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], true);
 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, true);
 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         virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload);
 585
 586         do {
 587                 start_dp[idx].addr  = rte_mbuf_data_iova(cookie);
 588                 start_dp[idx].len   = cookie->data_len;
 589                 if (prepend_header) {
 590                         start_dp[idx].addr -= head_size;
 591                         start_dp[idx].len += head_size;
 592                         prepend_header = false;
 593                 }
 594                 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 595                 idx = start_dp[idx].next;
 596         } while ((cookie = cookie->next) != NULL);
 597
 598         if (use_indirect)
 599                 idx = vq->vq_split.ring.desc[head_idx].next;
 600
 601         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 602
 603         vq->vq_desc_head_idx = idx;
 604         vq_update_avail_ring(vq, head_idx);
 605
 606         if (!in_order) {
 607                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 608                         vq->vq_desc_tail_idx = idx;
 609         }
 610 }
 611
 612 void
 613 virtio_dev_cq_start(struct rte_eth_dev *dev)
 614 {
 615         struct virtio_hw *hw = dev->data->dev_private;
 616
 617         if (hw->cvq) {
 618                 rte_spinlock_init(&hw->cvq->lock);
 619                 VIRTQUEUE_DUMP(virtnet_cq_to_vq(hw->cvq));
 620         }
 621 }
 622
 623 int
 624 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
 625                         uint16_t queue_idx,
 626                         uint16_t nb_desc,
 627                         unsigned int socket_id __rte_unused,
 628                         const struct rte_eth_rxconf *rx_conf,
 629                         struct rte_mempool *mp)
 630 {
 631         uint16_t vq_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 632         struct virtio_hw *hw = dev->data->dev_private;
 633         struct virtqueue *vq = hw->vqs[vq_idx];
 634         struct virtnet_rx *rxvq;
 635         uint16_t rx_free_thresh;
 636
 637         PMD_INIT_FUNC_TRACE();
 638
 639         if (rx_conf->rx_deferred_start) {
 640                 PMD_INIT_LOG(ERR, "Rx deferred start is not supported");
 641                 return -EINVAL;
 642         }
 643
 644         rx_free_thresh = rx_conf->rx_free_thresh;
 645         if (rx_free_thresh == 0)
 646                 rx_free_thresh =
 647                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_RX_FREE_THRESH);
 648
 649         if (rx_free_thresh & 0x3) {
 650                 RTE_LOG(ERR, PMD, "rx_free_thresh must be multiples of four."
 651                         " (rx_free_thresh=%u port=%u queue=%u)\n",
 652                         rx_free_thresh, dev->data->port_id, queue_idx);
 653                 return -EINVAL;
 654         }
 655
 656         if (rx_free_thresh >= vq->vq_nentries) {
 657                 RTE_LOG(ERR, PMD, "rx_free_thresh must be less than the "
 658                         "number of RX entries (%u)."
 659                         " (rx_free_thresh=%u port=%u queue=%u)\n",
 660                         vq->vq_nentries,
 661                         rx_free_thresh, dev->data->port_id, queue_idx);
 662                 return -EINVAL;
 663         }
 664         vq->vq_free_thresh = rx_free_thresh;
 665
 666         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 667                 nb_desc = vq->vq_nentries;
 668         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 669
 670         rxvq = &vq->rxq;
 671         rxvq->queue_id = queue_idx;
 672         rxvq->mpool = mp;
 673         dev->data->rx_queues[queue_idx] = rxvq;
 674
 675         return 0;
 676 }
 677
 678 int
 679 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
 680 {
 681         uint16_t vq_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 682         struct virtio_hw *hw = dev->data->dev_private;
 683         struct virtqueue *vq = hw->vqs[vq_idx];
 684         struct virtnet_rx *rxvq = &vq->rxq;
 685         struct rte_mbuf *m;
 686         uint16_t desc_idx;
 687         int error, nbufs, i;
 688         bool in_order = virtio_with_feature(hw, VIRTIO_F_IN_ORDER);
 689
 690         PMD_INIT_FUNC_TRACE();
 691
 692         /* Allocate blank mbufs for the each rx descriptor */
 693         nbufs = 0;
 694
 695         if (hw->use_vec_rx && !virtio_with_packed_queue(hw)) {
 696                 for (desc_idx = 0; desc_idx < vq->vq_nentries;
 697                      desc_idx++) {
 698                         vq->vq_split.ring.avail->ring[desc_idx] = desc_idx;
 699                         vq->vq_split.ring.desc[desc_idx].flags =
 700                                 VRING_DESC_F_WRITE;
 701                 }
 702
 703                 virtio_rxq_vec_setup(rxvq);
 704         }
 705
 706         memset(rxvq->fake_mbuf, 0, sizeof(*rxvq->fake_mbuf));
 707         for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST; desc_idx++)
 708                 vq->sw_ring[vq->vq_nentries + desc_idx] = rxvq->fake_mbuf;
 709
 710         if (hw->use_vec_rx && !virtio_with_packed_queue(hw)) {
 711                 while (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
 712                         virtio_rxq_rearm_vec(rxvq);
 713                         nbufs += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
 714                 }
 715         } else if (!virtio_with_packed_queue(vq->hw) && in_order) {
 716                 if ((!virtqueue_full(vq))) {
 717                         uint16_t free_cnt = vq->vq_free_cnt;
 718                         struct rte_mbuf *pkts[free_cnt];
 719
 720                         if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, pkts,
 721                                 free_cnt)) {
 722                                 error = virtqueue_enqueue_refill_inorder(vq,
 723                                                 pkts,
 724                                                 free_cnt);
 725                                 if (unlikely(error)) {
 726                                         for (i = 0; i < free_cnt; i++)
 727                                                 rte_pktmbuf_free(pkts[i]);
 728                                 }
 729                         }
 730
 731                         nbufs += free_cnt;
 732                         vq_update_avail_idx(vq);
 733                 }
 734         } else {
 735                 while (!virtqueue_full(vq)) {
 736                         m = rte_mbuf_raw_alloc(rxvq->mpool);
 737                         if (m == NULL)
 738                                 break;
 739
 740                         /* Enqueue allocated buffers */
 741                         if (virtio_with_packed_queue(vq->hw))
 742                                 error = virtqueue_enqueue_recv_refill_packed(vq,
 743                                                 &m, 1);
 744                         else
 745                                 error = virtqueue_enqueue_recv_refill(vq,
 746                                                 &m, 1);
 747                         if (error) {
 748                                 rte_pktmbuf_free(m);
 749                                 break;
 750                         }
 751                         nbufs++;
 752                 }
 753
 754                 if (!virtio_with_packed_queue(vq->hw))
 755                         vq_update_avail_idx(vq);
 756         }
 757
 758         PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
 759
 760         VIRTQUEUE_DUMP(vq);
 761
 762         return 0;
 763 }
 764
 765 /*
 766  * struct rte_eth_dev *dev: Used to update dev
 767  * uint16_t nb_desc: Defaults to values read from config space
 768  * unsigned int socket_id: Used to allocate memzone
 769  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
 770  * uint16_t queue_idx: Just used as an index in dev txq list
 771  */
 772 int
 773 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
 774                         uint16_t queue_idx,
 775                         uint16_t nb_desc,
 776                         unsigned int socket_id __rte_unused,
 777                         const struct rte_eth_txconf *tx_conf)
 778 {
 779         uint8_t vq_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 780         struct virtio_hw *hw = dev->data->dev_private;
 781         struct virtqueue *vq = hw->vqs[vq_idx];
 782         struct virtnet_tx *txvq;
 783         uint16_t tx_free_thresh;
 784
 785         PMD_INIT_FUNC_TRACE();
 786
 787         if (tx_conf->tx_deferred_start) {
 788                 PMD_INIT_LOG(ERR, "Tx deferred start is not supported");
 789                 return -EINVAL;
 790         }
 791
 792         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 793                 nb_desc = vq->vq_nentries;
 794         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 795
 796         txvq = &vq->txq;
 797         txvq->queue_id = queue_idx;
 798
 799         tx_free_thresh = tx_conf->tx_free_thresh;
 800         if (tx_free_thresh == 0)
 801                 tx_free_thresh =
 802                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
 803
 804         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
 805                 PMD_DRV_LOG(ERR, "tx_free_thresh must be less than the "
 806                         "number of TX entries minus 3 (%u)."
 807                         " (tx_free_thresh=%u port=%u queue=%u)\n",
 808                         vq->vq_nentries - 3,
 809                         tx_free_thresh, dev->data->port_id, queue_idx);
 810                 return -EINVAL;
 811         }
 812
 813         vq->vq_free_thresh = tx_free_thresh;
 814
 815         dev->data->tx_queues[queue_idx] = txvq;
 816         return 0;
 817 }
 818
 819 int
 820 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
 821                                 uint16_t queue_idx)
 822 {
 823         uint8_t vq_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 824         struct virtio_hw *hw = dev->data->dev_private;
 825         struct virtqueue *vq = hw->vqs[vq_idx];
 826
 827         PMD_INIT_FUNC_TRACE();
 828
 829         if (!virtio_with_packed_queue(hw)) {
 830                 if (virtio_with_feature(hw, VIRTIO_F_IN_ORDER))
 831                         vq->vq_split.ring.desc[vq->vq_nentries - 1].next = 0;
 832         }
 833
 834         VIRTQUEUE_DUMP(vq);
 835
 836         return 0;
 837 }
 838
 839 static inline void
 840 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
 841 {
 842         int error;
 843         /*
 844          * Requeue the discarded mbuf. This should always be
 845          * successful since it was just dequeued.
 846          */
 847         if (virtio_with_packed_queue(vq->hw))
 848                 error = virtqueue_enqueue_recv_refill_packed(vq, &m, 1);
 849         else
 850                 error = virtqueue_enqueue_recv_refill(vq, &m, 1);
 851
 852         if (unlikely(error)) {
 853                 PMD_DRV_LOG(ERR, "cannot requeue discarded mbuf");
 854                 rte_pktmbuf_free(m);
 855         }
 856 }
 857
 858 static inline void
 859 virtio_discard_rxbuf_inorder(struct virtqueue *vq, struct rte_mbuf *m)
 860 {
 861         int error;
 862
 863         error = virtqueue_enqueue_refill_inorder(vq, &m, 1);
 864         if (unlikely(error)) {
 865                 PMD_DRV_LOG(ERR, "cannot requeue discarded mbuf");
 866                 rte_pktmbuf_free(m);
 867         }
 868 }
 869
 870 /* Optionally fill offload information in structure */
 871 static inline int
 872 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
 873 {
 874         struct rte_net_hdr_lens hdr_lens;
 875         uint32_t hdrlen, ptype;
 876         int l4_supported = 0;
 877
 878         /* nothing to do */
 879         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 880                 return 0;
 881
 882         m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
 883
 884         ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
 885         m->packet_type = ptype;
 886         if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
 887             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
 888             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
 889                 l4_supported = 1;
 890
 891         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 892                 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
 893                 if (hdr->csum_start <= hdrlen && l4_supported) {
 894                         m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
 895                 } else {
 896                         /* Unknown proto or tunnel, do sw cksum. We can assume
 897                          * the cksum field is in the first segment since the
 898                          * buffers we provided to the host are large enough.
 899                          * In case of SCTP, this will be wrong since it's a CRC
 900                          * but there's nothing we can do.
 901                          */
 902                         uint16_t csum = 0, off;
 903
 904                         if (rte_raw_cksum_mbuf(m, hdr->csum_start,
 905                                 rte_pktmbuf_pkt_len(m) - hdr->csum_start,
 906                                 &csum) < 0)
 907                                 return -EINVAL;
 908                         if (likely(csum != 0xffff))
 909                                 csum = ~csum;
 910                         off = hdr->csum_offset + hdr->csum_start;
 911                         if (rte_pktmbuf_data_len(m) >= off + 1)
 912                                 *rte_pktmbuf_mtod_offset(m, uint16_t *,
 913                                         off) = csum;
 914                 }
 915         } else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
 916                 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
 917         }
 918
 919         /* GSO request, save required information in mbuf */
 920         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 921                 /* Check unsupported modes */
 922                 if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
 923                     (hdr->gso_size == 0)) {
 924                         return -EINVAL;
 925                 }
 926
 927                 /* Update mss lengthes in mbuf */
 928                 m->tso_segsz = hdr->gso_size;
 929                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 930                         case VIRTIO_NET_HDR_GSO_TCPV4:
 931                         case VIRTIO_NET_HDR_GSO_TCPV6:
 932                                 m->ol_flags |= PKT_RX_LRO | \
 933                                         PKT_RX_L4_CKSUM_NONE;
 934                                 break;
 935                         default:
 936                                 return -EINVAL;
 937                 }
 938         }
 939
 940         return 0;
 941 }
 942
 943 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
 944 uint16_t
 945 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 946 {
 947         struct virtnet_rx *rxvq = rx_queue;
 948         struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
 949         struct virtio_hw *hw = vq->hw;
 950         struct rte_mbuf *rxm;
 951         uint16_t nb_used, num, nb_rx;
 952         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 953         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 954         int error;
 955         uint32_t i, nb_enqueued;
 956         uint32_t hdr_size;
 957         struct virtio_net_hdr *hdr;
 958
 959         nb_rx = 0;
 960         if (unlikely(hw->started == 0))
 961                 return nb_rx;
 962
 963         nb_used = virtqueue_nused(vq);
 964
 965         num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
 966         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
 967                 num = VIRTIO_MBUF_BURST_SZ;
 968         if (likely(num > DESC_PER_CACHELINE))
 969                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
 970
 971         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
 972         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
 973
 974         nb_enqueued = 0;
 975         hdr_size = hw->vtnet_hdr_size;
 976
 977         for (i = 0; i < num ; i++) {
 978                 rxm = rcv_pkts[i];
 979
 980                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
 981
 982                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
 983                         PMD_RX_LOG(ERR, "Packet drop");
 984                         nb_enqueued++;
 985                         virtio_discard_rxbuf(vq, rxm);
 986                         rxvq->stats.errors++;
 987                         continue;
 988                 }
 989
 990                 rxm->port = rxvq->port_id;
 991                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 992                 rxm->ol_flags = 0;
 993                 rxm->vlan_tci = 0;
 994
 995                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
 996                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
 997
 998                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
 999                         RTE_PKTMBUF_HEADROOM - hdr_size);
1000
1001                 if (hw->vlan_strip)
1002                         rte_vlan_strip(rxm);
1003
1004                 if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1005                         virtio_discard_rxbuf(vq, rxm);
1006                         rxvq->stats.errors++;
1007                         continue;
1008                 }
1009
1010                 virtio_rx_stats_updated(rxvq, rxm);
1011
1012                 rx_pkts[nb_rx++] = rxm;
1013         }
1014
1015         rxvq->stats.packets += nb_rx;
1016
1017         /* Allocate new mbuf for the used descriptor */
1018         if (likely(!virtqueue_full(vq))) {
1019                 uint16_t free_cnt = vq->vq_free_cnt;
1020                 struct rte_mbuf *new_pkts[free_cnt];
1021
1022                 if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1023                                                 free_cnt) == 0)) {
1024                         error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1025                                         free_cnt);
1026                         if (unlikely(error)) {
1027                                 for (i = 0; i < free_cnt; i++)
1028                                         rte_pktmbuf_free(new_pkts[i]);
1029                         }
1030                         nb_enqueued += free_cnt;
1031                 } else {
1032                         struct rte_eth_dev *dev =
1033                                 &rte_eth_devices[rxvq->port_id];
1034                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1035                 }
1036         }
1037
1038         if (likely(nb_enqueued)) {
1039                 vq_update_avail_idx(vq);
1040
1041                 if (unlikely(virtqueue_kick_prepare(vq))) {
1042                         virtqueue_notify(vq);
1043                         PMD_RX_LOG(DEBUG, "Notified");
1044                 }
1045         }
1046
1047         return nb_rx;
1048 }
1049
1050 uint16_t
1051 virtio_recv_pkts_packed(void *rx_queue, struct rte_mbuf **rx_pkts,
1052                         uint16_t nb_pkts)
1053 {
1054         struct virtnet_rx *rxvq = rx_queue;
1055         struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1056         struct virtio_hw *hw = vq->hw;
1057         struct rte_mbuf *rxm;
1058         uint16_t num, nb_rx;
1059         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1060         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1061         int error;
1062         uint32_t i, nb_enqueued;
1063         uint32_t hdr_size;
1064         struct virtio_net_hdr *hdr;
1065
1066         nb_rx = 0;
1067         if (unlikely(hw->started == 0))
1068                 return nb_rx;
1069
1070         num = RTE_MIN(VIRTIO_MBUF_BURST_SZ, nb_pkts);
1071         if (likely(num > DESC_PER_CACHELINE))
1072                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1073
1074         num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1075         PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1076
1077         nb_enqueued = 0;
1078         hdr_size = hw->vtnet_hdr_size;
1079
1080         for (i = 0; i < num; i++) {
1081                 rxm = rcv_pkts[i];
1082
1083                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1084
1085                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1086                         PMD_RX_LOG(ERR, "Packet drop");
1087                         nb_enqueued++;
1088                         virtio_discard_rxbuf(vq, rxm);
1089                         rxvq->stats.errors++;
1090                         continue;
1091                 }
1092
1093                 rxm->port = rxvq->port_id;
1094                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1095                 rxm->ol_flags = 0;
1096                 rxm->vlan_tci = 0;
1097
1098                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1099                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1100
1101                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1102                         RTE_PKTMBUF_HEADROOM - hdr_size);
1103
1104                 if (hw->vlan_strip)
1105                         rte_vlan_strip(rxm);
1106
1107                 if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1108                         virtio_discard_rxbuf(vq, rxm);
1109                         rxvq->stats.errors++;
1110                         continue;
1111                 }
1112
1113                 virtio_rx_stats_updated(rxvq, rxm);
1114
1115                 rx_pkts[nb_rx++] = rxm;
1116         }
1117
1118         rxvq->stats.packets += nb_rx;
1119
1120         /* Allocate new mbuf for the used descriptor */
1121         if (likely(!virtqueue_full(vq))) {
1122                 uint16_t free_cnt = vq->vq_free_cnt;
1123                 struct rte_mbuf *new_pkts[free_cnt];
1124
1125                 if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1126                                                 free_cnt) == 0)) {
1127                         error = virtqueue_enqueue_recv_refill_packed(vq,
1128                                         new_pkts, free_cnt);
1129                         if (unlikely(error)) {
1130                                 for (i = 0; i < free_cnt; i++)
1131                                         rte_pktmbuf_free(new_pkts[i]);
1132                         }
1133                         nb_enqueued += free_cnt;
1134                 } else {
1135                         struct rte_eth_dev *dev =
1136                                 &rte_eth_devices[rxvq->port_id];
1137                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1138                 }
1139         }
1140
1141         if (likely(nb_enqueued)) {
1142                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1143                         virtqueue_notify(vq);
1144                         PMD_RX_LOG(DEBUG, "Notified");
1145                 }
1146         }
1147
1148         return nb_rx;
1149 }
1150
1151
1152 uint16_t
1153 virtio_recv_pkts_inorder(void *rx_queue,
1154                         struct rte_mbuf **rx_pkts,
1155                         uint16_t nb_pkts)
1156 {
1157         struct virtnet_rx *rxvq = rx_queue;
1158         struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1159         struct virtio_hw *hw = vq->hw;
1160         struct rte_mbuf *rxm;
1161         struct rte_mbuf *prev = NULL;
1162         uint16_t nb_used, num, nb_rx;
1163         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1164         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1165         int error;
1166         uint32_t nb_enqueued;
1167         uint32_t seg_num;
1168         uint32_t seg_res;
1169         uint32_t hdr_size;
1170         int32_t i;
1171
1172         nb_rx = 0;
1173         if (unlikely(hw->started == 0))
1174                 return nb_rx;
1175
1176         nb_used = virtqueue_nused(vq);
1177         nb_used = RTE_MIN(nb_used, nb_pkts);
1178         nb_used = RTE_MIN(nb_used, VIRTIO_MBUF_BURST_SZ);
1179
1180         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1181
1182         nb_enqueued = 0;
1183         seg_num = 1;
1184         seg_res = 0;
1185         hdr_size = hw->vtnet_hdr_size;
1186
1187         num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len, nb_used);
1188
1189         for (i = 0; i < num; i++) {
1190                 struct virtio_net_hdr_mrg_rxbuf *header;
1191
1192                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1193                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1194
1195                 rxm = rcv_pkts[i];
1196
1197                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1198                         PMD_RX_LOG(ERR, "Packet drop");
1199                         nb_enqueued++;
1200                         virtio_discard_rxbuf_inorder(vq, rxm);
1201                         rxvq->stats.errors++;
1202                         continue;
1203                 }
1204
1205                 header = (struct virtio_net_hdr_mrg_rxbuf *)
1206                          ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1207                          - hdr_size);
1208
1209                 if (virtio_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
1210                         seg_num = header->num_buffers;
1211                         if (seg_num == 0)
1212                                 seg_num = 1;
1213                 } else {
1214                         seg_num = 1;
1215                 }
1216
1217                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1218                 rxm->nb_segs = seg_num;
1219                 rxm->ol_flags = 0;
1220                 rxm->vlan_tci = 0;
1221                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1222                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1223
1224                 rxm->port = rxvq->port_id;
1225
1226                 rx_pkts[nb_rx] = rxm;
1227                 prev = rxm;
1228
1229                 if (vq->hw->has_rx_offload &&
1230                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1231                         virtio_discard_rxbuf_inorder(vq, rxm);
1232                         rxvq->stats.errors++;
1233                         continue;
1234                 }
1235
1236                 if (hw->vlan_strip)
1237                         rte_vlan_strip(rx_pkts[nb_rx]);
1238
1239                 seg_res = seg_num - 1;
1240
1241                 /* Merge remaining segments */
1242                 while (seg_res != 0 && i < (num - 1)) {
1243                         i++;
1244
1245                         rxm = rcv_pkts[i];
1246                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1247                         rxm->pkt_len = (uint32_t)(len[i]);
1248                         rxm->data_len = (uint16_t)(len[i]);
1249
1250                         rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1251
1252                         prev->next = rxm;
1253                         prev = rxm;
1254                         seg_res -= 1;
1255                 }
1256
1257                 if (!seg_res) {
1258                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1259                         nb_rx++;
1260                 }
1261         }
1262
1263         /* Last packet still need merge segments */
1264         while (seg_res != 0) {
1265                 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1266                                         VIRTIO_MBUF_BURST_SZ);
1267
1268                 if (likely(virtqueue_nused(vq) >= rcv_cnt)) {
1269                         num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len,
1270                                                            rcv_cnt);
1271                         uint16_t extra_idx = 0;
1272
1273                         rcv_cnt = num;
1274                         while (extra_idx < rcv_cnt) {
1275                                 rxm = rcv_pkts[extra_idx];
1276                                 rxm->data_off =
1277                                         RTE_PKTMBUF_HEADROOM - hdr_size;
1278                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1279                                 rxm->data_len = (uint16_t)(len[extra_idx]);
1280                                 prev->next = rxm;
1281                                 prev = rxm;
1282                                 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1283                                 extra_idx += 1;
1284                         };
1285                         seg_res -= rcv_cnt;
1286
1287                         if (!seg_res) {
1288                                 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1289                                 nb_rx++;
1290                         }
1291                 } else {
1292                         PMD_RX_LOG(ERR,
1293                                         "No enough segments for packet.");
1294                         rte_pktmbuf_free(rx_pkts[nb_rx]);
1295                         rxvq->stats.errors++;
1296                         break;
1297                 }
1298         }
1299
1300         rxvq->stats.packets += nb_rx;
1301
1302         /* Allocate new mbuf for the used descriptor */
1303
1304         if (likely(!virtqueue_full(vq))) {
1305                 /* free_cnt may include mrg descs */
1306                 uint16_t free_cnt = vq->vq_free_cnt;
1307                 struct rte_mbuf *new_pkts[free_cnt];
1308
1309                 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1310                         error = virtqueue_enqueue_refill_inorder(vq, new_pkts,
1311                                         free_cnt);
1312                         if (unlikely(error)) {
1313                                 for (i = 0; i < free_cnt; i++)
1314                                         rte_pktmbuf_free(new_pkts[i]);
1315                         }
1316                         nb_enqueued += free_cnt;
1317                 } else {
1318                         struct rte_eth_dev *dev =
1319                                 &rte_eth_devices[rxvq->port_id];
1320                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1321                 }
1322         }
1323
1324         if (likely(nb_enqueued)) {
1325                 vq_update_avail_idx(vq);
1326
1327                 if (unlikely(virtqueue_kick_prepare(vq))) {
1328                         virtqueue_notify(vq);
1329                         PMD_RX_LOG(DEBUG, "Notified");
1330                 }
1331         }
1332
1333         return nb_rx;
1334 }
1335
1336 uint16_t
1337 virtio_recv_mergeable_pkts(void *rx_queue,
1338                         struct rte_mbuf **rx_pkts,
1339                         uint16_t nb_pkts)
1340 {
1341         struct virtnet_rx *rxvq = rx_queue;
1342         struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1343         struct virtio_hw *hw = vq->hw;
1344         struct rte_mbuf *rxm;
1345         struct rte_mbuf *prev = NULL;
1346         uint16_t nb_used, num, nb_rx = 0;
1347         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1348         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1349         int error;
1350         uint32_t nb_enqueued = 0;
1351         uint32_t seg_num = 0;
1352         uint32_t seg_res = 0;
1353         uint32_t hdr_size = hw->vtnet_hdr_size;
1354         int32_t i;
1355
1356         if (unlikely(hw->started == 0))
1357                 return nb_rx;
1358
1359         nb_used = virtqueue_nused(vq);
1360
1361         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1362
1363         num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1364         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1365                 num = VIRTIO_MBUF_BURST_SZ;
1366         if (likely(num > DESC_PER_CACHELINE))
1367                 num = num - ((vq->vq_used_cons_idx + num) %
1368                                 DESC_PER_CACHELINE);
1369
1370
1371         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1372
1373         for (i = 0; i < num; i++) {
1374                 struct virtio_net_hdr_mrg_rxbuf *header;
1375
1376                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1377                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1378
1379                 rxm = rcv_pkts[i];
1380
1381                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1382                         PMD_RX_LOG(ERR, "Packet drop");
1383                         nb_enqueued++;
1384                         virtio_discard_rxbuf(vq, rxm);
1385                         rxvq->stats.errors++;
1386                         continue;
1387                 }
1388
1389                 header = (struct virtio_net_hdr_mrg_rxbuf *)
1390                          ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1391                          - hdr_size);
1392                 seg_num = header->num_buffers;
1393                 if (seg_num == 0)
1394                         seg_num = 1;
1395
1396                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1397                 rxm->nb_segs = seg_num;
1398                 rxm->ol_flags = 0;
1399                 rxm->vlan_tci = 0;
1400                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1401                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1402
1403                 rxm->port = rxvq->port_id;
1404
1405                 rx_pkts[nb_rx] = rxm;
1406                 prev = rxm;
1407
1408                 if (hw->has_rx_offload &&
1409                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1410                         virtio_discard_rxbuf(vq, rxm);
1411                         rxvq->stats.errors++;
1412                         continue;
1413                 }
1414
1415                 if (hw->vlan_strip)
1416                         rte_vlan_strip(rx_pkts[nb_rx]);
1417
1418                 seg_res = seg_num - 1;
1419
1420                 /* Merge remaining segments */
1421                 while (seg_res != 0 && i < (num - 1)) {
1422                         i++;
1423
1424                         rxm = rcv_pkts[i];
1425                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1426                         rxm->pkt_len = (uint32_t)(len[i]);
1427                         rxm->data_len = (uint16_t)(len[i]);
1428
1429                         rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1430
1431                         prev->next = rxm;
1432                         prev = rxm;
1433                         seg_res -= 1;
1434                 }
1435
1436                 if (!seg_res) {
1437                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1438                         nb_rx++;
1439                 }
1440         }
1441
1442         /* Last packet still need merge segments */
1443         while (seg_res != 0) {
1444                 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1445                                         VIRTIO_MBUF_BURST_SZ);
1446
1447                 if (likely(virtqueue_nused(vq) >= rcv_cnt)) {
1448                         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len,
1449                                                            rcv_cnt);
1450                         uint16_t extra_idx = 0;
1451
1452                         rcv_cnt = num;
1453                         while (extra_idx < rcv_cnt) {
1454                                 rxm = rcv_pkts[extra_idx];
1455                                 rxm->data_off =
1456                                         RTE_PKTMBUF_HEADROOM - hdr_size;
1457                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1458                                 rxm->data_len = (uint16_t)(len[extra_idx]);
1459                                 prev->next = rxm;
1460                                 prev = rxm;
1461                                 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1462                                 extra_idx += 1;
1463                         };
1464                         seg_res -= rcv_cnt;
1465
1466                         if (!seg_res) {
1467                                 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1468                                 nb_rx++;
1469                         }
1470                 } else {
1471                         PMD_RX_LOG(ERR,
1472                                         "No enough segments for packet.");
1473                         rte_pktmbuf_free(rx_pkts[nb_rx]);
1474                         rxvq->stats.errors++;
1475                         break;
1476                 }
1477         }
1478
1479         rxvq->stats.packets += nb_rx;
1480
1481         /* Allocate new mbuf for the used descriptor */
1482         if (likely(!virtqueue_full(vq))) {
1483                 /* free_cnt may include mrg descs */
1484                 uint16_t free_cnt = vq->vq_free_cnt;
1485                 struct rte_mbuf *new_pkts[free_cnt];
1486
1487                 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1488                         error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1489                                         free_cnt);
1490                         if (unlikely(error)) {
1491                                 for (i = 0; i < free_cnt; i++)
1492                                         rte_pktmbuf_free(new_pkts[i]);
1493                         }
1494                         nb_enqueued += free_cnt;
1495                 } else {
1496                         struct rte_eth_dev *dev =
1497                                 &rte_eth_devices[rxvq->port_id];
1498                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1499                 }
1500         }
1501
1502         if (likely(nb_enqueued)) {
1503                 vq_update_avail_idx(vq);
1504
1505                 if (unlikely(virtqueue_kick_prepare(vq))) {
1506                         virtqueue_notify(vq);
1507                         PMD_RX_LOG(DEBUG, "Notified");
1508                 }
1509         }
1510
1511         return nb_rx;
1512 }
1513
1514 uint16_t
1515 virtio_recv_mergeable_pkts_packed(void *rx_queue,
1516                         struct rte_mbuf **rx_pkts,
1517                         uint16_t nb_pkts)
1518 {
1519         struct virtnet_rx *rxvq = rx_queue;
1520         struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1521         struct virtio_hw *hw = vq->hw;
1522         struct rte_mbuf *rxm;
1523         struct rte_mbuf *prev = NULL;
1524         uint16_t num, nb_rx = 0;
1525         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1526         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1527         uint32_t nb_enqueued = 0;
1528         uint32_t seg_num = 0;
1529         uint32_t seg_res = 0;
1530         uint32_t hdr_size = hw->vtnet_hdr_size;
1531         int32_t i;
1532         int error;
1533
1534         if (unlikely(hw->started == 0))
1535                 return nb_rx;
1536
1537
1538         num = nb_pkts;
1539         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1540                 num = VIRTIO_MBUF_BURST_SZ;
1541         if (likely(num > DESC_PER_CACHELINE))
1542                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1543
1544         num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1545
1546         for (i = 0; i < num; i++) {
1547                 struct virtio_net_hdr_mrg_rxbuf *header;
1548
1549                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1550                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1551
1552                 rxm = rcv_pkts[i];
1553
1554                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1555                         PMD_RX_LOG(ERR, "Packet drop");
1556                         nb_enqueued++;
1557                         virtio_discard_rxbuf(vq, rxm);
1558                         rxvq->stats.errors++;
1559                         continue;
1560                 }
1561
1562                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)
1563                           rxm->buf_addr + RTE_PKTMBUF_HEADROOM - hdr_size);
1564                 seg_num = header->num_buffers;
1565
1566                 if (seg_num == 0)
1567                         seg_num = 1;
1568
1569                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1570                 rxm->nb_segs = seg_num;
1571                 rxm->ol_flags = 0;
1572                 rxm->vlan_tci = 0;
1573                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1574                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1575
1576                 rxm->port = rxvq->port_id;
1577                 rx_pkts[nb_rx] = rxm;
1578                 prev = rxm;
1579
1580                 if (hw->has_rx_offload &&
1581                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1582                         virtio_discard_rxbuf(vq, rxm);
1583                         rxvq->stats.errors++;
1584                         continue;
1585                 }
1586
1587                 if (hw->vlan_strip)
1588                         rte_vlan_strip(rx_pkts[nb_rx]);
1589
1590                 seg_res = seg_num - 1;
1591
1592                 /* Merge remaining segments */
1593                 while (seg_res != 0 && i < (num - 1)) {
1594                         i++;
1595
1596                         rxm = rcv_pkts[i];
1597                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1598                         rxm->pkt_len = (uint32_t)(len[i]);
1599                         rxm->data_len = (uint16_t)(len[i]);
1600
1601                         rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1602
1603                         prev->next = rxm;
1604                         prev = rxm;
1605                         seg_res -= 1;
1606                 }
1607
1608                 if (!seg_res) {
1609                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1610                         nb_rx++;
1611                 }
1612         }
1613
1614         /* Last packet still need merge segments */
1615         while (seg_res != 0) {
1616                 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1617                                         VIRTIO_MBUF_BURST_SZ);
1618                 uint16_t extra_idx = 0;
1619
1620                 rcv_cnt = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts,
1621                                 len, rcv_cnt);
1622                 if (unlikely(rcv_cnt == 0)) {
1623                         PMD_RX_LOG(ERR, "No enough segments for packet.");
1624                         rte_pktmbuf_free(rx_pkts[nb_rx]);
1625                         rxvq->stats.errors++;
1626                         break;
1627                 }
1628
1629                 while (extra_idx < rcv_cnt) {
1630                         rxm = rcv_pkts[extra_idx];
1631
1632                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1633                         rxm->pkt_len = (uint32_t)(len[extra_idx]);
1634                         rxm->data_len = (uint16_t)(len[extra_idx]);
1635
1636                         prev->next = rxm;
1637                         prev = rxm;
1638                         rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1639                         extra_idx += 1;
1640                 }
1641                 seg_res -= rcv_cnt;
1642                 if (!seg_res) {
1643                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1644                         nb_rx++;
1645                 }
1646         }
1647
1648         rxvq->stats.packets += nb_rx;
1649
1650         /* Allocate new mbuf for the used descriptor */
1651         if (likely(!virtqueue_full(vq))) {
1652                 /* free_cnt may include mrg descs */
1653                 uint16_t free_cnt = vq->vq_free_cnt;
1654                 struct rte_mbuf *new_pkts[free_cnt];
1655
1656                 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1657                         error = virtqueue_enqueue_recv_refill_packed(vq,
1658                                         new_pkts, free_cnt);
1659                         if (unlikely(error)) {
1660                                 for (i = 0; i < free_cnt; i++)
1661                                         rte_pktmbuf_free(new_pkts[i]);
1662                         }
1663                         nb_enqueued += free_cnt;
1664                 } else {
1665                         struct rte_eth_dev *dev =
1666                                 &rte_eth_devices[rxvq->port_id];
1667                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1668                 }
1669         }
1670
1671         if (likely(nb_enqueued)) {
1672                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1673                         virtqueue_notify(vq);
1674                         PMD_RX_LOG(DEBUG, "Notified");
1675                 }
1676         }
1677
1678         return nb_rx;
1679 }
1680
1681 uint16_t
1682 virtio_xmit_pkts_prepare(void *tx_queue __rte_unused, struct rte_mbuf **tx_pkts,
1683                         uint16_t nb_pkts)
1684 {
1685         uint16_t nb_tx;
1686         int error;
1687
1688         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1689                 struct rte_mbuf *m = tx_pkts[nb_tx];
1690
1691 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
1692                 error = rte_validate_tx_offload(m);
1693                 if (unlikely(error)) {
1694                         rte_errno = -error;
1695                         break;
1696                 }
1697 #endif
1698
1699                 /* Do VLAN tag insertion */
1700                 if (unlikely(m->ol_flags & PKT_TX_VLAN_PKT)) {
1701                         error = rte_vlan_insert(&m);
1702                         /* rte_vlan_insert() may change pointer
1703                          * even in the case of failure
1704                          */
1705                         tx_pkts[nb_tx] = m;
1706
1707                         if (unlikely(error)) {
1708                                 rte_errno = -error;
1709                                 break;
1710                         }
1711                 }
1712
1713                 error = rte_net_intel_cksum_prepare(m);
1714                 if (unlikely(error)) {
1715                         rte_errno = -error;
1716                         break;
1717                 }
1718
1719                 if (m->ol_flags & PKT_TX_TCP_SEG)
1720                         virtio_tso_fix_cksum(m);
1721         }
1722
1723         return nb_tx;
1724 }
1725
1726 uint16_t
1727 virtio_xmit_pkts_packed(void *tx_queue, struct rte_mbuf **tx_pkts,
1728                         uint16_t nb_pkts)
1729 {
1730         struct virtnet_tx *txvq = tx_queue;
1731         struct virtqueue *vq = virtnet_txq_to_vq(txvq);
1732         struct virtio_hw *hw = vq->hw;
1733         uint16_t hdr_size = hw->vtnet_hdr_size;
1734         uint16_t nb_tx = 0;
1735         bool in_order = virtio_with_feature(hw, VIRTIO_F_IN_ORDER);
1736
1737         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1738                 return nb_tx;
1739
1740         if (unlikely(nb_pkts < 1))
1741                 return nb_pkts;
1742
1743         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1744
1745         if (nb_pkts > vq->vq_free_cnt)
1746                 virtio_xmit_cleanup_packed(vq, nb_pkts - vq->vq_free_cnt,
1747                                            in_order);
1748
1749         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1750                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1751                 int can_push = 0, use_indirect = 0, slots, need;
1752
1753                 /* optimize ring usage */
1754                 if ((virtio_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1755                       virtio_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1756                     rte_mbuf_refcnt_read(txm) == 1 &&
1757                     RTE_MBUF_DIRECT(txm) &&
1758                     txm->nb_segs == 1 &&
1759                     rte_pktmbuf_headroom(txm) >= hdr_size &&
1760                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1761                            __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1762                         can_push = 1;
1763                 else if (virtio_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1764                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1765                         use_indirect = 1;
1766                 /* How many main ring entries are needed to this Tx?
1767                  * indirect   => 1
1768                  * any_layout => number of segments
1769                  * default    => number of segments + 1
1770                  */
1771                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1772                 need = slots - vq->vq_free_cnt;
1773
1774                 /* Positive value indicates it need free vring descriptors */
1775                 if (unlikely(need > 0)) {
1776                         virtio_xmit_cleanup_packed(vq, need, in_order);
1777                         need = slots - vq->vq_free_cnt;
1778                         if (unlikely(need > 0)) {
1779                                 PMD_TX_LOG(ERR,
1780                                            "No free tx descriptors to transmit");
1781                                 break;
1782                         }
1783                 }
1784
1785                 /* Enqueue Packet buffers */
1786                 if (can_push)
1787                         virtqueue_enqueue_xmit_packed_fast(txvq, txm, in_order);
1788                 else
1789                         virtqueue_enqueue_xmit_packed(txvq, txm, slots,
1790                                                       use_indirect, 0,
1791                                                       in_order);
1792
1793                 virtio_update_packet_stats(&txvq->stats, txm);
1794         }
1795
1796         txvq->stats.packets += nb_tx;
1797
1798         if (likely(nb_tx)) {
1799                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1800                         virtqueue_notify(vq);
1801                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1802                 }
1803         }
1804
1805         return nb_tx;
1806 }
1807
1808 uint16_t
1809 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1810 {
1811         struct virtnet_tx *txvq = tx_queue;
1812         struct virtqueue *vq = virtnet_txq_to_vq(txvq);
1813         struct virtio_hw *hw = vq->hw;
1814         uint16_t hdr_size = hw->vtnet_hdr_size;
1815         uint16_t nb_used, nb_tx = 0;
1816
1817         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1818                 return nb_tx;
1819
1820         if (unlikely(nb_pkts < 1))
1821                 return nb_pkts;
1822
1823         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1824
1825         nb_used = virtqueue_nused(vq);
1826
1827         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1828                 virtio_xmit_cleanup(vq, nb_used);
1829
1830         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1831                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1832                 int can_push = 0, use_indirect = 0, slots, need;
1833
1834                 /* optimize ring usage */
1835                 if ((virtio_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1836                       virtio_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1837                     rte_mbuf_refcnt_read(txm) == 1 &&
1838                     RTE_MBUF_DIRECT(txm) &&
1839                     txm->nb_segs == 1 &&
1840                     rte_pktmbuf_headroom(txm) >= hdr_size &&
1841                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1842                                    __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1843                         can_push = 1;
1844                 else if (virtio_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1845                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1846                         use_indirect = 1;
1847
1848                 /* How many main ring entries are needed to this Tx?
1849                  * any_layout => number of segments
1850                  * indirect   => 1
1851                  * default    => number of segments + 1
1852                  */
1853                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1854                 need = slots - vq->vq_free_cnt;
1855
1856                 /* Positive value indicates it need free vring descriptors */
1857                 if (unlikely(need > 0)) {
1858                         nb_used = virtqueue_nused(vq);
1859
1860                         need = RTE_MIN(need, (int)nb_used);
1861
1862                         virtio_xmit_cleanup(vq, need);
1863                         need = slots - vq->vq_free_cnt;
1864                         if (unlikely(need > 0)) {
1865                                 PMD_TX_LOG(ERR,
1866                                            "No free tx descriptors to transmit");
1867                                 break;
1868                         }
1869                 }
1870
1871                 /* Enqueue Packet buffers */
1872                 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
1873                         can_push, 0);
1874
1875                 virtio_update_packet_stats(&txvq->stats, txm);
1876         }
1877
1878         txvq->stats.packets += nb_tx;
1879
1880         if (likely(nb_tx)) {
1881                 vq_update_avail_idx(vq);
1882
1883                 if (unlikely(virtqueue_kick_prepare(vq))) {
1884                         virtqueue_notify(vq);
1885                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1886                 }
1887         }
1888
1889         return nb_tx;
1890 }
1891
1892 static __rte_always_inline int
1893 virtio_xmit_try_cleanup_inorder(struct virtqueue *vq, uint16_t need)
1894 {
1895         uint16_t nb_used, nb_clean, nb_descs;
1896
1897         nb_descs = vq->vq_free_cnt + need;
1898         nb_used = virtqueue_nused(vq);
1899         nb_clean = RTE_MIN(need, (int)nb_used);
1900
1901         virtio_xmit_cleanup_inorder(vq, nb_clean);
1902
1903         return nb_descs - vq->vq_free_cnt;
1904 }
1905
1906 uint16_t
1907 virtio_xmit_pkts_inorder(void *tx_queue,
1908                         struct rte_mbuf **tx_pkts,
1909                         uint16_t nb_pkts)
1910 {
1911         struct virtnet_tx *txvq = tx_queue;
1912         struct virtqueue *vq = virtnet_txq_to_vq(txvq);
1913         struct virtio_hw *hw = vq->hw;
1914         uint16_t hdr_size = hw->vtnet_hdr_size;
1915         uint16_t nb_used, nb_tx = 0, nb_inorder_pkts = 0;
1916         struct rte_mbuf *inorder_pkts[nb_pkts];
1917         int need;
1918
1919         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1920                 return nb_tx;
1921
1922         if (unlikely(nb_pkts < 1))
1923                 return nb_pkts;
1924
1925         VIRTQUEUE_DUMP(vq);
1926         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1927         nb_used = virtqueue_nused(vq);
1928
1929         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1930                 virtio_xmit_cleanup_inorder(vq, nb_used);
1931
1932         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1933                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1934                 int slots;
1935
1936                 /* optimize ring usage */
1937                 if ((virtio_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1938                      virtio_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1939                      rte_mbuf_refcnt_read(txm) == 1 &&
1940                      RTE_MBUF_DIRECT(txm) &&
1941                      txm->nb_segs == 1 &&
1942                      rte_pktmbuf_headroom(txm) >= hdr_size &&
1943                      rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1944                                 __alignof__(struct virtio_net_hdr_mrg_rxbuf))) {
1945                         inorder_pkts[nb_inorder_pkts] = txm;
1946                         nb_inorder_pkts++;
1947
1948                         continue;
1949                 }
1950
1951                 if (nb_inorder_pkts) {
1952                         need = nb_inorder_pkts - vq->vq_free_cnt;
1953                         if (unlikely(need > 0)) {
1954                                 need = virtio_xmit_try_cleanup_inorder(vq,
1955                                                                        need);
1956                                 if (unlikely(need > 0)) {
1957                                         PMD_TX_LOG(ERR,
1958                                                 "No free tx descriptors to "
1959                                                 "transmit");
1960                                         break;
1961                                 }
1962                         }
1963                         virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
1964                                                         nb_inorder_pkts);
1965                         nb_inorder_pkts = 0;
1966                 }
1967
1968                 slots = txm->nb_segs + 1;
1969                 need = slots - vq->vq_free_cnt;
1970                 if (unlikely(need > 0)) {
1971                         need = virtio_xmit_try_cleanup_inorder(vq, slots);
1972
1973                         if (unlikely(need > 0)) {
1974                                 PMD_TX_LOG(ERR,
1975                                         "No free tx descriptors to transmit");
1976                                 break;
1977                         }
1978                 }
1979                 /* Enqueue Packet buffers */
1980                 virtqueue_enqueue_xmit(txvq, txm, slots, 0, 0, 1);
1981
1982                 virtio_update_packet_stats(&txvq->stats, txm);
1983         }
1984
1985         /* Transmit all inorder packets */
1986         if (nb_inorder_pkts) {
1987                 need = nb_inorder_pkts - vq->vq_free_cnt;
1988                 if (unlikely(need > 0)) {
1989                         need = virtio_xmit_try_cleanup_inorder(vq,
1990                                                                   need);
1991                         if (unlikely(need > 0)) {
1992                                 PMD_TX_LOG(ERR,
1993                                         "No free tx descriptors to transmit");
1994                                 nb_inorder_pkts = vq->vq_free_cnt;
1995                                 nb_tx -= need;
1996                         }
1997                 }
1998
1999                 virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2000                                                 nb_inorder_pkts);
2001         }
2002
2003         txvq->stats.packets += nb_tx;
2004
2005         if (likely(nb_tx)) {
2006                 vq_update_avail_idx(vq);
2007
2008                 if (unlikely(virtqueue_kick_prepare(vq))) {
2009                         virtqueue_notify(vq);
2010                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2011                 }
2012         }
2013
2014         VIRTQUEUE_DUMP(vq);
2015
2016         return nb_tx;
2017 }
2018
2019 __rte_weak uint16_t
2020 virtio_recv_pkts_packed_vec(void *rx_queue __rte_unused,
2021                             struct rte_mbuf **rx_pkts __rte_unused,
2022                             uint16_t nb_pkts __rte_unused)
2023 {
2024         return 0;
2025 }
2026
2027 __rte_weak uint16_t
2028 virtio_xmit_pkts_packed_vec(void *tx_queue __rte_unused,
2029                             struct rte_mbuf **tx_pkts __rte_unused,
2030                             uint16_t nb_pkts __rte_unused)
2031 {
2032         return 0;
2033 }