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                                 uint16_t num)
 358 {
 359         struct vq_desc_extra *dxp;
 360         struct virtio_hw *hw = vq->hw;
 361         struct vring_desc *start_dp = vq->vq_ring.desc;
 362         uint16_t idx, i;
 363
 364         if (unlikely(vq->vq_free_cnt == 0))
 365                 return -ENOSPC;
 366         if (unlikely(vq->vq_free_cnt < num))
 367                 return -EMSGSIZE;
 368
 369         if (unlikely(vq->vq_desc_head_idx >= vq->vq_nentries))
 370                 return -EFAULT;
 371
 372         for (i = 0; i < num; i++) {
 373                 idx = vq->vq_desc_head_idx;
 374                 dxp = &vq->vq_descx[idx];
 375                 dxp->cookie = (void *)cookie[i];
 376                 dxp->ndescs = 1;
 377
 378                 start_dp[idx].addr =
 379                         VIRTIO_MBUF_ADDR(cookie[i], vq) +
 380                         RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 381                 start_dp[idx].len =
 382                         cookie[i]->buf_len - RTE_PKTMBUF_HEADROOM +
 383                         hw->vtnet_hdr_size;
 384                 start_dp[idx].flags = VRING_DESC_F_WRITE;
 385                 vq->vq_desc_head_idx = start_dp[idx].next;
 386                 vq_update_avail_ring(vq, idx);
 387                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END) {
 388                         vq->vq_desc_tail_idx = vq->vq_desc_head_idx;
 389                         break;
 390                 }
 391         }
 392
 393         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 394
 395         return 0;
 396 }
 397
 398 static inline int
 399 virtqueue_enqueue_recv_refill_packed(struct virtqueue *vq,
 400                                      struct rte_mbuf **cookie, uint16_t num)
 401 {
 402         struct vring_packed_desc *start_dp = vq->ring_packed.desc_packed;
 403         uint16_t flags = VRING_DESC_F_WRITE | vq->avail_used_flags;
 404         struct virtio_hw *hw = vq->hw;
 405         struct vq_desc_extra *dxp;
 406         uint16_t idx;
 407         int i;
 408
 409         if (unlikely(vq->vq_free_cnt == 0))
 410                 return -ENOSPC;
 411         if (unlikely(vq->vq_free_cnt < num))
 412                 return -EMSGSIZE;
 413
 414         for (i = 0; i < num; i++) {
 415                 idx = vq->vq_avail_idx;
 416                 dxp = &vq->vq_descx[idx];
 417                 dxp->cookie = (void *)cookie[i];
 418                 dxp->ndescs = 1;
 419
 420                 start_dp[idx].addr = VIRTIO_MBUF_ADDR(cookie[i], vq) +
 421                                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 422                 start_dp[idx].len = cookie[i]->buf_len - RTE_PKTMBUF_HEADROOM
 423                                         + hw->vtnet_hdr_size;
 424
 425                 vq->vq_desc_head_idx = dxp->next;
 426                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 427                         vq->vq_desc_tail_idx = vq->vq_desc_head_idx;
 428                 rte_smp_wmb();
 429                 start_dp[idx].flags = flags;
 430                 if (++vq->vq_avail_idx >= vq->vq_nentries) {
 431                         vq->vq_avail_idx -= vq->vq_nentries;
 432                         vq->avail_wrap_counter ^= 1;
 433                         vq->avail_used_flags =
 434                                 VRING_DESC_F_AVAIL(vq->avail_wrap_counter) |
 435                                 VRING_DESC_F_USED(!vq->avail_wrap_counter);
 436                         flags = VRING_DESC_F_WRITE | vq->avail_used_flags;
 437                 }
 438         }
 439         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 440         return 0;
 441 }
 442
 443 /* When doing TSO, the IP length is not included in the pseudo header
 444  * checksum of the packet given to the PMD, but for virtio it is
 445  * expected.
 446  */
 447 static void
 448 virtio_tso_fix_cksum(struct rte_mbuf *m)
 449 {
 450         /* common case: header is not fragmented */
 451         if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
 452                         m->l4_len)) {
 453                 struct ipv4_hdr *iph;
 454                 struct ipv6_hdr *ip6h;
 455                 struct tcp_hdr *th;
 456                 uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
 457                 uint32_t tmp;
 458
 459                 iph = rte_pktmbuf_mtod_offset(m, struct ipv4_hdr *, m->l2_len);
 460                 th = RTE_PTR_ADD(iph, m->l3_len);
 461                 if ((iph->version_ihl >> 4) == 4) {
 462                         iph->hdr_checksum = 0;
 463                         iph->hdr_checksum = rte_ipv4_cksum(iph);
 464                         ip_len = iph->total_length;
 465                         ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
 466                                 m->l3_len);
 467                 } else {
 468                         ip6h = (struct ipv6_hdr *)iph;
 469                         ip_paylen = ip6h->payload_len;
 470                 }
 471
 472                 /* calculate the new phdr checksum not including ip_paylen */
 473                 prev_cksum = th->cksum;
 474                 tmp = prev_cksum;
 475                 tmp += ip_paylen;
 476                 tmp = (tmp & 0xffff) + (tmp >> 16);
 477                 new_cksum = tmp;
 478
 479                 /* replace it in the packet */
 480                 th->cksum = new_cksum;
 481         }
 482 }
 483
 484
 485 /* avoid write operation when necessary, to lessen cache issues */
 486 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 487         if ((var) != (val))                     \
 488                 (var) = (val);                  \
 489 } while (0)
 490
 491 static inline void
 492 virtqueue_xmit_offload(struct virtio_net_hdr *hdr,
 493                         struct rte_mbuf *cookie,
 494                         bool offload)
 495 {
 496         if (offload) {
 497                 if (cookie->ol_flags & PKT_TX_TCP_SEG)
 498                         cookie->ol_flags |= PKT_TX_TCP_CKSUM;
 499
 500                 switch (cookie->ol_flags & PKT_TX_L4_MASK) {
 501                 case PKT_TX_UDP_CKSUM:
 502                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 503                         hdr->csum_offset = offsetof(struct udp_hdr,
 504                                 dgram_cksum);
 505                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 506                         break;
 507
 508                 case PKT_TX_TCP_CKSUM:
 509                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 510                         hdr->csum_offset = offsetof(struct tcp_hdr, cksum);
 511                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 512                         break;
 513
 514                 default:
 515                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 516                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 517                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 518                         break;
 519                 }
 520
 521                 /* TCP Segmentation Offload */
 522                 if (cookie->ol_flags & PKT_TX_TCP_SEG) {
 523                         virtio_tso_fix_cksum(cookie);
 524                         hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
 525                                 VIRTIO_NET_HDR_GSO_TCPV6 :
 526                                 VIRTIO_NET_HDR_GSO_TCPV4;
 527                         hdr->gso_size = cookie->tso_segsz;
 528                         hdr->hdr_len =
 529                                 cookie->l2_len +
 530                                 cookie->l3_len +
 531                                 cookie->l4_len;
 532                 } else {
 533                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 534                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 535                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 536                 }
 537         }
 538 }
 539
 540 static inline void
 541 virtqueue_enqueue_xmit_inorder(struct virtnet_tx *txvq,
 542                         struct rte_mbuf **cookies,
 543                         uint16_t num)
 544 {
 545         struct vq_desc_extra *dxp;
 546         struct virtqueue *vq = txvq->vq;
 547         struct vring_desc *start_dp;
 548         struct virtio_net_hdr *hdr;
 549         uint16_t idx;
 550         uint16_t head_size = vq->hw->vtnet_hdr_size;
 551         uint16_t i = 0;
 552
 553         idx = vq->vq_desc_head_idx;
 554         start_dp = vq->vq_ring.desc;
 555
 556         while (i < num) {
 557                 idx = idx & (vq->vq_nentries - 1);
 558                 dxp = &vq->vq_descx[idx];
 559                 dxp->cookie = (void *)cookies[i];
 560                 dxp->ndescs = 1;
 561
 562                 hdr = (struct virtio_net_hdr *)
 563                         rte_pktmbuf_prepend(cookies[i], head_size);
 564                 cookies[i]->pkt_len -= head_size;
 565
 566                 /* if offload disabled, it is not zeroed below, do it now */
 567                 if (!vq->hw->has_tx_offload) {
 568                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 569                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 570                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 571                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 572                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 573                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 574                 }
 575
 576                 virtqueue_xmit_offload(hdr, cookies[i],
 577                                 vq->hw->has_tx_offload);
 578
 579                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookies[i], vq);
 580                 start_dp[idx].len   = cookies[i]->data_len;
 581                 start_dp[idx].flags = 0;
 582
 583                 vq_update_avail_ring(vq, idx);
 584
 585                 idx++;
 586                 i++;
 587         };
 588
 589         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 590         vq->vq_desc_head_idx = idx & (vq->vq_nentries - 1);
 591 }
 592
 593 static inline void
 594 virtqueue_enqueue_xmit_packed(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
 595                               uint16_t needed, int can_push)
 596 {
 597         struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 598         struct vq_desc_extra *dxp;
 599         struct virtqueue *vq = txvq->vq;
 600         struct vring_packed_desc *start_dp, *head_dp;
 601         uint16_t idx, id, head_idx, head_flags;
 602         uint16_t head_size = vq->hw->vtnet_hdr_size;
 603         struct virtio_net_hdr *hdr;
 604         uint16_t prev;
 605
 606         id = vq->vq_desc_head_idx;
 607
 608         dxp = &vq->vq_descx[id];
 609         dxp->ndescs = needed;
 610         dxp->cookie = cookie;
 611
 612         head_idx = vq->vq_avail_idx;
 613         idx = head_idx;
 614         prev = head_idx;
 615         start_dp = vq->ring_packed.desc_packed;
 616
 617         head_dp = &vq->ring_packed.desc_packed[idx];
 618         head_flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 619         head_flags |= vq->avail_used_flags;
 620
 621         if (can_push) {
 622                 /* prepend cannot fail, checked by caller */
 623                 hdr = (struct virtio_net_hdr *)
 624                         rte_pktmbuf_prepend(cookie, head_size);
 625                 /* rte_pktmbuf_prepend() counts the hdr size to the pkt length,
 626                  * which is wrong. Below subtract restores correct pkt size.
 627                  */
 628                 cookie->pkt_len -= head_size;
 629
 630                 /* if offload disabled, it is not zeroed below, do it now */
 631                 if (!vq->hw->has_tx_offload) {
 632                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 633                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 634                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 635                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 636                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 637                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 638                 }
 639         } else {
 640                 /* setup first tx ring slot to point to header
 641                  * stored in reserved region.
 642                  */
 643                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 644                         RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
 645                 start_dp[idx].len   = vq->hw->vtnet_hdr_size;
 646                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 647                 idx++;
 648                 if (idx >= vq->vq_nentries) {
 649                         idx -= vq->vq_nentries;
 650                         vq->avail_wrap_counter ^= 1;
 651                         vq->avail_used_flags =
 652                                 VRING_DESC_F_AVAIL(vq->avail_wrap_counter) |
 653                                 VRING_DESC_F_USED(!vq->avail_wrap_counter);
 654                 }
 655         }
 656
 657         virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload);
 658
 659         do {
 660                 uint16_t flags;
 661
 662                 start_dp[idx].addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
 663                 start_dp[idx].len  = cookie->data_len;
 664                 if (likely(idx != head_idx)) {
 665                         flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 666                         flags |= vq->avail_used_flags;
 667                         start_dp[idx].flags = flags;
 668                 }
 669                 prev = idx;
 670                 idx++;
 671                 if (idx >= vq->vq_nentries) {
 672                         idx -= vq->vq_nentries;
 673                         vq->avail_wrap_counter ^= 1;
 674                         vq->avail_used_flags =
 675                                 VRING_DESC_F_AVAIL(vq->avail_wrap_counter) |
 676                                 VRING_DESC_F_USED(!vq->avail_wrap_counter);
 677                 }
 678         } while ((cookie = cookie->next) != NULL);
 679
 680         start_dp[prev].id = id;
 681
 682         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 683
 684         vq->vq_desc_head_idx = dxp->next;
 685         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 686                 vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
 687
 688         vq->vq_avail_idx = idx;
 689
 690         rte_smp_wmb();
 691         head_dp->flags = head_flags;
 692 }
 693
 694 static inline void
 695 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
 696                         uint16_t needed, int use_indirect, int can_push,
 697                         int in_order)
 698 {
 699         struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 700         struct vq_desc_extra *dxp;
 701         struct virtqueue *vq = txvq->vq;
 702         struct vring_desc *start_dp;
 703         uint16_t seg_num = cookie->nb_segs;
 704         uint16_t head_idx, idx;
 705         uint16_t head_size = vq->hw->vtnet_hdr_size;
 706         struct virtio_net_hdr *hdr;
 707
 708         head_idx = vq->vq_desc_head_idx;
 709         idx = head_idx;
 710         dxp = &vq->vq_descx[idx];
 711         dxp->cookie = (void *)cookie;
 712         dxp->ndescs = needed;
 713
 714         start_dp = vq->vq_ring.desc;
 715
 716         if (can_push) {
 717                 /* prepend cannot fail, checked by caller */
 718                 hdr = (struct virtio_net_hdr *)
 719                         rte_pktmbuf_prepend(cookie, head_size);
 720                 /* rte_pktmbuf_prepend() counts the hdr size to the pkt length,
 721                  * which is wrong. Below subtract restores correct pkt size.
 722                  */
 723                 cookie->pkt_len -= head_size;
 724
 725                 /* if offload disabled, it is not zeroed below, do it now */
 726                 if (!vq->hw->has_tx_offload) {
 727                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 728                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 729                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 730                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 731                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 732                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 733                 }
 734         } else if (use_indirect) {
 735                 /* setup tx ring slot to point to indirect
 736                  * descriptor list stored in reserved region.
 737                  *
 738                  * the first slot in indirect ring is already preset
 739                  * to point to the header in reserved region
 740                  */
 741                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 742                         RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
 743                 start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
 744                 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
 745                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 746
 747                 /* loop below will fill in rest of the indirect elements */
 748                 start_dp = txr[idx].tx_indir;
 749                 idx = 1;
 750         } else {
 751                 /* setup first tx ring slot to point to header
 752                  * stored in reserved region.
 753                  */
 754                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 755                         RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
 756                 start_dp[idx].len   = vq->hw->vtnet_hdr_size;
 757                 start_dp[idx].flags = VRING_DESC_F_NEXT;
 758                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 759
 760                 idx = start_dp[idx].next;
 761         }
 762
 763         virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload);
 764
 765         do {
 766                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
 767                 start_dp[idx].len   = cookie->data_len;
 768                 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 769                 idx = start_dp[idx].next;
 770         } while ((cookie = cookie->next) != NULL);
 771
 772         if (use_indirect)
 773                 idx = vq->vq_ring.desc[head_idx].next;
 774
 775         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 776
 777         vq->vq_desc_head_idx = idx;
 778         vq_update_avail_ring(vq, head_idx);
 779
 780         if (!in_order) {
 781                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 782                         vq->vq_desc_tail_idx = idx;
 783         }
 784 }
 785
 786 void
 787 virtio_dev_cq_start(struct rte_eth_dev *dev)
 788 {
 789         struct virtio_hw *hw = dev->data->dev_private;
 790
 791         if (hw->cvq && hw->cvq->vq) {
 792                 rte_spinlock_init(&hw->cvq->lock);
 793                 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
 794         }
 795 }
 796
 797 int
 798 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
 799                         uint16_t queue_idx,
 800                         uint16_t nb_desc,
 801                         unsigned int socket_id __rte_unused,
 802                         const struct rte_eth_rxconf *rx_conf __rte_unused,
 803                         struct rte_mempool *mp)
 804 {
 805         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 806         struct virtio_hw *hw = dev->data->dev_private;
 807         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 808         struct virtnet_rx *rxvq;
 809
 810         PMD_INIT_FUNC_TRACE();
 811
 812         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 813                 nb_desc = vq->vq_nentries;
 814         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 815
 816         rxvq = &vq->rxq;
 817         rxvq->queue_id = queue_idx;
 818         rxvq->mpool = mp;
 819         if (rxvq->mpool == NULL) {
 820                 rte_exit(EXIT_FAILURE,
 821                         "Cannot allocate mbufs for rx virtqueue");
 822         }
 823
 824         dev->data->rx_queues[queue_idx] = rxvq;
 825
 826         return 0;
 827 }
 828
 829 int
 830 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
 831 {
 832         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 833         struct virtio_hw *hw = dev->data->dev_private;
 834         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 835         struct virtnet_rx *rxvq = &vq->rxq;
 836         struct rte_mbuf *m;
 837         uint16_t desc_idx;
 838         int error, nbufs, i;
 839
 840         PMD_INIT_FUNC_TRACE();
 841
 842         /* Allocate blank mbufs for the each rx descriptor */
 843         nbufs = 0;
 844
 845         if (hw->use_simple_rx) {
 846                 for (desc_idx = 0; desc_idx < vq->vq_nentries;
 847                      desc_idx++) {
 848                         vq->vq_ring.avail->ring[desc_idx] = desc_idx;
 849                         vq->vq_ring.desc[desc_idx].flags =
 850                                 VRING_DESC_F_WRITE;
 851                 }
 852
 853                 virtio_rxq_vec_setup(rxvq);
 854         }
 855
 856         memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
 857         for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
 858              desc_idx++) {
 859                 vq->sw_ring[vq->vq_nentries + desc_idx] =
 860                         &rxvq->fake_mbuf;
 861         }
 862
 863         if (hw->use_simple_rx) {
 864                 while (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
 865                         virtio_rxq_rearm_vec(rxvq);
 866                         nbufs += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
 867                 }
 868         } else if (hw->use_inorder_rx) {
 869                 if ((!virtqueue_full(vq))) {
 870                         uint16_t free_cnt = vq->vq_free_cnt;
 871                         struct rte_mbuf *pkts[free_cnt];
 872
 873                         if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, pkts,
 874                                 free_cnt)) {
 875                                 error = virtqueue_enqueue_refill_inorder(vq,
 876                                                 pkts,
 877                                                 free_cnt);
 878                                 if (unlikely(error)) {
 879                                         for (i = 0; i < free_cnt; i++)
 880                                                 rte_pktmbuf_free(pkts[i]);
 881                                 }
 882                         }
 883
 884                         nbufs += free_cnt;
 885                         vq_update_avail_idx(vq);
 886                 }
 887         } else {
 888                 while (!virtqueue_full(vq)) {
 889                         m = rte_mbuf_raw_alloc(rxvq->mpool);
 890                         if (m == NULL)
 891                                 break;
 892
 893                         /* Enqueue allocated buffers */
 894                         if (vtpci_packed_queue(vq->hw))
 895                                 error = virtqueue_enqueue_recv_refill_packed(vq,
 896                                                 &m, 1);
 897                         else
 898                                 error = virtqueue_enqueue_recv_refill(vq,
 899                                                 &m, 1);
 900                         if (error) {
 901                                 rte_pktmbuf_free(m);
 902                                 break;
 903                         }
 904                         nbufs++;
 905                 }
 906
 907                 if (!vtpci_packed_queue(vq->hw))
 908                         vq_update_avail_idx(vq);
 909         }
 910
 911         PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
 912
 913         VIRTQUEUE_DUMP(vq);
 914
 915         return 0;
 916 }
 917
 918 /*
 919  * struct rte_eth_dev *dev: Used to update dev
 920  * uint16_t nb_desc: Defaults to values read from config space
 921  * unsigned int socket_id: Used to allocate memzone
 922  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
 923  * uint16_t queue_idx: Just used as an index in dev txq list
 924  */
 925 int
 926 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
 927                         uint16_t queue_idx,
 928                         uint16_t nb_desc,
 929                         unsigned int socket_id __rte_unused,
 930                         const struct rte_eth_txconf *tx_conf)
 931 {
 932         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 933         struct virtio_hw *hw = dev->data->dev_private;
 934         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 935         struct virtnet_tx *txvq;
 936         uint16_t tx_free_thresh;
 937
 938         PMD_INIT_FUNC_TRACE();
 939
 940         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 941                 nb_desc = vq->vq_nentries;
 942         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 943
 944         txvq = &vq->txq;
 945         txvq->queue_id = queue_idx;
 946
 947         tx_free_thresh = tx_conf->tx_free_thresh;
 948         if (tx_free_thresh == 0)
 949                 tx_free_thresh =
 950                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
 951
 952         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
 953                 RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
 954                         "number of TX entries minus 3 (%u)."
 955                         " (tx_free_thresh=%u port=%u queue=%u)\n",
 956                         vq->vq_nentries - 3,
 957                         tx_free_thresh, dev->data->port_id, queue_idx);
 958                 return -EINVAL;
 959         }
 960
 961         vq->vq_free_thresh = tx_free_thresh;
 962
 963         dev->data->tx_queues[queue_idx] = txvq;
 964         return 0;
 965 }
 966
 967 int
 968 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
 969                                 uint16_t queue_idx)
 970 {
 971         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 972         struct virtio_hw *hw = dev->data->dev_private;
 973         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 974
 975         PMD_INIT_FUNC_TRACE();
 976
 977         if (!vtpci_packed_queue(hw)) {
 978                 if (hw->use_inorder_tx)
 979                         vq->vq_ring.desc[vq->vq_nentries - 1].next = 0;
 980         }
 981
 982         VIRTQUEUE_DUMP(vq);
 983
 984         return 0;
 985 }
 986
 987 static inline void
 988 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
 989 {
 990         int error;
 991         /*
 992          * Requeue the discarded mbuf. This should always be
 993          * successful since it was just dequeued.
 994          */
 995         if (vtpci_packed_queue(vq->hw))
 996                 error = virtqueue_enqueue_recv_refill_packed(vq, &m, 1);
 997         else
 998                 error = virtqueue_enqueue_recv_refill(vq, &m, 1);
 999
1000         if (unlikely(error)) {
1001                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
1002                 rte_pktmbuf_free(m);
1003         }
1004 }
1005
1006 static inline void
1007 virtio_discard_rxbuf_inorder(struct virtqueue *vq, struct rte_mbuf *m)
1008 {
1009         int error;
1010
1011         error = virtqueue_enqueue_refill_inorder(vq, &m, 1);
1012         if (unlikely(error)) {
1013                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
1014                 rte_pktmbuf_free(m);
1015         }
1016 }
1017
1018 static inline void
1019 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
1020 {
1021         uint32_t s = mbuf->pkt_len;
1022         struct ether_addr *ea;
1023
1024         stats->bytes += s;
1025
1026         if (s == 64) {
1027                 stats->size_bins[1]++;
1028         } else if (s > 64 && s < 1024) {
1029                 uint32_t bin;
1030
1031                 /* count zeros, and offset into correct bin */
1032                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
1033                 stats->size_bins[bin]++;
1034         } else {
1035                 if (s < 64)
1036                         stats->size_bins[0]++;
1037                 else if (s < 1519)
1038                         stats->size_bins[6]++;
1039                 else if (s >= 1519)
1040                         stats->size_bins[7]++;
1041         }
1042
1043         ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
1044         if (is_multicast_ether_addr(ea)) {
1045                 if (is_broadcast_ether_addr(ea))
1046                         stats->broadcast++;
1047                 else
1048                         stats->multicast++;
1049         }
1050 }
1051
1052 static inline void
1053 virtio_rx_stats_updated(struct virtnet_rx *rxvq, struct rte_mbuf *m)
1054 {
1055         VIRTIO_DUMP_PACKET(m, m->data_len);
1056
1057         virtio_update_packet_stats(&rxvq->stats, m);
1058 }
1059
1060 /* Optionally fill offload information in structure */
1061 static inline int
1062 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
1063 {
1064         struct rte_net_hdr_lens hdr_lens;
1065         uint32_t hdrlen, ptype;
1066         int l4_supported = 0;
1067
1068         /* nothing to do */
1069         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1070                 return 0;
1071
1072         m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
1073
1074         ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
1075         m->packet_type = ptype;
1076         if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
1077             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
1078             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
1079                 l4_supported = 1;
1080
1081         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1082                 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
1083                 if (hdr->csum_start <= hdrlen && l4_supported) {
1084                         m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
1085                 } else {
1086                         /* Unknown proto or tunnel, do sw cksum. We can assume
1087                          * the cksum field is in the first segment since the
1088                          * buffers we provided to the host are large enough.
1089                          * In case of SCTP, this will be wrong since it's a CRC
1090                          * but there's nothing we can do.
1091                          */
1092                         uint16_t csum = 0, off;
1093
1094                         rte_raw_cksum_mbuf(m, hdr->csum_start,
1095                                 rte_pktmbuf_pkt_len(m) - hdr->csum_start,
1096                                 &csum);
1097                         if (likely(csum != 0xffff))
1098                                 csum = ~csum;
1099                         off = hdr->csum_offset + hdr->csum_start;
1100                         if (rte_pktmbuf_data_len(m) >= off + 1)
1101                                 *rte_pktmbuf_mtod_offset(m, uint16_t *,
1102                                         off) = csum;
1103                 }
1104         } else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
1105                 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
1106         }
1107
1108         /* GSO request, save required information in mbuf */
1109         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1110                 /* Check unsupported modes */
1111                 if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
1112                     (hdr->gso_size == 0)) {
1113                         return -EINVAL;
1114                 }
1115
1116                 /* Update mss lengthes in mbuf */
1117                 m->tso_segsz = hdr->gso_size;
1118                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1119                         case VIRTIO_NET_HDR_GSO_TCPV4:
1120                         case VIRTIO_NET_HDR_GSO_TCPV6:
1121                                 m->ol_flags |= PKT_RX_LRO | \
1122                                         PKT_RX_L4_CKSUM_NONE;
1123                                 break;
1124                         default:
1125                                 return -EINVAL;
1126                 }
1127         }
1128
1129         return 0;
1130 }
1131
1132 #define VIRTIO_MBUF_BURST_SZ 64
1133 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
1134 uint16_t
1135 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1136 {
1137         struct virtnet_rx *rxvq = rx_queue;
1138         struct virtqueue *vq = rxvq->vq;
1139         struct virtio_hw *hw = vq->hw;
1140         struct rte_mbuf *rxm, *new_mbuf;
1141         uint16_t nb_used, num, nb_rx;
1142         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1143         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1144         int error;
1145         uint32_t i, nb_enqueued;
1146         uint32_t hdr_size;
1147         struct virtio_net_hdr *hdr;
1148
1149         nb_rx = 0;
1150         if (unlikely(hw->started == 0))
1151                 return nb_rx;
1152
1153         nb_used = VIRTQUEUE_NUSED(vq);
1154
1155         virtio_rmb(hw->weak_barriers);
1156
1157         num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1158         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1159                 num = VIRTIO_MBUF_BURST_SZ;
1160         if (likely(num > DESC_PER_CACHELINE))
1161                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1162
1163         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1164         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
1165
1166         nb_enqueued = 0;
1167         hdr_size = hw->vtnet_hdr_size;
1168
1169         for (i = 0; i < num ; i++) {
1170                 rxm = rcv_pkts[i];
1171
1172                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1173
1174                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
1175                         PMD_RX_LOG(ERR, "Packet drop");
1176                         nb_enqueued++;
1177                         virtio_discard_rxbuf(vq, rxm);
1178                         rxvq->stats.errors++;
1179                         continue;
1180                 }
1181
1182                 rxm->port = rxvq->port_id;
1183                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1184                 rxm->ol_flags = 0;
1185                 rxm->vlan_tci = 0;
1186
1187                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1188                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1189
1190                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1191                         RTE_PKTMBUF_HEADROOM - hdr_size);
1192
1193                 if (hw->vlan_strip)
1194                         rte_vlan_strip(rxm);
1195
1196                 if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1197                         virtio_discard_rxbuf(vq, rxm);
1198                         rxvq->stats.errors++;
1199                         continue;
1200                 }
1201
1202                 virtio_rx_stats_updated(rxvq, rxm);
1203
1204                 rx_pkts[nb_rx++] = rxm;
1205         }
1206
1207         rxvq->stats.packets += nb_rx;
1208
1209         /* Allocate new mbuf for the used descriptor */
1210         while (likely(!virtqueue_full(vq))) {
1211                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
1212                 if (unlikely(new_mbuf == NULL)) {
1213                         struct rte_eth_dev *dev
1214                                 = &rte_eth_devices[rxvq->port_id];
1215                         dev->data->rx_mbuf_alloc_failed++;
1216                         break;
1217                 }
1218                 error = virtqueue_enqueue_recv_refill(vq, &new_mbuf, 1);
1219                 if (unlikely(error)) {
1220                         rte_pktmbuf_free(new_mbuf);
1221                         break;
1222                 }
1223                 nb_enqueued++;
1224         }
1225
1226         if (likely(nb_enqueued)) {
1227                 vq_update_avail_idx(vq);
1228
1229                 if (unlikely(virtqueue_kick_prepare(vq))) {
1230                         virtqueue_notify(vq);
1231                         PMD_RX_LOG(DEBUG, "Notified");
1232                 }
1233         }
1234
1235         return nb_rx;
1236 }
1237
1238 uint16_t
1239 virtio_recv_pkts_packed(void *rx_queue, struct rte_mbuf **rx_pkts,
1240                         uint16_t nb_pkts)
1241 {
1242         struct virtnet_rx *rxvq = rx_queue;
1243         struct virtqueue *vq = rxvq->vq;
1244         struct virtio_hw *hw = vq->hw;
1245         struct rte_mbuf *rxm, *new_mbuf;
1246         uint16_t num, nb_rx;
1247         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1248         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1249         int error;
1250         uint32_t i, nb_enqueued;
1251         uint32_t hdr_size;
1252         struct virtio_net_hdr *hdr;
1253
1254         nb_rx = 0;
1255         if (unlikely(hw->started == 0))
1256                 return nb_rx;
1257
1258         num = RTE_MIN(VIRTIO_MBUF_BURST_SZ, nb_pkts);
1259         if (likely(num > DESC_PER_CACHELINE))
1260                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1261
1262         num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1263         PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1264
1265         nb_enqueued = 0;
1266         hdr_size = hw->vtnet_hdr_size;
1267
1268         for (i = 0; i < num; i++) {
1269                 rxm = rcv_pkts[i];
1270
1271                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1272
1273                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
1274                         PMD_RX_LOG(ERR, "Packet drop");
1275                         nb_enqueued++;
1276                         virtio_discard_rxbuf(vq, rxm);
1277                         rxvq->stats.errors++;
1278                         continue;
1279                 }
1280
1281                 rxm->port = rxvq->port_id;
1282                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1283                 rxm->ol_flags = 0;
1284                 rxm->vlan_tci = 0;
1285
1286                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1287                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1288
1289                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1290                         RTE_PKTMBUF_HEADROOM - hdr_size);
1291
1292                 if (hw->vlan_strip)
1293                         rte_vlan_strip(rxm);
1294
1295                 if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1296                         virtio_discard_rxbuf(vq, rxm);
1297                         rxvq->stats.errors++;
1298                         continue;
1299                 }
1300
1301                 virtio_rx_stats_updated(rxvq, rxm);
1302
1303                 rx_pkts[nb_rx++] = rxm;
1304         }
1305
1306         rxvq->stats.packets += nb_rx;
1307
1308         /* Allocate new mbuf for the used descriptor */
1309         while (likely(!virtqueue_full(vq))) {
1310                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
1311                 if (unlikely(new_mbuf == NULL)) {
1312                         struct rte_eth_dev *dev =
1313                                 &rte_eth_devices[rxvq->port_id];
1314                         dev->data->rx_mbuf_alloc_failed++;
1315                         break;
1316                 }
1317                 error = virtqueue_enqueue_recv_refill_packed(vq, &new_mbuf, 1);
1318                 if (unlikely(error)) {
1319                         rte_pktmbuf_free(new_mbuf);
1320                         break;
1321                 }
1322                 nb_enqueued++;
1323         }
1324
1325         if (likely(nb_enqueued)) {
1326                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1327                         virtqueue_notify(vq);
1328                         PMD_RX_LOG(DEBUG, "Notified");
1329                 }
1330         }
1331
1332         return nb_rx;
1333 }
1334
1335
1336 uint16_t
1337 virtio_recv_pkts_inorder(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 = rxvq->vq;
1343         struct virtio_hw *hw = vq->hw;
1344         struct rte_mbuf *rxm;
1345         struct rte_mbuf *prev;
1346         uint16_t nb_used, num, nb_rx;
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;
1351         uint32_t seg_num;
1352         uint32_t seg_res;
1353         uint32_t hdr_size;
1354         int32_t i;
1355
1356         nb_rx = 0;
1357         if (unlikely(hw->started == 0))
1358                 return nb_rx;
1359
1360         nb_used = VIRTQUEUE_NUSED(vq);
1361         nb_used = RTE_MIN(nb_used, nb_pkts);
1362         nb_used = RTE_MIN(nb_used, VIRTIO_MBUF_BURST_SZ);
1363
1364         virtio_rmb(hw->weak_barriers);
1365
1366         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1367
1368         nb_enqueued = 0;
1369         seg_num = 1;
1370         seg_res = 0;
1371         hdr_size = hw->vtnet_hdr_size;
1372
1373         num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len, nb_used);
1374
1375         for (i = 0; i < num; i++) {
1376                 struct virtio_net_hdr_mrg_rxbuf *header;
1377
1378                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1379                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1380
1381                 rxm = rcv_pkts[i];
1382
1383                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
1384                         PMD_RX_LOG(ERR, "Packet drop");
1385                         nb_enqueued++;
1386                         virtio_discard_rxbuf_inorder(vq, rxm);
1387                         rxvq->stats.errors++;
1388                         continue;
1389                 }
1390
1391                 header = (struct virtio_net_hdr_mrg_rxbuf *)
1392                          ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1393                          - hdr_size);
1394
1395                 if (vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
1396                         seg_num = header->num_buffers;
1397                         if (seg_num == 0)
1398                                 seg_num = 1;
1399                 } else {
1400                         seg_num = 1;
1401                 }
1402
1403                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1404                 rxm->nb_segs = seg_num;
1405                 rxm->ol_flags = 0;
1406                 rxm->vlan_tci = 0;
1407                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1408                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1409
1410                 rxm->port = rxvq->port_id;
1411
1412                 rx_pkts[nb_rx] = rxm;
1413                 prev = rxm;
1414
1415                 if (vq->hw->has_rx_offload &&
1416                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1417                         virtio_discard_rxbuf_inorder(vq, rxm);
1418                         rxvq->stats.errors++;
1419                         continue;
1420                 }
1421
1422                 if (hw->vlan_strip)
1423                         rte_vlan_strip(rx_pkts[nb_rx]);
1424
1425                 seg_res = seg_num - 1;
1426
1427                 /* Merge remaining segments */
1428                 while (seg_res != 0 && i < (num - 1)) {
1429                         i++;
1430
1431                         rxm = rcv_pkts[i];
1432                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1433                         rxm->pkt_len = (uint32_t)(len[i]);
1434                         rxm->data_len = (uint16_t)(len[i]);
1435
1436                         rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1437                         rx_pkts[nb_rx]->data_len += (uint16_t)(len[i]);
1438
1439                         if (prev)
1440                                 prev->next = rxm;
1441
1442                         prev = rxm;
1443                         seg_res -= 1;
1444                 }
1445
1446                 if (!seg_res) {
1447                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1448                         nb_rx++;
1449                 }
1450         }
1451
1452         /* Last packet still need merge segments */
1453         while (seg_res != 0) {
1454                 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1455                                         VIRTIO_MBUF_BURST_SZ);
1456
1457                 prev = rcv_pkts[nb_rx];
1458                 if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
1459                         num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len,
1460                                                            rcv_cnt);
1461                         uint16_t extra_idx = 0;
1462
1463                         rcv_cnt = num;
1464                         while (extra_idx < rcv_cnt) {
1465                                 rxm = rcv_pkts[extra_idx];
1466                                 rxm->data_off =
1467                                         RTE_PKTMBUF_HEADROOM - hdr_size;
1468                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1469                                 rxm->data_len = (uint16_t)(len[extra_idx]);
1470                                 prev->next = rxm;
1471                                 prev = rxm;
1472                                 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1473                                 rx_pkts[nb_rx]->data_len += len[extra_idx];
1474                                 extra_idx += 1;
1475                         };
1476                         seg_res -= rcv_cnt;
1477
1478                         if (!seg_res) {
1479                                 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1480                                 nb_rx++;
1481                         }
1482                 } else {
1483                         PMD_RX_LOG(ERR,
1484                                         "No enough segments for packet.");
1485                         virtio_discard_rxbuf_inorder(vq, prev);
1486                         rxvq->stats.errors++;
1487                         break;
1488                 }
1489         }
1490
1491         rxvq->stats.packets += nb_rx;
1492
1493         /* Allocate new mbuf for the used descriptor */
1494
1495         if (likely(!virtqueue_full(vq))) {
1496                 /* free_cnt may include mrg descs */
1497                 uint16_t free_cnt = vq->vq_free_cnt;
1498                 struct rte_mbuf *new_pkts[free_cnt];
1499
1500                 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1501                         error = virtqueue_enqueue_refill_inorder(vq, new_pkts,
1502                                         free_cnt);
1503                         if (unlikely(error)) {
1504                                 for (i = 0; i < free_cnt; i++)
1505                                         rte_pktmbuf_free(new_pkts[i]);
1506                         }
1507                         nb_enqueued += free_cnt;
1508                 } else {
1509                         struct rte_eth_dev *dev =
1510                                 &rte_eth_devices[rxvq->port_id];
1511                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1512                 }
1513         }
1514
1515         if (likely(nb_enqueued)) {
1516                 vq_update_avail_idx(vq);
1517
1518                 if (unlikely(virtqueue_kick_prepare(vq))) {
1519                         virtqueue_notify(vq);
1520                         PMD_RX_LOG(DEBUG, "Notified");
1521                 }
1522         }
1523
1524         return nb_rx;
1525 }
1526
1527 uint16_t
1528 virtio_recv_mergeable_pkts(void *rx_queue,
1529                         struct rte_mbuf **rx_pkts,
1530                         uint16_t nb_pkts)
1531 {
1532         struct virtnet_rx *rxvq = rx_queue;
1533         struct virtqueue *vq = rxvq->vq;
1534         struct virtio_hw *hw = vq->hw;
1535         struct rte_mbuf *rxm;
1536         struct rte_mbuf *prev;
1537         uint16_t nb_used, num, nb_rx = 0;
1538         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1539         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1540         int error;
1541         uint32_t nb_enqueued = 0;
1542         uint32_t seg_num = 0;
1543         uint32_t seg_res = 0;
1544         uint32_t hdr_size = hw->vtnet_hdr_size;
1545         int32_t i;
1546
1547         if (unlikely(hw->started == 0))
1548                 return nb_rx;
1549
1550         nb_used = VIRTQUEUE_NUSED(vq);
1551
1552         virtio_rmb(hw->weak_barriers);
1553
1554         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1555
1556         num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1557         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1558                 num = VIRTIO_MBUF_BURST_SZ;
1559         if (likely(num > DESC_PER_CACHELINE))
1560                 num = num - ((vq->vq_used_cons_idx + num) %
1561                                 DESC_PER_CACHELINE);
1562
1563
1564         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1565
1566         for (i = 0; i < num; i++) {
1567                 struct virtio_net_hdr_mrg_rxbuf *header;
1568
1569                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1570                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1571
1572                 rxm = rcv_pkts[i];
1573
1574                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
1575                         PMD_RX_LOG(ERR, "Packet drop");
1576                         nb_enqueued++;
1577                         virtio_discard_rxbuf(vq, rxm);
1578                         rxvq->stats.errors++;
1579                         continue;
1580                 }
1581
1582                 header = (struct virtio_net_hdr_mrg_rxbuf *)
1583                          ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1584                          - hdr_size);
1585                 seg_num = header->num_buffers;
1586                 if (seg_num == 0)
1587                         seg_num = 1;
1588
1589                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1590                 rxm->nb_segs = seg_num;
1591                 rxm->ol_flags = 0;
1592                 rxm->vlan_tci = 0;
1593                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1594                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1595
1596                 rxm->port = rxvq->port_id;
1597
1598                 rx_pkts[nb_rx] = rxm;
1599                 prev = rxm;
1600
1601                 if (hw->has_rx_offload &&
1602                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1603                         virtio_discard_rxbuf(vq, rxm);
1604                         rxvq->stats.errors++;
1605                         continue;
1606                 }
1607
1608                 if (hw->vlan_strip)
1609                         rte_vlan_strip(rx_pkts[nb_rx]);
1610
1611                 seg_res = seg_num - 1;
1612
1613                 /* Merge remaining segments */
1614                 while (seg_res != 0 && i < (num - 1)) {
1615                         i++;
1616
1617                         rxm = rcv_pkts[i];
1618                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1619                         rxm->pkt_len = (uint32_t)(len[i]);
1620                         rxm->data_len = (uint16_t)(len[i]);
1621
1622                         rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1623                         rx_pkts[nb_rx]->data_len += (uint16_t)(len[i]);
1624
1625                         if (prev)
1626                                 prev->next = rxm;
1627
1628                         prev = rxm;
1629                         seg_res -= 1;
1630                 }
1631
1632                 if (!seg_res) {
1633                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1634                         nb_rx++;
1635                 }
1636         }
1637
1638         /* Last packet still need merge segments */
1639         while (seg_res != 0) {
1640                 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1641                                         VIRTIO_MBUF_BURST_SZ);
1642
1643                 prev = rcv_pkts[nb_rx];
1644                 if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
1645                         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len,
1646                                                            rcv_cnt);
1647                         uint16_t extra_idx = 0;
1648
1649                         rcv_cnt = num;
1650                         while (extra_idx < rcv_cnt) {
1651                                 rxm = rcv_pkts[extra_idx];
1652                                 rxm->data_off =
1653                                         RTE_PKTMBUF_HEADROOM - hdr_size;
1654                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1655                                 rxm->data_len = (uint16_t)(len[extra_idx]);
1656                                 prev->next = rxm;
1657                                 prev = rxm;
1658                                 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1659                                 rx_pkts[nb_rx]->data_len += len[extra_idx];
1660                                 extra_idx += 1;
1661                         };
1662                         seg_res -= rcv_cnt;
1663
1664                         if (!seg_res) {
1665                                 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1666                                 nb_rx++;
1667                         }
1668                 } else {
1669                         PMD_RX_LOG(ERR,
1670                                         "No enough segments for packet.");
1671                         virtio_discard_rxbuf(vq, prev);
1672                         rxvq->stats.errors++;
1673                         break;
1674                 }
1675         }
1676
1677         rxvq->stats.packets += nb_rx;
1678
1679         /* Allocate new mbuf for the used descriptor */
1680         if (likely(!virtqueue_full(vq))) {
1681                 /* free_cnt may include mrg descs */
1682                 uint16_t free_cnt = vq->vq_free_cnt;
1683                 struct rte_mbuf *new_pkts[free_cnt];
1684
1685                 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1686                         error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1687                                         free_cnt);
1688                         if (unlikely(error)) {
1689                                 for (i = 0; i < free_cnt; i++)
1690                                         rte_pktmbuf_free(new_pkts[i]);
1691                         }
1692                         nb_enqueued += free_cnt;
1693                 } else {
1694                         struct rte_eth_dev *dev =
1695                                 &rte_eth_devices[rxvq->port_id];
1696                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1697                 }
1698         }
1699
1700         if (likely(nb_enqueued)) {
1701                 vq_update_avail_idx(vq);
1702
1703                 if (unlikely(virtqueue_kick_prepare(vq))) {
1704                         virtqueue_notify(vq);
1705                         PMD_RX_LOG(DEBUG, "Notified");
1706                 }
1707         }
1708
1709         return nb_rx;
1710 }
1711
1712 uint16_t
1713 virtio_recv_mergeable_pkts_packed(void *rx_queue,
1714                         struct rte_mbuf **rx_pkts,
1715                         uint16_t nb_pkts)
1716 {
1717         struct virtnet_rx *rxvq = rx_queue;
1718         struct virtqueue *vq = rxvq->vq;
1719         struct virtio_hw *hw = vq->hw;
1720         struct rte_mbuf *rxm;
1721         struct rte_mbuf *prev = NULL;
1722         uint16_t num, nb_rx = 0;
1723         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1724         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1725         uint32_t nb_enqueued = 0;
1726         uint32_t seg_num = 0;
1727         uint32_t seg_res = 0;
1728         uint32_t hdr_size = hw->vtnet_hdr_size;
1729         int32_t i;
1730         int error;
1731
1732         if (unlikely(hw->started == 0))
1733                 return nb_rx;
1734
1735
1736         num = nb_pkts;
1737         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1738                 num = VIRTIO_MBUF_BURST_SZ;
1739         if (likely(num > DESC_PER_CACHELINE))
1740                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1741
1742         num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1743
1744         for (i = 0; i < num; i++) {
1745                 struct virtio_net_hdr_mrg_rxbuf *header;
1746
1747                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1748                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1749
1750                 rxm = rcv_pkts[i];
1751
1752                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
1753                         PMD_RX_LOG(ERR, "Packet drop");
1754                         nb_enqueued++;
1755                         virtio_discard_rxbuf(vq, rxm);
1756                         rxvq->stats.errors++;
1757                         continue;
1758                 }
1759
1760                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)
1761                           rxm->buf_addr + RTE_PKTMBUF_HEADROOM - hdr_size);
1762                 seg_num = header->num_buffers;
1763
1764                 if (seg_num == 0)
1765                         seg_num = 1;
1766
1767                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1768                 rxm->nb_segs = seg_num;
1769                 rxm->ol_flags = 0;
1770                 rxm->vlan_tci = 0;
1771                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1772                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1773
1774                 rxm->port = rxvq->port_id;
1775                 rx_pkts[nb_rx] = rxm;
1776                 prev = rxm;
1777
1778                 if (hw->has_rx_offload &&
1779                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1780                         virtio_discard_rxbuf(vq, rxm);
1781                         rxvq->stats.errors++;
1782                         continue;
1783                 }
1784
1785                 if (hw->vlan_strip)
1786                         rte_vlan_strip(rx_pkts[nb_rx]);
1787
1788                 seg_res = seg_num - 1;
1789
1790                 /* Merge remaining segments */
1791                 while (seg_res != 0 && i < (num - 1)) {
1792                         i++;
1793
1794                         rxm = rcv_pkts[i];
1795                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1796                         rxm->pkt_len = (uint32_t)(len[i]);
1797                         rxm->data_len = (uint16_t)(len[i]);
1798
1799                         rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1800                         rx_pkts[nb_rx]->data_len += (uint16_t)(len[i]);
1801
1802                         if (prev)
1803                                 prev->next = rxm;
1804
1805                         prev = rxm;
1806                         seg_res -= 1;
1807                 }
1808
1809                 if (!seg_res) {
1810                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1811                         nb_rx++;
1812                 }
1813         }
1814
1815         /* Last packet still need merge segments */
1816         while (seg_res != 0) {
1817                 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1818                                         VIRTIO_MBUF_BURST_SZ);
1819                 if (likely(vq->vq_free_cnt >= rcv_cnt)) {
1820                         num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts,
1821                                         len, rcv_cnt);
1822                         uint16_t extra_idx = 0;
1823
1824                         rcv_cnt = num;
1825
1826                         while (extra_idx < rcv_cnt) {
1827                                 rxm = rcv_pkts[extra_idx];
1828
1829                                 rxm->data_off =
1830                                         RTE_PKTMBUF_HEADROOM - hdr_size;
1831                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1832                                 rxm->data_len = (uint16_t)(len[extra_idx]);
1833
1834                                 prev->next = rxm;
1835                                 prev = rxm;
1836                                 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1837                                 rx_pkts[nb_rx]->data_len += len[extra_idx];
1838                                 extra_idx += 1;
1839                         }
1840                         seg_res -= rcv_cnt;
1841                         if (!seg_res) {
1842                                 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1843                                 nb_rx++;
1844                         }
1845                 } else {
1846                         PMD_RX_LOG(ERR,
1847                                         "No enough segments for packet.");
1848                         if (prev)
1849                                 virtio_discard_rxbuf(vq, prev);
1850                         rxvq->stats.errors++;
1851                         break;
1852                 }
1853         }
1854
1855         rxvq->stats.packets += nb_rx;
1856
1857         /* Allocate new mbuf for the used descriptor */
1858         if (likely(!virtqueue_full(vq))) {
1859                 /* free_cnt may include mrg descs */
1860                 uint16_t free_cnt = vq->vq_free_cnt;
1861                 struct rte_mbuf *new_pkts[free_cnt];
1862
1863                 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1864                         error = virtqueue_enqueue_recv_refill_packed(vq,
1865                                         new_pkts, free_cnt);
1866                         if (unlikely(error)) {
1867                                 for (i = 0; i < free_cnt; i++)
1868                                         rte_pktmbuf_free(new_pkts[i]);
1869                         }
1870                         nb_enqueued += free_cnt;
1871                 } else {
1872                         struct rte_eth_dev *dev =
1873                                 &rte_eth_devices[rxvq->port_id];
1874                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1875                 }
1876         }
1877
1878         if (likely(nb_enqueued)) {
1879                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1880                         virtqueue_notify(vq);
1881                         PMD_RX_LOG(DEBUG, "Notified");
1882                 }
1883         }
1884
1885         return nb_rx;
1886 }
1887
1888 uint16_t
1889 virtio_xmit_pkts_packed(void *tx_queue, struct rte_mbuf **tx_pkts,
1890                         uint16_t nb_pkts)
1891 {
1892         struct virtnet_tx *txvq = tx_queue;
1893         struct virtqueue *vq = txvq->vq;
1894         struct virtio_hw *hw = vq->hw;
1895         uint16_t hdr_size = hw->vtnet_hdr_size;
1896         uint16_t nb_tx = 0;
1897         int error;
1898
1899         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1900                 return nb_tx;
1901
1902         if (unlikely(nb_pkts < 1))
1903                 return nb_pkts;
1904
1905         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1906
1907         if (nb_pkts > vq->vq_free_cnt)
1908                 virtio_xmit_cleanup_packed(vq, nb_pkts - vq->vq_free_cnt);
1909
1910         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1911                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1912                 int can_push = 0, slots, need;
1913
1914                 /* Do VLAN tag insertion */
1915                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1916                         error = rte_vlan_insert(&txm);
1917                         if (unlikely(error)) {
1918                                 rte_pktmbuf_free(txm);
1919                                 continue;
1920                         }
1921                 }
1922
1923                 /* optimize ring usage */
1924                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1925                       vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1926                     rte_mbuf_refcnt_read(txm) == 1 &&
1927                     RTE_MBUF_DIRECT(txm) &&
1928                     txm->nb_segs == 1 &&
1929                     rte_pktmbuf_headroom(txm) >= hdr_size &&
1930                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1931                            __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1932                         can_push = 1;
1933
1934                 /* How many main ring entries are needed to this Tx?
1935                  * any_layout => number of segments
1936                  * default    => number of segments + 1
1937                  */
1938                 slots = txm->nb_segs + !can_push;
1939                 need = slots - vq->vq_free_cnt;
1940
1941                 /* Positive value indicates it need free vring descriptors */
1942                 if (unlikely(need > 0)) {
1943                         virtio_rmb(hw->weak_barriers);
1944                         need = RTE_MIN(need, (int)nb_pkts);
1945                         virtio_xmit_cleanup_packed(vq, need);
1946                         need = slots - vq->vq_free_cnt;
1947                         if (unlikely(need > 0)) {
1948                                 PMD_TX_LOG(ERR,
1949                                            "No free tx descriptors to transmit");
1950                                 break;
1951                         }
1952                 }
1953
1954                 /* Enqueue Packet buffers */
1955                 virtqueue_enqueue_xmit_packed(txvq, txm, slots, can_push);
1956
1957                 virtio_update_packet_stats(&txvq->stats, txm);
1958         }
1959
1960         txvq->stats.packets += nb_tx;
1961
1962         if (likely(nb_tx)) {
1963                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1964                         virtqueue_notify(vq);
1965                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1966                 }
1967         }
1968
1969         return nb_tx;
1970 }
1971
1972 uint16_t
1973 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1974 {
1975         struct virtnet_tx *txvq = tx_queue;
1976         struct virtqueue *vq = txvq->vq;
1977         struct virtio_hw *hw = vq->hw;
1978         uint16_t hdr_size = hw->vtnet_hdr_size;
1979         uint16_t nb_used, nb_tx = 0;
1980         int error;
1981
1982         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1983                 return nb_tx;
1984
1985         if (unlikely(nb_pkts < 1))
1986                 return nb_pkts;
1987
1988         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1989         nb_used = VIRTQUEUE_NUSED(vq);
1990
1991         virtio_rmb(hw->weak_barriers);
1992         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1993                 virtio_xmit_cleanup(vq, nb_used);
1994
1995         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1996                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1997                 int can_push = 0, use_indirect = 0, slots, need;
1998
1999                 /* Do VLAN tag insertion */
2000                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
2001                         error = rte_vlan_insert(&txm);
2002                         if (unlikely(error)) {
2003                                 rte_pktmbuf_free(txm);
2004                                 continue;
2005                         }
2006                 }
2007
2008                 /* optimize ring usage */
2009                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
2010                       vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
2011                     rte_mbuf_refcnt_read(txm) == 1 &&
2012                     RTE_MBUF_DIRECT(txm) &&
2013                     txm->nb_segs == 1 &&
2014                     rte_pktmbuf_headroom(txm) >= hdr_size &&
2015                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
2016                                    __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
2017                         can_push = 1;
2018                 else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
2019                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
2020                         use_indirect = 1;
2021
2022                 /* How many main ring entries are needed to this Tx?
2023                  * any_layout => number of segments
2024                  * indirect   => 1
2025                  * default    => number of segments + 1
2026                  */
2027                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
2028                 need = slots - vq->vq_free_cnt;
2029
2030                 /* Positive value indicates it need free vring descriptors */
2031                 if (unlikely(need > 0)) {
2032                         nb_used = VIRTQUEUE_NUSED(vq);
2033                         virtio_rmb(hw->weak_barriers);
2034                         need = RTE_MIN(need, (int)nb_used);
2035
2036                         virtio_xmit_cleanup(vq, need);
2037                         need = slots - vq->vq_free_cnt;
2038                         if (unlikely(need > 0)) {
2039                                 PMD_TX_LOG(ERR,
2040                                            "No free tx descriptors to transmit");
2041                                 break;
2042                         }
2043                 }
2044
2045                 /* Enqueue Packet buffers */
2046                 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
2047                         can_push, 0);
2048
2049                 virtio_update_packet_stats(&txvq->stats, txm);
2050         }
2051
2052         txvq->stats.packets += nb_tx;
2053
2054         if (likely(nb_tx)) {
2055                 vq_update_avail_idx(vq);
2056
2057                 if (unlikely(virtqueue_kick_prepare(vq))) {
2058                         virtqueue_notify(vq);
2059                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2060                 }
2061         }
2062
2063         return nb_tx;
2064 }
2065
2066 uint16_t
2067 virtio_xmit_pkts_inorder(void *tx_queue,
2068                         struct rte_mbuf **tx_pkts,
2069                         uint16_t nb_pkts)
2070 {
2071         struct virtnet_tx *txvq = tx_queue;
2072         struct virtqueue *vq = txvq->vq;
2073         struct virtio_hw *hw = vq->hw;
2074         uint16_t hdr_size = hw->vtnet_hdr_size;
2075         uint16_t nb_used, nb_avail, nb_tx = 0, nb_inorder_pkts = 0;
2076         struct rte_mbuf *inorder_pkts[nb_pkts];
2077         int error;
2078
2079         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
2080                 return nb_tx;
2081
2082         if (unlikely(nb_pkts < 1))
2083                 return nb_pkts;
2084
2085         VIRTQUEUE_DUMP(vq);
2086         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
2087         nb_used = VIRTQUEUE_NUSED(vq);
2088
2089         virtio_rmb(hw->weak_barriers);
2090         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
2091                 virtio_xmit_cleanup_inorder(vq, nb_used);
2092
2093         if (unlikely(!vq->vq_free_cnt))
2094                 virtio_xmit_cleanup_inorder(vq, nb_used);
2095
2096         nb_avail = RTE_MIN(vq->vq_free_cnt, nb_pkts);
2097
2098         for (nb_tx = 0; nb_tx < nb_avail; nb_tx++) {
2099                 struct rte_mbuf *txm = tx_pkts[nb_tx];
2100                 int slots, need;
2101
2102                 /* Do VLAN tag insertion */
2103                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
2104                         error = rte_vlan_insert(&txm);
2105                         if (unlikely(error)) {
2106                                 rte_pktmbuf_free(txm);
2107                                 continue;
2108                         }
2109                 }
2110
2111                 /* optimize ring usage */
2112                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
2113                      vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
2114                      rte_mbuf_refcnt_read(txm) == 1 &&
2115                      RTE_MBUF_DIRECT(txm) &&
2116                      txm->nb_segs == 1 &&
2117                      rte_pktmbuf_headroom(txm) >= hdr_size &&
2118                      rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
2119                                 __alignof__(struct virtio_net_hdr_mrg_rxbuf))) {
2120                         inorder_pkts[nb_inorder_pkts] = txm;
2121                         nb_inorder_pkts++;
2122
2123                         virtio_update_packet_stats(&txvq->stats, txm);
2124                         continue;
2125                 }
2126
2127                 if (nb_inorder_pkts) {
2128                         virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2129                                                         nb_inorder_pkts);
2130                         nb_inorder_pkts = 0;
2131                 }
2132
2133                 slots = txm->nb_segs + 1;
2134                 need = slots - vq->vq_free_cnt;
2135                 if (unlikely(need > 0)) {
2136                         nb_used = VIRTQUEUE_NUSED(vq);
2137                         virtio_rmb(hw->weak_barriers);
2138                         need = RTE_MIN(need, (int)nb_used);
2139
2140                         virtio_xmit_cleanup_inorder(vq, need);
2141
2142                         need = slots - vq->vq_free_cnt;
2143
2144                         if (unlikely(need > 0)) {
2145                                 PMD_TX_LOG(ERR,
2146                                         "No free tx descriptors to transmit");
2147                                 break;
2148                         }
2149                 }
2150                 /* Enqueue Packet buffers */
2151                 virtqueue_enqueue_xmit(txvq, txm, slots, 0, 0, 1);
2152
2153                 virtio_update_packet_stats(&txvq->stats, txm);
2154         }
2155
2156         /* Transmit all inorder packets */
2157         if (nb_inorder_pkts)
2158                 virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2159                                                 nb_inorder_pkts);
2160
2161         txvq->stats.packets += nb_tx;
2162
2163         if (likely(nb_tx)) {
2164                 vq_update_avail_idx(vq);
2165
2166                 if (unlikely(virtqueue_kick_prepare(vq))) {
2167                         virtqueue_notify(vq);
2168                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2169                 }
2170         }
2171
2172         VIRTQUEUE_DUMP(vq);
2173
2174         return nb_tx;
2175 }