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_split.ring.desc[desc_idx];
  66         dxp = &vq->vq_descx[desc_idx];
  67         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
  68         if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
  69                 while (dp->flags & VRING_DESC_F_NEXT) {
  70                         desc_idx_last = dp->next;
  71                         dp = &vq->vq_split.ring.desc[dp->next];
  72                 }
  73         }
  74         dxp->ndescs = 0;
  75
  76         /*
  77          * We must append the existing free chain, if any, to the end of
  78          * newly freed chain. If the virtqueue was completely used, then
  79          * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
  80          */
  81         if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
  82                 vq->vq_desc_head_idx = desc_idx;
  83         } else {
  84                 dp_tail = &vq->vq_split.ring.desc[vq->vq_desc_tail_idx];
  85                 dp_tail->next = desc_idx;
  86         }
  87
  88         vq->vq_desc_tail_idx = desc_idx_last;
  89         dp->next = VQ_RING_DESC_CHAIN_END;
  90 }
  91
  92 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->vq_packed.ring.desc;
 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                 virtio_rmb(vq->hw->weak_barriers);
 128                 len[i] = desc[used_idx].len;
 129                 id = desc[used_idx].id;
 130                 cookie = (struct rte_mbuf *)vq->vq_descx[id].cookie;
 131                 if (unlikely(cookie == NULL)) {
 132                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
 133                                 vq->vq_used_cons_idx);
 134                         break;
 135                 }
 136                 rte_prefetch0(cookie);
 137                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 138                 rx_pkts[i] = cookie;
 139
 140                 vq->vq_free_cnt++;
 141                 vq->vq_used_cons_idx++;
 142                 if (vq->vq_used_cons_idx >= vq->vq_nentries) {
 143                         vq->vq_used_cons_idx -= vq->vq_nentries;
 144                         vq->vq_packed.used_wrap_counter ^= 1;
 145                 }
 146         }
 147
 148         return i;
 149 }
 150
 151 static uint16_t
 152 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
 153                            uint32_t *len, uint16_t num)
 154 {
 155         struct vring_used_elem *uep;
 156         struct rte_mbuf *cookie;
 157         uint16_t used_idx, desc_idx;
 158         uint16_t i;
 159
 160         /*  Caller does the check */
 161         for (i = 0; i < num ; i++) {
 162                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 163                 uep = &vq->vq_split.ring.used->ring[used_idx];
 164                 desc_idx = (uint16_t) uep->id;
 165                 len[i] = uep->len;
 166                 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
 167
 168                 if (unlikely(cookie == NULL)) {
 169                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
 170                                 vq->vq_used_cons_idx);
 171                         break;
 172                 }
 173
 174                 rte_prefetch0(cookie);
 175                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 176                 rx_pkts[i]  = cookie;
 177                 vq->vq_used_cons_idx++;
 178                 vq_ring_free_chain(vq, desc_idx);
 179                 vq->vq_descx[desc_idx].cookie = NULL;
 180         }
 181
 182         return i;
 183 }
 184
 185 static uint16_t
 186 virtqueue_dequeue_rx_inorder(struct virtqueue *vq,
 187                         struct rte_mbuf **rx_pkts,
 188                         uint32_t *len,
 189                         uint16_t num)
 190 {
 191         struct vring_used_elem *uep;
 192         struct rte_mbuf *cookie;
 193         uint16_t used_idx = 0;
 194         uint16_t i;
 195
 196         if (unlikely(num == 0))
 197                 return 0;
 198
 199         for (i = 0; i < num; i++) {
 200                 used_idx = vq->vq_used_cons_idx & (vq->vq_nentries - 1);
 201                 /* Desc idx same as used idx */
 202                 uep = &vq->vq_split.ring.used->ring[used_idx];
 203                 len[i] = uep->len;
 204                 cookie = (struct rte_mbuf *)vq->vq_descx[used_idx].cookie;
 205
 206                 if (unlikely(cookie == NULL)) {
 207                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
 208                                 vq->vq_used_cons_idx);
 209                         break;
 210                 }
 211
 212                 rte_prefetch0(cookie);
 213                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 214                 rx_pkts[i]  = cookie;
 215                 vq->vq_used_cons_idx++;
 216                 vq->vq_descx[used_idx].cookie = NULL;
 217         }
 218
 219         vq_ring_free_inorder(vq, used_idx, i);
 220         return i;
 221 }
 222
 223 #ifndef DEFAULT_TX_FREE_THRESH
 224 #define DEFAULT_TX_FREE_THRESH 32
 225 #endif
 226
 227 static void
 228 virtio_xmit_cleanup_inorder_packed(struct virtqueue *vq, int num)
 229 {
 230         uint16_t used_idx, id, curr_id, free_cnt = 0;
 231         uint16_t size = vq->vq_nentries;
 232         struct vring_packed_desc *desc = vq->vq_packed.ring.desc;
 233         struct vq_desc_extra *dxp;
 234
 235         used_idx = vq->vq_used_cons_idx;
 236         while (num > 0 && desc_is_used(&desc[used_idx], vq)) {
 237                 virtio_rmb(vq->hw->weak_barriers);
 238                 id = desc[used_idx].id;
 239                 do {
 240                         curr_id = used_idx;
 241                         dxp = &vq->vq_descx[used_idx];
 242                         used_idx += dxp->ndescs;
 243                         free_cnt += dxp->ndescs;
 244                         num -= dxp->ndescs;
 245                         if (used_idx >= size) {
 246                                 used_idx -= size;
 247                                 vq->vq_packed.used_wrap_counter ^= 1;
 248                         }
 249                         if (dxp->cookie != NULL) {
 250                                 rte_pktmbuf_free(dxp->cookie);
 251                                 dxp->cookie = NULL;
 252                         }
 253                 } while (curr_id != id);
 254         }
 255         vq->vq_used_cons_idx = used_idx;
 256         vq->vq_free_cnt += free_cnt;
 257 }
 258
 259 static void
 260 virtio_xmit_cleanup_normal_packed(struct virtqueue *vq, int num)
 261 {
 262         uint16_t used_idx, id;
 263         uint16_t size = vq->vq_nentries;
 264         struct vring_packed_desc *desc = vq->vq_packed.ring.desc;
 265         struct vq_desc_extra *dxp;
 266
 267         used_idx = vq->vq_used_cons_idx;
 268         while (num-- && desc_is_used(&desc[used_idx], vq)) {
 269                 virtio_rmb(vq->hw->weak_barriers);
 270                 id = desc[used_idx].id;
 271                 dxp = &vq->vq_descx[id];
 272                 vq->vq_used_cons_idx += dxp->ndescs;
 273                 if (vq->vq_used_cons_idx >= size) {
 274                         vq->vq_used_cons_idx -= size;
 275                         vq->vq_packed.used_wrap_counter ^= 1;
 276                 }
 277                 vq_ring_free_id_packed(vq, id);
 278                 if (dxp->cookie != NULL) {
 279                         rte_pktmbuf_free(dxp->cookie);
 280                         dxp->cookie = NULL;
 281                 }
 282                 used_idx = vq->vq_used_cons_idx;
 283         }
 284 }
 285
 286 /* Cleanup from completed transmits. */
 287 static inline void
 288 virtio_xmit_cleanup_packed(struct virtqueue *vq, int num, int in_order)
 289 {
 290         if (in_order)
 291                 virtio_xmit_cleanup_inorder_packed(vq, num);
 292         else
 293                 virtio_xmit_cleanup_normal_packed(vq, num);
 294 }
 295
 296 static void
 297 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
 298 {
 299         uint16_t i, used_idx, desc_idx;
 300         for (i = 0; i < num; i++) {
 301                 struct vring_used_elem *uep;
 302                 struct vq_desc_extra *dxp;
 303
 304                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 305                 uep = &vq->vq_split.ring.used->ring[used_idx];
 306
 307                 desc_idx = (uint16_t) uep->id;
 308                 dxp = &vq->vq_descx[desc_idx];
 309                 vq->vq_used_cons_idx++;
 310                 vq_ring_free_chain(vq, desc_idx);
 311
 312                 if (dxp->cookie != NULL) {
 313                         rte_pktmbuf_free(dxp->cookie);
 314                         dxp->cookie = NULL;
 315                 }
 316         }
 317 }
 318
 319 /* Cleanup from completed inorder transmits. */
 320 static void
 321 virtio_xmit_cleanup_inorder(struct virtqueue *vq, uint16_t num)
 322 {
 323         uint16_t i, idx = vq->vq_used_cons_idx;
 324         int16_t free_cnt = 0;
 325         struct vq_desc_extra *dxp = NULL;
 326
 327         if (unlikely(num == 0))
 328                 return;
 329
 330         for (i = 0; i < num; i++) {
 331                 dxp = &vq->vq_descx[idx++ & (vq->vq_nentries - 1)];
 332                 free_cnt += dxp->ndescs;
 333                 if (dxp->cookie != NULL) {
 334                         rte_pktmbuf_free(dxp->cookie);
 335                         dxp->cookie = NULL;
 336                 }
 337         }
 338
 339         vq->vq_free_cnt += free_cnt;
 340         vq->vq_used_cons_idx = idx;
 341 }
 342
 343 static inline int
 344 virtqueue_enqueue_refill_inorder(struct virtqueue *vq,
 345                         struct rte_mbuf **cookies,
 346                         uint16_t num)
 347 {
 348         struct vq_desc_extra *dxp;
 349         struct virtio_hw *hw = vq->hw;
 350         struct vring_desc *start_dp;
 351         uint16_t head_idx, idx, i = 0;
 352
 353         if (unlikely(vq->vq_free_cnt == 0))
 354                 return -ENOSPC;
 355         if (unlikely(vq->vq_free_cnt < num))
 356                 return -EMSGSIZE;
 357
 358         head_idx = vq->vq_desc_head_idx & (vq->vq_nentries - 1);
 359         start_dp = vq->vq_split.ring.desc;
 360
 361         while (i < num) {
 362                 idx = head_idx & (vq->vq_nentries - 1);
 363                 dxp = &vq->vq_descx[idx];
 364                 dxp->cookie = (void *)cookies[i];
 365                 dxp->ndescs = 1;
 366
 367                 start_dp[idx].addr =
 368                                 VIRTIO_MBUF_ADDR(cookies[i], vq) +
 369                                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 370                 start_dp[idx].len =
 371                                 cookies[i]->buf_len -
 372                                 RTE_PKTMBUF_HEADROOM +
 373                                 hw->vtnet_hdr_size;
 374                 start_dp[idx].flags =  VRING_DESC_F_WRITE;
 375
 376                 vq_update_avail_ring(vq, idx);
 377                 head_idx++;
 378                 i++;
 379         }
 380
 381         vq->vq_desc_head_idx += num;
 382         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 383         return 0;
 384 }
 385
 386 static inline int
 387 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf **cookie,
 388                                 uint16_t num)
 389 {
 390         struct vq_desc_extra *dxp;
 391         struct virtio_hw *hw = vq->hw;
 392         struct vring_desc *start_dp = vq->vq_split.ring.desc;
 393         uint16_t idx, i;
 394
 395         if (unlikely(vq->vq_free_cnt == 0))
 396                 return -ENOSPC;
 397         if (unlikely(vq->vq_free_cnt < num))
 398                 return -EMSGSIZE;
 399
 400         if (unlikely(vq->vq_desc_head_idx >= vq->vq_nentries))
 401                 return -EFAULT;
 402
 403         for (i = 0; i < num; i++) {
 404                 idx = vq->vq_desc_head_idx;
 405                 dxp = &vq->vq_descx[idx];
 406                 dxp->cookie = (void *)cookie[i];
 407                 dxp->ndescs = 1;
 408
 409                 start_dp[idx].addr =
 410                         VIRTIO_MBUF_ADDR(cookie[i], vq) +
 411                         RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 412                 start_dp[idx].len =
 413                         cookie[i]->buf_len - RTE_PKTMBUF_HEADROOM +
 414                         hw->vtnet_hdr_size;
 415                 start_dp[idx].flags = VRING_DESC_F_WRITE;
 416                 vq->vq_desc_head_idx = start_dp[idx].next;
 417                 vq_update_avail_ring(vq, idx);
 418                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END) {
 419                         vq->vq_desc_tail_idx = vq->vq_desc_head_idx;
 420                         break;
 421                 }
 422         }
 423
 424         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 425
 426         return 0;
 427 }
 428
 429 static inline int
 430 virtqueue_enqueue_recv_refill_packed(struct virtqueue *vq,
 431                                      struct rte_mbuf **cookie, uint16_t num)
 432 {
 433         struct vring_packed_desc *start_dp = vq->vq_packed.ring.desc;
 434         uint16_t flags = vq->vq_packed.cached_flags;
 435         struct virtio_hw *hw = vq->hw;
 436         struct vq_desc_extra *dxp;
 437         uint16_t idx;
 438         int i;
 439
 440         if (unlikely(vq->vq_free_cnt == 0))
 441                 return -ENOSPC;
 442         if (unlikely(vq->vq_free_cnt < num))
 443                 return -EMSGSIZE;
 444
 445         for (i = 0; i < num; i++) {
 446                 idx = vq->vq_avail_idx;
 447                 dxp = &vq->vq_descx[idx];
 448                 dxp->cookie = (void *)cookie[i];
 449                 dxp->ndescs = 1;
 450
 451                 start_dp[idx].addr = VIRTIO_MBUF_ADDR(cookie[i], vq) +
 452                                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 453                 start_dp[idx].len = cookie[i]->buf_len - RTE_PKTMBUF_HEADROOM
 454                                         + hw->vtnet_hdr_size;
 455
 456                 vq->vq_desc_head_idx = dxp->next;
 457                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 458                         vq->vq_desc_tail_idx = vq->vq_desc_head_idx;
 459                 virtio_wmb(hw->weak_barriers);
 460                 start_dp[idx].flags = flags;
 461                 if (++vq->vq_avail_idx >= vq->vq_nentries) {
 462                         vq->vq_avail_idx -= vq->vq_nentries;
 463                         vq->vq_packed.cached_flags ^=
 464                                 VRING_PACKED_DESC_F_AVAIL_USED;
 465                         flags = vq->vq_packed.cached_flags;
 466                 }
 467         }
 468         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 469         return 0;
 470 }
 471
 472 /* When doing TSO, the IP length is not included in the pseudo header
 473  * checksum of the packet given to the PMD, but for virtio it is
 474  * expected.
 475  */
 476 static void
 477 virtio_tso_fix_cksum(struct rte_mbuf *m)
 478 {
 479         /* common case: header is not fragmented */
 480         if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
 481                         m->l4_len)) {
 482                 struct rte_ipv4_hdr *iph;
 483                 struct rte_ipv6_hdr *ip6h;
 484                 struct rte_tcp_hdr *th;
 485                 uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
 486                 uint32_t tmp;
 487
 488                 iph = rte_pktmbuf_mtod_offset(m,
 489                                         struct rte_ipv4_hdr *, m->l2_len);
 490                 th = RTE_PTR_ADD(iph, m->l3_len);
 491                 if ((iph->version_ihl >> 4) == 4) {
 492                         iph->hdr_checksum = 0;
 493                         iph->hdr_checksum = rte_ipv4_cksum(iph);
 494                         ip_len = iph->total_length;
 495                         ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
 496                                 m->l3_len);
 497                 } else {
 498                         ip6h = (struct rte_ipv6_hdr *)iph;
 499                         ip_paylen = ip6h->payload_len;
 500                 }
 501
 502                 /* calculate the new phdr checksum not including ip_paylen */
 503                 prev_cksum = th->cksum;
 504                 tmp = prev_cksum;
 505                 tmp += ip_paylen;
 506                 tmp = (tmp & 0xffff) + (tmp >> 16);
 507                 new_cksum = tmp;
 508
 509                 /* replace it in the packet */
 510                 th->cksum = new_cksum;
 511         }
 512 }
 513
 514
 515 /* avoid write operation when necessary, to lessen cache issues */
 516 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 517         if ((var) != (val))                     \
 518                 (var) = (val);                  \
 519 } while (0)
 520
 521 #define virtqueue_clear_net_hdr(_hdr) do {              \
 522         ASSIGN_UNLESS_EQUAL((_hdr)->csum_start, 0);     \
 523         ASSIGN_UNLESS_EQUAL((_hdr)->csum_offset, 0);    \
 524         ASSIGN_UNLESS_EQUAL((_hdr)->flags, 0);          \
 525         ASSIGN_UNLESS_EQUAL((_hdr)->gso_type, 0);       \
 526         ASSIGN_UNLESS_EQUAL((_hdr)->gso_size, 0);       \
 527         ASSIGN_UNLESS_EQUAL((_hdr)->hdr_len, 0);        \
 528 } while (0)
 529
 530 static inline void
 531 virtqueue_xmit_offload(struct virtio_net_hdr *hdr,
 532                         struct rte_mbuf *cookie,
 533                         bool offload)
 534 {
 535         if (offload) {
 536                 if (cookie->ol_flags & PKT_TX_TCP_SEG)
 537                         cookie->ol_flags |= PKT_TX_TCP_CKSUM;
 538
 539                 switch (cookie->ol_flags & PKT_TX_L4_MASK) {
 540                 case PKT_TX_UDP_CKSUM:
 541                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 542                         hdr->csum_offset = offsetof(struct rte_udp_hdr,
 543                                 dgram_cksum);
 544                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 545                         break;
 546
 547                 case PKT_TX_TCP_CKSUM:
 548                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 549                         hdr->csum_offset = offsetof(struct rte_tcp_hdr, cksum);
 550                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 551                         break;
 552
 553                 default:
 554                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 555                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 556                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 557                         break;
 558                 }
 559
 560                 /* TCP Segmentation Offload */
 561                 if (cookie->ol_flags & PKT_TX_TCP_SEG) {
 562                         virtio_tso_fix_cksum(cookie);
 563                         hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
 564                                 VIRTIO_NET_HDR_GSO_TCPV6 :
 565                                 VIRTIO_NET_HDR_GSO_TCPV4;
 566                         hdr->gso_size = cookie->tso_segsz;
 567                         hdr->hdr_len =
 568                                 cookie->l2_len +
 569                                 cookie->l3_len +
 570                                 cookie->l4_len;
 571                 } else {
 572                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 573                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 574                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 575                 }
 576         }
 577 }
 578
 579 static inline void
 580 virtqueue_enqueue_xmit_inorder(struct virtnet_tx *txvq,
 581                         struct rte_mbuf **cookies,
 582                         uint16_t num)
 583 {
 584         struct vq_desc_extra *dxp;
 585         struct virtqueue *vq = txvq->vq;
 586         struct vring_desc *start_dp;
 587         struct virtio_net_hdr *hdr;
 588         uint16_t idx;
 589         uint16_t head_size = vq->hw->vtnet_hdr_size;
 590         uint16_t i = 0;
 591
 592         idx = vq->vq_desc_head_idx;
 593         start_dp = vq->vq_split.ring.desc;
 594
 595         while (i < num) {
 596                 idx = idx & (vq->vq_nentries - 1);
 597                 dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
 598                 dxp->cookie = (void *)cookies[i];
 599                 dxp->ndescs = 1;
 600
 601                 hdr = (struct virtio_net_hdr *)
 602                         rte_pktmbuf_prepend(cookies[i], head_size);
 603                 cookies[i]->pkt_len -= head_size;
 604
 605                 /* if offload disabled, hdr is not zeroed yet, do it now */
 606                 if (!vq->hw->has_tx_offload)
 607                         virtqueue_clear_net_hdr(hdr);
 608                 else
 609                         virtqueue_xmit_offload(hdr, cookies[i], true);
 610
 611                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookies[i], vq);
 612                 start_dp[idx].len   = cookies[i]->data_len;
 613                 start_dp[idx].flags = 0;
 614
 615                 vq_update_avail_ring(vq, idx);
 616
 617                 idx++;
 618                 i++;
 619         };
 620
 621         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 622         vq->vq_desc_head_idx = idx & (vq->vq_nentries - 1);
 623 }
 624
 625 static inline void
 626 virtqueue_enqueue_xmit_packed_fast(struct virtnet_tx *txvq,
 627                                    struct rte_mbuf *cookie,
 628                                    int in_order)
 629 {
 630         struct virtqueue *vq = txvq->vq;
 631         struct vring_packed_desc *dp;
 632         struct vq_desc_extra *dxp;
 633         uint16_t idx, id, flags;
 634         uint16_t head_size = vq->hw->vtnet_hdr_size;
 635         struct virtio_net_hdr *hdr;
 636
 637         id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx;
 638         idx = vq->vq_avail_idx;
 639         dp = &vq->vq_packed.ring.desc[idx];
 640
 641         dxp = &vq->vq_descx[id];
 642         dxp->ndescs = 1;
 643         dxp->cookie = cookie;
 644
 645         flags = vq->vq_packed.cached_flags;
 646
 647         /* prepend cannot fail, checked by caller */
 648         hdr = (struct virtio_net_hdr *)
 649                 rte_pktmbuf_prepend(cookie, head_size);
 650         cookie->pkt_len -= head_size;
 651
 652         /* if offload disabled, hdr is not zeroed yet, do it now */
 653         if (!vq->hw->has_tx_offload)
 654                 virtqueue_clear_net_hdr(hdr);
 655         else
 656                 virtqueue_xmit_offload(hdr, cookie, true);
 657
 658         dp->addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
 659         dp->len  = cookie->data_len;
 660         dp->id   = id;
 661
 662         if (++vq->vq_avail_idx >= vq->vq_nentries) {
 663                 vq->vq_avail_idx -= vq->vq_nentries;
 664                 vq->vq_packed.cached_flags ^= VRING_PACKED_DESC_F_AVAIL_USED;
 665         }
 666
 667         vq->vq_free_cnt--;
 668
 669         if (!in_order) {
 670                 vq->vq_desc_head_idx = dxp->next;
 671                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 672                         vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
 673         }
 674
 675         virtio_wmb(vq->hw->weak_barriers);
 676         dp->flags = flags;
 677 }
 678
 679 static inline void
 680 virtqueue_enqueue_xmit_packed(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
 681                               uint16_t needed, int can_push, int in_order)
 682 {
 683         struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 684         struct vq_desc_extra *dxp;
 685         struct virtqueue *vq = txvq->vq;
 686         struct vring_packed_desc *start_dp, *head_dp;
 687         uint16_t idx, id, head_idx, head_flags;
 688         uint16_t head_size = vq->hw->vtnet_hdr_size;
 689         struct virtio_net_hdr *hdr;
 690         uint16_t prev;
 691
 692         id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx;
 693
 694         dxp = &vq->vq_descx[id];
 695         dxp->ndescs = needed;
 696         dxp->cookie = cookie;
 697
 698         head_idx = vq->vq_avail_idx;
 699         idx = head_idx;
 700         prev = head_idx;
 701         start_dp = vq->vq_packed.ring.desc;
 702
 703         head_dp = &vq->vq_packed.ring.desc[idx];
 704         head_flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 705         head_flags |= vq->vq_packed.cached_flags;
 706
 707         if (can_push) {
 708                 /* prepend cannot fail, checked by caller */
 709                 hdr = (struct virtio_net_hdr *)
 710                         rte_pktmbuf_prepend(cookie, head_size);
 711                 /* rte_pktmbuf_prepend() counts the hdr size to the pkt length,
 712                  * which is wrong. Below subtract restores correct pkt size.
 713                  */
 714                 cookie->pkt_len -= head_size;
 715
 716                 /* if offload disabled, it is not zeroed below, do it now */
 717                 if (!vq->hw->has_tx_offload)
 718                         virtqueue_clear_net_hdr(hdr);
 719         } else {
 720                 /* setup first tx ring slot to point to header
 721                  * stored in reserved region.
 722                  */
 723                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 724                         RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
 725                 start_dp[idx].len   = vq->hw->vtnet_hdr_size;
 726                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 727                 idx++;
 728                 if (idx >= vq->vq_nentries) {
 729                         idx -= vq->vq_nentries;
 730                         vq->vq_packed.cached_flags ^=
 731                                 VRING_PACKED_DESC_F_AVAIL_USED;
 732                 }
 733         }
 734
 735         virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload);
 736
 737         do {
 738                 uint16_t flags;
 739
 740                 start_dp[idx].addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
 741                 start_dp[idx].len  = cookie->data_len;
 742                 if (likely(idx != head_idx)) {
 743                         flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 744                         flags |= vq->vq_packed.cached_flags;
 745                         start_dp[idx].flags = flags;
 746                 }
 747                 prev = idx;
 748                 idx++;
 749                 if (idx >= vq->vq_nentries) {
 750                         idx -= vq->vq_nentries;
 751                         vq->vq_packed.cached_flags ^=
 752                                 VRING_PACKED_DESC_F_AVAIL_USED;
 753                 }
 754         } while ((cookie = cookie->next) != NULL);
 755
 756         start_dp[prev].id = id;
 757
 758         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 759         vq->vq_avail_idx = idx;
 760
 761         if (!in_order) {
 762                 vq->vq_desc_head_idx = dxp->next;
 763                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 764                         vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
 765         }
 766
 767         virtio_wmb(vq->hw->weak_barriers);
 768         head_dp->flags = head_flags;
 769 }
 770
 771 static inline void
 772 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
 773                         uint16_t needed, int use_indirect, int can_push,
 774                         int in_order)
 775 {
 776         struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 777         struct vq_desc_extra *dxp;
 778         struct virtqueue *vq = txvq->vq;
 779         struct vring_desc *start_dp;
 780         uint16_t seg_num = cookie->nb_segs;
 781         uint16_t head_idx, idx;
 782         uint16_t head_size = vq->hw->vtnet_hdr_size;
 783         struct virtio_net_hdr *hdr;
 784
 785         head_idx = vq->vq_desc_head_idx;
 786         idx = head_idx;
 787         if (in_order)
 788                 dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
 789         else
 790                 dxp = &vq->vq_descx[idx];
 791         dxp->cookie = (void *)cookie;
 792         dxp->ndescs = needed;
 793
 794         start_dp = vq->vq_split.ring.desc;
 795
 796         if (can_push) {
 797                 /* prepend cannot fail, checked by caller */
 798                 hdr = (struct virtio_net_hdr *)
 799                         rte_pktmbuf_prepend(cookie, head_size);
 800                 /* rte_pktmbuf_prepend() counts the hdr size to the pkt length,
 801                  * which is wrong. Below subtract restores correct pkt size.
 802                  */
 803                 cookie->pkt_len -= head_size;
 804
 805                 /* if offload disabled, it is not zeroed below, do it now */
 806                 if (!vq->hw->has_tx_offload)
 807                         virtqueue_clear_net_hdr(hdr);
 808         } else if (use_indirect) {
 809                 /* setup tx ring slot to point to indirect
 810                  * descriptor list stored in reserved region.
 811                  *
 812                  * the first slot in indirect ring is already preset
 813                  * to point to the header in reserved region
 814                  */
 815                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 816                         RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
 817                 start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
 818                 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
 819                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 820
 821                 /* loop below will fill in rest of the indirect elements */
 822                 start_dp = txr[idx].tx_indir;
 823                 idx = 1;
 824         } else {
 825                 /* setup first tx ring slot to point to header
 826                  * stored in reserved region.
 827                  */
 828                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 829                         RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
 830                 start_dp[idx].len   = vq->hw->vtnet_hdr_size;
 831                 start_dp[idx].flags = VRING_DESC_F_NEXT;
 832                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 833
 834                 idx = start_dp[idx].next;
 835         }
 836
 837         virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload);
 838
 839         do {
 840                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
 841                 start_dp[idx].len   = cookie->data_len;
 842                 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 843                 idx = start_dp[idx].next;
 844         } while ((cookie = cookie->next) != NULL);
 845
 846         if (use_indirect)
 847                 idx = vq->vq_split.ring.desc[head_idx].next;
 848
 849         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 850
 851         vq->vq_desc_head_idx = idx;
 852         vq_update_avail_ring(vq, head_idx);
 853
 854         if (!in_order) {
 855                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 856                         vq->vq_desc_tail_idx = idx;
 857         }
 858 }
 859
 860 void
 861 virtio_dev_cq_start(struct rte_eth_dev *dev)
 862 {
 863         struct virtio_hw *hw = dev->data->dev_private;
 864
 865         if (hw->cvq && hw->cvq->vq) {
 866                 rte_spinlock_init(&hw->cvq->lock);
 867                 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
 868         }
 869 }
 870
 871 int
 872 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
 873                         uint16_t queue_idx,
 874                         uint16_t nb_desc,
 875                         unsigned int socket_id __rte_unused,
 876                         const struct rte_eth_rxconf *rx_conf __rte_unused,
 877                         struct rte_mempool *mp)
 878 {
 879         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 880         struct virtio_hw *hw = dev->data->dev_private;
 881         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 882         struct virtnet_rx *rxvq;
 883
 884         PMD_INIT_FUNC_TRACE();
 885
 886         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 887                 nb_desc = vq->vq_nentries;
 888         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 889
 890         rxvq = &vq->rxq;
 891         rxvq->queue_id = queue_idx;
 892         rxvq->mpool = mp;
 893         dev->data->rx_queues[queue_idx] = rxvq;
 894
 895         return 0;
 896 }
 897
 898 int
 899 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
 900 {
 901         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 902         struct virtio_hw *hw = dev->data->dev_private;
 903         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 904         struct virtnet_rx *rxvq = &vq->rxq;
 905         struct rte_mbuf *m;
 906         uint16_t desc_idx;
 907         int error, nbufs, i;
 908
 909         PMD_INIT_FUNC_TRACE();
 910
 911         /* Allocate blank mbufs for the each rx descriptor */
 912         nbufs = 0;
 913
 914         if (hw->use_simple_rx) {
 915                 for (desc_idx = 0; desc_idx < vq->vq_nentries;
 916                      desc_idx++) {
 917                         vq->vq_split.ring.avail->ring[desc_idx] = desc_idx;
 918                         vq->vq_split.ring.desc[desc_idx].flags =
 919                                 VRING_DESC_F_WRITE;
 920                 }
 921
 922                 virtio_rxq_vec_setup(rxvq);
 923         }
 924
 925         memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
 926         for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
 927              desc_idx++) {
 928                 vq->sw_ring[vq->vq_nentries + desc_idx] =
 929                         &rxvq->fake_mbuf;
 930         }
 931
 932         if (hw->use_simple_rx) {
 933                 while (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
 934                         virtio_rxq_rearm_vec(rxvq);
 935                         nbufs += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
 936                 }
 937         } else if (hw->use_inorder_rx) {
 938                 if ((!virtqueue_full(vq))) {
 939                         uint16_t free_cnt = vq->vq_free_cnt;
 940                         struct rte_mbuf *pkts[free_cnt];
 941
 942                         if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, pkts,
 943                                 free_cnt)) {
 944                                 error = virtqueue_enqueue_refill_inorder(vq,
 945                                                 pkts,
 946                                                 free_cnt);
 947                                 if (unlikely(error)) {
 948                                         for (i = 0; i < free_cnt; i++)
 949                                                 rte_pktmbuf_free(pkts[i]);
 950                                 }
 951                         }
 952
 953                         nbufs += free_cnt;
 954                         vq_update_avail_idx(vq);
 955                 }
 956         } else {
 957                 while (!virtqueue_full(vq)) {
 958                         m = rte_mbuf_raw_alloc(rxvq->mpool);
 959                         if (m == NULL)
 960                                 break;
 961
 962                         /* Enqueue allocated buffers */
 963                         if (vtpci_packed_queue(vq->hw))
 964                                 error = virtqueue_enqueue_recv_refill_packed(vq,
 965                                                 &m, 1);
 966                         else
 967                                 error = virtqueue_enqueue_recv_refill(vq,
 968                                                 &m, 1);
 969                         if (error) {
 970                                 rte_pktmbuf_free(m);
 971                                 break;
 972                         }
 973                         nbufs++;
 974                 }
 975
 976                 if (!vtpci_packed_queue(vq->hw))
 977                         vq_update_avail_idx(vq);
 978         }
 979
 980         PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
 981
 982         VIRTQUEUE_DUMP(vq);
 983
 984         return 0;
 985 }
 986
 987 /*
 988  * struct rte_eth_dev *dev: Used to update dev
 989  * uint16_t nb_desc: Defaults to values read from config space
 990  * unsigned int socket_id: Used to allocate memzone
 991  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
 992  * uint16_t queue_idx: Just used as an index in dev txq list
 993  */
 994 int
 995 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
 996                         uint16_t queue_idx,
 997                         uint16_t nb_desc,
 998                         unsigned int socket_id __rte_unused,
 999                         const struct rte_eth_txconf *tx_conf)
1000 {
1001         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
1002         struct virtio_hw *hw = dev->data->dev_private;
1003         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
1004         struct virtnet_tx *txvq;
1005         uint16_t tx_free_thresh;
1006
1007         PMD_INIT_FUNC_TRACE();
1008
1009         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
1010                 nb_desc = vq->vq_nentries;
1011         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
1012
1013         txvq = &vq->txq;
1014         txvq->queue_id = queue_idx;
1015
1016         tx_free_thresh = tx_conf->tx_free_thresh;
1017         if (tx_free_thresh == 0)
1018                 tx_free_thresh =
1019                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
1020
1021         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
1022                 RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
1023                         "number of TX entries minus 3 (%u)."
1024                         " (tx_free_thresh=%u port=%u queue=%u)\n",
1025                         vq->vq_nentries - 3,
1026                         tx_free_thresh, dev->data->port_id, queue_idx);
1027                 return -EINVAL;
1028         }
1029
1030         vq->vq_free_thresh = tx_free_thresh;
1031
1032         dev->data->tx_queues[queue_idx] = txvq;
1033         return 0;
1034 }
1035
1036 int
1037 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
1038                                 uint16_t queue_idx)
1039 {
1040         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
1041         struct virtio_hw *hw = dev->data->dev_private;
1042         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
1043
1044         PMD_INIT_FUNC_TRACE();
1045
1046         if (!vtpci_packed_queue(hw)) {
1047                 if (hw->use_inorder_tx)
1048                         vq->vq_split.ring.desc[vq->vq_nentries - 1].next = 0;
1049         }
1050
1051         VIRTQUEUE_DUMP(vq);
1052
1053         return 0;
1054 }
1055
1056 static inline void
1057 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
1058 {
1059         int error;
1060         /*
1061          * Requeue the discarded mbuf. This should always be
1062          * successful since it was just dequeued.
1063          */
1064         if (vtpci_packed_queue(vq->hw))
1065                 error = virtqueue_enqueue_recv_refill_packed(vq, &m, 1);
1066         else
1067                 error = virtqueue_enqueue_recv_refill(vq, &m, 1);
1068
1069         if (unlikely(error)) {
1070                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
1071                 rte_pktmbuf_free(m);
1072         }
1073 }
1074
1075 static inline void
1076 virtio_discard_rxbuf_inorder(struct virtqueue *vq, struct rte_mbuf *m)
1077 {
1078         int error;
1079
1080         error = virtqueue_enqueue_refill_inorder(vq, &m, 1);
1081         if (unlikely(error)) {
1082                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
1083                 rte_pktmbuf_free(m);
1084         }
1085 }
1086
1087 static inline void
1088 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
1089 {
1090         uint32_t s = mbuf->pkt_len;
1091         struct rte_ether_addr *ea;
1092
1093         stats->bytes += s;
1094
1095         if (s == 64) {
1096                 stats->size_bins[1]++;
1097         } else if (s > 64 && s < 1024) {
1098                 uint32_t bin;
1099
1100                 /* count zeros, and offset into correct bin */
1101                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
1102                 stats->size_bins[bin]++;
1103         } else {
1104                 if (s < 64)
1105                         stats->size_bins[0]++;
1106                 else if (s < 1519)
1107                         stats->size_bins[6]++;
1108                 else
1109                         stats->size_bins[7]++;
1110         }
1111
1112         ea = rte_pktmbuf_mtod(mbuf, struct rte_ether_addr *);
1113         if (rte_is_multicast_ether_addr(ea)) {
1114                 if (rte_is_broadcast_ether_addr(ea))
1115                         stats->broadcast++;
1116                 else
1117                         stats->multicast++;
1118         }
1119 }
1120
1121 static inline void
1122 virtio_rx_stats_updated(struct virtnet_rx *rxvq, struct rte_mbuf *m)
1123 {
1124         VIRTIO_DUMP_PACKET(m, m->data_len);
1125
1126         virtio_update_packet_stats(&rxvq->stats, m);
1127 }
1128
1129 /* Optionally fill offload information in structure */
1130 static inline int
1131 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
1132 {
1133         struct rte_net_hdr_lens hdr_lens;
1134         uint32_t hdrlen, ptype;
1135         int l4_supported = 0;
1136
1137         /* nothing to do */
1138         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1139                 return 0;
1140
1141         m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
1142
1143         ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
1144         m->packet_type = ptype;
1145         if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
1146             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
1147             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
1148                 l4_supported = 1;
1149
1150         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1151                 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
1152                 if (hdr->csum_start <= hdrlen && l4_supported) {
1153                         m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
1154                 } else {
1155                         /* Unknown proto or tunnel, do sw cksum. We can assume
1156                          * the cksum field is in the first segment since the
1157                          * buffers we provided to the host are large enough.
1158                          * In case of SCTP, this will be wrong since it's a CRC
1159                          * but there's nothing we can do.
1160                          */
1161                         uint16_t csum = 0, off;
1162
1163                         rte_raw_cksum_mbuf(m, hdr->csum_start,
1164                                 rte_pktmbuf_pkt_len(m) - hdr->csum_start,
1165                                 &csum);
1166                         if (likely(csum != 0xffff))
1167                                 csum = ~csum;
1168                         off = hdr->csum_offset + hdr->csum_start;
1169                         if (rte_pktmbuf_data_len(m) >= off + 1)
1170                                 *rte_pktmbuf_mtod_offset(m, uint16_t *,
1171                                         off) = csum;
1172                 }
1173         } else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
1174                 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
1175         }
1176
1177         /* GSO request, save required information in mbuf */
1178         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1179                 /* Check unsupported modes */
1180                 if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
1181                     (hdr->gso_size == 0)) {
1182                         return -EINVAL;
1183                 }
1184
1185                 /* Update mss lengthes in mbuf */
1186                 m->tso_segsz = hdr->gso_size;
1187                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1188                         case VIRTIO_NET_HDR_GSO_TCPV4:
1189                         case VIRTIO_NET_HDR_GSO_TCPV6:
1190                                 m->ol_flags |= PKT_RX_LRO | \
1191                                         PKT_RX_L4_CKSUM_NONE;
1192                                 break;
1193                         default:
1194                                 return -EINVAL;
1195                 }
1196         }
1197
1198         return 0;
1199 }
1200
1201 #define VIRTIO_MBUF_BURST_SZ 64
1202 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
1203 uint16_t
1204 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1205 {
1206         struct virtnet_rx *rxvq = rx_queue;
1207         struct virtqueue *vq = rxvq->vq;
1208         struct virtio_hw *hw = vq->hw;
1209         struct rte_mbuf *rxm;
1210         uint16_t nb_used, num, nb_rx;
1211         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1212         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1213         int error;
1214         uint32_t i, nb_enqueued;
1215         uint32_t hdr_size;
1216         struct virtio_net_hdr *hdr;
1217
1218         nb_rx = 0;
1219         if (unlikely(hw->started == 0))
1220                 return nb_rx;
1221
1222         nb_used = VIRTQUEUE_NUSED(vq);
1223
1224         virtio_rmb(hw->weak_barriers);
1225
1226         num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1227         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1228                 num = VIRTIO_MBUF_BURST_SZ;
1229         if (likely(num > DESC_PER_CACHELINE))
1230                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1231
1232         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1233         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
1234
1235         nb_enqueued = 0;
1236         hdr_size = hw->vtnet_hdr_size;
1237
1238         for (i = 0; i < num ; i++) {
1239                 rxm = rcv_pkts[i];
1240
1241                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1242
1243                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1244                         PMD_RX_LOG(ERR, "Packet drop");
1245                         nb_enqueued++;
1246                         virtio_discard_rxbuf(vq, rxm);
1247                         rxvq->stats.errors++;
1248                         continue;
1249                 }
1250
1251                 rxm->port = rxvq->port_id;
1252                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1253                 rxm->ol_flags = 0;
1254                 rxm->vlan_tci = 0;
1255
1256                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1257                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1258
1259                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1260                         RTE_PKTMBUF_HEADROOM - hdr_size);
1261
1262                 if (hw->vlan_strip)
1263                         rte_vlan_strip(rxm);
1264
1265                 if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1266                         virtio_discard_rxbuf(vq, rxm);
1267                         rxvq->stats.errors++;
1268                         continue;
1269                 }
1270
1271                 virtio_rx_stats_updated(rxvq, rxm);
1272
1273                 rx_pkts[nb_rx++] = rxm;
1274         }
1275
1276         rxvq->stats.packets += nb_rx;
1277
1278         /* Allocate new mbuf for the used descriptor */
1279         if (likely(!virtqueue_full(vq))) {
1280                 uint16_t free_cnt = vq->vq_free_cnt;
1281                 struct rte_mbuf *new_pkts[free_cnt];
1282
1283                 if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1284                                                 free_cnt) == 0)) {
1285                         error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1286                                         free_cnt);
1287                         if (unlikely(error)) {
1288                                 for (i = 0; i < free_cnt; i++)
1289                                         rte_pktmbuf_free(new_pkts[i]);
1290                         }
1291                         nb_enqueued += free_cnt;
1292                 } else {
1293                         struct rte_eth_dev *dev =
1294                                 &rte_eth_devices[rxvq->port_id];
1295                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1296                 }
1297         }
1298
1299         if (likely(nb_enqueued)) {
1300                 vq_update_avail_idx(vq);
1301
1302                 if (unlikely(virtqueue_kick_prepare(vq))) {
1303                         virtqueue_notify(vq);
1304                         PMD_RX_LOG(DEBUG, "Notified");
1305                 }
1306         }
1307
1308         return nb_rx;
1309 }
1310
1311 uint16_t
1312 virtio_recv_pkts_packed(void *rx_queue, struct rte_mbuf **rx_pkts,
1313                         uint16_t nb_pkts)
1314 {
1315         struct virtnet_rx *rxvq = rx_queue;
1316         struct virtqueue *vq = rxvq->vq;
1317         struct virtio_hw *hw = vq->hw;
1318         struct rte_mbuf *rxm;
1319         uint16_t num, nb_rx;
1320         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1321         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1322         int error;
1323         uint32_t i, nb_enqueued;
1324         uint32_t hdr_size;
1325         struct virtio_net_hdr *hdr;
1326
1327         nb_rx = 0;
1328         if (unlikely(hw->started == 0))
1329                 return nb_rx;
1330
1331         num = RTE_MIN(VIRTIO_MBUF_BURST_SZ, nb_pkts);
1332         if (likely(num > DESC_PER_CACHELINE))
1333                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1334
1335         num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1336         PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1337
1338         nb_enqueued = 0;
1339         hdr_size = hw->vtnet_hdr_size;
1340
1341         for (i = 0; i < num; i++) {
1342                 rxm = rcv_pkts[i];
1343
1344                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1345
1346                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1347                         PMD_RX_LOG(ERR, "Packet drop");
1348                         nb_enqueued++;
1349                         virtio_discard_rxbuf(vq, rxm);
1350                         rxvq->stats.errors++;
1351                         continue;
1352                 }
1353
1354                 rxm->port = rxvq->port_id;
1355                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1356                 rxm->ol_flags = 0;
1357                 rxm->vlan_tci = 0;
1358
1359                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1360                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1361
1362                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1363                         RTE_PKTMBUF_HEADROOM - hdr_size);
1364
1365                 if (hw->vlan_strip)
1366                         rte_vlan_strip(rxm);
1367
1368                 if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1369                         virtio_discard_rxbuf(vq, rxm);
1370                         rxvq->stats.errors++;
1371                         continue;
1372                 }
1373
1374                 virtio_rx_stats_updated(rxvq, rxm);
1375
1376                 rx_pkts[nb_rx++] = rxm;
1377         }
1378
1379         rxvq->stats.packets += nb_rx;
1380
1381         /* Allocate new mbuf for the used descriptor */
1382         if (likely(!virtqueue_full(vq))) {
1383                 uint16_t free_cnt = vq->vq_free_cnt;
1384                 struct rte_mbuf *new_pkts[free_cnt];
1385
1386                 if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1387                                                 free_cnt) == 0)) {
1388                         error = virtqueue_enqueue_recv_refill_packed(vq,
1389                                         new_pkts, free_cnt);
1390                         if (unlikely(error)) {
1391                                 for (i = 0; i < free_cnt; i++)
1392                                         rte_pktmbuf_free(new_pkts[i]);
1393                         }
1394                         nb_enqueued += free_cnt;
1395                 } else {
1396                         struct rte_eth_dev *dev =
1397                                 &rte_eth_devices[rxvq->port_id];
1398                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1399                 }
1400         }
1401
1402         if (likely(nb_enqueued)) {
1403                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1404                         virtqueue_notify(vq);
1405                         PMD_RX_LOG(DEBUG, "Notified");
1406                 }
1407         }
1408
1409         return nb_rx;
1410 }
1411
1412
1413 uint16_t
1414 virtio_recv_pkts_inorder(void *rx_queue,
1415                         struct rte_mbuf **rx_pkts,
1416                         uint16_t nb_pkts)
1417 {
1418         struct virtnet_rx *rxvq = rx_queue;
1419         struct virtqueue *vq = rxvq->vq;
1420         struct virtio_hw *hw = vq->hw;
1421         struct rte_mbuf *rxm;
1422         struct rte_mbuf *prev;
1423         uint16_t nb_used, num, nb_rx;
1424         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1425         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1426         int error;
1427         uint32_t nb_enqueued;
1428         uint32_t seg_num;
1429         uint32_t seg_res;
1430         uint32_t hdr_size;
1431         int32_t i;
1432
1433         nb_rx = 0;
1434         if (unlikely(hw->started == 0))
1435                 return nb_rx;
1436
1437         nb_used = VIRTQUEUE_NUSED(vq);
1438         nb_used = RTE_MIN(nb_used, nb_pkts);
1439         nb_used = RTE_MIN(nb_used, VIRTIO_MBUF_BURST_SZ);
1440
1441         virtio_rmb(hw->weak_barriers);
1442
1443         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1444
1445         nb_enqueued = 0;
1446         seg_num = 1;
1447         seg_res = 0;
1448         hdr_size = hw->vtnet_hdr_size;
1449
1450         num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len, nb_used);
1451
1452         for (i = 0; i < num; i++) {
1453                 struct virtio_net_hdr_mrg_rxbuf *header;
1454
1455                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1456                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1457
1458                 rxm = rcv_pkts[i];
1459
1460                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1461                         PMD_RX_LOG(ERR, "Packet drop");
1462                         nb_enqueued++;
1463                         virtio_discard_rxbuf_inorder(vq, rxm);
1464                         rxvq->stats.errors++;
1465                         continue;
1466                 }
1467
1468                 header = (struct virtio_net_hdr_mrg_rxbuf *)
1469                          ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1470                          - hdr_size);
1471
1472                 if (vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
1473                         seg_num = header->num_buffers;
1474                         if (seg_num == 0)
1475                                 seg_num = 1;
1476                 } else {
1477                         seg_num = 1;
1478                 }
1479
1480                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1481                 rxm->nb_segs = seg_num;
1482                 rxm->ol_flags = 0;
1483                 rxm->vlan_tci = 0;
1484                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1485                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1486
1487                 rxm->port = rxvq->port_id;
1488
1489                 rx_pkts[nb_rx] = rxm;
1490                 prev = rxm;
1491
1492                 if (vq->hw->has_rx_offload &&
1493                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1494                         virtio_discard_rxbuf_inorder(vq, rxm);
1495                         rxvq->stats.errors++;
1496                         continue;
1497                 }
1498
1499                 if (hw->vlan_strip)
1500                         rte_vlan_strip(rx_pkts[nb_rx]);
1501
1502                 seg_res = seg_num - 1;
1503
1504                 /* Merge remaining segments */
1505                 while (seg_res != 0 && i < (num - 1)) {
1506                         i++;
1507
1508                         rxm = rcv_pkts[i];
1509                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1510                         rxm->pkt_len = (uint32_t)(len[i]);
1511                         rxm->data_len = (uint16_t)(len[i]);
1512
1513                         rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1514                         rx_pkts[nb_rx]->data_len += (uint16_t)(len[i]);
1515
1516                         if (prev)
1517                                 prev->next = rxm;
1518
1519                         prev = rxm;
1520                         seg_res -= 1;
1521                 }
1522
1523                 if (!seg_res) {
1524                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1525                         nb_rx++;
1526                 }
1527         }
1528
1529         /* Last packet still need merge segments */
1530         while (seg_res != 0) {
1531                 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1532                                         VIRTIO_MBUF_BURST_SZ);
1533
1534                 prev = rcv_pkts[nb_rx];
1535                 if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
1536                         virtio_rmb(hw->weak_barriers);
1537                         num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len,
1538                                                            rcv_cnt);
1539                         uint16_t extra_idx = 0;
1540
1541                         rcv_cnt = num;
1542                         while (extra_idx < rcv_cnt) {
1543                                 rxm = rcv_pkts[extra_idx];
1544                                 rxm->data_off =
1545                                         RTE_PKTMBUF_HEADROOM - hdr_size;
1546                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1547                                 rxm->data_len = (uint16_t)(len[extra_idx]);
1548                                 prev->next = rxm;
1549                                 prev = rxm;
1550                                 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1551                                 rx_pkts[nb_rx]->data_len += len[extra_idx];
1552                                 extra_idx += 1;
1553                         };
1554                         seg_res -= rcv_cnt;
1555
1556                         if (!seg_res) {
1557                                 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1558                                 nb_rx++;
1559                         }
1560                 } else {
1561                         PMD_RX_LOG(ERR,
1562                                         "No enough segments for packet.");
1563                         virtio_discard_rxbuf_inorder(vq, prev);
1564                         rxvq->stats.errors++;
1565                         break;
1566                 }
1567         }
1568
1569         rxvq->stats.packets += nb_rx;
1570
1571         /* Allocate new mbuf for the used descriptor */
1572
1573         if (likely(!virtqueue_full(vq))) {
1574                 /* free_cnt may include mrg descs */
1575                 uint16_t free_cnt = vq->vq_free_cnt;
1576                 struct rte_mbuf *new_pkts[free_cnt];
1577
1578                 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1579                         error = virtqueue_enqueue_refill_inorder(vq, new_pkts,
1580                                         free_cnt);
1581                         if (unlikely(error)) {
1582                                 for (i = 0; i < free_cnt; i++)
1583                                         rte_pktmbuf_free(new_pkts[i]);
1584                         }
1585                         nb_enqueued += free_cnt;
1586                 } else {
1587                         struct rte_eth_dev *dev =
1588                                 &rte_eth_devices[rxvq->port_id];
1589                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1590                 }
1591         }
1592
1593         if (likely(nb_enqueued)) {
1594                 vq_update_avail_idx(vq);
1595
1596                 if (unlikely(virtqueue_kick_prepare(vq))) {
1597                         virtqueue_notify(vq);
1598                         PMD_RX_LOG(DEBUG, "Notified");
1599                 }
1600         }
1601
1602         return nb_rx;
1603 }
1604
1605 uint16_t
1606 virtio_recv_mergeable_pkts(void *rx_queue,
1607                         struct rte_mbuf **rx_pkts,
1608                         uint16_t nb_pkts)
1609 {
1610         struct virtnet_rx *rxvq = rx_queue;
1611         struct virtqueue *vq = rxvq->vq;
1612         struct virtio_hw *hw = vq->hw;
1613         struct rte_mbuf *rxm;
1614         struct rte_mbuf *prev;
1615         uint16_t nb_used, num, nb_rx = 0;
1616         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1617         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1618         int error;
1619         uint32_t nb_enqueued = 0;
1620         uint32_t seg_num = 0;
1621         uint32_t seg_res = 0;
1622         uint32_t hdr_size = hw->vtnet_hdr_size;
1623         int32_t i;
1624
1625         if (unlikely(hw->started == 0))
1626                 return nb_rx;
1627
1628         nb_used = VIRTQUEUE_NUSED(vq);
1629
1630         virtio_rmb(hw->weak_barriers);
1631
1632         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1633
1634         num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1635         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1636                 num = VIRTIO_MBUF_BURST_SZ;
1637         if (likely(num > DESC_PER_CACHELINE))
1638                 num = num - ((vq->vq_used_cons_idx + num) %
1639                                 DESC_PER_CACHELINE);
1640
1641
1642         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1643
1644         for (i = 0; i < num; i++) {
1645                 struct virtio_net_hdr_mrg_rxbuf *header;
1646
1647                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1648                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1649
1650                 rxm = rcv_pkts[i];
1651
1652                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1653                         PMD_RX_LOG(ERR, "Packet drop");
1654                         nb_enqueued++;
1655                         virtio_discard_rxbuf(vq, rxm);
1656                         rxvq->stats.errors++;
1657                         continue;
1658                 }
1659
1660                 header = (struct virtio_net_hdr_mrg_rxbuf *)
1661                          ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1662                          - hdr_size);
1663                 seg_num = header->num_buffers;
1664                 if (seg_num == 0)
1665                         seg_num = 1;
1666
1667                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1668                 rxm->nb_segs = seg_num;
1669                 rxm->ol_flags = 0;
1670                 rxm->vlan_tci = 0;
1671                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1672                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1673
1674                 rxm->port = rxvq->port_id;
1675
1676                 rx_pkts[nb_rx] = rxm;
1677                 prev = rxm;
1678
1679                 if (hw->has_rx_offload &&
1680                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1681                         virtio_discard_rxbuf(vq, rxm);
1682                         rxvq->stats.errors++;
1683                         continue;
1684                 }
1685
1686                 if (hw->vlan_strip)
1687                         rte_vlan_strip(rx_pkts[nb_rx]);
1688
1689                 seg_res = seg_num - 1;
1690
1691                 /* Merge remaining segments */
1692                 while (seg_res != 0 && i < (num - 1)) {
1693                         i++;
1694
1695                         rxm = rcv_pkts[i];
1696                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1697                         rxm->pkt_len = (uint32_t)(len[i]);
1698                         rxm->data_len = (uint16_t)(len[i]);
1699
1700                         rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1701                         rx_pkts[nb_rx]->data_len += (uint16_t)(len[i]);
1702
1703                         if (prev)
1704                                 prev->next = rxm;
1705
1706                         prev = rxm;
1707                         seg_res -= 1;
1708                 }
1709
1710                 if (!seg_res) {
1711                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1712                         nb_rx++;
1713                 }
1714         }
1715
1716         /* Last packet still need merge segments */
1717         while (seg_res != 0) {
1718                 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1719                                         VIRTIO_MBUF_BURST_SZ);
1720
1721                 prev = rcv_pkts[nb_rx];
1722                 if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
1723                         virtio_rmb(hw->weak_barriers);
1724                         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len,
1725                                                            rcv_cnt);
1726                         uint16_t extra_idx = 0;
1727
1728                         rcv_cnt = num;
1729                         while (extra_idx < rcv_cnt) {
1730                                 rxm = rcv_pkts[extra_idx];
1731                                 rxm->data_off =
1732                                         RTE_PKTMBUF_HEADROOM - hdr_size;
1733                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1734                                 rxm->data_len = (uint16_t)(len[extra_idx]);
1735                                 prev->next = rxm;
1736                                 prev = rxm;
1737                                 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1738                                 rx_pkts[nb_rx]->data_len += len[extra_idx];
1739                                 extra_idx += 1;
1740                         };
1741                         seg_res -= rcv_cnt;
1742
1743                         if (!seg_res) {
1744                                 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1745                                 nb_rx++;
1746                         }
1747                 } else {
1748                         PMD_RX_LOG(ERR,
1749                                         "No enough segments for packet.");
1750                         virtio_discard_rxbuf(vq, prev);
1751                         rxvq->stats.errors++;
1752                         break;
1753                 }
1754         }
1755
1756         rxvq->stats.packets += nb_rx;
1757
1758         /* Allocate new mbuf for the used descriptor */
1759         if (likely(!virtqueue_full(vq))) {
1760                 /* free_cnt may include mrg descs */
1761                 uint16_t free_cnt = vq->vq_free_cnt;
1762                 struct rte_mbuf *new_pkts[free_cnt];
1763
1764                 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1765                         error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1766                                         free_cnt);
1767                         if (unlikely(error)) {
1768                                 for (i = 0; i < free_cnt; i++)
1769                                         rte_pktmbuf_free(new_pkts[i]);
1770                         }
1771                         nb_enqueued += free_cnt;
1772                 } else {
1773                         struct rte_eth_dev *dev =
1774                                 &rte_eth_devices[rxvq->port_id];
1775                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1776                 }
1777         }
1778
1779         if (likely(nb_enqueued)) {
1780                 vq_update_avail_idx(vq);
1781
1782                 if (unlikely(virtqueue_kick_prepare(vq))) {
1783                         virtqueue_notify(vq);
1784                         PMD_RX_LOG(DEBUG, "Notified");
1785                 }
1786         }
1787
1788         return nb_rx;
1789 }
1790
1791 uint16_t
1792 virtio_recv_mergeable_pkts_packed(void *rx_queue,
1793                         struct rte_mbuf **rx_pkts,
1794                         uint16_t nb_pkts)
1795 {
1796         struct virtnet_rx *rxvq = rx_queue;
1797         struct virtqueue *vq = rxvq->vq;
1798         struct virtio_hw *hw = vq->hw;
1799         struct rte_mbuf *rxm;
1800         struct rte_mbuf *prev = NULL;
1801         uint16_t num, nb_rx = 0;
1802         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1803         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1804         uint32_t nb_enqueued = 0;
1805         uint32_t seg_num = 0;
1806         uint32_t seg_res = 0;
1807         uint32_t hdr_size = hw->vtnet_hdr_size;
1808         int32_t i;
1809         int error;
1810
1811         if (unlikely(hw->started == 0))
1812                 return nb_rx;
1813
1814
1815         num = nb_pkts;
1816         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1817                 num = VIRTIO_MBUF_BURST_SZ;
1818         if (likely(num > DESC_PER_CACHELINE))
1819                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1820
1821         num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1822
1823         for (i = 0; i < num; i++) {
1824                 struct virtio_net_hdr_mrg_rxbuf *header;
1825
1826                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1827                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1828
1829                 rxm = rcv_pkts[i];
1830
1831                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1832                         PMD_RX_LOG(ERR, "Packet drop");
1833                         nb_enqueued++;
1834                         virtio_discard_rxbuf(vq, rxm);
1835                         rxvq->stats.errors++;
1836                         continue;
1837                 }
1838
1839                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)
1840                           rxm->buf_addr + RTE_PKTMBUF_HEADROOM - hdr_size);
1841                 seg_num = header->num_buffers;
1842
1843                 if (seg_num == 0)
1844                         seg_num = 1;
1845
1846                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1847                 rxm->nb_segs = seg_num;
1848                 rxm->ol_flags = 0;
1849                 rxm->vlan_tci = 0;
1850                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1851                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1852
1853                 rxm->port = rxvq->port_id;
1854                 rx_pkts[nb_rx] = rxm;
1855                 prev = rxm;
1856
1857                 if (hw->has_rx_offload &&
1858                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1859                         virtio_discard_rxbuf(vq, rxm);
1860                         rxvq->stats.errors++;
1861                         continue;
1862                 }
1863
1864                 if (hw->vlan_strip)
1865                         rte_vlan_strip(rx_pkts[nb_rx]);
1866
1867                 seg_res = seg_num - 1;
1868
1869                 /* Merge remaining segments */
1870                 while (seg_res != 0 && i < (num - 1)) {
1871                         i++;
1872
1873                         rxm = rcv_pkts[i];
1874                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1875                         rxm->pkt_len = (uint32_t)(len[i]);
1876                         rxm->data_len = (uint16_t)(len[i]);
1877
1878                         rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1879                         rx_pkts[nb_rx]->data_len += (uint16_t)(len[i]);
1880
1881                         if (prev)
1882                                 prev->next = rxm;
1883
1884                         prev = rxm;
1885                         seg_res -= 1;
1886                 }
1887
1888                 if (!seg_res) {
1889                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1890                         nb_rx++;
1891                 }
1892         }
1893
1894         /* Last packet still need merge segments */
1895         while (seg_res != 0) {
1896                 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1897                                         VIRTIO_MBUF_BURST_SZ);
1898                 if (likely(vq->vq_free_cnt >= rcv_cnt)) {
1899                         num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts,
1900                                         len, rcv_cnt);
1901                         uint16_t extra_idx = 0;
1902
1903                         rcv_cnt = num;
1904
1905                         while (extra_idx < rcv_cnt) {
1906                                 rxm = rcv_pkts[extra_idx];
1907
1908                                 rxm->data_off =
1909                                         RTE_PKTMBUF_HEADROOM - hdr_size;
1910                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1911                                 rxm->data_len = (uint16_t)(len[extra_idx]);
1912
1913                                 prev->next = rxm;
1914                                 prev = rxm;
1915                                 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1916                                 rx_pkts[nb_rx]->data_len += len[extra_idx];
1917                                 extra_idx += 1;
1918                         }
1919                         seg_res -= rcv_cnt;
1920                         if (!seg_res) {
1921                                 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1922                                 nb_rx++;
1923                         }
1924                 } else {
1925                         PMD_RX_LOG(ERR,
1926                                         "No enough segments for packet.");
1927                         if (prev)
1928                                 virtio_discard_rxbuf(vq, prev);
1929                         rxvq->stats.errors++;
1930                         break;
1931                 }
1932         }
1933
1934         rxvq->stats.packets += nb_rx;
1935
1936         /* Allocate new mbuf for the used descriptor */
1937         if (likely(!virtqueue_full(vq))) {
1938                 /* free_cnt may include mrg descs */
1939                 uint16_t free_cnt = vq->vq_free_cnt;
1940                 struct rte_mbuf *new_pkts[free_cnt];
1941
1942                 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1943                         error = virtqueue_enqueue_recv_refill_packed(vq,
1944                                         new_pkts, free_cnt);
1945                         if (unlikely(error)) {
1946                                 for (i = 0; i < free_cnt; i++)
1947                                         rte_pktmbuf_free(new_pkts[i]);
1948                         }
1949                         nb_enqueued += free_cnt;
1950                 } else {
1951                         struct rte_eth_dev *dev =
1952                                 &rte_eth_devices[rxvq->port_id];
1953                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1954                 }
1955         }
1956
1957         if (likely(nb_enqueued)) {
1958                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1959                         virtqueue_notify(vq);
1960                         PMD_RX_LOG(DEBUG, "Notified");
1961                 }
1962         }
1963
1964         return nb_rx;
1965 }
1966
1967 uint16_t
1968 virtio_xmit_pkts_packed(void *tx_queue, struct rte_mbuf **tx_pkts,
1969                         uint16_t nb_pkts)
1970 {
1971         struct virtnet_tx *txvq = tx_queue;
1972         struct virtqueue *vq = txvq->vq;
1973         struct virtio_hw *hw = vq->hw;
1974         uint16_t hdr_size = hw->vtnet_hdr_size;
1975         uint16_t nb_tx = 0;
1976         bool in_order = hw->use_inorder_tx;
1977         int error;
1978
1979         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1980                 return nb_tx;
1981
1982         if (unlikely(nb_pkts < 1))
1983                 return nb_pkts;
1984
1985         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1986
1987         if (nb_pkts > vq->vq_free_cnt)
1988                 virtio_xmit_cleanup_packed(vq, nb_pkts - vq->vq_free_cnt,
1989                                            in_order);
1990
1991         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1992                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1993                 int can_push = 0, slots, need;
1994
1995                 /* Do VLAN tag insertion */
1996                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1997                         error = rte_vlan_insert(&txm);
1998                         if (unlikely(error)) {
1999                                 rte_pktmbuf_free(txm);
2000                                 continue;
2001                         }
2002                         /* vlan_insert may add a header mbuf */
2003                         tx_pkts[nb_tx] = txm;
2004                 }
2005
2006                 /* optimize ring usage */
2007                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
2008                       vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
2009                     rte_mbuf_refcnt_read(txm) == 1 &&
2010                     RTE_MBUF_DIRECT(txm) &&
2011                     txm->nb_segs == 1 &&
2012                     rte_pktmbuf_headroom(txm) >= hdr_size &&
2013                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
2014                            __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
2015                         can_push = 1;
2016
2017                 /* How many main ring entries are needed to this Tx?
2018                  * any_layout => number of segments
2019                  * default    => number of segments + 1
2020                  */
2021                 slots = txm->nb_segs + !can_push;
2022                 need = slots - vq->vq_free_cnt;
2023
2024                 /* Positive value indicates it need free vring descriptors */
2025                 if (unlikely(need > 0)) {
2026                         virtio_xmit_cleanup_packed(vq, need, in_order);
2027                         need = slots - vq->vq_free_cnt;
2028                         if (unlikely(need > 0)) {
2029                                 PMD_TX_LOG(ERR,
2030                                            "No free tx descriptors to transmit");
2031                                 break;
2032                         }
2033                 }
2034
2035                 /* Enqueue Packet buffers */
2036                 if (can_push)
2037                         virtqueue_enqueue_xmit_packed_fast(txvq, txm, in_order);
2038                 else
2039                         virtqueue_enqueue_xmit_packed(txvq, txm, slots, 0,
2040                                                       in_order);
2041
2042                 virtio_update_packet_stats(&txvq->stats, txm);
2043         }
2044
2045         txvq->stats.packets += nb_tx;
2046
2047         if (likely(nb_tx)) {
2048                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
2049                         virtqueue_notify(vq);
2050                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2051                 }
2052         }
2053
2054         return nb_tx;
2055 }
2056
2057 uint16_t
2058 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
2059 {
2060         struct virtnet_tx *txvq = tx_queue;
2061         struct virtqueue *vq = txvq->vq;
2062         struct virtio_hw *hw = vq->hw;
2063         uint16_t hdr_size = hw->vtnet_hdr_size;
2064         uint16_t nb_used, nb_tx = 0;
2065         int error;
2066
2067         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
2068                 return nb_tx;
2069
2070         if (unlikely(nb_pkts < 1))
2071                 return nb_pkts;
2072
2073         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
2074         nb_used = VIRTQUEUE_NUSED(vq);
2075
2076         virtio_rmb(hw->weak_barriers);
2077         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
2078                 virtio_xmit_cleanup(vq, nb_used);
2079
2080         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
2081                 struct rte_mbuf *txm = tx_pkts[nb_tx];
2082                 int can_push = 0, use_indirect = 0, slots, need;
2083
2084                 /* Do VLAN tag insertion */
2085                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
2086                         error = rte_vlan_insert(&txm);
2087                         if (unlikely(error)) {
2088                                 rte_pktmbuf_free(txm);
2089                                 continue;
2090                         }
2091                         /* vlan_insert may add a header mbuf */
2092                         tx_pkts[nb_tx] = txm;
2093                 }
2094
2095                 /* optimize ring usage */
2096                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
2097                       vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
2098                     rte_mbuf_refcnt_read(txm) == 1 &&
2099                     RTE_MBUF_DIRECT(txm) &&
2100                     txm->nb_segs == 1 &&
2101                     rte_pktmbuf_headroom(txm) >= hdr_size &&
2102                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
2103                                    __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
2104                         can_push = 1;
2105                 else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
2106                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
2107                         use_indirect = 1;
2108
2109                 /* How many main ring entries are needed to this Tx?
2110                  * any_layout => number of segments
2111                  * indirect   => 1
2112                  * default    => number of segments + 1
2113                  */
2114                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
2115                 need = slots - vq->vq_free_cnt;
2116
2117                 /* Positive value indicates it need free vring descriptors */
2118                 if (unlikely(need > 0)) {
2119                         nb_used = VIRTQUEUE_NUSED(vq);
2120                         virtio_rmb(hw->weak_barriers);
2121                         need = RTE_MIN(need, (int)nb_used);
2122
2123                         virtio_xmit_cleanup(vq, need);
2124                         need = slots - vq->vq_free_cnt;
2125                         if (unlikely(need > 0)) {
2126                                 PMD_TX_LOG(ERR,
2127                                            "No free tx descriptors to transmit");
2128                                 break;
2129                         }
2130                 }
2131
2132                 /* Enqueue Packet buffers */
2133                 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
2134                         can_push, 0);
2135
2136                 virtio_update_packet_stats(&txvq->stats, txm);
2137         }
2138
2139         txvq->stats.packets += nb_tx;
2140
2141         if (likely(nb_tx)) {
2142                 vq_update_avail_idx(vq);
2143
2144                 if (unlikely(virtqueue_kick_prepare(vq))) {
2145                         virtqueue_notify(vq);
2146                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2147                 }
2148         }
2149
2150         return nb_tx;
2151 }
2152
2153 uint16_t
2154 virtio_xmit_pkts_inorder(void *tx_queue,
2155                         struct rte_mbuf **tx_pkts,
2156                         uint16_t nb_pkts)
2157 {
2158         struct virtnet_tx *txvq = tx_queue;
2159         struct virtqueue *vq = txvq->vq;
2160         struct virtio_hw *hw = vq->hw;
2161         uint16_t hdr_size = hw->vtnet_hdr_size;
2162         uint16_t nb_used, nb_avail, nb_tx = 0, nb_inorder_pkts = 0;
2163         struct rte_mbuf *inorder_pkts[nb_pkts];
2164         int error;
2165
2166         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
2167                 return nb_tx;
2168
2169         if (unlikely(nb_pkts < 1))
2170                 return nb_pkts;
2171
2172         VIRTQUEUE_DUMP(vq);
2173         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
2174         nb_used = VIRTQUEUE_NUSED(vq);
2175
2176         virtio_rmb(hw->weak_barriers);
2177         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
2178                 virtio_xmit_cleanup_inorder(vq, nb_used);
2179
2180         if (unlikely(!vq->vq_free_cnt))
2181                 virtio_xmit_cleanup_inorder(vq, nb_used);
2182
2183         nb_avail = RTE_MIN(vq->vq_free_cnt, nb_pkts);
2184
2185         for (nb_tx = 0; nb_tx < nb_avail; nb_tx++) {
2186                 struct rte_mbuf *txm = tx_pkts[nb_tx];
2187                 int slots, need;
2188
2189                 /* Do VLAN tag insertion */
2190                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
2191                         error = rte_vlan_insert(&txm);
2192                         if (unlikely(error)) {
2193                                 rte_pktmbuf_free(txm);
2194                                 continue;
2195                         }
2196                         /* vlan_insert may add a header mbuf */
2197                         tx_pkts[nb_tx] = txm;
2198                 }
2199
2200                 /* optimize ring usage */
2201                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
2202                      vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
2203                      rte_mbuf_refcnt_read(txm) == 1 &&
2204                      RTE_MBUF_DIRECT(txm) &&
2205                      txm->nb_segs == 1 &&
2206                      rte_pktmbuf_headroom(txm) >= hdr_size &&
2207                      rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
2208                                 __alignof__(struct virtio_net_hdr_mrg_rxbuf))) {
2209                         inorder_pkts[nb_inorder_pkts] = txm;
2210                         nb_inorder_pkts++;
2211
2212                         virtio_update_packet_stats(&txvq->stats, txm);
2213                         continue;
2214                 }
2215
2216                 if (nb_inorder_pkts) {
2217                         virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2218                                                         nb_inorder_pkts);
2219                         nb_inorder_pkts = 0;
2220                 }
2221
2222                 slots = txm->nb_segs + 1;
2223                 need = slots - vq->vq_free_cnt;
2224                 if (unlikely(need > 0)) {
2225                         nb_used = VIRTQUEUE_NUSED(vq);
2226                         virtio_rmb(hw->weak_barriers);
2227                         need = RTE_MIN(need, (int)nb_used);
2228
2229                         virtio_xmit_cleanup_inorder(vq, need);
2230
2231                         need = slots - vq->vq_free_cnt;
2232
2233                         if (unlikely(need > 0)) {
2234                                 PMD_TX_LOG(ERR,
2235                                         "No free tx descriptors to transmit");
2236                                 break;
2237                         }
2238                 }
2239                 /* Enqueue Packet buffers */
2240                 virtqueue_enqueue_xmit(txvq, txm, slots, 0, 0, 1);
2241
2242                 virtio_update_packet_stats(&txvq->stats, txm);
2243         }
2244
2245         /* Transmit all inorder packets */
2246         if (nb_inorder_pkts)
2247                 virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2248                                                 nb_inorder_pkts);
2249
2250         txvq->stats.packets += nb_tx;
2251
2252         if (likely(nb_tx)) {
2253                 vq_update_avail_idx(vq);
2254
2255                 if (unlikely(virtqueue_kick_prepare(vq))) {
2256                         virtqueue_notify(vq);
2257                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2258                 }
2259         }
2260
2261         VIRTQUEUE_DUMP(vq);
2262
2263         return nb_tx;
2264 }