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
  35 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
  36 #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
  37 #else
  38 #define  VIRTIO_DUMP_PACKET(m, len) do { } while (0)
  39 #endif
  40
  41 int
  42 virtio_dev_rx_queue_done(void *rxq, uint16_t offset)
  43 {
  44         struct virtnet_rx *rxvq = rxq;
  45         struct virtqueue *vq = rxvq->vq;
  46
  47         return VIRTQUEUE_NUSED(vq) >= offset;
  48 }
  49
  50 void
  51 vq_ring_free_inorder(struct virtqueue *vq, uint16_t desc_idx, uint16_t num)
  52 {
  53         vq->vq_free_cnt += num;
  54         vq->vq_desc_tail_idx = desc_idx & (vq->vq_nentries - 1);
  55 }
  56
  57 void
  58 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
  59 {
  60         struct vring_desc *dp, *dp_tail;
  61         struct vq_desc_extra *dxp;
  62         uint16_t desc_idx_last = desc_idx;
  63
  64         dp  = &vq->vq_ring.desc[desc_idx];
  65         dxp = &vq->vq_descx[desc_idx];
  66         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
  67         if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
  68                 while (dp->flags & VRING_DESC_F_NEXT) {
  69                         desc_idx_last = dp->next;
  70                         dp = &vq->vq_ring.desc[dp->next];
  71                 }
  72         }
  73         dxp->ndescs = 0;
  74
  75         /*
  76          * We must append the existing free chain, if any, to the end of
  77          * newly freed chain. If the virtqueue was completely used, then
  78          * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
  79          */
  80         if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
  81                 vq->vq_desc_head_idx = desc_idx;
  82         } else {
  83                 dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
  84                 dp_tail->next = desc_idx;
  85         }
  86
  87         vq->vq_desc_tail_idx = desc_idx_last;
  88         dp->next = VQ_RING_DESC_CHAIN_END;
  89 }
  90
  91 static uint16_t
  92 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
  93                            uint32_t *len, uint16_t num)
  94 {
  95         struct vring_used_elem *uep;
  96         struct rte_mbuf *cookie;
  97         uint16_t used_idx, desc_idx;
  98         uint16_t i;
  99
 100         /*  Caller does the check */
 101         for (i = 0; i < num ; i++) {
 102                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 103                 uep = &vq->vq_ring.used->ring[used_idx];
 104                 desc_idx = (uint16_t) uep->id;
 105                 len[i] = uep->len;
 106                 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
 107
 108                 if (unlikely(cookie == NULL)) {
 109                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
 110                                 vq->vq_used_cons_idx);
 111                         break;
 112                 }
 113
 114                 rte_prefetch0(cookie);
 115                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 116                 rx_pkts[i]  = cookie;
 117                 vq->vq_used_cons_idx++;
 118                 vq_ring_free_chain(vq, desc_idx);
 119                 vq->vq_descx[desc_idx].cookie = NULL;
 120         }
 121
 122         return i;
 123 }
 124
 125 static uint16_t
 126 virtqueue_dequeue_rx_inorder(struct virtqueue *vq,
 127                         struct rte_mbuf **rx_pkts,
 128                         uint32_t *len,
 129                         uint16_t num)
 130 {
 131         struct vring_used_elem *uep;
 132         struct rte_mbuf *cookie;
 133         uint16_t used_idx = 0;
 134         uint16_t i;
 135
 136         if (unlikely(num == 0))
 137                 return 0;
 138
 139         for (i = 0; i < num; i++) {
 140                 used_idx = vq->vq_used_cons_idx & (vq->vq_nentries - 1);
 141                 /* Desc idx same as used idx */
 142                 uep = &vq->vq_ring.used->ring[used_idx];
 143                 len[i] = uep->len;
 144                 cookie = (struct rte_mbuf *)vq->vq_descx[used_idx].cookie;
 145
 146                 if (unlikely(cookie == NULL)) {
 147                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
 148                                 vq->vq_used_cons_idx);
 149                         break;
 150                 }
 151
 152                 rte_prefetch0(cookie);
 153                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 154                 rx_pkts[i]  = cookie;
 155                 vq->vq_used_cons_idx++;
 156                 vq->vq_descx[used_idx].cookie = NULL;
 157         }
 158
 159         vq_ring_free_inorder(vq, used_idx, i);
 160         return i;
 161 }
 162
 163 #ifndef DEFAULT_TX_FREE_THRESH
 164 #define DEFAULT_TX_FREE_THRESH 32
 165 #endif
 166
 167 /* Cleanup from completed transmits. */
 168 static void
 169 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
 170 {
 171         uint16_t i, used_idx, desc_idx;
 172         for (i = 0; i < num; i++) {
 173                 struct vring_used_elem *uep;
 174                 struct vq_desc_extra *dxp;
 175
 176                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 177                 uep = &vq->vq_ring.used->ring[used_idx];
 178
 179                 desc_idx = (uint16_t) uep->id;
 180                 dxp = &vq->vq_descx[desc_idx];
 181                 vq->vq_used_cons_idx++;
 182                 vq_ring_free_chain(vq, desc_idx);
 183
 184                 if (dxp->cookie != NULL) {
 185                         rte_pktmbuf_free(dxp->cookie);
 186                         dxp->cookie = NULL;
 187                 }
 188         }
 189 }
 190
 191 /* Cleanup from completed inorder transmits. */
 192 static void
 193 virtio_xmit_cleanup_inorder(struct virtqueue *vq, uint16_t num)
 194 {
 195         uint16_t i, used_idx, desc_idx = 0, last_idx;
 196         int16_t free_cnt = 0;
 197         struct vq_desc_extra *dxp = NULL;
 198
 199         if (unlikely(num == 0))
 200                 return;
 201
 202         for (i = 0; i < num; i++) {
 203                 struct vring_used_elem *uep;
 204
 205                 used_idx = vq->vq_used_cons_idx & (vq->vq_nentries - 1);
 206                 uep = &vq->vq_ring.used->ring[used_idx];
 207                 desc_idx = (uint16_t)uep->id;
 208
 209                 dxp = &vq->vq_descx[desc_idx];
 210                 vq->vq_used_cons_idx++;
 211
 212                 if (dxp->cookie != NULL) {
 213                         rte_pktmbuf_free(dxp->cookie);
 214                         dxp->cookie = NULL;
 215                 }
 216         }
 217
 218         last_idx = desc_idx + dxp->ndescs - 1;
 219         free_cnt = last_idx - vq->vq_desc_tail_idx;
 220         if (free_cnt <= 0)
 221                 free_cnt += vq->vq_nentries;
 222
 223         vq_ring_free_inorder(vq, last_idx, free_cnt);
 224 }
 225
 226 static inline int
 227 virtqueue_enqueue_refill_inorder(struct virtqueue *vq,
 228                         struct rte_mbuf **cookies,
 229                         uint16_t num)
 230 {
 231         struct vq_desc_extra *dxp;
 232         struct virtio_hw *hw = vq->hw;
 233         struct vring_desc *start_dp;
 234         uint16_t head_idx, idx, i = 0;
 235
 236         if (unlikely(vq->vq_free_cnt == 0))
 237                 return -ENOSPC;
 238         if (unlikely(vq->vq_free_cnt < num))
 239                 return -EMSGSIZE;
 240
 241         head_idx = vq->vq_desc_head_idx & (vq->vq_nentries - 1);
 242         start_dp = vq->vq_ring.desc;
 243
 244         while (i < num) {
 245                 idx = head_idx & (vq->vq_nentries - 1);
 246                 dxp = &vq->vq_descx[idx];
 247                 dxp->cookie = (void *)cookies[i];
 248                 dxp->ndescs = 1;
 249
 250                 start_dp[idx].addr =
 251                                 VIRTIO_MBUF_ADDR(cookies[i], vq) +
 252                                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 253                 start_dp[idx].len =
 254                                 cookies[i]->buf_len -
 255                                 RTE_PKTMBUF_HEADROOM +
 256                                 hw->vtnet_hdr_size;
 257                 start_dp[idx].flags =  VRING_DESC_F_WRITE;
 258
 259                 vq_update_avail_ring(vq, idx);
 260                 head_idx++;
 261                 i++;
 262         }
 263
 264         vq->vq_desc_head_idx += num;
 265         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 266         return 0;
 267 }
 268
 269 static inline int
 270 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
 271 {
 272         struct vq_desc_extra *dxp;
 273         struct virtio_hw *hw = vq->hw;
 274         struct vring_desc *start_dp;
 275         uint16_t needed = 1;
 276         uint16_t head_idx, idx;
 277
 278         if (unlikely(vq->vq_free_cnt == 0))
 279                 return -ENOSPC;
 280         if (unlikely(vq->vq_free_cnt < needed))
 281                 return -EMSGSIZE;
 282
 283         head_idx = vq->vq_desc_head_idx;
 284         if (unlikely(head_idx >= vq->vq_nentries))
 285                 return -EFAULT;
 286
 287         idx = head_idx;
 288         dxp = &vq->vq_descx[idx];
 289         dxp->cookie = (void *)cookie;
 290         dxp->ndescs = needed;
 291
 292         start_dp = vq->vq_ring.desc;
 293         start_dp[idx].addr =
 294                 VIRTIO_MBUF_ADDR(cookie, vq) +
 295                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 296         start_dp[idx].len =
 297                 cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
 298         start_dp[idx].flags =  VRING_DESC_F_WRITE;
 299         idx = start_dp[idx].next;
 300         vq->vq_desc_head_idx = idx;
 301         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 302                 vq->vq_desc_tail_idx = idx;
 303         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 304         vq_update_avail_ring(vq, head_idx);
 305
 306         return 0;
 307 }
 308
 309 /* When doing TSO, the IP length is not included in the pseudo header
 310  * checksum of the packet given to the PMD, but for virtio it is
 311  * expected.
 312  */
 313 static void
 314 virtio_tso_fix_cksum(struct rte_mbuf *m)
 315 {
 316         /* common case: header is not fragmented */
 317         if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
 318                         m->l4_len)) {
 319                 struct ipv4_hdr *iph;
 320                 struct ipv6_hdr *ip6h;
 321                 struct tcp_hdr *th;
 322                 uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
 323                 uint32_t tmp;
 324
 325                 iph = rte_pktmbuf_mtod_offset(m, struct ipv4_hdr *, m->l2_len);
 326                 th = RTE_PTR_ADD(iph, m->l3_len);
 327                 if ((iph->version_ihl >> 4) == 4) {
 328                         iph->hdr_checksum = 0;
 329                         iph->hdr_checksum = rte_ipv4_cksum(iph);
 330                         ip_len = iph->total_length;
 331                         ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
 332                                 m->l3_len);
 333                 } else {
 334                         ip6h = (struct ipv6_hdr *)iph;
 335                         ip_paylen = ip6h->payload_len;
 336                 }
 337
 338                 /* calculate the new phdr checksum not including ip_paylen */
 339                 prev_cksum = th->cksum;
 340                 tmp = prev_cksum;
 341                 tmp += ip_paylen;
 342                 tmp = (tmp & 0xffff) + (tmp >> 16);
 343                 new_cksum = tmp;
 344
 345                 /* replace it in the packet */
 346                 th->cksum = new_cksum;
 347         }
 348 }
 349
 350
 351 /* avoid write operation when necessary, to lessen cache issues */
 352 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 353         if ((var) != (val))                     \
 354                 (var) = (val);                  \
 355 } while (0)
 356
 357 static inline void
 358 virtqueue_xmit_offload(struct virtio_net_hdr *hdr,
 359                         struct rte_mbuf *cookie,
 360                         bool offload)
 361 {
 362         if (offload) {
 363                 if (cookie->ol_flags & PKT_TX_TCP_SEG)
 364                         cookie->ol_flags |= PKT_TX_TCP_CKSUM;
 365
 366                 switch (cookie->ol_flags & PKT_TX_L4_MASK) {
 367                 case PKT_TX_UDP_CKSUM:
 368                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 369                         hdr->csum_offset = offsetof(struct udp_hdr,
 370                                 dgram_cksum);
 371                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 372                         break;
 373
 374                 case PKT_TX_TCP_CKSUM:
 375                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 376                         hdr->csum_offset = offsetof(struct tcp_hdr, cksum);
 377                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 378                         break;
 379
 380                 default:
 381                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 382                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 383                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 384                         break;
 385                 }
 386
 387                 /* TCP Segmentation Offload */
 388                 if (cookie->ol_flags & PKT_TX_TCP_SEG) {
 389                         virtio_tso_fix_cksum(cookie);
 390                         hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
 391                                 VIRTIO_NET_HDR_GSO_TCPV6 :
 392                                 VIRTIO_NET_HDR_GSO_TCPV4;
 393                         hdr->gso_size = cookie->tso_segsz;
 394                         hdr->hdr_len =
 395                                 cookie->l2_len +
 396                                 cookie->l3_len +
 397                                 cookie->l4_len;
 398                 } else {
 399                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 400                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 401                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 402                 }
 403         }
 404 }
 405
 406 static inline void
 407 virtqueue_enqueue_xmit_inorder(struct virtnet_tx *txvq,
 408                         struct rte_mbuf **cookies,
 409                         uint16_t num)
 410 {
 411         struct vq_desc_extra *dxp;
 412         struct virtqueue *vq = txvq->vq;
 413         struct vring_desc *start_dp;
 414         struct virtio_net_hdr *hdr;
 415         uint16_t idx;
 416         uint16_t head_size = vq->hw->vtnet_hdr_size;
 417         uint16_t i = 0;
 418
 419         idx = vq->vq_desc_head_idx;
 420         start_dp = vq->vq_ring.desc;
 421
 422         while (i < num) {
 423                 idx = idx & (vq->vq_nentries - 1);
 424                 dxp = &vq->vq_descx[idx];
 425                 dxp->cookie = (void *)cookies[i];
 426                 dxp->ndescs = 1;
 427
 428                 hdr = (struct virtio_net_hdr *)
 429                         rte_pktmbuf_prepend(cookies[i], head_size);
 430                 cookies[i]->pkt_len -= head_size;
 431
 432                 /* if offload disabled, it is not zeroed below, do it now */
 433                 if (!vq->hw->has_tx_offload) {
 434                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 435                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 436                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 437                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 438                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 439                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 440                 }
 441
 442                 virtqueue_xmit_offload(hdr, cookies[i],
 443                                 vq->hw->has_tx_offload);
 444
 445                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookies[i], vq);
 446                 start_dp[idx].len   = cookies[i]->data_len;
 447                 start_dp[idx].flags = 0;
 448
 449                 vq_update_avail_ring(vq, idx);
 450
 451                 idx++;
 452                 i++;
 453         };
 454
 455         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 456         vq->vq_desc_head_idx = idx & (vq->vq_nentries - 1);
 457 }
 458
 459 static inline void
 460 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
 461                         uint16_t needed, int use_indirect, int can_push,
 462                         int in_order)
 463 {
 464         struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 465         struct vq_desc_extra *dxp;
 466         struct virtqueue *vq = txvq->vq;
 467         struct vring_desc *start_dp;
 468         uint16_t seg_num = cookie->nb_segs;
 469         uint16_t head_idx, idx;
 470         uint16_t head_size = vq->hw->vtnet_hdr_size;
 471         struct virtio_net_hdr *hdr;
 472
 473         head_idx = vq->vq_desc_head_idx;
 474         idx = head_idx;
 475         dxp = &vq->vq_descx[idx];
 476         dxp->cookie = (void *)cookie;
 477         dxp->ndescs = needed;
 478
 479         start_dp = vq->vq_ring.desc;
 480
 481         if (can_push) {
 482                 /* prepend cannot fail, checked by caller */
 483                 hdr = (struct virtio_net_hdr *)
 484                         rte_pktmbuf_prepend(cookie, head_size);
 485                 /* rte_pktmbuf_prepend() counts the hdr size to the pkt length,
 486                  * which is wrong. Below subtract restores correct pkt size.
 487                  */
 488                 cookie->pkt_len -= head_size;
 489
 490                 /* if offload disabled, it is not zeroed below, do it now */
 491                 if (!vq->hw->has_tx_offload) {
 492                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 493                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 494                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 495                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 496                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 497                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 498                 }
 499         } else if (use_indirect) {
 500                 /* setup tx ring slot to point to indirect
 501                  * descriptor list stored in reserved region.
 502                  *
 503                  * the first slot in indirect ring is already preset
 504                  * to point to the header in reserved region
 505                  */
 506                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 507                         RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
 508                 start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
 509                 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
 510                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 511
 512                 /* loop below will fill in rest of the indirect elements */
 513                 start_dp = txr[idx].tx_indir;
 514                 idx = 1;
 515         } else {
 516                 /* setup first tx ring slot to point to header
 517                  * stored in reserved region.
 518                  */
 519                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 520                         RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
 521                 start_dp[idx].len   = vq->hw->vtnet_hdr_size;
 522                 start_dp[idx].flags = VRING_DESC_F_NEXT;
 523                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 524
 525                 idx = start_dp[idx].next;
 526         }
 527
 528         virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload);
 529
 530         do {
 531                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
 532                 start_dp[idx].len   = cookie->data_len;
 533                 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 534                 idx = start_dp[idx].next;
 535         } while ((cookie = cookie->next) != NULL);
 536
 537         if (use_indirect)
 538                 idx = vq->vq_ring.desc[head_idx].next;
 539
 540         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 541
 542         vq->vq_desc_head_idx = idx;
 543         vq_update_avail_ring(vq, head_idx);
 544
 545         if (!in_order) {
 546                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 547                         vq->vq_desc_tail_idx = idx;
 548         }
 549 }
 550
 551 void
 552 virtio_dev_cq_start(struct rte_eth_dev *dev)
 553 {
 554         struct virtio_hw *hw = dev->data->dev_private;
 555
 556         if (hw->cvq && hw->cvq->vq) {
 557                 rte_spinlock_init(&hw->cvq->lock);
 558                 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
 559         }
 560 }
 561
 562 int
 563 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
 564                         uint16_t queue_idx,
 565                         uint16_t nb_desc,
 566                         unsigned int socket_id __rte_unused,
 567                         const struct rte_eth_rxconf *rx_conf __rte_unused,
 568                         struct rte_mempool *mp)
 569 {
 570         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 571         struct virtio_hw *hw = dev->data->dev_private;
 572         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 573         struct virtnet_rx *rxvq;
 574
 575         PMD_INIT_FUNC_TRACE();
 576
 577         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 578                 nb_desc = vq->vq_nentries;
 579         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 580
 581         rxvq = &vq->rxq;
 582         rxvq->queue_id = queue_idx;
 583         rxvq->mpool = mp;
 584         if (rxvq->mpool == NULL) {
 585                 rte_exit(EXIT_FAILURE,
 586                         "Cannot allocate mbufs for rx virtqueue");
 587         }
 588
 589         dev->data->rx_queues[queue_idx] = rxvq;
 590
 591         return 0;
 592 }
 593
 594 int
 595 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
 596 {
 597         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 598         struct virtio_hw *hw = dev->data->dev_private;
 599         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 600         struct virtnet_rx *rxvq = &vq->rxq;
 601         struct rte_mbuf *m;
 602         uint16_t desc_idx;
 603         int error, nbufs, i;
 604
 605         PMD_INIT_FUNC_TRACE();
 606
 607         /* Allocate blank mbufs for the each rx descriptor */
 608         nbufs = 0;
 609
 610         if (hw->use_simple_rx) {
 611                 for (desc_idx = 0; desc_idx < vq->vq_nentries;
 612                      desc_idx++) {
 613                         vq->vq_ring.avail->ring[desc_idx] = desc_idx;
 614                         vq->vq_ring.desc[desc_idx].flags =
 615                                 VRING_DESC_F_WRITE;
 616                 }
 617
 618                 virtio_rxq_vec_setup(rxvq);
 619         }
 620
 621         memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
 622         for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
 623              desc_idx++) {
 624                 vq->sw_ring[vq->vq_nentries + desc_idx] =
 625                         &rxvq->fake_mbuf;
 626         }
 627
 628         if (hw->use_simple_rx) {
 629                 while (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
 630                         virtio_rxq_rearm_vec(rxvq);
 631                         nbufs += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
 632                 }
 633         } else if (hw->use_inorder_rx) {
 634                 if ((!virtqueue_full(vq))) {
 635                         uint16_t free_cnt = vq->vq_free_cnt;
 636                         struct rte_mbuf *pkts[free_cnt];
 637
 638                         if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, pkts,
 639                                 free_cnt)) {
 640                                 error = virtqueue_enqueue_refill_inorder(vq,
 641                                                 pkts,
 642                                                 free_cnt);
 643                                 if (unlikely(error)) {
 644                                         for (i = 0; i < free_cnt; i++)
 645                                                 rte_pktmbuf_free(pkts[i]);
 646                                 }
 647                         }
 648
 649                         nbufs += free_cnt;
 650                         vq_update_avail_idx(vq);
 651                 }
 652         } else {
 653                 while (!virtqueue_full(vq)) {
 654                         m = rte_mbuf_raw_alloc(rxvq->mpool);
 655                         if (m == NULL)
 656                                 break;
 657
 658                         /* Enqueue allocated buffers */
 659                         error = virtqueue_enqueue_recv_refill(vq, m);
 660                         if (error) {
 661                                 rte_pktmbuf_free(m);
 662                                 break;
 663                         }
 664                         nbufs++;
 665                 }
 666
 667                 vq_update_avail_idx(vq);
 668         }
 669
 670         PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
 671
 672         VIRTQUEUE_DUMP(vq);
 673
 674         return 0;
 675 }
 676
 677 /*
 678  * struct rte_eth_dev *dev: Used to update dev
 679  * uint16_t nb_desc: Defaults to values read from config space
 680  * unsigned int socket_id: Used to allocate memzone
 681  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
 682  * uint16_t queue_idx: Just used as an index in dev txq list
 683  */
 684 int
 685 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
 686                         uint16_t queue_idx,
 687                         uint16_t nb_desc,
 688                         unsigned int socket_id __rte_unused,
 689                         const struct rte_eth_txconf *tx_conf)
 690 {
 691         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 692         struct virtio_hw *hw = dev->data->dev_private;
 693         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 694         struct virtnet_tx *txvq;
 695         uint16_t tx_free_thresh;
 696
 697         PMD_INIT_FUNC_TRACE();
 698
 699         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 700                 nb_desc = vq->vq_nentries;
 701         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 702
 703         txvq = &vq->txq;
 704         txvq->queue_id = queue_idx;
 705
 706         tx_free_thresh = tx_conf->tx_free_thresh;
 707         if (tx_free_thresh == 0)
 708                 tx_free_thresh =
 709                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
 710
 711         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
 712                 RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
 713                         "number of TX entries minus 3 (%u)."
 714                         " (tx_free_thresh=%u port=%u queue=%u)\n",
 715                         vq->vq_nentries - 3,
 716                         tx_free_thresh, dev->data->port_id, queue_idx);
 717                 return -EINVAL;
 718         }
 719
 720         vq->vq_free_thresh = tx_free_thresh;
 721
 722         dev->data->tx_queues[queue_idx] = txvq;
 723         return 0;
 724 }
 725
 726 int
 727 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
 728                                 uint16_t queue_idx)
 729 {
 730         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 731         struct virtio_hw *hw = dev->data->dev_private;
 732         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 733
 734         PMD_INIT_FUNC_TRACE();
 735
 736         if (hw->use_inorder_tx)
 737                 vq->vq_ring.desc[vq->vq_nentries - 1].next = 0;
 738
 739         VIRTQUEUE_DUMP(vq);
 740
 741         return 0;
 742 }
 743
 744 static void
 745 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
 746 {
 747         int error;
 748         /*
 749          * Requeue the discarded mbuf. This should always be
 750          * successful since it was just dequeued.
 751          */
 752         error = virtqueue_enqueue_recv_refill(vq, m);
 753
 754         if (unlikely(error)) {
 755                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
 756                 rte_pktmbuf_free(m);
 757         }
 758 }
 759
 760 static void
 761 virtio_discard_rxbuf_inorder(struct virtqueue *vq, struct rte_mbuf *m)
 762 {
 763         int error;
 764
 765         error = virtqueue_enqueue_refill_inorder(vq, &m, 1);
 766         if (unlikely(error)) {
 767                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
 768                 rte_pktmbuf_free(m);
 769         }
 770 }
 771
 772 static void
 773 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
 774 {
 775         uint32_t s = mbuf->pkt_len;
 776         struct ether_addr *ea;
 777
 778         stats->bytes += s;
 779
 780         if (s == 64) {
 781                 stats->size_bins[1]++;
 782         } else if (s > 64 && s < 1024) {
 783                 uint32_t bin;
 784
 785                 /* count zeros, and offset into correct bin */
 786                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
 787                 stats->size_bins[bin]++;
 788         } else {
 789                 if (s < 64)
 790                         stats->size_bins[0]++;
 791                 else if (s < 1519)
 792                         stats->size_bins[6]++;
 793                 else if (s >= 1519)
 794                         stats->size_bins[7]++;
 795         }
 796
 797         ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
 798         if (is_multicast_ether_addr(ea)) {
 799                 if (is_broadcast_ether_addr(ea))
 800                         stats->broadcast++;
 801                 else
 802                         stats->multicast++;
 803         }
 804 }
 805
 806 static inline void
 807 virtio_rx_stats_updated(struct virtnet_rx *rxvq, struct rte_mbuf *m)
 808 {
 809         VIRTIO_DUMP_PACKET(m, m->data_len);
 810
 811         virtio_update_packet_stats(&rxvq->stats, m);
 812 }
 813
 814 /* Optionally fill offload information in structure */
 815 static int
 816 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
 817 {
 818         struct rte_net_hdr_lens hdr_lens;
 819         uint32_t hdrlen, ptype;
 820         int l4_supported = 0;
 821
 822         /* nothing to do */
 823         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 824                 return 0;
 825
 826         m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
 827
 828         ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
 829         m->packet_type = ptype;
 830         if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
 831             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
 832             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
 833                 l4_supported = 1;
 834
 835         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 836                 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
 837                 if (hdr->csum_start <= hdrlen && l4_supported) {
 838                         m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
 839                 } else {
 840                         /* Unknown proto or tunnel, do sw cksum. We can assume
 841                          * the cksum field is in the first segment since the
 842                          * buffers we provided to the host are large enough.
 843                          * In case of SCTP, this will be wrong since it's a CRC
 844                          * but there's nothing we can do.
 845                          */
 846                         uint16_t csum = 0, off;
 847
 848                         rte_raw_cksum_mbuf(m, hdr->csum_start,
 849                                 rte_pktmbuf_pkt_len(m) - hdr->csum_start,
 850                                 &csum);
 851                         if (likely(csum != 0xffff))
 852                                 csum = ~csum;
 853                         off = hdr->csum_offset + hdr->csum_start;
 854                         if (rte_pktmbuf_data_len(m) >= off + 1)
 855                                 *rte_pktmbuf_mtod_offset(m, uint16_t *,
 856                                         off) = csum;
 857                 }
 858         } else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
 859                 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
 860         }
 861
 862         /* GSO request, save required information in mbuf */
 863         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 864                 /* Check unsupported modes */
 865                 if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
 866                     (hdr->gso_size == 0)) {
 867                         return -EINVAL;
 868                 }
 869
 870                 /* Update mss lengthes in mbuf */
 871                 m->tso_segsz = hdr->gso_size;
 872                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 873                         case VIRTIO_NET_HDR_GSO_TCPV4:
 874                         case VIRTIO_NET_HDR_GSO_TCPV6:
 875                                 m->ol_flags |= PKT_RX_LRO | \
 876                                         PKT_RX_L4_CKSUM_NONE;
 877                                 break;
 878                         default:
 879                                 return -EINVAL;
 880                 }
 881         }
 882
 883         return 0;
 884 }
 885
 886 #define VIRTIO_MBUF_BURST_SZ 64
 887 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
 888 uint16_t
 889 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 890 {
 891         struct virtnet_rx *rxvq = rx_queue;
 892         struct virtqueue *vq = rxvq->vq;
 893         struct virtio_hw *hw = vq->hw;
 894         struct rte_mbuf *rxm, *new_mbuf;
 895         uint16_t nb_used, num, nb_rx;
 896         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 897         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 898         int error;
 899         uint32_t i, nb_enqueued;
 900         uint32_t hdr_size;
 901         struct virtio_net_hdr *hdr;
 902
 903         nb_rx = 0;
 904         if (unlikely(hw->started == 0))
 905                 return nb_rx;
 906
 907         nb_used = VIRTQUEUE_NUSED(vq);
 908
 909         virtio_rmb();
 910
 911         num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
 912         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
 913                 num = VIRTIO_MBUF_BURST_SZ;
 914         if (likely(num > DESC_PER_CACHELINE))
 915                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
 916
 917         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
 918         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
 919
 920         nb_enqueued = 0;
 921         hdr_size = hw->vtnet_hdr_size;
 922
 923         for (i = 0; i < num ; i++) {
 924                 rxm = rcv_pkts[i];
 925
 926                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
 927
 928                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
 929                         PMD_RX_LOG(ERR, "Packet drop");
 930                         nb_enqueued++;
 931                         virtio_discard_rxbuf(vq, rxm);
 932                         rxvq->stats.errors++;
 933                         continue;
 934                 }
 935
 936                 rxm->port = rxvq->port_id;
 937                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 938                 rxm->ol_flags = 0;
 939                 rxm->vlan_tci = 0;
 940
 941                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
 942                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
 943
 944                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
 945                         RTE_PKTMBUF_HEADROOM - hdr_size);
 946
 947                 if (hw->vlan_strip)
 948                         rte_vlan_strip(rxm);
 949
 950                 if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
 951                         virtio_discard_rxbuf(vq, rxm);
 952                         rxvq->stats.errors++;
 953                         continue;
 954                 }
 955
 956                 virtio_rx_stats_updated(rxvq, rxm);
 957
 958                 rx_pkts[nb_rx++] = rxm;
 959         }
 960
 961         rxvq->stats.packets += nb_rx;
 962
 963         /* Allocate new mbuf for the used descriptor */
 964         while (likely(!virtqueue_full(vq))) {
 965                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
 966                 if (unlikely(new_mbuf == NULL)) {
 967                         struct rte_eth_dev *dev
 968                                 = &rte_eth_devices[rxvq->port_id];
 969                         dev->data->rx_mbuf_alloc_failed++;
 970                         break;
 971                 }
 972                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
 973                 if (unlikely(error)) {
 974                         rte_pktmbuf_free(new_mbuf);
 975                         break;
 976                 }
 977                 nb_enqueued++;
 978         }
 979
 980         if (likely(nb_enqueued)) {
 981                 vq_update_avail_idx(vq);
 982
 983                 if (unlikely(virtqueue_kick_prepare(vq))) {
 984                         virtqueue_notify(vq);
 985                         PMD_RX_LOG(DEBUG, "Notified");
 986                 }
 987         }
 988
 989         return nb_rx;
 990 }
 991
 992 uint16_t
 993 virtio_recv_mergeable_pkts_inorder(void *rx_queue,
 994                         struct rte_mbuf **rx_pkts,
 995                         uint16_t nb_pkts)
 996 {
 997         struct virtnet_rx *rxvq = rx_queue;
 998         struct virtqueue *vq = rxvq->vq;
 999         struct virtio_hw *hw = vq->hw;
1000         struct rte_mbuf *rxm;
1001         struct rte_mbuf *prev;
1002         uint16_t nb_used, num, nb_rx;
1003         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1004         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1005         int error;
1006         uint32_t nb_enqueued;
1007         uint32_t seg_num;
1008         uint32_t seg_res;
1009         uint32_t hdr_size;
1010         int32_t i;
1011
1012         nb_rx = 0;
1013         if (unlikely(hw->started == 0))
1014                 return nb_rx;
1015
1016         nb_used = VIRTQUEUE_NUSED(vq);
1017         nb_used = RTE_MIN(nb_used, nb_pkts);
1018         nb_used = RTE_MIN(nb_used, VIRTIO_MBUF_BURST_SZ);
1019
1020         virtio_rmb();
1021
1022         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1023
1024         nb_enqueued = 0;
1025         seg_num = 1;
1026         seg_res = 0;
1027         hdr_size = hw->vtnet_hdr_size;
1028
1029         num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len, nb_used);
1030
1031         for (i = 0; i < num; i++) {
1032                 struct virtio_net_hdr_mrg_rxbuf *header;
1033
1034                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1035                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1036
1037                 rxm = rcv_pkts[i];
1038
1039                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
1040                         PMD_RX_LOG(ERR, "Packet drop");
1041                         nb_enqueued++;
1042                         virtio_discard_rxbuf_inorder(vq, rxm);
1043                         rxvq->stats.errors++;
1044                         continue;
1045                 }
1046
1047                 header = (struct virtio_net_hdr_mrg_rxbuf *)
1048                          ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1049                          - hdr_size);
1050                 seg_num = header->num_buffers;
1051
1052                 if (seg_num == 0)
1053                         seg_num = 1;
1054
1055                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1056                 rxm->nb_segs = seg_num;
1057                 rxm->ol_flags = 0;
1058                 rxm->vlan_tci = 0;
1059                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1060                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1061
1062                 rxm->port = rxvq->port_id;
1063
1064                 rx_pkts[nb_rx] = rxm;
1065                 prev = rxm;
1066
1067                 if (vq->hw->has_rx_offload &&
1068                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1069                         virtio_discard_rxbuf_inorder(vq, rxm);
1070                         rxvq->stats.errors++;
1071                         continue;
1072                 }
1073
1074                 if (hw->vlan_strip)
1075                         rte_vlan_strip(rx_pkts[nb_rx]);
1076
1077                 seg_res = seg_num - 1;
1078
1079                 /* Merge remaining segments */
1080                 while (seg_res != 0 && i < (num - 1)) {
1081                         i++;
1082
1083                         rxm = rcv_pkts[i];
1084                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1085                         rxm->pkt_len = (uint32_t)(len[i]);
1086                         rxm->data_len = (uint16_t)(len[i]);
1087
1088                         rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1089                         rx_pkts[nb_rx]->data_len += (uint16_t)(len[i]);
1090
1091                         if (prev)
1092                                 prev->next = rxm;
1093
1094                         prev = rxm;
1095                         seg_res -= 1;
1096                 }
1097
1098                 if (!seg_res) {
1099                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1100                         nb_rx++;
1101                 }
1102         }
1103
1104         /* Last packet still need merge segments */
1105         while (seg_res != 0) {
1106                 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1107                                         VIRTIO_MBUF_BURST_SZ);
1108
1109                 prev = rcv_pkts[nb_rx];
1110                 if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
1111                         num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len,
1112                                                            rcv_cnt);
1113                         uint16_t extra_idx = 0;
1114
1115                         rcv_cnt = num;
1116                         while (extra_idx < rcv_cnt) {
1117                                 rxm = rcv_pkts[extra_idx];
1118                                 rxm->data_off =
1119                                         RTE_PKTMBUF_HEADROOM - hdr_size;
1120                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1121                                 rxm->data_len = (uint16_t)(len[extra_idx]);
1122                                 prev->next = rxm;
1123                                 prev = rxm;
1124                                 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1125                                 rx_pkts[nb_rx]->data_len += len[extra_idx];
1126                                 extra_idx += 1;
1127                         };
1128                         seg_res -= rcv_cnt;
1129
1130                         if (!seg_res) {
1131                                 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1132                                 nb_rx++;
1133                         }
1134                 } else {
1135                         PMD_RX_LOG(ERR,
1136                                         "No enough segments for packet.");
1137                         virtio_discard_rxbuf_inorder(vq, prev);
1138                         rxvq->stats.errors++;
1139                         break;
1140                 }
1141         }
1142
1143         rxvq->stats.packets += nb_rx;
1144
1145         /* Allocate new mbuf for the used descriptor */
1146
1147         if (likely(!virtqueue_full(vq))) {
1148                 /* free_cnt may include mrg descs */
1149                 uint16_t free_cnt = vq->vq_free_cnt;
1150                 struct rte_mbuf *new_pkts[free_cnt];
1151
1152                 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1153                         error = virtqueue_enqueue_refill_inorder(vq, new_pkts,
1154                                         free_cnt);
1155                         if (unlikely(error)) {
1156                                 for (i = 0; i < free_cnt; i++)
1157                                         rte_pktmbuf_free(new_pkts[i]);
1158                         }
1159                         nb_enqueued += free_cnt;
1160                 } else {
1161                         struct rte_eth_dev *dev =
1162                                 &rte_eth_devices[rxvq->port_id];
1163                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1164                 }
1165         }
1166
1167         if (likely(nb_enqueued)) {
1168                 vq_update_avail_idx(vq);
1169
1170                 if (unlikely(virtqueue_kick_prepare(vq))) {
1171                         virtqueue_notify(vq);
1172                         PMD_RX_LOG(DEBUG, "Notified");
1173                 }
1174         }
1175
1176         return nb_rx;
1177 }
1178
1179 uint16_t
1180 virtio_recv_mergeable_pkts(void *rx_queue,
1181                         struct rte_mbuf **rx_pkts,
1182                         uint16_t nb_pkts)
1183 {
1184         struct virtnet_rx *rxvq = rx_queue;
1185         struct virtqueue *vq = rxvq->vq;
1186         struct virtio_hw *hw = vq->hw;
1187         struct rte_mbuf *rxm, *new_mbuf;
1188         uint16_t nb_used, num, nb_rx;
1189         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1190         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1191         struct rte_mbuf *prev;
1192         int error;
1193         uint32_t i, nb_enqueued;
1194         uint32_t seg_num;
1195         uint16_t extra_idx;
1196         uint32_t seg_res;
1197         uint32_t hdr_size;
1198
1199         nb_rx = 0;
1200         if (unlikely(hw->started == 0))
1201                 return nb_rx;
1202
1203         nb_used = VIRTQUEUE_NUSED(vq);
1204
1205         virtio_rmb();
1206
1207         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1208
1209         i = 0;
1210         nb_enqueued = 0;
1211         seg_num = 0;
1212         extra_idx = 0;
1213         seg_res = 0;
1214         hdr_size = hw->vtnet_hdr_size;
1215
1216         while (i < nb_used) {
1217                 struct virtio_net_hdr_mrg_rxbuf *header;
1218
1219                 if (nb_rx == nb_pkts)
1220                         break;
1221
1222                 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
1223                 if (num != 1)
1224                         continue;
1225
1226                 i++;
1227
1228                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1229                 PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
1230
1231                 rxm = rcv_pkts[0];
1232
1233                 if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
1234                         PMD_RX_LOG(ERR, "Packet drop");
1235                         nb_enqueued++;
1236                         virtio_discard_rxbuf(vq, rxm);
1237                         rxvq->stats.errors++;
1238                         continue;
1239                 }
1240
1241                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
1242                         RTE_PKTMBUF_HEADROOM - hdr_size);
1243                 seg_num = header->num_buffers;
1244
1245                 if (seg_num == 0)
1246                         seg_num = 1;
1247
1248                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1249                 rxm->nb_segs = seg_num;
1250                 rxm->ol_flags = 0;
1251                 rxm->vlan_tci = 0;
1252                 rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
1253                 rxm->data_len = (uint16_t)(len[0] - hdr_size);
1254
1255                 rxm->port = rxvq->port_id;
1256                 rx_pkts[nb_rx] = rxm;
1257                 prev = rxm;
1258
1259                 if (hw->has_rx_offload &&
1260                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1261                         virtio_discard_rxbuf(vq, rxm);
1262                         rxvq->stats.errors++;
1263                         continue;
1264                 }
1265
1266                 seg_res = seg_num - 1;
1267
1268                 while (seg_res != 0) {
1269                         /*
1270                          * Get extra segments for current uncompleted packet.
1271                          */
1272                         uint16_t  rcv_cnt =
1273                                 RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
1274                         if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
1275                                 uint32_t rx_num =
1276                                         virtqueue_dequeue_burst_rx(vq,
1277                                         rcv_pkts, len, rcv_cnt);
1278                                 i += rx_num;
1279                                 rcv_cnt = rx_num;
1280                         } else {
1281                                 PMD_RX_LOG(ERR,
1282                                            "No enough segments for packet.");
1283                                 nb_enqueued++;
1284                                 virtio_discard_rxbuf(vq, rxm);
1285                                 rxvq->stats.errors++;
1286                                 break;
1287                         }
1288
1289                         extra_idx = 0;
1290
1291                         while (extra_idx < rcv_cnt) {
1292                                 rxm = rcv_pkts[extra_idx];
1293
1294                                 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1295                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1296                                 rxm->data_len = (uint16_t)(len[extra_idx]);
1297
1298                                 if (prev)
1299                                         prev->next = rxm;
1300
1301                                 prev = rxm;
1302                                 rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
1303                                 extra_idx++;
1304                         };
1305                         seg_res -= rcv_cnt;
1306                 }
1307
1308                 if (hw->vlan_strip)
1309                         rte_vlan_strip(rx_pkts[nb_rx]);
1310
1311                 VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
1312                         rx_pkts[nb_rx]->data_len);
1313
1314                 virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
1315                 nb_rx++;
1316         }
1317
1318         rxvq->stats.packets += nb_rx;
1319
1320         /* Allocate new mbuf for the used descriptor */
1321         while (likely(!virtqueue_full(vq))) {
1322                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
1323                 if (unlikely(new_mbuf == NULL)) {
1324                         struct rte_eth_dev *dev
1325                                 = &rte_eth_devices[rxvq->port_id];
1326                         dev->data->rx_mbuf_alloc_failed++;
1327                         break;
1328                 }
1329                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
1330                 if (unlikely(error)) {
1331                         rte_pktmbuf_free(new_mbuf);
1332                         break;
1333                 }
1334                 nb_enqueued++;
1335         }
1336
1337         if (likely(nb_enqueued)) {
1338                 vq_update_avail_idx(vq);
1339
1340                 if (unlikely(virtqueue_kick_prepare(vq))) {
1341                         virtqueue_notify(vq);
1342                         PMD_RX_LOG(DEBUG, "Notified");
1343                 }
1344         }
1345
1346         return nb_rx;
1347 }
1348
1349 uint16_t
1350 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1351 {
1352         struct virtnet_tx *txvq = tx_queue;
1353         struct virtqueue *vq = txvq->vq;
1354         struct virtio_hw *hw = vq->hw;
1355         uint16_t hdr_size = hw->vtnet_hdr_size;
1356         uint16_t nb_used, nb_tx = 0;
1357         int error;
1358
1359         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1360                 return nb_tx;
1361
1362         if (unlikely(nb_pkts < 1))
1363                 return nb_pkts;
1364
1365         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1366         nb_used = VIRTQUEUE_NUSED(vq);
1367
1368         virtio_rmb();
1369         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1370                 virtio_xmit_cleanup(vq, nb_used);
1371
1372         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1373                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1374                 int can_push = 0, use_indirect = 0, slots, need;
1375
1376                 /* Do VLAN tag insertion */
1377                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1378                         error = rte_vlan_insert(&txm);
1379                         if (unlikely(error)) {
1380                                 rte_pktmbuf_free(txm);
1381                                 continue;
1382                         }
1383                 }
1384
1385                 /* optimize ring usage */
1386                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1387                       vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1388                     rte_mbuf_refcnt_read(txm) == 1 &&
1389                     RTE_MBUF_DIRECT(txm) &&
1390                     txm->nb_segs == 1 &&
1391                     rte_pktmbuf_headroom(txm) >= hdr_size &&
1392                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1393                                    __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1394                         can_push = 1;
1395                 else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1396                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1397                         use_indirect = 1;
1398
1399                 /* How many main ring entries are needed to this Tx?
1400                  * any_layout => number of segments
1401                  * indirect   => 1
1402                  * default    => number of segments + 1
1403                  */
1404                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1405                 need = slots - vq->vq_free_cnt;
1406
1407                 /* Positive value indicates it need free vring descriptors */
1408                 if (unlikely(need > 0)) {
1409                         nb_used = VIRTQUEUE_NUSED(vq);
1410                         virtio_rmb();
1411                         need = RTE_MIN(need, (int)nb_used);
1412
1413                         virtio_xmit_cleanup(vq, need);
1414                         need = slots - vq->vq_free_cnt;
1415                         if (unlikely(need > 0)) {
1416                                 PMD_TX_LOG(ERR,
1417                                            "No free tx descriptors to transmit");
1418                                 break;
1419                         }
1420                 }
1421
1422                 /* Enqueue Packet buffers */
1423                 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
1424                         can_push, 0);
1425
1426                 virtio_update_packet_stats(&txvq->stats, txm);
1427         }
1428
1429         txvq->stats.packets += nb_tx;
1430
1431         if (likely(nb_tx)) {
1432                 vq_update_avail_idx(vq);
1433
1434                 if (unlikely(virtqueue_kick_prepare(vq))) {
1435                         virtqueue_notify(vq);
1436                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1437                 }
1438         }
1439
1440         return nb_tx;
1441 }
1442
1443 uint16_t
1444 virtio_xmit_pkts_inorder(void *tx_queue,
1445                         struct rte_mbuf **tx_pkts,
1446                         uint16_t nb_pkts)
1447 {
1448         struct virtnet_tx *txvq = tx_queue;
1449         struct virtqueue *vq = txvq->vq;
1450         struct virtio_hw *hw = vq->hw;
1451         uint16_t hdr_size = hw->vtnet_hdr_size;
1452         uint16_t nb_used, nb_avail, nb_tx = 0, nb_inorder_pkts = 0;
1453         struct rte_mbuf *inorder_pkts[nb_pkts];
1454         int error;
1455
1456         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1457                 return nb_tx;
1458
1459         if (unlikely(nb_pkts < 1))
1460                 return nb_pkts;
1461
1462         VIRTQUEUE_DUMP(vq);
1463         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1464         nb_used = VIRTQUEUE_NUSED(vq);
1465
1466         virtio_rmb();
1467         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1468                 virtio_xmit_cleanup_inorder(vq, nb_used);
1469
1470         if (unlikely(!vq->vq_free_cnt))
1471                 virtio_xmit_cleanup_inorder(vq, nb_used);
1472
1473         nb_avail = RTE_MIN(vq->vq_free_cnt, nb_pkts);
1474
1475         for (nb_tx = 0; nb_tx < nb_avail; nb_tx++) {
1476                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1477                 int slots, need;
1478
1479                 /* Do VLAN tag insertion */
1480                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1481                         error = rte_vlan_insert(&txm);
1482                         if (unlikely(error)) {
1483                                 rte_pktmbuf_free(txm);
1484                                 continue;
1485                         }
1486                 }
1487
1488                 /* optimize ring usage */
1489                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1490                      vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1491                      rte_mbuf_refcnt_read(txm) == 1 &&
1492                      RTE_MBUF_DIRECT(txm) &&
1493                      txm->nb_segs == 1 &&
1494                      rte_pktmbuf_headroom(txm) >= hdr_size &&
1495                      rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1496                                 __alignof__(struct virtio_net_hdr_mrg_rxbuf))) {
1497                         inorder_pkts[nb_inorder_pkts] = txm;
1498                         nb_inorder_pkts++;
1499
1500                         virtio_update_packet_stats(&txvq->stats, txm);
1501                         continue;
1502                 }
1503
1504                 if (nb_inorder_pkts) {
1505                         virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
1506                                                         nb_inorder_pkts);
1507                         nb_inorder_pkts = 0;
1508                 }
1509
1510                 slots = txm->nb_segs + 1;
1511                 need = slots - vq->vq_free_cnt;
1512                 if (unlikely(need > 0)) {
1513                         nb_used = VIRTQUEUE_NUSED(vq);
1514                         virtio_rmb();
1515                         need = RTE_MIN(need, (int)nb_used);
1516
1517                         virtio_xmit_cleanup_inorder(vq, need);
1518
1519                         need = slots - vq->vq_free_cnt;
1520
1521                         if (unlikely(need > 0)) {
1522                                 PMD_TX_LOG(ERR,
1523                                         "No free tx descriptors to transmit");
1524                                 break;
1525                         }
1526                 }
1527                 /* Enqueue Packet buffers */
1528                 virtqueue_enqueue_xmit(txvq, txm, slots, 0, 0, 1);
1529
1530                 virtio_update_packet_stats(&txvq->stats, txm);
1531         }
1532
1533         /* Transmit all inorder packets */
1534         if (nb_inorder_pkts)
1535                 virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
1536                                                 nb_inorder_pkts);
1537
1538         txvq->stats.packets += nb_tx;
1539
1540         if (likely(nb_tx)) {
1541                 vq_update_avail_idx(vq);
1542
1543                 if (unlikely(virtqueue_kick_prepare(vq))) {
1544                         virtqueue_notify(vq);
1545                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1546                 }
1547         }
1548
1549         VIRTQUEUE_DUMP(vq);
1550
1551         return nb_tx;
1552 }