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 <rte_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_pci.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 = rxvq->vq;
  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_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_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_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 static void
  93 vq_ring_free_id_packed(struct virtqueue *vq, uint16_t id)
  94 {
  95         struct vq_desc_extra *dxp;
  96
  97         dxp = &vq->vq_descx[id];
  98         vq->vq_free_cnt += dxp->ndescs;
  99
 100         if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END)
 101                 vq->vq_desc_head_idx = id;
 102         else
 103                 vq->vq_descx[vq->vq_desc_tail_idx].next = id;
 104
 105         vq->vq_desc_tail_idx = id;
 106         dxp->next = VQ_RING_DESC_CHAIN_END;
 107 }
 108
 109 static uint16_t
 110 virtqueue_dequeue_burst_rx_packed(struct virtqueue *vq,
 111                                   struct rte_mbuf **rx_pkts,
 112                                   uint32_t *len,
 113                                   uint16_t num)
 114 {
 115         struct rte_mbuf *cookie;
 116         uint16_t used_idx;
 117         uint16_t id;
 118         struct vring_packed_desc *desc;
 119         uint16_t i;
 120
 121         desc = vq->ring_packed.desc_packed;
 122
 123         for (i = 0; i < num; i++) {
 124                 used_idx = vq->vq_used_cons_idx;
 125                 if (!desc_is_used(&desc[used_idx], vq))
 126                         return i;
 127                 len[i] = desc[used_idx].len;
 128                 id = desc[used_idx].id;
 129                 cookie = (struct rte_mbuf *)vq->vq_descx[id].cookie;
 130                 if (unlikely(cookie == NULL)) {
 131                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
 132                                 vq->vq_used_cons_idx);
 133                         break;
 134                 }
 135                 rte_prefetch0(cookie);
 136                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 137                 rx_pkts[i] = cookie;
 138
 139                 vq->vq_free_cnt++;
 140                 vq->vq_used_cons_idx++;
 141                 if (vq->vq_used_cons_idx >= vq->vq_nentries) {
 142                         vq->vq_used_cons_idx -= vq->vq_nentries;
 143                         vq->used_wrap_counter ^= 1;
 144                 }
 145         }
 146
 147         return i;
 148 }
 149
 150 static uint16_t
 151 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
 152                            uint32_t *len, uint16_t num)
 153 {
 154         struct vring_used_elem *uep;
 155         struct rte_mbuf *cookie;
 156         uint16_t used_idx, desc_idx;
 157         uint16_t i;
 158
 159         /*  Caller does the check */
 160         for (i = 0; i < num ; i++) {
 161                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 162                 uep = &vq->vq_ring.used->ring[used_idx];
 163                 desc_idx = (uint16_t) uep->id;
 164                 len[i] = uep->len;
 165                 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
 166
 167                 if (unlikely(cookie == NULL)) {
 168                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
 169                                 vq->vq_used_cons_idx);
 170                         break;
 171                 }
 172
 173                 rte_prefetch0(cookie);
 174                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 175                 rx_pkts[i]  = cookie;
 176                 vq->vq_used_cons_idx++;
 177                 vq_ring_free_chain(vq, desc_idx);
 178                 vq->vq_descx[desc_idx].cookie = NULL;
 179         }
 180
 181         return i;
 182 }
 183
 184 static uint16_t
 185 virtqueue_dequeue_rx_inorder(struct virtqueue *vq,
 186                         struct rte_mbuf **rx_pkts,
 187                         uint32_t *len,
 188                         uint16_t num)
 189 {
 190         struct vring_used_elem *uep;
 191         struct rte_mbuf *cookie;
 192         uint16_t used_idx = 0;
 193         uint16_t i;
 194
 195         if (unlikely(num == 0))
 196                 return 0;
 197
 198         for (i = 0; i < num; i++) {
 199                 used_idx = vq->vq_used_cons_idx & (vq->vq_nentries - 1);
 200                 /* Desc idx same as used idx */
 201                 uep = &vq->vq_ring.used->ring[used_idx];
 202                 len[i] = uep->len;
 203                 cookie = (struct rte_mbuf *)vq->vq_descx[used_idx].cookie;
 204
 205                 if (unlikely(cookie == NULL)) {
 206                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
 207                                 vq->vq_used_cons_idx);
 208                         break;
 209                 }
 210
 211                 rte_prefetch0(cookie);
 212                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 213                 rx_pkts[i]  = cookie;
 214                 vq->vq_used_cons_idx++;
 215                 vq->vq_descx[used_idx].cookie = NULL;
 216         }
 217
 218         vq_ring_free_inorder(vq, used_idx, i);
 219         return i;
 220 }
 221
 222 #ifndef DEFAULT_TX_FREE_THRESH
 223 #define DEFAULT_TX_FREE_THRESH 32
 224 #endif
 225
 226 /* Cleanup from completed transmits. */
 227 static void
 228 virtio_xmit_cleanup_packed(struct virtqueue *vq, int num)
 229 {
 230         uint16_t used_idx, id;
 231         uint16_t size = vq->vq_nentries;
 232         struct vring_packed_desc *desc = vq->ring_packed.desc_packed;
 233         struct vq_desc_extra *dxp;
 234
 235         used_idx = vq->vq_used_cons_idx;
 236         while (num-- && desc_is_used(&desc[used_idx], vq)) {
 237                 used_idx = vq->vq_used_cons_idx;
 238                 id = desc[used_idx].id;
 239                 dxp = &vq->vq_descx[id];
 240                 vq->vq_used_cons_idx += dxp->ndescs;
 241                 if (vq->vq_used_cons_idx >= size) {
 242                         vq->vq_used_cons_idx -= size;
 243                         vq->used_wrap_counter ^= 1;
 244                 }
 245                 vq_ring_free_id_packed(vq, id);
 246                 if (dxp->cookie != NULL) {
 247                         rte_pktmbuf_free(dxp->cookie);
 248                         dxp->cookie = NULL;
 249                 }
 250                 used_idx = vq->vq_used_cons_idx;
 251         }
 252 }
 253
 254 static void
 255 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
 256 {
 257         uint16_t i, used_idx, desc_idx;
 258         for (i = 0; i < num; i++) {
 259                 struct vring_used_elem *uep;
 260                 struct vq_desc_extra *dxp;
 261
 262                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 263                 uep = &vq->vq_ring.used->ring[used_idx];
 264
 265                 desc_idx = (uint16_t) uep->id;
 266                 dxp = &vq->vq_descx[desc_idx];
 267                 vq->vq_used_cons_idx++;
 268                 vq_ring_free_chain(vq, desc_idx);
 269
 270                 if (dxp->cookie != NULL) {
 271                         rte_pktmbuf_free(dxp->cookie);
 272                         dxp->cookie = NULL;
 273                 }
 274         }
 275 }
 276
 277 /* Cleanup from completed inorder transmits. */
 278 static void
 279 virtio_xmit_cleanup_inorder(struct virtqueue *vq, uint16_t num)
 280 {
 281         uint16_t i, used_idx, desc_idx = 0, last_idx;
 282         int16_t free_cnt = 0;
 283         struct vq_desc_extra *dxp = NULL;
 284
 285         if (unlikely(num == 0))
 286                 return;
 287
 288         for (i = 0; i < num; i++) {
 289                 struct vring_used_elem *uep;
 290
 291                 used_idx = vq->vq_used_cons_idx & (vq->vq_nentries - 1);
 292                 uep = &vq->vq_ring.used->ring[used_idx];
 293                 desc_idx = (uint16_t)uep->id;
 294
 295                 dxp = &vq->vq_descx[desc_idx];
 296                 vq->vq_used_cons_idx++;
 297
 298                 if (dxp->cookie != NULL) {
 299                         rte_pktmbuf_free(dxp->cookie);
 300                         dxp->cookie = NULL;
 301                 }
 302         }
 303
 304         last_idx = desc_idx + dxp->ndescs - 1;
 305         free_cnt = last_idx - vq->vq_desc_tail_idx;
 306         if (free_cnt <= 0)
 307                 free_cnt += vq->vq_nentries;
 308
 309         vq_ring_free_inorder(vq, last_idx, free_cnt);
 310 }
 311
 312 static inline int
 313 virtqueue_enqueue_refill_inorder(struct virtqueue *vq,
 314                         struct rte_mbuf **cookies,
 315                         uint16_t num)
 316 {
 317         struct vq_desc_extra *dxp;
 318         struct virtio_hw *hw = vq->hw;
 319         struct vring_desc *start_dp;
 320         uint16_t head_idx, idx, i = 0;
 321
 322         if (unlikely(vq->vq_free_cnt == 0))
 323                 return -ENOSPC;
 324         if (unlikely(vq->vq_free_cnt < num))
 325                 return -EMSGSIZE;
 326
 327         head_idx = vq->vq_desc_head_idx & (vq->vq_nentries - 1);
 328         start_dp = vq->vq_ring.desc;
 329
 330         while (i < num) {
 331                 idx = head_idx & (vq->vq_nentries - 1);
 332                 dxp = &vq->vq_descx[idx];
 333                 dxp->cookie = (void *)cookies[i];
 334                 dxp->ndescs = 1;
 335
 336                 start_dp[idx].addr =
 337                                 VIRTIO_MBUF_ADDR(cookies[i], vq) +
 338                                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 339                 start_dp[idx].len =
 340                                 cookies[i]->buf_len -
 341                                 RTE_PKTMBUF_HEADROOM +
 342                                 hw->vtnet_hdr_size;
 343                 start_dp[idx].flags =  VRING_DESC_F_WRITE;
 344
 345                 vq_update_avail_ring(vq, idx);
 346                 head_idx++;
 347                 i++;
 348         }
 349
 350         vq->vq_desc_head_idx += num;
 351         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 352         return 0;
 353 }
 354
 355 static inline int
 356 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
 357 {
 358         struct vq_desc_extra *dxp;
 359         struct virtio_hw *hw = vq->hw;
 360         struct vring_desc *start_dp;
 361         uint16_t needed = 1;
 362         uint16_t head_idx, idx;
 363
 364         if (unlikely(vq->vq_free_cnt == 0))
 365                 return -ENOSPC;
 366         if (unlikely(vq->vq_free_cnt < needed))
 367                 return -EMSGSIZE;
 368
 369         head_idx = vq->vq_desc_head_idx;
 370         if (unlikely(head_idx >= vq->vq_nentries))
 371                 return -EFAULT;
 372
 373         idx = head_idx;
 374         dxp = &vq->vq_descx[idx];
 375         dxp->cookie = (void *)cookie;
 376         dxp->ndescs = needed;
 377
 378         start_dp = vq->vq_ring.desc;
 379         start_dp[idx].addr =
 380                 VIRTIO_MBUF_ADDR(cookie, vq) +
 381                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 382         start_dp[idx].len =
 383                 cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
 384         start_dp[idx].flags =  VRING_DESC_F_WRITE;
 385         idx = start_dp[idx].next;
 386         vq->vq_desc_head_idx = idx;
 387         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 388                 vq->vq_desc_tail_idx = idx;
 389         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 390         vq_update_avail_ring(vq, head_idx);
 391
 392         return 0;
 393 }
 394
 395 static inline int
 396 virtqueue_enqueue_recv_refill_packed(struct virtqueue *vq,
 397                                      struct rte_mbuf **cookie, uint16_t num)
 398 {
 399         struct vring_packed_desc *start_dp = vq->ring_packed.desc_packed;
 400         uint16_t flags = VRING_DESC_F_WRITE | vq->avail_used_flags;
 401         struct virtio_hw *hw = vq->hw;
 402         struct vq_desc_extra *dxp;
 403         uint16_t idx;
 404         int i;
 405
 406         if (unlikely(vq->vq_free_cnt == 0))
 407                 return -ENOSPC;
 408         if (unlikely(vq->vq_free_cnt < num))
 409                 return -EMSGSIZE;
 410
 411         for (i = 0; i < num; i++) {
 412                 idx = vq->vq_avail_idx;
 413                 dxp = &vq->vq_descx[idx];
 414                 dxp->cookie = (void *)cookie[i];
 415                 dxp->ndescs = 1;
 416
 417                 start_dp[idx].addr = VIRTIO_MBUF_ADDR(cookie[i], vq) +
 418                                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 419                 start_dp[idx].len = cookie[i]->buf_len - RTE_PKTMBUF_HEADROOM
 420                                         + hw->vtnet_hdr_size;
 421
 422                 vq->vq_desc_head_idx = dxp->next;
 423                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 424                         vq->vq_desc_tail_idx = vq->vq_desc_head_idx;
 425                 rte_smp_wmb();
 426                 start_dp[idx].flags = flags;
 427                 if (++vq->vq_avail_idx >= vq->vq_nentries) {
 428                         vq->vq_avail_idx -= vq->vq_nentries;
 429                         vq->avail_wrap_counter ^= 1;
 430                         vq->avail_used_flags =
 431                                 VRING_DESC_F_AVAIL(vq->avail_wrap_counter) |
 432                                 VRING_DESC_F_USED(!vq->avail_wrap_counter);
 433                         flags = VRING_DESC_F_WRITE | vq->avail_used_flags;
 434                 }
 435         }
 436         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 437         return 0;
 438 }
 439
 440 /* When doing TSO, the IP length is not included in the pseudo header
 441  * checksum of the packet given to the PMD, but for virtio it is
 442  * expected.
 443  */
 444 static void
 445 virtio_tso_fix_cksum(struct rte_mbuf *m)
 446 {
 447         /* common case: header is not fragmented */
 448         if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
 449                         m->l4_len)) {
 450                 struct ipv4_hdr *iph;
 451                 struct ipv6_hdr *ip6h;
 452                 struct tcp_hdr *th;
 453                 uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
 454                 uint32_t tmp;
 455
 456                 iph = rte_pktmbuf_mtod_offset(m, struct ipv4_hdr *, m->l2_len);
 457                 th = RTE_PTR_ADD(iph, m->l3_len);
 458                 if ((iph->version_ihl >> 4) == 4) {
 459                         iph->hdr_checksum = 0;
 460                         iph->hdr_checksum = rte_ipv4_cksum(iph);
 461                         ip_len = iph->total_length;
 462                         ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
 463                                 m->l3_len);
 464                 } else {
 465                         ip6h = (struct ipv6_hdr *)iph;
 466                         ip_paylen = ip6h->payload_len;
 467                 }
 468
 469                 /* calculate the new phdr checksum not including ip_paylen */
 470                 prev_cksum = th->cksum;
 471                 tmp = prev_cksum;
 472                 tmp += ip_paylen;
 473                 tmp = (tmp & 0xffff) + (tmp >> 16);
 474                 new_cksum = tmp;
 475
 476                 /* replace it in the packet */
 477                 th->cksum = new_cksum;
 478         }
 479 }
 480
 481
 482 /* avoid write operation when necessary, to lessen cache issues */
 483 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 484         if ((var) != (val))                     \
 485                 (var) = (val);                  \
 486 } while (0)
 487
 488 static inline void
 489 virtqueue_xmit_offload(struct virtio_net_hdr *hdr,
 490                         struct rte_mbuf *cookie,
 491                         bool offload)
 492 {
 493         if (offload) {
 494                 if (cookie->ol_flags & PKT_TX_TCP_SEG)
 495                         cookie->ol_flags |= PKT_TX_TCP_CKSUM;
 496
 497                 switch (cookie->ol_flags & PKT_TX_L4_MASK) {
 498                 case PKT_TX_UDP_CKSUM:
 499                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 500                         hdr->csum_offset = offsetof(struct udp_hdr,
 501                                 dgram_cksum);
 502                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 503                         break;
 504
 505                 case PKT_TX_TCP_CKSUM:
 506                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 507                         hdr->csum_offset = offsetof(struct tcp_hdr, cksum);
 508                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 509                         break;
 510
 511                 default:
 512                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 513                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 514                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 515                         break;
 516                 }
 517
 518                 /* TCP Segmentation Offload */
 519                 if (cookie->ol_flags & PKT_TX_TCP_SEG) {
 520                         virtio_tso_fix_cksum(cookie);
 521                         hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
 522                                 VIRTIO_NET_HDR_GSO_TCPV6 :
 523                                 VIRTIO_NET_HDR_GSO_TCPV4;
 524                         hdr->gso_size = cookie->tso_segsz;
 525                         hdr->hdr_len =
 526                                 cookie->l2_len +
 527                                 cookie->l3_len +
 528                                 cookie->l4_len;
 529                 } else {
 530                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 531                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 532                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 533                 }
 534         }
 535 }
 536
 537 static inline void
 538 virtqueue_enqueue_xmit_inorder(struct virtnet_tx *txvq,
 539                         struct rte_mbuf **cookies,
 540                         uint16_t num)
 541 {
 542         struct vq_desc_extra *dxp;
 543         struct virtqueue *vq = txvq->vq;
 544         struct vring_desc *start_dp;
 545         struct virtio_net_hdr *hdr;
 546         uint16_t idx;
 547         uint16_t head_size = vq->hw->vtnet_hdr_size;
 548         uint16_t i = 0;
 549
 550         idx = vq->vq_desc_head_idx;
 551         start_dp = vq->vq_ring.desc;
 552
 553         while (i < num) {
 554                 idx = idx & (vq->vq_nentries - 1);
 555                 dxp = &vq->vq_descx[idx];
 556                 dxp->cookie = (void *)cookies[i];
 557                 dxp->ndescs = 1;
 558
 559                 hdr = (struct virtio_net_hdr *)
 560                         rte_pktmbuf_prepend(cookies[i], head_size);
 561                 cookies[i]->pkt_len -= head_size;
 562
 563                 /* if offload disabled, it is not zeroed below, do it now */
 564                 if (!vq->hw->has_tx_offload) {
 565                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 566                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 567                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 568                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 569                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 570                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 571                 }
 572
 573                 virtqueue_xmit_offload(hdr, cookies[i],
 574                                 vq->hw->has_tx_offload);
 575
 576                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookies[i], vq);
 577                 start_dp[idx].len   = cookies[i]->data_len;
 578                 start_dp[idx].flags = 0;
 579
 580                 vq_update_avail_ring(vq, idx);
 581
 582                 idx++;
 583                 i++;
 584         };
 585
 586         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 587         vq->vq_desc_head_idx = idx & (vq->vq_nentries - 1);
 588 }
 589
 590 static inline void
 591 virtqueue_enqueue_xmit_packed(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
 592                               uint16_t needed, int can_push)
 593 {
 594         struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 595         struct vq_desc_extra *dxp;
 596         struct virtqueue *vq = txvq->vq;
 597         struct vring_packed_desc *start_dp, *head_dp;
 598         uint16_t idx, id, head_idx, head_flags;
 599         uint16_t head_size = vq->hw->vtnet_hdr_size;
 600         struct virtio_net_hdr *hdr;
 601         uint16_t prev;
 602
 603         id = vq->vq_desc_head_idx;
 604
 605         dxp = &vq->vq_descx[id];
 606         dxp->ndescs = needed;
 607         dxp->cookie = cookie;
 608
 609         head_idx = vq->vq_avail_idx;
 610         idx = head_idx;
 611         prev = head_idx;
 612         start_dp = vq->ring_packed.desc_packed;
 613
 614         head_dp = &vq->ring_packed.desc_packed[idx];
 615         head_flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 616         head_flags |= vq->avail_used_flags;
 617
 618         if (can_push) {
 619                 /* prepend cannot fail, checked by caller */
 620                 hdr = (struct virtio_net_hdr *)
 621                         rte_pktmbuf_prepend(cookie, head_size);
 622                 /* rte_pktmbuf_prepend() counts the hdr size to the pkt length,
 623                  * which is wrong. Below subtract restores correct pkt size.
 624                  */
 625                 cookie->pkt_len -= head_size;
 626
 627                 /* if offload disabled, it is not zeroed below, do it now */
 628                 if (!vq->hw->has_tx_offload) {
 629                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 630                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 631                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 632                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 633                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 634                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 635                 }
 636         } else {
 637                 /* setup first tx ring slot to point to header
 638                  * stored in reserved region.
 639                  */
 640                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 641                         RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
 642                 start_dp[idx].len   = vq->hw->vtnet_hdr_size;
 643                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 644                 idx++;
 645                 if (idx >= vq->vq_nentries) {
 646                         idx -= vq->vq_nentries;
 647                         vq->avail_wrap_counter ^= 1;
 648                         vq->avail_used_flags =
 649                                 VRING_DESC_F_AVAIL(vq->avail_wrap_counter) |
 650                                 VRING_DESC_F_USED(!vq->avail_wrap_counter);
 651                 }
 652         }
 653
 654         virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload);
 655
 656         do {
 657                 uint16_t flags;
 658
 659                 start_dp[idx].addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
 660                 start_dp[idx].len  = cookie->data_len;
 661                 if (likely(idx != head_idx)) {
 662                         flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 663                         flags |= vq->avail_used_flags;
 664                         start_dp[idx].flags = flags;
 665                 }
 666                 prev = idx;
 667                 idx++;
 668                 if (idx >= vq->vq_nentries) {
 669                         idx -= vq->vq_nentries;
 670                         vq->avail_wrap_counter ^= 1;
 671                         vq->avail_used_flags =
 672                                 VRING_DESC_F_AVAIL(vq->avail_wrap_counter) |
 673                                 VRING_DESC_F_USED(!vq->avail_wrap_counter);
 674                 }
 675         } while ((cookie = cookie->next) != NULL);
 676
 677         start_dp[prev].id = id;
 678
 679         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 680
 681         vq->vq_desc_head_idx = dxp->next;
 682         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 683                 vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
 684
 685         vq->vq_avail_idx = idx;
 686
 687         rte_smp_wmb();
 688         head_dp->flags = head_flags;
 689 }
 690
 691 static inline void
 692 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
 693                         uint16_t needed, int use_indirect, int can_push,
 694                         int in_order)
 695 {
 696         struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 697         struct vq_desc_extra *dxp;
 698         struct virtqueue *vq = txvq->vq;
 699         struct vring_desc *start_dp;
 700         uint16_t seg_num = cookie->nb_segs;
 701         uint16_t head_idx, idx;
 702         uint16_t head_size = vq->hw->vtnet_hdr_size;
 703         struct virtio_net_hdr *hdr;
 704
 705         head_idx = vq->vq_desc_head_idx;
 706         idx = head_idx;
 707         dxp = &vq->vq_descx[idx];
 708         dxp->cookie = (void *)cookie;
 709         dxp->ndescs = needed;
 710
 711         start_dp = vq->vq_ring.desc;
 712
 713         if (can_push) {
 714                 /* prepend cannot fail, checked by caller */
 715                 hdr = (struct virtio_net_hdr *)
 716                         rte_pktmbuf_prepend(cookie, head_size);
 717                 /* rte_pktmbuf_prepend() counts the hdr size to the pkt length,
 718                  * which is wrong. Below subtract restores correct pkt size.
 719                  */
 720                 cookie->pkt_len -= head_size;
 721
 722                 /* if offload disabled, it is not zeroed below, do it now */
 723                 if (!vq->hw->has_tx_offload) {
 724                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 725                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 726                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 727                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 728                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 729                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 730                 }
 731         } else if (use_indirect) {
 732                 /* setup tx ring slot to point to indirect
 733                  * descriptor list stored in reserved region.
 734                  *
 735                  * the first slot in indirect ring is already preset
 736                  * to point to the header in reserved region
 737                  */
 738                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 739                         RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
 740                 start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
 741                 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
 742                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 743
 744                 /* loop below will fill in rest of the indirect elements */
 745                 start_dp = txr[idx].tx_indir;
 746                 idx = 1;
 747         } else {
 748                 /* setup first tx ring slot to point to header
 749                  * stored in reserved region.
 750                  */
 751                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 752                         RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
 753                 start_dp[idx].len   = vq->hw->vtnet_hdr_size;
 754                 start_dp[idx].flags = VRING_DESC_F_NEXT;
 755                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 756
 757                 idx = start_dp[idx].next;
 758         }
 759
 760         virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload);
 761
 762         do {
 763                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
 764                 start_dp[idx].len   = cookie->data_len;
 765                 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 766                 idx = start_dp[idx].next;
 767         } while ((cookie = cookie->next) != NULL);
 768
 769         if (use_indirect)
 770                 idx = vq->vq_ring.desc[head_idx].next;
 771
 772         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 773
 774         vq->vq_desc_head_idx = idx;
 775         vq_update_avail_ring(vq, head_idx);
 776
 777         if (!in_order) {
 778                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 779                         vq->vq_desc_tail_idx = idx;
 780         }
 781 }
 782
 783 void
 784 virtio_dev_cq_start(struct rte_eth_dev *dev)
 785 {
 786         struct virtio_hw *hw = dev->data->dev_private;
 787
 788         if (hw->cvq && hw->cvq->vq) {
 789                 rte_spinlock_init(&hw->cvq->lock);
 790                 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
 791         }
 792 }
 793
 794 int
 795 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
 796                         uint16_t queue_idx,
 797                         uint16_t nb_desc,
 798                         unsigned int socket_id __rte_unused,
 799                         const struct rte_eth_rxconf *rx_conf __rte_unused,
 800                         struct rte_mempool *mp)
 801 {
 802         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 803         struct virtio_hw *hw = dev->data->dev_private;
 804         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 805         struct virtnet_rx *rxvq;
 806
 807         PMD_INIT_FUNC_TRACE();
 808
 809         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 810                 nb_desc = vq->vq_nentries;
 811         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 812
 813         rxvq = &vq->rxq;
 814         rxvq->queue_id = queue_idx;
 815         rxvq->mpool = mp;
 816         if (rxvq->mpool == NULL) {
 817                 rte_exit(EXIT_FAILURE,
 818                         "Cannot allocate mbufs for rx virtqueue");
 819         }
 820
 821         dev->data->rx_queues[queue_idx] = rxvq;
 822
 823         return 0;
 824 }
 825
 826 int
 827 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
 828 {
 829         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 830         struct virtio_hw *hw = dev->data->dev_private;
 831         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 832         struct virtnet_rx *rxvq = &vq->rxq;
 833         struct rte_mbuf *m;
 834         uint16_t desc_idx;
 835         int error, nbufs, i;
 836
 837         PMD_INIT_FUNC_TRACE();
 838
 839         /* Allocate blank mbufs for the each rx descriptor */
 840         nbufs = 0;
 841
 842         if (hw->use_simple_rx) {
 843                 for (desc_idx = 0; desc_idx < vq->vq_nentries;
 844                      desc_idx++) {
 845                         vq->vq_ring.avail->ring[desc_idx] = desc_idx;
 846                         vq->vq_ring.desc[desc_idx].flags =
 847                                 VRING_DESC_F_WRITE;
 848                 }
 849
 850                 virtio_rxq_vec_setup(rxvq);
 851         }
 852
 853         memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
 854         for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
 855              desc_idx++) {
 856                 vq->sw_ring[vq->vq_nentries + desc_idx] =
 857                         &rxvq->fake_mbuf;
 858         }
 859
 860         if (hw->use_simple_rx) {
 861                 while (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
 862                         virtio_rxq_rearm_vec(rxvq);
 863                         nbufs += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
 864                 }
 865         } else if (hw->use_inorder_rx) {
 866                 if ((!virtqueue_full(vq))) {
 867                         uint16_t free_cnt = vq->vq_free_cnt;
 868                         struct rte_mbuf *pkts[free_cnt];
 869
 870                         if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, pkts,
 871                                 free_cnt)) {
 872                                 error = virtqueue_enqueue_refill_inorder(vq,
 873                                                 pkts,
 874                                                 free_cnt);
 875                                 if (unlikely(error)) {
 876                                         for (i = 0; i < free_cnt; i++)
 877                                                 rte_pktmbuf_free(pkts[i]);
 878                                 }
 879                         }
 880
 881                         nbufs += free_cnt;
 882                         vq_update_avail_idx(vq);
 883                 }
 884         } else {
 885                 while (!virtqueue_full(vq)) {
 886                         m = rte_mbuf_raw_alloc(rxvq->mpool);
 887                         if (m == NULL)
 888                                 break;
 889
 890                         /* Enqueue allocated buffers */
 891                         if (vtpci_packed_queue(vq->hw))
 892                                 error = virtqueue_enqueue_recv_refill_packed(vq,
 893                                                 &m, 1);
 894                         else
 895                                 error = virtqueue_enqueue_recv_refill(vq, m);
 896                         if (error) {
 897                                 rte_pktmbuf_free(m);
 898                                 break;
 899                         }
 900                         nbufs++;
 901                 }
 902
 903                 if (!vtpci_packed_queue(vq->hw))
 904                         vq_update_avail_idx(vq);
 905         }
 906
 907         PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
 908
 909         VIRTQUEUE_DUMP(vq);
 910
 911         return 0;
 912 }
 913
 914 /*
 915  * struct rte_eth_dev *dev: Used to update dev
 916  * uint16_t nb_desc: Defaults to values read from config space
 917  * unsigned int socket_id: Used to allocate memzone
 918  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
 919  * uint16_t queue_idx: Just used as an index in dev txq list
 920  */
 921 int
 922 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
 923                         uint16_t queue_idx,
 924                         uint16_t nb_desc,
 925                         unsigned int socket_id __rte_unused,
 926                         const struct rte_eth_txconf *tx_conf)
 927 {
 928         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 929         struct virtio_hw *hw = dev->data->dev_private;
 930         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 931         struct virtnet_tx *txvq;
 932         uint16_t tx_free_thresh;
 933
 934         PMD_INIT_FUNC_TRACE();
 935
 936         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 937                 nb_desc = vq->vq_nentries;
 938         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 939
 940         txvq = &vq->txq;
 941         txvq->queue_id = queue_idx;
 942
 943         tx_free_thresh = tx_conf->tx_free_thresh;
 944         if (tx_free_thresh == 0)
 945                 tx_free_thresh =
 946                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
 947
 948         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
 949                 RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
 950                         "number of TX entries minus 3 (%u)."
 951                         " (tx_free_thresh=%u port=%u queue=%u)\n",
 952                         vq->vq_nentries - 3,
 953                         tx_free_thresh, dev->data->port_id, queue_idx);
 954                 return -EINVAL;
 955         }
 956
 957         vq->vq_free_thresh = tx_free_thresh;
 958
 959         dev->data->tx_queues[queue_idx] = txvq;
 960         return 0;
 961 }
 962
 963 int
 964 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
 965                                 uint16_t queue_idx)
 966 {
 967         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 968         struct virtio_hw *hw = dev->data->dev_private;
 969         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 970
 971         PMD_INIT_FUNC_TRACE();
 972
 973         if (!vtpci_packed_queue(hw)) {
 974                 if (hw->use_inorder_tx)
 975                         vq->vq_ring.desc[vq->vq_nentries - 1].next = 0;
 976         }
 977
 978         VIRTQUEUE_DUMP(vq);
 979
 980         return 0;
 981 }
 982
 983 static inline void
 984 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
 985 {
 986         int error;
 987         /*
 988          * Requeue the discarded mbuf. This should always be
 989          * successful since it was just dequeued.
 990          */
 991         if (vtpci_packed_queue(vq->hw))
 992                 error = virtqueue_enqueue_recv_refill_packed(vq, &m, 1);
 993         else
 994                 error = virtqueue_enqueue_recv_refill(vq, m);
 995
 996         if (unlikely(error)) {
 997                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
 998                 rte_pktmbuf_free(m);
 999         }
1000 }
1001
1002 static inline void
1003 virtio_discard_rxbuf_inorder(struct virtqueue *vq, struct rte_mbuf *m)
1004 {
1005         int error;
1006
1007         error = virtqueue_enqueue_refill_inorder(vq, &m, 1);
1008         if (unlikely(error)) {
1009                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
1010                 rte_pktmbuf_free(m);
1011         }
1012 }
1013
1014 static inline void
1015 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
1016 {
1017         uint32_t s = mbuf->pkt_len;
1018         struct ether_addr *ea;
1019
1020         stats->bytes += s;
1021
1022         if (s == 64) {
1023                 stats->size_bins[1]++;
1024         } else if (s > 64 && s < 1024) {
1025                 uint32_t bin;
1026
1027                 /* count zeros, and offset into correct bin */
1028                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
1029                 stats->size_bins[bin]++;
1030         } else {
1031                 if (s < 64)
1032                         stats->size_bins[0]++;
1033                 else if (s < 1519)
1034                         stats->size_bins[6]++;
1035                 else if (s >= 1519)
1036                         stats->size_bins[7]++;
1037         }
1038
1039         ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
1040         if (is_multicast_ether_addr(ea)) {
1041                 if (is_broadcast_ether_addr(ea))
1042                         stats->broadcast++;
1043                 else
1044                         stats->multicast++;
1045         }
1046 }
1047
1048 static inline void
1049 virtio_rx_stats_updated(struct virtnet_rx *rxvq, struct rte_mbuf *m)
1050 {
1051         VIRTIO_DUMP_PACKET(m, m->data_len);
1052
1053         virtio_update_packet_stats(&rxvq->stats, m);
1054 }
1055
1056 /* Optionally fill offload information in structure */
1057 static inline int
1058 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
1059 {
1060         struct rte_net_hdr_lens hdr_lens;
1061         uint32_t hdrlen, ptype;
1062         int l4_supported = 0;
1063
1064         /* nothing to do */
1065         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1066                 return 0;
1067
1068         m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
1069
1070         ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
1071         m->packet_type = ptype;
1072         if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
1073             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
1074             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
1075                 l4_supported = 1;
1076
1077         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1078                 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
1079                 if (hdr->csum_start <= hdrlen && l4_supported) {
1080                         m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
1081                 } else {
1082                         /* Unknown proto or tunnel, do sw cksum. We can assume
1083                          * the cksum field is in the first segment since the
1084                          * buffers we provided to the host are large enough.
1085                          * In case of SCTP, this will be wrong since it's a CRC
1086                          * but there's nothing we can do.
1087                          */
1088                         uint16_t csum = 0, off;
1089
1090                         rte_raw_cksum_mbuf(m, hdr->csum_start,
1091                                 rte_pktmbuf_pkt_len(m) - hdr->csum_start,
1092                                 &csum);
1093                         if (likely(csum != 0xffff))
1094                                 csum = ~csum;
1095                         off = hdr->csum_offset + hdr->csum_start;
1096                         if (rte_pktmbuf_data_len(m) >= off + 1)
1097                                 *rte_pktmbuf_mtod_offset(m, uint16_t *,
1098                                         off) = csum;
1099                 }
1100         } else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
1101                 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
1102         }
1103
1104         /* GSO request, save required information in mbuf */
1105         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1106                 /* Check unsupported modes */
1107                 if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
1108                     (hdr->gso_size == 0)) {
1109                         return -EINVAL;
1110                 }
1111
1112                 /* Update mss lengthes in mbuf */
1113                 m->tso_segsz = hdr->gso_size;
1114                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1115                         case VIRTIO_NET_HDR_GSO_TCPV4:
1116                         case VIRTIO_NET_HDR_GSO_TCPV6:
1117                                 m->ol_flags |= PKT_RX_LRO | \
1118                                         PKT_RX_L4_CKSUM_NONE;
1119                                 break;
1120                         default:
1121                                 return -EINVAL;
1122                 }
1123         }
1124
1125         return 0;
1126 }
1127
1128 #define VIRTIO_MBUF_BURST_SZ 64
1129 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
1130 uint16_t
1131 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1132 {
1133         struct virtnet_rx *rxvq = rx_queue;
1134         struct virtqueue *vq = rxvq->vq;
1135         struct virtio_hw *hw = vq->hw;
1136         struct rte_mbuf *rxm, *new_mbuf;
1137         uint16_t nb_used, num, nb_rx;
1138         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1139         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1140         int error;
1141         uint32_t i, nb_enqueued;
1142         uint32_t hdr_size;
1143         struct virtio_net_hdr *hdr;
1144
1145         nb_rx = 0;
1146         if (unlikely(hw->started == 0))
1147                 return nb_rx;
1148
1149         nb_used = VIRTQUEUE_NUSED(vq);
1150
1151         virtio_rmb();
1152
1153         num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1154         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1155                 num = VIRTIO_MBUF_BURST_SZ;
1156         if (likely(num > DESC_PER_CACHELINE))
1157                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1158
1159         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1160         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
1161
1162         nb_enqueued = 0;
1163         hdr_size = hw->vtnet_hdr_size;
1164
1165         for (i = 0; i < num ; i++) {
1166                 rxm = rcv_pkts[i];
1167
1168                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1169
1170                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
1171                         PMD_RX_LOG(ERR, "Packet drop");
1172                         nb_enqueued++;
1173                         virtio_discard_rxbuf(vq, rxm);
1174                         rxvq->stats.errors++;
1175                         continue;
1176                 }
1177
1178                 rxm->port = rxvq->port_id;
1179                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1180                 rxm->ol_flags = 0;
1181                 rxm->vlan_tci = 0;
1182
1183                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1184                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1185
1186                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1187                         RTE_PKTMBUF_HEADROOM - hdr_size);
1188
1189                 if (hw->vlan_strip)
1190                         rte_vlan_strip(rxm);
1191
1192                 if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1193                         virtio_discard_rxbuf(vq, rxm);
1194                         rxvq->stats.errors++;
1195                         continue;
1196                 }
1197
1198                 virtio_rx_stats_updated(rxvq, rxm);
1199
1200                 rx_pkts[nb_rx++] = rxm;
1201         }
1202
1203         rxvq->stats.packets += nb_rx;
1204
1205         /* Allocate new mbuf for the used descriptor */
1206         while (likely(!virtqueue_full(vq))) {
1207                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
1208                 if (unlikely(new_mbuf == NULL)) {
1209                         struct rte_eth_dev *dev
1210                                 = &rte_eth_devices[rxvq->port_id];
1211                         dev->data->rx_mbuf_alloc_failed++;
1212                         break;
1213                 }
1214                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
1215                 if (unlikely(error)) {
1216                         rte_pktmbuf_free(new_mbuf);
1217                         break;
1218                 }
1219                 nb_enqueued++;
1220         }
1221
1222         if (likely(nb_enqueued)) {
1223                 vq_update_avail_idx(vq);
1224
1225                 if (unlikely(virtqueue_kick_prepare(vq))) {
1226                         virtqueue_notify(vq);
1227                         PMD_RX_LOG(DEBUG, "Notified");
1228                 }
1229         }
1230
1231         return nb_rx;
1232 }
1233
1234 uint16_t
1235 virtio_recv_pkts_packed(void *rx_queue, struct rte_mbuf **rx_pkts,
1236                         uint16_t nb_pkts)
1237 {
1238         struct virtnet_rx *rxvq = rx_queue;
1239         struct virtqueue *vq = rxvq->vq;
1240         struct virtio_hw *hw = vq->hw;
1241         struct rte_mbuf *rxm, *new_mbuf;
1242         uint16_t num, nb_rx;
1243         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1244         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1245         int error;
1246         uint32_t i, nb_enqueued;
1247         uint32_t hdr_size;
1248         struct virtio_net_hdr *hdr;
1249
1250         nb_rx = 0;
1251         if (unlikely(hw->started == 0))
1252                 return nb_rx;
1253
1254         num = RTE_MIN(VIRTIO_MBUF_BURST_SZ, nb_pkts);
1255         if (likely(num > DESC_PER_CACHELINE))
1256                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1257
1258         num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1259         PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1260
1261         nb_enqueued = 0;
1262         hdr_size = hw->vtnet_hdr_size;
1263
1264         for (i = 0; i < num; i++) {
1265                 rxm = rcv_pkts[i];
1266
1267                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1268
1269                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
1270                         PMD_RX_LOG(ERR, "Packet drop");
1271                         nb_enqueued++;
1272                         virtio_discard_rxbuf(vq, rxm);
1273                         rxvq->stats.errors++;
1274                         continue;
1275                 }
1276
1277                 rxm->port = rxvq->port_id;
1278                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1279                 rxm->ol_flags = 0;
1280                 rxm->vlan_tci = 0;
1281
1282                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1283                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1284
1285                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1286                         RTE_PKTMBUF_HEADROOM - hdr_size);
1287
1288                 if (hw->vlan_strip)
1289                         rte_vlan_strip(rxm);
1290
1291                 if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1292                         virtio_discard_rxbuf(vq, rxm);
1293                         rxvq->stats.errors++;
1294                         continue;
1295                 }
1296
1297                 virtio_rx_stats_updated(rxvq, rxm);
1298
1299                 rx_pkts[nb_rx++] = rxm;
1300         }
1301
1302         rxvq->stats.packets += nb_rx;
1303
1304         /* Allocate new mbuf for the used descriptor */
1305         while (likely(!virtqueue_full(vq))) {
1306                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
1307                 if (unlikely(new_mbuf == NULL)) {
1308                         struct rte_eth_dev *dev =
1309                                 &rte_eth_devices[rxvq->port_id];
1310                         dev->data->rx_mbuf_alloc_failed++;
1311                         break;
1312                 }
1313                 error = virtqueue_enqueue_recv_refill_packed(vq, &new_mbuf, 1);
1314                 if (unlikely(error)) {
1315                         rte_pktmbuf_free(new_mbuf);
1316                         break;
1317                 }
1318                 nb_enqueued++;
1319         }
1320
1321         if (likely(nb_enqueued)) {
1322                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1323                         virtqueue_notify(vq);
1324                         PMD_RX_LOG(DEBUG, "Notified");
1325                 }
1326         }
1327
1328         return nb_rx;
1329 }
1330
1331
1332 uint16_t
1333 virtio_recv_pkts_inorder(void *rx_queue,
1334                         struct rte_mbuf **rx_pkts,
1335                         uint16_t nb_pkts)
1336 {
1337         struct virtnet_rx *rxvq = rx_queue;
1338         struct virtqueue *vq = rxvq->vq;
1339         struct virtio_hw *hw = vq->hw;
1340         struct rte_mbuf *rxm;
1341         struct rte_mbuf *prev;
1342         uint16_t nb_used, num, nb_rx;
1343         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1344         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1345         int error;
1346         uint32_t nb_enqueued;
1347         uint32_t seg_num;
1348         uint32_t seg_res;
1349         uint32_t hdr_size;
1350         int32_t i;
1351
1352         nb_rx = 0;
1353         if (unlikely(hw->started == 0))
1354                 return nb_rx;
1355
1356         nb_used = VIRTQUEUE_NUSED(vq);
1357         nb_used = RTE_MIN(nb_used, nb_pkts);
1358         nb_used = RTE_MIN(nb_used, VIRTIO_MBUF_BURST_SZ);
1359
1360         virtio_rmb();
1361
1362         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1363
1364         nb_enqueued = 0;
1365         seg_num = 1;
1366         seg_res = 0;
1367         hdr_size = hw->vtnet_hdr_size;
1368
1369         num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len, nb_used);
1370
1371         for (i = 0; i < num; i++) {
1372                 struct virtio_net_hdr_mrg_rxbuf *header;
1373
1374                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1375                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1376
1377                 rxm = rcv_pkts[i];
1378
1379                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
1380                         PMD_RX_LOG(ERR, "Packet drop");
1381                         nb_enqueued++;
1382                         virtio_discard_rxbuf_inorder(vq, rxm);
1383                         rxvq->stats.errors++;
1384                         continue;
1385                 }
1386
1387                 header = (struct virtio_net_hdr_mrg_rxbuf *)
1388                          ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1389                          - hdr_size);
1390
1391                 if (vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
1392                         seg_num = header->num_buffers;
1393                         if (seg_num == 0)
1394                                 seg_num = 1;
1395                 } else {
1396                         seg_num = 1;
1397                 }
1398
1399                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1400                 rxm->nb_segs = seg_num;
1401                 rxm->ol_flags = 0;
1402                 rxm->vlan_tci = 0;
1403                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1404                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1405
1406                 rxm->port = rxvq->port_id;
1407
1408                 rx_pkts[nb_rx] = rxm;
1409                 prev = rxm;
1410
1411                 if (vq->hw->has_rx_offload &&
1412                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1413                         virtio_discard_rxbuf_inorder(vq, rxm);
1414                         rxvq->stats.errors++;
1415                         continue;
1416                 }
1417
1418                 if (hw->vlan_strip)
1419                         rte_vlan_strip(rx_pkts[nb_rx]);
1420
1421                 seg_res = seg_num - 1;
1422
1423                 /* Merge remaining segments */
1424                 while (seg_res != 0 && i < (num - 1)) {
1425                         i++;
1426
1427                         rxm = rcv_pkts[i];
1428                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1429                         rxm->pkt_len = (uint32_t)(len[i]);
1430                         rxm->data_len = (uint16_t)(len[i]);
1431
1432                         rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1433                         rx_pkts[nb_rx]->data_len += (uint16_t)(len[i]);
1434
1435                         if (prev)
1436                                 prev->next = rxm;
1437
1438                         prev = rxm;
1439                         seg_res -= 1;
1440                 }
1441
1442                 if (!seg_res) {
1443                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1444                         nb_rx++;
1445                 }
1446         }
1447
1448         /* Last packet still need merge segments */
1449         while (seg_res != 0) {
1450                 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1451                                         VIRTIO_MBUF_BURST_SZ);
1452
1453                 prev = rcv_pkts[nb_rx];
1454                 if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
1455                         num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len,
1456                                                            rcv_cnt);
1457                         uint16_t extra_idx = 0;
1458
1459                         rcv_cnt = num;
1460                         while (extra_idx < rcv_cnt) {
1461                                 rxm = rcv_pkts[extra_idx];
1462                                 rxm->data_off =
1463                                         RTE_PKTMBUF_HEADROOM - hdr_size;
1464                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1465                                 rxm->data_len = (uint16_t)(len[extra_idx]);
1466                                 prev->next = rxm;
1467                                 prev = rxm;
1468                                 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1469                                 rx_pkts[nb_rx]->data_len += len[extra_idx];
1470                                 extra_idx += 1;
1471                         };
1472                         seg_res -= rcv_cnt;
1473
1474                         if (!seg_res) {
1475                                 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1476                                 nb_rx++;
1477                         }
1478                 } else {
1479                         PMD_RX_LOG(ERR,
1480                                         "No enough segments for packet.");
1481                         virtio_discard_rxbuf_inorder(vq, prev);
1482                         rxvq->stats.errors++;
1483                         break;
1484                 }
1485         }
1486
1487         rxvq->stats.packets += nb_rx;
1488
1489         /* Allocate new mbuf for the used descriptor */
1490
1491         if (likely(!virtqueue_full(vq))) {
1492                 /* free_cnt may include mrg descs */
1493                 uint16_t free_cnt = vq->vq_free_cnt;
1494                 struct rte_mbuf *new_pkts[free_cnt];
1495
1496                 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1497                         error = virtqueue_enqueue_refill_inorder(vq, new_pkts,
1498                                         free_cnt);
1499                         if (unlikely(error)) {
1500                                 for (i = 0; i < free_cnt; i++)
1501                                         rte_pktmbuf_free(new_pkts[i]);
1502                         }
1503                         nb_enqueued += free_cnt;
1504                 } else {
1505                         struct rte_eth_dev *dev =
1506                                 &rte_eth_devices[rxvq->port_id];
1507                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1508                 }
1509         }
1510
1511         if (likely(nb_enqueued)) {
1512                 vq_update_avail_idx(vq);
1513
1514                 if (unlikely(virtqueue_kick_prepare(vq))) {
1515                         virtqueue_notify(vq);
1516                         PMD_RX_LOG(DEBUG, "Notified");
1517                 }
1518         }
1519
1520         return nb_rx;
1521 }
1522
1523 uint16_t
1524 virtio_recv_mergeable_pkts(void *rx_queue,
1525                         struct rte_mbuf **rx_pkts,
1526                         uint16_t nb_pkts)
1527 {
1528         struct virtnet_rx *rxvq = rx_queue;
1529         struct virtqueue *vq = rxvq->vq;
1530         struct virtio_hw *hw = vq->hw;
1531         struct rte_mbuf *rxm, *new_mbuf;
1532         uint16_t nb_used, num, nb_rx;
1533         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1534         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1535         struct rte_mbuf *prev;
1536         int error;
1537         uint32_t i, nb_enqueued;
1538         uint32_t seg_num;
1539         uint16_t extra_idx;
1540         uint32_t seg_res;
1541         uint32_t hdr_size;
1542
1543         nb_rx = 0;
1544         if (unlikely(hw->started == 0))
1545                 return nb_rx;
1546
1547         nb_used = VIRTQUEUE_NUSED(vq);
1548
1549         virtio_rmb();
1550
1551         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1552
1553         i = 0;
1554         nb_enqueued = 0;
1555         seg_num = 0;
1556         extra_idx = 0;
1557         seg_res = 0;
1558         hdr_size = hw->vtnet_hdr_size;
1559
1560         while (i < nb_used) {
1561                 struct virtio_net_hdr_mrg_rxbuf *header;
1562
1563                 if (nb_rx == nb_pkts)
1564                         break;
1565
1566                 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
1567                 if (num != 1)
1568                         continue;
1569
1570                 i++;
1571
1572                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1573                 PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
1574
1575                 rxm = rcv_pkts[0];
1576
1577                 if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
1578                         PMD_RX_LOG(ERR, "Packet drop");
1579                         nb_enqueued++;
1580                         virtio_discard_rxbuf(vq, rxm);
1581                         rxvq->stats.errors++;
1582                         continue;
1583                 }
1584
1585                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
1586                         RTE_PKTMBUF_HEADROOM - hdr_size);
1587                 seg_num = header->num_buffers;
1588
1589                 if (seg_num == 0)
1590                         seg_num = 1;
1591
1592                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1593                 rxm->nb_segs = seg_num;
1594                 rxm->ol_flags = 0;
1595                 rxm->vlan_tci = 0;
1596                 rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
1597                 rxm->data_len = (uint16_t)(len[0] - hdr_size);
1598
1599                 rxm->port = rxvq->port_id;
1600                 rx_pkts[nb_rx] = rxm;
1601                 prev = rxm;
1602
1603                 if (hw->has_rx_offload &&
1604                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1605                         virtio_discard_rxbuf(vq, rxm);
1606                         rxvq->stats.errors++;
1607                         continue;
1608                 }
1609
1610                 seg_res = seg_num - 1;
1611
1612                 while (seg_res != 0) {
1613                         /*
1614                          * Get extra segments for current uncompleted packet.
1615                          */
1616                         uint16_t  rcv_cnt =
1617                                 RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
1618                         if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
1619                                 uint32_t rx_num =
1620                                         virtqueue_dequeue_burst_rx(vq,
1621                                         rcv_pkts, len, rcv_cnt);
1622                                 i += rx_num;
1623                                 rcv_cnt = rx_num;
1624                         } else {
1625                                 PMD_RX_LOG(ERR,
1626                                            "No enough segments for packet.");
1627                                 nb_enqueued++;
1628                                 virtio_discard_rxbuf(vq, rxm);
1629                                 rxvq->stats.errors++;
1630                                 break;
1631                         }
1632
1633                         extra_idx = 0;
1634
1635                         while (extra_idx < rcv_cnt) {
1636                                 rxm = rcv_pkts[extra_idx];
1637
1638                                 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1639                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1640                                 rxm->data_len = (uint16_t)(len[extra_idx]);
1641
1642                                 if (prev)
1643                                         prev->next = rxm;
1644
1645                                 prev = rxm;
1646                                 rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
1647                                 extra_idx++;
1648                         };
1649                         seg_res -= rcv_cnt;
1650                 }
1651
1652                 if (hw->vlan_strip)
1653                         rte_vlan_strip(rx_pkts[nb_rx]);
1654
1655                 VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
1656                         rx_pkts[nb_rx]->data_len);
1657
1658                 virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
1659                 nb_rx++;
1660         }
1661
1662         rxvq->stats.packets += nb_rx;
1663
1664         /* Allocate new mbuf for the used descriptor */
1665         while (likely(!virtqueue_full(vq))) {
1666                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
1667                 if (unlikely(new_mbuf == NULL)) {
1668                         struct rte_eth_dev *dev
1669                                 = &rte_eth_devices[rxvq->port_id];
1670                         dev->data->rx_mbuf_alloc_failed++;
1671                         break;
1672                 }
1673                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
1674                 if (unlikely(error)) {
1675                         rte_pktmbuf_free(new_mbuf);
1676                         break;
1677                 }
1678                 nb_enqueued++;
1679         }
1680
1681         if (likely(nb_enqueued)) {
1682                 vq_update_avail_idx(vq);
1683
1684                 if (unlikely(virtqueue_kick_prepare(vq))) {
1685                         virtqueue_notify(vq);
1686                         PMD_RX_LOG(DEBUG, "Notified");
1687                 }
1688         }
1689
1690         return nb_rx;
1691 }
1692
1693 uint16_t
1694 virtio_recv_mergeable_pkts_packed(void *rx_queue,
1695                         struct rte_mbuf **rx_pkts,
1696                         uint16_t nb_pkts)
1697 {
1698         struct virtnet_rx *rxvq = rx_queue;
1699         struct virtqueue *vq = rxvq->vq;
1700         struct virtio_hw *hw = vq->hw;
1701         struct rte_mbuf *rxm;
1702         struct rte_mbuf *prev = NULL;
1703         uint16_t num, nb_rx = 0;
1704         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1705         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1706         uint32_t nb_enqueued = 0;
1707         uint32_t seg_num = 0;
1708         uint32_t seg_res = 0;
1709         uint32_t hdr_size = hw->vtnet_hdr_size;
1710         int32_t i;
1711         int error;
1712
1713         if (unlikely(hw->started == 0))
1714                 return nb_rx;
1715
1716
1717         num = nb_pkts;
1718         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1719                 num = VIRTIO_MBUF_BURST_SZ;
1720         if (likely(num > DESC_PER_CACHELINE))
1721                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1722
1723         num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1724
1725         for (i = 0; i < num; i++) {
1726                 struct virtio_net_hdr_mrg_rxbuf *header;
1727
1728                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1729                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1730
1731                 rxm = rcv_pkts[i];
1732
1733                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
1734                         PMD_RX_LOG(ERR, "Packet drop");
1735                         nb_enqueued++;
1736                         virtio_discard_rxbuf(vq, rxm);
1737                         rxvq->stats.errors++;
1738                         continue;
1739                 }
1740
1741                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)
1742                           rxm->buf_addr + RTE_PKTMBUF_HEADROOM - hdr_size);
1743                 seg_num = header->num_buffers;
1744
1745                 if (seg_num == 0)
1746                         seg_num = 1;
1747
1748                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1749                 rxm->nb_segs = seg_num;
1750                 rxm->ol_flags = 0;
1751                 rxm->vlan_tci = 0;
1752                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1753                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1754
1755                 rxm->port = rxvq->port_id;
1756                 rx_pkts[nb_rx] = rxm;
1757                 prev = rxm;
1758
1759                 if (hw->has_rx_offload &&
1760                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1761                         virtio_discard_rxbuf(vq, rxm);
1762                         rxvq->stats.errors++;
1763                         continue;
1764                 }
1765
1766                 if (hw->vlan_strip)
1767                         rte_vlan_strip(rx_pkts[nb_rx]);
1768
1769                 seg_res = seg_num - 1;
1770
1771                 /* Merge remaining segments */
1772                 while (seg_res != 0 && i < (num - 1)) {
1773                         i++;
1774
1775                         rxm = rcv_pkts[i];
1776                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1777                         rxm->pkt_len = (uint32_t)(len[i]);
1778                         rxm->data_len = (uint16_t)(len[i]);
1779
1780                         rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1781                         rx_pkts[nb_rx]->data_len += (uint16_t)(len[i]);
1782
1783                         if (prev)
1784                                 prev->next = rxm;
1785
1786                         prev = rxm;
1787                         seg_res -= 1;
1788                 }
1789
1790                 if (!seg_res) {
1791                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1792                         nb_rx++;
1793                 }
1794         }
1795
1796         /* Last packet still need merge segments */
1797         while (seg_res != 0) {
1798                 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1799                                         VIRTIO_MBUF_BURST_SZ);
1800                 if (likely(vq->vq_free_cnt >= rcv_cnt)) {
1801                         num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts,
1802                                         len, rcv_cnt);
1803                         uint16_t extra_idx = 0;
1804
1805                         rcv_cnt = num;
1806
1807                         while (extra_idx < rcv_cnt) {
1808                                 rxm = rcv_pkts[extra_idx];
1809
1810                                 rxm->data_off =
1811                                         RTE_PKTMBUF_HEADROOM - hdr_size;
1812                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1813                                 rxm->data_len = (uint16_t)(len[extra_idx]);
1814
1815                                 prev->next = rxm;
1816                                 prev = rxm;
1817                                 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1818                                 rx_pkts[nb_rx]->data_len += len[extra_idx];
1819                                 extra_idx += 1;
1820                         }
1821                         seg_res -= rcv_cnt;
1822                         if (!seg_res) {
1823                                 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1824                                 nb_rx++;
1825                         }
1826                 } else {
1827                         PMD_RX_LOG(ERR,
1828                                         "No enough segments for packet.");
1829                         if (prev)
1830                                 virtio_discard_rxbuf(vq, prev);
1831                         rxvq->stats.errors++;
1832                         break;
1833                 }
1834         }
1835
1836         rxvq->stats.packets += nb_rx;
1837
1838         /* Allocate new mbuf for the used descriptor */
1839         if (likely(!virtqueue_full(vq))) {
1840                 /* free_cnt may include mrg descs */
1841                 uint16_t free_cnt = vq->vq_free_cnt;
1842                 struct rte_mbuf *new_pkts[free_cnt];
1843
1844                 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1845                         error = virtqueue_enqueue_recv_refill_packed(vq,
1846                                         new_pkts, free_cnt);
1847                         if (unlikely(error)) {
1848                                 for (i = 0; i < free_cnt; i++)
1849                                         rte_pktmbuf_free(new_pkts[i]);
1850                         }
1851                         nb_enqueued += free_cnt;
1852                 } else {
1853                         struct rte_eth_dev *dev =
1854                                 &rte_eth_devices[rxvq->port_id];
1855                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1856                 }
1857         }
1858
1859         if (likely(nb_enqueued)) {
1860                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1861                         virtqueue_notify(vq);
1862                         PMD_RX_LOG(DEBUG, "Notified");
1863                 }
1864         }
1865
1866         return nb_rx;
1867 }
1868
1869 uint16_t
1870 virtio_xmit_pkts_packed(void *tx_queue, struct rte_mbuf **tx_pkts,
1871                         uint16_t nb_pkts)
1872 {
1873         struct virtnet_tx *txvq = tx_queue;
1874         struct virtqueue *vq = txvq->vq;
1875         struct virtio_hw *hw = vq->hw;
1876         uint16_t hdr_size = hw->vtnet_hdr_size;
1877         uint16_t nb_tx = 0;
1878         int error;
1879
1880         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1881                 return nb_tx;
1882
1883         if (unlikely(nb_pkts < 1))
1884                 return nb_pkts;
1885
1886         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1887
1888         if (nb_pkts > vq->vq_free_cnt)
1889                 virtio_xmit_cleanup_packed(vq, nb_pkts - vq->vq_free_cnt);
1890
1891         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1892                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1893                 int can_push = 0, slots, need;
1894
1895                 /* Do VLAN tag insertion */
1896                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1897                         error = rte_vlan_insert(&txm);
1898                         if (unlikely(error)) {
1899                                 rte_pktmbuf_free(txm);
1900                                 continue;
1901                         }
1902                 }
1903
1904                 /* optimize ring usage */
1905                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1906                       vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1907                     rte_mbuf_refcnt_read(txm) == 1 &&
1908                     RTE_MBUF_DIRECT(txm) &&
1909                     txm->nb_segs == 1 &&
1910                     rte_pktmbuf_headroom(txm) >= hdr_size &&
1911                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1912                            __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1913                         can_push = 1;
1914
1915                 /* How many main ring entries are needed to this Tx?
1916                  * any_layout => number of segments
1917                  * default    => number of segments + 1
1918                  */
1919                 slots = txm->nb_segs + !can_push;
1920                 need = slots - vq->vq_free_cnt;
1921
1922                 /* Positive value indicates it need free vring descriptors */
1923                 if (unlikely(need > 0)) {
1924                         virtio_rmb();
1925                         need = RTE_MIN(need, (int)nb_pkts);
1926                         virtio_xmit_cleanup_packed(vq, need);
1927                         need = slots - vq->vq_free_cnt;
1928                         if (unlikely(need > 0)) {
1929                                 PMD_TX_LOG(ERR,
1930                                            "No free tx descriptors to transmit");
1931                                 break;
1932                         }
1933                 }
1934
1935                 /* Enqueue Packet buffers */
1936                 virtqueue_enqueue_xmit_packed(txvq, txm, slots, can_push);
1937
1938                 virtio_update_packet_stats(&txvq->stats, txm);
1939         }
1940
1941         txvq->stats.packets += nb_tx;
1942
1943         if (likely(nb_tx)) {
1944                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1945                         virtqueue_notify(vq);
1946                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1947                 }
1948         }
1949
1950         return nb_tx;
1951 }
1952
1953 uint16_t
1954 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1955 {
1956         struct virtnet_tx *txvq = tx_queue;
1957         struct virtqueue *vq = txvq->vq;
1958         struct virtio_hw *hw = vq->hw;
1959         uint16_t hdr_size = hw->vtnet_hdr_size;
1960         uint16_t nb_used, nb_tx = 0;
1961         int error;
1962
1963         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1964                 return nb_tx;
1965
1966         if (unlikely(nb_pkts < 1))
1967                 return nb_pkts;
1968
1969         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1970         nb_used = VIRTQUEUE_NUSED(vq);
1971
1972         virtio_rmb();
1973         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1974                 virtio_xmit_cleanup(vq, nb_used);
1975
1976         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1977                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1978                 int can_push = 0, use_indirect = 0, slots, need;
1979
1980                 /* Do VLAN tag insertion */
1981                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1982                         error = rte_vlan_insert(&txm);
1983                         if (unlikely(error)) {
1984                                 rte_pktmbuf_free(txm);
1985                                 continue;
1986                         }
1987                 }
1988
1989                 /* optimize ring usage */
1990                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1991                       vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1992                     rte_mbuf_refcnt_read(txm) == 1 &&
1993                     RTE_MBUF_DIRECT(txm) &&
1994                     txm->nb_segs == 1 &&
1995                     rte_pktmbuf_headroom(txm) >= hdr_size &&
1996                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1997                                    __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1998                         can_push = 1;
1999                 else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
2000                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
2001                         use_indirect = 1;
2002
2003                 /* How many main ring entries are needed to this Tx?
2004                  * any_layout => number of segments
2005                  * indirect   => 1
2006                  * default    => number of segments + 1
2007                  */
2008                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
2009                 need = slots - vq->vq_free_cnt;
2010
2011                 /* Positive value indicates it need free vring descriptors */
2012                 if (unlikely(need > 0)) {
2013                         nb_used = VIRTQUEUE_NUSED(vq);
2014                         virtio_rmb();
2015                         need = RTE_MIN(need, (int)nb_used);
2016
2017                         virtio_xmit_cleanup(vq, need);
2018                         need = slots - vq->vq_free_cnt;
2019                         if (unlikely(need > 0)) {
2020                                 PMD_TX_LOG(ERR,
2021                                            "No free tx descriptors to transmit");
2022                                 break;
2023                         }
2024                 }
2025
2026                 /* Enqueue Packet buffers */
2027                 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
2028                         can_push, 0);
2029
2030                 virtio_update_packet_stats(&txvq->stats, txm);
2031         }
2032
2033         txvq->stats.packets += nb_tx;
2034
2035         if (likely(nb_tx)) {
2036                 vq_update_avail_idx(vq);
2037
2038                 if (unlikely(virtqueue_kick_prepare(vq))) {
2039                         virtqueue_notify(vq);
2040                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2041                 }
2042         }
2043
2044         return nb_tx;
2045 }
2046
2047 uint16_t
2048 virtio_xmit_pkts_inorder(void *tx_queue,
2049                         struct rte_mbuf **tx_pkts,
2050                         uint16_t nb_pkts)
2051 {
2052         struct virtnet_tx *txvq = tx_queue;
2053         struct virtqueue *vq = txvq->vq;
2054         struct virtio_hw *hw = vq->hw;
2055         uint16_t hdr_size = hw->vtnet_hdr_size;
2056         uint16_t nb_used, nb_avail, nb_tx = 0, nb_inorder_pkts = 0;
2057         struct rte_mbuf *inorder_pkts[nb_pkts];
2058         int error;
2059
2060         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
2061                 return nb_tx;
2062
2063         if (unlikely(nb_pkts < 1))
2064                 return nb_pkts;
2065
2066         VIRTQUEUE_DUMP(vq);
2067         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
2068         nb_used = VIRTQUEUE_NUSED(vq);
2069
2070         virtio_rmb();
2071         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
2072                 virtio_xmit_cleanup_inorder(vq, nb_used);
2073
2074         if (unlikely(!vq->vq_free_cnt))
2075                 virtio_xmit_cleanup_inorder(vq, nb_used);
2076
2077         nb_avail = RTE_MIN(vq->vq_free_cnt, nb_pkts);
2078
2079         for (nb_tx = 0; nb_tx < nb_avail; nb_tx++) {
2080                 struct rte_mbuf *txm = tx_pkts[nb_tx];
2081                 int slots, need;
2082
2083                 /* Do VLAN tag insertion */
2084                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
2085                         error = rte_vlan_insert(&txm);
2086                         if (unlikely(error)) {
2087                                 rte_pktmbuf_free(txm);
2088                                 continue;
2089                         }
2090                 }
2091
2092                 /* optimize ring usage */
2093                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
2094                      vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
2095                      rte_mbuf_refcnt_read(txm) == 1 &&
2096                      RTE_MBUF_DIRECT(txm) &&
2097                      txm->nb_segs == 1 &&
2098                      rte_pktmbuf_headroom(txm) >= hdr_size &&
2099                      rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
2100                                 __alignof__(struct virtio_net_hdr_mrg_rxbuf))) {
2101                         inorder_pkts[nb_inorder_pkts] = txm;
2102                         nb_inorder_pkts++;
2103
2104                         virtio_update_packet_stats(&txvq->stats, txm);
2105                         continue;
2106                 }
2107
2108                 if (nb_inorder_pkts) {
2109                         virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2110                                                         nb_inorder_pkts);
2111                         nb_inorder_pkts = 0;
2112                 }
2113
2114                 slots = txm->nb_segs + 1;
2115                 need = slots - vq->vq_free_cnt;
2116                 if (unlikely(need > 0)) {
2117                         nb_used = VIRTQUEUE_NUSED(vq);
2118                         virtio_rmb();
2119                         need = RTE_MIN(need, (int)nb_used);
2120
2121                         virtio_xmit_cleanup_inorder(vq, need);
2122
2123                         need = slots - vq->vq_free_cnt;
2124
2125                         if (unlikely(need > 0)) {
2126                                 PMD_TX_LOG(ERR,
2127                                         "No free tx descriptors to transmit");
2128                                 break;
2129                         }
2130                 }
2131                 /* Enqueue Packet buffers */
2132                 virtqueue_enqueue_xmit(txvq, txm, slots, 0, 0, 1);
2133
2134                 virtio_update_packet_stats(&txvq->stats, txm);
2135         }
2136
2137         /* Transmit all inorder packets */
2138         if (nb_inorder_pkts)
2139                 virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2140                                                 nb_inorder_pkts);
2141
2142         txvq->stats.packets += nb_tx;
2143
2144         if (likely(nb_tx)) {
2145                 vq_update_avail_idx(vq);
2146
2147                 if (unlikely(virtqueue_kick_prepare(vq))) {
2148                         virtqueue_notify(vq);
2149                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2150                 }
2151         }
2152
2153         VIRTQUEUE_DUMP(vq);
2154
2155         return nb_tx;
2156 }