drivers/net/virtio/virtio_rxtx.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2014 Intel Corporation
   3  */
   4
   5 #include <stdint.h>
   6 #include <stdio.h>
   7 #include <stdlib.h>
   8 #include <string.h>
   9 #include <errno.h>
  10
  11 #include <rte_cycles.h>
  12 #include <rte_memory.h>
  13 #include <rte_branch_prediction.h>
  14 #include <rte_mempool.h>
  15 #include <rte_malloc.h>
  16 #include <rte_mbuf.h>
  17 #include <rte_ether.h>
  18 #include <ethdev_driver.h>
  19 #include <rte_prefetch.h>
  20 #include <rte_string_fns.h>
  21 #include <rte_errno.h>
  22 #include <rte_byteorder.h>
  23 #include <rte_net.h>
  24 #include <rte_ip.h>
  25 #include <rte_udp.h>
  26 #include <rte_tcp.h>
  27
  28 #include "virtio_logs.h"
  29 #include "virtio_ethdev.h"
  30 #include "virtio.h"
  31 #include "virtqueue.h"
  32 #include "virtio_rxtx.h"
  33 #include "virtio_rxtx_simple.h"
  34 #include "virtio_ring.h"
  35
  36 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
  37 #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
  38 #else
  39 #define  VIRTIO_DUMP_PACKET(m, len) do { } while (0)
  40 #endif
  41
  42 int
  43 virtio_dev_rx_queue_done(void *rxq, uint16_t offset)
  44 {
  45         struct virtnet_rx *rxvq = rxq;
  46         struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
  47
  48         return virtqueue_nused(vq) >= offset;
  49 }
  50
  51 void
  52 vq_ring_free_inorder(struct virtqueue *vq, uint16_t desc_idx, uint16_t num)
  53 {
  54         vq->vq_free_cnt += num;
  55         vq->vq_desc_tail_idx = desc_idx & (vq->vq_nentries - 1);
  56 }
  57
  58 void
  59 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
  60 {
  61         struct vring_desc *dp, *dp_tail;
  62         struct vq_desc_extra *dxp;
  63         uint16_t desc_idx_last = desc_idx;
  64
  65         dp  = &vq->vq_split.ring.desc[desc_idx];
  66         dxp = &vq->vq_descx[desc_idx];
  67         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
  68         if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
  69                 while (dp->flags & VRING_DESC_F_NEXT) {
  70                         desc_idx_last = dp->next;
  71                         dp = &vq->vq_split.ring.desc[dp->next];
  72                 }
  73         }
  74         dxp->ndescs = 0;
  75
  76         /*
  77          * We must append the existing free chain, if any, to the end of
  78          * newly freed chain. If the virtqueue was completely used, then
  79          * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
  80          */
  81         if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
  82                 vq->vq_desc_head_idx = desc_idx;
  83         } else {
  84                 dp_tail = &vq->vq_split.ring.desc[vq->vq_desc_tail_idx];
  85                 dp_tail->next = desc_idx;
  86         }
  87
  88         vq->vq_desc_tail_idx = desc_idx_last;
  89         dp->next = VQ_RING_DESC_CHAIN_END;
  90 }
  91
  92 void
  93 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
  94 {
  95         uint32_t s = mbuf->pkt_len;
  96         struct rte_ether_addr *ea;
  97
  98         stats->bytes += s;
  99
 100         if (s == 64) {
 101                 stats->size_bins[1]++;
 102         } else if (s > 64 && s < 1024) {
 103                 uint32_t bin;
 104
 105                 /* count zeros, and offset into correct bin */
 106                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
 107                 stats->size_bins[bin]++;
 108         } else {
 109                 if (s < 64)
 110                         stats->size_bins[0]++;
 111                 else if (s < 1519)
 112                         stats->size_bins[6]++;
 113                 else
 114                         stats->size_bins[7]++;
 115         }
 116
 117         ea = rte_pktmbuf_mtod(mbuf, struct rte_ether_addr *);
 118         if (rte_is_multicast_ether_addr(ea)) {
 119                 if (rte_is_broadcast_ether_addr(ea))
 120                         stats->broadcast++;
 121                 else
 122                         stats->multicast++;
 123         }
 124 }
 125
 126 static inline void
 127 virtio_rx_stats_updated(struct virtnet_rx *rxvq, struct rte_mbuf *m)
 128 {
 129         VIRTIO_DUMP_PACKET(m, m->data_len);
 130
 131         virtio_update_packet_stats(&rxvq->stats, m);
 132 }
 133
 134 static uint16_t
 135 virtqueue_dequeue_burst_rx_packed(struct virtqueue *vq,
 136                                   struct rte_mbuf **rx_pkts,
 137                                   uint32_t *len,
 138                                   uint16_t num)
 139 {
 140         struct rte_mbuf *cookie;
 141         uint16_t used_idx;
 142         uint16_t id;
 143         struct vring_packed_desc *desc;
 144         uint16_t i;
 145
 146         desc = vq->vq_packed.ring.desc;
 147
 148         for (i = 0; i < num; i++) {
 149                 used_idx = vq->vq_used_cons_idx;
 150                 /* desc_is_used has a load-acquire or rte_io_rmb inside
 151                  * and wait for used desc in virtqueue.
 152                  */
 153                 if (!desc_is_used(&desc[used_idx], vq))
 154                         return i;
 155                 len[i] = desc[used_idx].len;
 156                 id = desc[used_idx].id;
 157                 cookie = (struct rte_mbuf *)vq->vq_descx[id].cookie;
 158                 if (unlikely(cookie == NULL)) {
 159                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
 160                                 vq->vq_used_cons_idx);
 161                         break;
 162                 }
 163                 rte_prefetch0(cookie);
 164                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 165                 rx_pkts[i] = cookie;
 166
 167                 vq->vq_free_cnt++;
 168                 vq->vq_used_cons_idx++;
 169                 if (vq->vq_used_cons_idx >= vq->vq_nentries) {
 170                         vq->vq_used_cons_idx -= vq->vq_nentries;
 171                         vq->vq_packed.used_wrap_counter ^= 1;
 172                 }
 173         }
 174
 175         return i;
 176 }
 177
 178 static uint16_t
 179 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
 180                            uint32_t *len, uint16_t num)
 181 {
 182         struct vring_used_elem *uep;
 183         struct rte_mbuf *cookie;
 184         uint16_t used_idx, desc_idx;
 185         uint16_t i;
 186
 187         /*  Caller does the check */
 188         for (i = 0; i < num ; i++) {
 189                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 190                 uep = &vq->vq_split.ring.used->ring[used_idx];
 191                 desc_idx = (uint16_t) uep->id;
 192                 len[i] = uep->len;
 193                 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
 194
 195                 if (unlikely(cookie == NULL)) {
 196                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
 197                                 vq->vq_used_cons_idx);
 198                         break;
 199                 }
 200
 201                 rte_prefetch0(cookie);
 202                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 203                 rx_pkts[i]  = cookie;
 204                 vq->vq_used_cons_idx++;
 205                 vq_ring_free_chain(vq, desc_idx);
 206                 vq->vq_descx[desc_idx].cookie = NULL;
 207         }
 208
 209         return i;
 210 }
 211
 212 static uint16_t
 213 virtqueue_dequeue_rx_inorder(struct virtqueue *vq,
 214                         struct rte_mbuf **rx_pkts,
 215                         uint32_t *len,
 216                         uint16_t num)
 217 {
 218         struct vring_used_elem *uep;
 219         struct rte_mbuf *cookie;
 220         uint16_t used_idx = 0;
 221         uint16_t i;
 222
 223         if (unlikely(num == 0))
 224                 return 0;
 225
 226         for (i = 0; i < num; i++) {
 227                 used_idx = vq->vq_used_cons_idx & (vq->vq_nentries - 1);
 228                 /* Desc idx same as used idx */
 229                 uep = &vq->vq_split.ring.used->ring[used_idx];
 230                 len[i] = uep->len;
 231                 cookie = (struct rte_mbuf *)vq->vq_descx[used_idx].cookie;
 232
 233                 if (unlikely(cookie == NULL)) {
 234                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
 235                                 vq->vq_used_cons_idx);
 236                         break;
 237                 }
 238
 239                 rte_prefetch0(cookie);
 240                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 241                 rx_pkts[i]  = cookie;
 242                 vq->vq_used_cons_idx++;
 243                 vq->vq_descx[used_idx].cookie = NULL;
 244         }
 245
 246         vq_ring_free_inorder(vq, used_idx, i);
 247         return i;
 248 }
 249
 250 static inline int
 251 virtqueue_enqueue_refill_inorder(struct virtqueue *vq,
 252                         struct rte_mbuf **cookies,
 253                         uint16_t num)
 254 {
 255         struct vq_desc_extra *dxp;
 256         struct virtio_hw *hw = vq->hw;
 257         struct vring_desc *start_dp;
 258         uint16_t head_idx, idx, i = 0;
 259
 260         if (unlikely(vq->vq_free_cnt == 0))
 261                 return -ENOSPC;
 262         if (unlikely(vq->vq_free_cnt < num))
 263                 return -EMSGSIZE;
 264
 265         head_idx = vq->vq_desc_head_idx & (vq->vq_nentries - 1);
 266         start_dp = vq->vq_split.ring.desc;
 267
 268         while (i < num) {
 269                 idx = head_idx & (vq->vq_nentries - 1);
 270                 dxp = &vq->vq_descx[idx];
 271                 dxp->cookie = (void *)cookies[i];
 272                 dxp->ndescs = 1;
 273
 274                 start_dp[idx].addr = cookies[i]->buf_iova +
 275                         RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 276                 start_dp[idx].len = cookies[i]->buf_len -
 277                         RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
 278                 start_dp[idx].flags =  VRING_DESC_F_WRITE;
 279
 280                 vq_update_avail_ring(vq, idx);
 281                 head_idx++;
 282                 i++;
 283         }
 284
 285         vq->vq_desc_head_idx += num;
 286         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 287         return 0;
 288 }
 289
 290 static inline int
 291 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf **cookie,
 292                                 uint16_t num)
 293 {
 294         struct vq_desc_extra *dxp;
 295         struct virtio_hw *hw = vq->hw;
 296         struct vring_desc *start_dp = vq->vq_split.ring.desc;
 297         uint16_t idx, i;
 298
 299         if (unlikely(vq->vq_free_cnt == 0))
 300                 return -ENOSPC;
 301         if (unlikely(vq->vq_free_cnt < num))
 302                 return -EMSGSIZE;
 303
 304         if (unlikely(vq->vq_desc_head_idx >= vq->vq_nentries))
 305                 return -EFAULT;
 306
 307         for (i = 0; i < num; i++) {
 308                 idx = vq->vq_desc_head_idx;
 309                 dxp = &vq->vq_descx[idx];
 310                 dxp->cookie = (void *)cookie[i];
 311                 dxp->ndescs = 1;
 312
 313                 start_dp[idx].addr = cookie[i]->buf_iova +
 314                         RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 315                 start_dp[idx].len = cookie[i]->buf_len -
 316                         RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
 317                 start_dp[idx].flags = VRING_DESC_F_WRITE;
 318                 vq->vq_desc_head_idx = start_dp[idx].next;
 319                 vq_update_avail_ring(vq, idx);
 320                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END) {
 321                         vq->vq_desc_tail_idx = vq->vq_desc_head_idx;
 322                         break;
 323                 }
 324         }
 325
 326         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 327
 328         return 0;
 329 }
 330
 331 static inline int
 332 virtqueue_enqueue_recv_refill_packed(struct virtqueue *vq,
 333                                      struct rte_mbuf **cookie, uint16_t num)
 334 {
 335         struct vring_packed_desc *start_dp = vq->vq_packed.ring.desc;
 336         uint16_t flags = vq->vq_packed.cached_flags;
 337         struct virtio_hw *hw = vq->hw;
 338         struct vq_desc_extra *dxp;
 339         uint16_t idx;
 340         int i;
 341
 342         if (unlikely(vq->vq_free_cnt == 0))
 343                 return -ENOSPC;
 344         if (unlikely(vq->vq_free_cnt < num))
 345                 return -EMSGSIZE;
 346
 347         for (i = 0; i < num; i++) {
 348                 idx = vq->vq_avail_idx;
 349                 dxp = &vq->vq_descx[idx];
 350                 dxp->cookie = (void *)cookie[i];
 351                 dxp->ndescs = 1;
 352
 353                 start_dp[idx].addr = cookie[i]->buf_iova +
 354                         RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 355                 start_dp[idx].len = cookie[i]->buf_len -
 356                         RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
 357
 358                 vq->vq_desc_head_idx = dxp->next;
 359                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 360                         vq->vq_desc_tail_idx = vq->vq_desc_head_idx;
 361
 362                 virtqueue_store_flags_packed(&start_dp[idx], flags,
 363                                              hw->weak_barriers);
 364
 365                 if (++vq->vq_avail_idx >= vq->vq_nentries) {
 366                         vq->vq_avail_idx -= vq->vq_nentries;
 367                         vq->vq_packed.cached_flags ^=
 368                                 VRING_PACKED_DESC_F_AVAIL_USED;
 369                         flags = vq->vq_packed.cached_flags;
 370                 }
 371         }
 372         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 373         return 0;
 374 }
 375
 376 /* When doing TSO, the IP length is not included in the pseudo header
 377  * checksum of the packet given to the PMD, but for virtio it is
 378  * expected.
 379  */
 380 static void
 381 virtio_tso_fix_cksum(struct rte_mbuf *m)
 382 {
 383         /* common case: header is not fragmented */
 384         if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
 385                         m->l4_len)) {
 386                 struct rte_ipv4_hdr *iph;
 387                 struct rte_ipv6_hdr *ip6h;
 388                 struct rte_tcp_hdr *th;
 389                 uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
 390                 uint32_t tmp;
 391
 392                 iph = rte_pktmbuf_mtod_offset(m,
 393                                         struct rte_ipv4_hdr *, m->l2_len);
 394                 th = RTE_PTR_ADD(iph, m->l3_len);
 395                 if ((iph->version_ihl >> 4) == 4) {
 396                         iph->hdr_checksum = 0;
 397                         iph->hdr_checksum = rte_ipv4_cksum(iph);
 398                         ip_len = iph->total_length;
 399                         ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
 400                                 m->l3_len);
 401                 } else {
 402                         ip6h = (struct rte_ipv6_hdr *)iph;
 403                         ip_paylen = ip6h->payload_len;
 404                 }
 405
 406                 /* calculate the new phdr checksum not including ip_paylen */
 407                 prev_cksum = th->cksum;
 408                 tmp = prev_cksum;
 409                 tmp += ip_paylen;
 410                 tmp = (tmp & 0xffff) + (tmp >> 16);
 411                 new_cksum = tmp;
 412
 413                 /* replace it in the packet */
 414                 th->cksum = new_cksum;
 415         }
 416 }
 417
 418
 419
 420
 421 static inline void
 422 virtqueue_enqueue_xmit_inorder(struct virtnet_tx *txvq,
 423                         struct rte_mbuf **cookies,
 424                         uint16_t num)
 425 {
 426         struct vq_desc_extra *dxp;
 427         struct virtqueue *vq = virtnet_txq_to_vq(txvq);
 428         struct vring_desc *start_dp;
 429         struct virtio_net_hdr *hdr;
 430         uint16_t idx;
 431         int16_t head_size = vq->hw->vtnet_hdr_size;
 432         uint16_t i = 0;
 433
 434         idx = vq->vq_desc_head_idx;
 435         start_dp = vq->vq_split.ring.desc;
 436
 437         while (i < num) {
 438                 idx = idx & (vq->vq_nentries - 1);
 439                 dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
 440                 dxp->cookie = (void *)cookies[i];
 441                 dxp->ndescs = 1;
 442                 virtio_update_packet_stats(&txvq->stats, cookies[i]);
 443
 444                 hdr = rte_pktmbuf_mtod_offset(cookies[i],
 445                                 struct virtio_net_hdr *, -head_size);
 446
 447                 /* if offload disabled, hdr is not zeroed yet, do it now */
 448                 if (!vq->hw->has_tx_offload)
 449                         virtqueue_clear_net_hdr(hdr);
 450                 else
 451                         virtqueue_xmit_offload(hdr, cookies[i]);
 452
 453                 start_dp[idx].addr  = rte_mbuf_data_iova(cookies[i]) - head_size;
 454                 start_dp[idx].len   = cookies[i]->data_len + head_size;
 455                 start_dp[idx].flags = 0;
 456
 457
 458                 vq_update_avail_ring(vq, idx);
 459
 460                 idx++;
 461                 i++;
 462         };
 463
 464         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 465         vq->vq_desc_head_idx = idx & (vq->vq_nentries - 1);
 466 }
 467
 468 static inline void
 469 virtqueue_enqueue_xmit_packed_fast(struct virtnet_tx *txvq,
 470                                    struct rte_mbuf *cookie,
 471                                    int in_order)
 472 {
 473         struct virtqueue *vq = virtnet_txq_to_vq(txvq);
 474         struct vring_packed_desc *dp;
 475         struct vq_desc_extra *dxp;
 476         uint16_t idx, id, flags;
 477         int16_t head_size = vq->hw->vtnet_hdr_size;
 478         struct virtio_net_hdr *hdr;
 479
 480         id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx;
 481         idx = vq->vq_avail_idx;
 482         dp = &vq->vq_packed.ring.desc[idx];
 483
 484         dxp = &vq->vq_descx[id];
 485         dxp->ndescs = 1;
 486         dxp->cookie = cookie;
 487
 488         flags = vq->vq_packed.cached_flags;
 489
 490         /* prepend cannot fail, checked by caller */
 491         hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
 492                                       -head_size);
 493
 494         /* if offload disabled, hdr is not zeroed yet, do it now */
 495         if (!vq->hw->has_tx_offload)
 496                 virtqueue_clear_net_hdr(hdr);
 497         else
 498                 virtqueue_xmit_offload(hdr, cookie);
 499
 500         dp->addr = rte_mbuf_data_iova(cookie) - head_size;
 501         dp->len  = cookie->data_len + head_size;
 502         dp->id   = id;
 503
 504         if (++vq->vq_avail_idx >= vq->vq_nentries) {
 505                 vq->vq_avail_idx -= vq->vq_nentries;
 506                 vq->vq_packed.cached_flags ^= VRING_PACKED_DESC_F_AVAIL_USED;
 507         }
 508
 509         vq->vq_free_cnt--;
 510
 511         if (!in_order) {
 512                 vq->vq_desc_head_idx = dxp->next;
 513                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 514                         vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
 515         }
 516
 517         virtqueue_store_flags_packed(dp, flags, vq->hw->weak_barriers);
 518 }
 519
 520 static inline void
 521 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
 522                         uint16_t needed, int use_indirect, int can_push,
 523                         int in_order)
 524 {
 525         struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 526         struct vq_desc_extra *dxp;
 527         struct virtqueue *vq = virtnet_txq_to_vq(txvq);
 528         struct vring_desc *start_dp;
 529         uint16_t seg_num = cookie->nb_segs;
 530         uint16_t head_idx, idx;
 531         int16_t head_size = vq->hw->vtnet_hdr_size;
 532         bool prepend_header = false;
 533         struct virtio_net_hdr *hdr;
 534
 535         head_idx = vq->vq_desc_head_idx;
 536         idx = head_idx;
 537         if (in_order)
 538                 dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
 539         else
 540                 dxp = &vq->vq_descx[idx];
 541         dxp->cookie = (void *)cookie;
 542         dxp->ndescs = needed;
 543
 544         start_dp = vq->vq_split.ring.desc;
 545
 546         if (can_push) {
 547                 /* prepend cannot fail, checked by caller */
 548                 hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
 549                                               -head_size);
 550                 prepend_header = true;
 551
 552                 /* if offload disabled, it is not zeroed below, do it now */
 553                 if (!vq->hw->has_tx_offload)
 554                         virtqueue_clear_net_hdr(hdr);
 555         } else if (use_indirect) {
 556                 /* setup tx ring slot to point to indirect
 557                  * descriptor list stored in reserved region.
 558                  *
 559                  * the first slot in indirect ring is already preset
 560                  * to point to the header in reserved region
 561                  */
 562                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 563                         RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
 564                 start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
 565                 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
 566                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 567
 568                 /* loop below will fill in rest of the indirect elements */
 569                 start_dp = txr[idx].tx_indir;
 570                 idx = 1;
 571         } else {
 572                 /* setup first tx ring slot to point to header
 573                  * stored in reserved region.
 574                  */
 575                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 576                         RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
 577                 start_dp[idx].len   = vq->hw->vtnet_hdr_size;
 578                 start_dp[idx].flags = VRING_DESC_F_NEXT;
 579                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 580
 581                 idx = start_dp[idx].next;
 582         }
 583
 584         if (vq->hw->has_tx_offload)
 585                 virtqueue_xmit_offload(hdr, cookie);
 586
 587         do {
 588                 start_dp[idx].addr  = rte_mbuf_data_iova(cookie);
 589                 start_dp[idx].len   = cookie->data_len;
 590                 if (prepend_header) {
 591                         start_dp[idx].addr -= head_size;
 592                         start_dp[idx].len += head_size;
 593                         prepend_header = false;
 594                 }
 595                 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 596                 idx = start_dp[idx].next;
 597         } while ((cookie = cookie->next) != NULL);
 598
 599         if (use_indirect)
 600                 idx = vq->vq_split.ring.desc[head_idx].next;
 601
 602         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 603
 604         vq->vq_desc_head_idx = idx;
 605         vq_update_avail_ring(vq, head_idx);
 606
 607         if (!in_order) {
 608                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 609                         vq->vq_desc_tail_idx = idx;
 610         }
 611 }
 612
 613 void
 614 virtio_dev_cq_start(struct rte_eth_dev *dev)
 615 {
 616         struct virtio_hw *hw = dev->data->dev_private;
 617
 618         if (hw->cvq) {
 619                 rte_spinlock_init(&hw->cvq->lock);
 620                 VIRTQUEUE_DUMP(virtnet_cq_to_vq(hw->cvq));
 621         }
 622 }
 623
 624 int
 625 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
 626                         uint16_t queue_idx,
 627                         uint16_t nb_desc,
 628                         unsigned int socket_id __rte_unused,
 629                         const struct rte_eth_rxconf *rx_conf,
 630                         struct rte_mempool *mp)
 631 {
 632         uint16_t vq_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 633         struct virtio_hw *hw = dev->data->dev_private;
 634         struct virtqueue *vq = hw->vqs[vq_idx];
 635         struct virtnet_rx *rxvq;
 636         uint16_t rx_free_thresh;
 637
 638         PMD_INIT_FUNC_TRACE();
 639
 640         if (rx_conf->rx_deferred_start) {
 641                 PMD_INIT_LOG(ERR, "Rx deferred start is not supported");
 642                 return -EINVAL;
 643         }
 644
 645         rx_free_thresh = rx_conf->rx_free_thresh;
 646         if (rx_free_thresh == 0)
 647                 rx_free_thresh =
 648                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_RX_FREE_THRESH);
 649
 650         if (rx_free_thresh & 0x3) {
 651                 RTE_LOG(ERR, PMD, "rx_free_thresh must be multiples of four."
 652                         " (rx_free_thresh=%u port=%u queue=%u)\n",
 653                         rx_free_thresh, dev->data->port_id, queue_idx);
 654                 return -EINVAL;
 655         }
 656
 657         if (rx_free_thresh >= vq->vq_nentries) {
 658                 RTE_LOG(ERR, PMD, "rx_free_thresh must be less than the "
 659                         "number of RX entries (%u)."
 660                         " (rx_free_thresh=%u port=%u queue=%u)\n",
 661                         vq->vq_nentries,
 662                         rx_free_thresh, dev->data->port_id, queue_idx);
 663                 return -EINVAL;
 664         }
 665         vq->vq_free_thresh = rx_free_thresh;
 666
 667         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 668                 nb_desc = vq->vq_nentries;
 669         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 670
 671         rxvq = &vq->rxq;
 672         rxvq->queue_id = queue_idx;
 673         rxvq->mpool = mp;
 674         dev->data->rx_queues[queue_idx] = rxvq;
 675
 676         return 0;
 677 }
 678
 679 int
 680 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
 681 {
 682         uint16_t vq_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 683         struct virtio_hw *hw = dev->data->dev_private;
 684         struct virtqueue *vq = hw->vqs[vq_idx];
 685         struct virtnet_rx *rxvq = &vq->rxq;
 686         struct rte_mbuf *m;
 687         uint16_t desc_idx;
 688         int error, nbufs, i;
 689         bool in_order = virtio_with_feature(hw, VIRTIO_F_IN_ORDER);
 690
 691         PMD_INIT_FUNC_TRACE();
 692
 693         /* Allocate blank mbufs for the each rx descriptor */
 694         nbufs = 0;
 695
 696         if (hw->use_vec_rx && !virtio_with_packed_queue(hw)) {
 697                 for (desc_idx = 0; desc_idx < vq->vq_nentries;
 698                      desc_idx++) {
 699                         vq->vq_split.ring.avail->ring[desc_idx] = desc_idx;
 700                         vq->vq_split.ring.desc[desc_idx].flags =
 701                                 VRING_DESC_F_WRITE;
 702                 }
 703
 704                 virtio_rxq_vec_setup(rxvq);
 705         }
 706
 707         memset(rxvq->fake_mbuf, 0, sizeof(*rxvq->fake_mbuf));
 708         for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST; desc_idx++)
 709                 vq->sw_ring[vq->vq_nentries + desc_idx] = rxvq->fake_mbuf;
 710
 711         if (hw->use_vec_rx && !virtio_with_packed_queue(hw)) {
 712                 while (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
 713                         virtio_rxq_rearm_vec(rxvq);
 714                         nbufs += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
 715                 }
 716         } else if (!virtio_with_packed_queue(vq->hw) && in_order) {
 717                 if ((!virtqueue_full(vq))) {
 718                         uint16_t free_cnt = vq->vq_free_cnt;
 719                         struct rte_mbuf *pkts[free_cnt];
 720
 721                         if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, pkts,
 722                                 free_cnt)) {
 723                                 error = virtqueue_enqueue_refill_inorder(vq,
 724                                                 pkts,
 725                                                 free_cnt);
 726                                 if (unlikely(error)) {
 727                                         for (i = 0; i < free_cnt; i++)
 728                                                 rte_pktmbuf_free(pkts[i]);
 729                                 }
 730                         }
 731
 732                         nbufs += free_cnt;
 733                         vq_update_avail_idx(vq);
 734                 }
 735         } else {
 736                 while (!virtqueue_full(vq)) {
 737                         m = rte_mbuf_raw_alloc(rxvq->mpool);
 738                         if (m == NULL)
 739                                 break;
 740
 741                         /* Enqueue allocated buffers */
 742                         if (virtio_with_packed_queue(vq->hw))
 743                                 error = virtqueue_enqueue_recv_refill_packed(vq,
 744                                                 &m, 1);
 745                         else
 746                                 error = virtqueue_enqueue_recv_refill(vq,
 747                                                 &m, 1);
 748                         if (error) {
 749                                 rte_pktmbuf_free(m);
 750                                 break;
 751                         }
 752                         nbufs++;
 753                 }
 754
 755                 if (!virtio_with_packed_queue(vq->hw))
 756                         vq_update_avail_idx(vq);
 757         }
 758
 759         PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
 760
 761         VIRTQUEUE_DUMP(vq);
 762
 763         return 0;
 764 }
 765
 766 /*
 767  * struct rte_eth_dev *dev: Used to update dev
 768  * uint16_t nb_desc: Defaults to values read from config space
 769  * unsigned int socket_id: Used to allocate memzone
 770  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
 771  * uint16_t queue_idx: Just used as an index in dev txq list
 772  */
 773 int
 774 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
 775                         uint16_t queue_idx,
 776                         uint16_t nb_desc,
 777                         unsigned int socket_id __rte_unused,
 778                         const struct rte_eth_txconf *tx_conf)
 779 {
 780         uint8_t vq_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 781         struct virtio_hw *hw = dev->data->dev_private;
 782         struct virtqueue *vq = hw->vqs[vq_idx];
 783         struct virtnet_tx *txvq;
 784         uint16_t tx_free_thresh;
 785
 786         PMD_INIT_FUNC_TRACE();
 787
 788         if (tx_conf->tx_deferred_start) {
 789                 PMD_INIT_LOG(ERR, "Tx deferred start is not supported");
 790                 return -EINVAL;
 791         }
 792
 793         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 794                 nb_desc = vq->vq_nentries;
 795         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 796
 797         txvq = &vq->txq;
 798         txvq->queue_id = queue_idx;
 799
 800         tx_free_thresh = tx_conf->tx_free_thresh;
 801         if (tx_free_thresh == 0)
 802                 tx_free_thresh =
 803                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
 804
 805         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
 806                 PMD_DRV_LOG(ERR, "tx_free_thresh must be less than the "
 807                         "number of TX entries minus 3 (%u)."
 808                         " (tx_free_thresh=%u port=%u queue=%u)\n",
 809                         vq->vq_nentries - 3,
 810                         tx_free_thresh, dev->data->port_id, queue_idx);
 811                 return -EINVAL;
 812         }
 813
 814         vq->vq_free_thresh = tx_free_thresh;
 815
 816         dev->data->tx_queues[queue_idx] = txvq;
 817         return 0;
 818 }
 819
 820 int
 821 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
 822                                 uint16_t queue_idx)
 823 {
 824         uint8_t vq_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 825         struct virtio_hw *hw = dev->data->dev_private;
 826         struct virtqueue *vq = hw->vqs[vq_idx];
 827
 828         PMD_INIT_FUNC_TRACE();
 829
 830         if (!virtio_with_packed_queue(hw)) {
 831                 if (virtio_with_feature(hw, VIRTIO_F_IN_ORDER))
 832                         vq->vq_split.ring.desc[vq->vq_nentries - 1].next = 0;
 833         }
 834
 835         VIRTQUEUE_DUMP(vq);
 836
 837         return 0;
 838 }
 839
 840 static inline void
 841 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
 842 {
 843         int error;
 844         /*
 845          * Requeue the discarded mbuf. This should always be
 846          * successful since it was just dequeued.
 847          */
 848         if (virtio_with_packed_queue(vq->hw))
 849                 error = virtqueue_enqueue_recv_refill_packed(vq, &m, 1);
 850         else
 851                 error = virtqueue_enqueue_recv_refill(vq, &m, 1);
 852
 853         if (unlikely(error)) {
 854                 PMD_DRV_LOG(ERR, "cannot requeue discarded mbuf");
 855                 rte_pktmbuf_free(m);
 856         }
 857 }
 858
 859 static inline void
 860 virtio_discard_rxbuf_inorder(struct virtqueue *vq, struct rte_mbuf *m)
 861 {
 862         int error;
 863
 864         error = virtqueue_enqueue_refill_inorder(vq, &m, 1);
 865         if (unlikely(error)) {
 866                 PMD_DRV_LOG(ERR, "cannot requeue discarded mbuf");
 867                 rte_pktmbuf_free(m);
 868         }
 869 }
 870
 871 /* Optionally fill offload information in structure */
 872 static inline int
 873 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
 874 {
 875         struct rte_net_hdr_lens hdr_lens;
 876         uint32_t hdrlen, ptype;
 877         int l4_supported = 0;
 878
 879         /* nothing to do */
 880         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 881                 return 0;
 882
 883         m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
 884
 885         ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
 886         m->packet_type = ptype;
 887         if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
 888             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
 889             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
 890                 l4_supported = 1;
 891
 892         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 893                 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
 894                 if (hdr->csum_start <= hdrlen && l4_supported) {
 895                         m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
 896                 } else {
 897                         /* Unknown proto or tunnel, do sw cksum. We can assume
 898                          * the cksum field is in the first segment since the
 899                          * buffers we provided to the host are large enough.
 900                          * In case of SCTP, this will be wrong since it's a CRC
 901                          * but there's nothing we can do.
 902                          */
 903                         uint16_t csum = 0, off;
 904
 905                         if (rte_raw_cksum_mbuf(m, hdr->csum_start,
 906                                 rte_pktmbuf_pkt_len(m) - hdr->csum_start,
 907                                 &csum) < 0)
 908                                 return -EINVAL;
 909                         if (likely(csum != 0xffff))
 910                                 csum = ~csum;
 911                         off = hdr->csum_offset + hdr->csum_start;
 912                         if (rte_pktmbuf_data_len(m) >= off + 1)
 913                                 *rte_pktmbuf_mtod_offset(m, uint16_t *,
 914                                         off) = csum;
 915                 }
 916         } else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
 917                 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
 918         }
 919
 920         /* GSO request, save required information in mbuf */
 921         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 922                 /* Check unsupported modes */
 923                 if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
 924                     (hdr->gso_size == 0)) {
 925                         return -EINVAL;
 926                 }
 927
 928                 /* Update mss lengthes in mbuf */
 929                 m->tso_segsz = hdr->gso_size;
 930                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 931                         case VIRTIO_NET_HDR_GSO_TCPV4:
 932                         case VIRTIO_NET_HDR_GSO_TCPV6:
 933                                 m->ol_flags |= PKT_RX_LRO | \
 934                                         PKT_RX_L4_CKSUM_NONE;
 935                                 break;
 936                         default:
 937                                 return -EINVAL;
 938                 }
 939         }
 940
 941         return 0;
 942 }
 943
 944 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
 945 uint16_t
 946 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 947 {
 948         struct virtnet_rx *rxvq = rx_queue;
 949         struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
 950         struct virtio_hw *hw = vq->hw;
 951         struct rte_mbuf *rxm;
 952         uint16_t nb_used, num, nb_rx;
 953         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 954         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 955         int error;
 956         uint32_t i, nb_enqueued;
 957         uint32_t hdr_size;
 958         struct virtio_net_hdr *hdr;
 959
 960         nb_rx = 0;
 961         if (unlikely(hw->started == 0))
 962                 return nb_rx;
 963
 964         nb_used = virtqueue_nused(vq);
 965
 966         num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
 967         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
 968                 num = VIRTIO_MBUF_BURST_SZ;
 969         if (likely(num > DESC_PER_CACHELINE))
 970                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
 971
 972         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
 973         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
 974
 975         nb_enqueued = 0;
 976         hdr_size = hw->vtnet_hdr_size;
 977
 978         for (i = 0; i < num ; i++) {
 979                 rxm = rcv_pkts[i];
 980
 981                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
 982
 983                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
 984                         PMD_RX_LOG(ERR, "Packet drop");
 985                         nb_enqueued++;
 986                         virtio_discard_rxbuf(vq, rxm);
 987                         rxvq->stats.errors++;
 988                         continue;
 989                 }
 990
 991                 rxm->port = rxvq->port_id;
 992                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 993                 rxm->ol_flags = 0;
 994                 rxm->vlan_tci = 0;
 995
 996                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
 997                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
 998
 999                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1000                         RTE_PKTMBUF_HEADROOM - hdr_size);
1001
1002                 if (hw->vlan_strip)
1003                         rte_vlan_strip(rxm);
1004
1005                 if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1006                         virtio_discard_rxbuf(vq, rxm);
1007                         rxvq->stats.errors++;
1008                         continue;
1009                 }
1010
1011                 virtio_rx_stats_updated(rxvq, rxm);
1012
1013                 rx_pkts[nb_rx++] = rxm;
1014         }
1015
1016         rxvq->stats.packets += nb_rx;
1017
1018         /* Allocate new mbuf for the used descriptor */
1019         if (likely(!virtqueue_full(vq))) {
1020                 uint16_t free_cnt = vq->vq_free_cnt;
1021                 struct rte_mbuf *new_pkts[free_cnt];
1022
1023                 if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1024                                                 free_cnt) == 0)) {
1025                         error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1026                                         free_cnt);
1027                         if (unlikely(error)) {
1028                                 for (i = 0; i < free_cnt; i++)
1029                                         rte_pktmbuf_free(new_pkts[i]);
1030                         }
1031                         nb_enqueued += free_cnt;
1032                 } else {
1033                         struct rte_eth_dev *dev =
1034                                 &rte_eth_devices[rxvq->port_id];
1035                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1036                 }
1037         }
1038
1039         if (likely(nb_enqueued)) {
1040                 vq_update_avail_idx(vq);
1041
1042                 if (unlikely(virtqueue_kick_prepare(vq))) {
1043                         virtqueue_notify(vq);
1044                         PMD_RX_LOG(DEBUG, "Notified");
1045                 }
1046         }
1047
1048         return nb_rx;
1049 }
1050
1051 uint16_t
1052 virtio_recv_pkts_packed(void *rx_queue, struct rte_mbuf **rx_pkts,
1053                         uint16_t nb_pkts)
1054 {
1055         struct virtnet_rx *rxvq = rx_queue;
1056         struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1057         struct virtio_hw *hw = vq->hw;
1058         struct rte_mbuf *rxm;
1059         uint16_t num, nb_rx;
1060         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1061         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1062         int error;
1063         uint32_t i, nb_enqueued;
1064         uint32_t hdr_size;
1065         struct virtio_net_hdr *hdr;
1066
1067         nb_rx = 0;
1068         if (unlikely(hw->started == 0))
1069                 return nb_rx;
1070
1071         num = RTE_MIN(VIRTIO_MBUF_BURST_SZ, nb_pkts);
1072         if (likely(num > DESC_PER_CACHELINE))
1073                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1074
1075         num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1076         PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1077
1078         nb_enqueued = 0;
1079         hdr_size = hw->vtnet_hdr_size;
1080
1081         for (i = 0; i < num; i++) {
1082                 rxm = rcv_pkts[i];
1083
1084                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1085
1086                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1087                         PMD_RX_LOG(ERR, "Packet drop");
1088                         nb_enqueued++;
1089                         virtio_discard_rxbuf(vq, rxm);
1090                         rxvq->stats.errors++;
1091                         continue;
1092                 }
1093
1094                 rxm->port = rxvq->port_id;
1095                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1096                 rxm->ol_flags = 0;
1097                 rxm->vlan_tci = 0;
1098
1099                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1100                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1101
1102                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1103                         RTE_PKTMBUF_HEADROOM - hdr_size);
1104
1105                 if (hw->vlan_strip)
1106                         rte_vlan_strip(rxm);
1107
1108                 if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1109                         virtio_discard_rxbuf(vq, rxm);
1110                         rxvq->stats.errors++;
1111                         continue;
1112                 }
1113
1114                 virtio_rx_stats_updated(rxvq, rxm);
1115
1116                 rx_pkts[nb_rx++] = rxm;
1117         }
1118
1119         rxvq->stats.packets += nb_rx;
1120
1121         /* Allocate new mbuf for the used descriptor */
1122         if (likely(!virtqueue_full(vq))) {
1123                 uint16_t free_cnt = vq->vq_free_cnt;
1124                 struct rte_mbuf *new_pkts[free_cnt];
1125
1126                 if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1127                                                 free_cnt) == 0)) {
1128                         error = virtqueue_enqueue_recv_refill_packed(vq,
1129                                         new_pkts, free_cnt);
1130                         if (unlikely(error)) {
1131                                 for (i = 0; i < free_cnt; i++)
1132                                         rte_pktmbuf_free(new_pkts[i]);
1133                         }
1134                         nb_enqueued += free_cnt;
1135                 } else {
1136                         struct rte_eth_dev *dev =
1137                                 &rte_eth_devices[rxvq->port_id];
1138                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1139                 }
1140         }
1141
1142         if (likely(nb_enqueued)) {
1143                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1144                         virtqueue_notify(vq);
1145                         PMD_RX_LOG(DEBUG, "Notified");
1146                 }
1147         }
1148
1149         return nb_rx;
1150 }
1151
1152
1153 uint16_t
1154 virtio_recv_pkts_inorder(void *rx_queue,
1155                         struct rte_mbuf **rx_pkts,
1156                         uint16_t nb_pkts)
1157 {
1158         struct virtnet_rx *rxvq = rx_queue;
1159         struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1160         struct virtio_hw *hw = vq->hw;
1161         struct rte_mbuf *rxm;
1162         struct rte_mbuf *prev = NULL;
1163         uint16_t nb_used, num, nb_rx;
1164         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1165         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1166         int error;
1167         uint32_t nb_enqueued;
1168         uint32_t seg_num;
1169         uint32_t seg_res;
1170         uint32_t hdr_size;
1171         int32_t i;
1172
1173         nb_rx = 0;
1174         if (unlikely(hw->started == 0))
1175                 return nb_rx;
1176
1177         nb_used = virtqueue_nused(vq);
1178         nb_used = RTE_MIN(nb_used, nb_pkts);
1179         nb_used = RTE_MIN(nb_used, VIRTIO_MBUF_BURST_SZ);
1180
1181         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1182
1183         nb_enqueued = 0;
1184         seg_num = 1;
1185         seg_res = 0;
1186         hdr_size = hw->vtnet_hdr_size;
1187
1188         num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len, nb_used);
1189
1190         for (i = 0; i < num; i++) {
1191                 struct virtio_net_hdr_mrg_rxbuf *header;
1192
1193                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1194                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1195
1196                 rxm = rcv_pkts[i];
1197
1198                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1199                         PMD_RX_LOG(ERR, "Packet drop");
1200                         nb_enqueued++;
1201                         virtio_discard_rxbuf_inorder(vq, rxm);
1202                         rxvq->stats.errors++;
1203                         continue;
1204                 }
1205
1206                 header = (struct virtio_net_hdr_mrg_rxbuf *)
1207                          ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1208                          - hdr_size);
1209
1210                 if (virtio_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
1211                         seg_num = header->num_buffers;
1212                         if (seg_num == 0)
1213                                 seg_num = 1;
1214                 } else {
1215                         seg_num = 1;
1216                 }
1217
1218                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1219                 rxm->nb_segs = seg_num;
1220                 rxm->ol_flags = 0;
1221                 rxm->vlan_tci = 0;
1222                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1223                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1224
1225                 rxm->port = rxvq->port_id;
1226
1227                 rx_pkts[nb_rx] = rxm;
1228                 prev = rxm;
1229
1230                 if (vq->hw->has_rx_offload &&
1231                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1232                         virtio_discard_rxbuf_inorder(vq, rxm);
1233                         rxvq->stats.errors++;
1234                         continue;
1235                 }
1236
1237                 if (hw->vlan_strip)
1238                         rte_vlan_strip(rx_pkts[nb_rx]);
1239
1240                 seg_res = seg_num - 1;
1241
1242                 /* Merge remaining segments */
1243                 while (seg_res != 0 && i < (num - 1)) {
1244                         i++;
1245
1246                         rxm = rcv_pkts[i];
1247                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1248                         rxm->pkt_len = (uint32_t)(len[i]);
1249                         rxm->data_len = (uint16_t)(len[i]);
1250
1251                         rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1252
1253                         prev->next = rxm;
1254                         prev = rxm;
1255                         seg_res -= 1;
1256                 }
1257
1258                 if (!seg_res) {
1259                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1260                         nb_rx++;
1261                 }
1262         }
1263
1264         /* Last packet still need merge segments */
1265         while (seg_res != 0) {
1266                 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1267                                         VIRTIO_MBUF_BURST_SZ);
1268
1269                 if (likely(virtqueue_nused(vq) >= rcv_cnt)) {
1270                         num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len,
1271                                                            rcv_cnt);
1272                         uint16_t extra_idx = 0;
1273
1274                         rcv_cnt = num;
1275                         while (extra_idx < rcv_cnt) {
1276                                 rxm = rcv_pkts[extra_idx];
1277                                 rxm->data_off =
1278                                         RTE_PKTMBUF_HEADROOM - hdr_size;
1279                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1280                                 rxm->data_len = (uint16_t)(len[extra_idx]);
1281                                 prev->next = rxm;
1282                                 prev = rxm;
1283                                 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1284                                 extra_idx += 1;
1285                         };
1286                         seg_res -= rcv_cnt;
1287
1288                         if (!seg_res) {
1289                                 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1290                                 nb_rx++;
1291                         }
1292                 } else {
1293                         PMD_RX_LOG(ERR,
1294                                         "No enough segments for packet.");
1295                         rte_pktmbuf_free(rx_pkts[nb_rx]);
1296                         rxvq->stats.errors++;
1297                         break;
1298                 }
1299         }
1300
1301         rxvq->stats.packets += nb_rx;
1302
1303         /* Allocate new mbuf for the used descriptor */
1304
1305         if (likely(!virtqueue_full(vq))) {
1306                 /* free_cnt may include mrg descs */
1307                 uint16_t free_cnt = vq->vq_free_cnt;
1308                 struct rte_mbuf *new_pkts[free_cnt];
1309
1310                 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1311                         error = virtqueue_enqueue_refill_inorder(vq, new_pkts,
1312                                         free_cnt);
1313                         if (unlikely(error)) {
1314                                 for (i = 0; i < free_cnt; i++)
1315                                         rte_pktmbuf_free(new_pkts[i]);
1316                         }
1317                         nb_enqueued += free_cnt;
1318                 } else {
1319                         struct rte_eth_dev *dev =
1320                                 &rte_eth_devices[rxvq->port_id];
1321                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1322                 }
1323         }
1324
1325         if (likely(nb_enqueued)) {
1326                 vq_update_avail_idx(vq);
1327
1328                 if (unlikely(virtqueue_kick_prepare(vq))) {
1329                         virtqueue_notify(vq);
1330                         PMD_RX_LOG(DEBUG, "Notified");
1331                 }
1332         }
1333
1334         return nb_rx;
1335 }
1336
1337 uint16_t
1338 virtio_recv_mergeable_pkts(void *rx_queue,
1339                         struct rte_mbuf **rx_pkts,
1340                         uint16_t nb_pkts)
1341 {
1342         struct virtnet_rx *rxvq = rx_queue;
1343         struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1344         struct virtio_hw *hw = vq->hw;
1345         struct rte_mbuf *rxm;
1346         struct rte_mbuf *prev = NULL;
1347         uint16_t nb_used, num, nb_rx = 0;
1348         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1349         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1350         int error;
1351         uint32_t nb_enqueued = 0;
1352         uint32_t seg_num = 0;
1353         uint32_t seg_res = 0;
1354         uint32_t hdr_size = hw->vtnet_hdr_size;
1355         int32_t i;
1356
1357         if (unlikely(hw->started == 0))
1358                 return nb_rx;
1359
1360         nb_used = virtqueue_nused(vq);
1361
1362         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1363
1364         num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1365         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1366                 num = VIRTIO_MBUF_BURST_SZ;
1367         if (likely(num > DESC_PER_CACHELINE))
1368                 num = num - ((vq->vq_used_cons_idx + num) %
1369                                 DESC_PER_CACHELINE);
1370
1371
1372         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1373
1374         for (i = 0; i < num; i++) {
1375                 struct virtio_net_hdr_mrg_rxbuf *header;
1376
1377                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1378                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1379
1380                 rxm = rcv_pkts[i];
1381
1382                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1383                         PMD_RX_LOG(ERR, "Packet drop");
1384                         nb_enqueued++;
1385                         virtio_discard_rxbuf(vq, rxm);
1386                         rxvq->stats.errors++;
1387                         continue;
1388                 }
1389
1390                 header = (struct virtio_net_hdr_mrg_rxbuf *)
1391                          ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1392                          - hdr_size);
1393                 seg_num = header->num_buffers;
1394                 if (seg_num == 0)
1395                         seg_num = 1;
1396
1397                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1398                 rxm->nb_segs = seg_num;
1399                 rxm->ol_flags = 0;
1400                 rxm->vlan_tci = 0;
1401                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1402                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1403
1404                 rxm->port = rxvq->port_id;
1405
1406                 rx_pkts[nb_rx] = rxm;
1407                 prev = rxm;
1408
1409                 if (hw->has_rx_offload &&
1410                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1411                         virtio_discard_rxbuf(vq, rxm);
1412                         rxvq->stats.errors++;
1413                         continue;
1414                 }
1415
1416                 if (hw->vlan_strip)
1417                         rte_vlan_strip(rx_pkts[nb_rx]);
1418
1419                 seg_res = seg_num - 1;
1420
1421                 /* Merge remaining segments */
1422                 while (seg_res != 0 && i < (num - 1)) {
1423                         i++;
1424
1425                         rxm = rcv_pkts[i];
1426                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1427                         rxm->pkt_len = (uint32_t)(len[i]);
1428                         rxm->data_len = (uint16_t)(len[i]);
1429
1430                         rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1431
1432                         prev->next = rxm;
1433                         prev = rxm;
1434                         seg_res -= 1;
1435                 }
1436
1437                 if (!seg_res) {
1438                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1439                         nb_rx++;
1440                 }
1441         }
1442
1443         /* Last packet still need merge segments */
1444         while (seg_res != 0) {
1445                 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1446                                         VIRTIO_MBUF_BURST_SZ);
1447
1448                 if (likely(virtqueue_nused(vq) >= rcv_cnt)) {
1449                         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len,
1450                                                            rcv_cnt);
1451                         uint16_t extra_idx = 0;
1452
1453                         rcv_cnt = num;
1454                         while (extra_idx < rcv_cnt) {
1455                                 rxm = rcv_pkts[extra_idx];
1456                                 rxm->data_off =
1457                                         RTE_PKTMBUF_HEADROOM - hdr_size;
1458                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1459                                 rxm->data_len = (uint16_t)(len[extra_idx]);
1460                                 prev->next = rxm;
1461                                 prev = rxm;
1462                                 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1463                                 extra_idx += 1;
1464                         };
1465                         seg_res -= rcv_cnt;
1466
1467                         if (!seg_res) {
1468                                 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1469                                 nb_rx++;
1470                         }
1471                 } else {
1472                         PMD_RX_LOG(ERR,
1473                                         "No enough segments for packet.");
1474                         rte_pktmbuf_free(rx_pkts[nb_rx]);
1475                         rxvq->stats.errors++;
1476                         break;
1477                 }
1478         }
1479
1480         rxvq->stats.packets += nb_rx;
1481
1482         /* Allocate new mbuf for the used descriptor */
1483         if (likely(!virtqueue_full(vq))) {
1484                 /* free_cnt may include mrg descs */
1485                 uint16_t free_cnt = vq->vq_free_cnt;
1486                 struct rte_mbuf *new_pkts[free_cnt];
1487
1488                 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1489                         error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1490                                         free_cnt);
1491                         if (unlikely(error)) {
1492                                 for (i = 0; i < free_cnt; i++)
1493                                         rte_pktmbuf_free(new_pkts[i]);
1494                         }
1495                         nb_enqueued += free_cnt;
1496                 } else {
1497                         struct rte_eth_dev *dev =
1498                                 &rte_eth_devices[rxvq->port_id];
1499                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1500                 }
1501         }
1502
1503         if (likely(nb_enqueued)) {
1504                 vq_update_avail_idx(vq);
1505
1506                 if (unlikely(virtqueue_kick_prepare(vq))) {
1507                         virtqueue_notify(vq);
1508                         PMD_RX_LOG(DEBUG, "Notified");
1509                 }
1510         }
1511
1512         return nb_rx;
1513 }
1514
1515 uint16_t
1516 virtio_recv_mergeable_pkts_packed(void *rx_queue,
1517                         struct rte_mbuf **rx_pkts,
1518                         uint16_t nb_pkts)
1519 {
1520         struct virtnet_rx *rxvq = rx_queue;
1521         struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1522         struct virtio_hw *hw = vq->hw;
1523         struct rte_mbuf *rxm;
1524         struct rte_mbuf *prev = NULL;
1525         uint16_t num, nb_rx = 0;
1526         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1527         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1528         uint32_t nb_enqueued = 0;
1529         uint32_t seg_num = 0;
1530         uint32_t seg_res = 0;
1531         uint32_t hdr_size = hw->vtnet_hdr_size;
1532         int32_t i;
1533         int error;
1534
1535         if (unlikely(hw->started == 0))
1536                 return nb_rx;
1537
1538
1539         num = nb_pkts;
1540         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1541                 num = VIRTIO_MBUF_BURST_SZ;
1542         if (likely(num > DESC_PER_CACHELINE))
1543                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1544
1545         num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1546
1547         for (i = 0; i < num; i++) {
1548                 struct virtio_net_hdr_mrg_rxbuf *header;
1549
1550                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1551                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1552
1553                 rxm = rcv_pkts[i];
1554
1555                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1556                         PMD_RX_LOG(ERR, "Packet drop");
1557                         nb_enqueued++;
1558                         virtio_discard_rxbuf(vq, rxm);
1559                         rxvq->stats.errors++;
1560                         continue;
1561                 }
1562
1563                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)
1564                           rxm->buf_addr + RTE_PKTMBUF_HEADROOM - hdr_size);
1565                 seg_num = header->num_buffers;
1566
1567                 if (seg_num == 0)
1568                         seg_num = 1;
1569
1570                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1571                 rxm->nb_segs = seg_num;
1572                 rxm->ol_flags = 0;
1573                 rxm->vlan_tci = 0;
1574                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1575                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1576
1577                 rxm->port = rxvq->port_id;
1578                 rx_pkts[nb_rx] = rxm;
1579                 prev = rxm;
1580
1581                 if (hw->has_rx_offload &&
1582                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1583                         virtio_discard_rxbuf(vq, rxm);
1584                         rxvq->stats.errors++;
1585                         continue;
1586                 }
1587
1588                 if (hw->vlan_strip)
1589                         rte_vlan_strip(rx_pkts[nb_rx]);
1590
1591                 seg_res = seg_num - 1;
1592
1593                 /* Merge remaining segments */
1594                 while (seg_res != 0 && i < (num - 1)) {
1595                         i++;
1596
1597                         rxm = rcv_pkts[i];
1598                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1599                         rxm->pkt_len = (uint32_t)(len[i]);
1600                         rxm->data_len = (uint16_t)(len[i]);
1601
1602                         rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1603
1604                         prev->next = rxm;
1605                         prev = rxm;
1606                         seg_res -= 1;
1607                 }
1608
1609                 if (!seg_res) {
1610                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1611                         nb_rx++;
1612                 }
1613         }
1614
1615         /* Last packet still need merge segments */
1616         while (seg_res != 0) {
1617                 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1618                                         VIRTIO_MBUF_BURST_SZ);
1619                 uint16_t extra_idx = 0;
1620
1621                 rcv_cnt = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts,
1622                                 len, rcv_cnt);
1623                 if (unlikely(rcv_cnt == 0)) {
1624                         PMD_RX_LOG(ERR, "No enough segments for packet.");
1625                         rte_pktmbuf_free(rx_pkts[nb_rx]);
1626                         rxvq->stats.errors++;
1627                         break;
1628                 }
1629
1630                 while (extra_idx < rcv_cnt) {
1631                         rxm = rcv_pkts[extra_idx];
1632
1633                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1634                         rxm->pkt_len = (uint32_t)(len[extra_idx]);
1635                         rxm->data_len = (uint16_t)(len[extra_idx]);
1636
1637                         prev->next = rxm;
1638                         prev = rxm;
1639                         rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1640                         extra_idx += 1;
1641                 }
1642                 seg_res -= rcv_cnt;
1643                 if (!seg_res) {
1644                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1645                         nb_rx++;
1646                 }
1647         }
1648
1649         rxvq->stats.packets += nb_rx;
1650
1651         /* Allocate new mbuf for the used descriptor */
1652         if (likely(!virtqueue_full(vq))) {
1653                 /* free_cnt may include mrg descs */
1654                 uint16_t free_cnt = vq->vq_free_cnt;
1655                 struct rte_mbuf *new_pkts[free_cnt];
1656
1657                 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1658                         error = virtqueue_enqueue_recv_refill_packed(vq,
1659                                         new_pkts, free_cnt);
1660                         if (unlikely(error)) {
1661                                 for (i = 0; i < free_cnt; i++)
1662                                         rte_pktmbuf_free(new_pkts[i]);
1663                         }
1664                         nb_enqueued += free_cnt;
1665                 } else {
1666                         struct rte_eth_dev *dev =
1667                                 &rte_eth_devices[rxvq->port_id];
1668                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1669                 }
1670         }
1671
1672         if (likely(nb_enqueued)) {
1673                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1674                         virtqueue_notify(vq);
1675                         PMD_RX_LOG(DEBUG, "Notified");
1676                 }
1677         }
1678
1679         return nb_rx;
1680 }
1681
1682 uint16_t
1683 virtio_xmit_pkts_prepare(void *tx_queue __rte_unused, struct rte_mbuf **tx_pkts,
1684                         uint16_t nb_pkts)
1685 {
1686         uint16_t nb_tx;
1687         int error;
1688
1689         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1690                 struct rte_mbuf *m = tx_pkts[nb_tx];
1691
1692 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
1693                 error = rte_validate_tx_offload(m);
1694                 if (unlikely(error)) {
1695                         rte_errno = -error;
1696                         break;
1697                 }
1698 #endif
1699
1700                 /* Do VLAN tag insertion */
1701                 if (unlikely(m->ol_flags & PKT_TX_VLAN_PKT)) {
1702                         error = rte_vlan_insert(&m);
1703                         /* rte_vlan_insert() may change pointer
1704                          * even in the case of failure
1705                          */
1706                         tx_pkts[nb_tx] = m;
1707
1708                         if (unlikely(error)) {
1709                                 rte_errno = -error;
1710                                 break;
1711                         }
1712                 }
1713
1714                 error = rte_net_intel_cksum_prepare(m);
1715                 if (unlikely(error)) {
1716                         rte_errno = -error;
1717                         break;
1718                 }
1719
1720                 if (m->ol_flags & PKT_TX_TCP_SEG)
1721                         virtio_tso_fix_cksum(m);
1722         }
1723
1724         return nb_tx;
1725 }
1726
1727 uint16_t
1728 virtio_xmit_pkts_packed(void *tx_queue, struct rte_mbuf **tx_pkts,
1729                         uint16_t nb_pkts)
1730 {
1731         struct virtnet_tx *txvq = tx_queue;
1732         struct virtqueue *vq = virtnet_txq_to_vq(txvq);
1733         struct virtio_hw *hw = vq->hw;
1734         uint16_t hdr_size = hw->vtnet_hdr_size;
1735         uint16_t nb_tx = 0;
1736         bool in_order = virtio_with_feature(hw, VIRTIO_F_IN_ORDER);
1737
1738         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1739                 return nb_tx;
1740
1741         if (unlikely(nb_pkts < 1))
1742                 return nb_pkts;
1743
1744         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1745
1746         if (nb_pkts > vq->vq_free_cnt)
1747                 virtio_xmit_cleanup_packed(vq, nb_pkts - vq->vq_free_cnt,
1748                                            in_order);
1749
1750         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1751                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1752                 int can_push = 0, use_indirect = 0, slots, need;
1753
1754                 /* optimize ring usage */
1755                 if ((virtio_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1756                       virtio_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1757                     rte_mbuf_refcnt_read(txm) == 1 &&
1758                     RTE_MBUF_DIRECT(txm) &&
1759                     txm->nb_segs == 1 &&
1760                     rte_pktmbuf_headroom(txm) >= hdr_size &&
1761                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1762                            __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1763                         can_push = 1;
1764                 else if (virtio_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1765                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1766                         use_indirect = 1;
1767                 /* How many main ring entries are needed to this Tx?
1768                  * indirect   => 1
1769                  * any_layout => number of segments
1770                  * default    => number of segments + 1
1771                  */
1772                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1773                 need = slots - vq->vq_free_cnt;
1774
1775                 /* Positive value indicates it need free vring descriptors */
1776                 if (unlikely(need > 0)) {
1777                         virtio_xmit_cleanup_packed(vq, need, in_order);
1778                         need = slots - vq->vq_free_cnt;
1779                         if (unlikely(need > 0)) {
1780                                 PMD_TX_LOG(ERR,
1781                                            "No free tx descriptors to transmit");
1782                                 break;
1783                         }
1784                 }
1785
1786                 /* Enqueue Packet buffers */
1787                 if (can_push)
1788                         virtqueue_enqueue_xmit_packed_fast(txvq, txm, in_order);
1789                 else
1790                         virtqueue_enqueue_xmit_packed(txvq, txm, slots,
1791                                                       use_indirect, 0,
1792                                                       in_order);
1793
1794                 virtio_update_packet_stats(&txvq->stats, txm);
1795         }
1796
1797         txvq->stats.packets += nb_tx;
1798
1799         if (likely(nb_tx)) {
1800                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1801                         virtqueue_notify(vq);
1802                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1803                 }
1804         }
1805
1806         return nb_tx;
1807 }
1808
1809 uint16_t
1810 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1811 {
1812         struct virtnet_tx *txvq = tx_queue;
1813         struct virtqueue *vq = virtnet_txq_to_vq(txvq);
1814         struct virtio_hw *hw = vq->hw;
1815         uint16_t hdr_size = hw->vtnet_hdr_size;
1816         uint16_t nb_used, nb_tx = 0;
1817
1818         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1819                 return nb_tx;
1820
1821         if (unlikely(nb_pkts < 1))
1822                 return nb_pkts;
1823
1824         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1825
1826         nb_used = virtqueue_nused(vq);
1827
1828         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1829                 virtio_xmit_cleanup(vq, nb_used);
1830
1831         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1832                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1833                 int can_push = 0, use_indirect = 0, slots, need;
1834
1835                 /* optimize ring usage */
1836                 if ((virtio_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1837                       virtio_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1838                     rte_mbuf_refcnt_read(txm) == 1 &&
1839                     RTE_MBUF_DIRECT(txm) &&
1840                     txm->nb_segs == 1 &&
1841                     rte_pktmbuf_headroom(txm) >= hdr_size &&
1842                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1843                                    __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1844                         can_push = 1;
1845                 else if (virtio_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1846                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1847                         use_indirect = 1;
1848
1849                 /* How many main ring entries are needed to this Tx?
1850                  * any_layout => number of segments
1851                  * indirect   => 1
1852                  * default    => number of segments + 1
1853                  */
1854                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1855                 need = slots - vq->vq_free_cnt;
1856
1857                 /* Positive value indicates it need free vring descriptors */
1858                 if (unlikely(need > 0)) {
1859                         nb_used = virtqueue_nused(vq);
1860
1861                         need = RTE_MIN(need, (int)nb_used);
1862
1863                         virtio_xmit_cleanup(vq, need);
1864                         need = slots - vq->vq_free_cnt;
1865                         if (unlikely(need > 0)) {
1866                                 PMD_TX_LOG(ERR,
1867                                            "No free tx descriptors to transmit");
1868                                 break;
1869                         }
1870                 }
1871
1872                 /* Enqueue Packet buffers */
1873                 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
1874                         can_push, 0);
1875
1876                 virtio_update_packet_stats(&txvq->stats, txm);
1877         }
1878
1879         txvq->stats.packets += nb_tx;
1880
1881         if (likely(nb_tx)) {
1882                 vq_update_avail_idx(vq);
1883
1884                 if (unlikely(virtqueue_kick_prepare(vq))) {
1885                         virtqueue_notify(vq);
1886                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1887                 }
1888         }
1889
1890         return nb_tx;
1891 }
1892
1893 static __rte_always_inline int
1894 virtio_xmit_try_cleanup_inorder(struct virtqueue *vq, uint16_t need)
1895 {
1896         uint16_t nb_used, nb_clean, nb_descs;
1897
1898         nb_descs = vq->vq_free_cnt + need;
1899         nb_used = virtqueue_nused(vq);
1900         nb_clean = RTE_MIN(need, (int)nb_used);
1901
1902         virtio_xmit_cleanup_inorder(vq, nb_clean);
1903
1904         return nb_descs - vq->vq_free_cnt;
1905 }
1906
1907 uint16_t
1908 virtio_xmit_pkts_inorder(void *tx_queue,
1909                         struct rte_mbuf **tx_pkts,
1910                         uint16_t nb_pkts)
1911 {
1912         struct virtnet_tx *txvq = tx_queue;
1913         struct virtqueue *vq = virtnet_txq_to_vq(txvq);
1914         struct virtio_hw *hw = vq->hw;
1915         uint16_t hdr_size = hw->vtnet_hdr_size;
1916         uint16_t nb_used, nb_tx = 0, nb_inorder_pkts = 0;
1917         struct rte_mbuf *inorder_pkts[nb_pkts];
1918         int need;
1919
1920         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1921                 return nb_tx;
1922
1923         if (unlikely(nb_pkts < 1))
1924                 return nb_pkts;
1925
1926         VIRTQUEUE_DUMP(vq);
1927         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1928         nb_used = virtqueue_nused(vq);
1929
1930         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1931                 virtio_xmit_cleanup_inorder(vq, nb_used);
1932
1933         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1934                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1935                 int slots;
1936
1937                 /* optimize ring usage */
1938                 if ((virtio_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1939                      virtio_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1940                      rte_mbuf_refcnt_read(txm) == 1 &&
1941                      RTE_MBUF_DIRECT(txm) &&
1942                      txm->nb_segs == 1 &&
1943                      rte_pktmbuf_headroom(txm) >= hdr_size &&
1944                      rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1945                                 __alignof__(struct virtio_net_hdr_mrg_rxbuf))) {
1946                         inorder_pkts[nb_inorder_pkts] = txm;
1947                         nb_inorder_pkts++;
1948
1949                         continue;
1950                 }
1951
1952                 if (nb_inorder_pkts) {
1953                         need = nb_inorder_pkts - vq->vq_free_cnt;
1954                         if (unlikely(need > 0)) {
1955                                 need = virtio_xmit_try_cleanup_inorder(vq,
1956                                                                        need);
1957                                 if (unlikely(need > 0)) {
1958                                         PMD_TX_LOG(ERR,
1959                                                 "No free tx descriptors to "
1960                                                 "transmit");
1961                                         break;
1962                                 }
1963                         }
1964                         virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
1965                                                         nb_inorder_pkts);
1966                         nb_inorder_pkts = 0;
1967                 }
1968
1969                 slots = txm->nb_segs + 1;
1970                 need = slots - vq->vq_free_cnt;
1971                 if (unlikely(need > 0)) {
1972                         need = virtio_xmit_try_cleanup_inorder(vq, slots);
1973
1974                         if (unlikely(need > 0)) {
1975                                 PMD_TX_LOG(ERR,
1976                                         "No free tx descriptors to transmit");
1977                                 break;
1978                         }
1979                 }
1980                 /* Enqueue Packet buffers */
1981                 virtqueue_enqueue_xmit(txvq, txm, slots, 0, 0, 1);
1982
1983                 virtio_update_packet_stats(&txvq->stats, txm);
1984         }
1985
1986         /* Transmit all inorder packets */
1987         if (nb_inorder_pkts) {
1988                 need = nb_inorder_pkts - vq->vq_free_cnt;
1989                 if (unlikely(need > 0)) {
1990                         need = virtio_xmit_try_cleanup_inorder(vq,
1991                                                                   need);
1992                         if (unlikely(need > 0)) {
1993                                 PMD_TX_LOG(ERR,
1994                                         "No free tx descriptors to transmit");
1995                                 nb_inorder_pkts = vq->vq_free_cnt;
1996                                 nb_tx -= need;
1997                         }
1998                 }
1999
2000                 virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2001                                                 nb_inorder_pkts);
2002         }
2003
2004         txvq->stats.packets += nb_tx;
2005
2006         if (likely(nb_tx)) {
2007                 vq_update_avail_idx(vq);
2008
2009                 if (unlikely(virtqueue_kick_prepare(vq))) {
2010                         virtqueue_notify(vq);
2011                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2012                 }
2013         }
2014
2015         VIRTQUEUE_DUMP(vq);
2016
2017         return nb_tx;
2018 }
2019
2020 __rte_weak uint16_t
2021 virtio_recv_pkts_packed_vec(void *rx_queue __rte_unused,
2022                             struct rte_mbuf **rx_pkts __rte_unused,
2023                             uint16_t nb_pkts __rte_unused)
2024 {
2025         return 0;
2026 }
2027
2028 __rte_weak uint16_t
2029 virtio_xmit_pkts_packed_vec(void *tx_queue __rte_unused,
2030                             struct rte_mbuf **tx_pkts __rte_unused,
2031                             uint16_t nb_pkts __rte_unused)
2032 {
2033         return 0;
2034 }