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