drivers/net/cnxk/cn9k_tx.h

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(C) 2021 Marvell.
   3  */
   4 #ifndef __CN9K_TX_H__
   5 #define __CN9K_TX_H__
   6
   7 #include <rte_vect.h>
   8
   9 #define NIX_TX_OFFLOAD_NONE           (0)
  10 #define NIX_TX_OFFLOAD_L3_L4_CSUM_F   BIT(0)
  11 #define NIX_TX_OFFLOAD_OL3_OL4_CSUM_F BIT(1)
  12 #define NIX_TX_OFFLOAD_VLAN_QINQ_F    BIT(2)
  13 #define NIX_TX_OFFLOAD_MBUF_NOFF_F    BIT(3)
  14 #define NIX_TX_OFFLOAD_TSO_F          BIT(4)
  15 #define NIX_TX_OFFLOAD_TSTAMP_F       BIT(5)
  16
  17 /* Flags to control xmit_prepare function.
  18  * Defining it from backwards to denote its been
  19  * not used as offload flags to pick function
  20  */
  21 #define NIX_TX_MULTI_SEG_F BIT(15)
  22
  23 #define NIX_TX_NEED_SEND_HDR_W1                                                \
  24         (NIX_TX_OFFLOAD_L3_L4_CSUM_F | NIX_TX_OFFLOAD_OL3_OL4_CSUM_F |         \
  25          NIX_TX_OFFLOAD_VLAN_QINQ_F | NIX_TX_OFFLOAD_TSO_F)
  26
  27 #define NIX_TX_NEED_EXT_HDR                                                    \
  28         (NIX_TX_OFFLOAD_VLAN_QINQ_F | NIX_TX_OFFLOAD_TSTAMP_F |                \
  29          NIX_TX_OFFLOAD_TSO_F)
  30
  31 #define NIX_XMIT_FC_OR_RETURN(txq, pkts)                                       \
  32         do {                                                                   \
  33                 /* Cached value is low, Update the fc_cache_pkts */            \
  34                 if (unlikely((txq)->fc_cache_pkts < (pkts))) {                 \
  35                         /* Multiply with sqe_per_sqb to express in pkts */     \
  36                         (txq)->fc_cache_pkts =                                 \
  37                                 ((txq)->nb_sqb_bufs_adj - *(txq)->fc_mem)      \
  38                                 << (txq)->sqes_per_sqb_log2;                   \
  39                         /* Check it again for the room */                      \
  40                         if (unlikely((txq)->fc_cache_pkts < (pkts)))           \
  41                                 return 0;                                      \
  42                 }                                                              \
  43         } while (0)
  44
  45 /* Function to determine no of tx subdesc required in case ext
  46  * sub desc is enabled.
  47  */
  48 static __rte_always_inline int
  49 cn9k_nix_tx_ext_subs(const uint16_t flags)
  50 {
  51         return (flags & NIX_TX_OFFLOAD_TSTAMP_F)
  52                        ? 2
  53                        : ((flags &
  54                            (NIX_TX_OFFLOAD_VLAN_QINQ_F | NIX_TX_OFFLOAD_TSO_F))
  55                                   ? 1
  56                                   : 0);
  57 }
  58
  59 static __rte_always_inline void
  60 cn9k_nix_xmit_prepare_tso(struct rte_mbuf *m, const uint64_t flags)
  61 {
  62         uint64_t mask, ol_flags = m->ol_flags;
  63
  64         if (flags & NIX_TX_OFFLOAD_TSO_F && (ol_flags & PKT_TX_TCP_SEG)) {
  65                 uintptr_t mdata = rte_pktmbuf_mtod(m, uintptr_t);
  66                 uint16_t *iplen, *oiplen, *oudplen;
  67                 uint16_t lso_sb, paylen;
  68
  69                 mask = -!!(ol_flags & (PKT_TX_OUTER_IPV4 | PKT_TX_OUTER_IPV6));
  70                 lso_sb = (mask & (m->outer_l2_len + m->outer_l3_len)) +
  71                          m->l2_len + m->l3_len + m->l4_len;
  72
  73                 /* Reduce payload len from base headers */
  74                 paylen = m->pkt_len - lso_sb;
  75
  76                 /* Get iplen position assuming no tunnel hdr */
  77                 iplen = (uint16_t *)(mdata + m->l2_len +
  78                                      (2 << !!(ol_flags & PKT_TX_IPV6)));
  79                 /* Handle tunnel tso */
  80                 if ((flags & NIX_TX_OFFLOAD_OL3_OL4_CSUM_F) &&
  81                     (ol_flags & PKT_TX_TUNNEL_MASK)) {
  82                         const uint8_t is_udp_tun =
  83                                 (CNXK_NIX_UDP_TUN_BITMASK >>
  84                                  ((ol_flags & PKT_TX_TUNNEL_MASK) >> 45)) &
  85                                 0x1;
  86
  87                         oiplen = (uint16_t *)(mdata + m->outer_l2_len +
  88                                               (2 << !!(ol_flags &
  89                                                        PKT_TX_OUTER_IPV6)));
  90                         *oiplen = rte_cpu_to_be_16(rte_be_to_cpu_16(*oiplen) -
  91                                                    paylen);
  92
  93                         /* Update format for UDP tunneled packet */
  94                         if (is_udp_tun) {
  95                                 oudplen = (uint16_t *)(mdata + m->outer_l2_len +
  96                                                        m->outer_l3_len + 4);
  97                                 *oudplen = rte_cpu_to_be_16(
  98                                         rte_be_to_cpu_16(*oudplen) - paylen);
  99                         }
 100
 101                         /* Update iplen position to inner ip hdr */
 102                         iplen = (uint16_t *)(mdata + lso_sb - m->l3_len -
 103                                              m->l4_len +
 104                                              (2 << !!(ol_flags & PKT_TX_IPV6)));
 105                 }
 106
 107                 *iplen = rte_cpu_to_be_16(rte_be_to_cpu_16(*iplen) - paylen);
 108         }
 109 }
 110
 111 static __rte_always_inline void
 112 cn9k_nix_xmit_prepare(struct rte_mbuf *m, uint64_t *cmd, const uint16_t flags,
 113                       const uint64_t lso_tun_fmt)
 114 {
 115         struct nix_send_ext_s *send_hdr_ext;
 116         struct nix_send_hdr_s *send_hdr;
 117         uint64_t ol_flags = 0, mask;
 118         union nix_send_hdr_w1_u w1;
 119         union nix_send_sg_s *sg;
 120
 121         send_hdr = (struct nix_send_hdr_s *)cmd;
 122         if (flags & NIX_TX_NEED_EXT_HDR) {
 123                 send_hdr_ext = (struct nix_send_ext_s *)(cmd + 2);
 124                 sg = (union nix_send_sg_s *)(cmd + 4);
 125                 /* Clear previous markings */
 126                 send_hdr_ext->w0.lso = 0;
 127                 send_hdr_ext->w1.u = 0;
 128         } else {
 129                 sg = (union nix_send_sg_s *)(cmd + 2);
 130         }
 131
 132         if (flags & NIX_TX_NEED_SEND_HDR_W1) {
 133                 ol_flags = m->ol_flags;
 134                 w1.u = 0;
 135         }
 136
 137         if (!(flags & NIX_TX_MULTI_SEG_F)) {
 138                 send_hdr->w0.total = m->data_len;
 139                 send_hdr->w0.aura =
 140                         roc_npa_aura_handle_to_aura(m->pool->pool_id);
 141         }
 142
 143         /*
 144          * L3type:  2 => IPV4
 145          *          3 => IPV4 with csum
 146          *          4 => IPV6
 147          * L3type and L3ptr needs to be set for either
 148          * L3 csum or L4 csum or LSO
 149          *
 150          */
 151
 152         if ((flags & NIX_TX_OFFLOAD_OL3_OL4_CSUM_F) &&
 153             (flags & NIX_TX_OFFLOAD_L3_L4_CSUM_F)) {
 154                 const uint8_t csum = !!(ol_flags & PKT_TX_OUTER_UDP_CKSUM);
 155                 const uint8_t ol3type =
 156                         ((!!(ol_flags & PKT_TX_OUTER_IPV4)) << 1) +
 157                         ((!!(ol_flags & PKT_TX_OUTER_IPV6)) << 2) +
 158                         !!(ol_flags & PKT_TX_OUTER_IP_CKSUM);
 159
 160                 /* Outer L3 */
 161                 w1.ol3type = ol3type;
 162                 mask = 0xffffull << ((!!ol3type) << 4);
 163                 w1.ol3ptr = ~mask & m->outer_l2_len;
 164                 w1.ol4ptr = ~mask & (w1.ol3ptr + m->outer_l3_len);
 165
 166                 /* Outer L4 */
 167                 w1.ol4type = csum + (csum << 1);
 168
 169                 /* Inner L3 */
 170                 w1.il3type = ((!!(ol_flags & PKT_TX_IPV4)) << 1) +
 171                              ((!!(ol_flags & PKT_TX_IPV6)) << 2);
 172                 w1.il3ptr = w1.ol4ptr + m->l2_len;
 173                 w1.il4ptr = w1.il3ptr + m->l3_len;
 174                 /* Increment it by 1 if it is IPV4 as 3 is with csum */
 175                 w1.il3type = w1.il3type + !!(ol_flags & PKT_TX_IP_CKSUM);
 176
 177                 /* Inner L4 */
 178                 w1.il4type = (ol_flags & PKT_TX_L4_MASK) >> 52;
 179
 180                 /* In case of no tunnel header use only
 181                  * shift IL3/IL4 fields a bit to use
 182                  * OL3/OL4 for header checksum
 183                  */
 184                 mask = !ol3type;
 185                 w1.u = ((w1.u & 0xFFFFFFFF00000000) >> (mask << 3)) |
 186                        ((w1.u & 0X00000000FFFFFFFF) >> (mask << 4));
 187
 188         } else if (flags & NIX_TX_OFFLOAD_OL3_OL4_CSUM_F) {
 189                 const uint8_t csum = !!(ol_flags & PKT_TX_OUTER_UDP_CKSUM);
 190                 const uint8_t outer_l2_len = m->outer_l2_len;
 191
 192                 /* Outer L3 */
 193                 w1.ol3ptr = outer_l2_len;
 194                 w1.ol4ptr = outer_l2_len + m->outer_l3_len;
 195                 /* Increment it by 1 if it is IPV4 as 3 is with csum */
 196                 w1.ol3type = ((!!(ol_flags & PKT_TX_OUTER_IPV4)) << 1) +
 197                              ((!!(ol_flags & PKT_TX_OUTER_IPV6)) << 2) +
 198                              !!(ol_flags & PKT_TX_OUTER_IP_CKSUM);
 199
 200                 /* Outer L4 */
 201                 w1.ol4type = csum + (csum << 1);
 202
 203         } else if (flags & NIX_TX_OFFLOAD_L3_L4_CSUM_F) {
 204                 const uint8_t l2_len = m->l2_len;
 205
 206                 /* Always use OLXPTR and OLXTYPE when only
 207                  * when one header is present
 208                  */
 209
 210                 /* Inner L3 */
 211                 w1.ol3ptr = l2_len;
 212                 w1.ol4ptr = l2_len + m->l3_len;
 213                 /* Increment it by 1 if it is IPV4 as 3 is with csum */
 214                 w1.ol3type = ((!!(ol_flags & PKT_TX_IPV4)) << 1) +
 215                              ((!!(ol_flags & PKT_TX_IPV6)) << 2) +
 216                              !!(ol_flags & PKT_TX_IP_CKSUM);
 217
 218                 /* Inner L4 */
 219                 w1.ol4type = (ol_flags & PKT_TX_L4_MASK) >> 52;
 220         }
 221
 222         if (flags & NIX_TX_NEED_EXT_HDR && flags & NIX_TX_OFFLOAD_VLAN_QINQ_F) {
 223                 send_hdr_ext->w1.vlan1_ins_ena = !!(ol_flags & PKT_TX_VLAN);
 224                 /* HW will update ptr after vlan0 update */
 225                 send_hdr_ext->w1.vlan1_ins_ptr = 12;
 226                 send_hdr_ext->w1.vlan1_ins_tci = m->vlan_tci;
 227
 228                 send_hdr_ext->w1.vlan0_ins_ena = !!(ol_flags & PKT_TX_QINQ);
 229                 /* 2B before end of l2 header */
 230                 send_hdr_ext->w1.vlan0_ins_ptr = 12;
 231                 send_hdr_ext->w1.vlan0_ins_tci = m->vlan_tci_outer;
 232         }
 233
 234         if (flags & NIX_TX_OFFLOAD_TSO_F && (ol_flags & PKT_TX_TCP_SEG)) {
 235                 uint16_t lso_sb;
 236                 uint64_t mask;
 237
 238                 mask = -(!w1.il3type);
 239                 lso_sb = (mask & w1.ol4ptr) + (~mask & w1.il4ptr) + m->l4_len;
 240
 241                 send_hdr_ext->w0.lso_sb = lso_sb;
 242                 send_hdr_ext->w0.lso = 1;
 243                 send_hdr_ext->w0.lso_mps = m->tso_segsz;
 244                 send_hdr_ext->w0.lso_format =
 245                         NIX_LSO_FORMAT_IDX_TSOV4 + !!(ol_flags & PKT_TX_IPV6);
 246                 w1.ol4type = NIX_SENDL4TYPE_TCP_CKSUM;
 247
 248                 /* Handle tunnel tso */
 249                 if ((flags & NIX_TX_OFFLOAD_OL3_OL4_CSUM_F) &&
 250                     (ol_flags & PKT_TX_TUNNEL_MASK)) {
 251                         const uint8_t is_udp_tun =
 252                                 (CNXK_NIX_UDP_TUN_BITMASK >>
 253                                  ((ol_flags & PKT_TX_TUNNEL_MASK) >> 45)) &
 254                                 0x1;
 255                         uint8_t shift = is_udp_tun ? 32 : 0;
 256
 257                         shift += (!!(ol_flags & PKT_TX_OUTER_IPV6) << 4);
 258                         shift += (!!(ol_flags & PKT_TX_IPV6) << 3);
 259
 260                         w1.il4type = NIX_SENDL4TYPE_TCP_CKSUM;
 261                         w1.ol4type = is_udp_tun ? NIX_SENDL4TYPE_UDP_CKSUM : 0;
 262                         /* Update format for UDP tunneled packet */
 263                         send_hdr_ext->w0.lso_format = (lso_tun_fmt >> shift);
 264                 }
 265         }
 266
 267         if (flags & NIX_TX_NEED_SEND_HDR_W1)
 268                 send_hdr->w1.u = w1.u;
 269
 270         if (!(flags & NIX_TX_MULTI_SEG_F)) {
 271                 sg->seg1_size = m->data_len;
 272                 *(rte_iova_t *)(++sg) = rte_mbuf_data_iova(m);
 273
 274                 if (flags & NIX_TX_OFFLOAD_MBUF_NOFF_F) {
 275                         /* DF bit = 1 if refcount of current mbuf or parent mbuf
 276                          *              is greater than 1
 277                          * DF bit = 0 otherwise
 278                          */
 279                         send_hdr->w0.df = cnxk_nix_prefree_seg(m);
 280                         /* Ensuring mbuf fields which got updated in
 281                          * cnxk_nix_prefree_seg are written before LMTST.
 282                          */
 283                         rte_io_wmb();
 284                 }
 285                 /* Mark mempool object as "put" since it is freed by NIX */
 286                 if (!send_hdr->w0.df)
 287                         __mempool_check_cookies(m->pool, (void **)&m, 1, 0);
 288         }
 289 }
 290
 291 static __rte_always_inline void
 292 cn9k_nix_xmit_prepare_tstamp(uint64_t *cmd, const uint64_t *send_mem_desc,
 293                              const uint64_t ol_flags, const uint16_t no_segdw,
 294                              const uint16_t flags)
 295 {
 296         if (flags & NIX_TX_OFFLOAD_TSTAMP_F) {
 297                 struct nix_send_mem_s *send_mem;
 298                 uint16_t off = (no_segdw - 1) << 1;
 299                 const uint8_t is_ol_tstamp = !(ol_flags & PKT_TX_IEEE1588_TMST);
 300
 301                 send_mem = (struct nix_send_mem_s *)(cmd + off);
 302                 if (flags & NIX_TX_MULTI_SEG_F) {
 303                         /* Retrieving the default desc values */
 304                         cmd[off] = send_mem_desc[6];
 305
 306                         /* Using compiler barier to avoid voilation of C
 307                          * aliasing rules.
 308                          */
 309                         rte_compiler_barrier();
 310                 }
 311
 312                 /* Packets for which PKT_TX_IEEE1588_TMST is not set, tx tstamp
 313                  * should not be recorded, hence changing the alg type to
 314                  * NIX_SENDMEMALG_SET and also changing send mem addr field to
 315                  * next 8 bytes as it corrpt the actual tx tstamp registered
 316                  * address.
 317                  */
 318                 send_mem->w0.cn9k.alg =
 319                         NIX_SENDMEMALG_SETTSTMP - (is_ol_tstamp);
 320
 321                 send_mem->addr = (rte_iova_t)((uint64_t *)send_mem_desc[7] +
 322                                               (is_ol_tstamp));
 323         }
 324 }
 325
 326 static __rte_always_inline void
 327 cn9k_nix_xmit_one(uint64_t *cmd, void *lmt_addr, const rte_iova_t io_addr,
 328                   const uint32_t flags)
 329 {
 330         uint64_t lmt_status;
 331
 332         do {
 333                 roc_lmt_mov(lmt_addr, cmd, cn9k_nix_tx_ext_subs(flags));
 334                 lmt_status = roc_lmt_submit_ldeor(io_addr);
 335         } while (lmt_status == 0);
 336 }
 337
 338 static __rte_always_inline void
 339 cn9k_nix_xmit_prep_lmt(uint64_t *cmd, void *lmt_addr, const uint32_t flags)
 340 {
 341         roc_lmt_mov(lmt_addr, cmd, cn9k_nix_tx_ext_subs(flags));
 342 }
 343
 344 static __rte_always_inline uint64_t
 345 cn9k_nix_xmit_submit_lmt(const rte_iova_t io_addr)
 346 {
 347         return roc_lmt_submit_ldeor(io_addr);
 348 }
 349
 350 static __rte_always_inline uint64_t
 351 cn9k_nix_xmit_submit_lmt_release(const rte_iova_t io_addr)
 352 {
 353         return roc_lmt_submit_ldeorl(io_addr);
 354 }
 355
 356 static __rte_always_inline uint16_t
 357 cn9k_nix_prepare_mseg(struct rte_mbuf *m, uint64_t *cmd, const uint16_t flags)
 358 {
 359         struct nix_send_hdr_s *send_hdr;
 360         union nix_send_sg_s *sg;
 361         struct rte_mbuf *m_next;
 362         uint64_t *slist, sg_u;
 363         uint64_t nb_segs;
 364         uint64_t segdw;
 365         uint8_t off, i;
 366
 367         send_hdr = (struct nix_send_hdr_s *)cmd;
 368         send_hdr->w0.total = m->pkt_len;
 369         send_hdr->w0.aura = roc_npa_aura_handle_to_aura(m->pool->pool_id);
 370
 371         if (flags & NIX_TX_NEED_EXT_HDR)
 372                 off = 2;
 373         else
 374                 off = 0;
 375
 376         sg = (union nix_send_sg_s *)&cmd[2 + off];
 377         /* Clear sg->u header before use */
 378         sg->u &= 0xFC00000000000000;
 379         sg_u = sg->u;
 380         slist = &cmd[3 + off];
 381
 382         i = 0;
 383         nb_segs = m->nb_segs;
 384
 385         /* Fill mbuf segments */
 386         do {
 387                 m_next = m->next;
 388                 sg_u = sg_u | ((uint64_t)m->data_len << (i << 4));
 389                 *slist = rte_mbuf_data_iova(m);
 390                 /* Set invert df if buffer is not to be freed by H/W */
 391                 if (flags & NIX_TX_OFFLOAD_MBUF_NOFF_F) {
 392                         sg_u |= (cnxk_nix_prefree_seg(m) << (i + 55));
 393                         /* Commit changes to mbuf */
 394                         rte_io_wmb();
 395                 }
 396                 /* Mark mempool object as "put" since it is freed by NIX */
 397 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
 398                 if (!(sg_u & (1ULL << (i + 55))))
 399                         __mempool_check_cookies(m->pool, (void **)&m, 1, 0);
 400                 rte_io_wmb();
 401 #endif
 402                 slist++;
 403                 i++;
 404                 nb_segs--;
 405                 if (i > 2 && nb_segs) {
 406                         i = 0;
 407                         /* Next SG subdesc */
 408                         *(uint64_t *)slist = sg_u & 0xFC00000000000000;
 409                         sg->u = sg_u;
 410                         sg->segs = 3;
 411                         sg = (union nix_send_sg_s *)slist;
 412                         sg_u = sg->u;
 413                         slist++;
 414                 }
 415                 m = m_next;
 416         } while (nb_segs);
 417
 418         sg->u = sg_u;
 419         sg->segs = i;
 420         segdw = (uint64_t *)slist - (uint64_t *)&cmd[2 + off];
 421         /* Roundup extra dwords to multiple of 2 */
 422         segdw = (segdw >> 1) + (segdw & 0x1);
 423         /* Default dwords */
 424         segdw += (off >> 1) + 1 + !!(flags & NIX_TX_OFFLOAD_TSTAMP_F);
 425         send_hdr->w0.sizem1 = segdw - 1;
 426
 427         return segdw;
 428 }
 429
 430 static __rte_always_inline void
 431 cn9k_nix_xmit_mseg_prep_lmt(uint64_t *cmd, void *lmt_addr, uint16_t segdw)
 432 {
 433         roc_lmt_mov_seg(lmt_addr, (const void *)cmd, segdw);
 434 }
 435
 436 static __rte_always_inline void
 437 cn9k_nix_xmit_mseg_one(uint64_t *cmd, void *lmt_addr, rte_iova_t io_addr,
 438                        uint16_t segdw)
 439 {
 440         uint64_t lmt_status;
 441
 442         do {
 443                 roc_lmt_mov_seg(lmt_addr, (const void *)cmd, segdw);
 444                 lmt_status = roc_lmt_submit_ldeor(io_addr);
 445         } while (lmt_status == 0);
 446 }
 447
 448 static __rte_always_inline void
 449 cn9k_nix_xmit_mseg_one_release(uint64_t *cmd, void *lmt_addr,
 450                                rte_iova_t io_addr, uint16_t segdw)
 451 {
 452         uint64_t lmt_status;
 453
 454         rte_io_wmb();
 455         do {
 456                 roc_lmt_mov_seg(lmt_addr, (const void *)cmd, segdw);
 457                 lmt_status = roc_lmt_submit_ldeor(io_addr);
 458         } while (lmt_status == 0);
 459 }
 460
 461 static __rte_always_inline uint16_t
 462 cn9k_nix_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t pkts,
 463                    uint64_t *cmd, const uint16_t flags)
 464 {
 465         struct cn9k_eth_txq *txq = tx_queue;
 466         const rte_iova_t io_addr = txq->io_addr;
 467         void *lmt_addr = txq->lmt_addr;
 468         uint64_t lso_tun_fmt;
 469         uint16_t i;
 470
 471         NIX_XMIT_FC_OR_RETURN(txq, pkts);
 472
 473         roc_lmt_mov(cmd, &txq->cmd[0], cn9k_nix_tx_ext_subs(flags));
 474
 475         /* Perform header writes before barrier for TSO */
 476         if (flags & NIX_TX_OFFLOAD_TSO_F) {
 477                 lso_tun_fmt = txq->lso_tun_fmt;
 478
 479                 for (i = 0; i < pkts; i++)
 480                         cn9k_nix_xmit_prepare_tso(tx_pkts[i], flags);
 481         }
 482
 483         /* Lets commit any changes in the packet here as no further changes
 484          * to the packet will be done unless no fast free is enabled.
 485          */
 486         if (!(flags & NIX_TX_OFFLOAD_MBUF_NOFF_F))
 487                 rte_io_wmb();
 488
 489         for (i = 0; i < pkts; i++) {
 490                 cn9k_nix_xmit_prepare(tx_pkts[i], cmd, flags, lso_tun_fmt);
 491                 cn9k_nix_xmit_prepare_tstamp(cmd, &txq->cmd[0],
 492                                              tx_pkts[i]->ol_flags, 4, flags);
 493                 cn9k_nix_xmit_one(cmd, lmt_addr, io_addr, flags);
 494         }
 495
 496         /* Reduce the cached count */
 497         txq->fc_cache_pkts -= pkts;
 498
 499         return pkts;
 500 }
 501
 502 static __rte_always_inline uint16_t
 503 cn9k_nix_xmit_pkts_mseg(void *tx_queue, struct rte_mbuf **tx_pkts,
 504                         uint16_t pkts, uint64_t *cmd, const uint16_t flags)
 505 {
 506         struct cn9k_eth_txq *txq = tx_queue;
 507         const rte_iova_t io_addr = txq->io_addr;
 508         void *lmt_addr = txq->lmt_addr;
 509         uint64_t lso_tun_fmt;
 510         uint16_t segdw;
 511         uint64_t i;
 512
 513         NIX_XMIT_FC_OR_RETURN(txq, pkts);
 514
 515         roc_lmt_mov(cmd, &txq->cmd[0], cn9k_nix_tx_ext_subs(flags));
 516
 517         /* Perform header writes before barrier for TSO */
 518         if (flags & NIX_TX_OFFLOAD_TSO_F) {
 519                 lso_tun_fmt = txq->lso_tun_fmt;
 520
 521                 for (i = 0; i < pkts; i++)
 522                         cn9k_nix_xmit_prepare_tso(tx_pkts[i], flags);
 523         }
 524
 525         /* Lets commit any changes in the packet here as no further changes
 526          * to the packet will be done unless no fast free is enabled.
 527          */
 528         if (!(flags & NIX_TX_OFFLOAD_MBUF_NOFF_F))
 529                 rte_io_wmb();
 530
 531         for (i = 0; i < pkts; i++) {
 532                 cn9k_nix_xmit_prepare(tx_pkts[i], cmd, flags, lso_tun_fmt);
 533                 segdw = cn9k_nix_prepare_mseg(tx_pkts[i], cmd, flags);
 534                 cn9k_nix_xmit_prepare_tstamp(cmd, &txq->cmd[0],
 535                                              tx_pkts[i]->ol_flags, segdw,
 536                                              flags);
 537                 cn9k_nix_xmit_mseg_one(cmd, lmt_addr, io_addr, segdw);
 538         }
 539
 540         /* Reduce the cached count */
 541         txq->fc_cache_pkts -= pkts;
 542
 543         return pkts;
 544 }
 545
 546 #if defined(RTE_ARCH_ARM64)
 547
 548 static __rte_always_inline void
 549 cn9k_nix_prepare_tso(struct rte_mbuf *m, union nix_send_hdr_w1_u *w1,
 550                      union nix_send_ext_w0_u *w0, uint64_t ol_flags,
 551                      uint64_t flags)
 552 {
 553         uint16_t lso_sb;
 554         uint64_t mask;
 555
 556         if (!(ol_flags & PKT_TX_TCP_SEG))
 557                 return;
 558
 559         mask = -(!w1->il3type);
 560         lso_sb = (mask & w1->ol4ptr) + (~mask & w1->il4ptr) + m->l4_len;
 561
 562         w0->u |= BIT(14);
 563         w0->lso_sb = lso_sb;
 564         w0->lso_mps = m->tso_segsz;
 565         w0->lso_format = NIX_LSO_FORMAT_IDX_TSOV4 + !!(ol_flags & PKT_TX_IPV6);
 566         w1->ol4type = NIX_SENDL4TYPE_TCP_CKSUM;
 567
 568         /* Handle tunnel tso */
 569         if ((flags & NIX_TX_OFFLOAD_OL3_OL4_CSUM_F) &&
 570             (ol_flags & PKT_TX_TUNNEL_MASK)) {
 571                 const uint8_t is_udp_tun =
 572                         (CNXK_NIX_UDP_TUN_BITMASK >>
 573                          ((ol_flags & PKT_TX_TUNNEL_MASK) >> 45)) &
 574                         0x1;
 575
 576                 w1->il4type = NIX_SENDL4TYPE_TCP_CKSUM;
 577                 w1->ol4type = is_udp_tun ? NIX_SENDL4TYPE_UDP_CKSUM : 0;
 578                 /* Update format for UDP tunneled packet */
 579                 w0->lso_format += is_udp_tun ? 2 : 6;
 580
 581                 w0->lso_format += !!(ol_flags & PKT_TX_OUTER_IPV6) << 1;
 582         }
 583 }
 584
 585 #define NIX_DESCS_PER_LOOP 4
 586 static __rte_always_inline uint16_t
 587 cn9k_nix_xmit_pkts_vector(void *tx_queue, struct rte_mbuf **tx_pkts,
 588                           uint16_t pkts, uint64_t *cmd, const uint16_t flags)
 589 {
 590         uint64x2_t dataoff_iova0, dataoff_iova1, dataoff_iova2, dataoff_iova3;
 591         uint64x2_t len_olflags0, len_olflags1, len_olflags2, len_olflags3;
 592         uint64x2_t cmd0[NIX_DESCS_PER_LOOP], cmd1[NIX_DESCS_PER_LOOP],
 593                 cmd2[NIX_DESCS_PER_LOOP], cmd3[NIX_DESCS_PER_LOOP];
 594         uint64_t *mbuf0, *mbuf1, *mbuf2, *mbuf3;
 595         uint64x2_t senddesc01_w0, senddesc23_w0;
 596         uint64x2_t senddesc01_w1, senddesc23_w1;
 597         uint64x2_t sendext01_w0, sendext23_w0;
 598         uint64x2_t sendext01_w1, sendext23_w1;
 599         uint64x2_t sendmem01_w0, sendmem23_w0;
 600         uint64x2_t sendmem01_w1, sendmem23_w1;
 601         uint64x2_t sgdesc01_w0, sgdesc23_w0;
 602         uint64x2_t sgdesc01_w1, sgdesc23_w1;
 603         struct cn9k_eth_txq *txq = tx_queue;
 604         uint64_t *lmt_addr = txq->lmt_addr;
 605         rte_iova_t io_addr = txq->io_addr;
 606         uint64x2_t ltypes01, ltypes23;
 607         uint64x2_t xtmp128, ytmp128;
 608         uint64x2_t xmask01, xmask23;
 609         uint64_t lmt_status, i;
 610         uint16_t pkts_left;
 611
 612         NIX_XMIT_FC_OR_RETURN(txq, pkts);
 613
 614         pkts_left = pkts & (NIX_DESCS_PER_LOOP - 1);
 615         pkts = RTE_ALIGN_FLOOR(pkts, NIX_DESCS_PER_LOOP);
 616
 617         /* Reduce the cached count */
 618         txq->fc_cache_pkts -= pkts;
 619
 620         /* Perform header writes before barrier for TSO */
 621         if (flags & NIX_TX_OFFLOAD_TSO_F) {
 622                 for (i = 0; i < pkts; i++)
 623                         cn9k_nix_xmit_prepare_tso(tx_pkts[i], flags);
 624         }
 625
 626         /* Lets commit any changes in the packet here as no further changes
 627          * to the packet will be done unless no fast free is enabled.
 628          */
 629         if (!(flags & NIX_TX_OFFLOAD_MBUF_NOFF_F))
 630                 rte_io_wmb();
 631
 632         senddesc01_w0 = vld1q_dup_u64(&txq->cmd[0]);
 633         senddesc23_w0 = senddesc01_w0;
 634         senddesc01_w1 = vdupq_n_u64(0);
 635         senddesc23_w1 = senddesc01_w1;
 636
 637         /* Load command defaults into vector variables. */
 638         if (flags & NIX_TX_NEED_EXT_HDR) {
 639                 sendext01_w0 = vld1q_dup_u64(&txq->cmd[2]);
 640                 sendext23_w0 = sendext01_w0;
 641                 sendext01_w1 = vdupq_n_u64(12 | 12U << 24);
 642                 sendext23_w1 = sendext01_w1;
 643                 sgdesc01_w0 = vld1q_dup_u64(&txq->cmd[4]);
 644                 sgdesc23_w0 = sgdesc01_w0;
 645                 if (flags & NIX_TX_OFFLOAD_TSTAMP_F) {
 646                         sendmem01_w0 = vld1q_dup_u64(&txq->cmd[6]);
 647                         sendmem23_w0 = sendmem01_w0;
 648                         sendmem01_w1 = vld1q_dup_u64(&txq->cmd[7]);
 649                         sendmem23_w1 = sendmem01_w1;
 650                 }
 651         } else {
 652                 sgdesc01_w0 = vld1q_dup_u64(&txq->cmd[2]);
 653                 sgdesc23_w0 = sgdesc01_w0;
 654         }
 655
 656         for (i = 0; i < pkts; i += NIX_DESCS_PER_LOOP) {
 657                 /* Clear lower 32bit of SEND_HDR_W0 and SEND_SG_W0 */
 658                 senddesc01_w0 =
 659                         vbicq_u64(senddesc01_w0, vdupq_n_u64(0xFFFFFFFF));
 660                 sgdesc01_w0 = vbicq_u64(sgdesc01_w0, vdupq_n_u64(0xFFFFFFFF));
 661
 662                 senddesc23_w0 = senddesc01_w0;
 663                 sgdesc23_w0 = sgdesc01_w0;
 664
 665                 /* Clear vlan enables. */
 666                 if (flags & NIX_TX_NEED_EXT_HDR) {
 667                         sendext01_w1 = vbicq_u64(sendext01_w1,
 668                                                  vdupq_n_u64(0x3FFFF00FFFF00));
 669                         sendext23_w1 = sendext01_w1;
 670                 }
 671
 672                 if (flags & NIX_TX_OFFLOAD_TSTAMP_F) {
 673                         /* Reset send mem alg to SETTSTMP from SUB*/
 674                         sendmem01_w0 = vbicq_u64(sendmem01_w0,
 675                                                  vdupq_n_u64(BIT_ULL(59)));
 676                         /* Reset send mem address to default. */
 677                         sendmem01_w1 =
 678                                 vbicq_u64(sendmem01_w1, vdupq_n_u64(0xF));
 679                         sendmem23_w0 = sendmem01_w0;
 680                         sendmem23_w1 = sendmem01_w1;
 681                 }
 682
 683                 if (flags & NIX_TX_OFFLOAD_TSO_F) {
 684                         /* Clear the LSO enable bit. */
 685                         sendext01_w0 = vbicq_u64(sendext01_w0,
 686                                                  vdupq_n_u64(BIT_ULL(14)));
 687                         sendext23_w0 = sendext01_w0;
 688                 }
 689
 690                 /* Move mbufs to iova */
 691                 mbuf0 = (uint64_t *)tx_pkts[0];
 692                 mbuf1 = (uint64_t *)tx_pkts[1];
 693                 mbuf2 = (uint64_t *)tx_pkts[2];
 694                 mbuf3 = (uint64_t *)tx_pkts[3];
 695
 696                 mbuf0 = (uint64_t *)((uintptr_t)mbuf0 +
 697                                      offsetof(struct rte_mbuf, buf_iova));
 698                 mbuf1 = (uint64_t *)((uintptr_t)mbuf1 +
 699                                      offsetof(struct rte_mbuf, buf_iova));
 700                 mbuf2 = (uint64_t *)((uintptr_t)mbuf2 +
 701                                      offsetof(struct rte_mbuf, buf_iova));
 702                 mbuf3 = (uint64_t *)((uintptr_t)mbuf3 +
 703                                      offsetof(struct rte_mbuf, buf_iova));
 704                 /*
 705                  * Get mbuf's, olflags, iova, pktlen, dataoff
 706                  * dataoff_iovaX.D[0] = iova,
 707                  * dataoff_iovaX.D[1](15:0) = mbuf->dataoff
 708                  * len_olflagsX.D[0] = ol_flags,
 709                  * len_olflagsX.D[1](63:32) = mbuf->pkt_len
 710                  */
 711                 dataoff_iova0 = vld1q_u64(mbuf0);
 712                 len_olflags0 = vld1q_u64(mbuf0 + 2);
 713                 dataoff_iova1 = vld1q_u64(mbuf1);
 714                 len_olflags1 = vld1q_u64(mbuf1 + 2);
 715                 dataoff_iova2 = vld1q_u64(mbuf2);
 716                 len_olflags2 = vld1q_u64(mbuf2 + 2);
 717                 dataoff_iova3 = vld1q_u64(mbuf3);
 718                 len_olflags3 = vld1q_u64(mbuf3 + 2);
 719
 720                 /* Move mbufs to point pool */
 721                 mbuf0 = (uint64_t *)((uintptr_t)mbuf0 +
 722                                      offsetof(struct rte_mbuf, pool) -
 723                                      offsetof(struct rte_mbuf, buf_iova));
 724                 mbuf1 = (uint64_t *)((uintptr_t)mbuf1 +
 725                                      offsetof(struct rte_mbuf, pool) -
 726                                      offsetof(struct rte_mbuf, buf_iova));
 727                 mbuf2 = (uint64_t *)((uintptr_t)mbuf2 +
 728                                      offsetof(struct rte_mbuf, pool) -
 729                                      offsetof(struct rte_mbuf, buf_iova));
 730                 mbuf3 = (uint64_t *)((uintptr_t)mbuf3 +
 731                                      offsetof(struct rte_mbuf, pool) -
 732                                      offsetof(struct rte_mbuf, buf_iova));
 733
 734                 if (flags & (NIX_TX_OFFLOAD_OL3_OL4_CSUM_F |
 735                              NIX_TX_OFFLOAD_L3_L4_CSUM_F)) {
 736                         /* Get tx_offload for ol2, ol3, l2, l3 lengths */
 737                         /*
 738                          * E(8):OL2_LEN(7):OL3_LEN(9):E(24):L3_LEN(9):L2_LEN(7)
 739                          * E(8):OL2_LEN(7):OL3_LEN(9):E(24):L3_LEN(9):L2_LEN(7)
 740                          */
 741
 742                         asm volatile("LD1 {%[a].D}[0],[%[in]]\n\t"
 743                                      : [a] "+w"(senddesc01_w1)
 744                                      : [in] "r"(mbuf0 + 2)
 745                                      : "memory");
 746
 747                         asm volatile("LD1 {%[a].D}[1],[%[in]]\n\t"
 748                                      : [a] "+w"(senddesc01_w1)
 749                                      : [in] "r"(mbuf1 + 2)
 750                                      : "memory");
 751
 752                         asm volatile("LD1 {%[b].D}[0],[%[in]]\n\t"
 753                                      : [b] "+w"(senddesc23_w1)
 754                                      : [in] "r"(mbuf2 + 2)
 755                                      : "memory");
 756
 757                         asm volatile("LD1 {%[b].D}[1],[%[in]]\n\t"
 758                                      : [b] "+w"(senddesc23_w1)
 759                                      : [in] "r"(mbuf3 + 2)
 760                                      : "memory");
 761
 762                         /* Get pool pointer alone */
 763                         mbuf0 = (uint64_t *)*mbuf0;
 764                         mbuf1 = (uint64_t *)*mbuf1;
 765                         mbuf2 = (uint64_t *)*mbuf2;
 766                         mbuf3 = (uint64_t *)*mbuf3;
 767                 } else {
 768                         /* Get pool pointer alone */
 769                         mbuf0 = (uint64_t *)*mbuf0;
 770                         mbuf1 = (uint64_t *)*mbuf1;
 771                         mbuf2 = (uint64_t *)*mbuf2;
 772                         mbuf3 = (uint64_t *)*mbuf3;
 773                 }
 774
 775                 const uint8x16_t shuf_mask2 = {
 776                         0x4, 0x5, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
 777                         0xc, 0xd, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
 778                 };
 779                 xtmp128 = vzip2q_u64(len_olflags0, len_olflags1);
 780                 ytmp128 = vzip2q_u64(len_olflags2, len_olflags3);
 781
 782                 /* Clear dataoff_iovaX.D[1] bits other than dataoff(15:0) */
 783                 const uint64x2_t and_mask0 = {
 784                         0xFFFFFFFFFFFFFFFF,
 785                         0x000000000000FFFF,
 786                 };
 787
 788                 dataoff_iova0 = vandq_u64(dataoff_iova0, and_mask0);
 789                 dataoff_iova1 = vandq_u64(dataoff_iova1, and_mask0);
 790                 dataoff_iova2 = vandq_u64(dataoff_iova2, and_mask0);
 791                 dataoff_iova3 = vandq_u64(dataoff_iova3, and_mask0);
 792
 793                 /*
 794                  * Pick only 16 bits of pktlen preset at bits 63:32
 795                  * and place them at bits 15:0.
 796                  */
 797                 xtmp128 = vqtbl1q_u8(xtmp128, shuf_mask2);
 798                 ytmp128 = vqtbl1q_u8(ytmp128, shuf_mask2);
 799
 800                 /* Add pairwise to get dataoff + iova in sgdesc_w1 */
 801                 sgdesc01_w1 = vpaddq_u64(dataoff_iova0, dataoff_iova1);
 802                 sgdesc23_w1 = vpaddq_u64(dataoff_iova2, dataoff_iova3);
 803
 804                 /* Orr both sgdesc_w0 and senddesc_w0 with 16 bits of
 805                  * pktlen at 15:0 position.
 806                  */
 807                 sgdesc01_w0 = vorrq_u64(sgdesc01_w0, xtmp128);
 808                 sgdesc23_w0 = vorrq_u64(sgdesc23_w0, ytmp128);
 809                 senddesc01_w0 = vorrq_u64(senddesc01_w0, xtmp128);
 810                 senddesc23_w0 = vorrq_u64(senddesc23_w0, ytmp128);
 811
 812                 /* Move mbuf to point to pool_id. */
 813                 mbuf0 = (uint64_t *)((uintptr_t)mbuf0 +
 814                                      offsetof(struct rte_mempool, pool_id));
 815                 mbuf1 = (uint64_t *)((uintptr_t)mbuf1 +
 816                                      offsetof(struct rte_mempool, pool_id));
 817                 mbuf2 = (uint64_t *)((uintptr_t)mbuf2 +
 818                                      offsetof(struct rte_mempool, pool_id));
 819                 mbuf3 = (uint64_t *)((uintptr_t)mbuf3 +
 820                                      offsetof(struct rte_mempool, pool_id));
 821
 822                 if ((flags & NIX_TX_OFFLOAD_L3_L4_CSUM_F) &&
 823                     !(flags & NIX_TX_OFFLOAD_OL3_OL4_CSUM_F)) {
 824                         /*
 825                          * Lookup table to translate ol_flags to
 826                          * il3/il4 types. But we still use ol3/ol4 types in
 827                          * senddesc_w1 as only one header processing is enabled.
 828                          */
 829                         const uint8x16_t tbl = {
 830                                 /* [0-15] = il4type:il3type */
 831                                 0x04, /* none (IPv6 assumed) */
 832                                 0x14, /* PKT_TX_TCP_CKSUM (IPv6 assumed) */
 833                                 0x24, /* PKT_TX_SCTP_CKSUM (IPv6 assumed) */
 834                                 0x34, /* PKT_TX_UDP_CKSUM (IPv6 assumed) */
 835                                 0x03, /* PKT_TX_IP_CKSUM */
 836                                 0x13, /* PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM */
 837                                 0x23, /* PKT_TX_IP_CKSUM | PKT_TX_SCTP_CKSUM */
 838                                 0x33, /* PKT_TX_IP_CKSUM | PKT_TX_UDP_CKSUM */
 839                                 0x02, /* PKT_TX_IPV4  */
 840                                 0x12, /* PKT_TX_IPV4 | PKT_TX_TCP_CKSUM */
 841                                 0x22, /* PKT_TX_IPV4 | PKT_TX_SCTP_CKSUM */
 842                                 0x32, /* PKT_TX_IPV4 | PKT_TX_UDP_CKSUM */
 843                                 0x03, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM */
 844                                 0x13, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM |
 845                                        * PKT_TX_TCP_CKSUM
 846                                        */
 847                                 0x23, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM |
 848                                        * PKT_TX_SCTP_CKSUM
 849                                        */
 850                                 0x33, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM |
 851                                        * PKT_TX_UDP_CKSUM
 852                                        */
 853                         };
 854
 855                         /* Extract olflags to translate to iltypes */
 856                         xtmp128 = vzip1q_u64(len_olflags0, len_olflags1);
 857                         ytmp128 = vzip1q_u64(len_olflags2, len_olflags3);
 858
 859                         /*
 860                          * E(47):L3_LEN(9):L2_LEN(7+z)
 861                          * E(47):L3_LEN(9):L2_LEN(7+z)
 862                          */
 863                         senddesc01_w1 = vshlq_n_u64(senddesc01_w1, 1);
 864                         senddesc23_w1 = vshlq_n_u64(senddesc23_w1, 1);
 865
 866                         /* Move OLFLAGS bits 55:52 to 51:48
 867                          * with zeros preprended on the byte and rest
 868                          * don't care
 869                          */
 870                         xtmp128 = vshrq_n_u8(xtmp128, 4);
 871                         ytmp128 = vshrq_n_u8(ytmp128, 4);
 872                         /*
 873                          * E(48):L3_LEN(8):L2_LEN(z+7)
 874                          * E(48):L3_LEN(8):L2_LEN(z+7)
 875                          */
 876                         const int8x16_t tshft3 = {
 877                                 -1, 0, 8, 8, 8, 8, 8, 8,
 878                                 -1, 0, 8, 8, 8, 8, 8, 8,
 879                         };
 880
 881                         senddesc01_w1 = vshlq_u8(senddesc01_w1, tshft3);
 882                         senddesc23_w1 = vshlq_u8(senddesc23_w1, tshft3);
 883
 884                         /* Do the lookup */
 885                         ltypes01 = vqtbl1q_u8(tbl, xtmp128);
 886                         ltypes23 = vqtbl1q_u8(tbl, ytmp128);
 887
 888                         /* Pick only relevant fields i.e Bit 48:55 of iltype
 889                          * and place it in ol3/ol4type of senddesc_w1
 890                          */
 891                         const uint8x16_t shuf_mask0 = {
 892                                 0xFF, 0xFF, 0xFF, 0xFF, 0x6, 0xFF, 0xFF, 0xFF,
 893                                 0xFF, 0xFF, 0xFF, 0xFF, 0xE, 0xFF, 0xFF, 0xFF,
 894                         };
 895
 896                         ltypes01 = vqtbl1q_u8(ltypes01, shuf_mask0);
 897                         ltypes23 = vqtbl1q_u8(ltypes23, shuf_mask0);
 898
 899                         /* Prepare ol4ptr, ol3ptr from ol3len, ol2len.
 900                          * a [E(32):E(16):OL3(8):OL2(8)]
 901                          * a = a + (a << 8)
 902                          * a [E(32):E(16):(OL3+OL2):OL2]
 903                          * => E(32):E(16)::OL4PTR(8):OL3PTR(8)
 904                          */
 905                         senddesc01_w1 = vaddq_u8(senddesc01_w1,
 906                                                  vshlq_n_u16(senddesc01_w1, 8));
 907                         senddesc23_w1 = vaddq_u8(senddesc23_w1,
 908                                                  vshlq_n_u16(senddesc23_w1, 8));
 909
 910                         /* Move ltypes to senddesc*_w1 */
 911                         senddesc01_w1 = vorrq_u64(senddesc01_w1, ltypes01);
 912                         senddesc23_w1 = vorrq_u64(senddesc23_w1, ltypes23);
 913                 } else if (!(flags & NIX_TX_OFFLOAD_L3_L4_CSUM_F) &&
 914                            (flags & NIX_TX_OFFLOAD_OL3_OL4_CSUM_F)) {
 915                         /*
 916                          * Lookup table to translate ol_flags to
 917                          * ol3/ol4 types.
 918                          */
 919
 920                         const uint8x16_t tbl = {
 921                                 /* [0-15] = ol4type:ol3type */
 922                                 0x00, /* none */
 923                                 0x03, /* OUTER_IP_CKSUM */
 924                                 0x02, /* OUTER_IPV4 */
 925                                 0x03, /* OUTER_IPV4 | OUTER_IP_CKSUM */
 926                                 0x04, /* OUTER_IPV6 */
 927                                 0x00, /* OUTER_IPV6 | OUTER_IP_CKSUM */
 928                                 0x00, /* OUTER_IPV6 | OUTER_IPV4 */
 929                                 0x00, /* OUTER_IPV6 | OUTER_IPV4 |
 930                                        * OUTER_IP_CKSUM
 931                                        */
 932                                 0x00, /* OUTER_UDP_CKSUM */
 933                                 0x33, /* OUTER_UDP_CKSUM | OUTER_IP_CKSUM */
 934                                 0x32, /* OUTER_UDP_CKSUM | OUTER_IPV4 */
 935                                 0x33, /* OUTER_UDP_CKSUM | OUTER_IPV4 |
 936                                        * OUTER_IP_CKSUM
 937                                        */
 938                                 0x34, /* OUTER_UDP_CKSUM | OUTER_IPV6 */
 939                                 0x00, /* OUTER_UDP_CKSUM | OUTER_IPV6 |
 940                                        * OUTER_IP_CKSUM
 941                                        */
 942                                 0x00, /* OUTER_UDP_CKSUM | OUTER_IPV6 |
 943                                        * OUTER_IPV4
 944                                        */
 945                                 0x00, /* OUTER_UDP_CKSUM | OUTER_IPV6 |
 946                                        * OUTER_IPV4 | OUTER_IP_CKSUM
 947                                        */
 948                         };
 949
 950                         /* Extract olflags to translate to iltypes */
 951                         xtmp128 = vzip1q_u64(len_olflags0, len_olflags1);
 952                         ytmp128 = vzip1q_u64(len_olflags2, len_olflags3);
 953
 954                         /*
 955                          * E(47):OL3_LEN(9):OL2_LEN(7+z)
 956                          * E(47):OL3_LEN(9):OL2_LEN(7+z)
 957                          */
 958                         const uint8x16_t shuf_mask5 = {
 959                                 0x6, 0x5, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
 960                                 0xE, 0xD, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
 961                         };
 962                         senddesc01_w1 = vqtbl1q_u8(senddesc01_w1, shuf_mask5);
 963                         senddesc23_w1 = vqtbl1q_u8(senddesc23_w1, shuf_mask5);
 964
 965                         /* Extract outer ol flags only */
 966                         const uint64x2_t o_cksum_mask = {
 967                                 0x1C00020000000000,
 968                                 0x1C00020000000000,
 969                         };
 970
 971                         xtmp128 = vandq_u64(xtmp128, o_cksum_mask);
 972                         ytmp128 = vandq_u64(ytmp128, o_cksum_mask);
 973
 974                         /* Extract OUTER_UDP_CKSUM bit 41 and
 975                          * move it to bit 61
 976                          */
 977
 978                         xtmp128 = xtmp128 | vshlq_n_u64(xtmp128, 20);
 979                         ytmp128 = ytmp128 | vshlq_n_u64(ytmp128, 20);
 980
 981                         /* Shift oltype by 2 to start nibble from BIT(56)
 982                          * instead of BIT(58)
 983                          */
 984                         xtmp128 = vshrq_n_u8(xtmp128, 2);
 985                         ytmp128 = vshrq_n_u8(ytmp128, 2);
 986                         /*
 987                          * E(48):L3_LEN(8):L2_LEN(z+7)
 988                          * E(48):L3_LEN(8):L2_LEN(z+7)
 989                          */
 990                         const int8x16_t tshft3 = {
 991                                 -1, 0, 8, 8, 8, 8, 8, 8,
 992                                 -1, 0, 8, 8, 8, 8, 8, 8,
 993                         };
 994
 995                         senddesc01_w1 = vshlq_u8(senddesc01_w1, tshft3);
 996                         senddesc23_w1 = vshlq_u8(senddesc23_w1, tshft3);
 997
 998                         /* Do the lookup */
 999                         ltypes01 = vqtbl1q_u8(tbl, xtmp128);
1000                         ltypes23 = vqtbl1q_u8(tbl, ytmp128);
1001
1002                         /* Pick only relevant fields i.e Bit 56:63 of oltype
1003                          * and place it in ol3/ol4type of senddesc_w1
1004                          */
1005                         const uint8x16_t shuf_mask0 = {
1006                                 0xFF, 0xFF, 0xFF, 0xFF, 0x7, 0xFF, 0xFF, 0xFF,
1007                                 0xFF, 0xFF, 0xFF, 0xFF, 0xF, 0xFF, 0xFF, 0xFF,
1008                         };
1009
1010                         ltypes01 = vqtbl1q_u8(ltypes01, shuf_mask0);
1011                         ltypes23 = vqtbl1q_u8(ltypes23, shuf_mask0);
1012
1013                         /* Prepare ol4ptr, ol3ptr from ol3len, ol2len.
1014                          * a [E(32):E(16):OL3(8):OL2(8)]
1015                          * a = a + (a << 8)
1016                          * a [E(32):E(16):(OL3+OL2):OL2]
1017                          * => E(32):E(16)::OL4PTR(8):OL3PTR(8)
1018                          */
1019                         senddesc01_w1 = vaddq_u8(senddesc01_w1,
1020                                                  vshlq_n_u16(senddesc01_w1, 8));
1021                         senddesc23_w1 = vaddq_u8(senddesc23_w1,
1022                                                  vshlq_n_u16(senddesc23_w1, 8));
1023
1024                         /* Move ltypes to senddesc*_w1 */
1025                         senddesc01_w1 = vorrq_u64(senddesc01_w1, ltypes01);
1026                         senddesc23_w1 = vorrq_u64(senddesc23_w1, ltypes23);
1027                 } else if ((flags & NIX_TX_OFFLOAD_L3_L4_CSUM_F) &&
1028                            (flags & NIX_TX_OFFLOAD_OL3_OL4_CSUM_F)) {
1029                         /* Lookup table to translate ol_flags to
1030                          * ol4type, ol3type, il4type, il3type of senddesc_w1
1031                          */
1032                         const uint8x16x2_t tbl = {{
1033                                 {
1034                                         /* [0-15] = il4type:il3type */
1035                                         0x04, /* none (IPv6) */
1036                                         0x14, /* PKT_TX_TCP_CKSUM (IPv6) */
1037                                         0x24, /* PKT_TX_SCTP_CKSUM (IPv6) */
1038                                         0x34, /* PKT_TX_UDP_CKSUM (IPv6) */
1039                                         0x03, /* PKT_TX_IP_CKSUM */
1040                                         0x13, /* PKT_TX_IP_CKSUM |
1041                                                * PKT_TX_TCP_CKSUM
1042                                                */
1043                                         0x23, /* PKT_TX_IP_CKSUM |
1044                                                * PKT_TX_SCTP_CKSUM
1045                                                */
1046                                         0x33, /* PKT_TX_IP_CKSUM |
1047                                                * PKT_TX_UDP_CKSUM
1048                                                */
1049                                         0x02, /* PKT_TX_IPV4 */
1050                                         0x12, /* PKT_TX_IPV4 |
1051                                                * PKT_TX_TCP_CKSUM
1052                                                */
1053                                         0x22, /* PKT_TX_IPV4 |
1054                                                * PKT_TX_SCTP_CKSUM
1055                                                */
1056                                         0x32, /* PKT_TX_IPV4 |
1057                                                * PKT_TX_UDP_CKSUM
1058                                                */
1059                                         0x03, /* PKT_TX_IPV4 |
1060                                                * PKT_TX_IP_CKSUM
1061                                                */
1062                                         0x13, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM |
1063                                                * PKT_TX_TCP_CKSUM
1064                                                */
1065                                         0x23, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM |
1066                                                * PKT_TX_SCTP_CKSUM
1067                                                */
1068                                         0x33, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM |
1069                                                * PKT_TX_UDP_CKSUM
1070                                                */
1071                                 },
1072
1073                                 {
1074                                         /* [16-31] = ol4type:ol3type */
1075                                         0x00, /* none */
1076                                         0x03, /* OUTER_IP_CKSUM */
1077                                         0x02, /* OUTER_IPV4 */
1078                                         0x03, /* OUTER_IPV4 | OUTER_IP_CKSUM */
1079                                         0x04, /* OUTER_IPV6 */
1080                                         0x00, /* OUTER_IPV6 | OUTER_IP_CKSUM */
1081                                         0x00, /* OUTER_IPV6 | OUTER_IPV4 */
1082                                         0x00, /* OUTER_IPV6 | OUTER_IPV4 |
1083                                                * OUTER_IP_CKSUM
1084                                                */
1085                                         0x00, /* OUTER_UDP_CKSUM */
1086                                         0x33, /* OUTER_UDP_CKSUM |
1087                                                * OUTER_IP_CKSUM
1088                                                */
1089                                         0x32, /* OUTER_UDP_CKSUM |
1090                                                * OUTER_IPV4
1091                                                */
1092                                         0x33, /* OUTER_UDP_CKSUM |
1093                                                * OUTER_IPV4 | OUTER_IP_CKSUM
1094                                                */
1095                                         0x34, /* OUTER_UDP_CKSUM |
1096                                                * OUTER_IPV6
1097                                                */
1098                                         0x00, /* OUTER_UDP_CKSUM | OUTER_IPV6 |
1099                                                * OUTER_IP_CKSUM
1100                                                */
1101                                         0x00, /* OUTER_UDP_CKSUM | OUTER_IPV6 |
1102                                                * OUTER_IPV4
1103                                                */
1104                                         0x00, /* OUTER_UDP_CKSUM | OUTER_IPV6 |
1105                                                * OUTER_IPV4 | OUTER_IP_CKSUM
1106                                                */
1107                                 },
1108                         }};
1109
1110                         /* Extract olflags to translate to oltype & iltype */
1111                         xtmp128 = vzip1q_u64(len_olflags0, len_olflags1);
1112                         ytmp128 = vzip1q_u64(len_olflags2, len_olflags3);
1113
1114                         /*
1115                          * E(8):OL2_LN(7):OL3_LN(9):E(23):L3_LN(9):L2_LN(7+z)
1116                          * E(8):OL2_LN(7):OL3_LN(9):E(23):L3_LN(9):L2_LN(7+z)
1117                          */
1118                         const uint32x4_t tshft_4 = {
1119                                 1,
1120                                 0,
1121                                 1,
1122                                 0,
1123                         };
1124                         senddesc01_w1 = vshlq_u32(senddesc01_w1, tshft_4);
1125                         senddesc23_w1 = vshlq_u32(senddesc23_w1, tshft_4);
1126
1127                         /*
1128                          * E(32):L3_LEN(8):L2_LEN(7+Z):OL3_LEN(8):OL2_LEN(7+Z)
1129                          * E(32):L3_LEN(8):L2_LEN(7+Z):OL3_LEN(8):OL2_LEN(7+Z)
1130                          */
1131                         const uint8x16_t shuf_mask5 = {
1132                                 0x6, 0x5, 0x0, 0x1, 0xFF, 0xFF, 0xFF, 0xFF,
1133                                 0xE, 0xD, 0x8, 0x9, 0xFF, 0xFF, 0xFF, 0xFF,
1134                         };
1135                         senddesc01_w1 = vqtbl1q_u8(senddesc01_w1, shuf_mask5);
1136                         senddesc23_w1 = vqtbl1q_u8(senddesc23_w1, shuf_mask5);
1137
1138                         /* Extract outer and inner header ol_flags */
1139                         const uint64x2_t oi_cksum_mask = {
1140                                 0x1CF0020000000000,
1141                                 0x1CF0020000000000,
1142                         };
1143
1144                         xtmp128 = vandq_u64(xtmp128, oi_cksum_mask);
1145                         ytmp128 = vandq_u64(ytmp128, oi_cksum_mask);
1146
1147                         /* Extract OUTER_UDP_CKSUM bit 41 and
1148                          * move it to bit 61
1149                          */
1150
1151                         xtmp128 = xtmp128 | vshlq_n_u64(xtmp128, 20);
1152                         ytmp128 = ytmp128 | vshlq_n_u64(ytmp128, 20);
1153
1154                         /* Shift right oltype by 2 and iltype by 4
1155                          * to start oltype nibble from BIT(58)
1156                          * instead of BIT(56) and iltype nibble from BIT(48)
1157                          * instead of BIT(52).
1158                          */
1159                         const int8x16_t tshft5 = {
1160                                 8, 8, 8, 8, 8, 8, -4, -2,
1161                                 8, 8, 8, 8, 8, 8, -4, -2,
1162                         };
1163
1164                         xtmp128 = vshlq_u8(xtmp128, tshft5);
1165                         ytmp128 = vshlq_u8(ytmp128, tshft5);
1166                         /*
1167                          * E(32):L3_LEN(8):L2_LEN(8):OL3_LEN(8):OL2_LEN(8)
1168                          * E(32):L3_LEN(8):L2_LEN(8):OL3_LEN(8):OL2_LEN(8)
1169                          */
1170                         const int8x16_t tshft3 = {
1171                                 -1, 0, -1, 0, 0, 0, 0, 0,
1172                                 -1, 0, -1, 0, 0, 0, 0, 0,
1173                         };
1174
1175                         senddesc01_w1 = vshlq_u8(senddesc01_w1, tshft3);
1176                         senddesc23_w1 = vshlq_u8(senddesc23_w1, tshft3);
1177
1178                         /* Mark Bit(4) of oltype */
1179                         const uint64x2_t oi_cksum_mask2 = {
1180                                 0x1000000000000000,
1181                                 0x1000000000000000,
1182                         };
1183
1184                         xtmp128 = vorrq_u64(xtmp128, oi_cksum_mask2);
1185                         ytmp128 = vorrq_u64(ytmp128, oi_cksum_mask2);
1186
1187                         /* Do the lookup */
1188                         ltypes01 = vqtbl2q_u8(tbl, xtmp128);
1189                         ltypes23 = vqtbl2q_u8(tbl, ytmp128);
1190
1191                         /* Pick only relevant fields i.e Bit 48:55 of iltype and
1192                          * Bit 56:63 of oltype and place it in corresponding
1193                          * place in senddesc_w1.
1194                          */
1195                         const uint8x16_t shuf_mask0 = {
1196                                 0xFF, 0xFF, 0xFF, 0xFF, 0x7, 0x6, 0xFF, 0xFF,
1197                                 0xFF, 0xFF, 0xFF, 0xFF, 0xF, 0xE, 0xFF, 0xFF,
1198                         };
1199
1200                         ltypes01 = vqtbl1q_u8(ltypes01, shuf_mask0);
1201                         ltypes23 = vqtbl1q_u8(ltypes23, shuf_mask0);
1202
1203                         /* Prepare l4ptr, l3ptr, ol4ptr, ol3ptr from
1204                          * l3len, l2len, ol3len, ol2len.
1205                          * a [E(32):L3(8):L2(8):OL3(8):OL2(8)]
1206                          * a = a + (a << 8)
1207                          * a [E:(L3+L2):(L2+OL3):(OL3+OL2):OL2]
1208                          * a = a + (a << 16)
1209                          * a [E:(L3+L2+OL3+OL2):(L2+OL3+OL2):(OL3+OL2):OL2]
1210                          * => E(32):IL4PTR(8):IL3PTR(8):OL4PTR(8):OL3PTR(8)
1211                          */
1212                         senddesc01_w1 = vaddq_u8(senddesc01_w1,
1213                                                  vshlq_n_u32(senddesc01_w1, 8));
1214                         senddesc23_w1 = vaddq_u8(senddesc23_w1,
1215                                                  vshlq_n_u32(senddesc23_w1, 8));
1216
1217                         /* Continue preparing l4ptr, l3ptr, ol4ptr, ol3ptr */
1218                         senddesc01_w1 = vaddq_u8(
1219                                 senddesc01_w1, vshlq_n_u32(senddesc01_w1, 16));
1220                         senddesc23_w1 = vaddq_u8(
1221                                 senddesc23_w1, vshlq_n_u32(senddesc23_w1, 16));
1222
1223                         /* Move ltypes to senddesc*_w1 */
1224                         senddesc01_w1 = vorrq_u64(senddesc01_w1, ltypes01);
1225                         senddesc23_w1 = vorrq_u64(senddesc23_w1, ltypes23);
1226                 }
1227
1228                 xmask01 = vdupq_n_u64(0);
1229                 xmask23 = xmask01;
1230                 asm volatile("LD1 {%[a].H}[0],[%[in]]\n\t"
1231                              : [a] "+w"(xmask01)
1232                              : [in] "r"(mbuf0)
1233                              : "memory");
1234
1235                 asm volatile("LD1 {%[a].H}[4],[%[in]]\n\t"
1236                              : [a] "+w"(xmask01)
1237                              : [in] "r"(mbuf1)
1238                              : "memory");
1239
1240                 asm volatile("LD1 {%[b].H}[0],[%[in]]\n\t"
1241                              : [b] "+w"(xmask23)
1242                              : [in] "r"(mbuf2)
1243                              : "memory");
1244
1245                 asm volatile("LD1 {%[b].H}[4],[%[in]]\n\t"
1246                              : [b] "+w"(xmask23)
1247                              : [in] "r"(mbuf3)
1248                              : "memory");
1249                 xmask01 = vshlq_n_u64(xmask01, 20);
1250                 xmask23 = vshlq_n_u64(xmask23, 20);
1251
1252                 senddesc01_w0 = vorrq_u64(senddesc01_w0, xmask01);
1253                 senddesc23_w0 = vorrq_u64(senddesc23_w0, xmask23);
1254
1255                 if (flags & NIX_TX_OFFLOAD_VLAN_QINQ_F) {
1256                         /* Tx ol_flag for vlan. */
1257                         const uint64x2_t olv = {PKT_TX_VLAN, PKT_TX_VLAN};
1258                         /* Bit enable for VLAN1 */
1259                         const uint64x2_t mlv = {BIT_ULL(49), BIT_ULL(49)};
1260                         /* Tx ol_flag for QnQ. */
1261                         const uint64x2_t olq = {PKT_TX_QINQ, PKT_TX_QINQ};
1262                         /* Bit enable for VLAN0 */
1263                         const uint64x2_t mlq = {BIT_ULL(48), BIT_ULL(48)};
1264                         /* Load vlan values from packet. outer is VLAN 0 */
1265                         uint64x2_t ext01 = {
1266                                 ((uint32_t)tx_pkts[0]->vlan_tci_outer) << 8 |
1267                                         ((uint64_t)tx_pkts[0]->vlan_tci) << 32,
1268                                 ((uint32_t)tx_pkts[1]->vlan_tci_outer) << 8 |
1269                                         ((uint64_t)tx_pkts[1]->vlan_tci) << 32,
1270                         };
1271                         uint64x2_t ext23 = {
1272                                 ((uint32_t)tx_pkts[2]->vlan_tci_outer) << 8 |
1273                                         ((uint64_t)tx_pkts[2]->vlan_tci) << 32,
1274                                 ((uint32_t)tx_pkts[3]->vlan_tci_outer) << 8 |
1275                                         ((uint64_t)tx_pkts[3]->vlan_tci) << 32,
1276                         };
1277
1278                         /* Get ol_flags of the packets. */
1279                         xtmp128 = vzip1q_u64(len_olflags0, len_olflags1);
1280                         ytmp128 = vzip1q_u64(len_olflags2, len_olflags3);
1281
1282                         /* ORR vlan outer/inner values into cmd. */
1283                         sendext01_w1 = vorrq_u64(sendext01_w1, ext01);
1284                         sendext23_w1 = vorrq_u64(sendext23_w1, ext23);
1285
1286                         /* Test for offload enable bits and generate masks. */
1287                         xtmp128 = vorrq_u64(vandq_u64(vtstq_u64(xtmp128, olv),
1288                                                       mlv),
1289                                             vandq_u64(vtstq_u64(xtmp128, olq),
1290                                                       mlq));
1291                         ytmp128 = vorrq_u64(vandq_u64(vtstq_u64(ytmp128, olv),
1292                                                       mlv),
1293                                             vandq_u64(vtstq_u64(ytmp128, olq),
1294                                                       mlq));
1295
1296                         /* Set vlan enable bits into cmd based on mask. */
1297                         sendext01_w1 = vorrq_u64(sendext01_w1, xtmp128);
1298                         sendext23_w1 = vorrq_u64(sendext23_w1, ytmp128);
1299                 }
1300
1301                 if (flags & NIX_TX_OFFLOAD_TSTAMP_F) {
1302                         /* Tx ol_flag for timestam. */
1303                         const uint64x2_t olf = {PKT_TX_IEEE1588_TMST,
1304                                                 PKT_TX_IEEE1588_TMST};
1305                         /* Set send mem alg to SUB. */
1306                         const uint64x2_t alg = {BIT_ULL(59), BIT_ULL(59)};
1307                         /* Increment send mem address by 8. */
1308                         const uint64x2_t addr = {0x8, 0x8};
1309
1310                         xtmp128 = vzip1q_u64(len_olflags0, len_olflags1);
1311                         ytmp128 = vzip1q_u64(len_olflags2, len_olflags3);
1312
1313                         /* Check if timestamp is requested and generate inverted
1314                          * mask as we need not make any changes to default cmd
1315                          * value.
1316                          */
1317                         xtmp128 = vmvnq_u32(vtstq_u64(olf, xtmp128));
1318                         ytmp128 = vmvnq_u32(vtstq_u64(olf, ytmp128));
1319
1320                         /* Change send mem address to an 8 byte offset when
1321                          * TSTMP is disabled.
1322                          */
1323                         sendmem01_w1 = vaddq_u64(sendmem01_w1,
1324                                                  vandq_u64(xtmp128, addr));
1325                         sendmem23_w1 = vaddq_u64(sendmem23_w1,
1326                                                  vandq_u64(ytmp128, addr));
1327                         /* Change send mem alg to SUB when TSTMP is disabled. */
1328                         sendmem01_w0 = vorrq_u64(sendmem01_w0,
1329                                                  vandq_u64(xtmp128, alg));
1330                         sendmem23_w0 = vorrq_u64(sendmem23_w0,
1331                                                  vandq_u64(ytmp128, alg));
1332
1333                         cmd3[0] = vzip1q_u64(sendmem01_w0, sendmem01_w1);
1334                         cmd3[1] = vzip2q_u64(sendmem01_w0, sendmem01_w1);
1335                         cmd3[2] = vzip1q_u64(sendmem23_w0, sendmem23_w1);
1336                         cmd3[3] = vzip2q_u64(sendmem23_w0, sendmem23_w1);
1337                 }
1338
1339                 if (flags & NIX_TX_OFFLOAD_TSO_F) {
1340                         uint64_t sx_w0[NIX_DESCS_PER_LOOP];
1341                         uint64_t sd_w1[NIX_DESCS_PER_LOOP];
1342
1343                         /* Extract SD W1 as we need to set L4 types. */
1344                         vst1q_u64(sd_w1, senddesc01_w1);
1345                         vst1q_u64(sd_w1 + 2, senddesc23_w1);
1346
1347                         /* Extract SX W0 as we need to set LSO fields. */
1348                         vst1q_u64(sx_w0, sendext01_w0);
1349                         vst1q_u64(sx_w0 + 2, sendext23_w0);
1350
1351                         /* Extract ol_flags. */
1352                         xtmp128 = vzip1q_u64(len_olflags0, len_olflags1);
1353                         ytmp128 = vzip1q_u64(len_olflags2, len_olflags3);
1354
1355                         /* Prepare individual mbufs. */
1356                         cn9k_nix_prepare_tso(tx_pkts[0],
1357                                 (union nix_send_hdr_w1_u *)&sd_w1[0],
1358                                 (union nix_send_ext_w0_u *)&sx_w0[0],
1359                                 vgetq_lane_u64(xtmp128, 0), flags);
1360
1361                         cn9k_nix_prepare_tso(tx_pkts[1],
1362                                 (union nix_send_hdr_w1_u *)&sd_w1[1],
1363                                 (union nix_send_ext_w0_u *)&sx_w0[1],
1364                                 vgetq_lane_u64(xtmp128, 1), flags);
1365
1366                         cn9k_nix_prepare_tso(tx_pkts[2],
1367                                 (union nix_send_hdr_w1_u *)&sd_w1[2],
1368                                 (union nix_send_ext_w0_u *)&sx_w0[2],
1369                                 vgetq_lane_u64(ytmp128, 0), flags);
1370
1371                         cn9k_nix_prepare_tso(tx_pkts[3],
1372                                 (union nix_send_hdr_w1_u *)&sd_w1[3],
1373                                 (union nix_send_ext_w0_u *)&sx_w0[3],
1374                                 vgetq_lane_u64(ytmp128, 1), flags);
1375
1376                         senddesc01_w1 = vld1q_u64(sd_w1);
1377                         senddesc23_w1 = vld1q_u64(sd_w1 + 2);
1378
1379                         sendext01_w0 = vld1q_u64(sx_w0);
1380                         sendext23_w0 = vld1q_u64(sx_w0 + 2);
1381                 }
1382
1383                 if (flags & NIX_TX_OFFLOAD_MBUF_NOFF_F) {
1384                         /* Set don't free bit if reference count > 1 */
1385                         xmask01 = vdupq_n_u64(0);
1386                         xmask23 = xmask01;
1387
1388                         /* Move mbufs to iova */
1389                         mbuf0 = (uint64_t *)tx_pkts[0];
1390                         mbuf1 = (uint64_t *)tx_pkts[1];
1391                         mbuf2 = (uint64_t *)tx_pkts[2];
1392                         mbuf3 = (uint64_t *)tx_pkts[3];
1393
1394                         if (cnxk_nix_prefree_seg((struct rte_mbuf *)mbuf0))
1395                                 vsetq_lane_u64(0x80000, xmask01, 0);
1396                         else
1397                                 __mempool_check_cookies(
1398                                         ((struct rte_mbuf *)mbuf0)->pool,
1399                                         (void **)&mbuf0, 1, 0);
1400
1401                         if (cnxk_nix_prefree_seg((struct rte_mbuf *)mbuf1))
1402                                 vsetq_lane_u64(0x80000, xmask01, 1);
1403                         else
1404                                 __mempool_check_cookies(
1405                                         ((struct rte_mbuf *)mbuf1)->pool,
1406                                         (void **)&mbuf1, 1, 0);
1407
1408                         if (cnxk_nix_prefree_seg((struct rte_mbuf *)mbuf2))
1409                                 vsetq_lane_u64(0x80000, xmask23, 0);
1410                         else
1411                                 __mempool_check_cookies(
1412                                         ((struct rte_mbuf *)mbuf2)->pool,
1413                                         (void **)&mbuf2, 1, 0);
1414
1415                         if (cnxk_nix_prefree_seg((struct rte_mbuf *)mbuf3))
1416                                 vsetq_lane_u64(0x80000, xmask23, 1);
1417                         else
1418                                 __mempool_check_cookies(
1419                                         ((struct rte_mbuf *)mbuf3)->pool,
1420                                         (void **)&mbuf3, 1, 0);
1421                         senddesc01_w0 = vorrq_u64(senddesc01_w0, xmask01);
1422                         senddesc23_w0 = vorrq_u64(senddesc23_w0, xmask23);
1423                         /* Ensuring mbuf fields which got updated in
1424                          * cnxk_nix_prefree_seg are written before LMTST.
1425                          */
1426                         rte_io_wmb();
1427                 } else {
1428                         /* Move mbufs to iova */
1429                         mbuf0 = (uint64_t *)tx_pkts[0];
1430                         mbuf1 = (uint64_t *)tx_pkts[1];
1431                         mbuf2 = (uint64_t *)tx_pkts[2];
1432                         mbuf3 = (uint64_t *)tx_pkts[3];
1433
1434                         /* Mark mempool object as "put" since
1435                          * it is freed by NIX
1436                          */
1437                         __mempool_check_cookies(
1438                                 ((struct rte_mbuf *)mbuf0)->pool,
1439                                 (void **)&mbuf0, 1, 0);
1440
1441                         __mempool_check_cookies(
1442                                 ((struct rte_mbuf *)mbuf1)->pool,
1443                                 (void **)&mbuf1, 1, 0);
1444
1445                         __mempool_check_cookies(
1446                                 ((struct rte_mbuf *)mbuf2)->pool,
1447                                 (void **)&mbuf2, 1, 0);
1448
1449                         __mempool_check_cookies(
1450                                 ((struct rte_mbuf *)mbuf3)->pool,
1451                                 (void **)&mbuf3, 1, 0);
1452 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
1453                         rte_io_wmb();
1454 #endif
1455                 }
1456
1457                 /* Create 4W cmd for 4 mbufs (sendhdr, sgdesc) */
1458                 cmd0[0] = vzip1q_u64(senddesc01_w0, senddesc01_w1);
1459                 cmd0[1] = vzip2q_u64(senddesc01_w0, senddesc01_w1);
1460                 cmd0[2] = vzip1q_u64(senddesc23_w0, senddesc23_w1);
1461                 cmd0[3] = vzip2q_u64(senddesc23_w0, senddesc23_w1);
1462
1463                 cmd1[0] = vzip1q_u64(sgdesc01_w0, sgdesc01_w1);
1464                 cmd1[1] = vzip2q_u64(sgdesc01_w0, sgdesc01_w1);
1465                 cmd1[2] = vzip1q_u64(sgdesc23_w0, sgdesc23_w1);
1466                 cmd1[3] = vzip2q_u64(sgdesc23_w0, sgdesc23_w1);
1467
1468                 if (flags & NIX_TX_NEED_EXT_HDR) {
1469                         cmd2[0] = vzip1q_u64(sendext01_w0, sendext01_w1);
1470                         cmd2[1] = vzip2q_u64(sendext01_w0, sendext01_w1);
1471                         cmd2[2] = vzip1q_u64(sendext23_w0, sendext23_w1);
1472                         cmd2[3] = vzip2q_u64(sendext23_w0, sendext23_w1);
1473                 }
1474
1475                 if (flags & NIX_TX_NEED_EXT_HDR) {
1476                         /* With ext header in the command we can no longer send
1477                          * all 4 packets together since LMTLINE is 128bytes.
1478                          * Split and Tx twice.
1479                          */
1480                         do {
1481                                 if (flags & NIX_TX_OFFLOAD_TSTAMP_F) {
1482                                         vst1q_u64(lmt_addr, cmd0[0]);
1483                                         vst1q_u64(lmt_addr + 2, cmd2[0]);
1484                                         vst1q_u64(lmt_addr + 4, cmd1[0]);
1485                                         vst1q_u64(lmt_addr + 6, cmd3[0]);
1486                                         vst1q_u64(lmt_addr + 8, cmd0[1]);
1487                                         vst1q_u64(lmt_addr + 10, cmd2[1]);
1488                                         vst1q_u64(lmt_addr + 12, cmd1[1]);
1489                                         vst1q_u64(lmt_addr + 14, cmd3[1]);
1490                                 } else {
1491                                         vst1q_u64(lmt_addr, cmd0[0]);
1492                                         vst1q_u64(lmt_addr + 2, cmd2[0]);
1493                                         vst1q_u64(lmt_addr + 4, cmd1[0]);
1494                                         vst1q_u64(lmt_addr + 6, cmd0[1]);
1495                                         vst1q_u64(lmt_addr + 8, cmd2[1]);
1496                                         vst1q_u64(lmt_addr + 10, cmd1[1]);
1497                                 }
1498                                 lmt_status = roc_lmt_submit_ldeor(io_addr);
1499                         } while (lmt_status == 0);
1500
1501                         do {
1502                                 if (flags & NIX_TX_OFFLOAD_TSTAMP_F) {
1503                                         vst1q_u64(lmt_addr, cmd0[2]);
1504                                         vst1q_u64(lmt_addr + 2, cmd2[2]);
1505                                         vst1q_u64(lmt_addr + 4, cmd1[2]);
1506                                         vst1q_u64(lmt_addr + 6, cmd3[2]);
1507                                         vst1q_u64(lmt_addr + 8, cmd0[3]);
1508                                         vst1q_u64(lmt_addr + 10, cmd2[3]);
1509                                         vst1q_u64(lmt_addr + 12, cmd1[3]);
1510                                         vst1q_u64(lmt_addr + 14, cmd3[3]);
1511                                 } else {
1512                                         vst1q_u64(lmt_addr, cmd0[2]);
1513                                         vst1q_u64(lmt_addr + 2, cmd2[2]);
1514                                         vst1q_u64(lmt_addr + 4, cmd1[2]);
1515                                         vst1q_u64(lmt_addr + 6, cmd0[3]);
1516                                         vst1q_u64(lmt_addr + 8, cmd2[3]);
1517                                         vst1q_u64(lmt_addr + 10, cmd1[3]);
1518                                 }
1519                                 lmt_status = roc_lmt_submit_ldeor(io_addr);
1520                         } while (lmt_status == 0);
1521                 } else {
1522                         do {
1523                                 vst1q_u64(lmt_addr, cmd0[0]);
1524                                 vst1q_u64(lmt_addr + 2, cmd1[0]);
1525                                 vst1q_u64(lmt_addr + 4, cmd0[1]);
1526                                 vst1q_u64(lmt_addr + 6, cmd1[1]);
1527                                 vst1q_u64(lmt_addr + 8, cmd0[2]);
1528                                 vst1q_u64(lmt_addr + 10, cmd1[2]);
1529                                 vst1q_u64(lmt_addr + 12, cmd0[3]);
1530                                 vst1q_u64(lmt_addr + 14, cmd1[3]);
1531                                 lmt_status = roc_lmt_submit_ldeor(io_addr);
1532                         } while (lmt_status == 0);
1533                 }
1534                 tx_pkts = tx_pkts + NIX_DESCS_PER_LOOP;
1535         }
1536
1537         if (unlikely(pkts_left))
1538                 pkts += cn9k_nix_xmit_pkts(tx_queue, tx_pkts, pkts_left, cmd,
1539                                            flags);
1540
1541         return pkts;
1542 }
1543
1544 #else
1545 static __rte_always_inline uint16_t
1546 cn9k_nix_xmit_pkts_vector(void *tx_queue, struct rte_mbuf **tx_pkts,
1547                           uint16_t pkts, uint64_t *cmd, const uint16_t flags)
1548 {
1549         RTE_SET_USED(tx_queue);
1550         RTE_SET_USED(tx_pkts);
1551         RTE_SET_USED(pkts);
1552         RTE_SET_USED(cmd);
1553         RTE_SET_USED(flags);
1554         return 0;
1555 }
1556 #endif
1557
1558 #define L3L4CSUM_F   NIX_TX_OFFLOAD_L3_L4_CSUM_F
1559 #define OL3OL4CSUM_F NIX_TX_OFFLOAD_OL3_OL4_CSUM_F
1560 #define VLAN_F       NIX_TX_OFFLOAD_VLAN_QINQ_F
1561 #define NOFF_F       NIX_TX_OFFLOAD_MBUF_NOFF_F
1562 #define TSO_F        NIX_TX_OFFLOAD_TSO_F
1563 #define TSP_F        NIX_TX_OFFLOAD_TSTAMP_F
1564
1565 /* [TSP] [TSO] [NOFF] [VLAN] [OL3OL4CSUM] [L3L4CSUM] */
1566 #define NIX_TX_FASTPATH_MODES                                                  \
1567 T(no_offload,                           0, 0, 0, 0, 0, 0,       4,             \
1568                 NIX_TX_OFFLOAD_NONE)                                           \
1569 T(l3l4csum,                             0, 0, 0, 0, 0, 1,       4,             \
1570                 L3L4CSUM_F)                                                    \
1571 T(ol3ol4csum,                           0, 0, 0, 0, 1, 0,       4,             \
1572                 OL3OL4CSUM_F)                                                  \
1573 T(ol3ol4csum_l3l4csum,                  0, 0, 0, 0, 1, 1,       4,             \
1574                 OL3OL4CSUM_F | L3L4CSUM_F)                                     \
1575 T(vlan,                                 0, 0, 0, 1, 0, 0,       6,             \
1576                 VLAN_F)                                                        \
1577 T(vlan_l3l4csum,                        0, 0, 0, 1, 0, 1,       6,             \
1578                 VLAN_F | L3L4CSUM_F)                                           \
1579 T(vlan_ol3ol4csum,                      0, 0, 0, 1, 1, 0,       6,             \
1580                 VLAN_F | OL3OL4CSUM_F)                                         \
1581 T(vlan_ol3ol4csum_l3l4csum,             0, 0, 0, 1, 1, 1,       6,             \
1582                 VLAN_F | OL3OL4CSUM_F | L3L4CSUM_F)                            \
1583 T(noff,                                 0, 0, 1, 0, 0, 0,       4,             \
1584                 NOFF_F)                                                        \
1585 T(noff_l3l4csum,                        0, 0, 1, 0, 0, 1,       4,             \
1586                 NOFF_F | L3L4CSUM_F)                                           \
1587 T(noff_ol3ol4csum,                      0, 0, 1, 0, 1, 0,       4,             \
1588                 NOFF_F | OL3OL4CSUM_F)                                         \
1589 T(noff_ol3ol4csum_l3l4csum,             0, 0, 1, 0, 1, 1,       4,             \
1590                 NOFF_F | OL3OL4CSUM_F | L3L4CSUM_F)                            \
1591 T(noff_vlan,                            0, 0, 1, 1, 0, 0,       6,             \
1592                 NOFF_F | VLAN_F)                                               \
1593 T(noff_vlan_l3l4csum,                   0, 0, 1, 1, 0, 1,       6,             \
1594                 NOFF_F | VLAN_F | L3L4CSUM_F)                                  \
1595 T(noff_vlan_ol3ol4csum,                 0, 0, 1, 1, 1, 0,       6,             \
1596                 NOFF_F | VLAN_F | OL3OL4CSUM_F)                                \
1597 T(noff_vlan_ol3ol4csum_l3l4csum,        0, 0, 1, 1, 1, 1,       6,             \
1598                 NOFF_F | VLAN_F | OL3OL4CSUM_F | L3L4CSUM_F)                   \
1599 T(tso,                                  0, 1, 0, 0, 0, 0,       6,             \
1600                 TSO_F)                                                         \
1601 T(tso_l3l4csum,                         0, 1, 0, 0, 0, 1,       6,             \
1602                 TSO_F | L3L4CSUM_F)                                            \
1603 T(tso_ol3ol4csum,                       0, 1, 0, 0, 1, 0,       6,             \
1604                 TSO_F | OL3OL4CSUM_F)                                          \
1605 T(tso_ol3ol4csum_l3l4csum,              0, 1, 0, 0, 1, 1,       6,             \
1606                 TSO_F | OL3OL4CSUM_F | L3L4CSUM_F)                             \
1607 T(tso_vlan,                             0, 1, 0, 1, 0, 0,       6,             \
1608                 TSO_F | VLAN_F)                                                \
1609 T(tso_vlan_l3l4csum,                    0, 1, 0, 1, 0, 1,       6,             \
1610                 TSO_F | VLAN_F | L3L4CSUM_F)                                   \
1611 T(tso_vlan_ol3ol4csum,                  0, 1, 0, 1, 1, 0,       6,             \
1612                 TSO_F | VLAN_F | OL3OL4CSUM_F)                                 \
1613 T(tso_vlan_ol3ol4csum_l3l4csum,         0, 1, 0, 1, 1, 1,       6,             \
1614                 TSO_F | VLAN_F | OL3OL4CSUM_F | L3L4CSUM_F)                    \
1615 T(tso_noff,                             0, 1, 1, 0, 0, 0,       6,             \
1616                 TSO_F | NOFF_F)                                                \
1617 T(tso_noff_l3l4csum,                    0, 1, 1, 0, 0, 1,       6,             \
1618                 TSO_F | NOFF_F | L3L4CSUM_F)                                   \
1619 T(tso_noff_ol3ol4csum,                  0, 1, 1, 0, 1, 0,       6,             \
1620                 TSO_F | NOFF_F | OL3OL4CSUM_F)                                 \
1621 T(tso_noff_ol3ol4csum_l3l4csum,         0, 1, 1, 0, 1, 1,       6,             \
1622                 TSO_F | NOFF_F | OL3OL4CSUM_F | L3L4CSUM_F)                    \
1623 T(tso_noff_vlan,                        0, 1, 1, 1, 0, 0,       6,             \
1624                 TSO_F | NOFF_F | VLAN_F)                                       \
1625 T(tso_noff_vlan_l3l4csum,               0, 1, 1, 1, 0, 1,       6,             \
1626                 TSO_F | NOFF_F | VLAN_F | L3L4CSUM_F)                          \
1627 T(tso_noff_vlan_ol3ol4csum,             0, 1, 1, 1, 1, 0,       6,             \
1628                 TSO_F | NOFF_F | VLAN_F | OL3OL4CSUM_F)                        \
1629 T(tso_noff_vlan_ol3ol4csum_l3l4csum,    0, 1, 1, 1, 1, 1,       6,             \
1630                 TSO_F | NOFF_F | VLAN_F | OL3OL4CSUM_F | L3L4CSUM_F)           \
1631 T(ts,                                   1, 0, 0, 0, 0, 0,       8,             \
1632                 TSP_F)                                                         \
1633 T(ts_l3l4csum,                          1, 0, 0, 0, 0, 1,       8,             \
1634                 TSP_F | L3L4CSUM_F)                                            \
1635 T(ts_ol3ol4csum,                        1, 0, 0, 0, 1, 0,       8,             \
1636                 TSP_F | OL3OL4CSUM_F)                                          \
1637 T(ts_ol3ol4csum_l3l4csum,               1, 0, 0, 0, 1, 1,       8,             \
1638                 TSP_F | OL3OL4CSUM_F | L3L4CSUM_F)                             \
1639 T(ts_vlan,                              1, 0, 0, 1, 0, 0,       8,             \
1640                 TSP_F | VLAN_F)                                                \
1641 T(ts_vlan_l3l4csum,                     1, 0, 0, 1, 0, 1,       8,             \
1642                 TSP_F | VLAN_F | L3L4CSUM_F)                                   \
1643 T(ts_vlan_ol3ol4csum,                   1, 0, 0, 1, 1, 0,       8,             \
1644                 TSP_F | VLAN_F | OL3OL4CSUM_F)                                 \
1645 T(ts_vlan_ol3ol4csum_l3l4csum,          1, 0, 0, 1, 1, 1,       8,             \
1646                 TSP_F | VLAN_F | OL3OL4CSUM_F | L3L4CSUM_F)                    \
1647 T(ts_noff,                              1, 0, 1, 0, 0, 0,       8,             \
1648                 TSP_F | NOFF_F)                                                \
1649 T(ts_noff_l3l4csum,                     1, 0, 1, 0, 0, 1,       8,             \
1650                 TSP_F | NOFF_F | L3L4CSUM_F)                                   \
1651 T(ts_noff_ol3ol4csum,                   1, 0, 1, 0, 1, 0,       8,             \
1652                 TSP_F | NOFF_F | OL3OL4CSUM_F)                                 \
1653 T(ts_noff_ol3ol4csum_l3l4csum,          1, 0, 1, 0, 1, 1,       8,             \
1654                 TSP_F | NOFF_F | OL3OL4CSUM_F | L3L4CSUM_F)                    \
1655 T(ts_noff_vlan,                         1, 0, 1, 1, 0, 0,       8,             \
1656                 TSP_F | NOFF_F | VLAN_F)                                       \
1657 T(ts_noff_vlan_l3l4csum,                1, 0, 1, 1, 0, 1,       8,             \
1658                 TSP_F | NOFF_F | VLAN_F | L3L4CSUM_F)                          \
1659 T(ts_noff_vlan_ol3ol4csum,              1, 0, 1, 1, 1, 0,       8,             \
1660                 TSP_F | NOFF_F | VLAN_F | OL3OL4CSUM_F)                        \
1661 T(ts_noff_vlan_ol3ol4csum_l3l4csum,     1, 0, 1, 1, 1, 1,       8,             \
1662                 TSP_F | NOFF_F | VLAN_F | OL3OL4CSUM_F | L3L4CSUM_F)           \
1663 T(ts_tso,                               1, 1, 0, 0, 0, 0,       8,             \
1664                 TSP_F | TSO_F)                                                 \
1665 T(ts_tso_l3l4csum,                      1, 1, 0, 0, 0, 1,       8,             \
1666                 TSP_F | TSO_F | L3L4CSUM_F)                                    \
1667 T(ts_tso_ol3ol4csum,                    1, 1, 0, 0, 1, 0,       8,             \
1668                 TSP_F | TSO_F | OL3OL4CSUM_F)                                  \
1669 T(ts_tso_ol3ol4csum_l3l4csum,           1, 1, 0, 0, 1, 1,       8,             \
1670                 TSP_F | TSO_F | OL3OL4CSUM_F | L3L4CSUM_F)                     \
1671 T(ts_tso_vlan,                          1, 1, 0, 1, 0, 0,       8,             \
1672                 TSP_F | TSO_F | VLAN_F)                                        \
1673 T(ts_tso_vlan_l3l4csum,                 1, 1, 0, 1, 0, 1,       8,             \
1674                 TSP_F | TSO_F | VLAN_F | L3L4CSUM_F)                           \
1675 T(ts_tso_vlan_ol3ol4csum,               1, 1, 0, 1, 1, 0,       8,             \
1676                 TSP_F | TSO_F | VLAN_F | OL3OL4CSUM_F)                         \
1677 T(ts_tso_vlan_ol3ol4csum_l3l4csum,      1, 1, 0, 1, 1, 1,       8,             \
1678                 TSP_F | TSO_F | VLAN_F | OL3OL4CSUM_F | L3L4CSUM_F)            \
1679 T(ts_tso_noff,                          1, 1, 1, 0, 0, 0,       8,             \
1680                 TSP_F | TSO_F | NOFF_F)                                        \
1681 T(ts_tso_noff_l3l4csum,                 1, 1, 1, 0, 0, 1,       8,             \
1682                 TSP_F | TSO_F | NOFF_F | L3L4CSUM_F)                           \
1683 T(ts_tso_noff_ol3ol4csum,               1, 1, 1, 0, 1, 0,       8,             \
1684                 TSP_F | TSO_F | NOFF_F | OL3OL4CSUM_F)                         \
1685 T(ts_tso_noff_ol3ol4csum_l3l4csum,      1, 1, 1, 0, 1, 1,       8,             \
1686                 TSP_F | TSO_F | NOFF_F | OL3OL4CSUM_F | L3L4CSUM_F)            \
1687 T(ts_tso_noff_vlan,                     1, 1, 1, 1, 0, 0,       8,             \
1688                 TSP_F | TSO_F | NOFF_F | VLAN_F)                               \
1689 T(ts_tso_noff_vlan_l3l4csum,            1, 1, 1, 1, 0, 1,       8,             \
1690                 TSP_F | TSO_F | NOFF_F | VLAN_F | L3L4CSUM_F)                  \
1691 T(ts_tso_noff_vlan_ol3ol4csum,          1, 1, 1, 1, 1, 0,       8,             \
1692                 TSP_F | TSO_F | NOFF_F | VLAN_F | OL3OL4CSUM_F)                \
1693 T(ts_tso_noff_vlan_ol3ol4csum_l3l4csum, 1, 1, 1, 1, 1, 1,       8,             \
1694                 TSP_F | TSO_F | NOFF_F | VLAN_F | OL3OL4CSUM_F | L3L4CSUM_F)
1695
1696 #define T(name, f5, f4, f3, f2, f1, f0, sz, flags)                             \
1697         uint16_t __rte_noinline __rte_hot cn9k_nix_xmit_pkts_##name(           \
1698                 void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t pkts);     \
1699                                                                                \
1700         uint16_t __rte_noinline __rte_hot cn9k_nix_xmit_pkts_mseg_##name(      \
1701                 void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t pkts);     \
1702                                                                                \
1703         uint16_t __rte_noinline __rte_hot cn9k_nix_xmit_pkts_vec_##name(       \
1704                 void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t pkts);
1705
1706 NIX_TX_FASTPATH_MODES
1707 #undef T
1708
1709 #endif /* __CN9K_TX_H__ */