drivers/net/thunderx/nicvf_rxtx.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2016 Cavium, Inc
   3  */
   4
   5 #include <unistd.h>
   6 #include <stdint.h>
   7 #include <stdio.h>
   8 #include <stdlib.h>
   9
  10 #include <rte_atomic.h>
  11 #include <rte_branch_prediction.h>
  12 #include <rte_byteorder.h>
  13 #include <rte_common.h>
  14 #include <rte_cycles.h>
  15 #include <rte_errno.h>
  16 #include <rte_ethdev_driver.h>
  17 #include <rte_ether.h>
  18 #include <rte_log.h>
  19 #include <rte_mbuf.h>
  20 #include <rte_prefetch.h>
  21
  22 #include "base/nicvf_plat.h"
  23
  24 #include "nicvf_ethdev.h"
  25 #include "nicvf_rxtx.h"
  26 #include "nicvf_logs.h"
  27
  28 static inline void __hot
  29 fill_sq_desc_header(union sq_entry_t *entry, struct rte_mbuf *pkt)
  30 {
  31         /* Local variable sqe to avoid read from sq desc memory*/
  32         union sq_entry_t sqe;
  33         uint64_t ol_flags;
  34
  35         /* Fill SQ header descriptor */
  36         sqe.buff[0] = 0;
  37         sqe.hdr.subdesc_type = SQ_DESC_TYPE_HEADER;
  38         /* Number of sub-descriptors following this one */
  39         sqe.hdr.subdesc_cnt = pkt->nb_segs;
  40         sqe.hdr.tot_len = pkt->pkt_len;
  41
  42         ol_flags = pkt->ol_flags & NICVF_TX_OFFLOAD_MASK;
  43         if (unlikely(ol_flags)) {
  44                 /* L4 cksum */
  45                 uint64_t l4_flags = ol_flags & PKT_TX_L4_MASK;
  46                 if (l4_flags == PKT_TX_TCP_CKSUM)
  47                         sqe.hdr.csum_l4 = SEND_L4_CSUM_TCP;
  48                 else if (l4_flags == PKT_TX_UDP_CKSUM)
  49                         sqe.hdr.csum_l4 = SEND_L4_CSUM_UDP;
  50                 else
  51                         sqe.hdr.csum_l4 = SEND_L4_CSUM_DISABLE;
  52
  53                 sqe.hdr.l3_offset = pkt->l2_len;
  54                 sqe.hdr.l4_offset = pkt->l3_len + pkt->l2_len;
  55
  56                 /* L3 cksum */
  57                 if (ol_flags & PKT_TX_IP_CKSUM)
  58                         sqe.hdr.csum_l3 = 1;
  59         }
  60
  61         entry->buff[0] = sqe.buff[0];
  62 }
  63
  64 static inline void __hot
  65 fill_sq_desc_header_zero_w1(union sq_entry_t *entry,
  66                                 struct rte_mbuf *pkt)
  67 {
  68         fill_sq_desc_header(entry, pkt);
  69         entry->buff[1] = 0ULL;
  70 }
  71
  72 void __hot
  73 nicvf_single_pool_free_xmited_buffers(struct nicvf_txq *sq)
  74 {
  75         int j = 0;
  76         uint32_t curr_head;
  77         uint32_t head = sq->head;
  78         struct rte_mbuf **txbuffs = sq->txbuffs;
  79         void *obj_p[NICVF_MAX_TX_FREE_THRESH] __rte_cache_aligned;
  80
  81         curr_head = nicvf_addr_read(sq->sq_head) >> 4;
  82         while (head != curr_head) {
  83                 if (txbuffs[head])
  84                         obj_p[j++] = txbuffs[head];
  85
  86                 head = (head + 1) & sq->qlen_mask;
  87         }
  88
  89         rte_mempool_put_bulk(sq->pool, obj_p, j);
  90         sq->head = curr_head;
  91         sq->xmit_bufs -= j;
  92         NICVF_TX_ASSERT(sq->xmit_bufs >= 0);
  93 }
  94
  95 void __hot
  96 nicvf_multi_pool_free_xmited_buffers(struct nicvf_txq *sq)
  97 {
  98         uint32_t n = 0;
  99         uint32_t curr_head;
 100         uint32_t head = sq->head;
 101         struct rte_mbuf **txbuffs = sq->txbuffs;
 102
 103         curr_head = nicvf_addr_read(sq->sq_head) >> 4;
 104         while (head != curr_head) {
 105                 if (txbuffs[head]) {
 106                         rte_pktmbuf_free_seg(txbuffs[head]);
 107                         n++;
 108                 }
 109
 110                 head = (head + 1) & sq->qlen_mask;
 111         }
 112
 113         sq->head = curr_head;
 114         sq->xmit_bufs -= n;
 115         NICVF_TX_ASSERT(sq->xmit_bufs >= 0);
 116 }
 117
 118 static inline uint32_t __hot
 119 nicvf_free_tx_desc(struct nicvf_txq *sq)
 120 {
 121         return ((sq->head - sq->tail - 1) & sq->qlen_mask);
 122 }
 123
 124 /* Send Header + Packet */
 125 #define TX_DESC_PER_PKT 2
 126
 127 static inline uint32_t __hot
 128 nicvf_free_xmitted_buffers(struct nicvf_txq *sq, struct rte_mbuf **tx_pkts,
 129                             uint16_t nb_pkts)
 130 {
 131         uint32_t free_desc = nicvf_free_tx_desc(sq);
 132
 133         if (free_desc < nb_pkts * TX_DESC_PER_PKT ||
 134                         sq->xmit_bufs > sq->tx_free_thresh) {
 135                 if (unlikely(sq->pool == NULL))
 136                         sq->pool = tx_pkts[0]->pool;
 137
 138                 sq->pool_free(sq);
 139                 /* Freed now, let see the number of free descs again */
 140                 free_desc = nicvf_free_tx_desc(sq);
 141         }
 142         return free_desc;
 143 }
 144
 145 uint16_t __hot
 146 nicvf_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 147 {
 148         int i;
 149         uint32_t free_desc;
 150         uint32_t tail;
 151         struct nicvf_txq *sq = tx_queue;
 152         union sq_entry_t *desc_ptr = sq->desc;
 153         struct rte_mbuf **txbuffs = sq->txbuffs;
 154         struct rte_mbuf *pkt;
 155         uint32_t qlen_mask = sq->qlen_mask;
 156
 157         tail = sq->tail;
 158         free_desc = nicvf_free_xmitted_buffers(sq, tx_pkts, nb_pkts);
 159
 160         for (i = 0; i < nb_pkts && (int)free_desc >= TX_DESC_PER_PKT; i++) {
 161                 pkt = tx_pkts[i];
 162
 163                 txbuffs[tail] = NULL;
 164                 fill_sq_desc_header(desc_ptr + tail, pkt);
 165                 tail = (tail + 1) & qlen_mask;
 166
 167                 txbuffs[tail] = pkt;
 168                 fill_sq_desc_gather(desc_ptr + tail, pkt);
 169                 tail = (tail + 1) & qlen_mask;
 170                 free_desc -= TX_DESC_PER_PKT;
 171         }
 172
 173         if (likely(i)) {
 174                 sq->tail = tail;
 175                 sq->xmit_bufs += i;
 176                 rte_wmb();
 177
 178                 /* Inform HW to xmit the packets */
 179                 nicvf_addr_write(sq->sq_door, i * TX_DESC_PER_PKT);
 180         }
 181         return i;
 182 }
 183
 184 uint16_t __hot
 185 nicvf_xmit_pkts_multiseg(void *tx_queue, struct rte_mbuf **tx_pkts,
 186                          uint16_t nb_pkts)
 187 {
 188         int i, k;
 189         uint32_t used_desc, next_used_desc, used_bufs, free_desc, tail;
 190         struct nicvf_txq *sq = tx_queue;
 191         union sq_entry_t *desc_ptr = sq->desc;
 192         struct rte_mbuf **txbuffs = sq->txbuffs;
 193         struct rte_mbuf *pkt, *seg;
 194         uint32_t qlen_mask = sq->qlen_mask;
 195         uint16_t nb_segs;
 196
 197         tail = sq->tail;
 198         used_desc = 0;
 199         used_bufs = 0;
 200
 201         free_desc = nicvf_free_xmitted_buffers(sq, tx_pkts, nb_pkts);
 202
 203         for (i = 0; i < nb_pkts; i++) {
 204                 pkt = tx_pkts[i];
 205
 206                 nb_segs = pkt->nb_segs;
 207
 208                 next_used_desc = used_desc + nb_segs + 1;
 209                 if (next_used_desc > free_desc)
 210                         break;
 211                 used_desc = next_used_desc;
 212                 used_bufs += nb_segs;
 213
 214                 txbuffs[tail] = NULL;
 215                 fill_sq_desc_header_zero_w1(desc_ptr + tail, pkt);
 216                 tail = (tail + 1) & qlen_mask;
 217
 218                 txbuffs[tail] = pkt;
 219                 fill_sq_desc_gather(desc_ptr + tail, pkt);
 220                 tail = (tail + 1) & qlen_mask;
 221
 222                 seg = pkt->next;
 223                 for (k = 1; k < nb_segs; k++) {
 224                         txbuffs[tail] = seg;
 225                         fill_sq_desc_gather(desc_ptr + tail, seg);
 226                         tail = (tail + 1) & qlen_mask;
 227                         seg = seg->next;
 228                 }
 229         }
 230
 231         if (likely(used_desc)) {
 232                 sq->tail = tail;
 233                 sq->xmit_bufs += used_bufs;
 234                 rte_wmb();
 235
 236                 /* Inform HW to xmit the packets */
 237                 nicvf_addr_write(sq->sq_door, used_desc);
 238         }
 239         return i;
 240 }
 241
 242 static const uint32_t ptype_table[16][16] __rte_cache_aligned = {
 243         [L3_NONE][L4_NONE] = RTE_PTYPE_UNKNOWN,
 244         [L3_NONE][L4_IPSEC_ESP] = RTE_PTYPE_UNKNOWN,
 245         [L3_NONE][L4_IPFRAG] = RTE_PTYPE_L4_FRAG,
 246         [L3_NONE][L4_IPCOMP] = RTE_PTYPE_UNKNOWN,
 247         [L3_NONE][L4_TCP] = RTE_PTYPE_L4_TCP,
 248         [L3_NONE][L4_UDP_PASS1] = RTE_PTYPE_L4_UDP,
 249         [L3_NONE][L4_GRE] = RTE_PTYPE_TUNNEL_GRE,
 250         [L3_NONE][L4_UDP_PASS2] = RTE_PTYPE_L4_UDP,
 251         [L3_NONE][L4_UDP_GENEVE] = RTE_PTYPE_TUNNEL_GENEVE,
 252         [L3_NONE][L4_UDP_VXLAN] = RTE_PTYPE_TUNNEL_VXLAN,
 253         [L3_NONE][L4_NVGRE] = RTE_PTYPE_TUNNEL_NVGRE,
 254
 255         [L3_IPV4][L4_NONE] = RTE_PTYPE_L3_IPV4 | RTE_PTYPE_UNKNOWN,
 256         [L3_IPV4][L4_IPSEC_ESP] = RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L3_IPV4,
 257         [L3_IPV4][L4_IPFRAG] = RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L4_FRAG,
 258         [L3_IPV4][L4_IPCOMP] = RTE_PTYPE_L3_IPV4 | RTE_PTYPE_UNKNOWN,
 259         [L3_IPV4][L4_TCP] = RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L4_TCP,
 260         [L3_IPV4][L4_UDP_PASS1] = RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L4_UDP,
 261         [L3_IPV4][L4_GRE] = RTE_PTYPE_L3_IPV4 | RTE_PTYPE_TUNNEL_GRE,
 262         [L3_IPV4][L4_UDP_PASS2] = RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L4_UDP,
 263         [L3_IPV4][L4_UDP_GENEVE] = RTE_PTYPE_L3_IPV4 | RTE_PTYPE_TUNNEL_GENEVE,
 264         [L3_IPV4][L4_UDP_VXLAN] = RTE_PTYPE_L3_IPV4 | RTE_PTYPE_TUNNEL_VXLAN,
 265         [L3_IPV4][L4_NVGRE] = RTE_PTYPE_L3_IPV4 | RTE_PTYPE_TUNNEL_NVGRE,
 266
 267         [L3_IPV4_OPT][L4_NONE] = RTE_PTYPE_L3_IPV4_EXT | RTE_PTYPE_UNKNOWN,
 268         [L3_IPV4_OPT][L4_IPSEC_ESP] =  RTE_PTYPE_L3_IPV4_EXT |
 269                                 RTE_PTYPE_L3_IPV4,
 270         [L3_IPV4_OPT][L4_IPFRAG] = RTE_PTYPE_L3_IPV4_EXT | RTE_PTYPE_L4_FRAG,
 271         [L3_IPV4_OPT][L4_IPCOMP] = RTE_PTYPE_L3_IPV4_EXT | RTE_PTYPE_UNKNOWN,
 272         [L3_IPV4_OPT][L4_TCP] = RTE_PTYPE_L3_IPV4_EXT | RTE_PTYPE_L4_TCP,
 273         [L3_IPV4_OPT][L4_UDP_PASS1] = RTE_PTYPE_L3_IPV4_EXT | RTE_PTYPE_L4_UDP,
 274         [L3_IPV4_OPT][L4_GRE] = RTE_PTYPE_L3_IPV4_EXT | RTE_PTYPE_TUNNEL_GRE,
 275         [L3_IPV4_OPT][L4_UDP_PASS2] = RTE_PTYPE_L3_IPV4_EXT | RTE_PTYPE_L4_UDP,
 276         [L3_IPV4_OPT][L4_UDP_GENEVE] = RTE_PTYPE_L3_IPV4_EXT |
 277                                 RTE_PTYPE_TUNNEL_GENEVE,
 278         [L3_IPV4_OPT][L4_UDP_VXLAN] = RTE_PTYPE_L3_IPV4_EXT |
 279                                 RTE_PTYPE_TUNNEL_VXLAN,
 280         [L3_IPV4_OPT][L4_NVGRE] = RTE_PTYPE_L3_IPV4_EXT |
 281                                 RTE_PTYPE_TUNNEL_NVGRE,
 282
 283         [L3_IPV6][L4_NONE] = RTE_PTYPE_L3_IPV6 | RTE_PTYPE_UNKNOWN,
 284         [L3_IPV6][L4_IPSEC_ESP] = RTE_PTYPE_L3_IPV6 | RTE_PTYPE_L3_IPV4,
 285         [L3_IPV6][L4_IPFRAG] = RTE_PTYPE_L3_IPV6 | RTE_PTYPE_L4_FRAG,
 286         [L3_IPV6][L4_IPCOMP] = RTE_PTYPE_L3_IPV6 | RTE_PTYPE_UNKNOWN,
 287         [L3_IPV6][L4_TCP] = RTE_PTYPE_L3_IPV6 | RTE_PTYPE_L4_TCP,
 288         [L3_IPV6][L4_UDP_PASS1] = RTE_PTYPE_L3_IPV6 | RTE_PTYPE_L4_UDP,
 289         [L3_IPV6][L4_GRE] = RTE_PTYPE_L3_IPV6 | RTE_PTYPE_TUNNEL_GRE,
 290         [L3_IPV6][L4_UDP_PASS2] = RTE_PTYPE_L3_IPV6 | RTE_PTYPE_L4_UDP,
 291         [L3_IPV6][L4_UDP_GENEVE] = RTE_PTYPE_L3_IPV6 | RTE_PTYPE_TUNNEL_GENEVE,
 292         [L3_IPV6][L4_UDP_VXLAN] = RTE_PTYPE_L3_IPV6 | RTE_PTYPE_TUNNEL_VXLAN,
 293         [L3_IPV6][L4_NVGRE] = RTE_PTYPE_L3_IPV6 | RTE_PTYPE_TUNNEL_NVGRE,
 294
 295         [L3_IPV6_OPT][L4_NONE] = RTE_PTYPE_L3_IPV6_EXT | RTE_PTYPE_UNKNOWN,
 296         [L3_IPV6_OPT][L4_IPSEC_ESP] =  RTE_PTYPE_L3_IPV6_EXT |
 297                                         RTE_PTYPE_L3_IPV4,
 298         [L3_IPV6_OPT][L4_IPFRAG] = RTE_PTYPE_L3_IPV6_EXT | RTE_PTYPE_L4_FRAG,
 299         [L3_IPV6_OPT][L4_IPCOMP] = RTE_PTYPE_L3_IPV6_EXT | RTE_PTYPE_UNKNOWN,
 300         [L3_IPV6_OPT][L4_TCP] = RTE_PTYPE_L3_IPV6_EXT | RTE_PTYPE_L4_TCP,
 301         [L3_IPV6_OPT][L4_UDP_PASS1] = RTE_PTYPE_L3_IPV6_EXT | RTE_PTYPE_L4_UDP,
 302         [L3_IPV6_OPT][L4_GRE] = RTE_PTYPE_L3_IPV6_EXT | RTE_PTYPE_TUNNEL_GRE,
 303         [L3_IPV6_OPT][L4_UDP_PASS2] = RTE_PTYPE_L3_IPV6_EXT | RTE_PTYPE_L4_UDP,
 304         [L3_IPV6_OPT][L4_UDP_GENEVE] = RTE_PTYPE_L3_IPV6_EXT |
 305                                         RTE_PTYPE_TUNNEL_GENEVE,
 306         [L3_IPV6_OPT][L4_UDP_VXLAN] = RTE_PTYPE_L3_IPV6_EXT |
 307                                         RTE_PTYPE_TUNNEL_VXLAN,
 308         [L3_IPV6_OPT][L4_NVGRE] = RTE_PTYPE_L3_IPV6_EXT |
 309                                         RTE_PTYPE_TUNNEL_NVGRE,
 310
 311         [L3_ET_STOP][L4_NONE] = RTE_PTYPE_UNKNOWN,
 312         [L3_ET_STOP][L4_IPSEC_ESP] = RTE_PTYPE_UNKNOWN,
 313         [L3_ET_STOP][L4_IPFRAG] = RTE_PTYPE_L4_FRAG,
 314         [L3_ET_STOP][L4_IPCOMP] = RTE_PTYPE_UNKNOWN,
 315         [L3_ET_STOP][L4_TCP] = RTE_PTYPE_L4_TCP,
 316         [L3_ET_STOP][L4_UDP_PASS1] = RTE_PTYPE_L4_UDP,
 317         [L3_ET_STOP][L4_GRE] = RTE_PTYPE_TUNNEL_GRE,
 318         [L3_ET_STOP][L4_UDP_PASS2] = RTE_PTYPE_L4_UDP,
 319         [L3_ET_STOP][L4_UDP_GENEVE] = RTE_PTYPE_TUNNEL_GENEVE,
 320         [L3_ET_STOP][L4_UDP_VXLAN] = RTE_PTYPE_TUNNEL_VXLAN,
 321         [L3_ET_STOP][L4_NVGRE] = RTE_PTYPE_TUNNEL_NVGRE,
 322
 323         [L3_OTHER][L4_NONE] = RTE_PTYPE_UNKNOWN,
 324         [L3_OTHER][L4_IPSEC_ESP] = RTE_PTYPE_UNKNOWN,
 325         [L3_OTHER][L4_IPFRAG] = RTE_PTYPE_L4_FRAG,
 326         [L3_OTHER][L4_IPCOMP] = RTE_PTYPE_UNKNOWN,
 327         [L3_OTHER][L4_TCP] = RTE_PTYPE_L4_TCP,
 328         [L3_OTHER][L4_UDP_PASS1] = RTE_PTYPE_L4_UDP,
 329         [L3_OTHER][L4_GRE] = RTE_PTYPE_TUNNEL_GRE,
 330         [L3_OTHER][L4_UDP_PASS2] = RTE_PTYPE_L4_UDP,
 331         [L3_OTHER][L4_UDP_GENEVE] = RTE_PTYPE_TUNNEL_GENEVE,
 332         [L3_OTHER][L4_UDP_VXLAN] = RTE_PTYPE_TUNNEL_VXLAN,
 333         [L3_OTHER][L4_NVGRE] = RTE_PTYPE_TUNNEL_NVGRE,
 334 };
 335
 336 static inline uint32_t __hot
 337 nicvf_rx_classify_pkt(cqe_rx_word0_t cqe_rx_w0)
 338 {
 339         return ptype_table[cqe_rx_w0.l3_type][cqe_rx_w0.l4_type];
 340 }
 341
 342 static inline uint64_t __hot
 343 nicvf_set_olflags(const cqe_rx_word0_t cqe_rx_w0)
 344 {
 345         static const uint64_t flag_table[3] __rte_cache_aligned = {
 346                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD,
 347                 PKT_RX_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_UNKNOWN,
 348                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD,
 349         };
 350
 351         const uint8_t idx = (cqe_rx_w0.err_opcode == CQE_RX_ERR_L4_CHK) << 1 |
 352                 (cqe_rx_w0.err_opcode == CQE_RX_ERR_IP_CHK);
 353         return flag_table[idx];
 354 }
 355
 356 static inline int __hot
 357 nicvf_fill_rbdr(struct nicvf_rxq *rxq, int to_fill)
 358 {
 359         int i;
 360         uint32_t ltail, next_tail;
 361         struct nicvf_rbdr *rbdr = rxq->shared_rbdr;
 362         uint64_t mbuf_phys_off = rxq->mbuf_phys_off;
 363         struct rbdr_entry_t *desc = rbdr->desc;
 364         uint32_t qlen_mask = rbdr->qlen_mask;
 365         uintptr_t door = rbdr->rbdr_door;
 366         void *obj_p[NICVF_MAX_RX_FREE_THRESH] __rte_cache_aligned;
 367
 368         if (unlikely(rte_mempool_get_bulk(rxq->pool, obj_p, to_fill) < 0)) {
 369                 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
 370                         to_fill;
 371                 return 0;
 372         }
 373
 374         NICVF_RX_ASSERT((unsigned int)to_fill <= (qlen_mask -
 375                 (nicvf_addr_read(rbdr->rbdr_status) & NICVF_RBDR_COUNT_MASK)));
 376
 377         next_tail = __atomic_fetch_add(&rbdr->next_tail, to_fill,
 378                                         __ATOMIC_ACQUIRE);
 379         ltail = next_tail;
 380         for (i = 0; i < to_fill; i++) {
 381                 struct rbdr_entry_t *entry = desc + (ltail & qlen_mask);
 382
 383                 entry->full_addr = nicvf_mbuff_virt2phy((uintptr_t)obj_p[i],
 384                                                         mbuf_phys_off);
 385                 ltail++;
 386         }
 387
 388         while (__atomic_load_n(&rbdr->tail, __ATOMIC_RELAXED) != next_tail)
 389                 rte_pause();
 390
 391         __atomic_store_n(&rbdr->tail, ltail, __ATOMIC_RELEASE);
 392         nicvf_addr_write(door, to_fill);
 393         return to_fill;
 394 }
 395
 396 static inline int32_t __hot
 397 nicvf_rx_pkts_to_process(struct nicvf_rxq *rxq, uint16_t nb_pkts,
 398                          int32_t available_space)
 399 {
 400         if (unlikely(available_space < nb_pkts))
 401                 rxq->available_space = nicvf_addr_read(rxq->cq_status)
 402                                                 & NICVF_CQ_CQE_COUNT_MASK;
 403
 404         return RTE_MIN(nb_pkts, available_space);
 405 }
 406
 407 static inline void __hot
 408 nicvf_rx_offload(cqe_rx_word0_t cqe_rx_w0, cqe_rx_word2_t cqe_rx_w2,
 409                  struct rte_mbuf *pkt)
 410 {
 411         if (likely(cqe_rx_w0.rss_alg)) {
 412                 pkt->hash.rss = cqe_rx_w2.rss_tag;
 413                 pkt->ol_flags |= PKT_RX_RSS_HASH;
 414
 415         }
 416 }
 417
 418 static __rte_always_inline uint16_t
 419 nicvf_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts,
 420                 const uint32_t flag)
 421 {
 422         uint32_t i, to_process;
 423         struct cqe_rx_t *cqe_rx;
 424         struct rte_mbuf *pkt;
 425         cqe_rx_word0_t cqe_rx_w0;
 426         cqe_rx_word1_t cqe_rx_w1;
 427         cqe_rx_word2_t cqe_rx_w2;
 428         cqe_rx_word3_t cqe_rx_w3;
 429         struct nicvf_rxq *rxq = rx_queue;
 430         union cq_entry_t *desc = rxq->desc;
 431         const uint64_t cqe_mask = rxq->qlen_mask;
 432         uint64_t rb0_ptr, mbuf_phys_off = rxq->mbuf_phys_off;
 433         const uint64_t mbuf_init = rxq->mbuf_initializer.value;
 434         uint32_t cqe_head = rxq->head & cqe_mask;
 435         int32_t available_space = rxq->available_space;
 436         const uint8_t rbptr_offset = rxq->rbptr_offset;
 437
 438         to_process = nicvf_rx_pkts_to_process(rxq, nb_pkts, available_space);
 439
 440         for (i = 0; i < to_process; i++) {
 441                 rte_prefetch_non_temporal(&desc[cqe_head + 2]);
 442                 cqe_rx = (struct cqe_rx_t *)&desc[cqe_head];
 443                 NICVF_RX_ASSERT(((struct cq_entry_type_t *)cqe_rx)->cqe_type
 444                                                  == CQE_TYPE_RX);
 445
 446                 NICVF_LOAD_PAIR(cqe_rx_w0.u64, cqe_rx_w1.u64, cqe_rx);
 447                 NICVF_LOAD_PAIR(cqe_rx_w2.u64, cqe_rx_w3.u64, &cqe_rx->word2);
 448                 rb0_ptr = *((uint64_t *)cqe_rx + rbptr_offset);
 449                 pkt = (struct rte_mbuf *)nicvf_mbuff_phy2virt
 450                                 (rb0_ptr - cqe_rx_w1.align_pad, mbuf_phys_off);
 451
 452                 if (flag & NICVF_RX_OFFLOAD_NONE)
 453                         pkt->ol_flags = 0;
 454                 if (flag & NICVF_RX_OFFLOAD_CKSUM)
 455                         pkt->ol_flags = nicvf_set_olflags(cqe_rx_w0);
 456                 if (flag & NICVF_RX_OFFLOAD_VLAN_STRIP) {
 457                         if (unlikely(cqe_rx_w0.vlan_stripped)) {
 458                                 pkt->ol_flags |= PKT_RX_VLAN
 459                                                         | PKT_RX_VLAN_STRIPPED;
 460                                 pkt->vlan_tci =
 461                                         rte_cpu_to_be_16(cqe_rx_w2.vlan_tci);
 462                         }
 463                 }
 464                 pkt->data_len = cqe_rx_w3.rb0_sz;
 465                 pkt->pkt_len = cqe_rx_w3.rb0_sz;
 466                 pkt->packet_type = nicvf_rx_classify_pkt(cqe_rx_w0);
 467                 nicvf_mbuff_init_update(pkt, mbuf_init, cqe_rx_w1.align_pad);
 468                 nicvf_rx_offload(cqe_rx_w0, cqe_rx_w2, pkt);
 469                 rx_pkts[i] = pkt;
 470                 cqe_head = (cqe_head + 1) & cqe_mask;
 471                 nicvf_prefetch_store_keep(pkt);
 472         }
 473
 474         if (likely(to_process)) {
 475                 rxq->available_space -= to_process;
 476                 rxq->head = cqe_head;
 477                 nicvf_addr_write(rxq->cq_door, to_process);
 478                 rxq->recv_buffers += to_process;
 479         }
 480         if (rxq->recv_buffers > rxq->rx_free_thresh) {
 481                 rxq->recv_buffers -= nicvf_fill_rbdr(rxq, rxq->rx_free_thresh);
 482                 NICVF_RX_ASSERT(rxq->recv_buffers >= 0);
 483         }
 484
 485         return to_process;
 486 }
 487
 488 uint16_t __hot
 489 nicvf_recv_pkts_no_offload(void *rx_queue, struct rte_mbuf **rx_pkts,
 490                 uint16_t nb_pkts)
 491 {
 492         return nicvf_recv_pkts(rx_queue, rx_pkts, nb_pkts,
 493                         NICVF_RX_OFFLOAD_NONE);
 494 }
 495
 496 uint16_t __hot
 497 nicvf_recv_pkts_cksum(void *rx_queue, struct rte_mbuf **rx_pkts,
 498                 uint16_t nb_pkts)
 499 {
 500         return nicvf_recv_pkts(rx_queue, rx_pkts, nb_pkts,
 501                         NICVF_RX_OFFLOAD_CKSUM);
 502 }
 503
 504 uint16_t __hot
 505 nicvf_recv_pkts_vlan_strip(void *rx_queue, struct rte_mbuf **rx_pkts,
 506                 uint16_t nb_pkts)
 507 {
 508         return nicvf_recv_pkts(rx_queue, rx_pkts, nb_pkts,
 509                         NICVF_RX_OFFLOAD_NONE | NICVF_RX_OFFLOAD_VLAN_STRIP);
 510 }
 511
 512 uint16_t __hot
 513 nicvf_recv_pkts_cksum_vlan_strip(void *rx_queue, struct rte_mbuf **rx_pkts,
 514                 uint16_t nb_pkts)
 515 {
 516         return nicvf_recv_pkts(rx_queue, rx_pkts, nb_pkts,
 517                         NICVF_RX_OFFLOAD_CKSUM | NICVF_RX_OFFLOAD_VLAN_STRIP);
 518 }
 519
 520 static __rte_always_inline uint16_t __hot
 521 nicvf_process_cq_mseg_entry(struct cqe_rx_t *cqe_rx,
 522                         uint64_t mbuf_phys_off,
 523                         struct rte_mbuf **rx_pkt, uint8_t rbptr_offset,
 524                         uint64_t mbuf_init, const uint32_t flag)
 525 {
 526         struct rte_mbuf *pkt, *seg, *prev;
 527         cqe_rx_word0_t cqe_rx_w0;
 528         cqe_rx_word1_t cqe_rx_w1;
 529         cqe_rx_word2_t cqe_rx_w2;
 530         uint16_t *rb_sz, nb_segs, seg_idx;
 531         uint64_t *rb_ptr;
 532
 533         NICVF_LOAD_PAIR(cqe_rx_w0.u64, cqe_rx_w1.u64, cqe_rx);
 534         NICVF_RX_ASSERT(cqe_rx_w0.cqe_type == CQE_TYPE_RX);
 535         cqe_rx_w2 = cqe_rx->word2;
 536         rb_sz = &cqe_rx->word3.rb0_sz;
 537         rb_ptr = (uint64_t *)cqe_rx + rbptr_offset;
 538         nb_segs = cqe_rx_w0.rb_cnt;
 539         pkt = (struct rte_mbuf *)nicvf_mbuff_phy2virt
 540                         (rb_ptr[0] - cqe_rx_w1.align_pad, mbuf_phys_off);
 541
 542         pkt->pkt_len = cqe_rx_w1.pkt_len;
 543         pkt->data_len = rb_sz[nicvf_frag_num(0)];
 544         nicvf_mbuff_init_mseg_update(
 545                                 pkt, mbuf_init, cqe_rx_w1.align_pad, nb_segs);
 546         pkt->packet_type = nicvf_rx_classify_pkt(cqe_rx_w0);
 547         if (flag & NICVF_RX_OFFLOAD_NONE)
 548                 pkt->ol_flags = 0;
 549         if (flag & NICVF_RX_OFFLOAD_CKSUM)
 550                 pkt->ol_flags = nicvf_set_olflags(cqe_rx_w0);
 551         if (flag & NICVF_RX_OFFLOAD_VLAN_STRIP) {
 552                 if (unlikely(cqe_rx_w0.vlan_stripped)) {
 553                         pkt->ol_flags |= PKT_RX_VLAN
 554                                 | PKT_RX_VLAN_STRIPPED;
 555                         pkt->vlan_tci = rte_cpu_to_be_16(cqe_rx_w2.vlan_tci);
 556                 }
 557         }
 558         nicvf_rx_offload(cqe_rx_w0, cqe_rx_w2, pkt);
 559
 560         *rx_pkt = pkt;
 561         prev = pkt;
 562         for (seg_idx = 1; seg_idx < nb_segs; seg_idx++) {
 563                 seg = (struct rte_mbuf *)nicvf_mbuff_phy2virt
 564                         (rb_ptr[seg_idx], mbuf_phys_off);
 565
 566                 prev->next = seg;
 567                 seg->data_len = rb_sz[nicvf_frag_num(seg_idx)];
 568                 nicvf_mbuff_init_update(seg, mbuf_init, 0);
 569
 570                 prev = seg;
 571         }
 572         prev->next = NULL;
 573         return nb_segs;
 574 }
 575
 576 static __rte_always_inline uint16_t __hot
 577 nicvf_recv_pkts_multiseg(void *rx_queue, struct rte_mbuf **rx_pkts,
 578                          uint16_t nb_pkts, const uint32_t flag)
 579 {
 580         union cq_entry_t *cq_entry;
 581         struct cqe_rx_t *cqe_rx;
 582         struct nicvf_rxq *rxq = rx_queue;
 583         union cq_entry_t *desc = rxq->desc;
 584         const uint64_t cqe_mask = rxq->qlen_mask;
 585         uint64_t mbuf_phys_off = rxq->mbuf_phys_off;
 586         uint32_t i, to_process, cqe_head, buffers_consumed = 0;
 587         int32_t available_space = rxq->available_space;
 588         uint16_t nb_segs;
 589         const uint64_t mbuf_init = rxq->mbuf_initializer.value;
 590         const uint8_t rbptr_offset = rxq->rbptr_offset;
 591
 592         cqe_head = rxq->head & cqe_mask;
 593         to_process = nicvf_rx_pkts_to_process(rxq, nb_pkts, available_space);
 594
 595         for (i = 0; i < to_process; i++) {
 596                 rte_prefetch_non_temporal(&desc[cqe_head + 2]);
 597                 cq_entry = &desc[cqe_head];
 598                 cqe_rx = (struct cqe_rx_t *)cq_entry;
 599                 nb_segs = nicvf_process_cq_mseg_entry(cqe_rx, mbuf_phys_off,
 600                         rx_pkts + i, rbptr_offset, mbuf_init, flag);
 601                 buffers_consumed += nb_segs;
 602                 cqe_head = (cqe_head + 1) & cqe_mask;
 603                 nicvf_prefetch_store_keep(rx_pkts[i]);
 604         }
 605
 606         if (likely(to_process)) {
 607                 rxq->available_space -= to_process;
 608                 rxq->head = cqe_head;
 609                 nicvf_addr_write(rxq->cq_door, to_process);
 610                 rxq->recv_buffers += buffers_consumed;
 611         }
 612         if (rxq->recv_buffers > rxq->rx_free_thresh) {
 613                 rxq->recv_buffers -= nicvf_fill_rbdr(rxq, rxq->rx_free_thresh);
 614                 NICVF_RX_ASSERT(rxq->recv_buffers >= 0);
 615         }
 616
 617         return to_process;
 618 }
 619
 620 uint16_t __hot
 621 nicvf_recv_pkts_multiseg_no_offload(void *rx_queue, struct rte_mbuf **rx_pkts,
 622                 uint16_t nb_pkts)
 623 {
 624         return nicvf_recv_pkts_multiseg(rx_queue, rx_pkts, nb_pkts,
 625                         NICVF_RX_OFFLOAD_NONE);
 626 }
 627
 628 uint16_t __hot
 629 nicvf_recv_pkts_multiseg_cksum(void *rx_queue, struct rte_mbuf **rx_pkts,
 630                 uint16_t nb_pkts)
 631 {
 632         return nicvf_recv_pkts_multiseg(rx_queue, rx_pkts, nb_pkts,
 633                         NICVF_RX_OFFLOAD_CKSUM);
 634 }
 635
 636 uint16_t __hot
 637 nicvf_recv_pkts_multiseg_vlan_strip(void *rx_queue, struct rte_mbuf **rx_pkts,
 638                 uint16_t nb_pkts)
 639 {
 640         return nicvf_recv_pkts_multiseg(rx_queue, rx_pkts, nb_pkts,
 641                         NICVF_RX_OFFLOAD_NONE | NICVF_RX_OFFLOAD_VLAN_STRIP);
 642 }
 643
 644 uint16_t __hot
 645 nicvf_recv_pkts_multiseg_cksum_vlan_strip(void *rx_queue,
 646                 struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 647 {
 648         return nicvf_recv_pkts_multiseg(rx_queue, rx_pkts, nb_pkts,
 649                         NICVF_RX_OFFLOAD_CKSUM | NICVF_RX_OFFLOAD_VLAN_STRIP);
 650 }
 651
 652 uint32_t
 653 nicvf_dev_rx_queue_count(struct rte_eth_dev *dev, uint16_t queue_idx)
 654 {
 655         struct nicvf_rxq *rxq;
 656
 657         rxq = dev->data->rx_queues[queue_idx];
 658         return nicvf_addr_read(rxq->cq_status) & NICVF_CQ_CQE_COUNT_MASK;
 659 }
 660
 661 uint32_t
 662 nicvf_dev_rbdr_refill(struct rte_eth_dev *dev, uint16_t queue_idx)
 663 {
 664         struct nicvf_rxq *rxq;
 665         uint32_t to_process;
 666         uint32_t rx_free;
 667
 668         rxq = dev->data->rx_queues[queue_idx];
 669         to_process = rxq->recv_buffers;
 670         while (rxq->recv_buffers > 0) {
 671                 rx_free = RTE_MIN(rxq->recv_buffers, NICVF_MAX_RX_FREE_THRESH);
 672                 rxq->recv_buffers -= nicvf_fill_rbdr(rxq, rx_free);
 673         }
 674
 675         assert(rxq->recv_buffers == 0);
 676         return to_process;
 677 }