drivers/net/cnxk/cn10k_rx.h

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(C) 2021 Marvell.
   3  */
   4 #ifndef __CN10K_RX_H__
   5 #define __CN10K_RX_H__
   6
   7 #include <rte_ether.h>
   8 #include <rte_vect.h>
   9
  10 #include <cnxk_ethdev.h>
  11
  12 #define NIX_RX_OFFLOAD_NONE          (0)
  13 #define NIX_RX_OFFLOAD_RSS_F         BIT(0)
  14 #define NIX_RX_OFFLOAD_PTYPE_F       BIT(1)
  15 #define NIX_RX_OFFLOAD_CHECKSUM_F    BIT(2)
  16 #define NIX_RX_OFFLOAD_MARK_UPDATE_F BIT(3)
  17 #define NIX_RX_OFFLOAD_TSTAMP_F      BIT(4)
  18 #define NIX_RX_OFFLOAD_VLAN_STRIP_F  BIT(5)
  19 #define NIX_RX_OFFLOAD_SECURITY_F    BIT(6)
  20 #define NIX_RX_OFFLOAD_MAX           (NIX_RX_OFFLOAD_SECURITY_F << 1)
  21
  22 /* Flags to control cqe_to_mbuf conversion function.
  23  * Defining it from backwards to denote its been
  24  * not used as offload flags to pick function
  25  */
  26 #define NIX_RX_VWQE_F      BIT(13)
  27 #define NIX_RX_MULTI_SEG_F BIT(14)
  28 #define CPT_RX_WQE_F       BIT(15)
  29
  30 #define CNXK_NIX_CQ_ENTRY_SZ 128
  31 #define NIX_DESCS_PER_LOOP   4
  32 #define CQE_CAST(x)          ((struct nix_cqe_hdr_s *)(x))
  33 #define CQE_SZ(x)            ((x) * CNXK_NIX_CQ_ENTRY_SZ)
  34
  35 #define CQE_PTR_OFF(b, i, o, f)                                                \
  36         (((f) & NIX_RX_VWQE_F) ?                                               \
  37                        (uint64_t *)(((uintptr_t)((uint64_t *)(b))[i]) + (o)) : \
  38                        (uint64_t *)(((uintptr_t)(b)) + CQE_SZ(i) + (o)))
  39
  40 union mbuf_initializer {
  41         struct {
  42                 uint16_t data_off;
  43                 uint16_t refcnt;
  44                 uint16_t nb_segs;
  45                 uint16_t port;
  46         } fields;
  47         uint64_t value;
  48 };
  49
  50 static __rte_always_inline uint64_t
  51 nix_clear_data_off(uint64_t oldval)
  52 {
  53         union mbuf_initializer mbuf_init = {.value = oldval};
  54
  55         mbuf_init.fields.data_off = 0;
  56         return mbuf_init.value;
  57 }
  58
  59 static __rte_always_inline struct rte_mbuf *
  60 nix_get_mbuf_from_cqe(void *cq, const uint64_t data_off)
  61 {
  62         rte_iova_t buff;
  63
  64         /* Skip CQE, NIX_RX_PARSE_S and SG HDR(9 DWORDs) and peek buff addr */
  65         buff = *((rte_iova_t *)((uint64_t *)cq + 9));
  66         return (struct rte_mbuf *)(buff - data_off);
  67 }
  68
  69 static __rte_always_inline void
  70 nix_sec_flush_meta_burst(uint16_t lmt_id, uint64_t data, uint16_t lnum,
  71                          uintptr_t aura_handle)
  72 {
  73         uint64_t pa;
  74
  75         /* Prepare PA and Data */
  76         pa = roc_npa_aura_handle_to_base(aura_handle) + NPA_LF_AURA_BATCH_FREE0;
  77         pa |= ((data & 0x7) << 4);
  78
  79         data >>= 3;
  80         data <<= 19;
  81         data |= (uint64_t)lmt_id;
  82         data |= (uint64_t)(lnum - 1) << 12;
  83
  84         roc_lmt_submit_steorl(data, pa);
  85 }
  86
  87 static __rte_always_inline void
  88 nix_sec_flush_meta(uintptr_t laddr, uint16_t lmt_id, uint8_t loff,
  89                    uintptr_t aura_handle)
  90 {
  91         uint64_t pa;
  92
  93         /* laddr is pointing to first pointer */
  94         laddr -= 8;
  95
  96         /* Trigger free either on lmtline full or different aura handle */
  97         pa = roc_npa_aura_handle_to_base(aura_handle) + NPA_LF_AURA_BATCH_FREE0;
  98
  99         /* Update aura handle */
 100         *(uint64_t *)laddr = (((uint64_t)(loff & 0x1) << 32) |
 101                               roc_npa_aura_handle_to_aura(aura_handle));
 102
 103         pa |= ((uint64_t)(loff >> 1) << 4);
 104         roc_lmt_submit_steorl(lmt_id, pa);
 105 }
 106
 107 static __rte_always_inline struct rte_mbuf *
 108 nix_sec_meta_to_mbuf_sc(uint64_t cq_w1, const uint64_t sa_base, uintptr_t laddr,
 109                         uint8_t *loff, struct rte_mbuf *mbuf, uint16_t data_off)
 110 {
 111         const void *__p = (void *)((uintptr_t)mbuf + (uint16_t)data_off);
 112         const struct cpt_parse_hdr_s *hdr = (const struct cpt_parse_hdr_s *)__p;
 113         struct cn10k_inb_priv_data *inb_priv;
 114         struct rte_mbuf *inner;
 115         uint32_t sa_idx;
 116         void *inb_sa;
 117         uint64_t w0;
 118
 119         if (cq_w1 & BIT(11)) {
 120                 inner = (struct rte_mbuf *)(rte_be_to_cpu_64(hdr->wqe_ptr) -
 121                                             sizeof(struct rte_mbuf));
 122
 123                 /* Get SPI from CPT_PARSE_S's cookie(already swapped) */
 124                 w0 = hdr->w0.u64;
 125                 sa_idx = w0 >> 32;
 126
 127                 inb_sa = roc_nix_inl_ot_ipsec_inb_sa(sa_base, sa_idx);
 128                 inb_priv = roc_nix_inl_ot_ipsec_inb_sa_sw_rsvd(inb_sa);
 129
 130                 /* Update dynamic field with userdata */
 131                 *rte_security_dynfield(inner) = (uint64_t)inb_priv->userdata;
 132
 133                 /* Update l2 hdr length first */
 134                 inner->pkt_len = (hdr->w2.il3_off -
 135                                   sizeof(struct cpt_parse_hdr_s) - (w0 & 0x7));
 136
 137                 /* Store meta in lmtline to free
 138                  * Assume all meta's from same aura.
 139                  */
 140                 *(uint64_t *)(laddr + (*loff << 3)) = (uint64_t)mbuf;
 141                 *loff = *loff + 1;
 142
 143                 /* Mark meta mbuf as put */
 144                 RTE_MEMPOOL_CHECK_COOKIES(mbuf->pool, (void **)&mbuf, 1, 0);
 145
 146                 /* Mark inner mbuf as get */
 147                 RTE_MEMPOOL_CHECK_COOKIES(inner->pool, (void **)&inner, 1, 1);
 148
 149                 return inner;
 150         }
 151         return mbuf;
 152 }
 153
 154 #if defined(RTE_ARCH_ARM64)
 155
 156 static __rte_always_inline struct rte_mbuf *
 157 nix_sec_meta_to_mbuf(uint64_t cq_w1, uintptr_t sa_base, uintptr_t laddr,
 158                      uint8_t *loff, struct rte_mbuf *mbuf, uint16_t data_off,
 159                      uint8x16_t *rx_desc_field1, uint64_t *ol_flags)
 160 {
 161         const void *__p = (void *)((uintptr_t)mbuf + (uint16_t)data_off);
 162         const struct cpt_parse_hdr_s *hdr = (const struct cpt_parse_hdr_s *)__p;
 163         struct cn10k_inb_priv_data *inb_priv;
 164         struct rte_mbuf *inner;
 165         uint64_t *sg, res_w1;
 166         uint32_t sa_idx;
 167         void *inb_sa;
 168         uint16_t len;
 169         uint64_t w0;
 170
 171         if (cq_w1 & BIT(11)) {
 172                 inner = (struct rte_mbuf *)(rte_be_to_cpu_64(hdr->wqe_ptr) -
 173                                             sizeof(struct rte_mbuf));
 174                 /* Get SPI from CPT_PARSE_S's cookie(already swapped) */
 175                 w0 = hdr->w0.u64;
 176                 sa_idx = w0 >> 32;
 177
 178                 inb_sa = roc_nix_inl_ot_ipsec_inb_sa(sa_base, sa_idx);
 179                 inb_priv = roc_nix_inl_ot_ipsec_inb_sa_sw_rsvd(inb_sa);
 180
 181                 /* Update dynamic field with userdata */
 182                 *rte_security_dynfield(inner) = (uint64_t)inb_priv->userdata;
 183
 184                 /* CPT result(struct cpt_cn10k_res_s) is at
 185                  * after first IOVA in meta
 186                  */
 187                 sg = (uint64_t *)(inner + 1);
 188                 res_w1 = sg[10];
 189
 190                 /* Clear checksum flags and update security flag */
 191                 *ol_flags &= ~(RTE_MBUF_F_RX_L4_CKSUM_MASK | RTE_MBUF_F_RX_IP_CKSUM_MASK);
 192                 *ol_flags |= (((res_w1 & 0xFF) == CPT_COMP_WARN) ?
 193                               RTE_MBUF_F_RX_SEC_OFFLOAD :
 194                               (RTE_MBUF_F_RX_SEC_OFFLOAD | RTE_MBUF_F_RX_SEC_OFFLOAD_FAILED));
 195                 /* Calculate inner packet length */
 196                 len = ((res_w1 >> 16) & 0xFFFF) + hdr->w2.il3_off -
 197                         sizeof(struct cpt_parse_hdr_s) - (w0 & 0x7);
 198                 /* Update pkt_len and data_len */
 199                 *rx_desc_field1 = vsetq_lane_u16(len, *rx_desc_field1, 2);
 200                 *rx_desc_field1 = vsetq_lane_u16(len, *rx_desc_field1, 4);
 201
 202                 /* Store meta in lmtline to free
 203                  * Assume all meta's from same aura.
 204                  */
 205                 *(uint64_t *)(laddr + (*loff << 3)) = (uint64_t)mbuf;
 206                 *loff = *loff + 1;
 207
 208                 /* Mark meta mbuf as put */
 209                 RTE_MEMPOOL_CHECK_COOKIES(mbuf->pool, (void **)&mbuf, 1, 0);
 210
 211                 /* Mark inner mbuf as get */
 212                 RTE_MEMPOOL_CHECK_COOKIES(inner->pool, (void **)&inner, 1, 1);
 213
 214                 /* Return inner mbuf */
 215                 return inner;
 216         }
 217
 218         /* Return same mbuf as it is not a decrypted pkt */
 219         return mbuf;
 220 }
 221 #endif
 222
 223 static __rte_always_inline uint32_t
 224 nix_ptype_get(const void *const lookup_mem, const uint64_t in)
 225 {
 226         const uint16_t *const ptype = lookup_mem;
 227         const uint16_t lh_lg_lf = (in & 0xFFF0000000000000) >> 52;
 228         const uint16_t tu_l2 = ptype[(in & 0x000FFFF000000000) >> 36];
 229         const uint16_t il4_tu = ptype[PTYPE_NON_TUNNEL_ARRAY_SZ + lh_lg_lf];
 230
 231         return (il4_tu << PTYPE_NON_TUNNEL_WIDTH) | tu_l2;
 232 }
 233
 234 static __rte_always_inline uint32_t
 235 nix_rx_olflags_get(const void *const lookup_mem, const uint64_t in)
 236 {
 237         const uint32_t *const ol_flags =
 238                 (const uint32_t *)((const uint8_t *)lookup_mem +
 239                                    PTYPE_ARRAY_SZ);
 240
 241         return ol_flags[(in & 0xfff00000) >> 20];
 242 }
 243
 244 static inline uint64_t
 245 nix_update_match_id(const uint16_t match_id, uint64_t ol_flags,
 246                     struct rte_mbuf *mbuf)
 247 {
 248         /* There is no separate bit to check match_id
 249          * is valid or not? and no flag to identify it is an
 250          * RTE_FLOW_ACTION_TYPE_FLAG vs RTE_FLOW_ACTION_TYPE_MARK
 251          * action. The former case addressed through 0 being invalid
 252          * value and inc/dec match_id pair when MARK is activated.
 253          * The later case addressed through defining
 254          * CNXK_FLOW_MARK_DEFAULT as value for
 255          * RTE_FLOW_ACTION_TYPE_MARK.
 256          * This would translate to not use
 257          * CNXK_FLOW_ACTION_FLAG_DEFAULT - 1 and
 258          * CNXK_FLOW_ACTION_FLAG_DEFAULT for match_id.
 259          * i.e valid mark_id's are from
 260          * 0 to CNXK_FLOW_ACTION_FLAG_DEFAULT - 2
 261          */
 262         if (likely(match_id)) {
 263                 ol_flags |= RTE_MBUF_F_RX_FDIR;
 264                 if (match_id != CNXK_FLOW_ACTION_FLAG_DEFAULT) {
 265                         ol_flags |= RTE_MBUF_F_RX_FDIR_ID;
 266                         mbuf->hash.fdir.hi = match_id - 1;
 267                 }
 268         }
 269
 270         return ol_flags;
 271 }
 272
 273 static __rte_always_inline void
 274 nix_cqe_xtract_mseg(const union nix_rx_parse_u *rx, struct rte_mbuf *mbuf,
 275                     uint64_t rearm, const uint16_t flags)
 276 {
 277         const rte_iova_t *iova_list;
 278         struct rte_mbuf *head;
 279         const rte_iova_t *eol;
 280         uint8_t nb_segs;
 281         uint64_t sg;
 282
 283         sg = *(const uint64_t *)(rx + 1);
 284         nb_segs = (sg >> 48) & 0x3;
 285
 286         if (nb_segs == 1) {
 287                 mbuf->next = NULL;
 288                 return;
 289         }
 290
 291         mbuf->pkt_len = (rx->pkt_lenm1 + 1) - (flags & NIX_RX_OFFLOAD_TSTAMP_F ?
 292                                                CNXK_NIX_TIMESYNC_RX_OFFSET : 0);
 293         mbuf->data_len = (sg & 0xFFFF) - (flags & NIX_RX_OFFLOAD_TSTAMP_F ?
 294                                           CNXK_NIX_TIMESYNC_RX_OFFSET : 0);
 295         mbuf->nb_segs = nb_segs;
 296         sg = sg >> 16;
 297
 298         eol = ((const rte_iova_t *)(rx + 1) + ((rx->desc_sizem1 + 1) << 1));
 299         /* Skip SG_S and first IOVA*/
 300         iova_list = ((const rte_iova_t *)(rx + 1)) + 2;
 301         nb_segs--;
 302
 303         rearm = rearm & ~0xFFFF;
 304
 305         head = mbuf;
 306         while (nb_segs) {
 307                 mbuf->next = ((struct rte_mbuf *)*iova_list) - 1;
 308                 mbuf = mbuf->next;
 309
 310                 RTE_MEMPOOL_CHECK_COOKIES(mbuf->pool, (void **)&mbuf, 1, 1);
 311
 312                 mbuf->data_len = sg & 0xFFFF;
 313                 sg = sg >> 16;
 314                 *(uint64_t *)(&mbuf->rearm_data) = rearm;
 315                 nb_segs--;
 316                 iova_list++;
 317
 318                 if (!nb_segs && (iova_list + 1 < eol)) {
 319                         sg = *(const uint64_t *)(iova_list);
 320                         nb_segs = (sg >> 48) & 0x3;
 321                         head->nb_segs += nb_segs;
 322                         iova_list = (const rte_iova_t *)(iova_list + 1);
 323                 }
 324         }
 325         mbuf->next = NULL;
 326 }
 327
 328 static __rte_always_inline void
 329 cn10k_nix_cqe_to_mbuf(const struct nix_cqe_hdr_s *cq, const uint32_t tag,
 330                       struct rte_mbuf *mbuf, const void *lookup_mem,
 331                       const uint64_t val, const uint16_t flag)
 332 {
 333         const union nix_rx_parse_u *rx =
 334                 (const union nix_rx_parse_u *)((const uint64_t *)cq + 1);
 335         const uint64_t w1 = *(const uint64_t *)rx;
 336         uint16_t len = rx->pkt_lenm1 + 1;
 337         uint64_t ol_flags = 0;
 338
 339         if (flag & NIX_RX_OFFLOAD_PTYPE_F)
 340                 mbuf->packet_type = nix_ptype_get(lookup_mem, w1);
 341         else
 342                 mbuf->packet_type = 0;
 343
 344         if (flag & NIX_RX_OFFLOAD_RSS_F) {
 345                 mbuf->hash.rss = tag;
 346                 ol_flags |= RTE_MBUF_F_RX_RSS_HASH;
 347         }
 348
 349         /* Process Security packets */
 350         if (flag & NIX_RX_OFFLOAD_SECURITY_F) {
 351                 if (w1 & BIT(11)) {
 352                         /* CPT result(struct cpt_cn10k_res_s) is at
 353                          * after first IOVA in meta
 354                          */
 355                         const uint64_t *sg = (const uint64_t *)(mbuf + 1);
 356                         const uint64_t res_w1 = sg[10];
 357                         const uint16_t uc_cc = res_w1 & 0xFF;
 358
 359                         /* Rlen */
 360                         len = ((res_w1 >> 16) & 0xFFFF) + mbuf->pkt_len;
 361                         ol_flags |= ((uc_cc == CPT_COMP_WARN) ?
 362                                                    RTE_MBUF_F_RX_SEC_OFFLOAD :
 363                                                    (RTE_MBUF_F_RX_SEC_OFFLOAD |
 364                                               RTE_MBUF_F_RX_SEC_OFFLOAD_FAILED));
 365                 } else {
 366                         if (flag & NIX_RX_OFFLOAD_CHECKSUM_F)
 367                                 ol_flags |= nix_rx_olflags_get(lookup_mem, w1);
 368                 }
 369         } else {
 370                 if (flag & NIX_RX_OFFLOAD_CHECKSUM_F)
 371                         ol_flags |= nix_rx_olflags_get(lookup_mem, w1);
 372         }
 373
 374         if (flag & NIX_RX_OFFLOAD_VLAN_STRIP_F) {
 375                 if (rx->vtag0_gone) {
 376                         ol_flags |= RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED;
 377                         mbuf->vlan_tci = rx->vtag0_tci;
 378                 }
 379                 if (rx->vtag1_gone) {
 380                         ol_flags |= RTE_MBUF_F_RX_QINQ | RTE_MBUF_F_RX_QINQ_STRIPPED;
 381                         mbuf->vlan_tci_outer = rx->vtag1_tci;
 382                 }
 383         }
 384
 385         if (flag & NIX_RX_OFFLOAD_MARK_UPDATE_F)
 386                 ol_flags = nix_update_match_id(rx->match_id, ol_flags, mbuf);
 387
 388         mbuf->ol_flags = ol_flags;
 389         mbuf->pkt_len = len;
 390         mbuf->data_len = len;
 391         *(uint64_t *)(&mbuf->rearm_data) = val;
 392
 393         if (flag & NIX_RX_MULTI_SEG_F)
 394                 /*
 395                  * For multi segment packets, mbuf length correction according
 396                  * to Rx timestamp length will be handled later during
 397                  * timestamp data process.
 398                  * Hence, flag argument is not required.
 399                  */
 400                 nix_cqe_xtract_mseg(rx, mbuf, val, 0);
 401         else
 402                 mbuf->next = NULL;
 403 }
 404
 405 static inline uint16_t
 406 nix_rx_nb_pkts(struct cn10k_eth_rxq *rxq, const uint64_t wdata,
 407                const uint16_t pkts, const uint32_t qmask)
 408 {
 409         uint32_t available = rxq->available;
 410
 411         /* Update the available count if cached value is not enough */
 412         if (unlikely(available < pkts)) {
 413                 uint64_t reg, head, tail;
 414
 415                 /* Use LDADDA version to avoid reorder */
 416                 reg = roc_atomic64_add_sync(wdata, rxq->cq_status);
 417                 /* CQ_OP_STATUS operation error */
 418                 if (reg & BIT_ULL(NIX_CQ_OP_STAT_OP_ERR) ||
 419                     reg & BIT_ULL(NIX_CQ_OP_STAT_CQ_ERR))
 420                         return 0;
 421
 422                 tail = reg & 0xFFFFF;
 423                 head = (reg >> 20) & 0xFFFFF;
 424                 if (tail < head)
 425                         available = tail - head + qmask + 1;
 426                 else
 427                         available = tail - head;
 428
 429                 rxq->available = available;
 430         }
 431
 432         return RTE_MIN(pkts, available);
 433 }
 434
 435 static __rte_always_inline uint16_t
 436 cn10k_nix_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t pkts,
 437                     const uint16_t flags)
 438 {
 439         struct cn10k_eth_rxq *rxq = rx_queue;
 440         const uint64_t mbuf_init = rxq->mbuf_initializer;
 441         const void *lookup_mem = rxq->lookup_mem;
 442         const uint64_t data_off = rxq->data_off;
 443         const uintptr_t desc = rxq->desc;
 444         const uint64_t wdata = rxq->wdata;
 445         const uint32_t qmask = rxq->qmask;
 446         uint64_t lbase = rxq->lmt_base;
 447         uint16_t packets = 0, nb_pkts;
 448         uint8_t loff = 0, lnum = 0;
 449         uint32_t head = rxq->head;
 450         struct nix_cqe_hdr_s *cq;
 451         struct rte_mbuf *mbuf;
 452         uint64_t aura_handle;
 453         uint64_t sa_base;
 454         uint16_t lmt_id;
 455         uint64_t laddr;
 456
 457         nb_pkts = nix_rx_nb_pkts(rxq, wdata, pkts, qmask);
 458
 459         if (flags & NIX_RX_OFFLOAD_SECURITY_F) {
 460                 aura_handle = rxq->aura_handle;
 461                 sa_base = rxq->sa_base;
 462                 sa_base &= ~(ROC_NIX_INL_SA_BASE_ALIGN - 1);
 463                 ROC_LMT_BASE_ID_GET(lbase, lmt_id);
 464                 laddr = lbase;
 465                 laddr += 8;
 466         }
 467
 468         while (packets < nb_pkts) {
 469                 /* Prefetch N desc ahead */
 470                 rte_prefetch_non_temporal(
 471                         (void *)(desc + (CQE_SZ((head + 2) & qmask))));
 472                 cq = (struct nix_cqe_hdr_s *)(desc + CQE_SZ(head));
 473
 474                 mbuf = nix_get_mbuf_from_cqe(cq, data_off);
 475
 476                 /* Mark mempool obj as "get" as it is alloc'ed by NIX */
 477                 RTE_MEMPOOL_CHECK_COOKIES(mbuf->pool, (void **)&mbuf, 1, 1);
 478
 479                 /* Translate meta to mbuf */
 480                 if (flags & NIX_RX_OFFLOAD_SECURITY_F) {
 481                         const uint64_t cq_w1 = *((const uint64_t *)cq + 1);
 482
 483                         mbuf = nix_sec_meta_to_mbuf_sc(cq_w1, sa_base, laddr,
 484                                                        &loff, mbuf, data_off);
 485                 }
 486
 487                 cn10k_nix_cqe_to_mbuf(cq, cq->tag, mbuf, lookup_mem, mbuf_init,
 488                                       flags);
 489                 cnxk_nix_mbuf_to_tstamp(mbuf, rxq->tstamp,
 490                                         (flags & NIX_RX_OFFLOAD_TSTAMP_F),
 491                                         (uint64_t *)((uint8_t *)mbuf
 492                                                                 + data_off));
 493                 rx_pkts[packets++] = mbuf;
 494                 roc_prefetch_store_keep(mbuf);
 495                 head++;
 496                 head &= qmask;
 497
 498                 if (flags & NIX_RX_OFFLOAD_SECURITY_F) {
 499                         /* Flush when we don't have space for 4 meta */
 500                         if ((15 - loff) < 1) {
 501                                 nix_sec_flush_meta(laddr, lmt_id + lnum, loff,
 502                                                    aura_handle);
 503                                 lnum++;
 504                                 lnum &= BIT_ULL(ROC_LMT_LINES_PER_CORE_LOG2) -
 505                                         1;
 506                                 /* First pointer starts at 8B offset */
 507                                 laddr = (uintptr_t)LMT_OFF(lbase, lnum, 8);
 508                                 loff = 0;
 509                         }
 510                 }
 511         }
 512
 513         rxq->head = head;
 514         rxq->available -= nb_pkts;
 515
 516         /* Free all the CQs that we've processed */
 517         plt_write64((wdata | nb_pkts), rxq->cq_door);
 518
 519         /* Free remaining meta buffers if any */
 520         if (flags & NIX_RX_OFFLOAD_SECURITY_F && loff) {
 521                 nix_sec_flush_meta(laddr, lmt_id + lnum, loff, aura_handle);
 522                 plt_io_wmb();
 523         }
 524
 525         return nb_pkts;
 526 }
 527
 528 #if defined(RTE_ARCH_ARM64)
 529
 530 static __rte_always_inline uint64_t
 531 nix_vlan_update(const uint64_t w2, uint64_t ol_flags, uint8x16_t *f)
 532 {
 533         if (w2 & BIT_ULL(21) /* vtag0_gone */) {
 534                 ol_flags |= RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED;
 535                 *f = vsetq_lane_u16((uint16_t)(w2 >> 32), *f, 5);
 536         }
 537
 538         return ol_flags;
 539 }
 540
 541 static __rte_always_inline uint64_t
 542 nix_qinq_update(const uint64_t w2, uint64_t ol_flags, struct rte_mbuf *mbuf)
 543 {
 544         if (w2 & BIT_ULL(23) /* vtag1_gone */) {
 545                 ol_flags |= RTE_MBUF_F_RX_QINQ | RTE_MBUF_F_RX_QINQ_STRIPPED;
 546                 mbuf->vlan_tci_outer = (uint16_t)(w2 >> 48);
 547         }
 548
 549         return ol_flags;
 550 }
 551
 552 static __rte_always_inline uint16_t
 553 cn10k_nix_recv_pkts_vector(void *args, struct rte_mbuf **mbufs, uint16_t pkts,
 554                            const uint16_t flags, void *lookup_mem,
 555                            struct cnxk_timesync_info *tstamp,
 556                            uintptr_t lmt_base)
 557 {
 558         struct cn10k_eth_rxq *rxq = args;
 559         const uint64_t mbuf_initializer = (flags & NIX_RX_VWQE_F) ?
 560                                                         *(uint64_t *)args :
 561                                                         rxq->mbuf_initializer;
 562         const uint64x2_t data_off = flags & NIX_RX_VWQE_F ?
 563                                                   vdupq_n_u64(0x80ULL) :
 564                                                   vdupq_n_u64(rxq->data_off);
 565         const uint32_t qmask = flags & NIX_RX_VWQE_F ? 0 : rxq->qmask;
 566         const uint64_t wdata = flags & NIX_RX_VWQE_F ? 0 : rxq->wdata;
 567         const uintptr_t desc = flags & NIX_RX_VWQE_F ? 0 : rxq->desc;
 568         uint64x2_t cq0_w8, cq1_w8, cq2_w8, cq3_w8, mbuf01, mbuf23;
 569         uint64_t ol_flags0, ol_flags1, ol_flags2, ol_flags3;
 570         uint64x2_t rearm0 = vdupq_n_u64(mbuf_initializer);
 571         uint64x2_t rearm1 = vdupq_n_u64(mbuf_initializer);
 572         uint64x2_t rearm2 = vdupq_n_u64(mbuf_initializer);
 573         uint64x2_t rearm3 = vdupq_n_u64(mbuf_initializer);
 574         struct rte_mbuf *mbuf0, *mbuf1, *mbuf2, *mbuf3;
 575         uint64_t aura_handle, lbase, laddr;
 576         uint8_t loff = 0, lnum = 0, shft = 0;
 577         uint8x16_t f0, f1, f2, f3;
 578         uint16_t lmt_id, d_off;
 579         uint16_t packets = 0;
 580         uint16_t pkts_left;
 581         uintptr_t sa_base;
 582         uint32_t head;
 583         uintptr_t cq0;
 584
 585         if (!(flags & NIX_RX_VWQE_F)) {
 586                 lookup_mem = rxq->lookup_mem;
 587                 head = rxq->head;
 588
 589                 pkts = nix_rx_nb_pkts(rxq, wdata, pkts, qmask);
 590                 pkts_left = pkts & (NIX_DESCS_PER_LOOP - 1);
 591                 /* Packets has to be floor-aligned to NIX_DESCS_PER_LOOP */
 592                 pkts = RTE_ALIGN_FLOOR(pkts, NIX_DESCS_PER_LOOP);
 593                 if (flags & NIX_RX_OFFLOAD_TSTAMP_F)
 594                         tstamp = rxq->tstamp;
 595         } else {
 596                 RTE_SET_USED(head);
 597         }
 598
 599         if (flags & NIX_RX_OFFLOAD_SECURITY_F) {
 600                 if (flags & NIX_RX_VWQE_F) {
 601                         uint16_t port;
 602
 603                         mbuf0 = (struct rte_mbuf *)((uintptr_t)mbufs[0] -
 604                                                     sizeof(struct rte_mbuf));
 605                         /* Pick first mbuf's aura handle assuming all
 606                          * mbufs are from a vec and are from same RQ.
 607                          */
 608                         aura_handle = mbuf0->pool->pool_id;
 609                         /* Calculate offset from mbuf to actual data area */
 610                         d_off = ((uintptr_t)mbuf0->buf_addr - (uintptr_t)mbuf0);
 611                         d_off += (mbuf_initializer & 0xFFFF);
 612
 613                         /* Get SA Base from lookup tbl using port_id */
 614                         port = mbuf_initializer >> 48;
 615                         sa_base = cnxk_nix_sa_base_get(port, lookup_mem);
 616
 617                         lbase = lmt_base;
 618                 } else {
 619                         aura_handle = rxq->aura_handle;
 620                         d_off = rxq->data_off;
 621                         sa_base = rxq->sa_base;
 622                         lbase = rxq->lmt_base;
 623                 }
 624                 sa_base &= ~(ROC_NIX_INL_SA_BASE_ALIGN - 1);
 625                 ROC_LMT_BASE_ID_GET(lbase, lmt_id);
 626                 lnum = 0;
 627                 laddr = lbase;
 628                 laddr += 8;
 629         }
 630
 631         while (packets < pkts) {
 632                 if (!(flags & NIX_RX_VWQE_F)) {
 633                         /* Exit loop if head is about to wrap and become
 634                          * unaligned.
 635                          */
 636                         if (((head + NIX_DESCS_PER_LOOP - 1) & qmask) <
 637                             NIX_DESCS_PER_LOOP) {
 638                                 pkts_left += (pkts - packets);
 639                                 break;
 640                         }
 641
 642                         cq0 = desc + CQE_SZ(head);
 643                 } else {
 644                         cq0 = (uintptr_t)&mbufs[packets];
 645                 }
 646
 647                 /* Prefetch N desc ahead */
 648                 rte_prefetch_non_temporal(CQE_PTR_OFF(cq0, 4, 64, flags));
 649                 rte_prefetch_non_temporal(CQE_PTR_OFF(cq0, 5, 64, flags));
 650                 rte_prefetch_non_temporal(CQE_PTR_OFF(cq0, 6, 64, flags));
 651                 rte_prefetch_non_temporal(CQE_PTR_OFF(cq0, 7, 64, flags));
 652
 653                 /* Get NIX_RX_SG_S for size and buffer pointer */
 654                 cq0_w8 = vld1q_u64(CQE_PTR_OFF(cq0, 0, 64, flags));
 655                 cq1_w8 = vld1q_u64(CQE_PTR_OFF(cq0, 1, 64, flags));
 656                 cq2_w8 = vld1q_u64(CQE_PTR_OFF(cq0, 2, 64, flags));
 657                 cq3_w8 = vld1q_u64(CQE_PTR_OFF(cq0, 3, 64, flags));
 658
 659                 if (!(flags & NIX_RX_VWQE_F)) {
 660                         /* Extract mbuf from NIX_RX_SG_S */
 661                         mbuf01 = vzip2q_u64(cq0_w8, cq1_w8);
 662                         mbuf23 = vzip2q_u64(cq2_w8, cq3_w8);
 663                         mbuf01 = vqsubq_u64(mbuf01, data_off);
 664                         mbuf23 = vqsubq_u64(mbuf23, data_off);
 665                 } else {
 666                         mbuf01 =
 667                                 vsubq_u64(vld1q_u64((uint64_t *)cq0), data_off);
 668                         mbuf23 = vsubq_u64(vld1q_u64((uint64_t *)(cq0 + 16)),
 669                                            data_off);
 670                 }
 671
 672                 /* Move mbufs to scalar registers for future use */
 673                 mbuf0 = (struct rte_mbuf *)vgetq_lane_u64(mbuf01, 0);
 674                 mbuf1 = (struct rte_mbuf *)vgetq_lane_u64(mbuf01, 1);
 675                 mbuf2 = (struct rte_mbuf *)vgetq_lane_u64(mbuf23, 0);
 676                 mbuf3 = (struct rte_mbuf *)vgetq_lane_u64(mbuf23, 1);
 677
 678                 /* Mask to get packet len from NIX_RX_SG_S */
 679                 const uint8x16_t shuf_msk = {
 680                         0xFF, 0xFF, /* pkt_type set as unknown */
 681                         0xFF, 0xFF, /* pkt_type set as unknown */
 682                         0,    1,    /* octet 1~0, low 16 bits pkt_len */
 683                         0xFF, 0xFF, /* skip high 16 bits pkt_len, zero out */
 684                         0,    1,    /* octet 1~0, 16 bits data_len */
 685                         0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF};
 686
 687                 /* Form the rx_descriptor_fields1 with pkt_len and data_len */
 688                 f0 = vqtbl1q_u8(cq0_w8, shuf_msk);
 689                 f1 = vqtbl1q_u8(cq1_w8, shuf_msk);
 690                 f2 = vqtbl1q_u8(cq2_w8, shuf_msk);
 691                 f3 = vqtbl1q_u8(cq3_w8, shuf_msk);
 692
 693                 if (flags & NIX_RX_OFFLOAD_SECURITY_F) {
 694                         /* Prefetch probable CPT parse header area */
 695                         rte_prefetch_non_temporal(RTE_PTR_ADD(mbuf0, d_off));
 696                         rte_prefetch_non_temporal(RTE_PTR_ADD(mbuf1, d_off));
 697                         rte_prefetch_non_temporal(RTE_PTR_ADD(mbuf2, d_off));
 698                         rte_prefetch_non_temporal(RTE_PTR_ADD(mbuf3, d_off));
 699                 }
 700
 701                 /* Load CQE word0 and word 1 */
 702                 const uint64_t cq0_w0 = *CQE_PTR_OFF(cq0, 0, 0, flags);
 703                 const uint64_t cq0_w1 = *CQE_PTR_OFF(cq0, 0, 8, flags);
 704                 const uint64_t cq1_w0 = *CQE_PTR_OFF(cq0, 1, 0, flags);
 705                 const uint64_t cq1_w1 = *CQE_PTR_OFF(cq0, 1, 8, flags);
 706                 const uint64_t cq2_w0 = *CQE_PTR_OFF(cq0, 2, 0, flags);
 707                 const uint64_t cq2_w1 = *CQE_PTR_OFF(cq0, 2, 8, flags);
 708                 const uint64_t cq3_w0 = *CQE_PTR_OFF(cq0, 3, 0, flags);
 709                 const uint64_t cq3_w1 = *CQE_PTR_OFF(cq0, 3, 8, flags);
 710
 711                 if (flags & NIX_RX_OFFLOAD_RSS_F) {
 712                         /* Fill rss in the rx_descriptor_fields1 */
 713                         f0 = vsetq_lane_u32(cq0_w0, f0, 3);
 714                         f1 = vsetq_lane_u32(cq1_w0, f1, 3);
 715                         f2 = vsetq_lane_u32(cq2_w0, f2, 3);
 716                         f3 = vsetq_lane_u32(cq3_w0, f3, 3);
 717                         ol_flags0 = RTE_MBUF_F_RX_RSS_HASH;
 718                         ol_flags1 = RTE_MBUF_F_RX_RSS_HASH;
 719                         ol_flags2 = RTE_MBUF_F_RX_RSS_HASH;
 720                         ol_flags3 = RTE_MBUF_F_RX_RSS_HASH;
 721                 } else {
 722                         ol_flags0 = 0;
 723                         ol_flags1 = 0;
 724                         ol_flags2 = 0;
 725                         ol_flags3 = 0;
 726                 }
 727
 728                 if (flags & NIX_RX_OFFLOAD_PTYPE_F) {
 729                         /* Fill packet_type in the rx_descriptor_fields1 */
 730                         f0 = vsetq_lane_u32(nix_ptype_get(lookup_mem, cq0_w1),
 731                                             f0, 0);
 732                         f1 = vsetq_lane_u32(nix_ptype_get(lookup_mem, cq1_w1),
 733                                             f1, 0);
 734                         f2 = vsetq_lane_u32(nix_ptype_get(lookup_mem, cq2_w1),
 735                                             f2, 0);
 736                         f3 = vsetq_lane_u32(nix_ptype_get(lookup_mem, cq3_w1),
 737                                             f3, 0);
 738                 }
 739
 740                 if (flags & NIX_RX_OFFLOAD_CHECKSUM_F) {
 741                         ol_flags0 |= nix_rx_olflags_get(lookup_mem, cq0_w1);
 742                         ol_flags1 |= nix_rx_olflags_get(lookup_mem, cq1_w1);
 743                         ol_flags2 |= nix_rx_olflags_get(lookup_mem, cq2_w1);
 744                         ol_flags3 |= nix_rx_olflags_get(lookup_mem, cq3_w1);
 745                 }
 746
 747                 /* Mark mempool obj as "get" as it is alloc'ed by NIX */
 748                 RTE_MEMPOOL_CHECK_COOKIES(mbuf0->pool, (void **)&mbuf0, 1, 1);
 749                 RTE_MEMPOOL_CHECK_COOKIES(mbuf1->pool, (void **)&mbuf1, 1, 1);
 750                 RTE_MEMPOOL_CHECK_COOKIES(mbuf2->pool, (void **)&mbuf2, 1, 1);
 751                 RTE_MEMPOOL_CHECK_COOKIES(mbuf3->pool, (void **)&mbuf3, 1, 1);
 752
 753                 /* Translate meta to mbuf */
 754                 if (flags & NIX_RX_OFFLOAD_SECURITY_F) {
 755                         /* Checksum ol_flags will be cleared if mbuf is meta */
 756                         mbuf0 = nix_sec_meta_to_mbuf(cq0_w1, sa_base, laddr,
 757                                                      &loff, mbuf0, d_off, &f0,
 758                                                      &ol_flags0);
 759                         mbuf01 = vsetq_lane_u64((uint64_t)mbuf0, mbuf01, 0);
 760
 761                         mbuf1 = nix_sec_meta_to_mbuf(cq1_w1, sa_base, laddr,
 762                                                      &loff, mbuf1, d_off, &f1,
 763                                                      &ol_flags1);
 764                         mbuf01 = vsetq_lane_u64((uint64_t)mbuf1, mbuf01, 1);
 765
 766                         mbuf2 = nix_sec_meta_to_mbuf(cq2_w1, sa_base, laddr,
 767                                                      &loff, mbuf2, d_off, &f2,
 768                                                      &ol_flags2);
 769                         mbuf23 = vsetq_lane_u64((uint64_t)mbuf2, mbuf23, 0);
 770
 771                         mbuf3 = nix_sec_meta_to_mbuf(cq3_w1, sa_base, laddr,
 772                                                      &loff, mbuf3, d_off, &f3,
 773                                                      &ol_flags3);
 774                         mbuf23 = vsetq_lane_u64((uint64_t)mbuf3, mbuf23, 1);
 775                 }
 776
 777                 if (flags & NIX_RX_OFFLOAD_VLAN_STRIP_F) {
 778                         uint64_t cq0_w2 = *(uint64_t *)(cq0 + CQE_SZ(0) + 16);
 779                         uint64_t cq1_w2 = *(uint64_t *)(cq0 + CQE_SZ(1) + 16);
 780                         uint64_t cq2_w2 = *(uint64_t *)(cq0 + CQE_SZ(2) + 16);
 781                         uint64_t cq3_w2 = *(uint64_t *)(cq0 + CQE_SZ(3) + 16);
 782
 783                         ol_flags0 = nix_vlan_update(cq0_w2, ol_flags0, &f0);
 784                         ol_flags1 = nix_vlan_update(cq1_w2, ol_flags1, &f1);
 785                         ol_flags2 = nix_vlan_update(cq2_w2, ol_flags2, &f2);
 786                         ol_flags3 = nix_vlan_update(cq3_w2, ol_flags3, &f3);
 787
 788                         ol_flags0 = nix_qinq_update(cq0_w2, ol_flags0, mbuf0);
 789                         ol_flags1 = nix_qinq_update(cq1_w2, ol_flags1, mbuf1);
 790                         ol_flags2 = nix_qinq_update(cq2_w2, ol_flags2, mbuf2);
 791                         ol_flags3 = nix_qinq_update(cq3_w2, ol_flags3, mbuf3);
 792                 }
 793
 794                 if (flags & NIX_RX_OFFLOAD_MARK_UPDATE_F) {
 795                         ol_flags0 = nix_update_match_id(
 796                                 *(uint16_t *)CQE_PTR_OFF(cq0, 0, 38, flags),
 797                                 ol_flags0, mbuf0);
 798                         ol_flags1 = nix_update_match_id(
 799                                 *(uint16_t *)CQE_PTR_OFF(cq0, 1, 38, flags),
 800                                 ol_flags1, mbuf1);
 801                         ol_flags2 = nix_update_match_id(
 802                                 *(uint16_t *)CQE_PTR_OFF(cq0, 2, 38, flags),
 803                                 ol_flags2, mbuf2);
 804                         ol_flags3 = nix_update_match_id(
 805                                 *(uint16_t *)CQE_PTR_OFF(cq0, 3, 38, flags),
 806                                 ol_flags3, mbuf3);
 807                 }
 808
 809                 if (flags & NIX_RX_OFFLOAD_TSTAMP_F) {
 810                         const uint16x8_t len_off = {
 811                                 0,                           /* ptype   0:15 */
 812                                 0,                           /* ptype  16:32 */
 813                                 CNXK_NIX_TIMESYNC_RX_OFFSET, /* pktlen  0:15*/
 814                                 0,                           /* pktlen 16:32 */
 815                                 CNXK_NIX_TIMESYNC_RX_OFFSET, /* datalen 0:15 */
 816                                 0,
 817                                 0,
 818                                 0};
 819                         const uint32x4_t ptype = {RTE_PTYPE_L2_ETHER_TIMESYNC,
 820                                                   RTE_PTYPE_L2_ETHER_TIMESYNC,
 821                                                   RTE_PTYPE_L2_ETHER_TIMESYNC,
 822                                                   RTE_PTYPE_L2_ETHER_TIMESYNC};
 823                         const uint64_t ts_olf = RTE_MBUF_F_RX_IEEE1588_PTP |
 824                                                 RTE_MBUF_F_RX_IEEE1588_TMST |
 825                                                 tstamp->rx_tstamp_dynflag;
 826                         const uint32x4_t and_mask = {0x1, 0x2, 0x4, 0x8};
 827                         uint64x2_t ts01, ts23, mask;
 828                         uint64_t ts[4];
 829                         uint8_t res;
 830
 831                         /* Subtract timesync length from total pkt length. */
 832                         f0 = vsubq_u16(f0, len_off);
 833                         f1 = vsubq_u16(f1, len_off);
 834                         f2 = vsubq_u16(f2, len_off);
 835                         f3 = vsubq_u16(f3, len_off);
 836
 837                         /* Get the address of actual timestamp. */
 838                         ts01 = vaddq_u64(mbuf01, data_off);
 839                         ts23 = vaddq_u64(mbuf23, data_off);
 840                         /* Load timestamp from address. */
 841                         ts01 = vsetq_lane_u64(*(uint64_t *)vgetq_lane_u64(ts01,
 842                                                                           0),
 843                                               ts01, 0);
 844                         ts01 = vsetq_lane_u64(*(uint64_t *)vgetq_lane_u64(ts01,
 845                                                                           1),
 846                                               ts01, 1);
 847                         ts23 = vsetq_lane_u64(*(uint64_t *)vgetq_lane_u64(ts23,
 848                                                                           0),
 849                                               ts23, 0);
 850                         ts23 = vsetq_lane_u64(*(uint64_t *)vgetq_lane_u64(ts23,
 851                                                                           1),
 852                                               ts23, 1);
 853                         /* Convert from be to cpu byteorder. */
 854                         ts01 = vrev64q_u8(ts01);
 855                         ts23 = vrev64q_u8(ts23);
 856                         /* Store timestamp into scalar for later use. */
 857                         ts[0] = vgetq_lane_u64(ts01, 0);
 858                         ts[1] = vgetq_lane_u64(ts01, 1);
 859                         ts[2] = vgetq_lane_u64(ts23, 0);
 860                         ts[3] = vgetq_lane_u64(ts23, 1);
 861
 862                         /* Store timestamp into dynfield. */
 863                         *cnxk_nix_timestamp_dynfield(mbuf0, tstamp) = ts[0];
 864                         *cnxk_nix_timestamp_dynfield(mbuf1, tstamp) = ts[1];
 865                         *cnxk_nix_timestamp_dynfield(mbuf2, tstamp) = ts[2];
 866                         *cnxk_nix_timestamp_dynfield(mbuf3, tstamp) = ts[3];
 867
 868                         /* Generate ptype mask to filter L2 ether timesync */
 869                         mask = vdupq_n_u32(vgetq_lane_u32(f0, 0));
 870                         mask = vsetq_lane_u32(vgetq_lane_u32(f1, 0), mask, 1);
 871                         mask = vsetq_lane_u32(vgetq_lane_u32(f2, 0), mask, 2);
 872                         mask = vsetq_lane_u32(vgetq_lane_u32(f3, 0), mask, 3);
 873
 874                         /* Match against L2 ether timesync. */
 875                         mask = vceqq_u32(mask, ptype);
 876                         /* Convert from vector from scalar mask */
 877                         res = vaddvq_u32(vandq_u32(mask, and_mask));
 878                         res &= 0xF;
 879
 880                         if (res) {
 881                                 /* Fill in the ol_flags for any packets that
 882                                  * matched.
 883                                  */
 884                                 ol_flags0 |= ((res & 0x1) ? ts_olf : 0);
 885                                 ol_flags1 |= ((res & 0x2) ? ts_olf : 0);
 886                                 ol_flags2 |= ((res & 0x4) ? ts_olf : 0);
 887                                 ol_flags3 |= ((res & 0x8) ? ts_olf : 0);
 888
 889                                 /* Update Rxq timestamp with the latest
 890                                  * timestamp.
 891                                  */
 892                                 tstamp->rx_ready = 1;
 893                                 tstamp->rx_tstamp = ts[31 - __builtin_clz(res)];
 894                         }
 895                 }
 896
 897                 /* Form rearm_data with ol_flags */
 898                 rearm0 = vsetq_lane_u64(ol_flags0, rearm0, 1);
 899                 rearm1 = vsetq_lane_u64(ol_flags1, rearm1, 1);
 900                 rearm2 = vsetq_lane_u64(ol_flags2, rearm2, 1);
 901                 rearm3 = vsetq_lane_u64(ol_flags3, rearm3, 1);
 902
 903                 /* Update rx_descriptor_fields1 */
 904                 vst1q_u64((uint64_t *)mbuf0->rx_descriptor_fields1, f0);
 905                 vst1q_u64((uint64_t *)mbuf1->rx_descriptor_fields1, f1);
 906                 vst1q_u64((uint64_t *)mbuf2->rx_descriptor_fields1, f2);
 907                 vst1q_u64((uint64_t *)mbuf3->rx_descriptor_fields1, f3);
 908
 909                 /* Update rearm_data */
 910                 vst1q_u64((uint64_t *)mbuf0->rearm_data, rearm0);
 911                 vst1q_u64((uint64_t *)mbuf1->rearm_data, rearm1);
 912                 vst1q_u64((uint64_t *)mbuf2->rearm_data, rearm2);
 913                 vst1q_u64((uint64_t *)mbuf3->rearm_data, rearm3);
 914
 915                 /* Store the mbufs to rx_pkts */
 916                 vst1q_u64((uint64_t *)&mbufs[packets], mbuf01);
 917                 vst1q_u64((uint64_t *)&mbufs[packets + 2], mbuf23);
 918
 919                 if (flags & NIX_RX_MULTI_SEG_F) {
 920                         /* Multi segment is enable build mseg list for
 921                          * individual mbufs in scalar mode.
 922                          */
 923                         nix_cqe_xtract_mseg((union nix_rx_parse_u *)
 924                                             (CQE_PTR_OFF(cq0, 0, 8, flags)),
 925                                             mbuf0, mbuf_initializer, flags);
 926                         nix_cqe_xtract_mseg((union nix_rx_parse_u *)
 927                                             (CQE_PTR_OFF(cq0, 1, 8, flags)),
 928                                             mbuf1, mbuf_initializer, flags);
 929                         nix_cqe_xtract_mseg((union nix_rx_parse_u *)
 930                                             (CQE_PTR_OFF(cq0, 2, 8, flags)),
 931                                             mbuf2, mbuf_initializer, flags);
 932                         nix_cqe_xtract_mseg((union nix_rx_parse_u *)
 933                                             (CQE_PTR_OFF(cq0, 3, 8, flags)),
 934                                             mbuf3, mbuf_initializer, flags);
 935                 } else {
 936                         /* Update that no more segments */
 937                         mbuf0->next = NULL;
 938                         mbuf1->next = NULL;
 939                         mbuf2->next = NULL;
 940                         mbuf3->next = NULL;
 941                 }
 942
 943                 /* Prefetch mbufs */
 944                 roc_prefetch_store_keep(mbuf0);
 945                 roc_prefetch_store_keep(mbuf1);
 946                 roc_prefetch_store_keep(mbuf2);
 947                 roc_prefetch_store_keep(mbuf3);
 948
 949                 packets += NIX_DESCS_PER_LOOP;
 950
 951                 if (!(flags & NIX_RX_VWQE_F)) {
 952                         /* Advance head pointer and packets */
 953                         head += NIX_DESCS_PER_LOOP;
 954                         head &= qmask;
 955                 }
 956
 957                 if (flags & NIX_RX_OFFLOAD_SECURITY_F) {
 958                         /* Check if lmtline border is crossed and adjust lnum */
 959                         if (loff > 15) {
 960                                 /* Update aura handle */
 961                                 *(uint64_t *)(laddr - 8) =
 962                                         (((uint64_t)(15 & 0x1) << 32) |
 963                                     roc_npa_aura_handle_to_aura(aura_handle));
 964                                 loff = loff - 15;
 965                                 shft += 3;
 966
 967                                 lnum++;
 968                                 laddr = (uintptr_t)LMT_OFF(lbase, lnum, 8);
 969                                 /* Pick the pointer from 16th index and put it
 970                                  * at end of this new line.
 971                                  */
 972                                 *(uint64_t *)(laddr + (loff << 3) - 8) =
 973                                         *(uint64_t *)(laddr - 8);
 974                         }
 975
 976                         /* Flush it when we are in 16th line and might
 977                          * overflow it
 978                          */
 979                         if (lnum >= 15 && loff >= 12) {
 980                                 /* 16 LMT Line size m1 */
 981                                 uint64_t data = BIT_ULL(48) - 1;
 982
 983                                 /* Update aura handle */
 984                                 *(uint64_t *)(laddr - 8) =
 985                                         (((uint64_t)(loff & 0x1) << 32) |
 986                                     roc_npa_aura_handle_to_aura(aura_handle));
 987
 988                                 data = (data & ~(0x7UL << shft)) |
 989                                        (((uint64_t)loff >> 1) << shft);
 990
 991                                 /* Send up to 16 lmt lines of pointers */
 992                                 nix_sec_flush_meta_burst(lmt_id, data, lnum + 1,
 993                                                          aura_handle);
 994                                 rte_io_wmb();
 995                                 lnum = 0;
 996                                 loff = 0;
 997                                 shft = 0;
 998                                 /* First pointer starts at 8B offset */
 999                                 laddr = (uintptr_t)LMT_OFF(lbase, lnum, 8);
1000                         }
1001                 }
1002         }
1003
1004         if (flags & NIX_RX_OFFLOAD_SECURITY_F && loff) {
1005                 /* 16 LMT Line size m1 */
1006                 uint64_t data = BIT_ULL(48) - 1;
1007
1008                 /* Update aura handle */
1009                 *(uint64_t *)(laddr - 8) =
1010                         (((uint64_t)(loff & 0x1) << 32) |
1011                          roc_npa_aura_handle_to_aura(aura_handle));
1012
1013                 data = (data & ~(0x7UL << shft)) |
1014                        (((uint64_t)loff >> 1) << shft);
1015
1016                 /* Send up to 16 lmt lines of pointers */
1017                 nix_sec_flush_meta_burst(lmt_id, data, lnum + 1, aura_handle);
1018                 if (flags & NIX_RX_VWQE_F)
1019                         plt_io_wmb();
1020         }
1021
1022         if (flags & NIX_RX_VWQE_F)
1023                 return packets;
1024
1025         rxq->head = head;
1026         rxq->available -= packets;
1027
1028         rte_io_wmb();
1029         /* Free all the CQs that we've processed */
1030         plt_write64((rxq->wdata | packets), rxq->cq_door);
1031
1032         if (unlikely(pkts_left))
1033                 packets += cn10k_nix_recv_pkts(args, &mbufs[packets], pkts_left,
1034                                                flags);
1035
1036         return packets;
1037 }
1038
1039 #else
1040
1041 static inline uint16_t
1042 cn10k_nix_recv_pkts_vector(void *args, struct rte_mbuf **mbufs, uint16_t pkts,
1043                            const uint16_t flags, void *lookup_mem,
1044                            struct cnxk_timesync_info *tstamp,
1045                            uintptr_t lmt_base)
1046 {
1047         RTE_SET_USED(args);
1048         RTE_SET_USED(mbufs);
1049         RTE_SET_USED(pkts);
1050         RTE_SET_USED(flags);
1051         RTE_SET_USED(lookup_mem);
1052         RTE_SET_USED(tstamp);
1053         RTE_SET_USED(lmt_base);
1054
1055         return 0;
1056 }
1057
1058 #endif
1059
1060
1061 #define RSS_F     NIX_RX_OFFLOAD_RSS_F
1062 #define PTYPE_F   NIX_RX_OFFLOAD_PTYPE_F
1063 #define CKSUM_F   NIX_RX_OFFLOAD_CHECKSUM_F
1064 #define MARK_F    NIX_RX_OFFLOAD_MARK_UPDATE_F
1065 #define TS_F      NIX_RX_OFFLOAD_TSTAMP_F
1066 #define RX_VLAN_F NIX_RX_OFFLOAD_VLAN_STRIP_F
1067 #define R_SEC_F   NIX_RX_OFFLOAD_SECURITY_F
1068
1069 /* [R_SEC_F] [RX_VLAN_F] [TS] [MARK] [CKSUM] [PTYPE] [RSS] */
1070 #define NIX_RX_FASTPATH_MODES_0_15                                             \
1071         R(no_offload, NIX_RX_OFFLOAD_NONE)                                     \
1072         R(rss, RSS_F)                                                          \
1073         R(ptype, PTYPE_F)                                                      \
1074         R(ptype_rss, PTYPE_F | RSS_F)                                          \
1075         R(cksum, CKSUM_F)                                                      \
1076         R(cksum_rss, CKSUM_F | RSS_F)                                          \
1077         R(cksum_ptype, CKSUM_F | PTYPE_F)                                      \
1078         R(cksum_ptype_rss, CKSUM_F | PTYPE_F | RSS_F)                          \
1079         R(mark, MARK_F)                                                        \
1080         R(mark_rss, MARK_F | RSS_F)                                            \
1081         R(mark_ptype, MARK_F | PTYPE_F)                                        \
1082         R(mark_ptype_rss, MARK_F | PTYPE_F | RSS_F)                            \
1083         R(mark_cksum, MARK_F | CKSUM_F)                                        \
1084         R(mark_cksum_rss, MARK_F | CKSUM_F | RSS_F)                            \
1085         R(mark_cksum_ptype, MARK_F | CKSUM_F | PTYPE_F)                        \
1086         R(mark_cksum_ptype_rss, MARK_F | CKSUM_F | PTYPE_F | RSS_F)
1087
1088 #define NIX_RX_FASTPATH_MODES_16_31                                            \
1089         R(ts, TS_F)                                                            \
1090         R(ts_rss, TS_F | RSS_F)                                                \
1091         R(ts_ptype, TS_F | PTYPE_F)                                            \
1092         R(ts_ptype_rss, TS_F | PTYPE_F | RSS_F)                                \
1093         R(ts_cksum, TS_F | CKSUM_F)                                            \
1094         R(ts_cksum_rss, TS_F | CKSUM_F | RSS_F)                                \
1095         R(ts_cksum_ptype, TS_F | CKSUM_F | PTYPE_F)                            \
1096         R(ts_cksum_ptype_rss, TS_F | CKSUM_F | PTYPE_F | RSS_F)                \
1097         R(ts_mark, TS_F | MARK_F)                                              \
1098         R(ts_mark_rss, TS_F | MARK_F | RSS_F)                                  \
1099         R(ts_mark_ptype, TS_F | MARK_F | PTYPE_F)                              \
1100         R(ts_mark_ptype_rss, TS_F | MARK_F | PTYPE_F | RSS_F)                  \
1101         R(ts_mark_cksum, TS_F | MARK_F | CKSUM_F)                              \
1102         R(ts_mark_cksum_rss, TS_F | MARK_F | CKSUM_F | RSS_F)                  \
1103         R(ts_mark_cksum_ptype, TS_F | MARK_F | CKSUM_F | PTYPE_F)              \
1104         R(ts_mark_cksum_ptype_rss, TS_F | MARK_F | CKSUM_F | PTYPE_F | RSS_F)
1105
1106 #define NIX_RX_FASTPATH_MODES_32_47                                            \
1107         R(vlan, RX_VLAN_F)                                                     \
1108         R(vlan_rss, RX_VLAN_F | RSS_F)                                         \
1109         R(vlan_ptype, RX_VLAN_F | PTYPE_F)                                     \
1110         R(vlan_ptype_rss, RX_VLAN_F | PTYPE_F | RSS_F)                         \
1111         R(vlan_cksum, RX_VLAN_F | CKSUM_F)                                     \
1112         R(vlan_cksum_rss, RX_VLAN_F | CKSUM_F | RSS_F)                         \
1113         R(vlan_cksum_ptype, RX_VLAN_F | CKSUM_F | PTYPE_F)                     \
1114         R(vlan_cksum_ptype_rss, RX_VLAN_F | CKSUM_F | PTYPE_F | RSS_F)         \
1115         R(vlan_mark, RX_VLAN_F | MARK_F)                                       \
1116         R(vlan_mark_rss, RX_VLAN_F | MARK_F | RSS_F)                           \
1117         R(vlan_mark_ptype, RX_VLAN_F | MARK_F | PTYPE_F)                       \
1118         R(vlan_mark_ptype_rss, RX_VLAN_F | MARK_F | PTYPE_F | RSS_F)           \
1119         R(vlan_mark_cksum, RX_VLAN_F | MARK_F | CKSUM_F)                       \
1120         R(vlan_mark_cksum_rss, RX_VLAN_F | MARK_F | CKSUM_F | RSS_F)           \
1121         R(vlan_mark_cksum_ptype, RX_VLAN_F | MARK_F | CKSUM_F | PTYPE_F)       \
1122         R(vlan_mark_cksum_ptype_rss,                                           \
1123           RX_VLAN_F | MARK_F | CKSUM_F | PTYPE_F | RSS_F)
1124
1125 #define NIX_RX_FASTPATH_MODES_48_63                                            \
1126         R(vlan_ts, RX_VLAN_F | TS_F)                                           \
1127         R(vlan_ts_rss, RX_VLAN_F | TS_F | RSS_F)                               \
1128         R(vlan_ts_ptype, RX_VLAN_F | TS_F | PTYPE_F)                           \
1129         R(vlan_ts_ptype_rss, RX_VLAN_F | TS_F | PTYPE_F | RSS_F)               \
1130         R(vlan_ts_cksum, RX_VLAN_F | TS_F | CKSUM_F)                           \
1131         R(vlan_ts_cksum_rss, RX_VLAN_F | TS_F | CKSUM_F | RSS_F)               \
1132         R(vlan_ts_cksum_ptype, RX_VLAN_F | TS_F | CKSUM_F | PTYPE_F)           \
1133         R(vlan_ts_cksum_ptype_rss,                                             \
1134           RX_VLAN_F | TS_F | CKSUM_F | PTYPE_F | RSS_F)                        \
1135         R(vlan_ts_mark, RX_VLAN_F | TS_F | MARK_F)                             \
1136         R(vlan_ts_mark_rss, RX_VLAN_F | TS_F | MARK_F | RSS_F)                 \
1137         R(vlan_ts_mark_ptype, RX_VLAN_F | TS_F | MARK_F | PTYPE_F)             \
1138         R(vlan_ts_mark_ptype_rss, RX_VLAN_F | TS_F | MARK_F | PTYPE_F | RSS_F) \
1139         R(vlan_ts_mark_cksum, RX_VLAN_F | TS_F | MARK_F | CKSUM_F)             \
1140         R(vlan_ts_mark_cksum_rss, RX_VLAN_F | TS_F | MARK_F | CKSUM_F | RSS_F) \
1141         R(vlan_ts_mark_cksum_ptype,                                            \
1142           RX_VLAN_F | TS_F | MARK_F | CKSUM_F | PTYPE_F)                       \
1143         R(vlan_ts_mark_cksum_ptype_rss,                                        \
1144           RX_VLAN_F | TS_F | MARK_F | CKSUM_F | PTYPE_F | RSS_F)
1145
1146 #define NIX_RX_FASTPATH_MODES_64_79                                            \
1147         R(sec, R_SEC_F)                                                        \
1148         R(sec_rss, R_SEC_F | RSS_F)                                            \
1149         R(sec_ptype, R_SEC_F | PTYPE_F)                                        \
1150         R(sec_ptype_rss, R_SEC_F | PTYPE_F | RSS_F)                            \
1151         R(sec_cksum, R_SEC_F | CKSUM_F)                                        \
1152         R(sec_cksum_rss, R_SEC_F | CKSUM_F | RSS_F)                            \
1153         R(sec_cksum_ptype, R_SEC_F | CKSUM_F | PTYPE_F)                        \
1154         R(sec_cksum_ptype_rss, R_SEC_F | CKSUM_F | PTYPE_F | RSS_F)            \
1155         R(sec_mark, R_SEC_F | MARK_F)                                          \
1156         R(sec_mark_rss, R_SEC_F | MARK_F | RSS_F)                              \
1157         R(sec_mark_ptype, R_SEC_F | MARK_F | PTYPE_F)                          \
1158         R(sec_mark_ptype_rss, R_SEC_F | MARK_F | PTYPE_F | RSS_F)              \
1159         R(sec_mark_cksum, R_SEC_F | MARK_F | CKSUM_F)                          \
1160         R(sec_mark_cksum_rss, R_SEC_F | MARK_F | CKSUM_F | RSS_F)              \
1161         R(sec_mark_cksum_ptype, R_SEC_F | MARK_F | CKSUM_F | PTYPE_F)          \
1162         R(sec_mark_cksum_ptype_rss,                                            \
1163           R_SEC_F | MARK_F | CKSUM_F | PTYPE_F | RSS_F)
1164
1165 #define NIX_RX_FASTPATH_MODES_80_95                                            \
1166         R(sec_ts, R_SEC_F | TS_F)                                              \
1167         R(sec_ts_rss, R_SEC_F | TS_F | RSS_F)                                  \
1168         R(sec_ts_ptype, R_SEC_F | TS_F | PTYPE_F)                              \
1169         R(sec_ts_ptype_rss, R_SEC_F | TS_F | PTYPE_F | RSS_F)                  \
1170         R(sec_ts_cksum, R_SEC_F | TS_F | CKSUM_F)                              \
1171         R(sec_ts_cksum_rss, R_SEC_F | TS_F | CKSUM_F | RSS_F)                  \
1172         R(sec_ts_cksum_ptype, R_SEC_F | TS_F | CKSUM_F | PTYPE_F)              \
1173         R(sec_ts_cksum_ptype_rss, R_SEC_F | TS_F | CKSUM_F | PTYPE_F | RSS_F)  \
1174         R(sec_ts_mark, R_SEC_F | TS_F | MARK_F)                                \
1175         R(sec_ts_mark_rss, R_SEC_F | TS_F | MARK_F | RSS_F)                    \
1176         R(sec_ts_mark_ptype, R_SEC_F | TS_F | MARK_F | PTYPE_F)                \
1177         R(sec_ts_mark_ptype_rss, R_SEC_F | TS_F | MARK_F | PTYPE_F | RSS_F)    \
1178         R(sec_ts_mark_cksum, R_SEC_F | TS_F | MARK_F | CKSUM_F)                \
1179         R(sec_ts_mark_cksum_rss, R_SEC_F | TS_F | MARK_F | CKSUM_F | RSS_F)    \
1180         R(sec_ts_mark_cksum_ptype,                                             \
1181           R_SEC_F | TS_F | MARK_F | CKSUM_F | PTYPE_F)                         \
1182         R(sec_ts_mark_cksum_ptype_rss,                                         \
1183           R_SEC_F | TS_F | MARK_F | CKSUM_F | PTYPE_F | RSS_F)
1184
1185 #define NIX_RX_FASTPATH_MODES_96_111                                           \
1186         R(sec_vlan, R_SEC_F | RX_VLAN_F)                                       \
1187         R(sec_vlan_rss, R_SEC_F | RX_VLAN_F | RSS_F)                           \
1188         R(sec_vlan_ptype, R_SEC_F | RX_VLAN_F | PTYPE_F)                       \
1189         R(sec_vlan_ptype_rss, R_SEC_F | RX_VLAN_F | PTYPE_F | RSS_F)           \
1190         R(sec_vlan_cksum, R_SEC_F | RX_VLAN_F | CKSUM_F)                       \
1191         R(sec_vlan_cksum_rss, R_SEC_F | RX_VLAN_F | CKSUM_F | RSS_F)           \
1192         R(sec_vlan_cksum_ptype, R_SEC_F | RX_VLAN_F | CKSUM_F | PTYPE_F)       \
1193         R(sec_vlan_cksum_ptype_rss,                                            \
1194           R_SEC_F | RX_VLAN_F | CKSUM_F | PTYPE_F | RSS_F)                     \
1195         R(sec_vlan_mark, R_SEC_F | RX_VLAN_F | MARK_F)                         \
1196         R(sec_vlan_mark_rss, R_SEC_F | RX_VLAN_F | MARK_F | RSS_F)             \
1197         R(sec_vlan_mark_ptype, R_SEC_F | RX_VLAN_F | MARK_F | PTYPE_F)         \
1198         R(sec_vlan_mark_ptype_rss,                                             \
1199           R_SEC_F | RX_VLAN_F | MARK_F | PTYPE_F | RSS_F)                      \
1200         R(sec_vlan_mark_cksum, R_SEC_F | RX_VLAN_F | MARK_F | CKSUM_F)         \
1201         R(sec_vlan_mark_cksum_rss,                                             \
1202           R_SEC_F | RX_VLAN_F | MARK_F | CKSUM_F | RSS_F)                      \
1203         R(sec_vlan_mark_cksum_ptype,                                           \
1204           R_SEC_F | RX_VLAN_F | MARK_F | CKSUM_F | PTYPE_F)                    \
1205         R(sec_vlan_mark_cksum_ptype_rss,                                       \
1206           R_SEC_F | RX_VLAN_F | MARK_F | CKSUM_F | PTYPE_F | RSS_F)
1207
1208 #define NIX_RX_FASTPATH_MODES_112_127                                          \
1209         R(sec_vlan_ts, R_SEC_F | RX_VLAN_F | TS_F)                             \
1210         R(sec_vlan_ts_rss, R_SEC_F | RX_VLAN_F | TS_F | RSS_F)                 \
1211         R(sec_vlan_ts_ptype, R_SEC_F | RX_VLAN_F | TS_F | PTYPE_F)             \
1212         R(sec_vlan_ts_ptype_rss, R_SEC_F | RX_VLAN_F | TS_F | PTYPE_F | RSS_F) \
1213         R(sec_vlan_ts_cksum, R_SEC_F | RX_VLAN_F | TS_F | CKSUM_F)             \
1214         R(sec_vlan_ts_cksum_rss, R_SEC_F | RX_VLAN_F | TS_F | CKSUM_F | RSS_F) \
1215         R(sec_vlan_ts_cksum_ptype,                                             \
1216           R_SEC_F | RX_VLAN_F | TS_F | CKSUM_F | PTYPE_F)                      \
1217         R(sec_vlan_ts_cksum_ptype_rss,                                         \
1218           R_SEC_F | RX_VLAN_F | TS_F | CKSUM_F | PTYPE_F | RSS_F)              \
1219         R(sec_vlan_ts_mark, R_SEC_F | RX_VLAN_F | TS_F | MARK_F)               \
1220         R(sec_vlan_ts_mark_rss, R_SEC_F | RX_VLAN_F | TS_F | MARK_F | RSS_F)   \
1221         R(sec_vlan_ts_mark_ptype,                                              \
1222           R_SEC_F | RX_VLAN_F | TS_F | MARK_F | PTYPE_F)                       \
1223         R(sec_vlan_ts_mark_ptype_rss,                                          \
1224           R_SEC_F | RX_VLAN_F | TS_F | MARK_F | PTYPE_F | RSS_F)               \
1225         R(sec_vlan_ts_mark_cksum,                                              \
1226           R_SEC_F | RX_VLAN_F | TS_F | MARK_F | CKSUM_F)                       \
1227         R(sec_vlan_ts_mark_cksum_rss,                                          \
1228           R_SEC_F | RX_VLAN_F | TS_F | MARK_F | CKSUM_F | RSS_F)               \
1229         R(sec_vlan_ts_mark_cksum_ptype,                                        \
1230           R_SEC_F | RX_VLAN_F | TS_F | MARK_F | CKSUM_F | PTYPE_F)             \
1231         R(sec_vlan_ts_mark_cksum_ptype_rss,                                    \
1232           R_SEC_F | RX_VLAN_F | TS_F | MARK_F | CKSUM_F | PTYPE_F | RSS_F)
1233
1234 #define NIX_RX_FASTPATH_MODES                                                  \
1235         NIX_RX_FASTPATH_MODES_0_15                                             \
1236         NIX_RX_FASTPATH_MODES_16_31                                            \
1237         NIX_RX_FASTPATH_MODES_32_47                                            \
1238         NIX_RX_FASTPATH_MODES_48_63                                            \
1239         NIX_RX_FASTPATH_MODES_64_79                                            \
1240         NIX_RX_FASTPATH_MODES_80_95                                            \
1241         NIX_RX_FASTPATH_MODES_96_111                                           \
1242         NIX_RX_FASTPATH_MODES_112_127
1243
1244 #define R(name, flags)                                                         \
1245         uint16_t __rte_noinline __rte_hot cn10k_nix_recv_pkts_##name(          \
1246                 void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t pkts);     \
1247         uint16_t __rte_noinline __rte_hot cn10k_nix_recv_pkts_mseg_##name(     \
1248                 void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t pkts);     \
1249         uint16_t __rte_noinline __rte_hot cn10k_nix_recv_pkts_vec_##name(      \
1250                 void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t pkts);     \
1251         uint16_t __rte_noinline __rte_hot cn10k_nix_recv_pkts_vec_mseg_##name( \
1252                 void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t pkts);
1253
1254 NIX_RX_FASTPATH_MODES
1255 #undef R
1256
1257 #define NIX_RX_RECV(fn, flags)                                                 \
1258         uint16_t __rte_noinline __rte_hot fn(                                  \
1259                 void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t pkts)      \
1260         {                                                                      \
1261                 return cn10k_nix_recv_pkts(rx_queue, rx_pkts, pkts, (flags));  \
1262         }
1263
1264 #define NIX_RX_RECV_MSEG(fn, flags) NIX_RX_RECV(fn, flags | NIX_RX_MULTI_SEG_F)
1265
1266 #define NIX_RX_RECV_VEC(fn, flags)                                             \
1267         uint16_t __rte_noinline __rte_hot fn(                                  \
1268                 void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t pkts)      \
1269         {                                                                      \
1270                 return cn10k_nix_recv_pkts_vector(rx_queue, rx_pkts, pkts,     \
1271                                                   (flags), NULL, NULL, 0);     \
1272         }
1273
1274 #define NIX_RX_RECV_VEC_MSEG(fn, flags)                                        \
1275         NIX_RX_RECV_VEC(fn, flags | NIX_RX_MULTI_SEG_F)
1276
1277 #endif /* __CN10K_RX_H__ */