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 <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 #include <rte_vect.h>
  27
  28 #include "i40e_logs.h"
  29 #include "base/i40e_prototype.h"
  30 #include "base/i40e_type.h"
  31 #include "i40e_ethdev.h"
  32 #include "i40e_rxtx.h"
  33
  34 #define DEFAULT_TX_RS_THRESH   32
  35 #define DEFAULT_TX_FREE_THRESH 32
  36
  37 #define I40E_TX_MAX_BURST  32
  38
  39 #define I40E_DMA_MEM_ALIGN 4096
  40
  41 /* Base address of the HW descriptor ring should be 128B aligned. */
  42 #define I40E_RING_BASE_ALIGN    128
  43
  44 #define I40E_TXD_CMD (I40E_TX_DESC_CMD_EOP | I40E_TX_DESC_CMD_RS)
  45
  46 #ifdef RTE_LIBRTE_IEEE1588
  47 #define I40E_TX_IEEE1588_TMST PKT_TX_IEEE1588_TMST
  48 #else
  49 #define I40E_TX_IEEE1588_TMST 0
  50 #endif
  51
  52 #define I40E_TX_CKSUM_OFFLOAD_MASK (             \
  53                 PKT_TX_IP_CKSUM |                \
  54                 PKT_TX_L4_MASK |                 \
  55                 PKT_TX_TCP_SEG |                 \
  56                 PKT_TX_OUTER_IP_CKSUM)
  57
  58 #define I40E_TX_OFFLOAD_MASK (  \
  59                 PKT_TX_OUTER_IPV4 |     \
  60                 PKT_TX_OUTER_IPV6 |     \
  61                 PKT_TX_IPV4 |           \
  62                 PKT_TX_IPV6 |           \
  63                 PKT_TX_IP_CKSUM |       \
  64                 PKT_TX_L4_MASK |        \
  65                 PKT_TX_OUTER_IP_CKSUM | \
  66                 PKT_TX_TCP_SEG |        \
  67                 PKT_TX_QINQ_PKT |       \
  68                 PKT_TX_VLAN_PKT |       \
  69                 PKT_TX_TUNNEL_MASK |    \
  70                 I40E_TX_IEEE1588_TMST)
  71
  72 #define I40E_TX_OFFLOAD_NOTSUP_MASK \
  73                 (PKT_TX_OFFLOAD_MASK ^ I40E_TX_OFFLOAD_MASK)
  74
  75 int
  76 i40e_get_monitor_addr(void *rx_queue, struct rte_power_monitor_cond *pmc)
  77 {
  78         struct i40e_rx_queue *rxq = rx_queue;
  79         volatile union i40e_rx_desc *rxdp;
  80         uint16_t desc;
  81
  82         desc = rxq->rx_tail;
  83         rxdp = &rxq->rx_ring[desc];
  84         /* watch for changes in status bit */
  85         pmc->addr = &rxdp->wb.qword1.status_error_len;
  86
  87         /*
  88          * we expect the DD bit to be set to 1 if this descriptor was already
  89          * written to.
  90          */
  91         pmc->val = rte_cpu_to_le_64(1 << I40E_RX_DESC_STATUS_DD_SHIFT);
  92         pmc->mask = rte_cpu_to_le_64(1 << I40E_RX_DESC_STATUS_DD_SHIFT);
  93
  94         /* registers are 64-bit */
  95         pmc->size = sizeof(uint64_t);
  96
  97         return 0;
  98 }
  99
 100 static inline void
 101 i40e_rxd_to_vlan_tci(struct rte_mbuf *mb, volatile union i40e_rx_desc *rxdp)
 102 {
 103         if (rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len) &
 104                 (1 << I40E_RX_DESC_STATUS_L2TAG1P_SHIFT)) {
 105                 mb->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
 106                 mb->vlan_tci =
 107                         rte_le_to_cpu_16(rxdp->wb.qword0.lo_dword.l2tag1);
 108                 PMD_RX_LOG(DEBUG, "Descriptor l2tag1: %u",
 109                            rte_le_to_cpu_16(rxdp->wb.qword0.lo_dword.l2tag1));
 110         } else {
 111                 mb->vlan_tci = 0;
 112         }
 113 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
 114         if (rte_le_to_cpu_16(rxdp->wb.qword2.ext_status) &
 115                 (1 << I40E_RX_DESC_EXT_STATUS_L2TAG2P_SHIFT)) {
 116                 mb->ol_flags |= PKT_RX_QINQ_STRIPPED | PKT_RX_QINQ |
 117                         PKT_RX_VLAN_STRIPPED | PKT_RX_VLAN;
 118                 mb->vlan_tci_outer = mb->vlan_tci;
 119                 mb->vlan_tci = rte_le_to_cpu_16(rxdp->wb.qword2.l2tag2_2);
 120                 PMD_RX_LOG(DEBUG, "Descriptor l2tag2_1: %u, l2tag2_2: %u",
 121                            rte_le_to_cpu_16(rxdp->wb.qword2.l2tag2_1),
 122                            rte_le_to_cpu_16(rxdp->wb.qword2.l2tag2_2));
 123         } else {
 124                 mb->vlan_tci_outer = 0;
 125         }
 126 #endif
 127         PMD_RX_LOG(DEBUG, "Mbuf vlan_tci: %u, vlan_tci_outer: %u",
 128                    mb->vlan_tci, mb->vlan_tci_outer);
 129 }
 130
 131 /* Translate the rx descriptor status to pkt flags */
 132 static inline uint64_t
 133 i40e_rxd_status_to_pkt_flags(uint64_t qword)
 134 {
 135         uint64_t flags;
 136
 137         /* Check if RSS_HASH */
 138         flags = (((qword >> I40E_RX_DESC_STATUS_FLTSTAT_SHIFT) &
 139                                         I40E_RX_DESC_FLTSTAT_RSS_HASH) ==
 140                         I40E_RX_DESC_FLTSTAT_RSS_HASH) ? PKT_RX_RSS_HASH : 0;
 141
 142         /* Check if FDIR Match */
 143         flags |= (qword & (1 << I40E_RX_DESC_STATUS_FLM_SHIFT) ?
 144                                                         PKT_RX_FDIR : 0);
 145
 146         return flags;
 147 }
 148
 149 static inline uint64_t
 150 i40e_rxd_error_to_pkt_flags(uint64_t qword)
 151 {
 152         uint64_t flags = 0;
 153         uint64_t error_bits = (qword >> I40E_RXD_QW1_ERROR_SHIFT);
 154
 155 #define I40E_RX_ERR_BITS 0x3f
 156         if (likely((error_bits & I40E_RX_ERR_BITS) == 0)) {
 157                 flags |= (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD);
 158                 return flags;
 159         }
 160
 161         if (unlikely(error_bits & (1 << I40E_RX_DESC_ERROR_IPE_SHIFT)))
 162                 flags |= PKT_RX_IP_CKSUM_BAD;
 163         else
 164                 flags |= PKT_RX_IP_CKSUM_GOOD;
 165
 166         if (unlikely(error_bits & (1 << I40E_RX_DESC_ERROR_L4E_SHIFT)))
 167                 flags |= PKT_RX_L4_CKSUM_BAD;
 168         else
 169                 flags |= PKT_RX_L4_CKSUM_GOOD;
 170
 171         if (unlikely(error_bits & (1 << I40E_RX_DESC_ERROR_EIPE_SHIFT)))
 172                 flags |= PKT_RX_OUTER_IP_CKSUM_BAD;
 173
 174         return flags;
 175 }
 176
 177 /* Function to check and set the ieee1588 timesync index and get the
 178  * appropriate flags.
 179  */
 180 #ifdef RTE_LIBRTE_IEEE1588
 181 static inline uint64_t
 182 i40e_get_iee15888_flags(struct rte_mbuf *mb, uint64_t qword)
 183 {
 184         uint64_t pkt_flags = 0;
 185         uint16_t tsyn = (qword & (I40E_RXD_QW1_STATUS_TSYNVALID_MASK
 186                                   | I40E_RXD_QW1_STATUS_TSYNINDX_MASK))
 187                                     >> I40E_RX_DESC_STATUS_TSYNINDX_SHIFT;
 188
 189         if ((mb->packet_type & RTE_PTYPE_L2_MASK)
 190                         == RTE_PTYPE_L2_ETHER_TIMESYNC)
 191                 pkt_flags = PKT_RX_IEEE1588_PTP;
 192         if (tsyn & 0x04) {
 193                 pkt_flags |= PKT_RX_IEEE1588_TMST;
 194                 mb->timesync = tsyn & 0x03;
 195         }
 196
 197         return pkt_flags;
 198 }
 199 #endif
 200
 201 static inline uint64_t
 202 i40e_rxd_build_fdir(volatile union i40e_rx_desc *rxdp, struct rte_mbuf *mb)
 203 {
 204         uint64_t flags = 0;
 205 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
 206         uint16_t flexbh, flexbl;
 207
 208         flexbh = (rte_le_to_cpu_32(rxdp->wb.qword2.ext_status) >>
 209                 I40E_RX_DESC_EXT_STATUS_FLEXBH_SHIFT) &
 210                 I40E_RX_DESC_EXT_STATUS_FLEXBH_MASK;
 211         flexbl = (rte_le_to_cpu_32(rxdp->wb.qword2.ext_status) >>
 212                 I40E_RX_DESC_EXT_STATUS_FLEXBL_SHIFT) &
 213                 I40E_RX_DESC_EXT_STATUS_FLEXBL_MASK;
 214
 215
 216         if (flexbh == I40E_RX_DESC_EXT_STATUS_FLEXBH_FD_ID) {
 217                 mb->hash.fdir.hi =
 218                         rte_le_to_cpu_32(rxdp->wb.qword3.hi_dword.fd_id);
 219                 flags |= PKT_RX_FDIR_ID;
 220         } else if (flexbh == I40E_RX_DESC_EXT_STATUS_FLEXBH_FLEX) {
 221                 mb->hash.fdir.hi =
 222                         rte_le_to_cpu_32(rxdp->wb.qword3.hi_dword.flex_bytes_hi);
 223                 flags |= PKT_RX_FDIR_FLX;
 224         }
 225         if (flexbl == I40E_RX_DESC_EXT_STATUS_FLEXBL_FLEX) {
 226                 mb->hash.fdir.lo =
 227                         rte_le_to_cpu_32(rxdp->wb.qword3.lo_dword.flex_bytes_lo);
 228                 flags |= PKT_RX_FDIR_FLX;
 229         }
 230 #else
 231         mb->hash.fdir.hi =
 232                 rte_le_to_cpu_32(rxdp->wb.qword0.hi_dword.fd_id);
 233         flags |= PKT_RX_FDIR_ID;
 234 #endif
 235         return flags;
 236 }
 237
 238 static inline void
 239 i40e_parse_tunneling_params(uint64_t ol_flags,
 240                             union i40e_tx_offload tx_offload,
 241                             uint32_t *cd_tunneling)
 242 {
 243         /* EIPT: External (outer) IP header type */
 244         if (ol_flags & PKT_TX_OUTER_IP_CKSUM)
 245                 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV4;
 246         else if (ol_flags & PKT_TX_OUTER_IPV4)
 247                 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV4_NO_CSUM;
 248         else if (ol_flags & PKT_TX_OUTER_IPV6)
 249                 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV6;
 250
 251         /* EIPLEN: External (outer) IP header length, in DWords */
 252         *cd_tunneling |= (tx_offload.outer_l3_len >> 2) <<
 253                 I40E_TXD_CTX_QW0_EXT_IPLEN_SHIFT;
 254
 255         /* L4TUNT: L4 Tunneling Type */
 256         switch (ol_flags & PKT_TX_TUNNEL_MASK) {
 257         case PKT_TX_TUNNEL_IPIP:
 258                 /* for non UDP / GRE tunneling, set to 00b */
 259                 break;
 260         case PKT_TX_TUNNEL_VXLAN:
 261         case PKT_TX_TUNNEL_GENEVE:
 262                 *cd_tunneling |= I40E_TXD_CTX_UDP_TUNNELING;
 263                 break;
 264         case PKT_TX_TUNNEL_GRE:
 265                 *cd_tunneling |= I40E_TXD_CTX_GRE_TUNNELING;
 266                 break;
 267         default:
 268                 PMD_TX_LOG(ERR, "Tunnel type not supported");
 269                 return;
 270         }
 271
 272         /* L4TUNLEN: L4 Tunneling Length, in Words
 273          *
 274          * We depend on app to set rte_mbuf.l2_len correctly.
 275          * For IP in GRE it should be set to the length of the GRE
 276          * header;
 277          * for MAC in GRE or MAC in UDP it should be set to the length
 278          * of the GRE or UDP headers plus the inner MAC up to including
 279          * its last Ethertype.
 280          */
 281         *cd_tunneling |= (tx_offload.l2_len >> 1) <<
 282                 I40E_TXD_CTX_QW0_NATLEN_SHIFT;
 283 }
 284
 285 static inline void
 286 i40e_txd_enable_checksum(uint64_t ol_flags,
 287                         uint32_t *td_cmd,
 288                         uint32_t *td_offset,
 289                         union i40e_tx_offload tx_offload)
 290 {
 291         /* Set MACLEN */
 292         if (ol_flags & PKT_TX_TUNNEL_MASK)
 293                 *td_offset |= (tx_offload.outer_l2_len >> 1)
 294                                 << I40E_TX_DESC_LENGTH_MACLEN_SHIFT;
 295         else
 296                 *td_offset |= (tx_offload.l2_len >> 1)
 297                         << I40E_TX_DESC_LENGTH_MACLEN_SHIFT;
 298
 299         /* Enable L3 checksum offloads */
 300         if (ol_flags & PKT_TX_IP_CKSUM) {
 301                 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV4_CSUM;
 302                 *td_offset |= (tx_offload.l3_len >> 2)
 303                                 << I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
 304         } else if (ol_flags & PKT_TX_IPV4) {
 305                 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV4;
 306                 *td_offset |= (tx_offload.l3_len >> 2)
 307                                 << I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
 308         } else if (ol_flags & PKT_TX_IPV6) {
 309                 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV6;
 310                 *td_offset |= (tx_offload.l3_len >> 2)
 311                                 << I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
 312         }
 313
 314         if (ol_flags & PKT_TX_TCP_SEG) {
 315                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_TCP;
 316                 *td_offset |= (tx_offload.l4_len >> 2)
 317                         << I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
 318                 return;
 319         }
 320
 321         /* Enable L4 checksum offloads */
 322         switch (ol_flags & PKT_TX_L4_MASK) {
 323         case PKT_TX_TCP_CKSUM:
 324                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_TCP;
 325                 *td_offset |= (sizeof(struct rte_tcp_hdr) >> 2) <<
 326                                 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
 327                 break;
 328         case PKT_TX_SCTP_CKSUM:
 329                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_SCTP;
 330                 *td_offset |= (sizeof(struct rte_sctp_hdr) >> 2) <<
 331                                 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
 332                 break;
 333         case PKT_TX_UDP_CKSUM:
 334                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_UDP;
 335                 *td_offset |= (sizeof(struct rte_udp_hdr) >> 2) <<
 336                                 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
 337                 break;
 338         default:
 339                 break;
 340         }
 341 }
 342
 343 /* Construct the tx flags */
 344 static inline uint64_t
 345 i40e_build_ctob(uint32_t td_cmd,
 346                 uint32_t td_offset,
 347                 unsigned int size,
 348                 uint32_t td_tag)
 349 {
 350         return rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DATA |
 351                         ((uint64_t)td_cmd  << I40E_TXD_QW1_CMD_SHIFT) |
 352                         ((uint64_t)td_offset << I40E_TXD_QW1_OFFSET_SHIFT) |
 353                         ((uint64_t)size  << I40E_TXD_QW1_TX_BUF_SZ_SHIFT) |
 354                         ((uint64_t)td_tag  << I40E_TXD_QW1_L2TAG1_SHIFT));
 355 }
 356
 357 static inline int
 358 i40e_xmit_cleanup(struct i40e_tx_queue *txq)
 359 {
 360         struct i40e_tx_entry *sw_ring = txq->sw_ring;
 361         volatile struct i40e_tx_desc *txd = txq->tx_ring;
 362         uint16_t last_desc_cleaned = txq->last_desc_cleaned;
 363         uint16_t nb_tx_desc = txq->nb_tx_desc;
 364         uint16_t desc_to_clean_to;
 365         uint16_t nb_tx_to_clean;
 366
 367         desc_to_clean_to = (uint16_t)(last_desc_cleaned + txq->tx_rs_thresh);
 368         if (desc_to_clean_to >= nb_tx_desc)
 369                 desc_to_clean_to = (uint16_t)(desc_to_clean_to - nb_tx_desc);
 370
 371         desc_to_clean_to = sw_ring[desc_to_clean_to].last_id;
 372         if ((txd[desc_to_clean_to].cmd_type_offset_bsz &
 373                         rte_cpu_to_le_64(I40E_TXD_QW1_DTYPE_MASK)) !=
 374                         rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE)) {
 375                 PMD_TX_LOG(DEBUG, "TX descriptor %4u is not done "
 376                            "(port=%d queue=%d)", desc_to_clean_to,
 377                            txq->port_id, txq->queue_id);
 378                 return -1;
 379         }
 380
 381         if (last_desc_cleaned > desc_to_clean_to)
 382                 nb_tx_to_clean = (uint16_t)((nb_tx_desc - last_desc_cleaned) +
 383                                                         desc_to_clean_to);
 384         else
 385                 nb_tx_to_clean = (uint16_t)(desc_to_clean_to -
 386                                         last_desc_cleaned);
 387
 388         txd[desc_to_clean_to].cmd_type_offset_bsz = 0;
 389
 390         txq->last_desc_cleaned = desc_to_clean_to;
 391         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + nb_tx_to_clean);
 392
 393         return 0;
 394 }
 395
 396 static inline int
 397 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
 398 check_rx_burst_bulk_alloc_preconditions(struct i40e_rx_queue *rxq)
 399 #else
 400 check_rx_burst_bulk_alloc_preconditions(__rte_unused struct i40e_rx_queue *rxq)
 401 #endif
 402 {
 403         int ret = 0;
 404
 405 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
 406         if (!(rxq->rx_free_thresh >= RTE_PMD_I40E_RX_MAX_BURST)) {
 407                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
 408                              "rxq->rx_free_thresh=%d, "
 409                              "RTE_PMD_I40E_RX_MAX_BURST=%d",
 410                              rxq->rx_free_thresh, RTE_PMD_I40E_RX_MAX_BURST);
 411                 ret = -EINVAL;
 412         } else if (!(rxq->rx_free_thresh < rxq->nb_rx_desc)) {
 413                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
 414                              "rxq->rx_free_thresh=%d, "
 415                              "rxq->nb_rx_desc=%d",
 416                              rxq->rx_free_thresh, rxq->nb_rx_desc);
 417                 ret = -EINVAL;
 418         } else if (rxq->nb_rx_desc % rxq->rx_free_thresh != 0) {
 419                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
 420                              "rxq->nb_rx_desc=%d, "
 421                              "rxq->rx_free_thresh=%d",
 422                              rxq->nb_rx_desc, rxq->rx_free_thresh);
 423                 ret = -EINVAL;
 424         }
 425 #else
 426         ret = -EINVAL;
 427 #endif
 428
 429         return ret;
 430 }
 431
 432 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
 433 #define I40E_LOOK_AHEAD 8
 434 #if (I40E_LOOK_AHEAD != 8)
 435 #error "PMD I40E: I40E_LOOK_AHEAD must be 8\n"
 436 #endif
 437 static inline int
 438 i40e_rx_scan_hw_ring(struct i40e_rx_queue *rxq)
 439 {
 440         volatile union i40e_rx_desc *rxdp;
 441         struct i40e_rx_entry *rxep;
 442         struct rte_mbuf *mb;
 443         uint16_t pkt_len;
 444         uint64_t qword1;
 445         uint32_t rx_status;
 446         int32_t s[I40E_LOOK_AHEAD], nb_dd;
 447         int32_t i, j, nb_rx = 0;
 448         uint64_t pkt_flags;
 449         uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
 450
 451         rxdp = &rxq->rx_ring[rxq->rx_tail];
 452         rxep = &rxq->sw_ring[rxq->rx_tail];
 453
 454         qword1 = rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len);
 455         rx_status = (qword1 & I40E_RXD_QW1_STATUS_MASK) >>
 456                                 I40E_RXD_QW1_STATUS_SHIFT;
 457
 458         /* Make sure there is at least 1 packet to receive */
 459         if (!(rx_status & (1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
 460                 return 0;
 461
 462         /**
 463          * Scan LOOK_AHEAD descriptors at a time to determine which
 464          * descriptors reference packets that are ready to be received.
 465          */
 466         for (i = 0; i < RTE_PMD_I40E_RX_MAX_BURST; i+=I40E_LOOK_AHEAD,
 467                         rxdp += I40E_LOOK_AHEAD, rxep += I40E_LOOK_AHEAD) {
 468                 /* Read desc statuses backwards to avoid race condition */
 469                 for (j = I40E_LOOK_AHEAD - 1; j >= 0; j--) {
 470                         qword1 = rte_le_to_cpu_64(\
 471                                 rxdp[j].wb.qword1.status_error_len);
 472                         s[j] = (qword1 & I40E_RXD_QW1_STATUS_MASK) >>
 473                                         I40E_RXD_QW1_STATUS_SHIFT;
 474                 }
 475
 476                 rte_smp_rmb();
 477
 478                 /* Compute how many status bits were set */
 479                 for (j = 0, nb_dd = 0; j < I40E_LOOK_AHEAD; j++)
 480                         nb_dd += s[j] & (1 << I40E_RX_DESC_STATUS_DD_SHIFT);
 481
 482                 nb_rx += nb_dd;
 483
 484                 /* Translate descriptor info to mbuf parameters */
 485                 for (j = 0; j < nb_dd; j++) {
 486                         mb = rxep[j].mbuf;
 487                         qword1 = rte_le_to_cpu_64(\
 488                                 rxdp[j].wb.qword1.status_error_len);
 489                         pkt_len = ((qword1 & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
 490                                 I40E_RXD_QW1_LENGTH_PBUF_SHIFT) - rxq->crc_len;
 491                         mb->data_len = pkt_len;
 492                         mb->pkt_len = pkt_len;
 493                         mb->ol_flags = 0;
 494                         i40e_rxd_to_vlan_tci(mb, &rxdp[j]);
 495                         pkt_flags = i40e_rxd_status_to_pkt_flags(qword1);
 496                         pkt_flags |= i40e_rxd_error_to_pkt_flags(qword1);
 497                         mb->packet_type =
 498                                 ptype_tbl[(uint8_t)((qword1 &
 499                                 I40E_RXD_QW1_PTYPE_MASK) >>
 500                                 I40E_RXD_QW1_PTYPE_SHIFT)];
 501                         if (pkt_flags & PKT_RX_RSS_HASH)
 502                                 mb->hash.rss = rte_le_to_cpu_32(\
 503                                         rxdp[j].wb.qword0.hi_dword.rss);
 504                         if (pkt_flags & PKT_RX_FDIR)
 505                                 pkt_flags |= i40e_rxd_build_fdir(&rxdp[j], mb);
 506
 507 #ifdef RTE_LIBRTE_IEEE1588
 508                         pkt_flags |= i40e_get_iee15888_flags(mb, qword1);
 509 #endif
 510                         mb->ol_flags |= pkt_flags;
 511
 512                 }
 513
 514                 for (j = 0; j < I40E_LOOK_AHEAD; j++)
 515                         rxq->rx_stage[i + j] = rxep[j].mbuf;
 516
 517                 if (nb_dd != I40E_LOOK_AHEAD)
 518                         break;
 519         }
 520
 521         /* Clear software ring entries */
 522         for (i = 0; i < nb_rx; i++)
 523                 rxq->sw_ring[rxq->rx_tail + i].mbuf = NULL;
 524
 525         return nb_rx;
 526 }
 527
 528 static inline uint16_t
 529 i40e_rx_fill_from_stage(struct i40e_rx_queue *rxq,
 530                         struct rte_mbuf **rx_pkts,
 531                         uint16_t nb_pkts)
 532 {
 533         uint16_t i;
 534         struct rte_mbuf **stage = &rxq->rx_stage[rxq->rx_next_avail];
 535
 536         nb_pkts = (uint16_t)RTE_MIN(nb_pkts, rxq->rx_nb_avail);
 537
 538         for (i = 0; i < nb_pkts; i++)
 539                 rx_pkts[i] = stage[i];
 540
 541         rxq->rx_nb_avail = (uint16_t)(rxq->rx_nb_avail - nb_pkts);
 542         rxq->rx_next_avail = (uint16_t)(rxq->rx_next_avail + nb_pkts);
 543
 544         return nb_pkts;
 545 }
 546
 547 static inline int
 548 i40e_rx_alloc_bufs(struct i40e_rx_queue *rxq)
 549 {
 550         volatile union i40e_rx_desc *rxdp;
 551         struct i40e_rx_entry *rxep;
 552         struct rte_mbuf *mb;
 553         uint16_t alloc_idx, i;
 554         uint64_t dma_addr;
 555         int diag;
 556
 557         /* Allocate buffers in bulk */
 558         alloc_idx = (uint16_t)(rxq->rx_free_trigger -
 559                                 (rxq->rx_free_thresh - 1));
 560         rxep = &(rxq->sw_ring[alloc_idx]);
 561         diag = rte_mempool_get_bulk(rxq->mp, (void *)rxep,
 562                                         rxq->rx_free_thresh);
 563         if (unlikely(diag != 0)) {
 564                 PMD_DRV_LOG(ERR, "Failed to get mbufs in bulk");
 565                 return -ENOMEM;
 566         }
 567
 568         rxdp = &rxq->rx_ring[alloc_idx];
 569         for (i = 0; i < rxq->rx_free_thresh; i++) {
 570                 if (likely(i < (rxq->rx_free_thresh - 1)))
 571                         /* Prefetch next mbuf */
 572                         rte_prefetch0(rxep[i + 1].mbuf);
 573
 574                 mb = rxep[i].mbuf;
 575                 rte_mbuf_refcnt_set(mb, 1);
 576                 mb->next = NULL;
 577                 mb->data_off = RTE_PKTMBUF_HEADROOM;
 578                 mb->nb_segs = 1;
 579                 mb->port = rxq->port_id;
 580                 dma_addr = rte_cpu_to_le_64(\
 581                         rte_mbuf_data_iova_default(mb));
 582                 rxdp[i].read.hdr_addr = 0;
 583                 rxdp[i].read.pkt_addr = dma_addr;
 584         }
 585
 586         /* Update rx tail regsiter */
 587         I40E_PCI_REG_WRITE(rxq->qrx_tail, rxq->rx_free_trigger);
 588
 589         rxq->rx_free_trigger =
 590                 (uint16_t)(rxq->rx_free_trigger + rxq->rx_free_thresh);
 591         if (rxq->rx_free_trigger >= rxq->nb_rx_desc)
 592                 rxq->rx_free_trigger = (uint16_t)(rxq->rx_free_thresh - 1);
 593
 594         return 0;
 595 }
 596
 597 static inline uint16_t
 598 rx_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 599 {
 600         struct i40e_rx_queue *rxq = (struct i40e_rx_queue *)rx_queue;
 601         struct rte_eth_dev *dev;
 602         uint16_t nb_rx = 0;
 603
 604         if (!nb_pkts)
 605                 return 0;
 606
 607         if (rxq->rx_nb_avail)
 608                 return i40e_rx_fill_from_stage(rxq, rx_pkts, nb_pkts);
 609
 610         nb_rx = (uint16_t)i40e_rx_scan_hw_ring(rxq);
 611         rxq->rx_next_avail = 0;
 612         rxq->rx_nb_avail = nb_rx;
 613         rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_rx);
 614
 615         if (rxq->rx_tail > rxq->rx_free_trigger) {
 616                 if (i40e_rx_alloc_bufs(rxq) != 0) {
 617                         uint16_t i, j;
 618
 619                         dev = I40E_VSI_TO_ETH_DEV(rxq->vsi);
 620                         dev->data->rx_mbuf_alloc_failed +=
 621                                 rxq->rx_free_thresh;
 622
 623                         rxq->rx_nb_avail = 0;
 624                         rxq->rx_tail = (uint16_t)(rxq->rx_tail - nb_rx);
 625                         for (i = 0, j = rxq->rx_tail; i < nb_rx; i++, j++)
 626                                 rxq->sw_ring[j].mbuf = rxq->rx_stage[i];
 627
 628                         return 0;
 629                 }
 630         }
 631
 632         if (rxq->rx_tail >= rxq->nb_rx_desc)
 633                 rxq->rx_tail = 0;
 634
 635         if (rxq->rx_nb_avail)
 636                 return i40e_rx_fill_from_stage(rxq, rx_pkts, nb_pkts);
 637
 638         return 0;
 639 }
 640
 641 static uint16_t
 642 i40e_recv_pkts_bulk_alloc(void *rx_queue,
 643                           struct rte_mbuf **rx_pkts,
 644                           uint16_t nb_pkts)
 645 {
 646         uint16_t nb_rx = 0, n, count;
 647
 648         if (unlikely(nb_pkts == 0))
 649                 return 0;
 650
 651         if (likely(nb_pkts <= RTE_PMD_I40E_RX_MAX_BURST))
 652                 return rx_recv_pkts(rx_queue, rx_pkts, nb_pkts);
 653
 654         while (nb_pkts) {
 655                 n = RTE_MIN(nb_pkts, RTE_PMD_I40E_RX_MAX_BURST);
 656                 count = rx_recv_pkts(rx_queue, &rx_pkts[nb_rx], n);
 657                 nb_rx = (uint16_t)(nb_rx + count);
 658                 nb_pkts = (uint16_t)(nb_pkts - count);
 659                 if (count < n)
 660                         break;
 661         }
 662
 663         return nb_rx;
 664 }
 665 #else
 666 static uint16_t
 667 i40e_recv_pkts_bulk_alloc(void __rte_unused *rx_queue,
 668                           struct rte_mbuf __rte_unused **rx_pkts,
 669                           uint16_t __rte_unused nb_pkts)
 670 {
 671         return 0;
 672 }
 673 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
 674
 675 uint16_t
 676 i40e_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 677 {
 678         struct i40e_rx_queue *rxq;
 679         volatile union i40e_rx_desc *rx_ring;
 680         volatile union i40e_rx_desc *rxdp;
 681         union i40e_rx_desc rxd;
 682         struct i40e_rx_entry *sw_ring;
 683         struct i40e_rx_entry *rxe;
 684         struct rte_eth_dev *dev;
 685         struct rte_mbuf *rxm;
 686         struct rte_mbuf *nmb;
 687         uint16_t nb_rx;
 688         uint32_t rx_status;
 689         uint64_t qword1;
 690         uint16_t rx_packet_len;
 691         uint16_t rx_id, nb_hold;
 692         uint64_t dma_addr;
 693         uint64_t pkt_flags;
 694         uint32_t *ptype_tbl;
 695
 696         nb_rx = 0;
 697         nb_hold = 0;
 698         rxq = rx_queue;
 699         rx_id = rxq->rx_tail;
 700         rx_ring = rxq->rx_ring;
 701         sw_ring = rxq->sw_ring;
 702         ptype_tbl = rxq->vsi->adapter->ptype_tbl;
 703
 704         while (nb_rx < nb_pkts) {
 705                 rxdp = &rx_ring[rx_id];
 706                 qword1 = rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len);
 707                 rx_status = (qword1 & I40E_RXD_QW1_STATUS_MASK)
 708                                 >> I40E_RXD_QW1_STATUS_SHIFT;
 709
 710                 /* Check the DD bit first */
 711                 if (!(rx_status & (1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
 712                         break;
 713
 714                 nmb = rte_mbuf_raw_alloc(rxq->mp);
 715                 if (unlikely(!nmb)) {
 716                         dev = I40E_VSI_TO_ETH_DEV(rxq->vsi);
 717                         dev->data->rx_mbuf_alloc_failed++;
 718                         break;
 719                 }
 720
 721                 rxd = *rxdp;
 722                 nb_hold++;
 723                 rxe = &sw_ring[rx_id];
 724                 rx_id++;
 725                 if (unlikely(rx_id == rxq->nb_rx_desc))
 726                         rx_id = 0;
 727
 728                 /* Prefetch next mbuf */
 729                 rte_prefetch0(sw_ring[rx_id].mbuf);
 730
 731                 /**
 732                  * When next RX descriptor is on a cache line boundary,
 733                  * prefetch the next 4 RX descriptors and next 8 pointers
 734                  * to mbufs.
 735                  */
 736                 if ((rx_id & 0x3) == 0) {
 737                         rte_prefetch0(&rx_ring[rx_id]);
 738                         rte_prefetch0(&sw_ring[rx_id]);
 739                 }
 740                 rxm = rxe->mbuf;
 741                 rxe->mbuf = nmb;
 742                 dma_addr =
 743                         rte_cpu_to_le_64(rte_mbuf_data_iova_default(nmb));
 744                 rxdp->read.hdr_addr = 0;
 745                 rxdp->read.pkt_addr = dma_addr;
 746
 747                 rx_packet_len = ((qword1 & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
 748                                 I40E_RXD_QW1_LENGTH_PBUF_SHIFT) - rxq->crc_len;
 749
 750                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 751                 rte_prefetch0(RTE_PTR_ADD(rxm->buf_addr, RTE_PKTMBUF_HEADROOM));
 752                 rxm->nb_segs = 1;
 753                 rxm->next = NULL;
 754                 rxm->pkt_len = rx_packet_len;
 755                 rxm->data_len = rx_packet_len;
 756                 rxm->port = rxq->port_id;
 757                 rxm->ol_flags = 0;
 758                 i40e_rxd_to_vlan_tci(rxm, &rxd);
 759                 pkt_flags = i40e_rxd_status_to_pkt_flags(qword1);
 760                 pkt_flags |= i40e_rxd_error_to_pkt_flags(qword1);
 761                 rxm->packet_type =
 762                         ptype_tbl[(uint8_t)((qword1 &
 763                         I40E_RXD_QW1_PTYPE_MASK) >> I40E_RXD_QW1_PTYPE_SHIFT)];
 764                 if (pkt_flags & PKT_RX_RSS_HASH)
 765                         rxm->hash.rss =
 766                                 rte_le_to_cpu_32(rxd.wb.qword0.hi_dword.rss);
 767                 if (pkt_flags & PKT_RX_FDIR)
 768                         pkt_flags |= i40e_rxd_build_fdir(&rxd, rxm);
 769
 770 #ifdef RTE_LIBRTE_IEEE1588
 771                 pkt_flags |= i40e_get_iee15888_flags(rxm, qword1);
 772 #endif
 773                 rxm->ol_flags |= pkt_flags;
 774
 775                 rx_pkts[nb_rx++] = rxm;
 776         }
 777         rxq->rx_tail = rx_id;
 778
 779         /**
 780          * If the number of free RX descriptors is greater than the RX free
 781          * threshold of the queue, advance the receive tail register of queue.
 782          * Update that register with the value of the last processed RX
 783          * descriptor minus 1.
 784          */
 785         nb_hold = (uint16_t)(nb_hold + rxq->nb_rx_hold);
 786         if (nb_hold > rxq->rx_free_thresh) {
 787                 rx_id = (uint16_t) ((rx_id == 0) ?
 788                         (rxq->nb_rx_desc - 1) : (rx_id - 1));
 789                 I40E_PCI_REG_WC_WRITE(rxq->qrx_tail, rx_id);
 790                 nb_hold = 0;
 791         }
 792         rxq->nb_rx_hold = nb_hold;
 793
 794         return nb_rx;
 795 }
 796
 797 uint16_t
 798 i40e_recv_scattered_pkts(void *rx_queue,
 799                          struct rte_mbuf **rx_pkts,
 800                          uint16_t nb_pkts)
 801 {
 802         struct i40e_rx_queue *rxq = rx_queue;
 803         volatile union i40e_rx_desc *rx_ring = rxq->rx_ring;
 804         volatile union i40e_rx_desc *rxdp;
 805         union i40e_rx_desc rxd;
 806         struct i40e_rx_entry *sw_ring = rxq->sw_ring;
 807         struct i40e_rx_entry *rxe;
 808         struct rte_mbuf *first_seg = rxq->pkt_first_seg;
 809         struct rte_mbuf *last_seg = rxq->pkt_last_seg;
 810         struct rte_mbuf *nmb, *rxm;
 811         uint16_t rx_id = rxq->rx_tail;
 812         uint16_t nb_rx = 0, nb_hold = 0, rx_packet_len;
 813         struct rte_eth_dev *dev;
 814         uint32_t rx_status;
 815         uint64_t qword1;
 816         uint64_t dma_addr;
 817         uint64_t pkt_flags;
 818         uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
 819
 820         while (nb_rx < nb_pkts) {
 821                 rxdp = &rx_ring[rx_id];
 822                 qword1 = rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len);
 823                 rx_status = (qword1 & I40E_RXD_QW1_STATUS_MASK) >>
 824                                         I40E_RXD_QW1_STATUS_SHIFT;
 825
 826                 /* Check the DD bit */
 827                 if (!(rx_status & (1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
 828                         break;
 829
 830                 nmb = rte_mbuf_raw_alloc(rxq->mp);
 831                 if (unlikely(!nmb)) {
 832                         dev = I40E_VSI_TO_ETH_DEV(rxq->vsi);
 833                         dev->data->rx_mbuf_alloc_failed++;
 834                         break;
 835                 }
 836
 837                 rxd = *rxdp;
 838                 nb_hold++;
 839                 rxe = &sw_ring[rx_id];
 840                 rx_id++;
 841                 if (rx_id == rxq->nb_rx_desc)
 842                         rx_id = 0;
 843
 844                 /* Prefetch next mbuf */
 845                 rte_prefetch0(sw_ring[rx_id].mbuf);
 846
 847                 /**
 848                  * When next RX descriptor is on a cache line boundary,
 849                  * prefetch the next 4 RX descriptors and next 8 pointers
 850                  * to mbufs.
 851                  */
 852                 if ((rx_id & 0x3) == 0) {
 853                         rte_prefetch0(&rx_ring[rx_id]);
 854                         rte_prefetch0(&sw_ring[rx_id]);
 855                 }
 856
 857                 rxm = rxe->mbuf;
 858                 rxe->mbuf = nmb;
 859                 dma_addr =
 860                         rte_cpu_to_le_64(rte_mbuf_data_iova_default(nmb));
 861
 862                 /* Set data buffer address and data length of the mbuf */
 863                 rxdp->read.hdr_addr = 0;
 864                 rxdp->read.pkt_addr = dma_addr;
 865                 rx_packet_len = (qword1 & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
 866                                         I40E_RXD_QW1_LENGTH_PBUF_SHIFT;
 867                 rxm->data_len = rx_packet_len;
 868                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 869
 870                 /**
 871                  * If this is the first buffer of the received packet, set the
 872                  * pointer to the first mbuf of the packet and initialize its
 873                  * context. Otherwise, update the total length and the number
 874                  * of segments of the current scattered packet, and update the
 875                  * pointer to the last mbuf of the current packet.
 876                  */
 877                 if (!first_seg) {
 878                         first_seg = rxm;
 879                         first_seg->nb_segs = 1;
 880                         first_seg->pkt_len = rx_packet_len;
 881                 } else {
 882                         first_seg->pkt_len =
 883                                 (uint16_t)(first_seg->pkt_len +
 884                                                 rx_packet_len);
 885                         first_seg->nb_segs++;
 886                         last_seg->next = rxm;
 887                 }
 888
 889                 /**
 890                  * If this is not the last buffer of the received packet,
 891                  * update the pointer to the last mbuf of the current scattered
 892                  * packet and continue to parse the RX ring.
 893                  */
 894                 if (!(rx_status & (1 << I40E_RX_DESC_STATUS_EOF_SHIFT))) {
 895                         last_seg = rxm;
 896                         continue;
 897                 }
 898
 899                 /**
 900                  * This is the last buffer of the received packet. If the CRC
 901                  * is not stripped by the hardware:
 902                  *  - Subtract the CRC length from the total packet length.
 903                  *  - If the last buffer only contains the whole CRC or a part
 904                  *  of it, free the mbuf associated to the last buffer. If part
 905                  *  of the CRC is also contained in the previous mbuf, subtract
 906                  *  the length of that CRC part from the data length of the
 907                  *  previous mbuf.
 908                  */
 909                 rxm->next = NULL;
 910                 if (unlikely(rxq->crc_len > 0)) {
 911                         first_seg->pkt_len -= RTE_ETHER_CRC_LEN;
 912                         if (rx_packet_len <= RTE_ETHER_CRC_LEN) {
 913                                 rte_pktmbuf_free_seg(rxm);
 914                                 first_seg->nb_segs--;
 915                                 last_seg->data_len =
 916                                         (uint16_t)(last_seg->data_len -
 917                                         (RTE_ETHER_CRC_LEN - rx_packet_len));
 918                                 last_seg->next = NULL;
 919                         } else
 920                                 rxm->data_len = (uint16_t)(rx_packet_len -
 921                                                         RTE_ETHER_CRC_LEN);
 922                 }
 923
 924                 first_seg->port = rxq->port_id;
 925                 first_seg->ol_flags = 0;
 926                 i40e_rxd_to_vlan_tci(first_seg, &rxd);
 927                 pkt_flags = i40e_rxd_status_to_pkt_flags(qword1);
 928                 pkt_flags |= i40e_rxd_error_to_pkt_flags(qword1);
 929                 first_seg->packet_type =
 930                         ptype_tbl[(uint8_t)((qword1 &
 931                         I40E_RXD_QW1_PTYPE_MASK) >> I40E_RXD_QW1_PTYPE_SHIFT)];
 932                 if (pkt_flags & PKT_RX_RSS_HASH)
 933                         first_seg->hash.rss =
 934                                 rte_le_to_cpu_32(rxd.wb.qword0.hi_dword.rss);
 935                 if (pkt_flags & PKT_RX_FDIR)
 936                         pkt_flags |= i40e_rxd_build_fdir(&rxd, first_seg);
 937
 938 #ifdef RTE_LIBRTE_IEEE1588
 939                 pkt_flags |= i40e_get_iee15888_flags(first_seg, qword1);
 940 #endif
 941                 first_seg->ol_flags |= pkt_flags;
 942
 943                 /* Prefetch data of first segment, if configured to do so. */
 944                 rte_prefetch0(RTE_PTR_ADD(first_seg->buf_addr,
 945                         first_seg->data_off));
 946                 rx_pkts[nb_rx++] = first_seg;
 947                 first_seg = NULL;
 948         }
 949
 950         /* Record index of the next RX descriptor to probe. */
 951         rxq->rx_tail = rx_id;
 952         rxq->pkt_first_seg = first_seg;
 953         rxq->pkt_last_seg = last_seg;
 954
 955         /**
 956          * If the number of free RX descriptors is greater than the RX free
 957          * threshold of the queue, advance the Receive Descriptor Tail (RDT)
 958          * register. Update the RDT with the value of the last processed RX
 959          * descriptor minus 1, to guarantee that the RDT register is never
 960          * equal to the RDH register, which creates a "full" ring situtation
 961          * from the hardware point of view.
 962          */
 963         nb_hold = (uint16_t)(nb_hold + rxq->nb_rx_hold);
 964         if (nb_hold > rxq->rx_free_thresh) {
 965                 rx_id = (uint16_t)(rx_id == 0 ?
 966                         (rxq->nb_rx_desc - 1) : (rx_id - 1));
 967                 I40E_PCI_REG_WC_WRITE(rxq->qrx_tail, rx_id);
 968                 nb_hold = 0;
 969         }
 970         rxq->nb_rx_hold = nb_hold;
 971
 972         return nb_rx;
 973 }
 974
 975 /* Check if the context descriptor is needed for TX offloading */
 976 static inline uint16_t
 977 i40e_calc_context_desc(uint64_t flags)
 978 {
 979         static uint64_t mask = PKT_TX_OUTER_IP_CKSUM |
 980                 PKT_TX_TCP_SEG |
 981                 PKT_TX_QINQ_PKT |
 982                 PKT_TX_TUNNEL_MASK;
 983
 984 #ifdef RTE_LIBRTE_IEEE1588
 985         mask |= PKT_TX_IEEE1588_TMST;
 986 #endif
 987
 988         return (flags & mask) ? 1 : 0;
 989 }
 990
 991 /* set i40e TSO context descriptor */
 992 static inline uint64_t
 993 i40e_set_tso_ctx(struct rte_mbuf *mbuf, union i40e_tx_offload tx_offload)
 994 {
 995         uint64_t ctx_desc = 0;
 996         uint32_t cd_cmd, hdr_len, cd_tso_len;
 997
 998         if (!tx_offload.l4_len) {
 999                 PMD_DRV_LOG(DEBUG, "L4 length set to 0");
1000                 return ctx_desc;
1001         }
1002
1003         hdr_len = tx_offload.l2_len + tx_offload.l3_len + tx_offload.l4_len;
1004         hdr_len += (mbuf->ol_flags & PKT_TX_TUNNEL_MASK) ?
1005                    tx_offload.outer_l2_len + tx_offload.outer_l3_len : 0;
1006
1007         cd_cmd = I40E_TX_CTX_DESC_TSO;
1008         cd_tso_len = mbuf->pkt_len - hdr_len;
1009         ctx_desc |= ((uint64_t)cd_cmd << I40E_TXD_CTX_QW1_CMD_SHIFT) |
1010                 ((uint64_t)cd_tso_len <<
1011                  I40E_TXD_CTX_QW1_TSO_LEN_SHIFT) |
1012                 ((uint64_t)mbuf->tso_segsz <<
1013                  I40E_TXD_CTX_QW1_MSS_SHIFT);
1014
1015         return ctx_desc;
1016 }
1017
1018 /* HW requires that Tx buffer size ranges from 1B up to (16K-1)B. */
1019 #define I40E_MAX_DATA_PER_TXD \
1020         (I40E_TXD_QW1_TX_BUF_SZ_MASK >> I40E_TXD_QW1_TX_BUF_SZ_SHIFT)
1021 /* Calculate the number of TX descriptors needed for each pkt */
1022 static inline uint16_t
1023 i40e_calc_pkt_desc(struct rte_mbuf *tx_pkt)
1024 {
1025         struct rte_mbuf *txd = tx_pkt;
1026         uint16_t count = 0;
1027
1028         while (txd != NULL) {
1029                 count += DIV_ROUND_UP(txd->data_len, I40E_MAX_DATA_PER_TXD);
1030                 txd = txd->next;
1031         }
1032
1033         return count;
1034 }
1035
1036 uint16_t
1037 i40e_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1038 {
1039         struct i40e_tx_queue *txq;
1040         struct i40e_tx_entry *sw_ring;
1041         struct i40e_tx_entry *txe, *txn;
1042         volatile struct i40e_tx_desc *txd;
1043         volatile struct i40e_tx_desc *txr;
1044         struct rte_mbuf *tx_pkt;
1045         struct rte_mbuf *m_seg;
1046         uint32_t cd_tunneling_params;
1047         uint16_t tx_id;
1048         uint16_t nb_tx;
1049         uint32_t td_cmd;
1050         uint32_t td_offset;
1051         uint32_t td_tag;
1052         uint64_t ol_flags;
1053         uint16_t nb_used;
1054         uint16_t nb_ctx;
1055         uint16_t tx_last;
1056         uint16_t slen;
1057         uint64_t buf_dma_addr;
1058         union i40e_tx_offload tx_offload = {0};
1059
1060         txq = tx_queue;
1061         sw_ring = txq->sw_ring;
1062         txr = txq->tx_ring;
1063         tx_id = txq->tx_tail;
1064         txe = &sw_ring[tx_id];
1065
1066         /* Check if the descriptor ring needs to be cleaned. */
1067         if (txq->nb_tx_free < txq->tx_free_thresh)
1068                 (void)i40e_xmit_cleanup(txq);
1069
1070         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1071                 td_cmd = 0;
1072                 td_tag = 0;
1073                 td_offset = 0;
1074
1075                 tx_pkt = *tx_pkts++;
1076                 RTE_MBUF_PREFETCH_TO_FREE(txe->mbuf);
1077
1078                 ol_flags = tx_pkt->ol_flags;
1079                 tx_offload.l2_len = tx_pkt->l2_len;
1080                 tx_offload.l3_len = tx_pkt->l3_len;
1081                 tx_offload.outer_l2_len = tx_pkt->outer_l2_len;
1082                 tx_offload.outer_l3_len = tx_pkt->outer_l3_len;
1083                 tx_offload.l4_len = tx_pkt->l4_len;
1084                 tx_offload.tso_segsz = tx_pkt->tso_segsz;
1085
1086                 /* Calculate the number of context descriptors needed. */
1087                 nb_ctx = i40e_calc_context_desc(ol_flags);
1088
1089                 /**
1090                  * The number of descriptors that must be allocated for
1091                  * a packet equals to the number of the segments of that
1092                  * packet plus 1 context descriptor if needed.
1093                  * Recalculate the needed tx descs when TSO enabled in case
1094                  * the mbuf data size exceeds max data size that hw allows
1095                  * per tx desc.
1096                  */
1097                 if (ol_flags & PKT_TX_TCP_SEG)
1098                         nb_used = (uint16_t)(i40e_calc_pkt_desc(tx_pkt) +
1099                                              nb_ctx);
1100                 else
1101                         nb_used = (uint16_t)(tx_pkt->nb_segs + nb_ctx);
1102                 tx_last = (uint16_t)(tx_id + nb_used - 1);
1103
1104                 /* Circular ring */
1105                 if (tx_last >= txq->nb_tx_desc)
1106                         tx_last = (uint16_t)(tx_last - txq->nb_tx_desc);
1107
1108                 if (nb_used > txq->nb_tx_free) {
1109                         if (i40e_xmit_cleanup(txq) != 0) {
1110                                 if (nb_tx == 0)
1111                                         return 0;
1112                                 goto end_of_tx;
1113                         }
1114                         if (unlikely(nb_used > txq->tx_rs_thresh)) {
1115                                 while (nb_used > txq->nb_tx_free) {
1116                                         if (i40e_xmit_cleanup(txq) != 0) {
1117                                                 if (nb_tx == 0)
1118                                                         return 0;
1119                                                 goto end_of_tx;
1120                                         }
1121                                 }
1122                         }
1123                 }
1124
1125                 /* Descriptor based VLAN insertion */
1126                 if (ol_flags & (PKT_TX_VLAN_PKT | PKT_TX_QINQ_PKT)) {
1127                         td_cmd |= I40E_TX_DESC_CMD_IL2TAG1;
1128                         td_tag = tx_pkt->vlan_tci;
1129                 }
1130
1131                 /* Always enable CRC offload insertion */
1132                 td_cmd |= I40E_TX_DESC_CMD_ICRC;
1133
1134                 /* Fill in tunneling parameters if necessary */
1135                 cd_tunneling_params = 0;
1136                 if (ol_flags & PKT_TX_TUNNEL_MASK)
1137                         i40e_parse_tunneling_params(ol_flags, tx_offload,
1138                                                     &cd_tunneling_params);
1139                 /* Enable checksum offloading */
1140                 if (ol_flags & I40E_TX_CKSUM_OFFLOAD_MASK)
1141                         i40e_txd_enable_checksum(ol_flags, &td_cmd,
1142                                                  &td_offset, tx_offload);
1143
1144                 if (nb_ctx) {
1145                         /* Setup TX context descriptor if required */
1146                         volatile struct i40e_tx_context_desc *ctx_txd =
1147                                 (volatile struct i40e_tx_context_desc *)\
1148                                                         &txr[tx_id];
1149                         uint16_t cd_l2tag2 = 0;
1150                         uint64_t cd_type_cmd_tso_mss =
1151                                 I40E_TX_DESC_DTYPE_CONTEXT;
1152
1153                         txn = &sw_ring[txe->next_id];
1154                         RTE_MBUF_PREFETCH_TO_FREE(txn->mbuf);
1155                         if (txe->mbuf != NULL) {
1156                                 rte_pktmbuf_free_seg(txe->mbuf);
1157                                 txe->mbuf = NULL;
1158                         }
1159
1160                         /* TSO enabled means no timestamp */
1161                         if (ol_flags & PKT_TX_TCP_SEG)
1162                                 cd_type_cmd_tso_mss |=
1163                                         i40e_set_tso_ctx(tx_pkt, tx_offload);
1164                         else {
1165 #ifdef RTE_LIBRTE_IEEE1588
1166                                 if (ol_flags & PKT_TX_IEEE1588_TMST)
1167                                         cd_type_cmd_tso_mss |=
1168                                                 ((uint64_t)I40E_TX_CTX_DESC_TSYN <<
1169                                                  I40E_TXD_CTX_QW1_CMD_SHIFT);
1170 #endif
1171                         }
1172
1173                         ctx_txd->tunneling_params =
1174                                 rte_cpu_to_le_32(cd_tunneling_params);
1175                         if (ol_flags & PKT_TX_QINQ_PKT) {
1176                                 cd_l2tag2 = tx_pkt->vlan_tci_outer;
1177                                 cd_type_cmd_tso_mss |=
1178                                         ((uint64_t)I40E_TX_CTX_DESC_IL2TAG2 <<
1179                                                 I40E_TXD_CTX_QW1_CMD_SHIFT);
1180                         }
1181                         ctx_txd->l2tag2 = rte_cpu_to_le_16(cd_l2tag2);
1182                         ctx_txd->type_cmd_tso_mss =
1183                                 rte_cpu_to_le_64(cd_type_cmd_tso_mss);
1184
1185                         PMD_TX_LOG(DEBUG, "mbuf: %p, TCD[%u]:\n"
1186                                 "tunneling_params: %#x;\n"
1187                                 "l2tag2: %#hx;\n"
1188                                 "rsvd: %#hx;\n"
1189                                 "type_cmd_tso_mss: %#"PRIx64";\n",
1190                                 tx_pkt, tx_id,
1191                                 ctx_txd->tunneling_params,
1192                                 ctx_txd->l2tag2,
1193                                 ctx_txd->rsvd,
1194                                 ctx_txd->type_cmd_tso_mss);
1195
1196                         txe->last_id = tx_last;
1197                         tx_id = txe->next_id;
1198                         txe = txn;
1199                 }
1200
1201                 m_seg = tx_pkt;
1202                 do {
1203                         txd = &txr[tx_id];
1204                         txn = &sw_ring[txe->next_id];
1205
1206                         if (txe->mbuf)
1207                                 rte_pktmbuf_free_seg(txe->mbuf);
1208                         txe->mbuf = m_seg;
1209
1210                         /* Setup TX Descriptor */
1211                         slen = m_seg->data_len;
1212                         buf_dma_addr = rte_mbuf_data_iova(m_seg);
1213
1214                         while ((ol_flags & PKT_TX_TCP_SEG) &&
1215                                 unlikely(slen > I40E_MAX_DATA_PER_TXD)) {
1216                                 txd->buffer_addr =
1217                                         rte_cpu_to_le_64(buf_dma_addr);
1218                                 txd->cmd_type_offset_bsz =
1219                                         i40e_build_ctob(td_cmd,
1220                                         td_offset, I40E_MAX_DATA_PER_TXD,
1221                                         td_tag);
1222
1223                                 buf_dma_addr += I40E_MAX_DATA_PER_TXD;
1224                                 slen -= I40E_MAX_DATA_PER_TXD;
1225
1226                                 txe->last_id = tx_last;
1227                                 tx_id = txe->next_id;
1228                                 txe = txn;
1229                                 txd = &txr[tx_id];
1230                                 txn = &sw_ring[txe->next_id];
1231                         }
1232                         PMD_TX_LOG(DEBUG, "mbuf: %p, TDD[%u]:\n"
1233                                 "buf_dma_addr: %#"PRIx64";\n"
1234                                 "td_cmd: %#x;\n"
1235                                 "td_offset: %#x;\n"
1236                                 "td_len: %u;\n"
1237                                 "td_tag: %#x;\n",
1238                                 tx_pkt, tx_id, buf_dma_addr,
1239                                 td_cmd, td_offset, slen, td_tag);
1240
1241                         txd->buffer_addr = rte_cpu_to_le_64(buf_dma_addr);
1242                         txd->cmd_type_offset_bsz = i40e_build_ctob(td_cmd,
1243                                                 td_offset, slen, td_tag);
1244                         txe->last_id = tx_last;
1245                         tx_id = txe->next_id;
1246                         txe = txn;
1247                         m_seg = m_seg->next;
1248                 } while (m_seg != NULL);
1249
1250                 /* The last packet data descriptor needs End Of Packet (EOP) */
1251                 td_cmd |= I40E_TX_DESC_CMD_EOP;
1252                 txq->nb_tx_used = (uint16_t)(txq->nb_tx_used + nb_used);
1253                 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_used);
1254
1255                 if (txq->nb_tx_used >= txq->tx_rs_thresh) {
1256                         PMD_TX_LOG(DEBUG,
1257                                    "Setting RS bit on TXD id="
1258                                    "%4u (port=%d queue=%d)",
1259                                    tx_last, txq->port_id, txq->queue_id);
1260
1261                         td_cmd |= I40E_TX_DESC_CMD_RS;
1262
1263                         /* Update txq RS bit counters */
1264                         txq->nb_tx_used = 0;
1265                 }
1266
1267                 txd->cmd_type_offset_bsz |=
1268                         rte_cpu_to_le_64(((uint64_t)td_cmd) <<
1269                                         I40E_TXD_QW1_CMD_SHIFT);
1270         }
1271
1272 end_of_tx:
1273         PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u tx_tail=%u nb_tx=%u",
1274                    (unsigned) txq->port_id, (unsigned) txq->queue_id,
1275                    (unsigned) tx_id, (unsigned) nb_tx);
1276
1277         rte_io_wmb();
1278         I40E_PCI_REG_WC_WRITE_RELAXED(txq->qtx_tail, tx_id);
1279         txq->tx_tail = tx_id;
1280
1281         return nb_tx;
1282 }
1283
1284 static __rte_always_inline int
1285 i40e_tx_free_bufs(struct i40e_tx_queue *txq)
1286 {
1287         struct i40e_tx_entry *txep;
1288         uint16_t i;
1289
1290         if ((txq->tx_ring[txq->tx_next_dd].cmd_type_offset_bsz &
1291                         rte_cpu_to_le_64(I40E_TXD_QW1_DTYPE_MASK)) !=
1292                         rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE))
1293                 return 0;
1294
1295         txep = &(txq->sw_ring[txq->tx_next_dd - (txq->tx_rs_thresh - 1)]);
1296
1297         for (i = 0; i < txq->tx_rs_thresh; i++)
1298                 rte_prefetch0((txep + i)->mbuf);
1299
1300         if (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE) {
1301                 for (i = 0; i < txq->tx_rs_thresh; ++i, ++txep) {
1302                         rte_mempool_put(txep->mbuf->pool, txep->mbuf);
1303                         txep->mbuf = NULL;
1304                 }
1305         } else {
1306                 for (i = 0; i < txq->tx_rs_thresh; ++i, ++txep) {
1307                         rte_pktmbuf_free_seg(txep->mbuf);
1308                         txep->mbuf = NULL;
1309                 }
1310         }
1311
1312         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + txq->tx_rs_thresh);
1313         txq->tx_next_dd = (uint16_t)(txq->tx_next_dd + txq->tx_rs_thresh);
1314         if (txq->tx_next_dd >= txq->nb_tx_desc)
1315                 txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
1316
1317         return txq->tx_rs_thresh;
1318 }
1319
1320 /* Populate 4 descriptors with data from 4 mbufs */
1321 static inline void
1322 tx4(volatile struct i40e_tx_desc *txdp, struct rte_mbuf **pkts)
1323 {
1324         uint64_t dma_addr;
1325         uint32_t i;
1326
1327         for (i = 0; i < 4; i++, txdp++, pkts++) {
1328                 dma_addr = rte_mbuf_data_iova(*pkts);
1329                 txdp->buffer_addr = rte_cpu_to_le_64(dma_addr);
1330                 txdp->cmd_type_offset_bsz =
1331                         i40e_build_ctob((uint32_t)I40E_TD_CMD, 0,
1332                                         (*pkts)->data_len, 0);
1333         }
1334 }
1335
1336 /* Populate 1 descriptor with data from 1 mbuf */
1337 static inline void
1338 tx1(volatile struct i40e_tx_desc *txdp, struct rte_mbuf **pkts)
1339 {
1340         uint64_t dma_addr;
1341
1342         dma_addr = rte_mbuf_data_iova(*pkts);
1343         txdp->buffer_addr = rte_cpu_to_le_64(dma_addr);
1344         txdp->cmd_type_offset_bsz =
1345                 i40e_build_ctob((uint32_t)I40E_TD_CMD, 0,
1346                                 (*pkts)->data_len, 0);
1347 }
1348
1349 /* Fill hardware descriptor ring with mbuf data */
1350 static inline void
1351 i40e_tx_fill_hw_ring(struct i40e_tx_queue *txq,
1352                      struct rte_mbuf **pkts,
1353                      uint16_t nb_pkts)
1354 {
1355         volatile struct i40e_tx_desc *txdp = &(txq->tx_ring[txq->tx_tail]);
1356         struct i40e_tx_entry *txep = &(txq->sw_ring[txq->tx_tail]);
1357         const int N_PER_LOOP = 4;
1358         const int N_PER_LOOP_MASK = N_PER_LOOP - 1;
1359         int mainpart, leftover;
1360         int i, j;
1361
1362         mainpart = (nb_pkts & ((uint32_t) ~N_PER_LOOP_MASK));
1363         leftover = (nb_pkts & ((uint32_t)  N_PER_LOOP_MASK));
1364         for (i = 0; i < mainpart; i += N_PER_LOOP) {
1365                 for (j = 0; j < N_PER_LOOP; ++j) {
1366                         (txep + i + j)->mbuf = *(pkts + i + j);
1367                 }
1368                 tx4(txdp + i, pkts + i);
1369         }
1370         if (unlikely(leftover > 0)) {
1371                 for (i = 0; i < leftover; ++i) {
1372                         (txep + mainpart + i)->mbuf = *(pkts + mainpart + i);
1373                         tx1(txdp + mainpart + i, pkts + mainpart + i);
1374                 }
1375         }
1376 }
1377
1378 static inline uint16_t
1379 tx_xmit_pkts(struct i40e_tx_queue *txq,
1380              struct rte_mbuf **tx_pkts,
1381              uint16_t nb_pkts)
1382 {
1383         volatile struct i40e_tx_desc *txr = txq->tx_ring;
1384         uint16_t n = 0;
1385
1386         /**
1387          * Begin scanning the H/W ring for done descriptors when the number
1388          * of available descriptors drops below tx_free_thresh. For each done
1389          * descriptor, free the associated buffer.
1390          */
1391         if (txq->nb_tx_free < txq->tx_free_thresh)
1392                 i40e_tx_free_bufs(txq);
1393
1394         /* Use available descriptor only */
1395         nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
1396         if (unlikely(!nb_pkts))
1397                 return 0;
1398
1399         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
1400         if ((txq->tx_tail + nb_pkts) > txq->nb_tx_desc) {
1401                 n = (uint16_t)(txq->nb_tx_desc - txq->tx_tail);
1402                 i40e_tx_fill_hw_ring(txq, tx_pkts, n);
1403                 txr[txq->tx_next_rs].cmd_type_offset_bsz |=
1404                         rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
1405                                                 I40E_TXD_QW1_CMD_SHIFT);
1406                 txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
1407                 txq->tx_tail = 0;
1408         }
1409
1410         /* Fill hardware descriptor ring with mbuf data */
1411         i40e_tx_fill_hw_ring(txq, tx_pkts + n, (uint16_t)(nb_pkts - n));
1412         txq->tx_tail = (uint16_t)(txq->tx_tail + (nb_pkts - n));
1413
1414         /* Determin if RS bit needs to be set */
1415         if (txq->tx_tail > txq->tx_next_rs) {
1416                 txr[txq->tx_next_rs].cmd_type_offset_bsz |=
1417                         rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
1418                                                 I40E_TXD_QW1_CMD_SHIFT);
1419                 txq->tx_next_rs =
1420                         (uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
1421                 if (txq->tx_next_rs >= txq->nb_tx_desc)
1422                         txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
1423         }
1424
1425         if (txq->tx_tail >= txq->nb_tx_desc)
1426                 txq->tx_tail = 0;
1427
1428         /* Update the tx tail register */
1429         I40E_PCI_REG_WC_WRITE(txq->qtx_tail, txq->tx_tail);
1430
1431         return nb_pkts;
1432 }
1433
1434 static uint16_t
1435 i40e_xmit_pkts_simple(void *tx_queue,
1436                       struct rte_mbuf **tx_pkts,
1437                       uint16_t nb_pkts)
1438 {
1439         uint16_t nb_tx = 0;
1440
1441         if (likely(nb_pkts <= I40E_TX_MAX_BURST))
1442                 return tx_xmit_pkts((struct i40e_tx_queue *)tx_queue,
1443                                                 tx_pkts, nb_pkts);
1444
1445         while (nb_pkts) {
1446                 uint16_t ret, num = (uint16_t)RTE_MIN(nb_pkts,
1447                                                 I40E_TX_MAX_BURST);
1448
1449                 ret = tx_xmit_pkts((struct i40e_tx_queue *)tx_queue,
1450                                                 &tx_pkts[nb_tx], num);
1451                 nb_tx = (uint16_t)(nb_tx + ret);
1452                 nb_pkts = (uint16_t)(nb_pkts - ret);
1453                 if (ret < num)
1454                         break;
1455         }
1456
1457         return nb_tx;
1458 }
1459
1460 static uint16_t
1461 i40e_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
1462                    uint16_t nb_pkts)
1463 {
1464         uint16_t nb_tx = 0;
1465         struct i40e_tx_queue *txq = (struct i40e_tx_queue *)tx_queue;
1466
1467         while (nb_pkts) {
1468                 uint16_t ret, num;
1469
1470                 num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh);
1471                 ret = i40e_xmit_fixed_burst_vec(tx_queue, &tx_pkts[nb_tx],
1472                                                 num);
1473                 nb_tx += ret;
1474                 nb_pkts -= ret;
1475                 if (ret < num)
1476                         break;
1477         }
1478
1479         return nb_tx;
1480 }
1481
1482 /*********************************************************************
1483  *
1484  *  TX prep functions
1485  *
1486  **********************************************************************/
1487 uint16_t
1488 i40e_prep_pkts(__rte_unused void *tx_queue, struct rte_mbuf **tx_pkts,
1489                 uint16_t nb_pkts)
1490 {
1491         int i, ret;
1492         uint64_t ol_flags;
1493         struct rte_mbuf *m;
1494
1495         for (i = 0; i < nb_pkts; i++) {
1496                 m = tx_pkts[i];
1497                 ol_flags = m->ol_flags;
1498
1499                 /* Check for m->nb_segs to not exceed the limits. */
1500                 if (!(ol_flags & PKT_TX_TCP_SEG)) {
1501                         if (m->nb_segs > I40E_TX_MAX_MTU_SEG ||
1502                             m->pkt_len > I40E_FRAME_SIZE_MAX) {
1503                                 rte_errno = EINVAL;
1504                                 return i;
1505                         }
1506                 } else if (m->nb_segs > I40E_TX_MAX_SEG ||
1507                            m->tso_segsz < I40E_MIN_TSO_MSS ||
1508                            m->tso_segsz > I40E_MAX_TSO_MSS ||
1509                            m->pkt_len > I40E_TSO_FRAME_SIZE_MAX) {
1510                         /* MSS outside the range (256B - 9674B) are considered
1511                          * malicious
1512                          */
1513                         rte_errno = EINVAL;
1514                         return i;
1515                 }
1516
1517                 if (ol_flags & I40E_TX_OFFLOAD_NOTSUP_MASK) {
1518                         rte_errno = ENOTSUP;
1519                         return i;
1520                 }
1521
1522                 /* check the size of packet */
1523                 if (m->pkt_len < I40E_TX_MIN_PKT_LEN) {
1524                         rte_errno = EINVAL;
1525                         return i;
1526                 }
1527
1528 #ifdef RTE_ETHDEV_DEBUG_TX
1529                 ret = rte_validate_tx_offload(m);
1530                 if (ret != 0) {
1531                         rte_errno = -ret;
1532                         return i;
1533                 }
1534 #endif
1535                 ret = rte_net_intel_cksum_prepare(m);
1536                 if (ret != 0) {
1537                         rte_errno = -ret;
1538                         return i;
1539                 }
1540         }
1541         return i;
1542 }
1543
1544 /*
1545  * Find the VSI the queue belongs to. 'queue_idx' is the queue index
1546  * application used, which assume having sequential ones. But from driver's
1547  * perspective, it's different. For example, q0 belongs to FDIR VSI, q1-q64
1548  * to MAIN VSI, , q65-96 to SRIOV VSIs, q97-128 to VMDQ VSIs. For application
1549  * running on host, q1-64 and q97-128 can be used, total 96 queues. They can
1550  * use queue_idx from 0 to 95 to access queues, while real queue would be
1551  * different. This function will do a queue mapping to find VSI the queue
1552  * belongs to.
1553  */
1554 static struct i40e_vsi*
1555 i40e_pf_get_vsi_by_qindex(struct i40e_pf *pf, uint16_t queue_idx)
1556 {
1557         /* the queue in MAIN VSI range */
1558         if (queue_idx < pf->main_vsi->nb_qps)
1559                 return pf->main_vsi;
1560
1561         queue_idx -= pf->main_vsi->nb_qps;
1562
1563         /* queue_idx is greater than VMDQ VSIs range */
1564         if (queue_idx > pf->nb_cfg_vmdq_vsi * pf->vmdq_nb_qps - 1) {
1565                 PMD_INIT_LOG(ERR, "queue_idx out of range. VMDQ configured?");
1566                 return NULL;
1567         }
1568
1569         return pf->vmdq[queue_idx / pf->vmdq_nb_qps].vsi;
1570 }
1571
1572 static uint16_t
1573 i40e_get_queue_offset_by_qindex(struct i40e_pf *pf, uint16_t queue_idx)
1574 {
1575         /* the queue in MAIN VSI range */
1576         if (queue_idx < pf->main_vsi->nb_qps)
1577                 return queue_idx;
1578
1579         /* It's VMDQ queues */
1580         queue_idx -= pf->main_vsi->nb_qps;
1581
1582         if (pf->nb_cfg_vmdq_vsi)
1583                 return queue_idx % pf->vmdq_nb_qps;
1584         else {
1585                 PMD_INIT_LOG(ERR, "Fail to get queue offset");
1586                 return (uint16_t)(-1);
1587         }
1588 }
1589
1590 int
1591 i40e_dev_rx_queue_start(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1592 {
1593         struct i40e_rx_queue *rxq;
1594         int err;
1595         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1596
1597         PMD_INIT_FUNC_TRACE();
1598
1599         rxq = dev->data->rx_queues[rx_queue_id];
1600         if (!rxq || !rxq->q_set) {
1601                 PMD_DRV_LOG(ERR, "RX queue %u not available or setup",
1602                             rx_queue_id);
1603                 return -EINVAL;
1604         }
1605
1606         if (rxq->rx_deferred_start)
1607                 PMD_DRV_LOG(WARNING, "RX queue %u is deferrd start",
1608                             rx_queue_id);
1609
1610         err = i40e_alloc_rx_queue_mbufs(rxq);
1611         if (err) {
1612                 PMD_DRV_LOG(ERR, "Failed to allocate RX queue mbuf");
1613                 return err;
1614         }
1615
1616         /* Init the RX tail regieter. */
1617         I40E_PCI_REG_WRITE(rxq->qrx_tail, rxq->nb_rx_desc - 1);
1618
1619         err = i40e_switch_rx_queue(hw, rxq->reg_idx, TRUE);
1620         if (err) {
1621                 PMD_DRV_LOG(ERR, "Failed to switch RX queue %u on",
1622                             rx_queue_id);
1623
1624                 i40e_rx_queue_release_mbufs(rxq);
1625                 i40e_reset_rx_queue(rxq);
1626                 return err;
1627         }
1628         dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
1629
1630         return 0;
1631 }
1632
1633 int
1634 i40e_dev_rx_queue_stop(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1635 {
1636         struct i40e_rx_queue *rxq;
1637         int err;
1638         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1639
1640         rxq = dev->data->rx_queues[rx_queue_id];
1641         if (!rxq || !rxq->q_set) {
1642                 PMD_DRV_LOG(ERR, "RX queue %u not available or setup",
1643                                 rx_queue_id);
1644                 return -EINVAL;
1645         }
1646
1647         /*
1648          * rx_queue_id is queue id application refers to, while
1649          * rxq->reg_idx is the real queue index.
1650          */
1651         err = i40e_switch_rx_queue(hw, rxq->reg_idx, FALSE);
1652         if (err) {
1653                 PMD_DRV_LOG(ERR, "Failed to switch RX queue %u off",
1654                             rx_queue_id);
1655                 return err;
1656         }
1657         i40e_rx_queue_release_mbufs(rxq);
1658         i40e_reset_rx_queue(rxq);
1659         dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
1660
1661         return 0;
1662 }
1663
1664 int
1665 i40e_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id)
1666 {
1667         int err;
1668         struct i40e_tx_queue *txq;
1669         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1670
1671         PMD_INIT_FUNC_TRACE();
1672
1673         txq = dev->data->tx_queues[tx_queue_id];
1674         if (!txq || !txq->q_set) {
1675                 PMD_DRV_LOG(ERR, "TX queue %u is not available or setup",
1676                             tx_queue_id);
1677                 return -EINVAL;
1678         }
1679
1680         if (txq->tx_deferred_start)
1681                 PMD_DRV_LOG(WARNING, "TX queue %u is deferrd start",
1682                             tx_queue_id);
1683
1684         /*
1685          * tx_queue_id is queue id application refers to, while
1686          * rxq->reg_idx is the real queue index.
1687          */
1688         err = i40e_switch_tx_queue(hw, txq->reg_idx, TRUE);
1689         if (err) {
1690                 PMD_DRV_LOG(ERR, "Failed to switch TX queue %u on",
1691                             tx_queue_id);
1692                 return err;
1693         }
1694         dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
1695
1696         return 0;
1697 }
1698
1699 int
1700 i40e_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id)
1701 {
1702         struct i40e_tx_queue *txq;
1703         int err;
1704         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1705
1706         txq = dev->data->tx_queues[tx_queue_id];
1707         if (!txq || !txq->q_set) {
1708                 PMD_DRV_LOG(ERR, "TX queue %u is not available or setup",
1709                         tx_queue_id);
1710                 return -EINVAL;
1711         }
1712
1713         /*
1714          * tx_queue_id is queue id application refers to, while
1715          * txq->reg_idx is the real queue index.
1716          */
1717         err = i40e_switch_tx_queue(hw, txq->reg_idx, FALSE);
1718         if (err) {
1719                 PMD_DRV_LOG(ERR, "Failed to switch TX queue %u of",
1720                             tx_queue_id);
1721                 return err;
1722         }
1723
1724         i40e_tx_queue_release_mbufs(txq);
1725         i40e_reset_tx_queue(txq);
1726         dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
1727
1728         return 0;
1729 }
1730
1731 const uint32_t *
1732 i40e_dev_supported_ptypes_get(struct rte_eth_dev *dev)
1733 {
1734         static const uint32_t ptypes[] = {
1735                 /* refers to i40e_rxd_pkt_type_mapping() */
1736                 RTE_PTYPE_L2_ETHER,
1737                 RTE_PTYPE_L2_ETHER_TIMESYNC,
1738                 RTE_PTYPE_L2_ETHER_LLDP,
1739                 RTE_PTYPE_L2_ETHER_ARP,
1740                 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
1741                 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
1742                 RTE_PTYPE_L4_FRAG,
1743                 RTE_PTYPE_L4_ICMP,
1744                 RTE_PTYPE_L4_NONFRAG,
1745                 RTE_PTYPE_L4_SCTP,
1746                 RTE_PTYPE_L4_TCP,
1747                 RTE_PTYPE_L4_UDP,
1748                 RTE_PTYPE_TUNNEL_GRENAT,
1749                 RTE_PTYPE_TUNNEL_IP,
1750                 RTE_PTYPE_INNER_L2_ETHER,
1751                 RTE_PTYPE_INNER_L2_ETHER_VLAN,
1752                 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
1753                 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
1754                 RTE_PTYPE_INNER_L4_FRAG,
1755                 RTE_PTYPE_INNER_L4_ICMP,
1756                 RTE_PTYPE_INNER_L4_NONFRAG,
1757                 RTE_PTYPE_INNER_L4_SCTP,
1758                 RTE_PTYPE_INNER_L4_TCP,
1759                 RTE_PTYPE_INNER_L4_UDP,
1760                 RTE_PTYPE_UNKNOWN
1761         };
1762
1763         if (dev->rx_pkt_burst == i40e_recv_pkts ||
1764 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
1765             dev->rx_pkt_burst == i40e_recv_pkts_bulk_alloc ||
1766 #endif
1767             dev->rx_pkt_burst == i40e_recv_scattered_pkts ||
1768             dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec ||
1769             dev->rx_pkt_burst == i40e_recv_pkts_vec ||
1770 #ifdef CC_AVX512_SUPPORT
1771             dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec_avx512 ||
1772             dev->rx_pkt_burst == i40e_recv_pkts_vec_avx512 ||
1773 #endif
1774             dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec_avx2 ||
1775             dev->rx_pkt_burst == i40e_recv_pkts_vec_avx2)
1776                 return ptypes;
1777         return NULL;
1778 }
1779
1780 static int
1781 i40e_dev_first_queue(uint16_t idx, void **queues, int num)
1782 {
1783         uint16_t i;
1784
1785         for (i = 0; i < num; i++) {
1786                 if (i != idx && queues[i])
1787                         return 0;
1788         }
1789
1790         return 1;
1791 }
1792
1793 static int
1794 i40e_dev_rx_queue_setup_runtime(struct rte_eth_dev *dev,
1795                                 struct i40e_rx_queue *rxq)
1796 {
1797         struct i40e_adapter *ad =
1798                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
1799         int use_def_burst_func =
1800                 check_rx_burst_bulk_alloc_preconditions(rxq);
1801         uint16_t buf_size =
1802                 (uint16_t)(rte_pktmbuf_data_room_size(rxq->mp) -
1803                            RTE_PKTMBUF_HEADROOM);
1804         int use_scattered_rx =
1805                 (rxq->max_pkt_len > buf_size);
1806
1807         if (i40e_rx_queue_init(rxq) != I40E_SUCCESS) {
1808                 PMD_DRV_LOG(ERR,
1809                             "Failed to do RX queue initialization");
1810                 return -EINVAL;
1811         }
1812
1813         if (i40e_dev_first_queue(rxq->queue_id,
1814                                  dev->data->rx_queues,
1815                                  dev->data->nb_rx_queues)) {
1816                 /**
1817                  * If it is the first queue to setup,
1818                  * set all flags to default and call
1819                  * i40e_set_rx_function.
1820                  */
1821                 ad->rx_bulk_alloc_allowed = true;
1822                 ad->rx_vec_allowed = true;
1823                 dev->data->scattered_rx = use_scattered_rx;
1824                 if (use_def_burst_func)
1825                         ad->rx_bulk_alloc_allowed = false;
1826                 i40e_set_rx_function(dev);
1827                 return 0;
1828         } else if (ad->rx_vec_allowed && !rte_is_power_of_2(rxq->nb_rx_desc)) {
1829                 PMD_DRV_LOG(ERR, "Vector mode is allowed, but descriptor"
1830                             " number %d of queue %d isn't power of 2",
1831                             rxq->nb_rx_desc, rxq->queue_id);
1832                 return -EINVAL;
1833         }
1834
1835         /* check bulk alloc conflict */
1836         if (ad->rx_bulk_alloc_allowed && use_def_burst_func) {
1837                 PMD_DRV_LOG(ERR, "Can't use default burst.");
1838                 return -EINVAL;
1839         }
1840         /* check scatterred conflict */
1841         if (!dev->data->scattered_rx && use_scattered_rx) {
1842                 PMD_DRV_LOG(ERR, "Scattered rx is required.");
1843                 return -EINVAL;
1844         }
1845         /* check vector conflict */
1846         if (ad->rx_vec_allowed && i40e_rxq_vec_setup(rxq)) {
1847                 PMD_DRV_LOG(ERR, "Failed vector rx setup.");
1848                 return -EINVAL;
1849         }
1850
1851         return 0;
1852 }
1853
1854 int
1855 i40e_dev_rx_queue_setup(struct rte_eth_dev *dev,
1856                         uint16_t queue_idx,
1857                         uint16_t nb_desc,
1858                         unsigned int socket_id,
1859                         const struct rte_eth_rxconf *rx_conf,
1860                         struct rte_mempool *mp)
1861 {
1862         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1863         struct i40e_adapter *ad =
1864                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
1865         struct i40e_vsi *vsi;
1866         struct i40e_pf *pf = NULL;
1867         struct i40e_vf *vf = NULL;
1868         struct i40e_rx_queue *rxq;
1869         const struct rte_memzone *rz;
1870         uint32_t ring_size;
1871         uint16_t len, i;
1872         uint16_t reg_idx, base, bsf, tc_mapping;
1873         int q_offset, use_def_burst_func = 1;
1874         uint64_t offloads;
1875
1876         offloads = rx_conf->offloads | dev->data->dev_conf.rxmode.offloads;
1877
1878         if (hw->mac.type == I40E_MAC_VF || hw->mac.type == I40E_MAC_X722_VF) {
1879                 vf = I40EVF_DEV_PRIVATE_TO_VF(dev->data->dev_private);
1880                 vsi = &vf->vsi;
1881                 if (!vsi)
1882                         return -EINVAL;
1883                 reg_idx = queue_idx;
1884         } else {
1885                 pf = I40E_DEV_PRIVATE_TO_PF(dev->data->dev_private);
1886                 vsi = i40e_pf_get_vsi_by_qindex(pf, queue_idx);
1887                 if (!vsi)
1888                         return -EINVAL;
1889                 q_offset = i40e_get_queue_offset_by_qindex(pf, queue_idx);
1890                 if (q_offset < 0)
1891                         return -EINVAL;
1892                 reg_idx = vsi->base_queue + q_offset;
1893         }
1894
1895         if (nb_desc % I40E_ALIGN_RING_DESC != 0 ||
1896             (nb_desc > I40E_MAX_RING_DESC) ||
1897             (nb_desc < I40E_MIN_RING_DESC)) {
1898                 PMD_DRV_LOG(ERR, "Number (%u) of receive descriptors is "
1899                             "invalid", nb_desc);
1900                 return -EINVAL;
1901         }
1902
1903         /* Free memory if needed */
1904         if (dev->data->rx_queues[queue_idx]) {
1905                 i40e_dev_rx_queue_release(dev->data->rx_queues[queue_idx]);
1906                 dev->data->rx_queues[queue_idx] = NULL;
1907         }
1908
1909         /* Allocate the rx queue data structure */
1910         rxq = rte_zmalloc_socket("i40e rx queue",
1911                                  sizeof(struct i40e_rx_queue),
1912                                  RTE_CACHE_LINE_SIZE,
1913                                  socket_id);
1914         if (!rxq) {
1915                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
1916                             "rx queue data structure");
1917                 return -ENOMEM;
1918         }
1919         rxq->mp = mp;
1920         rxq->nb_rx_desc = nb_desc;
1921         rxq->rx_free_thresh = rx_conf->rx_free_thresh;
1922         rxq->queue_id = queue_idx;
1923         rxq->reg_idx = reg_idx;
1924         rxq->port_id = dev->data->port_id;
1925         if (dev->data->dev_conf.rxmode.offloads & DEV_RX_OFFLOAD_KEEP_CRC)
1926                 rxq->crc_len = RTE_ETHER_CRC_LEN;
1927         else
1928                 rxq->crc_len = 0;
1929         rxq->drop_en = rx_conf->rx_drop_en;
1930         rxq->vsi = vsi;
1931         rxq->rx_deferred_start = rx_conf->rx_deferred_start;
1932         rxq->offloads = offloads;
1933
1934         /* Allocate the maximun number of RX ring hardware descriptor. */
1935         len = I40E_MAX_RING_DESC;
1936
1937         /**
1938          * Allocating a little more memory because vectorized/bulk_alloc Rx
1939          * functions doesn't check boundaries each time.
1940          */
1941         len += RTE_PMD_I40E_RX_MAX_BURST;
1942
1943         ring_size = RTE_ALIGN(len * sizeof(union i40e_rx_desc),
1944                               I40E_DMA_MEM_ALIGN);
1945
1946         rz = rte_eth_dma_zone_reserve(dev, "rx_ring", queue_idx,
1947                               ring_size, I40E_RING_BASE_ALIGN, socket_id);
1948         if (!rz) {
1949                 i40e_dev_rx_queue_release(rxq);
1950                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for RX");
1951                 return -ENOMEM;
1952         }
1953
1954         /* Zero all the descriptors in the ring. */
1955         memset(rz->addr, 0, ring_size);
1956
1957         rxq->rx_ring_phys_addr = rz->iova;
1958         rxq->rx_ring = (union i40e_rx_desc *)rz->addr;
1959
1960         len = (uint16_t)(nb_desc + RTE_PMD_I40E_RX_MAX_BURST);
1961
1962         /* Allocate the software ring. */
1963         rxq->sw_ring =
1964                 rte_zmalloc_socket("i40e rx sw ring",
1965                                    sizeof(struct i40e_rx_entry) * len,
1966                                    RTE_CACHE_LINE_SIZE,
1967                                    socket_id);
1968         if (!rxq->sw_ring) {
1969                 i40e_dev_rx_queue_release(rxq);
1970                 PMD_DRV_LOG(ERR, "Failed to allocate memory for SW ring");
1971                 return -ENOMEM;
1972         }
1973
1974         i40e_reset_rx_queue(rxq);
1975         rxq->q_set = TRUE;
1976
1977         for (i = 0; i < I40E_MAX_TRAFFIC_CLASS; i++) {
1978                 if (!(vsi->enabled_tc & (1 << i)))
1979                         continue;
1980                 tc_mapping = rte_le_to_cpu_16(vsi->info.tc_mapping[i]);
1981                 base = (tc_mapping & I40E_AQ_VSI_TC_QUE_OFFSET_MASK) >>
1982                         I40E_AQ_VSI_TC_QUE_OFFSET_SHIFT;
1983                 bsf = (tc_mapping & I40E_AQ_VSI_TC_QUE_NUMBER_MASK) >>
1984                         I40E_AQ_VSI_TC_QUE_NUMBER_SHIFT;
1985
1986                 if (queue_idx >= base && queue_idx < (base + BIT(bsf)))
1987                         rxq->dcb_tc = i;
1988         }
1989
1990         if (dev->data->dev_started) {
1991                 if (i40e_dev_rx_queue_setup_runtime(dev, rxq)) {
1992                         i40e_dev_rx_queue_release(rxq);
1993                         return -EINVAL;
1994                 }
1995         } else {
1996                 use_def_burst_func =
1997                         check_rx_burst_bulk_alloc_preconditions(rxq);
1998                 if (!use_def_burst_func) {
1999 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2000                         PMD_INIT_LOG(DEBUG,
2001                           "Rx Burst Bulk Alloc Preconditions are "
2002                           "satisfied. Rx Burst Bulk Alloc function will be "
2003                           "used on port=%d, queue=%d.",
2004                           rxq->port_id, rxq->queue_id);
2005 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2006                 } else {
2007                         PMD_INIT_LOG(DEBUG,
2008                           "Rx Burst Bulk Alloc Preconditions are "
2009                           "not satisfied, Scattered Rx is requested, "
2010                           "or RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC is "
2011                           "not enabled on port=%d, queue=%d.",
2012                           rxq->port_id, rxq->queue_id);
2013                         ad->rx_bulk_alloc_allowed = false;
2014                 }
2015         }
2016
2017         dev->data->rx_queues[queue_idx] = rxq;
2018         return 0;
2019 }
2020
2021 void
2022 i40e_dev_rx_queue_release(void *rxq)
2023 {
2024         struct i40e_rx_queue *q = (struct i40e_rx_queue *)rxq;
2025
2026         if (!q) {
2027                 PMD_DRV_LOG(DEBUG, "Pointer to rxq is NULL");
2028                 return;
2029         }
2030
2031         i40e_rx_queue_release_mbufs(q);
2032         rte_free(q->sw_ring);
2033         rte_free(q);
2034 }
2035
2036 uint32_t
2037 i40e_dev_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
2038 {
2039 #define I40E_RXQ_SCAN_INTERVAL 4
2040         volatile union i40e_rx_desc *rxdp;
2041         struct i40e_rx_queue *rxq;
2042         uint16_t desc = 0;
2043
2044         rxq = dev->data->rx_queues[rx_queue_id];
2045         rxdp = &(rxq->rx_ring[rxq->rx_tail]);
2046         while ((desc < rxq->nb_rx_desc) &&
2047                 ((rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len) &
2048                 I40E_RXD_QW1_STATUS_MASK) >> I40E_RXD_QW1_STATUS_SHIFT) &
2049                                 (1 << I40E_RX_DESC_STATUS_DD_SHIFT)) {
2050                 /**
2051                  * Check the DD bit of a rx descriptor of each 4 in a group,
2052                  * to avoid checking too frequently and downgrading performance
2053                  * too much.
2054                  */
2055                 desc += I40E_RXQ_SCAN_INTERVAL;
2056                 rxdp += I40E_RXQ_SCAN_INTERVAL;
2057                 if (rxq->rx_tail + desc >= rxq->nb_rx_desc)
2058                         rxdp = &(rxq->rx_ring[rxq->rx_tail +
2059                                         desc - rxq->nb_rx_desc]);
2060         }
2061
2062         return desc;
2063 }
2064
2065 int
2066 i40e_dev_rx_descriptor_done(void *rx_queue, uint16_t offset)
2067 {
2068         volatile union i40e_rx_desc *rxdp;
2069         struct i40e_rx_queue *rxq = rx_queue;
2070         uint16_t desc;
2071         int ret;
2072
2073         if (unlikely(offset >= rxq->nb_rx_desc)) {
2074                 PMD_DRV_LOG(ERR, "Invalid RX descriptor id %u", offset);
2075                 return 0;
2076         }
2077
2078         desc = rxq->rx_tail + offset;
2079         if (desc >= rxq->nb_rx_desc)
2080                 desc -= rxq->nb_rx_desc;
2081
2082         rxdp = &(rxq->rx_ring[desc]);
2083
2084         ret = !!(((rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len) &
2085                 I40E_RXD_QW1_STATUS_MASK) >> I40E_RXD_QW1_STATUS_SHIFT) &
2086                                 (1 << I40E_RX_DESC_STATUS_DD_SHIFT));
2087
2088         return ret;
2089 }
2090
2091 int
2092 i40e_dev_rx_descriptor_status(void *rx_queue, uint16_t offset)
2093 {
2094         struct i40e_rx_queue *rxq = rx_queue;
2095         volatile uint64_t *status;
2096         uint64_t mask;
2097         uint32_t desc;
2098
2099         if (unlikely(offset >= rxq->nb_rx_desc))
2100                 return -EINVAL;
2101
2102         if (offset >= rxq->nb_rx_desc - rxq->nb_rx_hold)
2103                 return RTE_ETH_RX_DESC_UNAVAIL;
2104
2105         desc = rxq->rx_tail + offset;
2106         if (desc >= rxq->nb_rx_desc)
2107                 desc -= rxq->nb_rx_desc;
2108
2109         status = &rxq->rx_ring[desc].wb.qword1.status_error_len;
2110         mask = rte_le_to_cpu_64((1ULL << I40E_RX_DESC_STATUS_DD_SHIFT)
2111                 << I40E_RXD_QW1_STATUS_SHIFT);
2112         if (*status & mask)
2113                 return RTE_ETH_RX_DESC_DONE;
2114
2115         return RTE_ETH_RX_DESC_AVAIL;
2116 }
2117
2118 int
2119 i40e_dev_tx_descriptor_status(void *tx_queue, uint16_t offset)
2120 {
2121         struct i40e_tx_queue *txq = tx_queue;
2122         volatile uint64_t *status;
2123         uint64_t mask, expect;
2124         uint32_t desc;
2125
2126         if (unlikely(offset >= txq->nb_tx_desc))
2127                 return -EINVAL;
2128
2129         desc = txq->tx_tail + offset;
2130         /* go to next desc that has the RS bit */
2131         desc = ((desc + txq->tx_rs_thresh - 1) / txq->tx_rs_thresh) *
2132                 txq->tx_rs_thresh;
2133         if (desc >= txq->nb_tx_desc) {
2134                 desc -= txq->nb_tx_desc;
2135                 if (desc >= txq->nb_tx_desc)
2136                         desc -= txq->nb_tx_desc;
2137         }
2138
2139         status = &txq->tx_ring[desc].cmd_type_offset_bsz;
2140         mask = rte_le_to_cpu_64(I40E_TXD_QW1_DTYPE_MASK);
2141         expect = rte_cpu_to_le_64(
2142                 I40E_TX_DESC_DTYPE_DESC_DONE << I40E_TXD_QW1_DTYPE_SHIFT);
2143         if ((*status & mask) == expect)
2144                 return RTE_ETH_TX_DESC_DONE;
2145
2146         return RTE_ETH_TX_DESC_FULL;
2147 }
2148
2149 static int
2150 i40e_dev_tx_queue_setup_runtime(struct rte_eth_dev *dev,
2151                                 struct i40e_tx_queue *txq)
2152 {
2153         struct i40e_adapter *ad =
2154                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
2155
2156         if (i40e_tx_queue_init(txq) != I40E_SUCCESS) {
2157                 PMD_DRV_LOG(ERR,
2158                             "Failed to do TX queue initialization");
2159                 return -EINVAL;
2160         }
2161
2162         if (i40e_dev_first_queue(txq->queue_id,
2163                                  dev->data->tx_queues,
2164                                  dev->data->nb_tx_queues)) {
2165                 /**
2166                  * If it is the first queue to setup,
2167                  * set all flags and call
2168                  * i40e_set_tx_function.
2169                  */
2170                 i40e_set_tx_function_flag(dev, txq);
2171                 i40e_set_tx_function(dev);
2172                 return 0;
2173         }
2174
2175         /* check vector conflict */
2176         if (ad->tx_vec_allowed) {
2177                 if (txq->tx_rs_thresh > RTE_I40E_TX_MAX_FREE_BUF_SZ ||
2178                     i40e_txq_vec_setup(txq)) {
2179                         PMD_DRV_LOG(ERR, "Failed vector tx setup.");
2180                         return -EINVAL;
2181                 }
2182         }
2183         /* check simple tx conflict */
2184         if (ad->tx_simple_allowed) {
2185                 if ((txq->offloads & ~DEV_TX_OFFLOAD_MBUF_FAST_FREE) != 0 ||
2186                                 txq->tx_rs_thresh < RTE_PMD_I40E_TX_MAX_BURST) {
2187                         PMD_DRV_LOG(ERR, "No-simple tx is required.");
2188                         return -EINVAL;
2189                 }
2190         }
2191
2192         return 0;
2193 }
2194
2195 int
2196 i40e_dev_tx_queue_setup(struct rte_eth_dev *dev,
2197                         uint16_t queue_idx,
2198                         uint16_t nb_desc,
2199                         unsigned int socket_id,
2200                         const struct rte_eth_txconf *tx_conf)
2201 {
2202         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2203         struct i40e_vsi *vsi;
2204         struct i40e_pf *pf = NULL;
2205         struct i40e_vf *vf = NULL;
2206         struct i40e_tx_queue *txq;
2207         const struct rte_memzone *tz;
2208         uint32_t ring_size;
2209         uint16_t tx_rs_thresh, tx_free_thresh;
2210         uint16_t reg_idx, i, base, bsf, tc_mapping;
2211         int q_offset;
2212         uint64_t offloads;
2213
2214         offloads = tx_conf->offloads | dev->data->dev_conf.txmode.offloads;
2215
2216         if (hw->mac.type == I40E_MAC_VF || hw->mac.type == I40E_MAC_X722_VF) {
2217                 vf = I40EVF_DEV_PRIVATE_TO_VF(dev->data->dev_private);
2218                 vsi = &vf->vsi;
2219                 if (!vsi)
2220                         return -EINVAL;
2221                 reg_idx = queue_idx;
2222         } else {
2223                 pf = I40E_DEV_PRIVATE_TO_PF(dev->data->dev_private);
2224                 vsi = i40e_pf_get_vsi_by_qindex(pf, queue_idx);
2225                 if (!vsi)
2226                         return -EINVAL;
2227                 q_offset = i40e_get_queue_offset_by_qindex(pf, queue_idx);
2228                 if (q_offset < 0)
2229                         return -EINVAL;
2230                 reg_idx = vsi->base_queue + q_offset;
2231         }
2232
2233         if (nb_desc % I40E_ALIGN_RING_DESC != 0 ||
2234             (nb_desc > I40E_MAX_RING_DESC) ||
2235             (nb_desc < I40E_MIN_RING_DESC)) {
2236                 PMD_DRV_LOG(ERR, "Number (%u) of transmit descriptors is "
2237                             "invalid", nb_desc);
2238                 return -EINVAL;
2239         }
2240
2241         /**
2242          * The following two parameters control the setting of the RS bit on
2243          * transmit descriptors. TX descriptors will have their RS bit set
2244          * after txq->tx_rs_thresh descriptors have been used. The TX
2245          * descriptor ring will be cleaned after txq->tx_free_thresh
2246          * descriptors are used or if the number of descriptors required to
2247          * transmit a packet is greater than the number of free TX descriptors.
2248          *
2249          * The following constraints must be satisfied:
2250          *  - tx_rs_thresh must be greater than 0.
2251          *  - tx_rs_thresh must be less than the size of the ring minus 2.
2252          *  - tx_rs_thresh must be less than or equal to tx_free_thresh.
2253          *  - tx_rs_thresh must be a divisor of the ring size.
2254          *  - tx_free_thresh must be greater than 0.
2255          *  - tx_free_thresh must be less than the size of the ring minus 3.
2256          *  - tx_free_thresh + tx_rs_thresh must not exceed nb_desc.
2257          *
2258          * One descriptor in the TX ring is used as a sentinel to avoid a H/W
2259          * race condition, hence the maximum threshold constraints. When set
2260          * to zero use default values.
2261          */
2262         tx_free_thresh = (uint16_t)((tx_conf->tx_free_thresh) ?
2263                 tx_conf->tx_free_thresh : DEFAULT_TX_FREE_THRESH);
2264         /* force tx_rs_thresh to adapt an aggresive tx_free_thresh */
2265         tx_rs_thresh = (DEFAULT_TX_RS_THRESH + tx_free_thresh > nb_desc) ?
2266                 nb_desc - tx_free_thresh : DEFAULT_TX_RS_THRESH;
2267         if (tx_conf->tx_rs_thresh > 0)
2268                 tx_rs_thresh = tx_conf->tx_rs_thresh;
2269         if (tx_rs_thresh + tx_free_thresh > nb_desc) {
2270                 PMD_INIT_LOG(ERR, "tx_rs_thresh + tx_free_thresh must not "
2271                                 "exceed nb_desc. (tx_rs_thresh=%u "
2272                                 "tx_free_thresh=%u nb_desc=%u port=%d queue=%d)",
2273                                 (unsigned int)tx_rs_thresh,
2274                                 (unsigned int)tx_free_thresh,
2275                                 (unsigned int)nb_desc,
2276                                 (int)dev->data->port_id,
2277                                 (int)queue_idx);
2278                 return I40E_ERR_PARAM;
2279         }
2280         if (tx_rs_thresh >= (nb_desc - 2)) {
2281                 PMD_INIT_LOG(ERR, "tx_rs_thresh must be less than the "
2282                              "number of TX descriptors minus 2. "
2283                              "(tx_rs_thresh=%u port=%d queue=%d)",
2284                              (unsigned int)tx_rs_thresh,
2285                              (int)dev->data->port_id,
2286                              (int)queue_idx);
2287                 return I40E_ERR_PARAM;
2288         }
2289         if (tx_free_thresh >= (nb_desc - 3)) {
2290                 PMD_INIT_LOG(ERR, "tx_free_thresh must be less than the "
2291                              "number of TX descriptors minus 3. "
2292                              "(tx_free_thresh=%u port=%d queue=%d)",
2293                              (unsigned int)tx_free_thresh,
2294                              (int)dev->data->port_id,
2295                              (int)queue_idx);
2296                 return I40E_ERR_PARAM;
2297         }
2298         if (tx_rs_thresh > tx_free_thresh) {
2299                 PMD_INIT_LOG(ERR, "tx_rs_thresh must be less than or "
2300                              "equal to tx_free_thresh. (tx_free_thresh=%u"
2301                              " tx_rs_thresh=%u port=%d queue=%d)",
2302                              (unsigned int)tx_free_thresh,
2303                              (unsigned int)tx_rs_thresh,
2304                              (int)dev->data->port_id,
2305                              (int)queue_idx);
2306                 return I40E_ERR_PARAM;
2307         }
2308         if ((nb_desc % tx_rs_thresh) != 0) {
2309                 PMD_INIT_LOG(ERR, "tx_rs_thresh must be a divisor of the "
2310                              "number of TX descriptors. (tx_rs_thresh=%u"
2311                              " port=%d queue=%d)",
2312                              (unsigned int)tx_rs_thresh,
2313                              (int)dev->data->port_id,
2314                              (int)queue_idx);
2315                 return I40E_ERR_PARAM;
2316         }
2317         if ((tx_rs_thresh > 1) && (tx_conf->tx_thresh.wthresh != 0)) {
2318                 PMD_INIT_LOG(ERR, "TX WTHRESH must be set to 0 if "
2319                              "tx_rs_thresh is greater than 1. "
2320                              "(tx_rs_thresh=%u port=%d queue=%d)",
2321                              (unsigned int)tx_rs_thresh,
2322                              (int)dev->data->port_id,
2323                              (int)queue_idx);
2324                 return I40E_ERR_PARAM;
2325         }
2326
2327         /* Free memory if needed. */
2328         if (dev->data->tx_queues[queue_idx]) {
2329                 i40e_dev_tx_queue_release(dev->data->tx_queues[queue_idx]);
2330                 dev->data->tx_queues[queue_idx] = NULL;
2331         }
2332
2333         /* Allocate the TX queue data structure. */
2334         txq = rte_zmalloc_socket("i40e tx queue",
2335                                   sizeof(struct i40e_tx_queue),
2336                                   RTE_CACHE_LINE_SIZE,
2337                                   socket_id);
2338         if (!txq) {
2339                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
2340                             "tx queue structure");
2341                 return -ENOMEM;
2342         }
2343
2344         /* Allocate TX hardware ring descriptors. */
2345         ring_size = sizeof(struct i40e_tx_desc) * I40E_MAX_RING_DESC;
2346         ring_size = RTE_ALIGN(ring_size, I40E_DMA_MEM_ALIGN);
2347         tz = rte_eth_dma_zone_reserve(dev, "tx_ring", queue_idx,
2348                               ring_size, I40E_RING_BASE_ALIGN, socket_id);
2349         if (!tz) {
2350                 i40e_dev_tx_queue_release(txq);
2351                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for TX");
2352                 return -ENOMEM;
2353         }
2354
2355         txq->nb_tx_desc = nb_desc;
2356         txq->tx_rs_thresh = tx_rs_thresh;
2357         txq->tx_free_thresh = tx_free_thresh;
2358         txq->pthresh = tx_conf->tx_thresh.pthresh;
2359         txq->hthresh = tx_conf->tx_thresh.hthresh;
2360         txq->wthresh = tx_conf->tx_thresh.wthresh;
2361         txq->queue_id = queue_idx;
2362         txq->reg_idx = reg_idx;
2363         txq->port_id = dev->data->port_id;
2364         txq->offloads = offloads;
2365         txq->vsi = vsi;
2366         txq->tx_deferred_start = tx_conf->tx_deferred_start;
2367
2368         txq->tx_ring_phys_addr = tz->iova;
2369         txq->tx_ring = (struct i40e_tx_desc *)tz->addr;
2370
2371         /* Allocate software ring */
2372         txq->sw_ring =
2373                 rte_zmalloc_socket("i40e tx sw ring",
2374                                    sizeof(struct i40e_tx_entry) * nb_desc,
2375                                    RTE_CACHE_LINE_SIZE,
2376                                    socket_id);
2377         if (!txq->sw_ring) {
2378                 i40e_dev_tx_queue_release(txq);
2379                 PMD_DRV_LOG(ERR, "Failed to allocate memory for SW TX ring");
2380                 return -ENOMEM;
2381         }
2382
2383         i40e_reset_tx_queue(txq);
2384         txq->q_set = TRUE;
2385
2386         for (i = 0; i < I40E_MAX_TRAFFIC_CLASS; i++) {
2387                 if (!(vsi->enabled_tc & (1 << i)))
2388                         continue;
2389                 tc_mapping = rte_le_to_cpu_16(vsi->info.tc_mapping[i]);
2390                 base = (tc_mapping & I40E_AQ_VSI_TC_QUE_OFFSET_MASK) >>
2391                         I40E_AQ_VSI_TC_QUE_OFFSET_SHIFT;
2392                 bsf = (tc_mapping & I40E_AQ_VSI_TC_QUE_NUMBER_MASK) >>
2393                         I40E_AQ_VSI_TC_QUE_NUMBER_SHIFT;
2394
2395                 if (queue_idx >= base && queue_idx < (base + BIT(bsf)))
2396                         txq->dcb_tc = i;
2397         }
2398
2399         if (dev->data->dev_started) {
2400                 if (i40e_dev_tx_queue_setup_runtime(dev, txq)) {
2401                         i40e_dev_tx_queue_release(txq);
2402                         return -EINVAL;
2403                 }
2404         } else {
2405                 /**
2406                  * Use a simple TX queue without offloads or
2407                  * multi segs if possible
2408                  */
2409                 i40e_set_tx_function_flag(dev, txq);
2410         }
2411         dev->data->tx_queues[queue_idx] = txq;
2412
2413         return 0;
2414 }
2415
2416 void
2417 i40e_dev_tx_queue_release(void *txq)
2418 {
2419         struct i40e_tx_queue *q = (struct i40e_tx_queue *)txq;
2420
2421         if (!q) {
2422                 PMD_DRV_LOG(DEBUG, "Pointer to TX queue is NULL");
2423                 return;
2424         }
2425
2426         i40e_tx_queue_release_mbufs(q);
2427         rte_free(q->sw_ring);
2428         rte_free(q);
2429 }
2430
2431 const struct rte_memzone *
2432 i40e_memzone_reserve(const char *name, uint32_t len, int socket_id)
2433 {
2434         const struct rte_memzone *mz;
2435
2436         mz = rte_memzone_lookup(name);
2437         if (mz)
2438                 return mz;
2439
2440         mz = rte_memzone_reserve_aligned(name, len, socket_id,
2441                         RTE_MEMZONE_IOVA_CONTIG, I40E_RING_BASE_ALIGN);
2442         return mz;
2443 }
2444
2445 void
2446 i40e_rx_queue_release_mbufs(struct i40e_rx_queue *rxq)
2447 {
2448         uint16_t i;
2449
2450         /* SSE Vector driver has a different way of releasing mbufs. */
2451         if (rxq->rx_using_sse) {
2452                 i40e_rx_queue_release_mbufs_vec(rxq);
2453                 return;
2454         }
2455
2456         if (!rxq->sw_ring) {
2457                 PMD_DRV_LOG(DEBUG, "Pointer to sw_ring is NULL");
2458                 return;
2459         }
2460
2461         for (i = 0; i < rxq->nb_rx_desc; i++) {
2462                 if (rxq->sw_ring[i].mbuf) {
2463                         rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
2464                         rxq->sw_ring[i].mbuf = NULL;
2465                 }
2466         }
2467 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2468         if (rxq->rx_nb_avail == 0)
2469                 return;
2470         for (i = 0; i < rxq->rx_nb_avail; i++) {
2471                 struct rte_mbuf *mbuf;
2472
2473                 mbuf = rxq->rx_stage[rxq->rx_next_avail + i];
2474                 rte_pktmbuf_free_seg(mbuf);
2475         }
2476         rxq->rx_nb_avail = 0;
2477 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2478 }
2479
2480 void
2481 i40e_reset_rx_queue(struct i40e_rx_queue *rxq)
2482 {
2483         unsigned i;
2484         uint16_t len;
2485
2486         if (!rxq) {
2487                 PMD_DRV_LOG(DEBUG, "Pointer to rxq is NULL");
2488                 return;
2489         }
2490
2491 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2492         if (check_rx_burst_bulk_alloc_preconditions(rxq) == 0)
2493                 len = (uint16_t)(rxq->nb_rx_desc + RTE_PMD_I40E_RX_MAX_BURST);
2494         else
2495 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2496                 len = rxq->nb_rx_desc;
2497
2498         for (i = 0; i < len * sizeof(union i40e_rx_desc); i++)
2499                 ((volatile char *)rxq->rx_ring)[i] = 0;
2500
2501         memset(&rxq->fake_mbuf, 0x0, sizeof(rxq->fake_mbuf));
2502         for (i = 0; i < RTE_PMD_I40E_RX_MAX_BURST; ++i)
2503                 rxq->sw_ring[rxq->nb_rx_desc + i].mbuf = &rxq->fake_mbuf;
2504
2505 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2506         rxq->rx_nb_avail = 0;
2507         rxq->rx_next_avail = 0;
2508         rxq->rx_free_trigger = (uint16_t)(rxq->rx_free_thresh - 1);
2509 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2510         rxq->rx_tail = 0;
2511         rxq->nb_rx_hold = 0;
2512         rxq->pkt_first_seg = NULL;
2513         rxq->pkt_last_seg = NULL;
2514
2515         rxq->rxrearm_start = 0;
2516         rxq->rxrearm_nb = 0;
2517 }
2518
2519 void
2520 i40e_tx_queue_release_mbufs(struct i40e_tx_queue *txq)
2521 {
2522         struct rte_eth_dev *dev;
2523         uint16_t i;
2524
2525         if (!txq || !txq->sw_ring) {
2526                 PMD_DRV_LOG(DEBUG, "Pointer to txq or sw_ring is NULL");
2527                 return;
2528         }
2529
2530         dev = &rte_eth_devices[txq->port_id];
2531
2532         /**
2533          *  vPMD tx will not set sw_ring's mbuf to NULL after free,
2534          *  so need to free remains more carefully.
2535          */
2536 #ifdef CC_AVX512_SUPPORT
2537         if (dev->tx_pkt_burst == i40e_xmit_pkts_vec_avx512) {
2538                 struct i40e_vec_tx_entry *swr = (void *)txq->sw_ring;
2539
2540                 i = txq->tx_next_dd - txq->tx_rs_thresh + 1;
2541                 if (txq->tx_tail < i) {
2542                         for (; i < txq->nb_tx_desc; i++) {
2543                                 rte_pktmbuf_free_seg(swr[i].mbuf);
2544                                 swr[i].mbuf = NULL;
2545                         }
2546                         i = 0;
2547                 }
2548                 for (; i < txq->tx_tail; i++) {
2549                         rte_pktmbuf_free_seg(swr[i].mbuf);
2550                         swr[i].mbuf = NULL;
2551                 }
2552                 return;
2553         }
2554 #endif
2555         if (dev->tx_pkt_burst == i40e_xmit_pkts_vec_avx2 ||
2556                         dev->tx_pkt_burst == i40e_xmit_pkts_vec) {
2557                 i = txq->tx_next_dd - txq->tx_rs_thresh + 1;
2558                 if (txq->tx_tail < i) {
2559                         for (; i < txq->nb_tx_desc; i++) {
2560                                 rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
2561                                 txq->sw_ring[i].mbuf = NULL;
2562                         }
2563                         i = 0;
2564                 }
2565                 for (; i < txq->tx_tail; i++) {
2566                         rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
2567                         txq->sw_ring[i].mbuf = NULL;
2568                 }
2569         } else {
2570                 for (i = 0; i < txq->nb_tx_desc; i++) {
2571                         if (txq->sw_ring[i].mbuf) {
2572                                 rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
2573                                 txq->sw_ring[i].mbuf = NULL;
2574                         }
2575                 }
2576         }
2577 }
2578
2579 static int
2580 i40e_tx_done_cleanup_full(struct i40e_tx_queue *txq,
2581                         uint32_t free_cnt)
2582 {
2583         struct i40e_tx_entry *swr_ring = txq->sw_ring;
2584         uint16_t i, tx_last, tx_id;
2585         uint16_t nb_tx_free_last;
2586         uint16_t nb_tx_to_clean;
2587         uint32_t pkt_cnt;
2588
2589         /* Start free mbuf from the next of tx_tail */
2590         tx_last = txq->tx_tail;
2591         tx_id  = swr_ring[tx_last].next_id;
2592
2593         if (txq->nb_tx_free == 0 && i40e_xmit_cleanup(txq))
2594                 return 0;
2595
2596         nb_tx_to_clean = txq->nb_tx_free;
2597         nb_tx_free_last = txq->nb_tx_free;
2598         if (!free_cnt)
2599                 free_cnt = txq->nb_tx_desc;
2600
2601         /* Loop through swr_ring to count the amount of
2602          * freeable mubfs and packets.
2603          */
2604         for (pkt_cnt = 0; pkt_cnt < free_cnt; ) {
2605                 for (i = 0; i < nb_tx_to_clean &&
2606                         pkt_cnt < free_cnt &&
2607                         tx_id != tx_last; i++) {
2608                         if (swr_ring[tx_id].mbuf != NULL) {
2609                                 rte_pktmbuf_free_seg(swr_ring[tx_id].mbuf);
2610                                 swr_ring[tx_id].mbuf = NULL;
2611
2612                                 /*
2613                                  * last segment in the packet,
2614                                  * increment packet count
2615                                  */
2616                                 pkt_cnt += (swr_ring[tx_id].last_id == tx_id);
2617                         }
2618
2619                         tx_id = swr_ring[tx_id].next_id;
2620                 }
2621
2622                 if (txq->tx_rs_thresh > txq->nb_tx_desc -
2623                         txq->nb_tx_free || tx_id == tx_last)
2624                         break;
2625
2626                 if (pkt_cnt < free_cnt) {
2627                         if (i40e_xmit_cleanup(txq))
2628                                 break;
2629
2630                         nb_tx_to_clean = txq->nb_tx_free - nb_tx_free_last;
2631                         nb_tx_free_last = txq->nb_tx_free;
2632                 }
2633         }
2634
2635         return (int)pkt_cnt;
2636 }
2637
2638 static int
2639 i40e_tx_done_cleanup_simple(struct i40e_tx_queue *txq,
2640                         uint32_t free_cnt)
2641 {
2642         int i, n, cnt;
2643
2644         if (free_cnt == 0 || free_cnt > txq->nb_tx_desc)
2645                 free_cnt = txq->nb_tx_desc;
2646
2647         cnt = free_cnt - free_cnt % txq->tx_rs_thresh;
2648
2649         for (i = 0; i < cnt; i += n) {
2650                 if (txq->nb_tx_desc - txq->nb_tx_free < txq->tx_rs_thresh)
2651                         break;
2652
2653                 n = i40e_tx_free_bufs(txq);
2654
2655                 if (n == 0)
2656                         break;
2657         }
2658
2659         return i;
2660 }
2661
2662 static int
2663 i40e_tx_done_cleanup_vec(struct i40e_tx_queue *txq __rte_unused,
2664                         uint32_t free_cnt __rte_unused)
2665 {
2666         return -ENOTSUP;
2667 }
2668 int
2669 i40e_tx_done_cleanup(void *txq, uint32_t free_cnt)
2670 {
2671         struct i40e_tx_queue *q = (struct i40e_tx_queue *)txq;
2672         struct rte_eth_dev *dev = &rte_eth_devices[q->port_id];
2673         struct i40e_adapter *ad =
2674                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
2675
2676         if (ad->tx_simple_allowed) {
2677                 if (ad->tx_vec_allowed)
2678                         return i40e_tx_done_cleanup_vec(q, free_cnt);
2679                 else
2680                         return i40e_tx_done_cleanup_simple(q, free_cnt);
2681         } else {
2682                 return i40e_tx_done_cleanup_full(q, free_cnt);
2683         }
2684 }
2685
2686 void
2687 i40e_reset_tx_queue(struct i40e_tx_queue *txq)
2688 {
2689         struct i40e_tx_entry *txe;
2690         uint16_t i, prev, size;
2691
2692         if (!txq) {
2693                 PMD_DRV_LOG(DEBUG, "Pointer to txq is NULL");
2694                 return;
2695         }
2696
2697         txe = txq->sw_ring;
2698         size = sizeof(struct i40e_tx_desc) * txq->nb_tx_desc;
2699         for (i = 0; i < size; i++)
2700                 ((volatile char *)txq->tx_ring)[i] = 0;
2701
2702         prev = (uint16_t)(txq->nb_tx_desc - 1);
2703         for (i = 0; i < txq->nb_tx_desc; i++) {
2704                 volatile struct i40e_tx_desc *txd = &txq->tx_ring[i];
2705
2706                 txd->cmd_type_offset_bsz =
2707                         rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE);
2708                 txe[i].mbuf =  NULL;
2709                 txe[i].last_id = i;
2710                 txe[prev].next_id = i;
2711                 prev = i;
2712         }
2713
2714         txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
2715         txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
2716
2717         txq->tx_tail = 0;
2718         txq->nb_tx_used = 0;
2719
2720         txq->last_desc_cleaned = (uint16_t)(txq->nb_tx_desc - 1);
2721         txq->nb_tx_free = (uint16_t)(txq->nb_tx_desc - 1);
2722 }
2723
2724 /* Init the TX queue in hardware */
2725 int
2726 i40e_tx_queue_init(struct i40e_tx_queue *txq)
2727 {
2728         enum i40e_status_code err = I40E_SUCCESS;
2729         struct i40e_vsi *vsi = txq->vsi;
2730         struct i40e_hw *hw = I40E_VSI_TO_HW(vsi);
2731         uint16_t pf_q = txq->reg_idx;
2732         struct i40e_hmc_obj_txq tx_ctx;
2733         uint32_t qtx_ctl;
2734
2735         /* clear the context structure first */
2736         memset(&tx_ctx, 0, sizeof(tx_ctx));
2737         tx_ctx.new_context = 1;
2738         tx_ctx.base = txq->tx_ring_phys_addr / I40E_QUEUE_BASE_ADDR_UNIT;
2739         tx_ctx.qlen = txq->nb_tx_desc;
2740
2741 #ifdef RTE_LIBRTE_IEEE1588
2742         tx_ctx.timesync_ena = 1;
2743 #endif
2744         tx_ctx.rdylist = rte_le_to_cpu_16(vsi->info.qs_handle[txq->dcb_tc]);
2745         if (vsi->type == I40E_VSI_FDIR)
2746                 tx_ctx.fd_ena = TRUE;
2747
2748         err = i40e_clear_lan_tx_queue_context(hw, pf_q);
2749         if (err != I40E_SUCCESS) {
2750                 PMD_DRV_LOG(ERR, "Failure of clean lan tx queue context");
2751                 return err;
2752         }
2753
2754         err = i40e_set_lan_tx_queue_context(hw, pf_q, &tx_ctx);
2755         if (err != I40E_SUCCESS) {
2756                 PMD_DRV_LOG(ERR, "Failure of set lan tx queue context");
2757                 return err;
2758         }
2759
2760         /* Now associate this queue with this PCI function */
2761         qtx_ctl = I40E_QTX_CTL_PF_QUEUE;
2762         qtx_ctl |= ((hw->pf_id << I40E_QTX_CTL_PF_INDX_SHIFT) &
2763                                         I40E_QTX_CTL_PF_INDX_MASK);
2764         I40E_WRITE_REG(hw, I40E_QTX_CTL(pf_q), qtx_ctl);
2765         I40E_WRITE_FLUSH(hw);
2766
2767         txq->qtx_tail = hw->hw_addr + I40E_QTX_TAIL(pf_q);
2768
2769         return err;
2770 }
2771
2772 int
2773 i40e_alloc_rx_queue_mbufs(struct i40e_rx_queue *rxq)
2774 {
2775         struct i40e_rx_entry *rxe = rxq->sw_ring;
2776         uint64_t dma_addr;
2777         uint16_t i;
2778
2779         for (i = 0; i < rxq->nb_rx_desc; i++) {
2780                 volatile union i40e_rx_desc *rxd;
2781                 struct rte_mbuf *mbuf = rte_mbuf_raw_alloc(rxq->mp);
2782
2783                 if (unlikely(!mbuf)) {
2784                         PMD_DRV_LOG(ERR, "Failed to allocate mbuf for RX");
2785                         return -ENOMEM;
2786                 }
2787
2788                 rte_mbuf_refcnt_set(mbuf, 1);
2789                 mbuf->next = NULL;
2790                 mbuf->data_off = RTE_PKTMBUF_HEADROOM;
2791                 mbuf->nb_segs = 1;
2792                 mbuf->port = rxq->port_id;
2793
2794                 dma_addr =
2795                         rte_cpu_to_le_64(rte_mbuf_data_iova_default(mbuf));
2796
2797                 rxd = &rxq->rx_ring[i];
2798                 rxd->read.pkt_addr = dma_addr;
2799                 rxd->read.hdr_addr = 0;
2800 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
2801                 rxd->read.rsvd1 = 0;
2802                 rxd->read.rsvd2 = 0;
2803 #endif /* RTE_LIBRTE_I40E_16BYTE_RX_DESC */
2804
2805                 rxe[i].mbuf = mbuf;
2806         }
2807
2808         return 0;
2809 }
2810
2811 /*
2812  * Calculate the buffer length, and check the jumbo frame
2813  * and maximum packet length.
2814  */
2815 static int
2816 i40e_rx_queue_config(struct i40e_rx_queue *rxq)
2817 {
2818         struct i40e_pf *pf = I40E_VSI_TO_PF(rxq->vsi);
2819         struct i40e_hw *hw = I40E_VSI_TO_HW(rxq->vsi);
2820         struct rte_eth_dev_data *data = pf->dev_data;
2821         uint16_t buf_size;
2822
2823         buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mp) -
2824                 RTE_PKTMBUF_HEADROOM);
2825
2826         switch (pf->flags & (I40E_FLAG_HEADER_SPLIT_DISABLED |
2827                         I40E_FLAG_HEADER_SPLIT_ENABLED)) {
2828         case I40E_FLAG_HEADER_SPLIT_ENABLED: /* Not supported */
2829                 rxq->rx_hdr_len = RTE_ALIGN(I40E_RXBUF_SZ_1024,
2830                                 (1 << I40E_RXQ_CTX_HBUFF_SHIFT));
2831                 rxq->rx_buf_len = RTE_ALIGN(I40E_RXBUF_SZ_2048,
2832                                 (1 << I40E_RXQ_CTX_DBUFF_SHIFT));
2833                 rxq->hs_mode = i40e_header_split_enabled;
2834                 break;
2835         case I40E_FLAG_HEADER_SPLIT_DISABLED:
2836         default:
2837                 rxq->rx_hdr_len = 0;
2838                 rxq->rx_buf_len = RTE_ALIGN_FLOOR(buf_size,
2839                         (1 << I40E_RXQ_CTX_DBUFF_SHIFT));
2840                 rxq->hs_mode = i40e_header_split_none;
2841                 break;
2842         }
2843
2844         rxq->max_pkt_len =
2845                 RTE_MIN((uint32_t)(hw->func_caps.rx_buf_chain_len *
2846                         rxq->rx_buf_len), data->dev_conf.rxmode.max_rx_pkt_len);
2847         if (data->dev_conf.rxmode.offloads & DEV_RX_OFFLOAD_JUMBO_FRAME) {
2848                 if (rxq->max_pkt_len <= I40E_ETH_MAX_LEN ||
2849                         rxq->max_pkt_len > I40E_FRAME_SIZE_MAX) {
2850                         PMD_DRV_LOG(ERR, "maximum packet length must "
2851                                     "be larger than %u and smaller than %u,"
2852                                     "as jumbo frame is enabled",
2853                                     (uint32_t)I40E_ETH_MAX_LEN,
2854                                     (uint32_t)I40E_FRAME_SIZE_MAX);
2855                         return I40E_ERR_CONFIG;
2856                 }
2857         } else {
2858                 if (rxq->max_pkt_len < RTE_ETHER_MIN_LEN ||
2859                         rxq->max_pkt_len > I40E_ETH_MAX_LEN) {
2860                         PMD_DRV_LOG(ERR, "maximum packet length must be "
2861                                     "larger than %u and smaller than %u, "
2862                                     "as jumbo frame is disabled",
2863                                     (uint32_t)RTE_ETHER_MIN_LEN,
2864                                     (uint32_t)I40E_ETH_MAX_LEN);
2865                         return I40E_ERR_CONFIG;
2866                 }
2867         }
2868
2869         return 0;
2870 }
2871
2872 /* Init the RX queue in hardware */
2873 int
2874 i40e_rx_queue_init(struct i40e_rx_queue *rxq)
2875 {
2876         int err = I40E_SUCCESS;
2877         struct i40e_hw *hw = I40E_VSI_TO_HW(rxq->vsi);
2878         struct rte_eth_dev_data *dev_data = I40E_VSI_TO_DEV_DATA(rxq->vsi);
2879         uint16_t pf_q = rxq->reg_idx;
2880         uint16_t buf_size;
2881         struct i40e_hmc_obj_rxq rx_ctx;
2882
2883         err = i40e_rx_queue_config(rxq);
2884         if (err < 0) {
2885                 PMD_DRV_LOG(ERR, "Failed to config RX queue");
2886                 return err;
2887         }
2888
2889         /* Clear the context structure first */
2890         memset(&rx_ctx, 0, sizeof(struct i40e_hmc_obj_rxq));
2891         rx_ctx.dbuff = rxq->rx_buf_len >> I40E_RXQ_CTX_DBUFF_SHIFT;
2892         rx_ctx.hbuff = rxq->rx_hdr_len >> I40E_RXQ_CTX_HBUFF_SHIFT;
2893
2894         rx_ctx.base = rxq->rx_ring_phys_addr / I40E_QUEUE_BASE_ADDR_UNIT;
2895         rx_ctx.qlen = rxq->nb_rx_desc;
2896 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
2897         rx_ctx.dsize = 1;
2898 #endif
2899         rx_ctx.dtype = rxq->hs_mode;
2900         if (rxq->hs_mode)
2901                 rx_ctx.hsplit_0 = I40E_HEADER_SPLIT_ALL;
2902         else
2903                 rx_ctx.hsplit_0 = I40E_HEADER_SPLIT_NONE;
2904         rx_ctx.rxmax = rxq->max_pkt_len;
2905         rx_ctx.tphrdesc_ena = 1;
2906         rx_ctx.tphwdesc_ena = 1;
2907         rx_ctx.tphdata_ena = 1;
2908         rx_ctx.tphhead_ena = 1;
2909         rx_ctx.lrxqthresh = 2;
2910         rx_ctx.crcstrip = (rxq->crc_len == 0) ? 1 : 0;
2911         rx_ctx.l2tsel = 1;
2912         /* showiv indicates if inner VLAN is stripped inside of tunnel
2913          * packet. When set it to 1, vlan information is stripped from
2914          * the inner header, but the hardware does not put it in the
2915          * descriptor. So set it zero by default.
2916          */
2917         rx_ctx.showiv = 0;
2918         rx_ctx.prefena = 1;
2919
2920         err = i40e_clear_lan_rx_queue_context(hw, pf_q);
2921         if (err != I40E_SUCCESS) {
2922                 PMD_DRV_LOG(ERR, "Failed to clear LAN RX queue context");
2923                 return err;
2924         }
2925         err = i40e_set_lan_rx_queue_context(hw, pf_q, &rx_ctx);
2926         if (err != I40E_SUCCESS) {
2927                 PMD_DRV_LOG(ERR, "Failed to set LAN RX queue context");
2928                 return err;
2929         }
2930
2931         rxq->qrx_tail = hw->hw_addr + I40E_QRX_TAIL(pf_q);
2932
2933         buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mp) -
2934                 RTE_PKTMBUF_HEADROOM);
2935
2936         /* Check if scattered RX needs to be used. */
2937         if (rxq->max_pkt_len > buf_size)
2938                 dev_data->scattered_rx = 1;
2939
2940         /* Init the RX tail regieter. */
2941         I40E_PCI_REG_WRITE(rxq->qrx_tail, rxq->nb_rx_desc - 1);
2942
2943         return 0;
2944 }
2945
2946 void
2947 i40e_dev_clear_queues(struct rte_eth_dev *dev)
2948 {
2949         uint16_t i;
2950
2951         PMD_INIT_FUNC_TRACE();
2952
2953         for (i = 0; i < dev->data->nb_tx_queues; i++) {
2954                 if (!dev->data->tx_queues[i])
2955                         continue;
2956                 i40e_tx_queue_release_mbufs(dev->data->tx_queues[i]);
2957                 i40e_reset_tx_queue(dev->data->tx_queues[i]);
2958         }
2959
2960         for (i = 0; i < dev->data->nb_rx_queues; i++) {
2961                 if (!dev->data->rx_queues[i])
2962                         continue;
2963                 i40e_rx_queue_release_mbufs(dev->data->rx_queues[i]);
2964                 i40e_reset_rx_queue(dev->data->rx_queues[i]);
2965         }
2966 }
2967
2968 void
2969 i40e_dev_free_queues(struct rte_eth_dev *dev)
2970 {
2971         uint16_t i;
2972
2973         PMD_INIT_FUNC_TRACE();
2974
2975         for (i = 0; i < dev->data->nb_rx_queues; i++) {
2976                 if (!dev->data->rx_queues[i])
2977                         continue;
2978                 i40e_dev_rx_queue_release(dev->data->rx_queues[i]);
2979                 dev->data->rx_queues[i] = NULL;
2980                 rte_eth_dma_zone_free(dev, "rx_ring", i);
2981         }
2982
2983         for (i = 0; i < dev->data->nb_tx_queues; i++) {
2984                 if (!dev->data->tx_queues[i])
2985                         continue;
2986                 i40e_dev_tx_queue_release(dev->data->tx_queues[i]);
2987                 dev->data->tx_queues[i] = NULL;
2988                 rte_eth_dma_zone_free(dev, "tx_ring", i);
2989         }
2990 }
2991
2992 enum i40e_status_code
2993 i40e_fdir_setup_tx_resources(struct i40e_pf *pf)
2994 {
2995         struct i40e_tx_queue *txq;
2996         const struct rte_memzone *tz = NULL;
2997         struct rte_eth_dev *dev;
2998         uint32_t ring_size;
2999
3000         if (!pf) {
3001                 PMD_DRV_LOG(ERR, "PF is not available");
3002                 return I40E_ERR_BAD_PTR;
3003         }
3004
3005         dev = pf->adapter->eth_dev;
3006
3007         /* Allocate the TX queue data structure. */
3008         txq = rte_zmalloc_socket("i40e fdir tx queue",
3009                                   sizeof(struct i40e_tx_queue),
3010                                   RTE_CACHE_LINE_SIZE,
3011                                   SOCKET_ID_ANY);
3012         if (!txq) {
3013                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
3014                                         "tx queue structure.");
3015                 return I40E_ERR_NO_MEMORY;
3016         }
3017
3018         /* Allocate TX hardware ring descriptors. */
3019         ring_size = sizeof(struct i40e_tx_desc) * I40E_FDIR_NUM_TX_DESC;
3020         ring_size = RTE_ALIGN(ring_size, I40E_DMA_MEM_ALIGN);
3021
3022         tz = rte_eth_dma_zone_reserve(dev, "fdir_tx_ring",
3023                                       I40E_FDIR_QUEUE_ID, ring_size,
3024                                       I40E_RING_BASE_ALIGN, SOCKET_ID_ANY);
3025         if (!tz) {
3026                 i40e_dev_tx_queue_release(txq);
3027                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for TX.");
3028                 return I40E_ERR_NO_MEMORY;
3029         }
3030
3031         txq->nb_tx_desc = I40E_FDIR_NUM_TX_DESC;
3032         txq->queue_id = I40E_FDIR_QUEUE_ID;
3033         txq->reg_idx = pf->fdir.fdir_vsi->base_queue;
3034         txq->vsi = pf->fdir.fdir_vsi;
3035
3036         txq->tx_ring_phys_addr = tz->iova;
3037         txq->tx_ring = (struct i40e_tx_desc *)tz->addr;
3038
3039         /*
3040          * don't need to allocate software ring and reset for the fdir
3041          * program queue just set the queue has been configured.
3042          */
3043         txq->q_set = TRUE;
3044         pf->fdir.txq = txq;
3045         pf->fdir.txq_available_buf_count = I40E_FDIR_PRG_PKT_CNT;
3046
3047         return I40E_SUCCESS;
3048 }
3049
3050 enum i40e_status_code
3051 i40e_fdir_setup_rx_resources(struct i40e_pf *pf)
3052 {
3053         struct i40e_rx_queue *rxq;
3054         const struct rte_memzone *rz = NULL;
3055         uint32_t ring_size;
3056         struct rte_eth_dev *dev;
3057
3058         if (!pf) {
3059                 PMD_DRV_LOG(ERR, "PF is not available");
3060                 return I40E_ERR_BAD_PTR;
3061         }
3062
3063         dev = pf->adapter->eth_dev;
3064
3065         /* Allocate the RX queue data structure. */
3066         rxq = rte_zmalloc_socket("i40e fdir rx queue",
3067                                   sizeof(struct i40e_rx_queue),
3068                                   RTE_CACHE_LINE_SIZE,
3069                                   SOCKET_ID_ANY);
3070         if (!rxq) {
3071                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
3072                                         "rx queue structure.");
3073                 return I40E_ERR_NO_MEMORY;
3074         }
3075
3076         /* Allocate RX hardware ring descriptors. */
3077         ring_size = sizeof(union i40e_rx_desc) * I40E_FDIR_NUM_RX_DESC;
3078         ring_size = RTE_ALIGN(ring_size, I40E_DMA_MEM_ALIGN);
3079
3080         rz = rte_eth_dma_zone_reserve(dev, "fdir_rx_ring",
3081                                       I40E_FDIR_QUEUE_ID, ring_size,
3082                                       I40E_RING_BASE_ALIGN, SOCKET_ID_ANY);
3083         if (!rz) {
3084                 i40e_dev_rx_queue_release(rxq);
3085                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for RX.");
3086                 return I40E_ERR_NO_MEMORY;
3087         }
3088
3089         rxq->nb_rx_desc = I40E_FDIR_NUM_RX_DESC;
3090         rxq->queue_id = I40E_FDIR_QUEUE_ID;
3091         rxq->reg_idx = pf->fdir.fdir_vsi->base_queue;
3092         rxq->vsi = pf->fdir.fdir_vsi;
3093
3094         rxq->rx_ring_phys_addr = rz->iova;
3095         memset(rz->addr, 0, I40E_FDIR_NUM_RX_DESC * sizeof(union i40e_rx_desc));
3096         rxq->rx_ring = (union i40e_rx_desc *)rz->addr;
3097
3098         /*
3099          * Don't need to allocate software ring and reset for the fdir
3100          * rx queue, just set the queue has been configured.
3101          */
3102         rxq->q_set = TRUE;
3103         pf->fdir.rxq = rxq;
3104
3105         return I40E_SUCCESS;
3106 }
3107
3108 void
3109 i40e_rxq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
3110         struct rte_eth_rxq_info *qinfo)
3111 {
3112         struct i40e_rx_queue *rxq;
3113
3114         rxq = dev->data->rx_queues[queue_id];
3115
3116         qinfo->mp = rxq->mp;
3117         qinfo->scattered_rx = dev->data->scattered_rx;
3118         qinfo->nb_desc = rxq->nb_rx_desc;
3119
3120         qinfo->conf.rx_free_thresh = rxq->rx_free_thresh;
3121         qinfo->conf.rx_drop_en = rxq->drop_en;
3122         qinfo->conf.rx_deferred_start = rxq->rx_deferred_start;
3123         qinfo->conf.offloads = rxq->offloads;
3124 }
3125
3126 void
3127 i40e_txq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
3128         struct rte_eth_txq_info *qinfo)
3129 {
3130         struct i40e_tx_queue *txq;
3131
3132         txq = dev->data->tx_queues[queue_id];
3133
3134         qinfo->nb_desc = txq->nb_tx_desc;
3135
3136         qinfo->conf.tx_thresh.pthresh = txq->pthresh;
3137         qinfo->conf.tx_thresh.hthresh = txq->hthresh;
3138         qinfo->conf.tx_thresh.wthresh = txq->wthresh;
3139
3140         qinfo->conf.tx_free_thresh = txq->tx_free_thresh;
3141         qinfo->conf.tx_rs_thresh = txq->tx_rs_thresh;
3142         qinfo->conf.tx_deferred_start = txq->tx_deferred_start;
3143         qinfo->conf.offloads = txq->offloads;
3144 }
3145
3146 static inline bool
3147 get_avx_supported(bool request_avx512)
3148 {
3149 #ifdef RTE_ARCH_X86
3150         if (request_avx512) {
3151                 if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_512 &&
3152                 rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1 &&
3153                 rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512BW) == 1)
3154 #ifdef CC_AVX512_SUPPORT
3155                         return true;
3156 #else
3157                 PMD_DRV_LOG(NOTICE,
3158                         "AVX512 is not supported in build env");
3159                 return false;
3160 #endif
3161         } else {
3162                 if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_256 &&
3163                 rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 &&
3164                 rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
3165 #ifdef CC_AVX2_SUPPORT
3166                         return true;
3167 #else
3168                 PMD_DRV_LOG(NOTICE,
3169                         "AVX2 is not supported in build env");
3170                 return false;
3171 #endif
3172         }
3173 #else
3174         RTE_SET_USED(request_avx512);
3175 #endif /* RTE_ARCH_X86 */
3176
3177         return false;
3178 }
3179
3180
3181 void __rte_cold
3182 i40e_set_rx_function(struct rte_eth_dev *dev)
3183 {
3184         struct i40e_adapter *ad =
3185                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3186         uint16_t rx_using_sse, i;
3187         bool use_avx2 = false;
3188         bool use_avx512 = false;
3189         /* In order to allow Vector Rx there are a few configuration
3190          * conditions to be met and Rx Bulk Allocation should be allowed.
3191          */
3192         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
3193                 if (i40e_rx_vec_dev_conf_condition_check(dev) ||
3194                     !ad->rx_bulk_alloc_allowed) {
3195                         PMD_INIT_LOG(DEBUG, "Port[%d] doesn't meet"
3196                                      " Vector Rx preconditions",
3197                                      dev->data->port_id);
3198
3199                         ad->rx_vec_allowed = false;
3200                 }
3201                 if (ad->rx_vec_allowed) {
3202                         for (i = 0; i < dev->data->nb_rx_queues; i++) {
3203                                 struct i40e_rx_queue *rxq =
3204                                         dev->data->rx_queues[i];
3205
3206                                 if (rxq && i40e_rxq_vec_setup(rxq)) {
3207                                         ad->rx_vec_allowed = false;
3208                                         break;
3209                                 }
3210                         }
3211
3212                         use_avx512 = get_avx_supported(1);
3213
3214                         if (!use_avx512)
3215                                 use_avx2 = get_avx_supported(0);
3216                 }
3217         }
3218
3219         if (ad->rx_vec_allowed  &&
3220                         rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
3221                 if (dev->data->scattered_rx) {
3222                         if (use_avx512) {
3223 #ifdef CC_AVX512_SUPPORT
3224                                 PMD_DRV_LOG(NOTICE,
3225                                         "Using AVX512 Vector Scattered Rx (port %d).",
3226                                         dev->data->port_id);
3227                                 dev->rx_pkt_burst =
3228                                         i40e_recv_scattered_pkts_vec_avx512;
3229 #endif
3230                         } else {
3231                                 PMD_INIT_LOG(DEBUG,
3232                                         "Using %sVector Scattered Rx (port %d).",
3233                                         use_avx2 ? "avx2 " : "",
3234                                         dev->data->port_id);
3235                                 dev->rx_pkt_burst = use_avx2 ?
3236                                         i40e_recv_scattered_pkts_vec_avx2 :
3237                                         i40e_recv_scattered_pkts_vec;
3238                         }
3239                 } else {
3240                         if (use_avx512) {
3241 #ifdef CC_AVX512_SUPPORT
3242                                 PMD_DRV_LOG(NOTICE,
3243                                         "Using AVX512 Vector Rx (port %d).",
3244                                         dev->data->port_id);
3245                                 dev->rx_pkt_burst =
3246                                         i40e_recv_pkts_vec_avx512;
3247 #endif
3248                         } else {
3249                                 PMD_INIT_LOG(DEBUG,
3250                                         "Using %sVector Rx (port %d).",
3251                                         use_avx2 ? "avx2 " : "",
3252                                         dev->data->port_id);
3253                                 dev->rx_pkt_burst = use_avx2 ?
3254                                         i40e_recv_pkts_vec_avx2 :
3255                                         i40e_recv_pkts_vec;
3256                         }
3257                 }
3258         } else if (!dev->data->scattered_rx && ad->rx_bulk_alloc_allowed) {
3259                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are "
3260                                     "satisfied. Rx Burst Bulk Alloc function "
3261                                     "will be used on port=%d.",
3262                              dev->data->port_id);
3263
3264                 dev->rx_pkt_burst = i40e_recv_pkts_bulk_alloc;
3265         } else {
3266                 /* Simple Rx Path. */
3267                 PMD_INIT_LOG(DEBUG, "Simple Rx path will be used on port=%d.",
3268                              dev->data->port_id);
3269                 dev->rx_pkt_burst = dev->data->scattered_rx ?
3270                                         i40e_recv_scattered_pkts :
3271                                         i40e_recv_pkts;
3272         }
3273
3274         /* Propagate information about RX function choice through all queues. */
3275         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
3276                 rx_using_sse =
3277                         (dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec ||
3278                          dev->rx_pkt_burst == i40e_recv_pkts_vec ||
3279 #ifdef CC_AVX512_SUPPORT
3280                          dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec_avx512 ||
3281                          dev->rx_pkt_burst == i40e_recv_pkts_vec_avx512 ||
3282 #endif
3283                          dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec_avx2 ||
3284                          dev->rx_pkt_burst == i40e_recv_pkts_vec_avx2);
3285
3286                 for (i = 0; i < dev->data->nb_rx_queues; i++) {
3287                         struct i40e_rx_queue *rxq = dev->data->rx_queues[i];
3288
3289                         if (rxq)
3290                                 rxq->rx_using_sse = rx_using_sse;
3291                 }
3292         }
3293 }
3294
3295 static const struct {
3296         eth_rx_burst_t pkt_burst;
3297         const char *info;
3298 } i40e_rx_burst_infos[] = {
3299         { i40e_recv_scattered_pkts,          "Scalar Scattered" },
3300         { i40e_recv_pkts_bulk_alloc,         "Scalar Bulk Alloc" },
3301         { i40e_recv_pkts,                    "Scalar" },
3302 #ifdef RTE_ARCH_X86
3303 #ifdef CC_AVX512_SUPPORT
3304         { i40e_recv_scattered_pkts_vec_avx512, "Vector AVX512 Scattered" },
3305         { i40e_recv_pkts_vec_avx512,           "Vector AVX512" },
3306 #endif
3307         { i40e_recv_scattered_pkts_vec_avx2, "Vector AVX2 Scattered" },
3308         { i40e_recv_pkts_vec_avx2,           "Vector AVX2" },
3309         { i40e_recv_scattered_pkts_vec,      "Vector SSE Scattered" },
3310         { i40e_recv_pkts_vec,                "Vector SSE" },
3311 #elif defined(RTE_ARCH_ARM64)
3312         { i40e_recv_scattered_pkts_vec,      "Vector Neon Scattered" },
3313         { i40e_recv_pkts_vec,                "Vector Neon" },
3314 #elif defined(RTE_ARCH_PPC_64)
3315         { i40e_recv_scattered_pkts_vec,      "Vector AltiVec Scattered" },
3316         { i40e_recv_pkts_vec,                "Vector AltiVec" },
3317 #endif
3318 };
3319
3320 int
3321 i40e_rx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id,
3322                        struct rte_eth_burst_mode *mode)
3323 {
3324         eth_rx_burst_t pkt_burst = dev->rx_pkt_burst;
3325         int ret = -EINVAL;
3326         unsigned int i;
3327
3328         for (i = 0; i < RTE_DIM(i40e_rx_burst_infos); ++i) {
3329                 if (pkt_burst == i40e_rx_burst_infos[i].pkt_burst) {
3330                         snprintf(mode->info, sizeof(mode->info), "%s",
3331                                  i40e_rx_burst_infos[i].info);
3332                         ret = 0;
3333                         break;
3334                 }
3335         }
3336
3337         return ret;
3338 }
3339
3340 void __rte_cold
3341 i40e_set_tx_function_flag(struct rte_eth_dev *dev, struct i40e_tx_queue *txq)
3342 {
3343         struct i40e_adapter *ad =
3344                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3345
3346         /* Use a simple Tx queue if possible (only fast free is allowed) */
3347         ad->tx_simple_allowed =
3348                 (txq->offloads ==
3349                  (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE) &&
3350                  txq->tx_rs_thresh >= RTE_PMD_I40E_TX_MAX_BURST);
3351         ad->tx_vec_allowed = (ad->tx_simple_allowed &&
3352                         txq->tx_rs_thresh <= RTE_I40E_TX_MAX_FREE_BUF_SZ);
3353
3354         if (ad->tx_vec_allowed)
3355                 PMD_INIT_LOG(DEBUG, "Vector Tx can be enabled on Tx queue %u.",
3356                                 txq->queue_id);
3357         else if (ad->tx_simple_allowed)
3358                 PMD_INIT_LOG(DEBUG, "Simple Tx can be enabled on Tx queue %u.",
3359                                 txq->queue_id);
3360         else
3361                 PMD_INIT_LOG(DEBUG,
3362                                 "Neither simple nor vector Tx enabled on Tx queue %u\n",
3363                                 txq->queue_id);
3364 }
3365
3366 void __rte_cold
3367 i40e_set_tx_function(struct rte_eth_dev *dev)
3368 {
3369         struct i40e_adapter *ad =
3370                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3371         int i;
3372         bool use_avx2 = false;
3373         bool use_avx512 = false;
3374
3375         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
3376                 if (ad->tx_vec_allowed) {
3377                         for (i = 0; i < dev->data->nb_tx_queues; i++) {
3378                                 struct i40e_tx_queue *txq =
3379                                         dev->data->tx_queues[i];
3380
3381                                 if (txq && i40e_txq_vec_setup(txq)) {
3382                                         ad->tx_vec_allowed = false;
3383                                         break;
3384                                 }
3385                         }
3386
3387                         use_avx512 = get_avx_supported(1);
3388
3389                         if (!use_avx512)
3390                                 use_avx2 = get_avx_supported(0);
3391                 }
3392         }
3393
3394         if (ad->tx_simple_allowed) {
3395                 if (ad->tx_vec_allowed &&
3396                                 rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
3397                         if (use_avx512) {
3398 #ifdef CC_AVX512_SUPPORT
3399                                 PMD_DRV_LOG(NOTICE, "Using AVX512 Vector Tx (port %d).",
3400                                             dev->data->port_id);
3401                                 dev->tx_pkt_burst = i40e_xmit_pkts_vec_avx512;
3402 #endif
3403                         } else {
3404                                 PMD_INIT_LOG(DEBUG, "Using %sVector Tx (port %d).",
3405                                              use_avx2 ? "avx2 " : "",
3406                                              dev->data->port_id);
3407                                 dev->tx_pkt_burst = use_avx2 ?
3408                                                     i40e_xmit_pkts_vec_avx2 :
3409                                                     i40e_xmit_pkts_vec;
3410                         }
3411                 } else {
3412                         PMD_INIT_LOG(DEBUG, "Simple tx finally be used.");
3413                         dev->tx_pkt_burst = i40e_xmit_pkts_simple;
3414                 }
3415                 dev->tx_pkt_prepare = NULL;
3416         } else {
3417                 PMD_INIT_LOG(DEBUG, "Xmit tx finally be used.");
3418                 dev->tx_pkt_burst = i40e_xmit_pkts;
3419                 dev->tx_pkt_prepare = i40e_prep_pkts;
3420         }
3421 }
3422
3423 static const struct {
3424         eth_tx_burst_t pkt_burst;
3425         const char *info;
3426 } i40e_tx_burst_infos[] = {
3427         { i40e_xmit_pkts_simple,   "Scalar Simple" },
3428         { i40e_xmit_pkts,          "Scalar" },
3429 #ifdef RTE_ARCH_X86
3430 #ifdef CC_AVX512_SUPPORT
3431         { i40e_xmit_pkts_vec_avx512, "Vector AVX512" },
3432 #endif
3433         { i40e_xmit_pkts_vec_avx2, "Vector AVX2" },
3434         { i40e_xmit_pkts_vec,      "Vector SSE" },
3435 #elif defined(RTE_ARCH_ARM64)
3436         { i40e_xmit_pkts_vec,      "Vector Neon" },
3437 #elif defined(RTE_ARCH_PPC_64)
3438         { i40e_xmit_pkts_vec,      "Vector AltiVec" },
3439 #endif
3440 };
3441
3442 int
3443 i40e_tx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id,
3444                        struct rte_eth_burst_mode *mode)
3445 {
3446         eth_tx_burst_t pkt_burst = dev->tx_pkt_burst;
3447         int ret = -EINVAL;
3448         unsigned int i;
3449
3450         for (i = 0; i < RTE_DIM(i40e_tx_burst_infos); ++i) {
3451                 if (pkt_burst == i40e_tx_burst_infos[i].pkt_burst) {
3452                         snprintf(mode->info, sizeof(mode->info), "%s",
3453                                  i40e_tx_burst_infos[i].info);
3454                         ret = 0;
3455                         break;
3456                 }
3457         }
3458
3459         return ret;
3460 }
3461
3462 void __rte_cold
3463 i40e_set_default_ptype_table(struct rte_eth_dev *dev)
3464 {
3465         struct i40e_adapter *ad =
3466                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3467         int i;
3468
3469         for (i = 0; i < I40E_MAX_PKT_TYPE; i++)
3470                 ad->ptype_tbl[i] = i40e_get_default_pkt_type(i);
3471 }
3472
3473 void __rte_cold
3474 i40e_set_default_pctype_table(struct rte_eth_dev *dev)
3475 {
3476         struct i40e_adapter *ad =
3477                         I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3478         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3479         int i;
3480
3481         for (i = 0; i < I40E_FLOW_TYPE_MAX; i++)
3482                 ad->pctypes_tbl[i] = 0ULL;
3483         ad->flow_types_mask = 0ULL;
3484         ad->pctypes_mask = 0ULL;
3485
3486         ad->pctypes_tbl[RTE_ETH_FLOW_FRAG_IPV4] =
3487                                 (1ULL << I40E_FILTER_PCTYPE_FRAG_IPV4);
3488         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_UDP] =
3489                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_UDP);
3490         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_TCP] =
3491                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_TCP);
3492         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_SCTP] =
3493                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_SCTP);
3494         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_OTHER] =
3495                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_OTHER);
3496         ad->pctypes_tbl[RTE_ETH_FLOW_FRAG_IPV6] =
3497                                 (1ULL << I40E_FILTER_PCTYPE_FRAG_IPV6);
3498         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_UDP] =
3499                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_UDP);
3500         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_TCP] =
3501                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_TCP);
3502         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_SCTP] =
3503                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_SCTP);
3504         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_OTHER] =
3505                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_OTHER);
3506         ad->pctypes_tbl[RTE_ETH_FLOW_L2_PAYLOAD] =
3507                                 (1ULL << I40E_FILTER_PCTYPE_L2_PAYLOAD);
3508
3509         if (hw->mac.type == I40E_MAC_X722 ||
3510                 hw->mac.type == I40E_MAC_X722_VF) {
3511                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_UDP] |=
3512                         (1ULL << I40E_FILTER_PCTYPE_NONF_UNICAST_IPV4_UDP);
3513                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_UDP] |=
3514                         (1ULL << I40E_FILTER_PCTYPE_NONF_MULTICAST_IPV4_UDP);
3515                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_TCP] |=
3516                         (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_TCP_SYN_NO_ACK);
3517                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_UDP] |=
3518                         (1ULL << I40E_FILTER_PCTYPE_NONF_UNICAST_IPV6_UDP);
3519                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_UDP] |=
3520                         (1ULL << I40E_FILTER_PCTYPE_NONF_MULTICAST_IPV6_UDP);
3521                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_TCP] |=
3522                         (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_TCP_SYN_NO_ACK);
3523         }
3524
3525         for (i = 0; i < I40E_FLOW_TYPE_MAX; i++) {
3526                 if (ad->pctypes_tbl[i])
3527                         ad->flow_types_mask |= (1ULL << i);
3528                 ad->pctypes_mask |= ad->pctypes_tbl[i];
3529         }
3530 }
3531
3532 #ifndef CC_AVX2_SUPPORT
3533 uint16_t
3534 i40e_recv_pkts_vec_avx2(void __rte_unused *rx_queue,
3535                         struct rte_mbuf __rte_unused **rx_pkts,
3536                         uint16_t __rte_unused nb_pkts)
3537 {
3538         return 0;
3539 }
3540
3541 uint16_t
3542 i40e_recv_scattered_pkts_vec_avx2(void __rte_unused *rx_queue,
3543                         struct rte_mbuf __rte_unused **rx_pkts,
3544                         uint16_t __rte_unused nb_pkts)
3545 {
3546         return 0;
3547 }
3548
3549 uint16_t
3550 i40e_xmit_pkts_vec_avx2(void __rte_unused * tx_queue,
3551                           struct rte_mbuf __rte_unused **tx_pkts,
3552                           uint16_t __rte_unused nb_pkts)
3553 {
3554         return 0;
3555 }
3556 #endif /* ifndef CC_AVX2_SUPPORT */