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 #define I40E_TX_OFFLOAD_SIMPLE_SUP_MASK ( \
  76                 PKT_TX_IPV4 | \
  77                 PKT_TX_IPV6 | \
  78                 PKT_TX_OUTER_IPV4 | \
  79                 PKT_TX_OUTER_IPV6)
  80
  81 #define I40E_TX_OFFLOAD_SIMPLE_NOTSUP_MASK \
  82                 (PKT_TX_OFFLOAD_MASK ^ I40E_TX_OFFLOAD_SIMPLE_SUP_MASK)
  83
  84 static int
  85 i40e_monitor_callback(const uint64_t value,
  86                 const uint64_t arg[RTE_POWER_MONITOR_OPAQUE_SZ] __rte_unused)
  87 {
  88         const uint64_t m = rte_cpu_to_le_64(1 << I40E_RX_DESC_STATUS_DD_SHIFT);
  89         /*
  90          * we expect the DD bit to be set to 1 if this descriptor was already
  91          * written to.
  92          */
  93         return (value & m) == m ? -1 : 0;
  94 }
  95
  96 int
  97 i40e_get_monitor_addr(void *rx_queue, struct rte_power_monitor_cond *pmc)
  98 {
  99         struct i40e_rx_queue *rxq = rx_queue;
 100         volatile union i40e_rx_desc *rxdp;
 101         uint16_t desc;
 102
 103         desc = rxq->rx_tail;
 104         rxdp = &rxq->rx_ring[desc];
 105         /* watch for changes in status bit */
 106         pmc->addr = &rxdp->wb.qword1.status_error_len;
 107
 108         /* comparison callback */
 109         pmc->fn = i40e_monitor_callback;
 110
 111         /* registers are 64-bit */
 112         pmc->size = sizeof(uint64_t);
 113
 114         return 0;
 115 }
 116
 117 static inline void
 118 i40e_rxd_to_vlan_tci(struct rte_mbuf *mb, volatile union i40e_rx_desc *rxdp)
 119 {
 120         if (rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len) &
 121                 (1 << I40E_RX_DESC_STATUS_L2TAG1P_SHIFT)) {
 122                 mb->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
 123                 mb->vlan_tci =
 124                         rte_le_to_cpu_16(rxdp->wb.qword0.lo_dword.l2tag1);
 125                 PMD_RX_LOG(DEBUG, "Descriptor l2tag1: %u",
 126                            rte_le_to_cpu_16(rxdp->wb.qword0.lo_dword.l2tag1));
 127         } else {
 128                 mb->vlan_tci = 0;
 129         }
 130 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
 131         if (rte_le_to_cpu_16(rxdp->wb.qword2.ext_status) &
 132                 (1 << I40E_RX_DESC_EXT_STATUS_L2TAG2P_SHIFT)) {
 133                 mb->ol_flags |= PKT_RX_QINQ_STRIPPED | PKT_RX_QINQ |
 134                         PKT_RX_VLAN_STRIPPED | PKT_RX_VLAN;
 135                 mb->vlan_tci_outer = mb->vlan_tci;
 136                 mb->vlan_tci = rte_le_to_cpu_16(rxdp->wb.qword2.l2tag2_2);
 137                 PMD_RX_LOG(DEBUG, "Descriptor l2tag2_1: %u, l2tag2_2: %u",
 138                            rte_le_to_cpu_16(rxdp->wb.qword2.l2tag2_1),
 139                            rte_le_to_cpu_16(rxdp->wb.qword2.l2tag2_2));
 140         } else {
 141                 mb->vlan_tci_outer = 0;
 142         }
 143 #endif
 144         PMD_RX_LOG(DEBUG, "Mbuf vlan_tci: %u, vlan_tci_outer: %u",
 145                    mb->vlan_tci, mb->vlan_tci_outer);
 146 }
 147
 148 /* Translate the rx descriptor status to pkt flags */
 149 static inline uint64_t
 150 i40e_rxd_status_to_pkt_flags(uint64_t qword)
 151 {
 152         uint64_t flags;
 153
 154         /* Check if RSS_HASH */
 155         flags = (((qword >> I40E_RX_DESC_STATUS_FLTSTAT_SHIFT) &
 156                                         I40E_RX_DESC_FLTSTAT_RSS_HASH) ==
 157                         I40E_RX_DESC_FLTSTAT_RSS_HASH) ? PKT_RX_RSS_HASH : 0;
 158
 159         /* Check if FDIR Match */
 160         flags |= (qword & (1 << I40E_RX_DESC_STATUS_FLM_SHIFT) ?
 161                                                         PKT_RX_FDIR : 0);
 162
 163         return flags;
 164 }
 165
 166 static inline uint64_t
 167 i40e_rxd_error_to_pkt_flags(uint64_t qword)
 168 {
 169         uint64_t flags = 0;
 170         uint64_t error_bits = (qword >> I40E_RXD_QW1_ERROR_SHIFT);
 171
 172 #define I40E_RX_ERR_BITS 0x3f
 173         if (likely((error_bits & I40E_RX_ERR_BITS) == 0)) {
 174                 flags |= (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD);
 175                 return flags;
 176         }
 177
 178         if (unlikely(error_bits & (1 << I40E_RX_DESC_ERROR_IPE_SHIFT)))
 179                 flags |= PKT_RX_IP_CKSUM_BAD;
 180         else
 181                 flags |= PKT_RX_IP_CKSUM_GOOD;
 182
 183         if (unlikely(error_bits & (1 << I40E_RX_DESC_ERROR_L4E_SHIFT)))
 184                 flags |= PKT_RX_L4_CKSUM_BAD;
 185         else
 186                 flags |= PKT_RX_L4_CKSUM_GOOD;
 187
 188         if (unlikely(error_bits & (1 << I40E_RX_DESC_ERROR_EIPE_SHIFT)))
 189                 flags |= PKT_RX_OUTER_IP_CKSUM_BAD;
 190
 191         return flags;
 192 }
 193
 194 /* Function to check and set the ieee1588 timesync index and get the
 195  * appropriate flags.
 196  */
 197 #ifdef RTE_LIBRTE_IEEE1588
 198 static inline uint64_t
 199 i40e_get_iee15888_flags(struct rte_mbuf *mb, uint64_t qword)
 200 {
 201         uint64_t pkt_flags = 0;
 202         uint16_t tsyn = (qword & (I40E_RXD_QW1_STATUS_TSYNVALID_MASK
 203                                   | I40E_RXD_QW1_STATUS_TSYNINDX_MASK))
 204                                     >> I40E_RX_DESC_STATUS_TSYNINDX_SHIFT;
 205
 206         if ((mb->packet_type & RTE_PTYPE_L2_MASK)
 207                         == RTE_PTYPE_L2_ETHER_TIMESYNC)
 208                 pkt_flags = PKT_RX_IEEE1588_PTP;
 209         if (tsyn & 0x04) {
 210                 pkt_flags |= PKT_RX_IEEE1588_TMST;
 211                 mb->timesync = tsyn & 0x03;
 212         }
 213
 214         return pkt_flags;
 215 }
 216 #endif
 217
 218 static inline uint64_t
 219 i40e_rxd_build_fdir(volatile union i40e_rx_desc *rxdp, struct rte_mbuf *mb)
 220 {
 221         uint64_t flags = 0;
 222 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
 223         uint16_t flexbh, flexbl;
 224
 225         flexbh = (rte_le_to_cpu_32(rxdp->wb.qword2.ext_status) >>
 226                 I40E_RX_DESC_EXT_STATUS_FLEXBH_SHIFT) &
 227                 I40E_RX_DESC_EXT_STATUS_FLEXBH_MASK;
 228         flexbl = (rte_le_to_cpu_32(rxdp->wb.qword2.ext_status) >>
 229                 I40E_RX_DESC_EXT_STATUS_FLEXBL_SHIFT) &
 230                 I40E_RX_DESC_EXT_STATUS_FLEXBL_MASK;
 231
 232
 233         if (flexbh == I40E_RX_DESC_EXT_STATUS_FLEXBH_FD_ID) {
 234                 mb->hash.fdir.hi =
 235                         rte_le_to_cpu_32(rxdp->wb.qword3.hi_dword.fd_id);
 236                 flags |= PKT_RX_FDIR_ID;
 237         } else if (flexbh == I40E_RX_DESC_EXT_STATUS_FLEXBH_FLEX) {
 238                 mb->hash.fdir.hi =
 239                         rte_le_to_cpu_32(rxdp->wb.qword3.hi_dword.flex_bytes_hi);
 240                 flags |= PKT_RX_FDIR_FLX;
 241         }
 242         if (flexbl == I40E_RX_DESC_EXT_STATUS_FLEXBL_FLEX) {
 243                 mb->hash.fdir.lo =
 244                         rte_le_to_cpu_32(rxdp->wb.qword3.lo_dword.flex_bytes_lo);
 245                 flags |= PKT_RX_FDIR_FLX;
 246         }
 247 #else
 248         mb->hash.fdir.hi =
 249                 rte_le_to_cpu_32(rxdp->wb.qword0.hi_dword.fd_id);
 250         flags |= PKT_RX_FDIR_ID;
 251 #endif
 252         return flags;
 253 }
 254
 255 static inline void
 256 i40e_parse_tunneling_params(uint64_t ol_flags,
 257                             union i40e_tx_offload tx_offload,
 258                             uint32_t *cd_tunneling)
 259 {
 260         /* EIPT: External (outer) IP header type */
 261         if (ol_flags & PKT_TX_OUTER_IP_CKSUM)
 262                 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV4;
 263         else if (ol_flags & PKT_TX_OUTER_IPV4)
 264                 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV4_NO_CSUM;
 265         else if (ol_flags & PKT_TX_OUTER_IPV6)
 266                 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV6;
 267
 268         /* EIPLEN: External (outer) IP header length, in DWords */
 269         *cd_tunneling |= (tx_offload.outer_l3_len >> 2) <<
 270                 I40E_TXD_CTX_QW0_EXT_IPLEN_SHIFT;
 271
 272         /* L4TUNT: L4 Tunneling Type */
 273         switch (ol_flags & PKT_TX_TUNNEL_MASK) {
 274         case PKT_TX_TUNNEL_IPIP:
 275                 /* for non UDP / GRE tunneling, set to 00b */
 276                 break;
 277         case PKT_TX_TUNNEL_VXLAN:
 278         case PKT_TX_TUNNEL_GENEVE:
 279                 *cd_tunneling |= I40E_TXD_CTX_UDP_TUNNELING;
 280                 break;
 281         case PKT_TX_TUNNEL_GRE:
 282                 *cd_tunneling |= I40E_TXD_CTX_GRE_TUNNELING;
 283                 break;
 284         default:
 285                 PMD_TX_LOG(ERR, "Tunnel type not supported");
 286                 return;
 287         }
 288
 289         /* L4TUNLEN: L4 Tunneling Length, in Words
 290          *
 291          * We depend on app to set rte_mbuf.l2_len correctly.
 292          * For IP in GRE it should be set to the length of the GRE
 293          * header;
 294          * for MAC in GRE or MAC in UDP it should be set to the length
 295          * of the GRE or UDP headers plus the inner MAC up to including
 296          * its last Ethertype.
 297          */
 298         *cd_tunneling |= (tx_offload.l2_len >> 1) <<
 299                 I40E_TXD_CTX_QW0_NATLEN_SHIFT;
 300 }
 301
 302 static inline void
 303 i40e_txd_enable_checksum(uint64_t ol_flags,
 304                         uint32_t *td_cmd,
 305                         uint32_t *td_offset,
 306                         union i40e_tx_offload tx_offload)
 307 {
 308         /* Set MACLEN */
 309         if (ol_flags & PKT_TX_TUNNEL_MASK)
 310                 *td_offset |= (tx_offload.outer_l2_len >> 1)
 311                                 << I40E_TX_DESC_LENGTH_MACLEN_SHIFT;
 312         else
 313                 *td_offset |= (tx_offload.l2_len >> 1)
 314                         << I40E_TX_DESC_LENGTH_MACLEN_SHIFT;
 315
 316         /* Enable L3 checksum offloads */
 317         if (ol_flags & PKT_TX_IP_CKSUM) {
 318                 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV4_CSUM;
 319                 *td_offset |= (tx_offload.l3_len >> 2)
 320                                 << I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
 321         } else if (ol_flags & PKT_TX_IPV4) {
 322                 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV4;
 323                 *td_offset |= (tx_offload.l3_len >> 2)
 324                                 << I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
 325         } else if (ol_flags & PKT_TX_IPV6) {
 326                 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV6;
 327                 *td_offset |= (tx_offload.l3_len >> 2)
 328                                 << I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
 329         }
 330
 331         if (ol_flags & PKT_TX_TCP_SEG) {
 332                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_TCP;
 333                 *td_offset |= (tx_offload.l4_len >> 2)
 334                         << I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
 335                 return;
 336         }
 337
 338         /* Enable L4 checksum offloads */
 339         switch (ol_flags & PKT_TX_L4_MASK) {
 340         case PKT_TX_TCP_CKSUM:
 341                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_TCP;
 342                 *td_offset |= (sizeof(struct rte_tcp_hdr) >> 2) <<
 343                                 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
 344                 break;
 345         case PKT_TX_SCTP_CKSUM:
 346                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_SCTP;
 347                 *td_offset |= (sizeof(struct rte_sctp_hdr) >> 2) <<
 348                                 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
 349                 break;
 350         case PKT_TX_UDP_CKSUM:
 351                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_UDP;
 352                 *td_offset |= (sizeof(struct rte_udp_hdr) >> 2) <<
 353                                 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
 354                 break;
 355         default:
 356                 break;
 357         }
 358 }
 359
 360 /* Construct the tx flags */
 361 static inline uint64_t
 362 i40e_build_ctob(uint32_t td_cmd,
 363                 uint32_t td_offset,
 364                 unsigned int size,
 365                 uint32_t td_tag)
 366 {
 367         return rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DATA |
 368                         ((uint64_t)td_cmd  << I40E_TXD_QW1_CMD_SHIFT) |
 369                         ((uint64_t)td_offset << I40E_TXD_QW1_OFFSET_SHIFT) |
 370                         ((uint64_t)size  << I40E_TXD_QW1_TX_BUF_SZ_SHIFT) |
 371                         ((uint64_t)td_tag  << I40E_TXD_QW1_L2TAG1_SHIFT));
 372 }
 373
 374 static inline int
 375 i40e_xmit_cleanup(struct i40e_tx_queue *txq)
 376 {
 377         struct i40e_tx_entry *sw_ring = txq->sw_ring;
 378         volatile struct i40e_tx_desc *txd = txq->tx_ring;
 379         uint16_t last_desc_cleaned = txq->last_desc_cleaned;
 380         uint16_t nb_tx_desc = txq->nb_tx_desc;
 381         uint16_t desc_to_clean_to;
 382         uint16_t nb_tx_to_clean;
 383
 384         desc_to_clean_to = (uint16_t)(last_desc_cleaned + txq->tx_rs_thresh);
 385         if (desc_to_clean_to >= nb_tx_desc)
 386                 desc_to_clean_to = (uint16_t)(desc_to_clean_to - nb_tx_desc);
 387
 388         desc_to_clean_to = sw_ring[desc_to_clean_to].last_id;
 389         if ((txd[desc_to_clean_to].cmd_type_offset_bsz &
 390                         rte_cpu_to_le_64(I40E_TXD_QW1_DTYPE_MASK)) !=
 391                         rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE)) {
 392                 PMD_TX_LOG(DEBUG, "TX descriptor %4u is not done "
 393                            "(port=%d queue=%d)", desc_to_clean_to,
 394                            txq->port_id, txq->queue_id);
 395                 return -1;
 396         }
 397
 398         if (last_desc_cleaned > desc_to_clean_to)
 399                 nb_tx_to_clean = (uint16_t)((nb_tx_desc - last_desc_cleaned) +
 400                                                         desc_to_clean_to);
 401         else
 402                 nb_tx_to_clean = (uint16_t)(desc_to_clean_to -
 403                                         last_desc_cleaned);
 404
 405         txd[desc_to_clean_to].cmd_type_offset_bsz = 0;
 406
 407         txq->last_desc_cleaned = desc_to_clean_to;
 408         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + nb_tx_to_clean);
 409
 410         return 0;
 411 }
 412
 413 static inline int
 414 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
 415 check_rx_burst_bulk_alloc_preconditions(struct i40e_rx_queue *rxq)
 416 #else
 417 check_rx_burst_bulk_alloc_preconditions(__rte_unused struct i40e_rx_queue *rxq)
 418 #endif
 419 {
 420         int ret = 0;
 421
 422 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
 423         if (!(rxq->rx_free_thresh >= RTE_PMD_I40E_RX_MAX_BURST)) {
 424                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
 425                              "rxq->rx_free_thresh=%d, "
 426                              "RTE_PMD_I40E_RX_MAX_BURST=%d",
 427                              rxq->rx_free_thresh, RTE_PMD_I40E_RX_MAX_BURST);
 428                 ret = -EINVAL;
 429         } else if (!(rxq->rx_free_thresh < rxq->nb_rx_desc)) {
 430                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
 431                              "rxq->rx_free_thresh=%d, "
 432                              "rxq->nb_rx_desc=%d",
 433                              rxq->rx_free_thresh, rxq->nb_rx_desc);
 434                 ret = -EINVAL;
 435         } else if (rxq->nb_rx_desc % rxq->rx_free_thresh != 0) {
 436                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
 437                              "rxq->nb_rx_desc=%d, "
 438                              "rxq->rx_free_thresh=%d",
 439                              rxq->nb_rx_desc, rxq->rx_free_thresh);
 440                 ret = -EINVAL;
 441         }
 442 #else
 443         ret = -EINVAL;
 444 #endif
 445
 446         return ret;
 447 }
 448
 449 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
 450 #define I40E_LOOK_AHEAD 8
 451 #if (I40E_LOOK_AHEAD != 8)
 452 #error "PMD I40E: I40E_LOOK_AHEAD must be 8\n"
 453 #endif
 454 static inline int
 455 i40e_rx_scan_hw_ring(struct i40e_rx_queue *rxq)
 456 {
 457         volatile union i40e_rx_desc *rxdp;
 458         struct i40e_rx_entry *rxep;
 459         struct rte_mbuf *mb;
 460         uint16_t pkt_len;
 461         uint64_t qword1;
 462         uint32_t rx_status;
 463         int32_t s[I40E_LOOK_AHEAD], var, nb_dd;
 464         int32_t i, j, nb_rx = 0;
 465         uint64_t pkt_flags;
 466         uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
 467
 468         rxdp = &rxq->rx_ring[rxq->rx_tail];
 469         rxep = &rxq->sw_ring[rxq->rx_tail];
 470
 471         qword1 = rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len);
 472         rx_status = (qword1 & I40E_RXD_QW1_STATUS_MASK) >>
 473                                 I40E_RXD_QW1_STATUS_SHIFT;
 474
 475         /* Make sure there is at least 1 packet to receive */
 476         if (!(rx_status & (1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
 477                 return 0;
 478
 479         /**
 480          * Scan LOOK_AHEAD descriptors at a time to determine which
 481          * descriptors reference packets that are ready to be received.
 482          */
 483         for (i = 0; i < RTE_PMD_I40E_RX_MAX_BURST; i+=I40E_LOOK_AHEAD,
 484                         rxdp += I40E_LOOK_AHEAD, rxep += I40E_LOOK_AHEAD) {
 485                 /* Read desc statuses backwards to avoid race condition */
 486                 for (j = I40E_LOOK_AHEAD - 1; j >= 0; j--) {
 487                         qword1 = rte_le_to_cpu_64(\
 488                                 rxdp[j].wb.qword1.status_error_len);
 489                         s[j] = (qword1 & I40E_RXD_QW1_STATUS_MASK) >>
 490                                         I40E_RXD_QW1_STATUS_SHIFT;
 491                 }
 492
 493                 /* This barrier is to order loads of different words in the descriptor */
 494                 rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
 495
 496                 /* Compute how many status bits were set */
 497                 for (j = 0, nb_dd = 0; j < I40E_LOOK_AHEAD; j++) {
 498                         var = s[j] & (1 << I40E_RX_DESC_STATUS_DD_SHIFT);
 499 #ifdef RTE_ARCH_ARM
 500                         /* For Arm platforms, only compute continuous status bits */
 501                         if (var)
 502                                 nb_dd += 1;
 503                         else
 504                                 break;
 505 #else
 506                         nb_dd += var;
 507 #endif
 508                 }
 509
 510                 nb_rx += nb_dd;
 511
 512                 /* Translate descriptor info to mbuf parameters */
 513                 for (j = 0; j < nb_dd; j++) {
 514                         mb = rxep[j].mbuf;
 515                         qword1 = rte_le_to_cpu_64(\
 516                                 rxdp[j].wb.qword1.status_error_len);
 517                         pkt_len = ((qword1 & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
 518                                 I40E_RXD_QW1_LENGTH_PBUF_SHIFT) - rxq->crc_len;
 519                         mb->data_len = pkt_len;
 520                         mb->pkt_len = pkt_len;
 521                         mb->ol_flags = 0;
 522                         i40e_rxd_to_vlan_tci(mb, &rxdp[j]);
 523                         pkt_flags = i40e_rxd_status_to_pkt_flags(qword1);
 524                         pkt_flags |= i40e_rxd_error_to_pkt_flags(qword1);
 525                         mb->packet_type =
 526                                 ptype_tbl[(uint8_t)((qword1 &
 527                                 I40E_RXD_QW1_PTYPE_MASK) >>
 528                                 I40E_RXD_QW1_PTYPE_SHIFT)];
 529                         if (pkt_flags & PKT_RX_RSS_HASH)
 530                                 mb->hash.rss = rte_le_to_cpu_32(\
 531                                         rxdp[j].wb.qword0.hi_dword.rss);
 532                         if (pkt_flags & PKT_RX_FDIR)
 533                                 pkt_flags |= i40e_rxd_build_fdir(&rxdp[j], mb);
 534
 535 #ifdef RTE_LIBRTE_IEEE1588
 536                         pkt_flags |= i40e_get_iee15888_flags(mb, qword1);
 537 #endif
 538                         mb->ol_flags |= pkt_flags;
 539
 540                 }
 541
 542                 for (j = 0; j < I40E_LOOK_AHEAD; j++)
 543                         rxq->rx_stage[i + j] = rxep[j].mbuf;
 544
 545                 if (nb_dd != I40E_LOOK_AHEAD)
 546                         break;
 547         }
 548
 549         /* Clear software ring entries */
 550         for (i = 0; i < nb_rx; i++)
 551                 rxq->sw_ring[rxq->rx_tail + i].mbuf = NULL;
 552
 553         return nb_rx;
 554 }
 555
 556 static inline uint16_t
 557 i40e_rx_fill_from_stage(struct i40e_rx_queue *rxq,
 558                         struct rte_mbuf **rx_pkts,
 559                         uint16_t nb_pkts)
 560 {
 561         uint16_t i;
 562         struct rte_mbuf **stage = &rxq->rx_stage[rxq->rx_next_avail];
 563
 564         nb_pkts = (uint16_t)RTE_MIN(nb_pkts, rxq->rx_nb_avail);
 565
 566         for (i = 0; i < nb_pkts; i++)
 567                 rx_pkts[i] = stage[i];
 568
 569         rxq->rx_nb_avail = (uint16_t)(rxq->rx_nb_avail - nb_pkts);
 570         rxq->rx_next_avail = (uint16_t)(rxq->rx_next_avail + nb_pkts);
 571
 572         return nb_pkts;
 573 }
 574
 575 static inline int
 576 i40e_rx_alloc_bufs(struct i40e_rx_queue *rxq)
 577 {
 578         volatile union i40e_rx_desc *rxdp;
 579         struct i40e_rx_entry *rxep;
 580         struct rte_mbuf *mb;
 581         uint16_t alloc_idx, i;
 582         uint64_t dma_addr;
 583         int diag;
 584
 585         /* Allocate buffers in bulk */
 586         alloc_idx = (uint16_t)(rxq->rx_free_trigger -
 587                                 (rxq->rx_free_thresh - 1));
 588         rxep = &(rxq->sw_ring[alloc_idx]);
 589         diag = rte_mempool_get_bulk(rxq->mp, (void *)rxep,
 590                                         rxq->rx_free_thresh);
 591         if (unlikely(diag != 0)) {
 592                 PMD_DRV_LOG(ERR, "Failed to get mbufs in bulk");
 593                 return -ENOMEM;
 594         }
 595
 596         rxdp = &rxq->rx_ring[alloc_idx];
 597         for (i = 0; i < rxq->rx_free_thresh; i++) {
 598                 if (likely(i < (rxq->rx_free_thresh - 1)))
 599                         /* Prefetch next mbuf */
 600                         rte_prefetch0(rxep[i + 1].mbuf);
 601
 602                 mb = rxep[i].mbuf;
 603                 rte_mbuf_refcnt_set(mb, 1);
 604                 mb->next = NULL;
 605                 mb->data_off = RTE_PKTMBUF_HEADROOM;
 606                 mb->nb_segs = 1;
 607                 mb->port = rxq->port_id;
 608                 dma_addr = rte_cpu_to_le_64(\
 609                         rte_mbuf_data_iova_default(mb));
 610                 rxdp[i].read.hdr_addr = 0;
 611                 rxdp[i].read.pkt_addr = dma_addr;
 612         }
 613
 614         /* Update rx tail regsiter */
 615         I40E_PCI_REG_WRITE(rxq->qrx_tail, rxq->rx_free_trigger);
 616
 617         rxq->rx_free_trigger =
 618                 (uint16_t)(rxq->rx_free_trigger + rxq->rx_free_thresh);
 619         if (rxq->rx_free_trigger >= rxq->nb_rx_desc)
 620                 rxq->rx_free_trigger = (uint16_t)(rxq->rx_free_thresh - 1);
 621
 622         return 0;
 623 }
 624
 625 static inline uint16_t
 626 rx_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 627 {
 628         struct i40e_rx_queue *rxq = (struct i40e_rx_queue *)rx_queue;
 629         struct rte_eth_dev *dev;
 630         uint16_t nb_rx = 0;
 631
 632         if (!nb_pkts)
 633                 return 0;
 634
 635         if (rxq->rx_nb_avail)
 636                 return i40e_rx_fill_from_stage(rxq, rx_pkts, nb_pkts);
 637
 638         nb_rx = (uint16_t)i40e_rx_scan_hw_ring(rxq);
 639         rxq->rx_next_avail = 0;
 640         rxq->rx_nb_avail = nb_rx;
 641         rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_rx);
 642
 643         if (rxq->rx_tail > rxq->rx_free_trigger) {
 644                 if (i40e_rx_alloc_bufs(rxq) != 0) {
 645                         uint16_t i, j;
 646
 647                         dev = I40E_VSI_TO_ETH_DEV(rxq->vsi);
 648                         dev->data->rx_mbuf_alloc_failed +=
 649                                 rxq->rx_free_thresh;
 650
 651                         rxq->rx_nb_avail = 0;
 652                         rxq->rx_tail = (uint16_t)(rxq->rx_tail - nb_rx);
 653                         for (i = 0, j = rxq->rx_tail; i < nb_rx; i++, j++)
 654                                 rxq->sw_ring[j].mbuf = rxq->rx_stage[i];
 655
 656                         return 0;
 657                 }
 658         }
 659
 660         if (rxq->rx_tail >= rxq->nb_rx_desc)
 661                 rxq->rx_tail = 0;
 662
 663         if (rxq->rx_nb_avail)
 664                 return i40e_rx_fill_from_stage(rxq, rx_pkts, nb_pkts);
 665
 666         return 0;
 667 }
 668
 669 static uint16_t
 670 i40e_recv_pkts_bulk_alloc(void *rx_queue,
 671                           struct rte_mbuf **rx_pkts,
 672                           uint16_t nb_pkts)
 673 {
 674         uint16_t nb_rx = 0, n, count;
 675
 676         if (unlikely(nb_pkts == 0))
 677                 return 0;
 678
 679         if (likely(nb_pkts <= RTE_PMD_I40E_RX_MAX_BURST))
 680                 return rx_recv_pkts(rx_queue, rx_pkts, nb_pkts);
 681
 682         while (nb_pkts) {
 683                 n = RTE_MIN(nb_pkts, RTE_PMD_I40E_RX_MAX_BURST);
 684                 count = rx_recv_pkts(rx_queue, &rx_pkts[nb_rx], n);
 685                 nb_rx = (uint16_t)(nb_rx + count);
 686                 nb_pkts = (uint16_t)(nb_pkts - count);
 687                 if (count < n)
 688                         break;
 689         }
 690
 691         return nb_rx;
 692 }
 693 #else
 694 static uint16_t
 695 i40e_recv_pkts_bulk_alloc(void __rte_unused *rx_queue,
 696                           struct rte_mbuf __rte_unused **rx_pkts,
 697                           uint16_t __rte_unused nb_pkts)
 698 {
 699         return 0;
 700 }
 701 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
 702
 703 uint16_t
 704 i40e_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 705 {
 706         struct i40e_rx_queue *rxq;
 707         volatile union i40e_rx_desc *rx_ring;
 708         volatile union i40e_rx_desc *rxdp;
 709         union i40e_rx_desc rxd;
 710         struct i40e_rx_entry *sw_ring;
 711         struct i40e_rx_entry *rxe;
 712         struct rte_eth_dev *dev;
 713         struct rte_mbuf *rxm;
 714         struct rte_mbuf *nmb;
 715         uint16_t nb_rx;
 716         uint32_t rx_status;
 717         uint64_t qword1;
 718         uint16_t rx_packet_len;
 719         uint16_t rx_id, nb_hold;
 720         uint64_t dma_addr;
 721         uint64_t pkt_flags;
 722         uint32_t *ptype_tbl;
 723
 724         nb_rx = 0;
 725         nb_hold = 0;
 726         rxq = rx_queue;
 727         rx_id = rxq->rx_tail;
 728         rx_ring = rxq->rx_ring;
 729         sw_ring = rxq->sw_ring;
 730         ptype_tbl = rxq->vsi->adapter->ptype_tbl;
 731
 732         while (nb_rx < nb_pkts) {
 733                 rxdp = &rx_ring[rx_id];
 734                 qword1 = rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len);
 735                 rx_status = (qword1 & I40E_RXD_QW1_STATUS_MASK)
 736                                 >> I40E_RXD_QW1_STATUS_SHIFT;
 737
 738                 /* Check the DD bit first */
 739                 if (!(rx_status & (1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
 740                         break;
 741
 742                 nmb = rte_mbuf_raw_alloc(rxq->mp);
 743                 if (unlikely(!nmb)) {
 744                         dev = I40E_VSI_TO_ETH_DEV(rxq->vsi);
 745                         dev->data->rx_mbuf_alloc_failed++;
 746                         break;
 747                 }
 748
 749                 rxd = *rxdp;
 750                 nb_hold++;
 751                 rxe = &sw_ring[rx_id];
 752                 rx_id++;
 753                 if (unlikely(rx_id == rxq->nb_rx_desc))
 754                         rx_id = 0;
 755
 756                 /* Prefetch next mbuf */
 757                 rte_prefetch0(sw_ring[rx_id].mbuf);
 758
 759                 /**
 760                  * When next RX descriptor is on a cache line boundary,
 761                  * prefetch the next 4 RX descriptors and next 8 pointers
 762                  * to mbufs.
 763                  */
 764                 if ((rx_id & 0x3) == 0) {
 765                         rte_prefetch0(&rx_ring[rx_id]);
 766                         rte_prefetch0(&sw_ring[rx_id]);
 767                 }
 768                 rxm = rxe->mbuf;
 769                 rxe->mbuf = nmb;
 770                 dma_addr =
 771                         rte_cpu_to_le_64(rte_mbuf_data_iova_default(nmb));
 772                 rxdp->read.hdr_addr = 0;
 773                 rxdp->read.pkt_addr = dma_addr;
 774
 775                 rx_packet_len = ((qword1 & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
 776                                 I40E_RXD_QW1_LENGTH_PBUF_SHIFT) - rxq->crc_len;
 777
 778                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 779                 rte_prefetch0(RTE_PTR_ADD(rxm->buf_addr, RTE_PKTMBUF_HEADROOM));
 780                 rxm->nb_segs = 1;
 781                 rxm->next = NULL;
 782                 rxm->pkt_len = rx_packet_len;
 783                 rxm->data_len = rx_packet_len;
 784                 rxm->port = rxq->port_id;
 785                 rxm->ol_flags = 0;
 786                 i40e_rxd_to_vlan_tci(rxm, &rxd);
 787                 pkt_flags = i40e_rxd_status_to_pkt_flags(qword1);
 788                 pkt_flags |= i40e_rxd_error_to_pkt_flags(qword1);
 789                 rxm->packet_type =
 790                         ptype_tbl[(uint8_t)((qword1 &
 791                         I40E_RXD_QW1_PTYPE_MASK) >> I40E_RXD_QW1_PTYPE_SHIFT)];
 792                 if (pkt_flags & PKT_RX_RSS_HASH)
 793                         rxm->hash.rss =
 794                                 rte_le_to_cpu_32(rxd.wb.qword0.hi_dword.rss);
 795                 if (pkt_flags & PKT_RX_FDIR)
 796                         pkt_flags |= i40e_rxd_build_fdir(&rxd, rxm);
 797
 798 #ifdef RTE_LIBRTE_IEEE1588
 799                 pkt_flags |= i40e_get_iee15888_flags(rxm, qword1);
 800 #endif
 801                 rxm->ol_flags |= pkt_flags;
 802
 803                 rx_pkts[nb_rx++] = rxm;
 804         }
 805         rxq->rx_tail = rx_id;
 806
 807         /**
 808          * If the number of free RX descriptors is greater than the RX free
 809          * threshold of the queue, advance the receive tail register of queue.
 810          * Update that register with the value of the last processed RX
 811          * descriptor minus 1.
 812          */
 813         nb_hold = (uint16_t)(nb_hold + rxq->nb_rx_hold);
 814         if (nb_hold > rxq->rx_free_thresh) {
 815                 rx_id = (uint16_t) ((rx_id == 0) ?
 816                         (rxq->nb_rx_desc - 1) : (rx_id - 1));
 817                 I40E_PCI_REG_WC_WRITE(rxq->qrx_tail, rx_id);
 818                 nb_hold = 0;
 819         }
 820         rxq->nb_rx_hold = nb_hold;
 821
 822         return nb_rx;
 823 }
 824
 825 uint16_t
 826 i40e_recv_scattered_pkts(void *rx_queue,
 827                          struct rte_mbuf **rx_pkts,
 828                          uint16_t nb_pkts)
 829 {
 830         struct i40e_rx_queue *rxq = rx_queue;
 831         volatile union i40e_rx_desc *rx_ring = rxq->rx_ring;
 832         volatile union i40e_rx_desc *rxdp;
 833         union i40e_rx_desc rxd;
 834         struct i40e_rx_entry *sw_ring = rxq->sw_ring;
 835         struct i40e_rx_entry *rxe;
 836         struct rte_mbuf *first_seg = rxq->pkt_first_seg;
 837         struct rte_mbuf *last_seg = rxq->pkt_last_seg;
 838         struct rte_mbuf *nmb, *rxm;
 839         uint16_t rx_id = rxq->rx_tail;
 840         uint16_t nb_rx = 0, nb_hold = 0, rx_packet_len;
 841         struct rte_eth_dev *dev;
 842         uint32_t rx_status;
 843         uint64_t qword1;
 844         uint64_t dma_addr;
 845         uint64_t pkt_flags;
 846         uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
 847
 848         while (nb_rx < nb_pkts) {
 849                 rxdp = &rx_ring[rx_id];
 850                 qword1 = rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len);
 851                 rx_status = (qword1 & I40E_RXD_QW1_STATUS_MASK) >>
 852                                         I40E_RXD_QW1_STATUS_SHIFT;
 853
 854                 /* Check the DD bit */
 855                 if (!(rx_status & (1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
 856                         break;
 857
 858                 nmb = rte_mbuf_raw_alloc(rxq->mp);
 859                 if (unlikely(!nmb)) {
 860                         dev = I40E_VSI_TO_ETH_DEV(rxq->vsi);
 861                         dev->data->rx_mbuf_alloc_failed++;
 862                         break;
 863                 }
 864
 865                 rxd = *rxdp;
 866                 nb_hold++;
 867                 rxe = &sw_ring[rx_id];
 868                 rx_id++;
 869                 if (rx_id == rxq->nb_rx_desc)
 870                         rx_id = 0;
 871
 872                 /* Prefetch next mbuf */
 873                 rte_prefetch0(sw_ring[rx_id].mbuf);
 874
 875                 /**
 876                  * When next RX descriptor is on a cache line boundary,
 877                  * prefetch the next 4 RX descriptors and next 8 pointers
 878                  * to mbufs.
 879                  */
 880                 if ((rx_id & 0x3) == 0) {
 881                         rte_prefetch0(&rx_ring[rx_id]);
 882                         rte_prefetch0(&sw_ring[rx_id]);
 883                 }
 884
 885                 rxm = rxe->mbuf;
 886                 rxe->mbuf = nmb;
 887                 dma_addr =
 888                         rte_cpu_to_le_64(rte_mbuf_data_iova_default(nmb));
 889
 890                 /* Set data buffer address and data length of the mbuf */
 891                 rxdp->read.hdr_addr = 0;
 892                 rxdp->read.pkt_addr = dma_addr;
 893                 rx_packet_len = (qword1 & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
 894                                         I40E_RXD_QW1_LENGTH_PBUF_SHIFT;
 895                 rxm->data_len = rx_packet_len;
 896                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 897
 898                 /**
 899                  * If this is the first buffer of the received packet, set the
 900                  * pointer to the first mbuf of the packet and initialize its
 901                  * context. Otherwise, update the total length and the number
 902                  * of segments of the current scattered packet, and update the
 903                  * pointer to the last mbuf of the current packet.
 904                  */
 905                 if (!first_seg) {
 906                         first_seg = rxm;
 907                         first_seg->nb_segs = 1;
 908                         first_seg->pkt_len = rx_packet_len;
 909                 } else {
 910                         first_seg->pkt_len =
 911                                 (uint16_t)(first_seg->pkt_len +
 912                                                 rx_packet_len);
 913                         first_seg->nb_segs++;
 914                         last_seg->next = rxm;
 915                 }
 916
 917                 /**
 918                  * If this is not the last buffer of the received packet,
 919                  * update the pointer to the last mbuf of the current scattered
 920                  * packet and continue to parse the RX ring.
 921                  */
 922                 if (!(rx_status & (1 << I40E_RX_DESC_STATUS_EOF_SHIFT))) {
 923                         last_seg = rxm;
 924                         continue;
 925                 }
 926
 927                 /**
 928                  * This is the last buffer of the received packet. If the CRC
 929                  * is not stripped by the hardware:
 930                  *  - Subtract the CRC length from the total packet length.
 931                  *  - If the last buffer only contains the whole CRC or a part
 932                  *  of it, free the mbuf associated to the last buffer. If part
 933                  *  of the CRC is also contained in the previous mbuf, subtract
 934                  *  the length of that CRC part from the data length of the
 935                  *  previous mbuf.
 936                  */
 937                 rxm->next = NULL;
 938                 if (unlikely(rxq->crc_len > 0)) {
 939                         first_seg->pkt_len -= RTE_ETHER_CRC_LEN;
 940                         if (rx_packet_len <= RTE_ETHER_CRC_LEN) {
 941                                 rte_pktmbuf_free_seg(rxm);
 942                                 first_seg->nb_segs--;
 943                                 last_seg->data_len =
 944                                         (uint16_t)(last_seg->data_len -
 945                                         (RTE_ETHER_CRC_LEN - rx_packet_len));
 946                                 last_seg->next = NULL;
 947                         } else
 948                                 rxm->data_len = (uint16_t)(rx_packet_len -
 949                                                         RTE_ETHER_CRC_LEN);
 950                 }
 951
 952                 first_seg->port = rxq->port_id;
 953                 first_seg->ol_flags = 0;
 954                 i40e_rxd_to_vlan_tci(first_seg, &rxd);
 955                 pkt_flags = i40e_rxd_status_to_pkt_flags(qword1);
 956                 pkt_flags |= i40e_rxd_error_to_pkt_flags(qword1);
 957                 first_seg->packet_type =
 958                         ptype_tbl[(uint8_t)((qword1 &
 959                         I40E_RXD_QW1_PTYPE_MASK) >> I40E_RXD_QW1_PTYPE_SHIFT)];
 960                 if (pkt_flags & PKT_RX_RSS_HASH)
 961                         first_seg->hash.rss =
 962                                 rte_le_to_cpu_32(rxd.wb.qword0.hi_dword.rss);
 963                 if (pkt_flags & PKT_RX_FDIR)
 964                         pkt_flags |= i40e_rxd_build_fdir(&rxd, first_seg);
 965
 966 #ifdef RTE_LIBRTE_IEEE1588
 967                 pkt_flags |= i40e_get_iee15888_flags(first_seg, qword1);
 968 #endif
 969                 first_seg->ol_flags |= pkt_flags;
 970
 971                 /* Prefetch data of first segment, if configured to do so. */
 972                 rte_prefetch0(RTE_PTR_ADD(first_seg->buf_addr,
 973                         first_seg->data_off));
 974                 rx_pkts[nb_rx++] = first_seg;
 975                 first_seg = NULL;
 976         }
 977
 978         /* Record index of the next RX descriptor to probe. */
 979         rxq->rx_tail = rx_id;
 980         rxq->pkt_first_seg = first_seg;
 981         rxq->pkt_last_seg = last_seg;
 982
 983         /**
 984          * If the number of free RX descriptors is greater than the RX free
 985          * threshold of the queue, advance the Receive Descriptor Tail (RDT)
 986          * register. Update the RDT with the value of the last processed RX
 987          * descriptor minus 1, to guarantee that the RDT register is never
 988          * equal to the RDH register, which creates a "full" ring situtation
 989          * from the hardware point of view.
 990          */
 991         nb_hold = (uint16_t)(nb_hold + rxq->nb_rx_hold);
 992         if (nb_hold > rxq->rx_free_thresh) {
 993                 rx_id = (uint16_t)(rx_id == 0 ?
 994                         (rxq->nb_rx_desc - 1) : (rx_id - 1));
 995                 I40E_PCI_REG_WC_WRITE(rxq->qrx_tail, rx_id);
 996                 nb_hold = 0;
 997         }
 998         rxq->nb_rx_hold = nb_hold;
 999
1000         return nb_rx;
1001 }
1002
1003 /* Check if the context descriptor is needed for TX offloading */
1004 static inline uint16_t
1005 i40e_calc_context_desc(uint64_t flags)
1006 {
1007         static uint64_t mask = PKT_TX_OUTER_IP_CKSUM |
1008                 PKT_TX_TCP_SEG |
1009                 PKT_TX_QINQ_PKT |
1010                 PKT_TX_TUNNEL_MASK;
1011
1012 #ifdef RTE_LIBRTE_IEEE1588
1013         mask |= PKT_TX_IEEE1588_TMST;
1014 #endif
1015
1016         return (flags & mask) ? 1 : 0;
1017 }
1018
1019 /* set i40e TSO context descriptor */
1020 static inline uint64_t
1021 i40e_set_tso_ctx(struct rte_mbuf *mbuf, union i40e_tx_offload tx_offload)
1022 {
1023         uint64_t ctx_desc = 0;
1024         uint32_t cd_cmd, hdr_len, cd_tso_len;
1025
1026         if (!tx_offload.l4_len) {
1027                 PMD_DRV_LOG(DEBUG, "L4 length set to 0");
1028                 return ctx_desc;
1029         }
1030
1031         hdr_len = tx_offload.l2_len + tx_offload.l3_len + tx_offload.l4_len;
1032         hdr_len += (mbuf->ol_flags & PKT_TX_TUNNEL_MASK) ?
1033                    tx_offload.outer_l2_len + tx_offload.outer_l3_len : 0;
1034
1035         cd_cmd = I40E_TX_CTX_DESC_TSO;
1036         cd_tso_len = mbuf->pkt_len - hdr_len;
1037         ctx_desc |= ((uint64_t)cd_cmd << I40E_TXD_CTX_QW1_CMD_SHIFT) |
1038                 ((uint64_t)cd_tso_len <<
1039                  I40E_TXD_CTX_QW1_TSO_LEN_SHIFT) |
1040                 ((uint64_t)mbuf->tso_segsz <<
1041                  I40E_TXD_CTX_QW1_MSS_SHIFT);
1042
1043         return ctx_desc;
1044 }
1045
1046 /* HW requires that Tx buffer size ranges from 1B up to (16K-1)B. */
1047 #define I40E_MAX_DATA_PER_TXD \
1048         (I40E_TXD_QW1_TX_BUF_SZ_MASK >> I40E_TXD_QW1_TX_BUF_SZ_SHIFT)
1049 /* Calculate the number of TX descriptors needed for each pkt */
1050 static inline uint16_t
1051 i40e_calc_pkt_desc(struct rte_mbuf *tx_pkt)
1052 {
1053         struct rte_mbuf *txd = tx_pkt;
1054         uint16_t count = 0;
1055
1056         while (txd != NULL) {
1057                 count += DIV_ROUND_UP(txd->data_len, I40E_MAX_DATA_PER_TXD);
1058                 txd = txd->next;
1059         }
1060
1061         return count;
1062 }
1063
1064 uint16_t
1065 i40e_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1066 {
1067         struct i40e_tx_queue *txq;
1068         struct i40e_tx_entry *sw_ring;
1069         struct i40e_tx_entry *txe, *txn;
1070         volatile struct i40e_tx_desc *txd;
1071         volatile struct i40e_tx_desc *txr;
1072         struct rte_mbuf *tx_pkt;
1073         struct rte_mbuf *m_seg;
1074         uint32_t cd_tunneling_params;
1075         uint16_t tx_id;
1076         uint16_t nb_tx;
1077         uint32_t td_cmd;
1078         uint32_t td_offset;
1079         uint32_t td_tag;
1080         uint64_t ol_flags;
1081         uint16_t nb_used;
1082         uint16_t nb_ctx;
1083         uint16_t tx_last;
1084         uint16_t slen;
1085         uint64_t buf_dma_addr;
1086         union i40e_tx_offload tx_offload = {0};
1087
1088         txq = tx_queue;
1089         sw_ring = txq->sw_ring;
1090         txr = txq->tx_ring;
1091         tx_id = txq->tx_tail;
1092         txe = &sw_ring[tx_id];
1093
1094         /* Check if the descriptor ring needs to be cleaned. */
1095         if (txq->nb_tx_free < txq->tx_free_thresh)
1096                 (void)i40e_xmit_cleanup(txq);
1097
1098         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1099                 td_cmd = 0;
1100                 td_tag = 0;
1101                 td_offset = 0;
1102
1103                 tx_pkt = *tx_pkts++;
1104                 RTE_MBUF_PREFETCH_TO_FREE(txe->mbuf);
1105
1106                 ol_flags = tx_pkt->ol_flags;
1107                 tx_offload.l2_len = tx_pkt->l2_len;
1108                 tx_offload.l3_len = tx_pkt->l3_len;
1109                 tx_offload.outer_l2_len = tx_pkt->outer_l2_len;
1110                 tx_offload.outer_l3_len = tx_pkt->outer_l3_len;
1111                 tx_offload.l4_len = tx_pkt->l4_len;
1112                 tx_offload.tso_segsz = tx_pkt->tso_segsz;
1113
1114                 /* Calculate the number of context descriptors needed. */
1115                 nb_ctx = i40e_calc_context_desc(ol_flags);
1116
1117                 /**
1118                  * The number of descriptors that must be allocated for
1119                  * a packet equals to the number of the segments of that
1120                  * packet plus 1 context descriptor if needed.
1121                  * Recalculate the needed tx descs when TSO enabled in case
1122                  * the mbuf data size exceeds max data size that hw allows
1123                  * per tx desc.
1124                  */
1125                 if (ol_flags & PKT_TX_TCP_SEG)
1126                         nb_used = (uint16_t)(i40e_calc_pkt_desc(tx_pkt) +
1127                                              nb_ctx);
1128                 else
1129                         nb_used = (uint16_t)(tx_pkt->nb_segs + nb_ctx);
1130                 tx_last = (uint16_t)(tx_id + nb_used - 1);
1131
1132                 /* Circular ring */
1133                 if (tx_last >= txq->nb_tx_desc)
1134                         tx_last = (uint16_t)(tx_last - txq->nb_tx_desc);
1135
1136                 if (nb_used > txq->nb_tx_free) {
1137                         if (i40e_xmit_cleanup(txq) != 0) {
1138                                 if (nb_tx == 0)
1139                                         return 0;
1140                                 goto end_of_tx;
1141                         }
1142                         if (unlikely(nb_used > txq->tx_rs_thresh)) {
1143                                 while (nb_used > txq->nb_tx_free) {
1144                                         if (i40e_xmit_cleanup(txq) != 0) {
1145                                                 if (nb_tx == 0)
1146                                                         return 0;
1147                                                 goto end_of_tx;
1148                                         }
1149                                 }
1150                         }
1151                 }
1152
1153                 /* Descriptor based VLAN insertion */
1154                 if (ol_flags & (PKT_TX_VLAN_PKT | PKT_TX_QINQ_PKT)) {
1155                         td_cmd |= I40E_TX_DESC_CMD_IL2TAG1;
1156                         td_tag = tx_pkt->vlan_tci;
1157                 }
1158
1159                 /* Always enable CRC offload insertion */
1160                 td_cmd |= I40E_TX_DESC_CMD_ICRC;
1161
1162                 /* Fill in tunneling parameters if necessary */
1163                 cd_tunneling_params = 0;
1164                 if (ol_flags & PKT_TX_TUNNEL_MASK)
1165                         i40e_parse_tunneling_params(ol_flags, tx_offload,
1166                                                     &cd_tunneling_params);
1167                 /* Enable checksum offloading */
1168                 if (ol_flags & I40E_TX_CKSUM_OFFLOAD_MASK)
1169                         i40e_txd_enable_checksum(ol_flags, &td_cmd,
1170                                                  &td_offset, tx_offload);
1171
1172                 if (nb_ctx) {
1173                         /* Setup TX context descriptor if required */
1174                         volatile struct i40e_tx_context_desc *ctx_txd =
1175                                 (volatile struct i40e_tx_context_desc *)\
1176                                                         &txr[tx_id];
1177                         uint16_t cd_l2tag2 = 0;
1178                         uint64_t cd_type_cmd_tso_mss =
1179                                 I40E_TX_DESC_DTYPE_CONTEXT;
1180
1181                         txn = &sw_ring[txe->next_id];
1182                         RTE_MBUF_PREFETCH_TO_FREE(txn->mbuf);
1183                         if (txe->mbuf != NULL) {
1184                                 rte_pktmbuf_free_seg(txe->mbuf);
1185                                 txe->mbuf = NULL;
1186                         }
1187
1188                         /* TSO enabled means no timestamp */
1189                         if (ol_flags & PKT_TX_TCP_SEG)
1190                                 cd_type_cmd_tso_mss |=
1191                                         i40e_set_tso_ctx(tx_pkt, tx_offload);
1192                         else {
1193 #ifdef RTE_LIBRTE_IEEE1588
1194                                 if (ol_flags & PKT_TX_IEEE1588_TMST)
1195                                         cd_type_cmd_tso_mss |=
1196                                                 ((uint64_t)I40E_TX_CTX_DESC_TSYN <<
1197                                                  I40E_TXD_CTX_QW1_CMD_SHIFT);
1198 #endif
1199                         }
1200
1201                         ctx_txd->tunneling_params =
1202                                 rte_cpu_to_le_32(cd_tunneling_params);
1203                         if (ol_flags & PKT_TX_QINQ_PKT) {
1204                                 cd_l2tag2 = tx_pkt->vlan_tci_outer;
1205                                 cd_type_cmd_tso_mss |=
1206                                         ((uint64_t)I40E_TX_CTX_DESC_IL2TAG2 <<
1207                                                 I40E_TXD_CTX_QW1_CMD_SHIFT);
1208                         }
1209                         ctx_txd->l2tag2 = rte_cpu_to_le_16(cd_l2tag2);
1210                         ctx_txd->type_cmd_tso_mss =
1211                                 rte_cpu_to_le_64(cd_type_cmd_tso_mss);
1212
1213                         PMD_TX_LOG(DEBUG, "mbuf: %p, TCD[%u]:\n"
1214                                 "tunneling_params: %#x;\n"
1215                                 "l2tag2: %#hx;\n"
1216                                 "rsvd: %#hx;\n"
1217                                 "type_cmd_tso_mss: %#"PRIx64";\n",
1218                                 tx_pkt, tx_id,
1219                                 ctx_txd->tunneling_params,
1220                                 ctx_txd->l2tag2,
1221                                 ctx_txd->rsvd,
1222                                 ctx_txd->type_cmd_tso_mss);
1223
1224                         txe->last_id = tx_last;
1225                         tx_id = txe->next_id;
1226                         txe = txn;
1227                 }
1228
1229                 m_seg = tx_pkt;
1230                 do {
1231                         txd = &txr[tx_id];
1232                         txn = &sw_ring[txe->next_id];
1233
1234                         if (txe->mbuf)
1235                                 rte_pktmbuf_free_seg(txe->mbuf);
1236                         txe->mbuf = m_seg;
1237
1238                         /* Setup TX Descriptor */
1239                         slen = m_seg->data_len;
1240                         buf_dma_addr = rte_mbuf_data_iova(m_seg);
1241
1242                         while ((ol_flags & PKT_TX_TCP_SEG) &&
1243                                 unlikely(slen > I40E_MAX_DATA_PER_TXD)) {
1244                                 txd->buffer_addr =
1245                                         rte_cpu_to_le_64(buf_dma_addr);
1246                                 txd->cmd_type_offset_bsz =
1247                                         i40e_build_ctob(td_cmd,
1248                                         td_offset, I40E_MAX_DATA_PER_TXD,
1249                                         td_tag);
1250
1251                                 buf_dma_addr += I40E_MAX_DATA_PER_TXD;
1252                                 slen -= I40E_MAX_DATA_PER_TXD;
1253
1254                                 txe->last_id = tx_last;
1255                                 tx_id = txe->next_id;
1256                                 txe = txn;
1257                                 txd = &txr[tx_id];
1258                                 txn = &sw_ring[txe->next_id];
1259                         }
1260                         PMD_TX_LOG(DEBUG, "mbuf: %p, TDD[%u]:\n"
1261                                 "buf_dma_addr: %#"PRIx64";\n"
1262                                 "td_cmd: %#x;\n"
1263                                 "td_offset: %#x;\n"
1264                                 "td_len: %u;\n"
1265                                 "td_tag: %#x;\n",
1266                                 tx_pkt, tx_id, buf_dma_addr,
1267                                 td_cmd, td_offset, slen, td_tag);
1268
1269                         txd->buffer_addr = rte_cpu_to_le_64(buf_dma_addr);
1270                         txd->cmd_type_offset_bsz = i40e_build_ctob(td_cmd,
1271                                                 td_offset, slen, td_tag);
1272                         txe->last_id = tx_last;
1273                         tx_id = txe->next_id;
1274                         txe = txn;
1275                         m_seg = m_seg->next;
1276                 } while (m_seg != NULL);
1277
1278                 /* The last packet data descriptor needs End Of Packet (EOP) */
1279                 td_cmd |= I40E_TX_DESC_CMD_EOP;
1280                 txq->nb_tx_used = (uint16_t)(txq->nb_tx_used + nb_used);
1281                 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_used);
1282
1283                 if (txq->nb_tx_used >= txq->tx_rs_thresh) {
1284                         PMD_TX_LOG(DEBUG,
1285                                    "Setting RS bit on TXD id="
1286                                    "%4u (port=%d queue=%d)",
1287                                    tx_last, txq->port_id, txq->queue_id);
1288
1289                         td_cmd |= I40E_TX_DESC_CMD_RS;
1290
1291                         /* Update txq RS bit counters */
1292                         txq->nb_tx_used = 0;
1293                 }
1294
1295                 txd->cmd_type_offset_bsz |=
1296                         rte_cpu_to_le_64(((uint64_t)td_cmd) <<
1297                                         I40E_TXD_QW1_CMD_SHIFT);
1298         }
1299
1300 end_of_tx:
1301         PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u tx_tail=%u nb_tx=%u",
1302                    (unsigned) txq->port_id, (unsigned) txq->queue_id,
1303                    (unsigned) tx_id, (unsigned) nb_tx);
1304
1305         rte_io_wmb();
1306         I40E_PCI_REG_WC_WRITE_RELAXED(txq->qtx_tail, tx_id);
1307         txq->tx_tail = tx_id;
1308
1309         return nb_tx;
1310 }
1311
1312 static __rte_always_inline int
1313 i40e_tx_free_bufs(struct i40e_tx_queue *txq)
1314 {
1315         struct i40e_tx_entry *txep;
1316         uint16_t tx_rs_thresh = txq->tx_rs_thresh;
1317         uint16_t i = 0, j = 0;
1318         struct rte_mbuf *free[RTE_I40E_TX_MAX_FREE_BUF_SZ];
1319         const uint16_t k = RTE_ALIGN_FLOOR(tx_rs_thresh, RTE_I40E_TX_MAX_FREE_BUF_SZ);
1320         const uint16_t m = tx_rs_thresh % RTE_I40E_TX_MAX_FREE_BUF_SZ;
1321
1322         if ((txq->tx_ring[txq->tx_next_dd].cmd_type_offset_bsz &
1323                         rte_cpu_to_le_64(I40E_TXD_QW1_DTYPE_MASK)) !=
1324                         rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE))
1325                 return 0;
1326
1327         txep = &txq->sw_ring[txq->tx_next_dd - (tx_rs_thresh - 1)];
1328
1329         for (i = 0; i < tx_rs_thresh; i++)
1330                 rte_prefetch0((txep + i)->mbuf);
1331
1332         if (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE) {
1333                 if (k) {
1334                         for (j = 0; j != k; j += RTE_I40E_TX_MAX_FREE_BUF_SZ) {
1335                                 for (i = 0; i < RTE_I40E_TX_MAX_FREE_BUF_SZ; ++i, ++txep) {
1336                                         free[i] = txep->mbuf;
1337                                         txep->mbuf = NULL;
1338                                 }
1339                                 rte_mempool_put_bulk(free[0]->pool, (void **)free,
1340                                                 RTE_I40E_TX_MAX_FREE_BUF_SZ);
1341                         }
1342                 }
1343
1344                 if (m) {
1345                         for (i = 0; i < m; ++i, ++txep) {
1346                                 free[i] = txep->mbuf;
1347                                 txep->mbuf = NULL;
1348                         }
1349                         rte_mempool_put_bulk(free[0]->pool, (void **)free, m);
1350                 }
1351         } else {
1352                 for (i = 0; i < txq->tx_rs_thresh; ++i, ++txep) {
1353                         rte_pktmbuf_free_seg(txep->mbuf);
1354                         txep->mbuf = NULL;
1355                 }
1356         }
1357
1358         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + txq->tx_rs_thresh);
1359         txq->tx_next_dd = (uint16_t)(txq->tx_next_dd + txq->tx_rs_thresh);
1360         if (txq->tx_next_dd >= txq->nb_tx_desc)
1361                 txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
1362
1363         return txq->tx_rs_thresh;
1364 }
1365
1366 /* Populate 4 descriptors with data from 4 mbufs */
1367 static inline void
1368 tx4(volatile struct i40e_tx_desc *txdp, struct rte_mbuf **pkts)
1369 {
1370         uint64_t dma_addr;
1371         uint32_t i;
1372
1373         for (i = 0; i < 4; i++, txdp++, pkts++) {
1374                 dma_addr = rte_mbuf_data_iova(*pkts);
1375                 txdp->buffer_addr = rte_cpu_to_le_64(dma_addr);
1376                 txdp->cmd_type_offset_bsz =
1377                         i40e_build_ctob((uint32_t)I40E_TD_CMD, 0,
1378                                         (*pkts)->data_len, 0);
1379         }
1380 }
1381
1382 /* Populate 1 descriptor with data from 1 mbuf */
1383 static inline void
1384 tx1(volatile struct i40e_tx_desc *txdp, struct rte_mbuf **pkts)
1385 {
1386         uint64_t dma_addr;
1387
1388         dma_addr = rte_mbuf_data_iova(*pkts);
1389         txdp->buffer_addr = rte_cpu_to_le_64(dma_addr);
1390         txdp->cmd_type_offset_bsz =
1391                 i40e_build_ctob((uint32_t)I40E_TD_CMD, 0,
1392                                 (*pkts)->data_len, 0);
1393 }
1394
1395 /* Fill hardware descriptor ring with mbuf data */
1396 static inline void
1397 i40e_tx_fill_hw_ring(struct i40e_tx_queue *txq,
1398                      struct rte_mbuf **pkts,
1399                      uint16_t nb_pkts)
1400 {
1401         volatile struct i40e_tx_desc *txdp = &(txq->tx_ring[txq->tx_tail]);
1402         struct i40e_tx_entry *txep = &(txq->sw_ring[txq->tx_tail]);
1403         const int N_PER_LOOP = 4;
1404         const int N_PER_LOOP_MASK = N_PER_LOOP - 1;
1405         int mainpart, leftover;
1406         int i, j;
1407
1408         mainpart = (nb_pkts & ((uint32_t) ~N_PER_LOOP_MASK));
1409         leftover = (nb_pkts & ((uint32_t)  N_PER_LOOP_MASK));
1410         for (i = 0; i < mainpart; i += N_PER_LOOP) {
1411                 for (j = 0; j < N_PER_LOOP; ++j) {
1412                         (txep + i + j)->mbuf = *(pkts + i + j);
1413                 }
1414                 tx4(txdp + i, pkts + i);
1415         }
1416         if (unlikely(leftover > 0)) {
1417                 for (i = 0; i < leftover; ++i) {
1418                         (txep + mainpart + i)->mbuf = *(pkts + mainpart + i);
1419                         tx1(txdp + mainpart + i, pkts + mainpart + i);
1420                 }
1421         }
1422 }
1423
1424 static inline uint16_t
1425 tx_xmit_pkts(struct i40e_tx_queue *txq,
1426              struct rte_mbuf **tx_pkts,
1427              uint16_t nb_pkts)
1428 {
1429         volatile struct i40e_tx_desc *txr = txq->tx_ring;
1430         uint16_t n = 0;
1431
1432         /**
1433          * Begin scanning the H/W ring for done descriptors when the number
1434          * of available descriptors drops below tx_free_thresh. For each done
1435          * descriptor, free the associated buffer.
1436          */
1437         if (txq->nb_tx_free < txq->tx_free_thresh)
1438                 i40e_tx_free_bufs(txq);
1439
1440         /* Use available descriptor only */
1441         nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
1442         if (unlikely(!nb_pkts))
1443                 return 0;
1444
1445         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
1446         if ((txq->tx_tail + nb_pkts) > txq->nb_tx_desc) {
1447                 n = (uint16_t)(txq->nb_tx_desc - txq->tx_tail);
1448                 i40e_tx_fill_hw_ring(txq, tx_pkts, n);
1449                 txr[txq->tx_next_rs].cmd_type_offset_bsz |=
1450                         rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
1451                                                 I40E_TXD_QW1_CMD_SHIFT);
1452                 txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
1453                 txq->tx_tail = 0;
1454         }
1455
1456         /* Fill hardware descriptor ring with mbuf data */
1457         i40e_tx_fill_hw_ring(txq, tx_pkts + n, (uint16_t)(nb_pkts - n));
1458         txq->tx_tail = (uint16_t)(txq->tx_tail + (nb_pkts - n));
1459
1460         /* Determin if RS bit needs to be set */
1461         if (txq->tx_tail > txq->tx_next_rs) {
1462                 txr[txq->tx_next_rs].cmd_type_offset_bsz |=
1463                         rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
1464                                                 I40E_TXD_QW1_CMD_SHIFT);
1465                 txq->tx_next_rs =
1466                         (uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
1467                 if (txq->tx_next_rs >= txq->nb_tx_desc)
1468                         txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
1469         }
1470
1471         if (txq->tx_tail >= txq->nb_tx_desc)
1472                 txq->tx_tail = 0;
1473
1474         /* Update the tx tail register */
1475         I40E_PCI_REG_WC_WRITE(txq->qtx_tail, txq->tx_tail);
1476
1477         return nb_pkts;
1478 }
1479
1480 static uint16_t
1481 i40e_xmit_pkts_simple(void *tx_queue,
1482                       struct rte_mbuf **tx_pkts,
1483                       uint16_t nb_pkts)
1484 {
1485         uint16_t nb_tx = 0;
1486
1487         if (likely(nb_pkts <= I40E_TX_MAX_BURST))
1488                 return tx_xmit_pkts((struct i40e_tx_queue *)tx_queue,
1489                                                 tx_pkts, nb_pkts);
1490
1491         while (nb_pkts) {
1492                 uint16_t ret, num = (uint16_t)RTE_MIN(nb_pkts,
1493                                                 I40E_TX_MAX_BURST);
1494
1495                 ret = tx_xmit_pkts((struct i40e_tx_queue *)tx_queue,
1496                                                 &tx_pkts[nb_tx], num);
1497                 nb_tx = (uint16_t)(nb_tx + ret);
1498                 nb_pkts = (uint16_t)(nb_pkts - ret);
1499                 if (ret < num)
1500                         break;
1501         }
1502
1503         return nb_tx;
1504 }
1505
1506 static uint16_t
1507 i40e_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
1508                    uint16_t nb_pkts)
1509 {
1510         uint16_t nb_tx = 0;
1511         struct i40e_tx_queue *txq = (struct i40e_tx_queue *)tx_queue;
1512
1513         while (nb_pkts) {
1514                 uint16_t ret, num;
1515
1516                 num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh);
1517                 ret = i40e_xmit_fixed_burst_vec(tx_queue, &tx_pkts[nb_tx],
1518                                                 num);
1519                 nb_tx += ret;
1520                 nb_pkts -= ret;
1521                 if (ret < num)
1522                         break;
1523         }
1524
1525         return nb_tx;
1526 }
1527
1528 /*********************************************************************
1529  *
1530  *  TX simple prep functions
1531  *
1532  **********************************************************************/
1533 uint16_t
1534 i40e_simple_prep_pkts(__rte_unused void *tx_queue, struct rte_mbuf **tx_pkts,
1535                       uint16_t nb_pkts)
1536 {
1537         int i;
1538         uint64_t ol_flags;
1539         struct rte_mbuf *m;
1540
1541         for (i = 0; i < nb_pkts; i++) {
1542                 m = tx_pkts[i];
1543                 ol_flags = m->ol_flags;
1544
1545                 if (m->nb_segs != 1) {
1546                         rte_errno = EINVAL;
1547                         return i;
1548                 }
1549
1550                 if (ol_flags & I40E_TX_OFFLOAD_SIMPLE_NOTSUP_MASK) {
1551                         rte_errno = ENOTSUP;
1552                         return i;
1553                 }
1554
1555                 /* check the size of packet */
1556                 if (m->pkt_len < I40E_TX_MIN_PKT_LEN ||
1557                     m->pkt_len > I40E_FRAME_SIZE_MAX) {
1558                         rte_errno = EINVAL;
1559                         return i;
1560                 }
1561         }
1562         return i;
1563 }
1564
1565 /*********************************************************************
1566  *
1567  *  TX prep functions
1568  *
1569  **********************************************************************/
1570 uint16_t
1571 i40e_prep_pkts(__rte_unused void *tx_queue, struct rte_mbuf **tx_pkts,
1572                 uint16_t nb_pkts)
1573 {
1574         int i, ret;
1575         uint64_t ol_flags;
1576         struct rte_mbuf *m;
1577
1578         for (i = 0; i < nb_pkts; i++) {
1579                 m = tx_pkts[i];
1580                 ol_flags = m->ol_flags;
1581
1582                 /* Check for m->nb_segs to not exceed the limits. */
1583                 if (!(ol_flags & PKT_TX_TCP_SEG)) {
1584                         if (m->nb_segs > I40E_TX_MAX_MTU_SEG ||
1585                             m->pkt_len > I40E_FRAME_SIZE_MAX) {
1586                                 rte_errno = EINVAL;
1587                                 return i;
1588                         }
1589                 } else if (m->nb_segs > I40E_TX_MAX_SEG ||
1590                            m->tso_segsz < I40E_MIN_TSO_MSS ||
1591                            m->tso_segsz > I40E_MAX_TSO_MSS ||
1592                            m->pkt_len > I40E_TSO_FRAME_SIZE_MAX) {
1593                         /* MSS outside the range (256B - 9674B) are considered
1594                          * malicious
1595                          */
1596                         rte_errno = EINVAL;
1597                         return i;
1598                 }
1599
1600                 if (ol_flags & I40E_TX_OFFLOAD_NOTSUP_MASK) {
1601                         rte_errno = ENOTSUP;
1602                         return i;
1603                 }
1604
1605                 /* check the size of packet */
1606                 if (m->pkt_len < I40E_TX_MIN_PKT_LEN) {
1607                         rte_errno = EINVAL;
1608                         return i;
1609                 }
1610
1611 #ifdef RTE_ETHDEV_DEBUG_TX
1612                 ret = rte_validate_tx_offload(m);
1613                 if (ret != 0) {
1614                         rte_errno = -ret;
1615                         return i;
1616                 }
1617 #endif
1618                 ret = rte_net_intel_cksum_prepare(m);
1619                 if (ret != 0) {
1620                         rte_errno = -ret;
1621                         return i;
1622                 }
1623         }
1624         return i;
1625 }
1626
1627 /*
1628  * Find the VSI the queue belongs to. 'queue_idx' is the queue index
1629  * application used, which assume having sequential ones. But from driver's
1630  * perspective, it's different. For example, q0 belongs to FDIR VSI, q1-q64
1631  * to MAIN VSI, , q65-96 to SRIOV VSIs, q97-128 to VMDQ VSIs. For application
1632  * running on host, q1-64 and q97-128 can be used, total 96 queues. They can
1633  * use queue_idx from 0 to 95 to access queues, while real queue would be
1634  * different. This function will do a queue mapping to find VSI the queue
1635  * belongs to.
1636  */
1637 static struct i40e_vsi*
1638 i40e_pf_get_vsi_by_qindex(struct i40e_pf *pf, uint16_t queue_idx)
1639 {
1640         /* the queue in MAIN VSI range */
1641         if (queue_idx < pf->main_vsi->nb_qps)
1642                 return pf->main_vsi;
1643
1644         queue_idx -= pf->main_vsi->nb_qps;
1645
1646         /* queue_idx is greater than VMDQ VSIs range */
1647         if (queue_idx > pf->nb_cfg_vmdq_vsi * pf->vmdq_nb_qps - 1) {
1648                 PMD_INIT_LOG(ERR, "queue_idx out of range. VMDQ configured?");
1649                 return NULL;
1650         }
1651
1652         return pf->vmdq[queue_idx / pf->vmdq_nb_qps].vsi;
1653 }
1654
1655 static uint16_t
1656 i40e_get_queue_offset_by_qindex(struct i40e_pf *pf, uint16_t queue_idx)
1657 {
1658         /* the queue in MAIN VSI range */
1659         if (queue_idx < pf->main_vsi->nb_qps)
1660                 return queue_idx;
1661
1662         /* It's VMDQ queues */
1663         queue_idx -= pf->main_vsi->nb_qps;
1664
1665         if (pf->nb_cfg_vmdq_vsi)
1666                 return queue_idx % pf->vmdq_nb_qps;
1667         else {
1668                 PMD_INIT_LOG(ERR, "Fail to get queue offset");
1669                 return (uint16_t)(-1);
1670         }
1671 }
1672
1673 int
1674 i40e_dev_rx_queue_start(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1675 {
1676         struct i40e_rx_queue *rxq;
1677         int err;
1678         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1679
1680         PMD_INIT_FUNC_TRACE();
1681
1682         rxq = dev->data->rx_queues[rx_queue_id];
1683         if (!rxq || !rxq->q_set) {
1684                 PMD_DRV_LOG(ERR, "RX queue %u not available or setup",
1685                             rx_queue_id);
1686                 return -EINVAL;
1687         }
1688
1689         if (rxq->rx_deferred_start)
1690                 PMD_DRV_LOG(WARNING, "RX queue %u is deferrd start",
1691                             rx_queue_id);
1692
1693         err = i40e_alloc_rx_queue_mbufs(rxq);
1694         if (err) {
1695                 PMD_DRV_LOG(ERR, "Failed to allocate RX queue mbuf");
1696                 return err;
1697         }
1698
1699         /* Init the RX tail regieter. */
1700         I40E_PCI_REG_WRITE(rxq->qrx_tail, rxq->nb_rx_desc - 1);
1701
1702         err = i40e_switch_rx_queue(hw, rxq->reg_idx, TRUE);
1703         if (err) {
1704                 PMD_DRV_LOG(ERR, "Failed to switch RX queue %u on",
1705                             rx_queue_id);
1706
1707                 i40e_rx_queue_release_mbufs(rxq);
1708                 i40e_reset_rx_queue(rxq);
1709                 return err;
1710         }
1711         dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
1712
1713         return 0;
1714 }
1715
1716 int
1717 i40e_dev_rx_queue_stop(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1718 {
1719         struct i40e_rx_queue *rxq;
1720         int err;
1721         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1722
1723         rxq = dev->data->rx_queues[rx_queue_id];
1724         if (!rxq || !rxq->q_set) {
1725                 PMD_DRV_LOG(ERR, "RX queue %u not available or setup",
1726                                 rx_queue_id);
1727                 return -EINVAL;
1728         }
1729
1730         /*
1731          * rx_queue_id is queue id application refers to, while
1732          * rxq->reg_idx is the real queue index.
1733          */
1734         err = i40e_switch_rx_queue(hw, rxq->reg_idx, FALSE);
1735         if (err) {
1736                 PMD_DRV_LOG(ERR, "Failed to switch RX queue %u off",
1737                             rx_queue_id);
1738                 return err;
1739         }
1740         i40e_rx_queue_release_mbufs(rxq);
1741         i40e_reset_rx_queue(rxq);
1742         dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
1743
1744         return 0;
1745 }
1746
1747 int
1748 i40e_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id)
1749 {
1750         int err;
1751         struct i40e_tx_queue *txq;
1752         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1753
1754         PMD_INIT_FUNC_TRACE();
1755
1756         txq = dev->data->tx_queues[tx_queue_id];
1757         if (!txq || !txq->q_set) {
1758                 PMD_DRV_LOG(ERR, "TX queue %u is not available or setup",
1759                             tx_queue_id);
1760                 return -EINVAL;
1761         }
1762
1763         if (txq->tx_deferred_start)
1764                 PMD_DRV_LOG(WARNING, "TX queue %u is deferrd start",
1765                             tx_queue_id);
1766
1767         /*
1768          * tx_queue_id is queue id application refers to, while
1769          * rxq->reg_idx is the real queue index.
1770          */
1771         err = i40e_switch_tx_queue(hw, txq->reg_idx, TRUE);
1772         if (err) {
1773                 PMD_DRV_LOG(ERR, "Failed to switch TX queue %u on",
1774                             tx_queue_id);
1775                 return err;
1776         }
1777         dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
1778
1779         return 0;
1780 }
1781
1782 int
1783 i40e_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id)
1784 {
1785         struct i40e_tx_queue *txq;
1786         int err;
1787         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1788
1789         txq = dev->data->tx_queues[tx_queue_id];
1790         if (!txq || !txq->q_set) {
1791                 PMD_DRV_LOG(ERR, "TX queue %u is not available or setup",
1792                         tx_queue_id);
1793                 return -EINVAL;
1794         }
1795
1796         /*
1797          * tx_queue_id is queue id application refers to, while
1798          * txq->reg_idx is the real queue index.
1799          */
1800         err = i40e_switch_tx_queue(hw, txq->reg_idx, FALSE);
1801         if (err) {
1802                 PMD_DRV_LOG(ERR, "Failed to switch TX queue %u of",
1803                             tx_queue_id);
1804                 return err;
1805         }
1806
1807         i40e_tx_queue_release_mbufs(txq);
1808         i40e_reset_tx_queue(txq);
1809         dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
1810
1811         return 0;
1812 }
1813
1814 const uint32_t *
1815 i40e_dev_supported_ptypes_get(struct rte_eth_dev *dev)
1816 {
1817         static const uint32_t ptypes[] = {
1818                 /* refers to i40e_rxd_pkt_type_mapping() */
1819                 RTE_PTYPE_L2_ETHER,
1820                 RTE_PTYPE_L2_ETHER_TIMESYNC,
1821                 RTE_PTYPE_L2_ETHER_LLDP,
1822                 RTE_PTYPE_L2_ETHER_ARP,
1823                 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
1824                 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
1825                 RTE_PTYPE_L4_FRAG,
1826                 RTE_PTYPE_L4_ICMP,
1827                 RTE_PTYPE_L4_NONFRAG,
1828                 RTE_PTYPE_L4_SCTP,
1829                 RTE_PTYPE_L4_TCP,
1830                 RTE_PTYPE_L4_UDP,
1831                 RTE_PTYPE_TUNNEL_GRENAT,
1832                 RTE_PTYPE_TUNNEL_IP,
1833                 RTE_PTYPE_INNER_L2_ETHER,
1834                 RTE_PTYPE_INNER_L2_ETHER_VLAN,
1835                 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
1836                 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
1837                 RTE_PTYPE_INNER_L4_FRAG,
1838                 RTE_PTYPE_INNER_L4_ICMP,
1839                 RTE_PTYPE_INNER_L4_NONFRAG,
1840                 RTE_PTYPE_INNER_L4_SCTP,
1841                 RTE_PTYPE_INNER_L4_TCP,
1842                 RTE_PTYPE_INNER_L4_UDP,
1843                 RTE_PTYPE_UNKNOWN
1844         };
1845
1846         if (dev->rx_pkt_burst == i40e_recv_pkts ||
1847 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
1848             dev->rx_pkt_burst == i40e_recv_pkts_bulk_alloc ||
1849 #endif
1850             dev->rx_pkt_burst == i40e_recv_scattered_pkts ||
1851             dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec ||
1852             dev->rx_pkt_burst == i40e_recv_pkts_vec ||
1853 #ifdef CC_AVX512_SUPPORT
1854             dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec_avx512 ||
1855             dev->rx_pkt_burst == i40e_recv_pkts_vec_avx512 ||
1856 #endif
1857             dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec_avx2 ||
1858             dev->rx_pkt_burst == i40e_recv_pkts_vec_avx2)
1859                 return ptypes;
1860         return NULL;
1861 }
1862
1863 static int
1864 i40e_dev_first_queue(uint16_t idx, void **queues, int num)
1865 {
1866         uint16_t i;
1867
1868         for (i = 0; i < num; i++) {
1869                 if (i != idx && queues[i])
1870                         return 0;
1871         }
1872
1873         return 1;
1874 }
1875
1876 static int
1877 i40e_dev_rx_queue_setup_runtime(struct rte_eth_dev *dev,
1878                                 struct i40e_rx_queue *rxq)
1879 {
1880         struct i40e_adapter *ad =
1881                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
1882         int use_def_burst_func =
1883                 check_rx_burst_bulk_alloc_preconditions(rxq);
1884         uint16_t buf_size =
1885                 (uint16_t)(rte_pktmbuf_data_room_size(rxq->mp) -
1886                            RTE_PKTMBUF_HEADROOM);
1887         int use_scattered_rx =
1888                 (rxq->max_pkt_len > buf_size);
1889
1890         if (i40e_rx_queue_init(rxq) != I40E_SUCCESS) {
1891                 PMD_DRV_LOG(ERR,
1892                             "Failed to do RX queue initialization");
1893                 return -EINVAL;
1894         }
1895
1896         if (i40e_dev_first_queue(rxq->queue_id,
1897                                  dev->data->rx_queues,
1898                                  dev->data->nb_rx_queues)) {
1899                 /**
1900                  * If it is the first queue to setup,
1901                  * set all flags to default and call
1902                  * i40e_set_rx_function.
1903                  */
1904                 ad->rx_bulk_alloc_allowed = true;
1905                 ad->rx_vec_allowed = true;
1906                 dev->data->scattered_rx = use_scattered_rx;
1907                 if (use_def_burst_func)
1908                         ad->rx_bulk_alloc_allowed = false;
1909                 i40e_set_rx_function(dev);
1910                 return 0;
1911         } else if (ad->rx_vec_allowed && !rte_is_power_of_2(rxq->nb_rx_desc)) {
1912                 PMD_DRV_LOG(ERR, "Vector mode is allowed, but descriptor"
1913                             " number %d of queue %d isn't power of 2",
1914                             rxq->nb_rx_desc, rxq->queue_id);
1915                 return -EINVAL;
1916         }
1917
1918         /* check bulk alloc conflict */
1919         if (ad->rx_bulk_alloc_allowed && use_def_burst_func) {
1920                 PMD_DRV_LOG(ERR, "Can't use default burst.");
1921                 return -EINVAL;
1922         }
1923         /* check scatterred conflict */
1924         if (!dev->data->scattered_rx && use_scattered_rx) {
1925                 PMD_DRV_LOG(ERR, "Scattered rx is required.");
1926                 return -EINVAL;
1927         }
1928         /* check vector conflict */
1929         if (ad->rx_vec_allowed && i40e_rxq_vec_setup(rxq)) {
1930                 PMD_DRV_LOG(ERR, "Failed vector rx setup.");
1931                 return -EINVAL;
1932         }
1933
1934         return 0;
1935 }
1936
1937 int
1938 i40e_dev_rx_queue_setup(struct rte_eth_dev *dev,
1939                         uint16_t queue_idx,
1940                         uint16_t nb_desc,
1941                         unsigned int socket_id,
1942                         const struct rte_eth_rxconf *rx_conf,
1943                         struct rte_mempool *mp)
1944 {
1945         struct i40e_adapter *ad =
1946                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
1947         struct i40e_vsi *vsi;
1948         struct i40e_pf *pf = NULL;
1949         struct i40e_rx_queue *rxq;
1950         const struct rte_memzone *rz;
1951         uint32_t ring_size;
1952         uint16_t len, i;
1953         uint16_t reg_idx, base, bsf, tc_mapping;
1954         int q_offset, use_def_burst_func = 1;
1955         uint64_t offloads;
1956
1957         offloads = rx_conf->offloads | dev->data->dev_conf.rxmode.offloads;
1958
1959         pf = I40E_DEV_PRIVATE_TO_PF(dev->data->dev_private);
1960         vsi = i40e_pf_get_vsi_by_qindex(pf, queue_idx);
1961         if (!vsi)
1962                 return -EINVAL;
1963         q_offset = i40e_get_queue_offset_by_qindex(pf, queue_idx);
1964         if (q_offset < 0)
1965                 return -EINVAL;
1966         reg_idx = vsi->base_queue + q_offset;
1967
1968         if (nb_desc % I40E_ALIGN_RING_DESC != 0 ||
1969             (nb_desc > I40E_MAX_RING_DESC) ||
1970             (nb_desc < I40E_MIN_RING_DESC)) {
1971                 PMD_DRV_LOG(ERR, "Number (%u) of receive descriptors is "
1972                             "invalid", nb_desc);
1973                 return -EINVAL;
1974         }
1975
1976         /* Free memory if needed */
1977         if (dev->data->rx_queues[queue_idx]) {
1978                 i40e_rx_queue_release(dev->data->rx_queues[queue_idx]);
1979                 dev->data->rx_queues[queue_idx] = NULL;
1980         }
1981
1982         /* Allocate the rx queue data structure */
1983         rxq = rte_zmalloc_socket("i40e rx queue",
1984                                  sizeof(struct i40e_rx_queue),
1985                                  RTE_CACHE_LINE_SIZE,
1986                                  socket_id);
1987         if (!rxq) {
1988                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
1989                             "rx queue data structure");
1990                 return -ENOMEM;
1991         }
1992         rxq->mp = mp;
1993         rxq->nb_rx_desc = nb_desc;
1994         rxq->rx_free_thresh = rx_conf->rx_free_thresh;
1995         rxq->queue_id = queue_idx;
1996         rxq->reg_idx = reg_idx;
1997         rxq->port_id = dev->data->port_id;
1998         if (dev->data->dev_conf.rxmode.offloads & DEV_RX_OFFLOAD_KEEP_CRC)
1999                 rxq->crc_len = RTE_ETHER_CRC_LEN;
2000         else
2001                 rxq->crc_len = 0;
2002         rxq->drop_en = rx_conf->rx_drop_en;
2003         rxq->vsi = vsi;
2004         rxq->rx_deferred_start = rx_conf->rx_deferred_start;
2005         rxq->offloads = offloads;
2006
2007         /* Allocate the maximun number of RX ring hardware descriptor. */
2008         len = I40E_MAX_RING_DESC;
2009
2010         /**
2011          * Allocating a little more memory because vectorized/bulk_alloc Rx
2012          * functions doesn't check boundaries each time.
2013          */
2014         len += RTE_PMD_I40E_RX_MAX_BURST;
2015
2016         ring_size = RTE_ALIGN(len * sizeof(union i40e_rx_desc),
2017                               I40E_DMA_MEM_ALIGN);
2018
2019         rz = rte_eth_dma_zone_reserve(dev, "rx_ring", queue_idx,
2020                               ring_size, I40E_RING_BASE_ALIGN, socket_id);
2021         if (!rz) {
2022                 i40e_rx_queue_release(rxq);
2023                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for RX");
2024                 return -ENOMEM;
2025         }
2026
2027         rxq->mz = rz;
2028         /* Zero all the descriptors in the ring. */
2029         memset(rz->addr, 0, ring_size);
2030
2031         rxq->rx_ring_phys_addr = rz->iova;
2032         rxq->rx_ring = (union i40e_rx_desc *)rz->addr;
2033
2034         len = (uint16_t)(nb_desc + RTE_PMD_I40E_RX_MAX_BURST);
2035
2036         /* Allocate the software ring. */
2037         rxq->sw_ring =
2038                 rte_zmalloc_socket("i40e rx sw ring",
2039                                    sizeof(struct i40e_rx_entry) * len,
2040                                    RTE_CACHE_LINE_SIZE,
2041                                    socket_id);
2042         if (!rxq->sw_ring) {
2043                 i40e_rx_queue_release(rxq);
2044                 PMD_DRV_LOG(ERR, "Failed to allocate memory for SW ring");
2045                 return -ENOMEM;
2046         }
2047
2048         i40e_reset_rx_queue(rxq);
2049         rxq->q_set = TRUE;
2050
2051         for (i = 0; i < I40E_MAX_TRAFFIC_CLASS; i++) {
2052                 if (!(vsi->enabled_tc & (1 << i)))
2053                         continue;
2054                 tc_mapping = rte_le_to_cpu_16(vsi->info.tc_mapping[i]);
2055                 base = (tc_mapping & I40E_AQ_VSI_TC_QUE_OFFSET_MASK) >>
2056                         I40E_AQ_VSI_TC_QUE_OFFSET_SHIFT;
2057                 bsf = (tc_mapping & I40E_AQ_VSI_TC_QUE_NUMBER_MASK) >>
2058                         I40E_AQ_VSI_TC_QUE_NUMBER_SHIFT;
2059
2060                 if (queue_idx >= base && queue_idx < (base + BIT(bsf)))
2061                         rxq->dcb_tc = i;
2062         }
2063
2064         if (dev->data->dev_started) {
2065                 if (i40e_dev_rx_queue_setup_runtime(dev, rxq)) {
2066                         i40e_rx_queue_release(rxq);
2067                         return -EINVAL;
2068                 }
2069         } else {
2070                 use_def_burst_func =
2071                         check_rx_burst_bulk_alloc_preconditions(rxq);
2072                 if (!use_def_burst_func) {
2073 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2074                         PMD_INIT_LOG(DEBUG,
2075                           "Rx Burst Bulk Alloc Preconditions are "
2076                           "satisfied. Rx Burst Bulk Alloc function will be "
2077                           "used on port=%d, queue=%d.",
2078                           rxq->port_id, rxq->queue_id);
2079 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2080                 } else {
2081                         PMD_INIT_LOG(DEBUG,
2082                           "Rx Burst Bulk Alloc Preconditions are "
2083                           "not satisfied, Scattered Rx is requested, "
2084                           "or RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC is "
2085                           "not enabled on port=%d, queue=%d.",
2086                           rxq->port_id, rxq->queue_id);
2087                         ad->rx_bulk_alloc_allowed = false;
2088                 }
2089         }
2090
2091         dev->data->rx_queues[queue_idx] = rxq;
2092         return 0;
2093 }
2094
2095 void
2096 i40e_dev_rx_queue_release(struct rte_eth_dev *dev, uint16_t qid)
2097 {
2098         i40e_rx_queue_release(dev->data->rx_queues[qid]);
2099 }
2100
2101 void
2102 i40e_dev_tx_queue_release(struct rte_eth_dev *dev, uint16_t qid)
2103 {
2104         i40e_tx_queue_release(dev->data->tx_queues[qid]);
2105 }
2106
2107 void
2108 i40e_rx_queue_release(void *rxq)
2109 {
2110         struct i40e_rx_queue *q = (struct i40e_rx_queue *)rxq;
2111
2112         if (!q) {
2113                 PMD_DRV_LOG(DEBUG, "Pointer to rxq is NULL");
2114                 return;
2115         }
2116
2117         i40e_rx_queue_release_mbufs(q);
2118         rte_free(q->sw_ring);
2119         rte_memzone_free(q->mz);
2120         rte_free(q);
2121 }
2122
2123 uint32_t
2124 i40e_dev_rx_queue_count(void *rx_queue)
2125 {
2126 #define I40E_RXQ_SCAN_INTERVAL 4
2127         volatile union i40e_rx_desc *rxdp;
2128         struct i40e_rx_queue *rxq;
2129         uint16_t desc = 0;
2130
2131         rxq = rx_queue;
2132         rxdp = &(rxq->rx_ring[rxq->rx_tail]);
2133         while ((desc < rxq->nb_rx_desc) &&
2134                 ((rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len) &
2135                 I40E_RXD_QW1_STATUS_MASK) >> I40E_RXD_QW1_STATUS_SHIFT) &
2136                                 (1 << I40E_RX_DESC_STATUS_DD_SHIFT)) {
2137                 /**
2138                  * Check the DD bit of a rx descriptor of each 4 in a group,
2139                  * to avoid checking too frequently and downgrading performance
2140                  * too much.
2141                  */
2142                 desc += I40E_RXQ_SCAN_INTERVAL;
2143                 rxdp += I40E_RXQ_SCAN_INTERVAL;
2144                 if (rxq->rx_tail + desc >= rxq->nb_rx_desc)
2145                         rxdp = &(rxq->rx_ring[rxq->rx_tail +
2146                                         desc - rxq->nb_rx_desc]);
2147         }
2148
2149         return desc;
2150 }
2151
2152 int
2153 i40e_dev_rx_descriptor_status(void *rx_queue, uint16_t offset)
2154 {
2155         struct i40e_rx_queue *rxq = rx_queue;
2156         volatile uint64_t *status;
2157         uint64_t mask;
2158         uint32_t desc;
2159
2160         if (unlikely(offset >= rxq->nb_rx_desc))
2161                 return -EINVAL;
2162
2163         if (offset >= rxq->nb_rx_desc - rxq->nb_rx_hold)
2164                 return RTE_ETH_RX_DESC_UNAVAIL;
2165
2166         desc = rxq->rx_tail + offset;
2167         if (desc >= rxq->nb_rx_desc)
2168                 desc -= rxq->nb_rx_desc;
2169
2170         status = &rxq->rx_ring[desc].wb.qword1.status_error_len;
2171         mask = rte_le_to_cpu_64((1ULL << I40E_RX_DESC_STATUS_DD_SHIFT)
2172                 << I40E_RXD_QW1_STATUS_SHIFT);
2173         if (*status & mask)
2174                 return RTE_ETH_RX_DESC_DONE;
2175
2176         return RTE_ETH_RX_DESC_AVAIL;
2177 }
2178
2179 int
2180 i40e_dev_tx_descriptor_status(void *tx_queue, uint16_t offset)
2181 {
2182         struct i40e_tx_queue *txq = tx_queue;
2183         volatile uint64_t *status;
2184         uint64_t mask, expect;
2185         uint32_t desc;
2186
2187         if (unlikely(offset >= txq->nb_tx_desc))
2188                 return -EINVAL;
2189
2190         desc = txq->tx_tail + offset;
2191         /* go to next desc that has the RS bit */
2192         desc = ((desc + txq->tx_rs_thresh - 1) / txq->tx_rs_thresh) *
2193                 txq->tx_rs_thresh;
2194         if (desc >= txq->nb_tx_desc) {
2195                 desc -= txq->nb_tx_desc;
2196                 if (desc >= txq->nb_tx_desc)
2197                         desc -= txq->nb_tx_desc;
2198         }
2199
2200         status = &txq->tx_ring[desc].cmd_type_offset_bsz;
2201         mask = rte_le_to_cpu_64(I40E_TXD_QW1_DTYPE_MASK);
2202         expect = rte_cpu_to_le_64(
2203                 I40E_TX_DESC_DTYPE_DESC_DONE << I40E_TXD_QW1_DTYPE_SHIFT);
2204         if ((*status & mask) == expect)
2205                 return RTE_ETH_TX_DESC_DONE;
2206
2207         return RTE_ETH_TX_DESC_FULL;
2208 }
2209
2210 static int
2211 i40e_dev_tx_queue_setup_runtime(struct rte_eth_dev *dev,
2212                                 struct i40e_tx_queue *txq)
2213 {
2214         struct i40e_adapter *ad =
2215                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
2216
2217         if (i40e_tx_queue_init(txq) != I40E_SUCCESS) {
2218                 PMD_DRV_LOG(ERR,
2219                             "Failed to do TX queue initialization");
2220                 return -EINVAL;
2221         }
2222
2223         if (i40e_dev_first_queue(txq->queue_id,
2224                                  dev->data->tx_queues,
2225                                  dev->data->nb_tx_queues)) {
2226                 /**
2227                  * If it is the first queue to setup,
2228                  * set all flags and call
2229                  * i40e_set_tx_function.
2230                  */
2231                 i40e_set_tx_function_flag(dev, txq);
2232                 i40e_set_tx_function(dev);
2233                 return 0;
2234         }
2235
2236         /* check vector conflict */
2237         if (ad->tx_vec_allowed) {
2238                 if (txq->tx_rs_thresh > RTE_I40E_TX_MAX_FREE_BUF_SZ ||
2239                     i40e_txq_vec_setup(txq)) {
2240                         PMD_DRV_LOG(ERR, "Failed vector tx setup.");
2241                         return -EINVAL;
2242                 }
2243         }
2244         /* check simple tx conflict */
2245         if (ad->tx_simple_allowed) {
2246                 if ((txq->offloads & ~DEV_TX_OFFLOAD_MBUF_FAST_FREE) != 0 ||
2247                                 txq->tx_rs_thresh < RTE_PMD_I40E_TX_MAX_BURST) {
2248                         PMD_DRV_LOG(ERR, "No-simple tx is required.");
2249                         return -EINVAL;
2250                 }
2251         }
2252
2253         return 0;
2254 }
2255
2256 int
2257 i40e_dev_tx_queue_setup(struct rte_eth_dev *dev,
2258                         uint16_t queue_idx,
2259                         uint16_t nb_desc,
2260                         unsigned int socket_id,
2261                         const struct rte_eth_txconf *tx_conf)
2262 {
2263         struct i40e_vsi *vsi;
2264         struct i40e_pf *pf = NULL;
2265         struct i40e_tx_queue *txq;
2266         const struct rte_memzone *tz;
2267         uint32_t ring_size;
2268         uint16_t tx_rs_thresh, tx_free_thresh;
2269         uint16_t reg_idx, i, base, bsf, tc_mapping;
2270         int q_offset;
2271         uint64_t offloads;
2272
2273         offloads = tx_conf->offloads | dev->data->dev_conf.txmode.offloads;
2274
2275         pf = I40E_DEV_PRIVATE_TO_PF(dev->data->dev_private);
2276         vsi = i40e_pf_get_vsi_by_qindex(pf, queue_idx);
2277         if (!vsi)
2278                 return -EINVAL;
2279         q_offset = i40e_get_queue_offset_by_qindex(pf, queue_idx);
2280         if (q_offset < 0)
2281                 return -EINVAL;
2282         reg_idx = vsi->base_queue + q_offset;
2283
2284         if (nb_desc % I40E_ALIGN_RING_DESC != 0 ||
2285             (nb_desc > I40E_MAX_RING_DESC) ||
2286             (nb_desc < I40E_MIN_RING_DESC)) {
2287                 PMD_DRV_LOG(ERR, "Number (%u) of transmit descriptors is "
2288                             "invalid", nb_desc);
2289                 return -EINVAL;
2290         }
2291
2292         /**
2293          * The following two parameters control the setting of the RS bit on
2294          * transmit descriptors. TX descriptors will have their RS bit set
2295          * after txq->tx_rs_thresh descriptors have been used. The TX
2296          * descriptor ring will be cleaned after txq->tx_free_thresh
2297          * descriptors are used or if the number of descriptors required to
2298          * transmit a packet is greater than the number of free TX descriptors.
2299          *
2300          * The following constraints must be satisfied:
2301          *  - tx_rs_thresh must be greater than 0.
2302          *  - tx_rs_thresh must be less than the size of the ring minus 2.
2303          *  - tx_rs_thresh must be less than or equal to tx_free_thresh.
2304          *  - tx_rs_thresh must be a divisor of the ring size.
2305          *  - tx_free_thresh must be greater than 0.
2306          *  - tx_free_thresh must be less than the size of the ring minus 3.
2307          *  - tx_free_thresh + tx_rs_thresh must not exceed nb_desc.
2308          *
2309          * One descriptor in the TX ring is used as a sentinel to avoid a H/W
2310          * race condition, hence the maximum threshold constraints. When set
2311          * to zero use default values.
2312          */
2313         tx_free_thresh = (uint16_t)((tx_conf->tx_free_thresh) ?
2314                 tx_conf->tx_free_thresh : DEFAULT_TX_FREE_THRESH);
2315         /* force tx_rs_thresh to adapt an aggresive tx_free_thresh */
2316         tx_rs_thresh = (DEFAULT_TX_RS_THRESH + tx_free_thresh > nb_desc) ?
2317                 nb_desc - tx_free_thresh : DEFAULT_TX_RS_THRESH;
2318         if (tx_conf->tx_rs_thresh > 0)
2319                 tx_rs_thresh = tx_conf->tx_rs_thresh;
2320         if (tx_rs_thresh + tx_free_thresh > nb_desc) {
2321                 PMD_INIT_LOG(ERR, "tx_rs_thresh + tx_free_thresh must not "
2322                                 "exceed nb_desc. (tx_rs_thresh=%u "
2323                                 "tx_free_thresh=%u nb_desc=%u port=%d queue=%d)",
2324                                 (unsigned int)tx_rs_thresh,
2325                                 (unsigned int)tx_free_thresh,
2326                                 (unsigned int)nb_desc,
2327                                 (int)dev->data->port_id,
2328                                 (int)queue_idx);
2329                 return I40E_ERR_PARAM;
2330         }
2331         if (tx_rs_thresh >= (nb_desc - 2)) {
2332                 PMD_INIT_LOG(ERR, "tx_rs_thresh must be less than the "
2333                              "number of TX descriptors minus 2. "
2334                              "(tx_rs_thresh=%u port=%d queue=%d)",
2335                              (unsigned int)tx_rs_thresh,
2336                              (int)dev->data->port_id,
2337                              (int)queue_idx);
2338                 return I40E_ERR_PARAM;
2339         }
2340         if (tx_free_thresh >= (nb_desc - 3)) {
2341                 PMD_INIT_LOG(ERR, "tx_free_thresh must be less than the "
2342                              "number of TX descriptors minus 3. "
2343                              "(tx_free_thresh=%u port=%d queue=%d)",
2344                              (unsigned int)tx_free_thresh,
2345                              (int)dev->data->port_id,
2346                              (int)queue_idx);
2347                 return I40E_ERR_PARAM;
2348         }
2349         if (tx_rs_thresh > tx_free_thresh) {
2350                 PMD_INIT_LOG(ERR, "tx_rs_thresh must be less than or "
2351                              "equal to tx_free_thresh. (tx_free_thresh=%u"
2352                              " tx_rs_thresh=%u port=%d queue=%d)",
2353                              (unsigned int)tx_free_thresh,
2354                              (unsigned int)tx_rs_thresh,
2355                              (int)dev->data->port_id,
2356                              (int)queue_idx);
2357                 return I40E_ERR_PARAM;
2358         }
2359         if ((nb_desc % tx_rs_thresh) != 0) {
2360                 PMD_INIT_LOG(ERR, "tx_rs_thresh must be a divisor of the "
2361                              "number of TX descriptors. (tx_rs_thresh=%u"
2362                              " port=%d queue=%d)",
2363                              (unsigned int)tx_rs_thresh,
2364                              (int)dev->data->port_id,
2365                              (int)queue_idx);
2366                 return I40E_ERR_PARAM;
2367         }
2368         if ((tx_rs_thresh > 1) && (tx_conf->tx_thresh.wthresh != 0)) {
2369                 PMD_INIT_LOG(ERR, "TX WTHRESH must be set to 0 if "
2370                              "tx_rs_thresh is greater than 1. "
2371                              "(tx_rs_thresh=%u port=%d queue=%d)",
2372                              (unsigned int)tx_rs_thresh,
2373                              (int)dev->data->port_id,
2374                              (int)queue_idx);
2375                 return I40E_ERR_PARAM;
2376         }
2377
2378         /* Free memory if needed. */
2379         if (dev->data->tx_queues[queue_idx]) {
2380                 i40e_tx_queue_release(dev->data->tx_queues[queue_idx]);
2381                 dev->data->tx_queues[queue_idx] = NULL;
2382         }
2383
2384         /* Allocate the TX queue data structure. */
2385         txq = rte_zmalloc_socket("i40e tx queue",
2386                                   sizeof(struct i40e_tx_queue),
2387                                   RTE_CACHE_LINE_SIZE,
2388                                   socket_id);
2389         if (!txq) {
2390                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
2391                             "tx queue structure");
2392                 return -ENOMEM;
2393         }
2394
2395         /* Allocate TX hardware ring descriptors. */
2396         ring_size = sizeof(struct i40e_tx_desc) * I40E_MAX_RING_DESC;
2397         ring_size = RTE_ALIGN(ring_size, I40E_DMA_MEM_ALIGN);
2398         tz = rte_eth_dma_zone_reserve(dev, "tx_ring", queue_idx,
2399                               ring_size, I40E_RING_BASE_ALIGN, socket_id);
2400         if (!tz) {
2401                 i40e_tx_queue_release(txq);
2402                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for TX");
2403                 return -ENOMEM;
2404         }
2405
2406         txq->mz = tz;
2407         txq->nb_tx_desc = nb_desc;
2408         txq->tx_rs_thresh = tx_rs_thresh;
2409         txq->tx_free_thresh = tx_free_thresh;
2410         txq->pthresh = tx_conf->tx_thresh.pthresh;
2411         txq->hthresh = tx_conf->tx_thresh.hthresh;
2412         txq->wthresh = tx_conf->tx_thresh.wthresh;
2413         txq->queue_id = queue_idx;
2414         txq->reg_idx = reg_idx;
2415         txq->port_id = dev->data->port_id;
2416         txq->offloads = offloads;
2417         txq->vsi = vsi;
2418         txq->tx_deferred_start = tx_conf->tx_deferred_start;
2419
2420         txq->tx_ring_phys_addr = tz->iova;
2421         txq->tx_ring = (struct i40e_tx_desc *)tz->addr;
2422
2423         /* Allocate software ring */
2424         txq->sw_ring =
2425                 rte_zmalloc_socket("i40e tx sw ring",
2426                                    sizeof(struct i40e_tx_entry) * nb_desc,
2427                                    RTE_CACHE_LINE_SIZE,
2428                                    socket_id);
2429         if (!txq->sw_ring) {
2430                 i40e_tx_queue_release(txq);
2431                 PMD_DRV_LOG(ERR, "Failed to allocate memory for SW TX ring");
2432                 return -ENOMEM;
2433         }
2434
2435         i40e_reset_tx_queue(txq);
2436         txq->q_set = TRUE;
2437
2438         for (i = 0; i < I40E_MAX_TRAFFIC_CLASS; i++) {
2439                 if (!(vsi->enabled_tc & (1 << i)))
2440                         continue;
2441                 tc_mapping = rte_le_to_cpu_16(vsi->info.tc_mapping[i]);
2442                 base = (tc_mapping & I40E_AQ_VSI_TC_QUE_OFFSET_MASK) >>
2443                         I40E_AQ_VSI_TC_QUE_OFFSET_SHIFT;
2444                 bsf = (tc_mapping & I40E_AQ_VSI_TC_QUE_NUMBER_MASK) >>
2445                         I40E_AQ_VSI_TC_QUE_NUMBER_SHIFT;
2446
2447                 if (queue_idx >= base && queue_idx < (base + BIT(bsf)))
2448                         txq->dcb_tc = i;
2449         }
2450
2451         if (dev->data->dev_started) {
2452                 if (i40e_dev_tx_queue_setup_runtime(dev, txq)) {
2453                         i40e_tx_queue_release(txq);
2454                         return -EINVAL;
2455                 }
2456         } else {
2457                 /**
2458                  * Use a simple TX queue without offloads or
2459                  * multi segs if possible
2460                  */
2461                 i40e_set_tx_function_flag(dev, txq);
2462         }
2463         dev->data->tx_queues[queue_idx] = txq;
2464
2465         return 0;
2466 }
2467
2468 void
2469 i40e_tx_queue_release(void *txq)
2470 {
2471         struct i40e_tx_queue *q = (struct i40e_tx_queue *)txq;
2472
2473         if (!q) {
2474                 PMD_DRV_LOG(DEBUG, "Pointer to TX queue is NULL");
2475                 return;
2476         }
2477
2478         i40e_tx_queue_release_mbufs(q);
2479         rte_free(q->sw_ring);
2480         rte_memzone_free(q->mz);
2481         rte_free(q);
2482 }
2483
2484 const struct rte_memzone *
2485 i40e_memzone_reserve(const char *name, uint32_t len, int socket_id)
2486 {
2487         const struct rte_memzone *mz;
2488
2489         mz = rte_memzone_lookup(name);
2490         if (mz)
2491                 return mz;
2492
2493         mz = rte_memzone_reserve_aligned(name, len, socket_id,
2494                         RTE_MEMZONE_IOVA_CONTIG, I40E_RING_BASE_ALIGN);
2495         return mz;
2496 }
2497
2498 void
2499 i40e_rx_queue_release_mbufs(struct i40e_rx_queue *rxq)
2500 {
2501         uint16_t i;
2502
2503         /* SSE Vector driver has a different way of releasing mbufs. */
2504         if (rxq->rx_using_sse) {
2505                 i40e_rx_queue_release_mbufs_vec(rxq);
2506                 return;
2507         }
2508
2509         if (!rxq->sw_ring) {
2510                 PMD_DRV_LOG(DEBUG, "Pointer to sw_ring is NULL");
2511                 return;
2512         }
2513
2514         for (i = 0; i < rxq->nb_rx_desc; i++) {
2515                 if (rxq->sw_ring[i].mbuf) {
2516                         rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
2517                         rxq->sw_ring[i].mbuf = NULL;
2518                 }
2519         }
2520 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2521         if (rxq->rx_nb_avail == 0)
2522                 return;
2523         for (i = 0; i < rxq->rx_nb_avail; i++) {
2524                 struct rte_mbuf *mbuf;
2525
2526                 mbuf = rxq->rx_stage[rxq->rx_next_avail + i];
2527                 rte_pktmbuf_free_seg(mbuf);
2528         }
2529         rxq->rx_nb_avail = 0;
2530 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2531 }
2532
2533 void
2534 i40e_reset_rx_queue(struct i40e_rx_queue *rxq)
2535 {
2536         unsigned i;
2537         uint16_t len;
2538
2539         if (!rxq) {
2540                 PMD_DRV_LOG(DEBUG, "Pointer to rxq is NULL");
2541                 return;
2542         }
2543
2544 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2545         if (check_rx_burst_bulk_alloc_preconditions(rxq) == 0)
2546                 len = (uint16_t)(rxq->nb_rx_desc + RTE_PMD_I40E_RX_MAX_BURST);
2547         else
2548 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2549                 len = rxq->nb_rx_desc;
2550
2551         for (i = 0; i < len * sizeof(union i40e_rx_desc); i++)
2552                 ((volatile char *)rxq->rx_ring)[i] = 0;
2553
2554         memset(&rxq->fake_mbuf, 0x0, sizeof(rxq->fake_mbuf));
2555         for (i = 0; i < RTE_PMD_I40E_RX_MAX_BURST; ++i)
2556                 rxq->sw_ring[rxq->nb_rx_desc + i].mbuf = &rxq->fake_mbuf;
2557
2558 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2559         rxq->rx_nb_avail = 0;
2560         rxq->rx_next_avail = 0;
2561         rxq->rx_free_trigger = (uint16_t)(rxq->rx_free_thresh - 1);
2562 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2563         rxq->rx_tail = 0;
2564         rxq->nb_rx_hold = 0;
2565
2566         if (rxq->pkt_first_seg != NULL)
2567                 rte_pktmbuf_free(rxq->pkt_first_seg);
2568
2569         rxq->pkt_first_seg = NULL;
2570         rxq->pkt_last_seg = NULL;
2571
2572         rxq->rxrearm_start = 0;
2573         rxq->rxrearm_nb = 0;
2574 }
2575
2576 void
2577 i40e_tx_queue_release_mbufs(struct i40e_tx_queue *txq)
2578 {
2579         struct rte_eth_dev *dev;
2580         uint16_t i;
2581
2582         if (!txq || !txq->sw_ring) {
2583                 PMD_DRV_LOG(DEBUG, "Pointer to txq or sw_ring is NULL");
2584                 return;
2585         }
2586
2587         dev = &rte_eth_devices[txq->port_id];
2588
2589         /**
2590          *  vPMD tx will not set sw_ring's mbuf to NULL after free,
2591          *  so need to free remains more carefully.
2592          */
2593 #ifdef CC_AVX512_SUPPORT
2594         if (dev->tx_pkt_burst == i40e_xmit_pkts_vec_avx512) {
2595                 struct i40e_vec_tx_entry *swr = (void *)txq->sw_ring;
2596
2597                 i = txq->tx_next_dd - txq->tx_rs_thresh + 1;
2598                 if (txq->tx_tail < i) {
2599                         for (; i < txq->nb_tx_desc; i++) {
2600                                 rte_pktmbuf_free_seg(swr[i].mbuf);
2601                                 swr[i].mbuf = NULL;
2602                         }
2603                         i = 0;
2604                 }
2605                 for (; i < txq->tx_tail; i++) {
2606                         rte_pktmbuf_free_seg(swr[i].mbuf);
2607                         swr[i].mbuf = NULL;
2608                 }
2609                 return;
2610         }
2611 #endif
2612         if (dev->tx_pkt_burst == i40e_xmit_pkts_vec_avx2 ||
2613                         dev->tx_pkt_burst == i40e_xmit_pkts_vec) {
2614                 i = txq->tx_next_dd - txq->tx_rs_thresh + 1;
2615                 if (txq->tx_tail < i) {
2616                         for (; i < txq->nb_tx_desc; i++) {
2617                                 rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
2618                                 txq->sw_ring[i].mbuf = NULL;
2619                         }
2620                         i = 0;
2621                 }
2622                 for (; i < txq->tx_tail; i++) {
2623                         rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
2624                         txq->sw_ring[i].mbuf = NULL;
2625                 }
2626         } else {
2627                 for (i = 0; i < txq->nb_tx_desc; i++) {
2628                         if (txq->sw_ring[i].mbuf) {
2629                                 rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
2630                                 txq->sw_ring[i].mbuf = NULL;
2631                         }
2632                 }
2633         }
2634 }
2635
2636 static int
2637 i40e_tx_done_cleanup_full(struct i40e_tx_queue *txq,
2638                         uint32_t free_cnt)
2639 {
2640         struct i40e_tx_entry *swr_ring = txq->sw_ring;
2641         uint16_t i, tx_last, tx_id;
2642         uint16_t nb_tx_free_last;
2643         uint16_t nb_tx_to_clean;
2644         uint32_t pkt_cnt;
2645
2646         /* Start free mbuf from the next of tx_tail */
2647         tx_last = txq->tx_tail;
2648         tx_id  = swr_ring[tx_last].next_id;
2649
2650         if (txq->nb_tx_free == 0 && i40e_xmit_cleanup(txq))
2651                 return 0;
2652
2653         nb_tx_to_clean = txq->nb_tx_free;
2654         nb_tx_free_last = txq->nb_tx_free;
2655         if (!free_cnt)
2656                 free_cnt = txq->nb_tx_desc;
2657
2658         /* Loop through swr_ring to count the amount of
2659          * freeable mubfs and packets.
2660          */
2661         for (pkt_cnt = 0; pkt_cnt < free_cnt; ) {
2662                 for (i = 0; i < nb_tx_to_clean &&
2663                         pkt_cnt < free_cnt &&
2664                         tx_id != tx_last; i++) {
2665                         if (swr_ring[tx_id].mbuf != NULL) {
2666                                 rte_pktmbuf_free_seg(swr_ring[tx_id].mbuf);
2667                                 swr_ring[tx_id].mbuf = NULL;
2668
2669                                 /*
2670                                  * last segment in the packet,
2671                                  * increment packet count
2672                                  */
2673                                 pkt_cnt += (swr_ring[tx_id].last_id == tx_id);
2674                         }
2675
2676                         tx_id = swr_ring[tx_id].next_id;
2677                 }
2678
2679                 if (txq->tx_rs_thresh > txq->nb_tx_desc -
2680                         txq->nb_tx_free || tx_id == tx_last)
2681                         break;
2682
2683                 if (pkt_cnt < free_cnt) {
2684                         if (i40e_xmit_cleanup(txq))
2685                                 break;
2686
2687                         nb_tx_to_clean = txq->nb_tx_free - nb_tx_free_last;
2688                         nb_tx_free_last = txq->nb_tx_free;
2689                 }
2690         }
2691
2692         return (int)pkt_cnt;
2693 }
2694
2695 static int
2696 i40e_tx_done_cleanup_simple(struct i40e_tx_queue *txq,
2697                         uint32_t free_cnt)
2698 {
2699         int i, n, cnt;
2700
2701         if (free_cnt == 0 || free_cnt > txq->nb_tx_desc)
2702                 free_cnt = txq->nb_tx_desc;
2703
2704         cnt = free_cnt - free_cnt % txq->tx_rs_thresh;
2705
2706         for (i = 0; i < cnt; i += n) {
2707                 if (txq->nb_tx_desc - txq->nb_tx_free < txq->tx_rs_thresh)
2708                         break;
2709
2710                 n = i40e_tx_free_bufs(txq);
2711
2712                 if (n == 0)
2713                         break;
2714         }
2715
2716         return i;
2717 }
2718
2719 static int
2720 i40e_tx_done_cleanup_vec(struct i40e_tx_queue *txq __rte_unused,
2721                         uint32_t free_cnt __rte_unused)
2722 {
2723         return -ENOTSUP;
2724 }
2725 int
2726 i40e_tx_done_cleanup(void *txq, uint32_t free_cnt)
2727 {
2728         struct i40e_tx_queue *q = (struct i40e_tx_queue *)txq;
2729         struct rte_eth_dev *dev = &rte_eth_devices[q->port_id];
2730         struct i40e_adapter *ad =
2731                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
2732
2733         if (ad->tx_simple_allowed) {
2734                 if (ad->tx_vec_allowed)
2735                         return i40e_tx_done_cleanup_vec(q, free_cnt);
2736                 else
2737                         return i40e_tx_done_cleanup_simple(q, free_cnt);
2738         } else {
2739                 return i40e_tx_done_cleanup_full(q, free_cnt);
2740         }
2741 }
2742
2743 void
2744 i40e_reset_tx_queue(struct i40e_tx_queue *txq)
2745 {
2746         struct i40e_tx_entry *txe;
2747         uint16_t i, prev, size;
2748
2749         if (!txq) {
2750                 PMD_DRV_LOG(DEBUG, "Pointer to txq is NULL");
2751                 return;
2752         }
2753
2754         txe = txq->sw_ring;
2755         size = sizeof(struct i40e_tx_desc) * txq->nb_tx_desc;
2756         for (i = 0; i < size; i++)
2757                 ((volatile char *)txq->tx_ring)[i] = 0;
2758
2759         prev = (uint16_t)(txq->nb_tx_desc - 1);
2760         for (i = 0; i < txq->nb_tx_desc; i++) {
2761                 volatile struct i40e_tx_desc *txd = &txq->tx_ring[i];
2762
2763                 txd->cmd_type_offset_bsz =
2764                         rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE);
2765                 txe[i].mbuf =  NULL;
2766                 txe[i].last_id = i;
2767                 txe[prev].next_id = i;
2768                 prev = i;
2769         }
2770
2771         txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
2772         txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
2773
2774         txq->tx_tail = 0;
2775         txq->nb_tx_used = 0;
2776
2777         txq->last_desc_cleaned = (uint16_t)(txq->nb_tx_desc - 1);
2778         txq->nb_tx_free = (uint16_t)(txq->nb_tx_desc - 1);
2779 }
2780
2781 /* Init the TX queue in hardware */
2782 int
2783 i40e_tx_queue_init(struct i40e_tx_queue *txq)
2784 {
2785         enum i40e_status_code err = I40E_SUCCESS;
2786         struct i40e_vsi *vsi = txq->vsi;
2787         struct i40e_hw *hw = I40E_VSI_TO_HW(vsi);
2788         uint16_t pf_q = txq->reg_idx;
2789         struct i40e_hmc_obj_txq tx_ctx;
2790         uint32_t qtx_ctl;
2791
2792         /* clear the context structure first */
2793         memset(&tx_ctx, 0, sizeof(tx_ctx));
2794         tx_ctx.new_context = 1;
2795         tx_ctx.base = txq->tx_ring_phys_addr / I40E_QUEUE_BASE_ADDR_UNIT;
2796         tx_ctx.qlen = txq->nb_tx_desc;
2797
2798 #ifdef RTE_LIBRTE_IEEE1588
2799         tx_ctx.timesync_ena = 1;
2800 #endif
2801         tx_ctx.rdylist = rte_le_to_cpu_16(vsi->info.qs_handle[txq->dcb_tc]);
2802         if (vsi->type == I40E_VSI_FDIR)
2803                 tx_ctx.fd_ena = TRUE;
2804
2805         err = i40e_clear_lan_tx_queue_context(hw, pf_q);
2806         if (err != I40E_SUCCESS) {
2807                 PMD_DRV_LOG(ERR, "Failure of clean lan tx queue context");
2808                 return err;
2809         }
2810
2811         err = i40e_set_lan_tx_queue_context(hw, pf_q, &tx_ctx);
2812         if (err != I40E_SUCCESS) {
2813                 PMD_DRV_LOG(ERR, "Failure of set lan tx queue context");
2814                 return err;
2815         }
2816
2817         /* Now associate this queue with this PCI function */
2818         qtx_ctl = I40E_QTX_CTL_PF_QUEUE;
2819         qtx_ctl |= ((hw->pf_id << I40E_QTX_CTL_PF_INDX_SHIFT) &
2820                                         I40E_QTX_CTL_PF_INDX_MASK);
2821         I40E_WRITE_REG(hw, I40E_QTX_CTL(pf_q), qtx_ctl);
2822         I40E_WRITE_FLUSH(hw);
2823
2824         txq->qtx_tail = hw->hw_addr + I40E_QTX_TAIL(pf_q);
2825
2826         return err;
2827 }
2828
2829 int
2830 i40e_alloc_rx_queue_mbufs(struct i40e_rx_queue *rxq)
2831 {
2832         struct i40e_rx_entry *rxe = rxq->sw_ring;
2833         uint64_t dma_addr;
2834         uint16_t i;
2835
2836         for (i = 0; i < rxq->nb_rx_desc; i++) {
2837                 volatile union i40e_rx_desc *rxd;
2838                 struct rte_mbuf *mbuf = rte_mbuf_raw_alloc(rxq->mp);
2839
2840                 if (unlikely(!mbuf)) {
2841                         PMD_DRV_LOG(ERR, "Failed to allocate mbuf for RX");
2842                         return -ENOMEM;
2843                 }
2844
2845                 rte_mbuf_refcnt_set(mbuf, 1);
2846                 mbuf->next = NULL;
2847                 mbuf->data_off = RTE_PKTMBUF_HEADROOM;
2848                 mbuf->nb_segs = 1;
2849                 mbuf->port = rxq->port_id;
2850
2851                 dma_addr =
2852                         rte_cpu_to_le_64(rte_mbuf_data_iova_default(mbuf));
2853
2854                 rxd = &rxq->rx_ring[i];
2855                 rxd->read.pkt_addr = dma_addr;
2856                 rxd->read.hdr_addr = 0;
2857 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
2858                 rxd->read.rsvd1 = 0;
2859                 rxd->read.rsvd2 = 0;
2860 #endif /* RTE_LIBRTE_I40E_16BYTE_RX_DESC */
2861
2862                 rxe[i].mbuf = mbuf;
2863         }
2864
2865         return 0;
2866 }
2867
2868 /*
2869  * Calculate the buffer length, and check the jumbo frame
2870  * and maximum packet length.
2871  */
2872 static int
2873 i40e_rx_queue_config(struct i40e_rx_queue *rxq)
2874 {
2875         struct i40e_pf *pf = I40E_VSI_TO_PF(rxq->vsi);
2876         struct i40e_hw *hw = I40E_VSI_TO_HW(rxq->vsi);
2877         struct rte_eth_dev_data *data = pf->dev_data;
2878         uint16_t buf_size;
2879
2880         buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mp) -
2881                 RTE_PKTMBUF_HEADROOM);
2882
2883         switch (pf->flags & (I40E_FLAG_HEADER_SPLIT_DISABLED |
2884                         I40E_FLAG_HEADER_SPLIT_ENABLED)) {
2885         case I40E_FLAG_HEADER_SPLIT_ENABLED: /* Not supported */
2886                 rxq->rx_hdr_len = RTE_ALIGN(I40E_RXBUF_SZ_1024,
2887                                 (1 << I40E_RXQ_CTX_HBUFF_SHIFT));
2888                 rxq->rx_buf_len = RTE_ALIGN(I40E_RXBUF_SZ_2048,
2889                                 (1 << I40E_RXQ_CTX_DBUFF_SHIFT));
2890                 rxq->hs_mode = i40e_header_split_enabled;
2891                 break;
2892         case I40E_FLAG_HEADER_SPLIT_DISABLED:
2893         default:
2894                 rxq->rx_hdr_len = 0;
2895                 rxq->rx_buf_len = RTE_ALIGN_FLOOR(buf_size,
2896                         (1 << I40E_RXQ_CTX_DBUFF_SHIFT));
2897                 rxq->hs_mode = i40e_header_split_none;
2898                 break;
2899         }
2900
2901         rxq->max_pkt_len =
2902                 RTE_MIN(hw->func_caps.rx_buf_chain_len * rxq->rx_buf_len,
2903                                 data->mtu + I40E_ETH_OVERHEAD);
2904         if (rxq->max_pkt_len < RTE_ETHER_MIN_LEN ||
2905                 rxq->max_pkt_len > I40E_FRAME_SIZE_MAX) {
2906                 PMD_DRV_LOG(ERR, "maximum packet length must be "
2907                             "larger than %u and smaller than %u",
2908                             (uint32_t)RTE_ETHER_MIN_LEN,
2909                             (uint32_t)I40E_FRAME_SIZE_MAX);
2910                 return I40E_ERR_CONFIG;
2911         }
2912
2913         return 0;
2914 }
2915
2916 /* Init the RX queue in hardware */
2917 int
2918 i40e_rx_queue_init(struct i40e_rx_queue *rxq)
2919 {
2920         int err = I40E_SUCCESS;
2921         struct i40e_hw *hw = I40E_VSI_TO_HW(rxq->vsi);
2922         struct rte_eth_dev_data *dev_data = I40E_VSI_TO_DEV_DATA(rxq->vsi);
2923         uint16_t pf_q = rxq->reg_idx;
2924         uint16_t buf_size;
2925         struct i40e_hmc_obj_rxq rx_ctx;
2926
2927         err = i40e_rx_queue_config(rxq);
2928         if (err < 0) {
2929                 PMD_DRV_LOG(ERR, "Failed to config RX queue");
2930                 return err;
2931         }
2932
2933         /* Clear the context structure first */
2934         memset(&rx_ctx, 0, sizeof(struct i40e_hmc_obj_rxq));
2935         rx_ctx.dbuff = rxq->rx_buf_len >> I40E_RXQ_CTX_DBUFF_SHIFT;
2936         rx_ctx.hbuff = rxq->rx_hdr_len >> I40E_RXQ_CTX_HBUFF_SHIFT;
2937
2938         rx_ctx.base = rxq->rx_ring_phys_addr / I40E_QUEUE_BASE_ADDR_UNIT;
2939         rx_ctx.qlen = rxq->nb_rx_desc;
2940 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
2941         rx_ctx.dsize = 1;
2942 #endif
2943         rx_ctx.dtype = rxq->hs_mode;
2944         if (rxq->hs_mode)
2945                 rx_ctx.hsplit_0 = I40E_HEADER_SPLIT_ALL;
2946         else
2947                 rx_ctx.hsplit_0 = I40E_HEADER_SPLIT_NONE;
2948         rx_ctx.rxmax = rxq->max_pkt_len;
2949         rx_ctx.tphrdesc_ena = 1;
2950         rx_ctx.tphwdesc_ena = 1;
2951         rx_ctx.tphdata_ena = 1;
2952         rx_ctx.tphhead_ena = 1;
2953         rx_ctx.lrxqthresh = 2;
2954         rx_ctx.crcstrip = (rxq->crc_len == 0) ? 1 : 0;
2955         rx_ctx.l2tsel = 1;
2956         /* showiv indicates if inner VLAN is stripped inside of tunnel
2957          * packet. When set it to 1, vlan information is stripped from
2958          * the inner header, but the hardware does not put it in the
2959          * descriptor. So set it zero by default.
2960          */
2961         rx_ctx.showiv = 0;
2962         rx_ctx.prefena = 1;
2963
2964         err = i40e_clear_lan_rx_queue_context(hw, pf_q);
2965         if (err != I40E_SUCCESS) {
2966                 PMD_DRV_LOG(ERR, "Failed to clear LAN RX queue context");
2967                 return err;
2968         }
2969         err = i40e_set_lan_rx_queue_context(hw, pf_q, &rx_ctx);
2970         if (err != I40E_SUCCESS) {
2971                 PMD_DRV_LOG(ERR, "Failed to set LAN RX queue context");
2972                 return err;
2973         }
2974
2975         rxq->qrx_tail = hw->hw_addr + I40E_QRX_TAIL(pf_q);
2976
2977         buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mp) -
2978                 RTE_PKTMBUF_HEADROOM);
2979
2980         /* Check if scattered RX needs to be used. */
2981         if (rxq->max_pkt_len > buf_size)
2982                 dev_data->scattered_rx = 1;
2983
2984         /* Init the RX tail regieter. */
2985         I40E_PCI_REG_WRITE(rxq->qrx_tail, rxq->nb_rx_desc - 1);
2986
2987         return 0;
2988 }
2989
2990 void
2991 i40e_dev_clear_queues(struct rte_eth_dev *dev)
2992 {
2993         uint16_t i;
2994
2995         PMD_INIT_FUNC_TRACE();
2996
2997         for (i = 0; i < dev->data->nb_tx_queues; i++) {
2998                 if (!dev->data->tx_queues[i])
2999                         continue;
3000                 i40e_tx_queue_release_mbufs(dev->data->tx_queues[i]);
3001                 i40e_reset_tx_queue(dev->data->tx_queues[i]);
3002         }
3003
3004         for (i = 0; i < dev->data->nb_rx_queues; i++) {
3005                 if (!dev->data->rx_queues[i])
3006                         continue;
3007                 i40e_rx_queue_release_mbufs(dev->data->rx_queues[i]);
3008                 i40e_reset_rx_queue(dev->data->rx_queues[i]);
3009         }
3010 }
3011
3012 void
3013 i40e_dev_free_queues(struct rte_eth_dev *dev)
3014 {
3015         uint16_t i;
3016
3017         PMD_INIT_FUNC_TRACE();
3018
3019         for (i = 0; i < dev->data->nb_rx_queues; i++) {
3020                 if (!dev->data->rx_queues[i])
3021                         continue;
3022                 i40e_rx_queue_release(dev->data->rx_queues[i]);
3023                 dev->data->rx_queues[i] = NULL;
3024         }
3025
3026         for (i = 0; i < dev->data->nb_tx_queues; i++) {
3027                 if (!dev->data->tx_queues[i])
3028                         continue;
3029                 i40e_tx_queue_release(dev->data->tx_queues[i]);
3030                 dev->data->tx_queues[i] = NULL;
3031         }
3032 }
3033
3034 enum i40e_status_code
3035 i40e_fdir_setup_tx_resources(struct i40e_pf *pf)
3036 {
3037         struct i40e_tx_queue *txq;
3038         const struct rte_memzone *tz = NULL;
3039         struct rte_eth_dev *dev;
3040         uint32_t ring_size;
3041
3042         if (!pf) {
3043                 PMD_DRV_LOG(ERR, "PF is not available");
3044                 return I40E_ERR_BAD_PTR;
3045         }
3046
3047         dev = &rte_eth_devices[pf->dev_data->port_id];
3048
3049         /* Allocate the TX queue data structure. */
3050         txq = rte_zmalloc_socket("i40e fdir tx queue",
3051                                   sizeof(struct i40e_tx_queue),
3052                                   RTE_CACHE_LINE_SIZE,
3053                                   SOCKET_ID_ANY);
3054         if (!txq) {
3055                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
3056                                         "tx queue structure.");
3057                 return I40E_ERR_NO_MEMORY;
3058         }
3059
3060         /* Allocate TX hardware ring descriptors. */
3061         ring_size = sizeof(struct i40e_tx_desc) * I40E_FDIR_NUM_TX_DESC;
3062         ring_size = RTE_ALIGN(ring_size, I40E_DMA_MEM_ALIGN);
3063
3064         tz = rte_eth_dma_zone_reserve(dev, "fdir_tx_ring",
3065                                       I40E_FDIR_QUEUE_ID, ring_size,
3066                                       I40E_RING_BASE_ALIGN, SOCKET_ID_ANY);
3067         if (!tz) {
3068                 i40e_tx_queue_release(txq);
3069                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for TX.");
3070                 return I40E_ERR_NO_MEMORY;
3071         }
3072
3073         txq->mz = tz;
3074         txq->nb_tx_desc = I40E_FDIR_NUM_TX_DESC;
3075         txq->queue_id = I40E_FDIR_QUEUE_ID;
3076         txq->reg_idx = pf->fdir.fdir_vsi->base_queue;
3077         txq->vsi = pf->fdir.fdir_vsi;
3078
3079         txq->tx_ring_phys_addr = tz->iova;
3080         txq->tx_ring = (struct i40e_tx_desc *)tz->addr;
3081
3082         /*
3083          * don't need to allocate software ring and reset for the fdir
3084          * program queue just set the queue has been configured.
3085          */
3086         txq->q_set = TRUE;
3087         pf->fdir.txq = txq;
3088         pf->fdir.txq_available_buf_count = I40E_FDIR_PRG_PKT_CNT;
3089
3090         return I40E_SUCCESS;
3091 }
3092
3093 enum i40e_status_code
3094 i40e_fdir_setup_rx_resources(struct i40e_pf *pf)
3095 {
3096         struct i40e_rx_queue *rxq;
3097         const struct rte_memzone *rz = NULL;
3098         uint32_t ring_size;
3099         struct rte_eth_dev *dev;
3100
3101         if (!pf) {
3102                 PMD_DRV_LOG(ERR, "PF is not available");
3103                 return I40E_ERR_BAD_PTR;
3104         }
3105
3106         dev = &rte_eth_devices[pf->dev_data->port_id];
3107
3108         /* Allocate the RX queue data structure. */
3109         rxq = rte_zmalloc_socket("i40e fdir rx queue",
3110                                   sizeof(struct i40e_rx_queue),
3111                                   RTE_CACHE_LINE_SIZE,
3112                                   SOCKET_ID_ANY);
3113         if (!rxq) {
3114                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
3115                                         "rx queue structure.");
3116                 return I40E_ERR_NO_MEMORY;
3117         }
3118
3119         /* Allocate RX hardware ring descriptors. */
3120         ring_size = sizeof(union i40e_rx_desc) * I40E_FDIR_NUM_RX_DESC;
3121         ring_size = RTE_ALIGN(ring_size, I40E_DMA_MEM_ALIGN);
3122
3123         rz = rte_eth_dma_zone_reserve(dev, "fdir_rx_ring",
3124                                       I40E_FDIR_QUEUE_ID, ring_size,
3125                                       I40E_RING_BASE_ALIGN, SOCKET_ID_ANY);
3126         if (!rz) {
3127                 i40e_rx_queue_release(rxq);
3128                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for RX.");
3129                 return I40E_ERR_NO_MEMORY;
3130         }
3131
3132         rxq->mz = rz;
3133         rxq->nb_rx_desc = I40E_FDIR_NUM_RX_DESC;
3134         rxq->queue_id = I40E_FDIR_QUEUE_ID;
3135         rxq->reg_idx = pf->fdir.fdir_vsi->base_queue;
3136         rxq->vsi = pf->fdir.fdir_vsi;
3137
3138         rxq->rx_ring_phys_addr = rz->iova;
3139         memset(rz->addr, 0, I40E_FDIR_NUM_RX_DESC * sizeof(union i40e_rx_desc));
3140         rxq->rx_ring = (union i40e_rx_desc *)rz->addr;
3141
3142         /*
3143          * Don't need to allocate software ring and reset for the fdir
3144          * rx queue, just set the queue has been configured.
3145          */
3146         rxq->q_set = TRUE;
3147         pf->fdir.rxq = rxq;
3148
3149         return I40E_SUCCESS;
3150 }
3151
3152 void
3153 i40e_rxq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
3154         struct rte_eth_rxq_info *qinfo)
3155 {
3156         struct i40e_rx_queue *rxq;
3157
3158         rxq = dev->data->rx_queues[queue_id];
3159
3160         qinfo->mp = rxq->mp;
3161         qinfo->scattered_rx = dev->data->scattered_rx;
3162         qinfo->nb_desc = rxq->nb_rx_desc;
3163
3164         qinfo->conf.rx_free_thresh = rxq->rx_free_thresh;
3165         qinfo->conf.rx_drop_en = rxq->drop_en;
3166         qinfo->conf.rx_deferred_start = rxq->rx_deferred_start;
3167         qinfo->conf.offloads = rxq->offloads;
3168 }
3169
3170 void
3171 i40e_txq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
3172         struct rte_eth_txq_info *qinfo)
3173 {
3174         struct i40e_tx_queue *txq;
3175
3176         txq = dev->data->tx_queues[queue_id];
3177
3178         qinfo->nb_desc = txq->nb_tx_desc;
3179
3180         qinfo->conf.tx_thresh.pthresh = txq->pthresh;
3181         qinfo->conf.tx_thresh.hthresh = txq->hthresh;
3182         qinfo->conf.tx_thresh.wthresh = txq->wthresh;
3183
3184         qinfo->conf.tx_free_thresh = txq->tx_free_thresh;
3185         qinfo->conf.tx_rs_thresh = txq->tx_rs_thresh;
3186         qinfo->conf.tx_deferred_start = txq->tx_deferred_start;
3187         qinfo->conf.offloads = txq->offloads;
3188 }
3189
3190 #ifdef RTE_ARCH_X86
3191 static inline bool
3192 get_avx_supported(bool request_avx512)
3193 {
3194         if (request_avx512) {
3195                 if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_512 &&
3196                 rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1 &&
3197                 rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512BW) == 1)
3198 #ifdef CC_AVX512_SUPPORT
3199                         return true;
3200 #else
3201                 PMD_DRV_LOG(NOTICE,
3202                         "AVX512 is not supported in build env");
3203                 return false;
3204 #endif
3205         } else {
3206                 if (rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_256 &&
3207                 rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 &&
3208                 rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1)
3209 #ifdef CC_AVX2_SUPPORT
3210                         return true;
3211 #else
3212                 PMD_DRV_LOG(NOTICE,
3213                         "AVX2 is not supported in build env");
3214                 return false;
3215 #endif
3216         }
3217
3218         return false;
3219 }
3220 #endif /* RTE_ARCH_X86 */
3221
3222
3223 void __rte_cold
3224 i40e_set_rx_function(struct rte_eth_dev *dev)
3225 {
3226         struct i40e_adapter *ad =
3227                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3228         uint16_t rx_using_sse, i;
3229         /* In order to allow Vector Rx there are a few configuration
3230          * conditions to be met and Rx Bulk Allocation should be allowed.
3231          */
3232         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
3233 #ifdef RTE_ARCH_X86
3234                 ad->rx_use_avx512 = false;
3235                 ad->rx_use_avx2 = false;
3236 #endif
3237                 if (i40e_rx_vec_dev_conf_condition_check(dev) ||
3238                     !ad->rx_bulk_alloc_allowed) {
3239                         PMD_INIT_LOG(DEBUG, "Port[%d] doesn't meet"
3240                                      " Vector Rx preconditions",
3241                                      dev->data->port_id);
3242
3243                         ad->rx_vec_allowed = false;
3244                 }
3245                 if (ad->rx_vec_allowed) {
3246                         for (i = 0; i < dev->data->nb_rx_queues; i++) {
3247                                 struct i40e_rx_queue *rxq =
3248                                         dev->data->rx_queues[i];
3249
3250                                 if (rxq && i40e_rxq_vec_setup(rxq)) {
3251                                         ad->rx_vec_allowed = false;
3252                                         break;
3253                                 }
3254                         }
3255 #ifdef RTE_ARCH_X86
3256                         ad->rx_use_avx512 = get_avx_supported(1);
3257
3258                         if (!ad->rx_use_avx512)
3259                                 ad->rx_use_avx2 = get_avx_supported(0);
3260 #endif
3261                 }
3262         }
3263
3264         if (ad->rx_vec_allowed  &&
3265             rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
3266 #ifdef RTE_ARCH_X86
3267                 if (dev->data->scattered_rx) {
3268                         if (ad->rx_use_avx512) {
3269 #ifdef CC_AVX512_SUPPORT
3270                                 PMD_DRV_LOG(NOTICE,
3271                                         "Using AVX512 Vector Scattered Rx (port %d).",
3272                                         dev->data->port_id);
3273                                 dev->rx_pkt_burst =
3274                                         i40e_recv_scattered_pkts_vec_avx512;
3275 #endif
3276                         } else {
3277                                 PMD_INIT_LOG(DEBUG,
3278                                         "Using %sVector Scattered Rx (port %d).",
3279                                         ad->rx_use_avx2 ? "avx2 " : "",
3280                                         dev->data->port_id);
3281                                 dev->rx_pkt_burst = ad->rx_use_avx2 ?
3282                                         i40e_recv_scattered_pkts_vec_avx2 :
3283                                         i40e_recv_scattered_pkts_vec;
3284                         }
3285                 } else {
3286                         if (ad->rx_use_avx512) {
3287 #ifdef CC_AVX512_SUPPORT
3288                                 PMD_DRV_LOG(NOTICE,
3289                                         "Using AVX512 Vector Rx (port %d).",
3290                                         dev->data->port_id);
3291                                 dev->rx_pkt_burst =
3292                                         i40e_recv_pkts_vec_avx512;
3293 #endif
3294                         } else {
3295                                 PMD_INIT_LOG(DEBUG,
3296                                         "Using %sVector Rx (port %d).",
3297                                         ad->rx_use_avx2 ? "avx2 " : "",
3298                                         dev->data->port_id);
3299                                 dev->rx_pkt_burst = ad->rx_use_avx2 ?
3300                                         i40e_recv_pkts_vec_avx2 :
3301                                         i40e_recv_pkts_vec;
3302                         }
3303                 }
3304 #else /* RTE_ARCH_X86 */
3305                 if (dev->data->scattered_rx) {
3306                         PMD_INIT_LOG(DEBUG,
3307                                      "Using Vector Scattered Rx (port %d).",
3308                                      dev->data->port_id);
3309                         dev->rx_pkt_burst = i40e_recv_scattered_pkts_vec;
3310                 } else {
3311                         PMD_INIT_LOG(DEBUG, "Using Vector Rx (port %d).",
3312                                      dev->data->port_id);
3313                         dev->rx_pkt_burst = i40e_recv_pkts_vec;
3314                 }
3315 #endif /* RTE_ARCH_X86 */
3316         } else if (!dev->data->scattered_rx && ad->rx_bulk_alloc_allowed) {
3317                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are "
3318                                     "satisfied. Rx Burst Bulk Alloc function "
3319                                     "will be used on port=%d.",
3320                              dev->data->port_id);
3321
3322                 dev->rx_pkt_burst = i40e_recv_pkts_bulk_alloc;
3323         } else {
3324                 /* Simple Rx Path. */
3325                 PMD_INIT_LOG(DEBUG, "Simple Rx path will be used on port=%d.",
3326                              dev->data->port_id);
3327                 dev->rx_pkt_burst = dev->data->scattered_rx ?
3328                                         i40e_recv_scattered_pkts :
3329                                         i40e_recv_pkts;
3330         }
3331
3332         /* Propagate information about RX function choice through all queues. */
3333         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
3334                 rx_using_sse =
3335                         (dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec ||
3336                          dev->rx_pkt_burst == i40e_recv_pkts_vec ||
3337 #ifdef CC_AVX512_SUPPORT
3338                          dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec_avx512 ||
3339                          dev->rx_pkt_burst == i40e_recv_pkts_vec_avx512 ||
3340 #endif
3341                          dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec_avx2 ||
3342                          dev->rx_pkt_burst == i40e_recv_pkts_vec_avx2);
3343
3344                 for (i = 0; i < dev->data->nb_rx_queues; i++) {
3345                         struct i40e_rx_queue *rxq = dev->data->rx_queues[i];
3346
3347                         if (rxq)
3348                                 rxq->rx_using_sse = rx_using_sse;
3349                 }
3350         }
3351 }
3352
3353 static const struct {
3354         eth_rx_burst_t pkt_burst;
3355         const char *info;
3356 } i40e_rx_burst_infos[] = {
3357         { i40e_recv_scattered_pkts,          "Scalar Scattered" },
3358         { i40e_recv_pkts_bulk_alloc,         "Scalar Bulk Alloc" },
3359         { i40e_recv_pkts,                    "Scalar" },
3360 #ifdef RTE_ARCH_X86
3361 #ifdef CC_AVX512_SUPPORT
3362         { i40e_recv_scattered_pkts_vec_avx512, "Vector AVX512 Scattered" },
3363         { i40e_recv_pkts_vec_avx512,           "Vector AVX512" },
3364 #endif
3365         { i40e_recv_scattered_pkts_vec_avx2, "Vector AVX2 Scattered" },
3366         { i40e_recv_pkts_vec_avx2,           "Vector AVX2" },
3367         { i40e_recv_scattered_pkts_vec,      "Vector SSE Scattered" },
3368         { i40e_recv_pkts_vec,                "Vector SSE" },
3369 #elif defined(RTE_ARCH_ARM64)
3370         { i40e_recv_scattered_pkts_vec,      "Vector Neon Scattered" },
3371         { i40e_recv_pkts_vec,                "Vector Neon" },
3372 #elif defined(RTE_ARCH_PPC_64)
3373         { i40e_recv_scattered_pkts_vec,      "Vector AltiVec Scattered" },
3374         { i40e_recv_pkts_vec,                "Vector AltiVec" },
3375 #endif
3376 };
3377
3378 int
3379 i40e_rx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id,
3380                        struct rte_eth_burst_mode *mode)
3381 {
3382         eth_rx_burst_t pkt_burst = dev->rx_pkt_burst;
3383         int ret = -EINVAL;
3384         unsigned int i;
3385
3386         for (i = 0; i < RTE_DIM(i40e_rx_burst_infos); ++i) {
3387                 if (pkt_burst == i40e_rx_burst_infos[i].pkt_burst) {
3388                         snprintf(mode->info, sizeof(mode->info), "%s",
3389                                  i40e_rx_burst_infos[i].info);
3390                         ret = 0;
3391                         break;
3392                 }
3393         }
3394
3395         return ret;
3396 }
3397
3398 void __rte_cold
3399 i40e_set_tx_function_flag(struct rte_eth_dev *dev, struct i40e_tx_queue *txq)
3400 {
3401         struct i40e_adapter *ad =
3402                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3403
3404         /* Use a simple Tx queue if possible (only fast free is allowed) */
3405         ad->tx_simple_allowed =
3406                 (txq->offloads ==
3407                  (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE) &&
3408                  txq->tx_rs_thresh >= RTE_PMD_I40E_TX_MAX_BURST);
3409         ad->tx_vec_allowed = (ad->tx_simple_allowed &&
3410                         txq->tx_rs_thresh <= RTE_I40E_TX_MAX_FREE_BUF_SZ);
3411
3412         if (ad->tx_vec_allowed)
3413                 PMD_INIT_LOG(DEBUG, "Vector Tx can be enabled on Tx queue %u.",
3414                                 txq->queue_id);
3415         else if (ad->tx_simple_allowed)
3416                 PMD_INIT_LOG(DEBUG, "Simple Tx can be enabled on Tx queue %u.",
3417                                 txq->queue_id);
3418         else
3419                 PMD_INIT_LOG(DEBUG,
3420                                 "Neither simple nor vector Tx enabled on Tx queue %u\n",
3421                                 txq->queue_id);
3422 }
3423
3424 void __rte_cold
3425 i40e_set_tx_function(struct rte_eth_dev *dev)
3426 {
3427         struct i40e_adapter *ad =
3428                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3429         int i;
3430
3431         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
3432 #ifdef RTE_ARCH_X86
3433                 ad->tx_use_avx2 = false;
3434                 ad->tx_use_avx512 = false;
3435 #endif
3436                 if (ad->tx_vec_allowed) {
3437                         for (i = 0; i < dev->data->nb_tx_queues; i++) {
3438                                 struct i40e_tx_queue *txq =
3439                                         dev->data->tx_queues[i];
3440
3441                                 if (txq && i40e_txq_vec_setup(txq)) {
3442                                         ad->tx_vec_allowed = false;
3443                                         break;
3444                                 }
3445                         }
3446 #ifdef RTE_ARCH_X86
3447                         ad->tx_use_avx512 = get_avx_supported(1);
3448
3449                         if (!ad->tx_use_avx512)
3450                                 ad->tx_use_avx2 = get_avx_supported(0);
3451 #endif
3452                 }
3453         }
3454
3455         if (ad->tx_simple_allowed) {
3456                 if (ad->tx_vec_allowed &&
3457                     rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
3458 #ifdef RTE_ARCH_X86
3459                         if (ad->tx_use_avx512) {
3460 #ifdef CC_AVX512_SUPPORT
3461                                 PMD_DRV_LOG(NOTICE, "Using AVX512 Vector Tx (port %d).",
3462                                             dev->data->port_id);
3463                                 dev->tx_pkt_burst = i40e_xmit_pkts_vec_avx512;
3464 #endif
3465                         } else {
3466                                 PMD_INIT_LOG(DEBUG, "Using %sVector Tx (port %d).",
3467                                              ad->tx_use_avx2 ? "avx2 " : "",
3468                                              dev->data->port_id);
3469                                 dev->tx_pkt_burst = ad->tx_use_avx2 ?
3470                                                     i40e_xmit_pkts_vec_avx2 :
3471                                                     i40e_xmit_pkts_vec;
3472                         }
3473 #else /* RTE_ARCH_X86 */
3474                         PMD_INIT_LOG(DEBUG, "Using Vector Tx (port %d).",
3475                                      dev->data->port_id);
3476                         dev->tx_pkt_burst = i40e_xmit_pkts_vec;
3477 #endif /* RTE_ARCH_X86 */
3478                 } else {
3479                         PMD_INIT_LOG(DEBUG, "Simple tx finally be used.");
3480                         dev->tx_pkt_burst = i40e_xmit_pkts_simple;
3481                 }
3482                 dev->tx_pkt_prepare = i40e_simple_prep_pkts;
3483         } else {
3484                 PMD_INIT_LOG(DEBUG, "Xmit tx finally be used.");
3485                 dev->tx_pkt_burst = i40e_xmit_pkts;
3486                 dev->tx_pkt_prepare = i40e_prep_pkts;
3487         }
3488 }
3489
3490 static const struct {
3491         eth_tx_burst_t pkt_burst;
3492         const char *info;
3493 } i40e_tx_burst_infos[] = {
3494         { i40e_xmit_pkts_simple,   "Scalar Simple" },
3495         { i40e_xmit_pkts,          "Scalar" },
3496 #ifdef RTE_ARCH_X86
3497 #ifdef CC_AVX512_SUPPORT
3498         { i40e_xmit_pkts_vec_avx512, "Vector AVX512" },
3499 #endif
3500         { i40e_xmit_pkts_vec_avx2, "Vector AVX2" },
3501         { i40e_xmit_pkts_vec,      "Vector SSE" },
3502 #elif defined(RTE_ARCH_ARM64)
3503         { i40e_xmit_pkts_vec,      "Vector Neon" },
3504 #elif defined(RTE_ARCH_PPC_64)
3505         { i40e_xmit_pkts_vec,      "Vector AltiVec" },
3506 #endif
3507 };
3508
3509 int
3510 i40e_tx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id,
3511                        struct rte_eth_burst_mode *mode)
3512 {
3513         eth_tx_burst_t pkt_burst = dev->tx_pkt_burst;
3514         int ret = -EINVAL;
3515         unsigned int i;
3516
3517         for (i = 0; i < RTE_DIM(i40e_tx_burst_infos); ++i) {
3518                 if (pkt_burst == i40e_tx_burst_infos[i].pkt_burst) {
3519                         snprintf(mode->info, sizeof(mode->info), "%s",
3520                                  i40e_tx_burst_infos[i].info);
3521                         ret = 0;
3522                         break;
3523                 }
3524         }
3525
3526         return ret;
3527 }
3528
3529 void __rte_cold
3530 i40e_set_default_ptype_table(struct rte_eth_dev *dev)
3531 {
3532         struct i40e_adapter *ad =
3533                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3534         int i;
3535
3536         for (i = 0; i < I40E_MAX_PKT_TYPE; i++)
3537                 ad->ptype_tbl[i] = i40e_get_default_pkt_type(i);
3538 }
3539
3540 void __rte_cold
3541 i40e_set_default_pctype_table(struct rte_eth_dev *dev)
3542 {
3543         struct i40e_adapter *ad =
3544                         I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3545         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3546         int i;
3547
3548         for (i = 0; i < I40E_FLOW_TYPE_MAX; i++)
3549                 ad->pctypes_tbl[i] = 0ULL;
3550         ad->flow_types_mask = 0ULL;
3551         ad->pctypes_mask = 0ULL;
3552
3553         ad->pctypes_tbl[RTE_ETH_FLOW_FRAG_IPV4] =
3554                                 (1ULL << I40E_FILTER_PCTYPE_FRAG_IPV4);
3555         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_UDP] =
3556                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_UDP);
3557         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_TCP] =
3558                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_TCP);
3559         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_SCTP] =
3560                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_SCTP);
3561         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_OTHER] =
3562                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_OTHER);
3563         ad->pctypes_tbl[RTE_ETH_FLOW_FRAG_IPV6] =
3564                                 (1ULL << I40E_FILTER_PCTYPE_FRAG_IPV6);
3565         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_UDP] =
3566                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_UDP);
3567         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_TCP] =
3568                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_TCP);
3569         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_SCTP] =
3570                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_SCTP);
3571         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_OTHER] =
3572                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_OTHER);
3573         ad->pctypes_tbl[RTE_ETH_FLOW_L2_PAYLOAD] =
3574                                 (1ULL << I40E_FILTER_PCTYPE_L2_PAYLOAD);
3575
3576         if (hw->mac.type == I40E_MAC_X722 ||
3577                 hw->mac.type == I40E_MAC_X722_VF) {
3578                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_UDP] |=
3579                         (1ULL << I40E_FILTER_PCTYPE_NONF_UNICAST_IPV4_UDP);
3580                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_UDP] |=
3581                         (1ULL << I40E_FILTER_PCTYPE_NONF_MULTICAST_IPV4_UDP);
3582                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_TCP] |=
3583                         (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_TCP_SYN_NO_ACK);
3584                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_UDP] |=
3585                         (1ULL << I40E_FILTER_PCTYPE_NONF_UNICAST_IPV6_UDP);
3586                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_UDP] |=
3587                         (1ULL << I40E_FILTER_PCTYPE_NONF_MULTICAST_IPV6_UDP);
3588                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_TCP] |=
3589                         (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_TCP_SYN_NO_ACK);
3590         }
3591
3592         for (i = 0; i < I40E_FLOW_TYPE_MAX; i++) {
3593                 if (ad->pctypes_tbl[i])
3594                         ad->flow_types_mask |= (1ULL << i);
3595                 ad->pctypes_mask |= ad->pctypes_tbl[i];
3596         }
3597 }
3598
3599 #ifndef CC_AVX2_SUPPORT
3600 uint16_t
3601 i40e_recv_pkts_vec_avx2(void __rte_unused *rx_queue,
3602                         struct rte_mbuf __rte_unused **rx_pkts,
3603                         uint16_t __rte_unused nb_pkts)
3604 {
3605         return 0;
3606 }
3607
3608 uint16_t
3609 i40e_recv_scattered_pkts_vec_avx2(void __rte_unused *rx_queue,
3610                         struct rte_mbuf __rte_unused **rx_pkts,
3611                         uint16_t __rte_unused nb_pkts)
3612 {
3613         return 0;
3614 }
3615
3616 uint16_t
3617 i40e_xmit_pkts_vec_avx2(void __rte_unused * tx_queue,
3618                           struct rte_mbuf __rte_unused **tx_pkts,
3619                           uint16_t __rte_unused nb_pkts)
3620 {
3621         return 0;
3622 }
3623 #endif /* ifndef CC_AVX2_SUPPORT */