int
1736 i40e_dev_rx_queue_setup_runtime(struct rte_eth_dev *dev,
1737                                 struct i40e_rx_queue *rxq)
1738 {
1739         struct i40e_adapter *ad =
1740                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
1741         int use_def_burst_func =
1742                 check_rx_burst_bulk_alloc_preconditions(rxq);
1743         uint16_t buf_size =
1744                 (uint16_t)(rte_pktmbuf_data_room_size(rxq->mp) -
1745                            RTE_PKTMBUF_HEADROOM);
1746         int use_scattered_rx =
1747                 (rxq->max_pkt_len > buf_size);
1748
1749         if (i40e_rx_queue_init(rxq) != I40E_SUCCESS) {
1750                 PMD_DRV_LOG(ERR,
1751                             "Failed to do RX queue initialization");
1752                 return -EINVAL;
1753         }
1754
1755         if (i40e_dev_first_queue(rxq->queue_id,
1756                                  dev->data->rx_queues,
1757                                  dev->data->nb_rx_queues)) {
1758                 /**
1759                  * If it is the first queue to setup,
1760                  * set all flags to default and call
1761                  * i40e_set_rx_function.
1762                  */
1763                 ad->rx_bulk_alloc_allowed = true;
1764                 ad->rx_vec_allowed = true;
1765                 dev->data->scattered_rx = use_scattered_rx;
1766                 if (use_def_burst_func)
1767                         ad->rx_bulk_alloc_allowed = false;
1768                 i40e_set_rx_function(dev);
1769                 return 0;
1770         } else if (ad->rx_vec_allowed && !rte_is_power_of_2(rxq->nb_rx_desc)) {
1771                 PMD_DRV_LOG(ERR, "Vector mode is allowed, but descriptor"
1772                             " number %d of queue %d isn't power of 2",
1773                             rxq->nb_rx_desc, rxq->queue_id);
1774                 return -EINVAL;
1775         }
1776
1777         /* check bulk alloc conflict */
1778         if (ad->rx_bulk_alloc_allowed && use_def_burst_func) {
1779                 PMD_DRV_LOG(ERR, "Can't use default burst.");
1780                 return -EINVAL;
1781         }
1782         /* check scatterred conflict */
1783         if (!dev->data->scattered_rx && use_scattered_rx) {
1784                 PMD_DRV_LOG(ERR, "Scattered rx is required.");
1785                 return -EINVAL;
1786         }
1787         /* check vector conflict */
1788         if (ad->rx_vec_allowed && i40e_rxq_vec_setup(rxq)) {
1789                 PMD_DRV_LOG(ERR, "Failed vector rx setup.");
1790                 return -EINVAL;
1791         }
1792
1793         return 0;
1794 }
1795
1796 int
1797 i40e_dev_rx_queue_setup(struct rte_eth_dev *dev,
1798                         uint16_t queue_idx,
1799                         uint16_t nb_desc,
1800                         unsigned int socket_id,
1801                         const struct rte_eth_rxconf *rx_conf,
1802                         struct rte_mempool *mp)
1803 {
1804         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1805         struct i40e_adapter *ad =
1806                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
1807         struct i40e_vsi *vsi;
1808         struct i40e_pf *pf = NULL;
1809         struct i40e_vf *vf = NULL;
1810         struct i40e_rx_queue *rxq;
1811         const struct rte_memzone *rz;
1812         uint32_t ring_size;
1813         uint16_t len, i;
1814         uint16_t reg_idx, base, bsf, tc_mapping;
1815         int q_offset, use_def_burst_func = 1;
1816         uint64_t offloads;
1817
1818         offloads = rx_conf->offloads | dev->data->dev_conf.rxmode.offloads;
1819
1820         if (hw->mac.type == I40E_MAC_VF || hw->mac.type == I40E_MAC_X722_VF) {
1821                 vf = I40EVF_DEV_PRIVATE_TO_VF(dev->data->dev_private);
1822                 vsi = &vf->vsi;
1823                 if (!vsi)
1824                         return -EINVAL;
1825                 reg_idx = queue_idx;
1826         } else {
1827                 pf = I40E_DEV_PRIVATE_TO_PF(dev->data->dev_private);
1828                 vsi = i40e_pf_get_vsi_by_qindex(pf, queue_idx);
1829                 if (!vsi)
1830                         return -EINVAL;
1831                 q_offset = i40e_get_queue_offset_by_qindex(pf, queue_idx);
1832                 if (q_offset < 0)
1833                         return -EINVAL;
1834                 reg_idx = vsi->base_queue + q_offset;
1835         }
1836
1837         if (nb_desc % I40E_ALIGN_RING_DESC != 0 ||
1838             (nb_desc > I40E_MAX_RING_DESC) ||
1839             (nb_desc < I40E_MIN_RING_DESC)) {
1840                 PMD_DRV_LOG(ERR, "Number (%u) of receive descriptors is "
1841                             "invalid", nb_desc);
1842                 return -EINVAL;
1843         }
1844
1845         /* Free memory if needed */
1846         if (dev->data->rx_queues[queue_idx]) {
1847                 i40e_dev_rx_queue_release(dev->data->rx_queues[queue_idx]);
1848                 dev->data->rx_queues[queue_idx] = NULL;
1849         }
1850
1851         /* Allocate the rx queue data structure */
1852         rxq = rte_zmalloc_socket("i40e rx queue",
1853                                  sizeof(struct i40e_rx_queue),
1854                                  RTE_CACHE_LINE_SIZE,
1855                                  socket_id);
1856         if (!rxq) {
1857                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
1858                             "rx queue data structure");
1859                 return -ENOMEM;
1860         }
1861         rxq->mp = mp;
1862         rxq->nb_rx_desc = nb_desc;
1863         rxq->rx_free_thresh = rx_conf->rx_free_thresh;
1864         rxq->queue_id = queue_idx;
1865         rxq->reg_idx = reg_idx;
1866         rxq->port_id = dev->data->port_id;
1867         if (dev->data->dev_conf.rxmode.offloads & DEV_RX_OFFLOAD_KEEP_CRC)
1868                 rxq->crc_len = RTE_ETHER_CRC_LEN;
1869         else
1870                 rxq->crc_len = 0;
1871         rxq->drop_en = rx_conf->rx_drop_en;
1872         rxq->vsi = vsi;
1873         rxq->rx_deferred_start = rx_conf->rx_deferred_start;
1874         rxq->offloads = offloads;
1875
1876         /* Allocate the maximun number of RX ring hardware descriptor. */
1877         len = I40E_MAX_RING_DESC;
1878
1879         /**
1880          * Allocating a little more memory because vectorized/bulk_alloc Rx
1881          * functions doesn't check boundaries each time.
1882          */
1883         len += RTE_PMD_I40E_RX_MAX_BURST;
1884
1885         ring_size = RTE_ALIGN(len * sizeof(union i40e_rx_desc),
1886                               I40E_DMA_MEM_ALIGN);
1887
1888         rz = rte_eth_dma_zone_reserve(dev, "rx_ring", queue_idx,
1889                               ring_size, I40E_RING_BASE_ALIGN, socket_id);
1890         if (!rz) {
1891                 i40e_dev_rx_queue_release(rxq);
1892                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for RX");
1893                 return -ENOMEM;
1894         }
1895
1896         /* Zero all the descriptors in the ring. */
1897         memset(rz->addr, 0, ring_size);
1898
1899         rxq->rx_ring_phys_addr = rz->iova;
1900         rxq->rx_ring = (union i40e_rx_desc *)rz->addr;
1901
1902         len = (uint16_t)(nb_desc + RTE_PMD_I40E_RX_MAX_BURST);
1903
1904         /* Allocate the software ring. */
1905         rxq->sw_ring =
1906                 rte_zmalloc_socket("i40e rx sw ring",
1907                                    sizeof(struct i40e_rx_entry) * len,
1908                                    RTE_CACHE_LINE_SIZE,
1909                                    socket_id);
1910         if (!rxq->sw_ring) {
1911                 i40e_dev_rx_queue_release(rxq);
1912                 PMD_DRV_LOG(ERR, "Failed to allocate memory for SW ring");
1913                 return -ENOMEM;
1914         }
1915
1916         i40e_reset_rx_queue(rxq);
1917         rxq->q_set = TRUE;
1918
1919         for (i = 0; i < I40E_MAX_TRAFFIC_CLASS; i++) {
1920                 if (!(vsi->enabled_tc & (1 << i)))
1921                         continue;
1922                 tc_mapping = rte_le_to_cpu_16(vsi->info.tc_mapping[i]);
1923                 base = (tc_mapping & I40E_AQ_VSI_TC_QUE_OFFSET_MASK) >>
1924                         I40E_AQ_VSI_TC_QUE_OFFSET_SHIFT;
1925                 bsf = (tc_mapping & I40E_AQ_VSI_TC_QUE_NUMBER_MASK) >>
1926                         I40E_AQ_VSI_TC_QUE_NUMBER_SHIFT;
1927
1928                 if (queue_idx >= base && queue_idx < (base + BIT(bsf)))
1929                         rxq->dcb_tc = i;
1930         }
1931
1932         if (dev->data->dev_started) {
1933                 if (i40e_dev_rx_queue_setup_runtime(dev, rxq)) {
1934                         i40e_dev_rx_queue_release(rxq);
1935                         return -EINVAL;
1936                 }
1937         } else {
1938                 use_def_burst_func =
1939                         check_rx_burst_bulk_alloc_preconditions(rxq);
1940                 if (!use_def_burst_func) {
1941 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
1942                         PMD_INIT_LOG(DEBUG,
1943                           "Rx Burst Bulk Alloc Preconditions are "
1944                           "satisfied. Rx Burst Bulk Alloc function will be "
1945                           "used on port=%d, queue=%d.",
1946                           rxq->port_id, rxq->queue_id);
1947 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
1948                 } else {
1949                         PMD_INIT_LOG(DEBUG,
1950                           "Rx Burst Bulk Alloc Preconditions are "
1951                           "not satisfied, Scattered Rx is requested, "
1952                           "or RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC is "
1953                           "not enabled on port=%d, queue=%d.",
1954                           rxq->port_id, rxq->queue_id);
1955                         ad->rx_bulk_alloc_allowed = false;
1956                 }
1957         }
1958
1959         dev->data->rx_queues[queue_idx] = rxq;
1960         return 0;
1961 }
1962
1963 void
1964 i40e_dev_rx_queue_release(void *rxq)
1965 {
1966         struct i40e_rx_queue *q = (struct i40e_rx_queue *)rxq;
1967
1968         if (!q) {
1969                 PMD_DRV_LOG(DEBUG, "Pointer to rxq is NULL");
1970                 return;
1971         }
1972
1973         i40e_rx_queue_release_mbufs(q);
1974         rte_free(q->sw_ring);
1975         rte_free(q);
1976 }
1977
1978 uint32_t
1979 i40e_dev_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1980 {
1981 #define I40E_RXQ_SCAN_INTERVAL 4
1982         volatile union i40e_rx_desc *rxdp;
1983         struct i40e_rx_queue *rxq;
1984         uint16_t desc = 0;
1985
1986         rxq = dev->data->rx_queues[rx_queue_id];
1987         rxdp = &(rxq->rx_ring[rxq->rx_tail]);
1988         while ((desc < rxq->nb_rx_desc) &&
1989                 ((rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len) &
1990                 I40E_RXD_QW1_STATUS_MASK) >> I40E_RXD_QW1_STATUS_SHIFT) &
1991                                 (1 << I40E_RX_DESC_STATUS_DD_SHIFT)) {
1992                 /**
1993                  * Check the DD bit of a rx descriptor of each 4 in a group,
1994                  * to avoid checking too frequently and downgrading performance
1995                  * too much.
1996                  */
1997                 desc += I40E_RXQ_SCAN_INTERVAL;
1998                 rxdp += I40E_RXQ_SCAN_INTERVAL;
1999                 if (rxq->rx_tail + desc >= rxq->nb_rx_desc)
2000                         rxdp = &(rxq->rx_ring[rxq->rx_tail +
2001                                         desc - rxq->nb_rx_desc]);
2002         }
2003
2004         return desc;
2005 }
2006
2007 int
2008 i40e_dev_rx_descriptor_done(void *rx_queue, uint16_t offset)
2009 {
2010         volatile union i40e_rx_desc *rxdp;
2011         struct i40e_rx_queue *rxq = rx_queue;
2012         uint16_t desc;
2013         int ret;
2014
2015         if (unlikely(offset >= rxq->nb_rx_desc)) {
2016                 PMD_DRV_LOG(ERR, "Invalid RX descriptor id %u", offset);
2017                 return 0;
2018         }
2019
2020         desc = rxq->rx_tail + offset;
2021         if (desc >= rxq->nb_rx_desc)
2022                 desc -= rxq->nb_rx_desc;
2023
2024         rxdp = &(rxq->rx_ring[desc]);
2025
2026         ret = !!(((rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len) &
2027                 I40E_RXD_QW1_STATUS_MASK) >> I40E_RXD_QW1_STATUS_SHIFT) &
2028                                 (1 << I40E_RX_DESC_STATUS_DD_SHIFT));
2029
2030         return ret;
2031 }
2032
2033 int
2034 i40e_dev_rx_descriptor_status(void *rx_queue, uint16_t offset)
2035 {
2036         struct i40e_rx_queue *rxq = rx_queue;
2037         volatile uint64_t *status;
2038         uint64_t mask;
2039         uint32_t desc;
2040
2041         if (unlikely(offset >= rxq->nb_rx_desc))
2042                 return -EINVAL;
2043
2044         if (offset >= rxq->nb_rx_desc - rxq->nb_rx_hold)
2045                 return RTE_ETH_RX_DESC_UNAVAIL;
2046
2047         desc = rxq->rx_tail + offset;
2048         if (desc >= rxq->nb_rx_desc)
2049                 desc -= rxq->nb_rx_desc;
2050
2051         status = &rxq->rx_ring[desc].wb.qword1.status_error_len;
2052         mask = rte_le_to_cpu_64((1ULL << I40E_RX_DESC_STATUS_DD_SHIFT)
2053                 << I40E_RXD_QW1_STATUS_SHIFT);
2054         if (*status & mask)
2055                 return RTE_ETH_RX_DESC_DONE;
2056
2057         return RTE_ETH_RX_DESC_AVAIL;
2058 }
2059
2060 int
2061 i40e_dev_tx_descriptor_status(void *tx_queue, uint16_t offset)
2062 {
2063         struct i40e_tx_queue *txq = tx_queue;
2064         volatile uint64_t *status;
2065         uint64_t mask, expect;
2066         uint32_t desc;
2067
2068         if (unlikely(offset >= txq->nb_tx_desc))
2069                 return -EINVAL;
2070
2071         desc = txq->tx_tail + offset;
2072         /* go to next desc that has the RS bit */
2073         desc = ((desc + txq->tx_rs_thresh - 1) / txq->tx_rs_thresh) *
2074                 txq->tx_rs_thresh;
2075         if (desc >= txq->nb_tx_desc) {
2076                 desc -= txq->nb_tx_desc;
2077                 if (desc >= txq->nb_tx_desc)
2078                         desc -= txq->nb_tx_desc;
2079         }
2080
2081         status = &txq->tx_ring[desc].cmd_type_offset_bsz;
2082         mask = rte_le_to_cpu_64(I40E_TXD_QW1_DTYPE_MASK);
2083         expect = rte_cpu_to_le_64(
2084                 I40E_TX_DESC_DTYPE_DESC_DONE << I40E_TXD_QW1_DTYPE_SHIFT);
2085         if ((*status & mask) == expect)
2086                 return RTE_ETH_TX_DESC_DONE;
2087
2088         return RTE_ETH_TX_DESC_FULL;
2089 }
2090
2091 static int
2092 i40e_dev_tx_queue_setup_runtime(struct rte_eth_dev *dev,
2093                                 struct i40e_tx_queue *txq)
2094 {
2095         struct i40e_adapter *ad =
2096                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
2097
2098         if (i40e_tx_queue_init(txq) != I40E_SUCCESS) {
2099                 PMD_DRV_LOG(ERR,
2100                             "Failed to do TX queue initialization");
2101                 return -EINVAL;
2102         }
2103
2104         if (i40e_dev_first_queue(txq->queue_id,
2105                                  dev->data->tx_queues,
2106                                  dev->data->nb_tx_queues)) {
2107                 /**
2108                  * If it is the first queue to setup,
2109                  * set all flags and call
2110                  * i40e_set_tx_function.
2111                  */
2112                 i40e_set_tx_function_flag(dev, txq);
2113                 i40e_set_tx_function(dev);
2114                 return 0;
2115         }
2116
2117         /* check vector conflict */
2118         if (ad->tx_vec_allowed) {
2119                 if (txq->tx_rs_thresh > RTE_I40E_TX_MAX_FREE_BUF_SZ ||
2120                     i40e_txq_vec_setup(txq)) {
2121                         PMD_DRV_LOG(ERR, "Failed vector tx setup.");
2122                         return -EINVAL;
2123                 }
2124         }
2125         /* check simple tx conflict */
2126         if (ad->tx_simple_allowed) {
2127                 if ((txq->offloads & ~DEV_TX_OFFLOAD_MBUF_FAST_FREE) != 0 ||
2128                                 txq->tx_rs_thresh < RTE_PMD_I40E_TX_MAX_BURST) {
2129                         PMD_DRV_LOG(ERR, "No-simple tx is required.");
2130                         return -EINVAL;
2131                 }
2132         }
2133
2134         return 0;
2135 }
2136
2137 int
2138 i40e_dev_tx_queue_setup(struct rte_eth_dev *dev,
2139                         uint16_t queue_idx,
2140                         uint16_t nb_desc,
2141                         unsigned int socket_id,
2142                         const struct rte_eth_txconf *tx_conf)
2143 {
2144         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2145         struct i40e_vsi *vsi;
2146         struct i40e_pf *pf = NULL;
2147         struct i40e_vf *vf = NULL;
2148         struct i40e_tx_queue *txq;
2149         const struct rte_memzone *tz;
2150         uint32_t ring_size;
2151         uint16_t tx_rs_thresh, tx_free_thresh;
2152         uint16_t reg_idx, i, base, bsf, tc_mapping;
2153         int q_offset;
2154         uint64_t offloads;
2155
2156         offloads = tx_conf->offloads | dev->data->dev_conf.txmode.offloads;
2157
2158         if (hw->mac.type == I40E_MAC_VF || hw->mac.type == I40E_MAC_X722_VF) {
2159                 vf = I40EVF_DEV_PRIVATE_TO_VF(dev->data->dev_private);
2160                 vsi = &vf->vsi;
2161                 if (!vsi)
2162                         return -EINVAL;
2163                 reg_idx = queue_idx;
2164         } else {
2165                 pf = I40E_DEV_PRIVATE_TO_PF(dev->data->dev_private);
2166                 vsi = i40e_pf_get_vsi_by_qindex(pf, queue_idx);
2167                 if (!vsi)
2168                         return -EINVAL;
2169                 q_offset = i40e_get_queue_offset_by_qindex(pf, queue_idx);
2170                 if (q_offset < 0)
2171                         return -EINVAL;
2172                 reg_idx = vsi->base_queue + q_offset;
2173         }
2174
2175         if (nb_desc % I40E_ALIGN_RING_DESC != 0 ||
2176             (nb_desc > I40E_MAX_RING_DESC) ||
2177             (nb_desc < I40E_MIN_RING_DESC)) {
2178                 PMD_DRV_LOG(ERR, "Number (%u) of transmit descriptors is "
2179                             "invalid", nb_desc);
2180                 return -EINVAL;
2181         }
2182
2183         /**
2184          * The following two parameters control the setting of the RS bit on
2185          * transmit descriptors. TX descriptors will have their RS bit set
2186          * after txq->tx_rs_thresh descriptors have been used. The TX
2187          * descriptor ring will be cleaned after txq->tx_free_thresh
2188          * descriptors are used or if the number of descriptors required to
2189          * transmit a packet is greater than the number of free TX descriptors.
2190          *
2191          * The following constraints must be satisfied:
2192          *  - tx_rs_thresh must be greater than 0.
2193          *  - tx_rs_thresh must be less than the size of the ring minus 2.
2194          *  - tx_rs_thresh must be less than or equal to tx_free_thresh.
2195          *  - tx_rs_thresh must be a divisor of the ring size.
2196          *  - tx_free_thresh must be greater than 0.
2197          *  - tx_free_thresh must be less than the size of the ring minus 3.
2198          *  - tx_free_thresh + tx_rs_thresh must not exceed nb_desc.
2199          *
2200          * One descriptor in the TX ring is used as a sentinel to avoid a H/W
2201          * race condition, hence the maximum threshold constraints. When set
2202          * to zero use default values.
2203          */
2204         tx_free_thresh = (uint16_t)((tx_conf->tx_free_thresh) ?
2205                 tx_conf->tx_free_thresh : DEFAULT_TX_FREE_THRESH);
2206         /* force tx_rs_thresh to adapt an aggresive tx_free_thresh */
2207         tx_rs_thresh = (DEFAULT_TX_RS_THRESH + tx_free_thresh > nb_desc) ?
2208                 nb_desc - tx_free_thresh : DEFAULT_TX_RS_THRESH;
2209         if (tx_conf->tx_rs_thresh > 0)
2210                 tx_rs_thresh = tx_conf->tx_rs_thresh;
2211         if (tx_rs_thresh + tx_free_thresh > nb_desc) {
2212                 PMD_INIT_LOG(ERR, "tx_rs_thresh + tx_free_thresh must not "
2213                                 "exceed nb_desc. (tx_rs_thresh=%u "
2214                                 "tx_free_thresh=%u nb_desc=%u port=%d queue=%d)",
2215                                 (unsigned int)tx_rs_thresh,
2216                                 (unsigned int)tx_free_thresh,
2217                                 (unsigned int)nb_desc,
2218                                 (int)dev->data->port_id,
2219                                 (int)queue_idx);
2220                 return I40E_ERR_PARAM;
2221         }
2222         if (tx_rs_thresh >= (nb_desc - 2)) {
2223                 PMD_INIT_LOG(ERR, "tx_rs_thresh must be less than the "
2224                              "number of TX descriptors minus 2. "
2225                              "(tx_rs_thresh=%u port=%d queue=%d)",
2226                              (unsigned int)tx_rs_thresh,
2227                              (int)dev->data->port_id,
2228                              (int)queue_idx);
2229                 return I40E_ERR_PARAM;
2230         }
2231         if (tx_free_thresh >= (nb_desc - 3)) {
2232                 PMD_INIT_LOG(ERR, "tx_free_thresh must be less than the "
2233                              "number of TX descriptors minus 3. "
2234                              "(tx_free_thresh=%u port=%d queue=%d)",
2235                              (unsigned int)tx_free_thresh,
2236                              (int)dev->data->port_id,
2237                              (int)queue_idx);
2238                 return I40E_ERR_PARAM;
2239         }
2240         if (tx_rs_thresh > tx_free_thresh) {
2241                 PMD_INIT_LOG(ERR, "tx_rs_thresh must be less than or "
2242                              "equal to tx_free_thresh. (tx_free_thresh=%u"
2243                              " tx_rs_thresh=%u port=%d queue=%d)",
2244                              (unsigned int)tx_free_thresh,
2245                              (unsigned int)tx_rs_thresh,
2246                              (int)dev->data->port_id,
2247                              (int)queue_idx);
2248                 return I40E_ERR_PARAM;
2249         }
2250         if ((nb_desc % tx_rs_thresh) != 0) {
2251                 PMD_INIT_LOG(ERR, "tx_rs_thresh must be a divisor of the "
2252                              "number of TX descriptors. (tx_rs_thresh=%u"
2253                              " port=%d queue=%d)",
2254                              (unsigned int)tx_rs_thresh,
2255                              (int)dev->data->port_id,
2256                              (int)queue_idx);
2257                 return I40E_ERR_PARAM;
2258         }
2259         if ((tx_rs_thresh > 1) && (tx_conf->tx_thresh.wthresh != 0)) {
2260                 PMD_INIT_LOG(ERR, "TX WTHRESH must be set to 0 if "
2261                              "tx_rs_thresh is greater than 1. "
2262                              "(tx_rs_thresh=%u port=%d queue=%d)",
2263                              (unsigned int)tx_rs_thresh,
2264                              (int)dev->data->port_id,
2265                              (int)queue_idx);
2266                 return I40E_ERR_PARAM;
2267         }
2268
2269         /* Free memory if needed. */
2270         if (dev->data->tx_queues[queue_idx]) {
2271                 i40e_dev_tx_queue_release(dev->data->tx_queues[queue_idx]);
2272                 dev->data->tx_queues[queue_idx] = NULL;
2273         }
2274
2275         /* Allocate the TX queue data structure. */
2276         txq = rte_zmalloc_socket("i40e tx queue",
2277                                   sizeof(struct i40e_tx_queue),
2278                                   RTE_CACHE_LINE_SIZE,
2279                                   socket_id);
2280         if (!txq) {
2281                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
2282                             "tx queue structure");
2283                 return -ENOMEM;
2284         }
2285
2286         /* Allocate TX hardware ring descriptors. */
2287         ring_size = sizeof(struct i40e_tx_desc) * I40E_MAX_RING_DESC;
2288         ring_size = RTE_ALIGN(ring_size, I40E_DMA_MEM_ALIGN);
2289         tz = rte_eth_dma_zone_reserve(dev, "tx_ring", queue_idx,
2290                               ring_size, I40E_RING_BASE_ALIGN, socket_id);
2291         if (!tz) {
2292                 i40e_dev_tx_queue_release(txq);
2293                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for TX");
2294                 return -ENOMEM;
2295         }
2296
2297         txq->nb_tx_desc = nb_desc;
2298         txq->tx_rs_thresh = tx_rs_thresh;
2299         txq->tx_free_thresh = tx_free_thresh;
2300         txq->pthresh = tx_conf->tx_thresh.pthresh;
2301         txq->hthresh = tx_conf->tx_thresh.hthresh;
2302         txq->wthresh = tx_conf->tx_thresh.wthresh;
2303         txq->queue_id = queue_idx;
2304         txq->reg_idx = reg_idx;
2305         txq->port_id = dev->data->port_id;
2306         txq->offloads = offloads;
2307         txq->vsi = vsi;
2308         txq->tx_deferred_start = tx_conf->tx_deferred_start;
2309
2310         txq->tx_ring_phys_addr = tz->iova;
2311         txq->tx_ring = (struct i40e_tx_desc *)tz->addr;
2312
2313         /* Allocate software ring */
2314         txq->sw_ring =
2315                 rte_zmalloc_socket("i40e tx sw ring",
2316                                    sizeof(struct i40e_tx_entry) * nb_desc,
2317                                    RTE_CACHE_LINE_SIZE,
2318                                    socket_id);
2319         if (!txq->sw_ring) {
2320                 i40e_dev_tx_queue_release(txq);
2321                 PMD_DRV_LOG(ERR, "Failed to allocate memory for SW TX ring");
2322                 return -ENOMEM;
2323         }
2324
2325         i40e_reset_tx_queue(txq);
2326         txq->q_set = TRUE;
2327
2328         for (i = 0; i < I40E_MAX_TRAFFIC_CLASS; i++) {
2329                 if (!(vsi->enabled_tc & (1 << i)))
2330                         continue;
2331                 tc_mapping = rte_le_to_cpu_16(vsi->info.tc_mapping[i]);
2332                 base = (tc_mapping & I40E_AQ_VSI_TC_QUE_OFFSET_MASK) >>
2333                         I40E_AQ_VSI_TC_QUE_OFFSET_SHIFT;
2334                 bsf = (tc_mapping & I40E_AQ_VSI_TC_QUE_NUMBER_MASK) >>
2335                         I40E_AQ_VSI_TC_QUE_NUMBER_SHIFT;
2336
2337                 if (queue_idx >= base && queue_idx < (base + BIT(bsf)))
2338                         txq->dcb_tc = i;
2339         }
2340
2341         if (dev->data->dev_started) {
2342                 if (i40e_dev_tx_queue_setup_runtime(dev, txq)) {
2343                         i40e_dev_tx_queue_release(txq);
2344                         return -EINVAL;
2345                 }
2346         } else {
2347                 /**
2348                  * Use a simple TX queue without offloads or
2349                  * multi segs if possible
2350                  */
2351                 i40e_set_tx_function_flag(dev, txq);
2352         }
2353         dev->data->tx_queues[queue_idx] = txq;
2354
2355         return 0;
2356 }
2357
2358 void
2359 i40e_dev_tx_queue_release(void *txq)
2360 {
2361         struct i40e_tx_queue *q = (struct i40e_tx_queue *)txq;
2362
2363         if (!q) {
2364                 PMD_DRV_LOG(DEBUG, "Pointer to TX queue is NULL");
2365                 return;
2366         }
2367
2368         i40e_tx_queue_release_mbufs(q);
2369         rte_free(q->sw_ring);
2370         rte_free(q);
2371 }
2372
2373 const struct rte_memzone *
2374 i40e_memzone_reserve(const char *name, uint32_t len, int socket_id)
2375 {
2376         const struct rte_memzone *mz;
2377
2378         mz = rte_memzone_lookup(name);
2379         if (mz)
2380                 return mz;
2381
2382         mz = rte_memzone_reserve_aligned(name, len, socket_id,
2383                         RTE_MEMZONE_IOVA_CONTIG, I40E_RING_BASE_ALIGN);
2384         return mz;
2385 }
2386
2387 void
2388 i40e_rx_queue_release_mbufs(struct i40e_rx_queue *rxq)
2389 {
2390         uint16_t i;
2391
2392         /* SSE Vector driver has a different way of releasing mbufs. */
2393         if (rxq->rx_using_sse) {
2394                 i40e_rx_queue_release_mbufs_vec(rxq);
2395                 return;
2396         }
2397
2398         if (!rxq->sw_ring) {
2399                 PMD_DRV_LOG(DEBUG, "Pointer to sw_ring is NULL");
2400                 return;
2401         }
2402
2403         for (i = 0; i < rxq->nb_rx_desc; i++) {
2404                 if (rxq->sw_ring[i].mbuf) {
2405                         rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
2406                         rxq->sw_ring[i].mbuf = NULL;
2407                 }
2408         }
2409 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2410         if (rxq->rx_nb_avail == 0)
2411                 return;
2412         for (i = 0; i < rxq->rx_nb_avail; i++) {
2413                 struct rte_mbuf *mbuf;
2414
2415                 mbuf = rxq->rx_stage[rxq->rx_next_avail + i];
2416                 rte_pktmbuf_free_seg(mbuf);
2417         }
2418         rxq->rx_nb_avail = 0;
2419 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2420 }
2421
2422 void
2423 i40e_reset_rx_queue(struct i40e_rx_queue *rxq)
2424 {
2425         unsigned i;
2426         uint16_t len;
2427
2428         if (!rxq) {
2429                 PMD_DRV_LOG(DEBUG, "Pointer to rxq is NULL");
2430                 return;
2431         }
2432
2433 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2434         if (check_rx_burst_bulk_alloc_preconditions(rxq) == 0)
2435                 len = (uint16_t)(rxq->nb_rx_desc + RTE_PMD_I40E_RX_MAX_BURST);
2436         else
2437 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2438                 len = rxq->nb_rx_desc;
2439
2440         for (i = 0; i < len * sizeof(union i40e_rx_desc); i++)
2441                 ((volatile char *)rxq->rx_ring)[i] = 0;
2442
2443         memset(&rxq->fake_mbuf, 0x0, sizeof(rxq->fake_mbuf));
2444         for (i = 0; i < RTE_PMD_I40E_RX_MAX_BURST; ++i)
2445                 rxq->sw_ring[rxq->nb_rx_desc + i].mbuf = &rxq->fake_mbuf;
2446
2447 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2448         rxq->rx_nb_avail = 0;
2449         rxq->rx_next_avail = 0;
2450         rxq->rx_free_trigger = (uint16_t)(rxq->rx_free_thresh - 1);
2451 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2452         rxq->rx_tail = 0;
2453         rxq->nb_rx_hold = 0;
2454         rxq->pkt_first_seg = NULL;
2455         rxq->pkt_last_seg = NULL;
2456
2457         rxq->rxrearm_start = 0;
2458         rxq->rxrearm_nb = 0;
2459 }
2460
2461 void
2462 i40e_tx_queue_release_mbufs(struct i40e_tx_queue *txq)
2463 {
2464         struct rte_eth_dev *dev;
2465         uint16_t i;
2466
2467         if (!txq || !txq->sw_ring) {
2468                 PMD_DRV_LOG(DEBUG, "Pointer to txq or sw_ring is NULL");
2469                 return;
2470         }
2471
2472         dev = &rte_eth_devices[txq->port_id];
2473
2474         /**
2475          *  vPMD tx will not set sw_ring's mbuf to NULL after free,
2476          *  so need to free remains more carefully.
2477          */
2478         if (dev->tx_pkt_burst == i40e_xmit_pkts_vec_avx2 ||
2479                         dev->tx_pkt_burst == i40e_xmit_pkts_vec) {
2480                 i = txq->tx_next_dd - txq->tx_rs_thresh + 1;
2481                 if (txq->tx_tail < i) {
2482                         for (; i < txq->nb_tx_desc; i++) {
2483                                 rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
2484                                 txq->sw_ring[i].mbuf = NULL;
2485                         }
2486                         i = 0;
2487                 }
2488                 for (; i < txq->tx_tail; i++) {
2489                         rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
2490                         txq->sw_ring[i].mbuf = NULL;
2491                 }
2492         } else {
2493                 for (i = 0; i < txq->nb_tx_desc; i++) {
2494                         if (txq->sw_ring[i].mbuf) {
2495                                 rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
2496                                 txq->sw_ring[i].mbuf = NULL;
2497                         }
2498                 }
2499         }
2500 }
2501
2502 static int
2503 i40e_tx_done_cleanup_full(struct i40e_tx_queue *txq,
2504                         uint32_t free_cnt)
2505 {
2506         struct i40e_tx_entry *swr_ring = txq->sw_ring;
2507         uint16_t i, tx_last, tx_id;
2508         uint16_t nb_tx_free_last;
2509         uint16_t nb_tx_to_clean;
2510         uint32_t pkt_cnt;
2511
2512         /* Start free mbuf from the next of tx_tail */
2513         tx_last = txq->tx_tail;
2514         tx_id  = swr_ring[tx_last].next_id;
2515
2516         if (txq->nb_tx_free == 0 && i40e_xmit_cleanup(txq))
2517                 return 0;
2518
2519         nb_tx_to_clean = txq->nb_tx_free;
2520         nb_tx_free_last = txq->nb_tx_free;
2521         if (!free_cnt)
2522                 free_cnt = txq->nb_tx_desc;
2523
2524         /* Loop through swr_ring to count the amount of
2525          * freeable mubfs and packets.
2526          */
2527         for (pkt_cnt = 0; pkt_cnt < free_cnt; ) {
2528                 for (i = 0; i < nb_tx_to_clean &&
2529                         pkt_cnt < free_cnt &&
2530                         tx_id != tx_last; i++) {
2531                         if (swr_ring[tx_id].mbuf != NULL) {
2532                                 rte_pktmbuf_free_seg(swr_ring[tx_id].mbuf);
2533                                 swr_ring[tx_id].mbuf = NULL;
2534
2535                                 /*
2536                                  * last segment in the packet,
2537                                  * increment packet count
2538                                  */
2539                                 pkt_cnt += (swr_ring[tx_id].last_id == tx_id);
2540                         }
2541
2542                         tx_id = swr_ring[tx_id].next_id;
2543                 }
2544
2545                 if (txq->tx_rs_thresh > txq->nb_tx_desc -
2546                         txq->nb_tx_free || tx_id == tx_last)
2547                         break;
2548
2549                 if (pkt_cnt < free_cnt) {
2550                         if (i40e_xmit_cleanup(txq))
2551                                 break;
2552
2553                         nb_tx_to_clean = txq->nb_tx_free - nb_tx_free_last;
2554                         nb_tx_free_last = txq->nb_tx_free;
2555                 }
2556         }
2557
2558         return (int)pkt_cnt;
2559 }
2560
2561 static int
2562 i40e_tx_done_cleanup_simple(struct i40e_tx_queue *txq,
2563                         uint32_t free_cnt)
2564 {
2565         int i, n, cnt;
2566
2567         if (free_cnt == 0 || free_cnt > txq->nb_tx_desc)
2568                 free_cnt = txq->nb_tx_desc;
2569
2570         cnt = free_cnt - free_cnt % txq->tx_rs_thresh;
2571
2572         for (i = 0; i < cnt; i += n) {
2573                 if (txq->nb_tx_desc - txq->nb_tx_free < txq->tx_rs_thresh)
2574                         break;
2575
2576                 n = i40e_tx_free_bufs(txq);
2577
2578                 if (n == 0)
2579                         break;
2580         }
2581
2582         return i;
2583 }
2584
2585 static int
2586 i40e_tx_done_cleanup_vec(struct i40e_tx_queue *txq __rte_unused,
2587                         uint32_t free_cnt __rte_unused)
2588 {
2589         return -ENOTSUP;
2590 }
2591 int
2592 i40e_tx_done_cleanup(void *txq, uint32_t free_cnt)
2593 {
2594         struct i40e_tx_queue *q = (struct i40e_tx_queue *)txq;
2595         struct rte_eth_dev *dev = &rte_eth_devices[q->port_id];
2596         struct i40e_adapter *ad =
2597                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
2598
2599         if (ad->tx_simple_allowed) {
2600                 if (ad->tx_vec_allowed)
2601                         return i40e_tx_done_cleanup_vec(q, free_cnt);
2602                 else
2603                         return i40e_tx_done_cleanup_simple(q, free_cnt);
2604         } else {
2605                 return i40e_tx_done_cleanup_full(q, free_cnt);
2606         }
2607 }
2608
2609 void
2610 i40e_reset_tx_queue(struct i40e_tx_queue *txq)
2611 {
2612         struct i40e_tx_entry *txe;
2613         uint16_t i, prev, size;
2614
2615         if (!txq) {
2616                 PMD_DRV_LOG(DEBUG, "Pointer to txq is NULL");
2617                 return;
2618         }
2619
2620         txe = txq->sw_ring;
2621         size = sizeof(struct i40e_tx_desc) * txq->nb_tx_desc;
2622         for (i = 0; i < size; i++)
2623                 ((volatile char *)txq->tx_ring)[i] = 0;
2624
2625         prev = (uint16_t)(txq->nb_tx_desc - 1);
2626         for (i = 0; i < txq->nb_tx_desc; i++) {
2627                 volatile struct i40e_tx_desc *txd = &txq->tx_ring[i];
2628
2629                 txd->cmd_type_offset_bsz =
2630                         rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE);
2631                 txe[i].mbuf =  NULL;
2632                 txe[i].last_id = i;
2633                 txe[prev].next_id = i;
2634                 prev = i;
2635         }
2636
2637         txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
2638         txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
2639
2640         txq->tx_tail = 0;
2641         txq->nb_tx_used = 0;
2642
2643         txq->last_desc_cleaned = (uint16_t)(txq->nb_tx_desc - 1);
2644         txq->nb_tx_free = (uint16_t)(txq->nb_tx_desc - 1);
2645 }
2646
2647 /* Init the TX queue in hardware */
2648 int
2649 i40e_tx_queue_init(struct i40e_tx_queue *txq)
2650 {
2651         enum i40e_status_code err = I40E_SUCCESS;
2652         struct i40e_vsi *vsi = txq->vsi;
2653         struct i40e_hw *hw = I40E_VSI_TO_HW(vsi);
2654         uint16_t pf_q = txq->reg_idx;
2655         struct i40e_hmc_obj_txq tx_ctx;
2656         uint32_t qtx_ctl;
2657
2658         /* clear the context structure first */
2659         memset(&tx_ctx, 0, sizeof(tx_ctx));
2660         tx_ctx.new_context = 1;
2661         tx_ctx.base = txq->tx_ring_phys_addr / I40E_QUEUE_BASE_ADDR_UNIT;
2662         tx_ctx.qlen = txq->nb_tx_desc;
2663
2664 #ifdef RTE_LIBRTE_IEEE1588
2665         tx_ctx.timesync_ena = 1;
2666 #endif
2667         tx_ctx.rdylist = rte_le_to_cpu_16(vsi->info.qs_handle[txq->dcb_tc]);
2668         if (vsi->type == I40E_VSI_FDIR)
2669                 tx_ctx.fd_ena = TRUE;
2670
2671         err = i40e_clear_lan_tx_queue_context(hw, pf_q);
2672         if (err != I40E_SUCCESS) {
2673                 PMD_DRV_LOG(ERR, "Failure of clean lan tx queue context");
2674                 return err;
2675         }
2676
2677         err = i40e_set_lan_tx_queue_context(hw, pf_q, &tx_ctx);
2678         if (err != I40E_SUCCESS) {
2679                 PMD_DRV_LOG(ERR, "Failure of set lan tx queue context");
2680                 return err;
2681         }
2682
2683         /* Now associate this queue with this PCI function */
2684         qtx_ctl = I40E_QTX_CTL_PF_QUEUE;
2685         qtx_ctl |= ((hw->pf_id << I40E_QTX_CTL_PF_INDX_SHIFT) &
2686                                         I40E_QTX_CTL_PF_INDX_MASK);
2687         I40E_WRITE_REG(hw, I40E_QTX_CTL(pf_q), qtx_ctl);
2688         I40E_WRITE_FLUSH(hw);
2689
2690         txq->qtx_tail = hw->hw_addr + I40E_QTX_TAIL(pf_q);
2691
2692         return err;
2693 }
2694
2695 int
2696 i40e_alloc_rx_queue_mbufs(struct i40e_rx_queue *rxq)
2697 {
2698         struct i40e_rx_entry *rxe = rxq->sw_ring;
2699         uint64_t dma_addr;
2700         uint16_t i;
2701
2702         for (i = 0; i < rxq->nb_rx_desc; i++) {
2703                 volatile union i40e_rx_desc *rxd;
2704                 struct rte_mbuf *mbuf = rte_mbuf_raw_alloc(rxq->mp);
2705
2706                 if (unlikely(!mbuf)) {
2707                         PMD_DRV_LOG(ERR, "Failed to allocate mbuf for RX");
2708                         return -ENOMEM;
2709                 }
2710
2711                 rte_mbuf_refcnt_set(mbuf, 1);
2712                 mbuf->next = NULL;
2713                 mbuf->data_off = RTE_PKTMBUF_HEADROOM;
2714                 mbuf->nb_segs = 1;
2715                 mbuf->port = rxq->port_id;
2716
2717                 dma_addr =
2718                         rte_cpu_to_le_64(rte_mbuf_data_iova_default(mbuf));
2719
2720                 rxd = &rxq->rx_ring[i];
2721                 rxd->read.pkt_addr = dma_addr;
2722                 rxd->read.hdr_addr = 0;
2723 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
2724                 rxd->read.rsvd1 = 0;
2725                 rxd->read.rsvd2 = 0;
2726 #endif /* RTE_LIBRTE_I40E_16BYTE_RX_DESC */
2727
2728                 rxe[i].mbuf = mbuf;
2729         }
2730
2731         return 0;
2732 }
2733
2734 /*
2735  * Calculate the buffer length, and check the jumbo frame
2736  * and maximum packet length.
2737  */
2738 static int
2739 i40e_rx_queue_config(struct i40e_rx_queue *rxq)
2740 {
2741         struct i40e_pf *pf = I40E_VSI_TO_PF(rxq->vsi);
2742         struct i40e_hw *hw = I40E_VSI_TO_HW(rxq->vsi);
2743         struct rte_eth_dev_data *data = pf->dev_data;
2744         uint16_t buf_size;
2745
2746         buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mp) -
2747                 RTE_PKTMBUF_HEADROOM);
2748
2749         switch (pf->flags & (I40E_FLAG_HEADER_SPLIT_DISABLED |
2750                         I40E_FLAG_HEADER_SPLIT_ENABLED)) {
2751         case I40E_FLAG_HEADER_SPLIT_ENABLED: /* Not supported */
2752                 rxq->rx_hdr_len = RTE_ALIGN(I40E_RXBUF_SZ_1024,
2753                                 (1 << I40E_RXQ_CTX_HBUFF_SHIFT));
2754                 rxq->rx_buf_len = RTE_ALIGN(I40E_RXBUF_SZ_2048,
2755                                 (1 << I40E_RXQ_CTX_DBUFF_SHIFT));
2756                 rxq->hs_mode = i40e_header_split_enabled;
2757                 break;
2758         case I40E_FLAG_HEADER_SPLIT_DISABLED:
2759         default:
2760                 rxq->rx_hdr_len = 0;
2761                 rxq->rx_buf_len = RTE_ALIGN_FLOOR(buf_size,
2762                         (1 << I40E_RXQ_CTX_DBUFF_SHIFT));
2763                 rxq->hs_mode = i40e_header_split_none;
2764                 break;
2765         }
2766
2767         rxq->max_pkt_len =
2768                 RTE_MIN((uint32_t)(hw->func_caps.rx_buf_chain_len *
2769                         rxq->rx_buf_len), data->dev_conf.rxmode.max_rx_pkt_len);
2770         if (data->dev_conf.rxmode.offloads & DEV_RX_OFFLOAD_JUMBO_FRAME) {
2771                 if (rxq->max_pkt_len <= RTE_ETHER_MAX_LEN ||
2772                         rxq->max_pkt_len > I40E_FRAME_SIZE_MAX) {
2773                         PMD_DRV_LOG(ERR, "maximum packet length must "
2774                                     "be larger than %u and smaller than %u,"
2775                                     "as jumbo frame is enabled",
2776                                     (uint32_t)RTE_ETHER_MAX_LEN,
2777                                     (uint32_t)I40E_FRAME_SIZE_MAX);
2778                         return I40E_ERR_CONFIG;
2779                 }
2780         } else {
2781                 if (rxq->max_pkt_len < RTE_ETHER_MIN_LEN ||
2782                         rxq->max_pkt_len > RTE_ETHER_MAX_LEN) {
2783                         PMD_DRV_LOG(ERR, "maximum packet length must be "
2784                                     "larger than %u and smaller than %u, "
2785                                     "as jumbo frame is disabled",
2786                                     (uint32_t)RTE_ETHER_MIN_LEN,
2787                                     (uint32_t)RTE_ETHER_MAX_LEN);
2788                         return I40E_ERR_CONFIG;
2789                 }
2790         }
2791
2792         return 0;
2793 }
2794
2795 /* Init the RX queue in hardware */
2796 int
2797 i40e_rx_queue_init(struct i40e_rx_queue *rxq)
2798 {
2799         int err = I40E_SUCCESS;
2800         struct i40e_hw *hw = I40E_VSI_TO_HW(rxq->vsi);
2801         struct rte_eth_dev_data *dev_data = I40E_VSI_TO_DEV_DATA(rxq->vsi);
2802         uint16_t pf_q = rxq->reg_idx;
2803         uint16_t buf_size;
2804         struct i40e_hmc_obj_rxq rx_ctx;
2805
2806         err = i40e_rx_queue_config(rxq);
2807         if (err < 0) {
2808                 PMD_DRV_LOG(ERR, "Failed to config RX queue");
2809                 return err;
2810         }
2811
2812         /* Clear the context structure first */
2813         memset(&rx_ctx, 0, sizeof(struct i40e_hmc_obj_rxq));
2814         rx_ctx.dbuff = rxq->rx_buf_len >> I40E_RXQ_CTX_DBUFF_SHIFT;
2815         rx_ctx.hbuff = rxq->rx_hdr_len >> I40E_RXQ_CTX_HBUFF_SHIFT;
2816
2817         rx_ctx.base = rxq->rx_ring_phys_addr / I40E_QUEUE_BASE_ADDR_UNIT;
2818         rx_ctx.qlen = rxq->nb_rx_desc;
2819 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
2820         rx_ctx.dsize = 1;
2821 #endif
2822         rx_ctx.dtype = rxq->hs_mode;
2823         if (rxq->hs_mode)
2824                 rx_ctx.hsplit_0 = I40E_HEADER_SPLIT_ALL;
2825         else
2826                 rx_ctx.hsplit_0 = I40E_HEADER_SPLIT_NONE;
2827         rx_ctx.rxmax = rxq->max_pkt_len;
2828         rx_ctx.tphrdesc_ena = 1;
2829         rx_ctx.tphwdesc_ena = 1;
2830         rx_ctx.tphdata_ena = 1;
2831         rx_ctx.tphhead_ena = 1;
2832         rx_ctx.lrxqthresh = 2;
2833         rx_ctx.crcstrip = (rxq->crc_len == 0) ? 1 : 0;
2834         rx_ctx.l2tsel = 1;
2835         /* showiv indicates if inner VLAN is stripped inside of tunnel
2836          * packet. When set it to 1, vlan information is stripped from
2837          * the inner header, but the hardware does not put it in the
2838          * descriptor. So set it zero by default.
2839          */
2840         rx_ctx.showiv = 0;
2841         rx_ctx.prefena = 1;
2842
2843         err = i40e_clear_lan_rx_queue_context(hw, pf_q);
2844         if (err != I40E_SUCCESS) {
2845                 PMD_DRV_LOG(ERR, "Failed to clear LAN RX queue context");
2846                 return err;
2847         }
2848         err = i40e_set_lan_rx_queue_context(hw, pf_q, &rx_ctx);
2849         if (err != I40E_SUCCESS) {
2850                 PMD_DRV_LOG(ERR, "Failed to set LAN RX queue context");
2851                 return err;
2852         }
2853
2854         rxq->qrx_tail = hw->hw_addr + I40E_QRX_TAIL(pf_q);
2855
2856         buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mp) -
2857                 RTE_PKTMBUF_HEADROOM);
2858
2859         /* Check if scattered RX needs to be used. */
2860         if (rxq->max_pkt_len > buf_size)
2861                 dev_data->scattered_rx = 1;
2862
2863         /* Init the RX tail regieter. */
2864         I40E_PCI_REG_WRITE(rxq->qrx_tail, rxq->nb_rx_desc - 1);
2865
2866         return 0;
2867 }
2868
2869 void
2870 i40e_dev_clear_queues(struct rte_eth_dev *dev)
2871 {
2872         uint16_t i;
2873
2874         PMD_INIT_FUNC_TRACE();
2875
2876         for (i = 0; i < dev->data->nb_tx_queues; i++) {
2877                 if (!dev->data->tx_queues[i])
2878                         continue;
2879                 i40e_tx_queue_release_mbufs(dev->data->tx_queues[i]);
2880                 i40e_reset_tx_queue(dev->data->tx_queues[i]);
2881         }
2882
2883         for (i = 0; i < dev->data->nb_rx_queues; i++) {
2884                 if (!dev->data->rx_queues[i])
2885                         continue;
2886                 i40e_rx_queue_release_mbufs(dev->data->rx_queues[i]);
2887                 i40e_reset_rx_queue(dev->data->rx_queues[i]);
2888         }
2889 }
2890
2891 void
2892 i40e_dev_free_queues(struct rte_eth_dev *dev)
2893 {
2894         uint16_t i;
2895
2896         PMD_INIT_FUNC_TRACE();
2897
2898         for (i = 0; i < dev->data->nb_rx_queues; i++) {
2899                 if (!dev->data->rx_queues[i])
2900                         continue;
2901                 i40e_dev_rx_queue_release(dev->data->rx_queues[i]);
2902                 dev->data->rx_queues[i] = NULL;
2903         }
2904
2905         for (i = 0; i < dev->data->nb_tx_queues; i++) {
2906                 if (!dev->data->tx_queues[i])
2907                         continue;
2908                 i40e_dev_tx_queue_release(dev->data->tx_queues[i]);
2909                 dev->data->tx_queues[i] = NULL;
2910         }
2911 }
2912
2913 #define I40E_FDIR_NUM_TX_DESC  I40E_MIN_RING_DESC
2914 #define I40E_FDIR_NUM_RX_DESC  I40E_MIN_RING_DESC
2915
2916 enum i40e_status_code
2917 i40e_fdir_setup_tx_resources(struct i40e_pf *pf)
2918 {
2919         struct i40e_tx_queue *txq;
2920         const struct rte_memzone *tz = NULL;
2921         uint32_t ring_size;
2922         struct rte_eth_dev *dev;
2923
2924         if (!pf) {
2925                 PMD_DRV_LOG(ERR, "PF is not available");
2926                 return I40E_ERR_BAD_PTR;
2927         }
2928
2929         dev = pf->adapter->eth_dev;
2930
2931         /* Allocate the TX queue data structure. */
2932         txq = rte_zmalloc_socket("i40e fdir tx queue",
2933                                   sizeof(struct i40e_tx_queue),
2934                                   RTE_CACHE_LINE_SIZE,
2935                                   SOCKET_ID_ANY);
2936         if (!txq) {
2937                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
2938                                         "tx queue structure.");
2939                 return I40E_ERR_NO_MEMORY;
2940         }
2941
2942         /* Allocate TX hardware ring descriptors. */
2943         ring_size = sizeof(struct i40e_tx_desc) * I40E_FDIR_NUM_TX_DESC;
2944         ring_size = RTE_ALIGN(ring_size, I40E_DMA_MEM_ALIGN);
2945
2946         tz = rte_eth_dma_zone_reserve(dev, "fdir_tx_ring",
2947                                       I40E_FDIR_QUEUE_ID, ring_size,
2948                                       I40E_RING_BASE_ALIGN, SOCKET_ID_ANY);
2949         if (!tz) {
2950                 i40e_dev_tx_queue_release(txq);
2951                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for TX.");
2952                 return I40E_ERR_NO_MEMORY;
2953         }
2954
2955         txq->nb_tx_desc = I40E_FDIR_NUM_TX_DESC;
2956         txq->queue_id = I40E_FDIR_QUEUE_ID;
2957         txq->reg_idx = pf->fdir.fdir_vsi->base_queue;
2958         txq->vsi = pf->fdir.fdir_vsi;
2959
2960         txq->tx_ring_phys_addr = tz->iova;
2961         txq->tx_ring = (struct i40e_tx_desc *)tz->addr;
2962         /*
2963          * don't need to allocate software ring and reset for the fdir
2964          * program queue just set the queue has been configured.
2965          */
2966         txq->q_set = TRUE;
2967         pf->fdir.txq = txq;
2968
2969         return I40E_SUCCESS;
2970 }
2971
2972 enum i40e_status_code
2973 i40e_fdir_setup_rx_resources(struct i40e_pf *pf)
2974 {
2975         struct i40e_rx_queue *rxq;
2976         const struct rte_memzone *rz = NULL;
2977         uint32_t ring_size;
2978         struct rte_eth_dev *dev;
2979
2980         if (!pf) {
2981                 PMD_DRV_LOG(ERR, "PF is not available");
2982                 return I40E_ERR_BAD_PTR;
2983         }
2984
2985         dev = pf->adapter->eth_dev;
2986
2987         /* Allocate the RX queue data structure. */
2988         rxq = rte_zmalloc_socket("i40e fdir rx queue",
2989                                   sizeof(struct i40e_rx_queue),
2990                                   RTE_CACHE_LINE_SIZE,
2991                                   SOCKET_ID_ANY);
2992         if (!rxq) {
2993                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
2994                                         "rx queue structure.");
2995                 return I40E_ERR_NO_MEMORY;
2996         }
2997
2998         /* Allocate RX hardware ring descriptors. */
2999         ring_size = sizeof(union i40e_rx_desc) * I40E_FDIR_NUM_RX_DESC;
3000         ring_size = RTE_ALIGN(ring_size, I40E_DMA_MEM_ALIGN);
3001
3002         rz = rte_eth_dma_zone_reserve(dev, "fdir_rx_ring",
3003                                       I40E_FDIR_QUEUE_ID, ring_size,
3004                                       I40E_RING_BASE_ALIGN, SOCKET_ID_ANY);
3005         if (!rz) {
3006                 i40e_dev_rx_queue_release(rxq);
3007                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for RX.");
3008                 return I40E_ERR_NO_MEMORY;
3009         }
3010
3011         rxq->nb_rx_desc = I40E_FDIR_NUM_RX_DESC;
3012         rxq->queue_id = I40E_FDIR_QUEUE_ID;
3013         rxq->reg_idx = pf->fdir.fdir_vsi->base_queue;
3014         rxq->vsi = pf->fdir.fdir_vsi;
3015
3016         rxq->rx_ring_phys_addr = rz->iova;
3017         memset(rz->addr, 0, I40E_FDIR_NUM_RX_DESC * sizeof(union i40e_rx_desc));
3018         rxq->rx_ring = (union i40e_rx_desc *)rz->addr;
3019
3020         /*
3021          * Don't need to allocate software ring and reset for the fdir
3022          * rx queue, just set the queue has been configured.
3023          */
3024         rxq->q_set = TRUE;
3025         pf->fdir.rxq = rxq;
3026
3027         return I40E_SUCCESS;
3028 }
3029
3030 void
3031 i40e_rxq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
3032         struct rte_eth_rxq_info *qinfo)
3033 {
3034         struct i40e_rx_queue *rxq;
3035
3036         rxq = dev->data->rx_queues[queue_id];
3037
3038         qinfo->mp = rxq->mp;
3039         qinfo->scattered_rx = dev->data->scattered_rx;
3040         qinfo->nb_desc = rxq->nb_rx_desc;
3041
3042         qinfo->conf.rx_free_thresh = rxq->rx_free_thresh;
3043         qinfo->conf.rx_drop_en = rxq->drop_en;
3044         qinfo->conf.rx_deferred_start = rxq->rx_deferred_start;
3045         qinfo->conf.offloads = rxq->offloads;
3046 }
3047
3048 void
3049 i40e_txq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
3050         struct rte_eth_txq_info *qinfo)
3051 {
3052         struct i40e_tx_queue *txq;
3053
3054         txq = dev->data->tx_queues[queue_id];
3055
3056         qinfo->nb_desc = txq->nb_tx_desc;
3057
3058         qinfo->conf.tx_thresh.pthresh = txq->pthresh;
3059         qinfo->conf.tx_thresh.hthresh = txq->hthresh;
3060         qinfo->conf.tx_thresh.wthresh = txq->wthresh;
3061
3062         qinfo->conf.tx_free_thresh = txq->tx_free_thresh;
3063         qinfo->conf.tx_rs_thresh = txq->tx_rs_thresh;
3064         qinfo->conf.tx_deferred_start = txq->tx_deferred_start;
3065         qinfo->conf.offloads = txq->offloads;
3066 }
3067
3068 static eth_rx_burst_t
3069 i40e_get_latest_rx_vec(bool scatter)
3070 {
3071 #if defined(RTE_ARCH_X86) && defined(CC_AVX2_SUPPORT)
3072         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2))
3073                 return scatter ? i40e_recv_scattered_pkts_vec_avx2 :
3074                                  i40e_recv_pkts_vec_avx2;
3075 #endif
3076         return scatter ? i40e_recv_scattered_pkts_vec :
3077                          i40e_recv_pkts_vec;
3078 }
3079
3080 static eth_rx_burst_t
3081 i40e_get_recommend_rx_vec(bool scatter)
3082 {
3083 #if defined(RTE_ARCH_X86) && defined(CC_AVX2_SUPPORT)
3084         /*
3085          * since AVX frequency can be different to base frequency, limit
3086          * use of AVX2 version to later plaforms, not all those that could
3087          * theoretically run it.
3088          */
3089         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F))
3090                 return scatter ? i40e_recv_scattered_pkts_vec_avx2 :
3091                                  i40e_recv_pkts_vec_avx2;
3092 #endif
3093         return scatter ? i40e_recv_scattered_pkts_vec :
3094                          i40e_recv_pkts_vec;
3095 }
3096
3097 void __attribute__((cold))
3098 i40e_set_rx_function(struct rte_eth_dev *dev)
3099 {
3100         struct i40e_adapter *ad =
3101                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3102         uint16_t rx_using_sse, i;
3103         /* In order to allow Vector Rx there are a few configuration
3104          * conditions to be met and Rx Bulk Allocation should be allowed.
3105          */
3106         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
3107                 if (i40e_rx_vec_dev_conf_condition_check(dev) ||
3108                     !ad->rx_bulk_alloc_allowed) {
3109                         PMD_INIT_LOG(DEBUG, "Port[%d] doesn't meet"
3110                                      " Vector Rx preconditions",
3111                                      dev->data->port_id);
3112
3113                         ad->rx_vec_allowed = false;
3114                 }
3115                 if (ad->rx_vec_allowed) {
3116                         for (i = 0; i < dev->data->nb_rx_queues; i++) {
3117                                 struct i40e_rx_queue *rxq =
3118                                         dev->data->rx_queues[i];
3119
3120                                 if (rxq && i40e_rxq_vec_setup(rxq)) {
3121                                         ad->rx_vec_allowed = false;
3122                                         break;
3123                                 }
3124                         }
3125                 }
3126         }
3127
3128         if (ad->rx_vec_allowed) {
3129                 /* Vec Rx path */
3130                 PMD_INIT_LOG(DEBUG, "Vector Rx path will be used on port=%d.",
3131                                 dev->data->port_id);
3132                 if (ad->use_latest_vec)
3133                         dev->rx_pkt_burst =
3134                         i40e_get_latest_rx_vec(dev->data->scattered_rx);
3135                 else
3136                         dev->rx_pkt_burst =
3137                         i40e_get_recommend_rx_vec(dev->data->scattered_rx);
3138         } else if (!dev->data->scattered_rx && ad->rx_bulk_alloc_allowed) {
3139                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are "
3140                                     "satisfied. Rx Burst Bulk Alloc function "
3141                                     "will be used on port=%d.",
3142                              dev->data->port_id);
3143
3144                 dev->rx_pkt_burst = i40e_recv_pkts_bulk_alloc;
3145         } else {
3146                 /* Simple Rx Path. */
3147                 PMD_INIT_LOG(DEBUG, "Simple Rx path will be used on port=%d.",
3148                              dev->data->port_id);
3149                 dev->rx_pkt_burst = dev->data->scattered_rx ?
3150                                         i40e_recv_scattered_pkts :
3151                                         i40e_recv_pkts;
3152         }
3153
3154         /* Propagate information about RX function choice through all queues. */
3155         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
3156                 rx_using_sse =
3157                         (dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec ||
3158                          dev->rx_pkt_burst == i40e_recv_pkts_vec ||
3159                          dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec_avx2 ||
3160                          dev->rx_pkt_burst == i40e_recv_pkts_vec_avx2);
3161
3162                 for (i = 0; i < dev->data->nb_rx_queues; i++) {
3163                         struct i40e_rx_queue *rxq = dev->data->rx_queues[i];
3164
3165                         if (rxq)
3166                                 rxq->rx_using_sse = rx_using_sse;
3167                 }
3168         }
3169 }
3170
3171 static const struct {
3172         eth_rx_burst_t pkt_burst;
3173         const char *info;
3174 } i40e_rx_burst_infos[] = {
3175         { i40e_recv_scattered_pkts,          "Scalar Scattered" },
3176         { i40e_recv_pkts_bulk_alloc,         "Scalar Bulk Alloc" },
3177         { i40e_recv_pkts,                    "Scalar" },
3178 #ifdef RTE_ARCH_X86
3179         { i40e_recv_scattered_pkts_vec_avx2, "Vector AVX2 Scattered" },
3180         { i40e_recv_pkts_vec_avx2,           "Vector AVX2" },
3181         { i40e_recv_scattered_pkts_vec,      "Vector SSE Scattered" },
3182         { i40e_recv_pkts_vec,                "Vector SSE" },
3183 #elif defined(RTE_ARCH_ARM64)
3184         { i40e_recv_scattered_pkts_vec,      "Vector Neon Scattered" },
3185         { i40e_recv_pkts_vec,                "Vector Neon" },
3186 #elif defined(RTE_ARCH_PPC_64)
3187         { i40e_recv_scattered_pkts_vec,      "Vector AltiVec Scattered" },
3188         { i40e_recv_pkts_vec,                "Vector AltiVec" },
3189 #endif
3190 };
3191
3192 int
3193 i40e_rx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id,
3194                        struct rte_eth_burst_mode *mode)
3195 {
3196         eth_rx_burst_t pkt_burst = dev->rx_pkt_burst;
3197         int ret = -EINVAL;
3198         unsigned int i;
3199
3200         for (i = 0; i < RTE_DIM(i40e_rx_burst_infos); ++i) {
3201                 if (pkt_burst == i40e_rx_burst_infos[i].pkt_burst) {
3202                         snprintf(mode->info, sizeof(mode->info), "%s",
3203                                  i40e_rx_burst_infos[i].info);
3204                         ret = 0;
3205                         break;
3206                 }
3207         }
3208
3209         return ret;
3210 }
3211
3212 void __attribute__((cold))
3213 i40e_set_tx_function_flag(struct rte_eth_dev *dev, struct i40e_tx_queue *txq)
3214 {
3215         struct i40e_adapter *ad =
3216                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3217
3218         /* Use a simple Tx queue if possible (only fast free is allowed) */
3219         ad->tx_simple_allowed =
3220                 (txq->offloads ==
3221                  (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE) &&
3222                  txq->tx_rs_thresh >= RTE_PMD_I40E_TX_MAX_BURST);
3223         ad->tx_vec_allowed = (ad->tx_simple_allowed &&
3224                         txq->tx_rs_thresh <= RTE_I40E_TX_MAX_FREE_BUF_SZ);
3225
3226         if (ad->tx_vec_allowed)
3227                 PMD_INIT_LOG(DEBUG, "Vector Tx can be enabled on Tx queue %u.",
3228                                 txq->queue_id);
3229         else if (ad->tx_simple_allowed)
3230                 PMD_INIT_LOG(DEBUG, "Simple Tx can be enabled on Tx queue %u.",
3231                                 txq->queue_id);
3232         else
3233                 PMD_INIT_LOG(DEBUG,
3234                                 "Neither simple nor vector Tx enabled on Tx queue %u\n",
3235                                 txq->queue_id);
3236 }
3237
3238 static eth_tx_burst_t
3239 i40e_get_latest_tx_vec(void)
3240 {
3241 #if defined(RTE_ARCH_X86) && defined(CC_AVX2_SUPPORT)
3242         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2))
3243                 return i40e_xmit_pkts_vec_avx2;
3244 #endif
3245         return i40e_xmit_pkts_vec;
3246 }
3247
3248 static eth_tx_burst_t
3249 i40e_get_recommend_tx_vec(void)
3250 {
3251 #if defined(RTE_ARCH_X86) && defined(CC_AVX2_SUPPORT)
3252         /*
3253          * since AVX frequency can be different to base frequency, limit
3254          * use of AVX2 version to later plaforms, not all those that could
3255          * theoretically run it.
3256          */
3257         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F))
3258                 return i40e_xmit_pkts_vec_avx2;
3259 #endif
3260         return i40e_xmit_pkts_vec;
3261 }
3262
3263 void __attribute__((cold))
3264 i40e_set_tx_function(struct rte_eth_dev *dev)
3265 {
3266         struct i40e_adapter *ad =
3267                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3268         int i;
3269
3270         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
3271                 if (ad->tx_vec_allowed) {
3272                         for (i = 0; i < dev->data->nb_tx_queues; i++) {
3273                                 struct i40e_tx_queue *txq =
3274                                         dev->data->tx_queues[i];
3275
3276                                 if (txq && i40e_txq_vec_setup(txq)) {
3277                                         ad->tx_vec_allowed = false;
3278                                         break;
3279                                 }
3280                         }
3281                 }
3282         }
3283
3284         if (ad->tx_simple_allowed) {
3285                 if (ad->tx_vec_allowed) {
3286                         PMD_INIT_LOG(DEBUG, "Vector tx finally be used.");
3287                         if (ad->use_latest_vec)
3288                                 dev->tx_pkt_burst =
3289                                         i40e_get_latest_tx_vec();
3290                         else
3291                                 dev->tx_pkt_burst =
3292                                         i40e_get_recommend_tx_vec();
3293                 } else {
3294                         PMD_INIT_LOG(DEBUG, "Simple tx finally be used.");
3295                         dev->tx_pkt_burst = i40e_xmit_pkts_simple;
3296                 }
3297                 dev->tx_pkt_prepare = NULL;
3298         } else {
3299                 PMD_INIT_LOG(DEBUG, "Xmit tx finally be used.");
3300                 dev->tx_pkt_burst = i40e_xmit_pkts;
3301                 dev->tx_pkt_prepare = i40e_prep_pkts;
3302         }
3303 }
3304
3305 static const struct {
3306         eth_tx_burst_t pkt_burst;
3307         const char *info;
3308 } i40e_tx_burst_infos[] = {
3309         { i40e_xmit_pkts_simple,   "Scalar Simple" },
3310         { i40e_xmit_pkts,          "Scalar" },
3311 #ifdef RTE_ARCH_X86
3312         { i40e_xmit_pkts_vec_avx2, "Vector AVX2" },
3313         { i40e_xmit_pkts_vec,      "Vector SSE" },
3314 #elif defined(RTE_ARCH_ARM64)
3315         { i40e_xmit_pkts_vec,      "Vector Neon" },
3316 #elif defined(RTE_ARCH_PPC_64)
3317         { i40e_xmit_pkts_vec,      "Vector AltiVec" },
3318 #endif
3319 };
3320
3321 int
3322 i40e_tx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id,
3323                        struct rte_eth_burst_mode *mode)
3324 {
3325         eth_tx_burst_t pkt_burst = dev->tx_pkt_burst;
3326         int ret = -EINVAL;
3327         unsigned int i;
3328
3329         for (i = 0; i < RTE_DIM(i40e_tx_burst_infos); ++i) {
3330                 if (pkt_burst == i40e_tx_burst_infos[i].pkt_burst) {
3331                         snprintf(mode->info, sizeof(mode->info), "%s",
3332                                  i40e_tx_burst_infos[i].info);
3333                         ret = 0;
3334                         break;
3335                 }
3336         }
3337
3338         return ret;
3339 }
3340
3341 void __attribute__((cold))
3342 i40e_set_default_ptype_table(struct rte_eth_dev *dev)
3343 {
3344         struct i40e_adapter *ad =
3345                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3346         int i;
3347
3348         for (i = 0; i < I40E_MAX_PKT_TYPE; i++)
3349                 ad->ptype_tbl[i] = i40e_get_default_pkt_type(i);
3350 }
3351
3352 void __attribute__((cold))
3353 i40e_set_default_pctype_table(struct rte_eth_dev *dev)
3354 {
3355         struct i40e_adapter *ad =
3356                         I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3357         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3358         int i;
3359
3360         for (i = 0; i < I40E_FLOW_TYPE_MAX; i++)
3361                 ad->pctypes_tbl[i] = 0ULL;
3362         ad->flow_types_mask = 0ULL;
3363         ad->pctypes_mask = 0ULL;
3364
3365         ad->pctypes_tbl[RTE_ETH_FLOW_FRAG_IPV4] =
3366                                 (1ULL << I40E_FILTER_PCTYPE_FRAG_IPV4);
3367         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_UDP] =
3368                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_UDP);
3369         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_TCP] =
3370                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_TCP);
3371         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_SCTP] =
3372                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_SCTP);
3373         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_OTHER] =
3374                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_OTHER);
3375         ad->pctypes_tbl[RTE_ETH_FLOW_FRAG_IPV6] =
3376                                 (1ULL << I40E_FILTER_PCTYPE_FRAG_IPV6);
3377         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_UDP] =
3378                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_UDP);
3379         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_TCP] =
3380                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_TCP);
3381         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_SCTP] =
3382                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_SCTP);
3383         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_OTHER] =
3384                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_OTHER);
3385         ad->pctypes_tbl[RTE_ETH_FLOW_L2_PAYLOAD] =
3386                                 (1ULL << I40E_FILTER_PCTYPE_L2_PAYLOAD);
3387
3388         if (hw->mac.type == I40E_MAC_X722 ||
3389                 hw->mac.type == I40E_MAC_X722_VF) {
3390                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_UDP] |=
3391                         (1ULL << I40E_FILTER_PCTYPE_NONF_UNICAST_IPV4_UDP);
3392                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_UDP] |=
3393                         (1ULL << I40E_FILTER_PCTYPE_NONF_MULTICAST_IPV4_UDP);
3394                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_TCP] |=
3395                         (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_TCP_SYN_NO_ACK);
3396                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_UDP] |=
3397                         (1ULL << I40E_FILTER_PCTYPE_NONF_UNICAST_IPV6_UDP);
3398                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_UDP] |=
3399                         (1ULL << I40E_FILTER_PCTYPE_NONF_MULTICAST_IPV6_UDP);
3400                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_TCP] |=
3401                         (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_TCP_SYN_NO_ACK);
3402         }
3403
3404         for (i = 0; i < I40E_FLOW_TYPE_MAX; i++) {
3405                 if (ad->pctypes_tbl[i])
3406                         ad->flow_types_mask |= (1ULL << i);
3407                 ad->pctypes_mask |= ad->pctypes_tbl[i];
3408         }
3409 }
3410
3411 #ifndef RTE_LIBRTE_I40E_INC_VECTOR
3412 /* Stubs needed for linkage when CONFIG_RTE_LIBRTE_I40E_INC_VECTOR is set to 'n' */
3413 int
3414 i40e_rx_vec_dev_conf_condition_check(struct rte_eth_dev __rte_unused *dev)
3415 {
3416         return -1;
3417 }
3418
3419 uint16_t
3420 i40e_recv_pkts_vec(
3421         void __rte_unused *rx_queue,
3422         struct rte_mbuf __rte_unused **rx_pkts,
3423         uint16_t __rte_unused nb_pkts)
3424 {
3425         return 0;
3426 }
3427
3428 uint16_t
3429 i40e_recv_scattered_pkts_vec(
3430         void __rte_unused *rx_queue,
3431         struct rte_mbuf __rte_unused **rx_pkts,
3432         uint16_t __rte_unused nb_pkts)
3433 {
3434         return 0;
3435 }
3436
3437 int
3438 i40e_rxq_vec_setup(struct i40e_rx_queue __rte_unused *rxq)
3439 {
3440         return -1;
3441 }
3442
3443 int
3444 i40e_txq_vec_setup(struct i40e_tx_queue __rte_unused *txq)
3445 {
3446         return -1;
3447 }
3448
3449 void
3450 i40e_rx_queue_release_mbufs_vec(struct i40e_rx_queue __rte_unused*rxq)
3451 {
3452         return;
3453 }
3454
3455 uint16_t
3456 i40e_xmit_fixed_burst_vec(void __rte_unused * tx_queue,
3457                           struct rte_mbuf __rte_unused **tx_pkts,
3458                           uint16_t __rte_unused nb_pkts)
3459 {
3460         return 0;
3461 }
3462 #endif /* ifndef RTE_LIBRTE_I40E_INC_VECTOR */
3463
3464 #ifndef CC_AVX2_SUPPORT
3465 uint16_t
3466 i40e_recv_pkts_vec_avx2(void __rte_unused *rx_queue,
3467                         struct rte_mbuf __rte_unused **rx_pkts,
3468                         uint16_t __rte_unused nb_pkts)
3469 {
3470         return 0;
3471 }
3472
3473 uint16_t
3474 i40e_recv_scattered_pkts_vec_avx2(void __rte_unused *rx_queue,
3475                         struct rte_mbuf __rte_unused **rx_pkts,
3476                         uint16_t __rte_unused nb_pkts)
3477 {
3478         return 0;
3479 }
3480
3481 uint16_t
3482 i40e_xmit_pkts_vec_avx2(void __rte_unused * tx_queue,
3483                           struct rte_mbuf __rte_unused **tx_pkts,
3484                           uint16_t __rte_unused nb_pkts)
3485 {
3486         return 0;
3487 }
3488 #endif /* ifndef CC_AVX2_SUPPORT */