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 void
 110 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
 111 {
 112         uint32_t s = mbuf->pkt_len;
 113         struct rte_ether_addr *ea;
 114
 115         stats->bytes += s;
 116
 117         if (s == 64) {
 118                 stats->size_bins[1]++;
 119         } else if (s > 64 && s < 1024) {
 120                 uint32_t bin;
 121
 122                 /* count zeros, and offset into correct bin */
 123                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
 124                 stats->size_bins[bin]++;
 125         } else {
 126                 if (s < 64)
 127                         stats->size_bins[0]++;
 128                 else if (s < 1519)
 129                         stats->size_bins[6]++;
 130                 else
 131                         stats->size_bins[7]++;
 132         }
 133
 134         ea = rte_pktmbuf_mtod(mbuf, struct rte_ether_addr *);
 135         if (rte_is_multicast_ether_addr(ea)) {
 136                 if (rte_is_broadcast_ether_addr(ea))
 137                         stats->broadcast++;
 138                 else
 139                         stats->multicast++;
 140         }
 141 }
 142
 143 static inline void
 144 virtio_rx_stats_updated(struct virtnet_rx *rxvq, struct rte_mbuf *m)
 145 {
 146         VIRTIO_DUMP_PACKET(m, m->data_len);
 147
 148         virtio_update_packet_stats(&rxvq->stats, m);
 149 }
 150
 151 static uint16_t
 152 virtqueue_dequeue_burst_rx_packed(struct virtqueue *vq,
 153                                   struct rte_mbuf **rx_pkts,
 154                                   uint32_t *len,
 155                                   uint16_t num)
 156 {
 157         struct rte_mbuf *cookie;
 158         uint16_t used_idx;
 159         uint16_t id;
 160         struct vring_packed_desc *desc;
 161         uint16_t i;
 162
 163         desc = vq->vq_packed.ring.desc;
 164
 165         for (i = 0; i < num; i++) {
 166                 used_idx = vq->vq_used_cons_idx;
 167                 if (!desc_is_used(&desc[used_idx], vq))
 168                         return i;
 169                 virtio_rmb(vq->hw->weak_barriers);
 170                 len[i] = desc[used_idx].len;
 171                 id = desc[used_idx].id;
 172                 cookie = (struct rte_mbuf *)vq->vq_descx[id].cookie;
 173                 if (unlikely(cookie == NULL)) {
 174                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
 175                                 vq->vq_used_cons_idx);
 176                         break;
 177                 }
 178                 rte_prefetch0(cookie);
 179                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 180                 rx_pkts[i] = cookie;
 181
 182                 vq->vq_free_cnt++;
 183                 vq->vq_used_cons_idx++;
 184                 if (vq->vq_used_cons_idx >= vq->vq_nentries) {
 185                         vq->vq_used_cons_idx -= vq->vq_nentries;
 186                         vq->vq_packed.used_wrap_counter ^= 1;
 187                 }
 188         }
 189
 190         return i;
 191 }
 192
 193 static uint16_t
 194 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
 195                            uint32_t *len, uint16_t num)
 196 {
 197         struct vring_used_elem *uep;
 198         struct rte_mbuf *cookie;
 199         uint16_t used_idx, desc_idx;
 200         uint16_t i;
 201
 202         /*  Caller does the check */
 203         for (i = 0; i < num ; i++) {
 204                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 205                 uep = &vq->vq_split.ring.used->ring[used_idx];
 206                 desc_idx = (uint16_t) uep->id;
 207                 len[i] = uep->len;
 208                 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
 209
 210                 if (unlikely(cookie == NULL)) {
 211                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
 212                                 vq->vq_used_cons_idx);
 213                         break;
 214                 }
 215
 216                 rte_prefetch0(cookie);
 217                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 218                 rx_pkts[i]  = cookie;
 219                 vq->vq_used_cons_idx++;
 220                 vq_ring_free_chain(vq, desc_idx);
 221                 vq->vq_descx[desc_idx].cookie = NULL;
 222         }
 223
 224         return i;
 225 }
 226
 227 static uint16_t
 228 virtqueue_dequeue_rx_inorder(struct virtqueue *vq,
 229                         struct rte_mbuf **rx_pkts,
 230                         uint32_t *len,
 231                         uint16_t num)
 232 {
 233         struct vring_used_elem *uep;
 234         struct rte_mbuf *cookie;
 235         uint16_t used_idx = 0;
 236         uint16_t i;
 237
 238         if (unlikely(num == 0))
 239                 return 0;
 240
 241         for (i = 0; i < num; i++) {
 242                 used_idx = vq->vq_used_cons_idx & (vq->vq_nentries - 1);
 243                 /* Desc idx same as used idx */
 244                 uep = &vq->vq_split.ring.used->ring[used_idx];
 245                 len[i] = uep->len;
 246                 cookie = (struct rte_mbuf *)vq->vq_descx[used_idx].cookie;
 247
 248                 if (unlikely(cookie == NULL)) {
 249                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
 250                                 vq->vq_used_cons_idx);
 251                         break;
 252                 }
 253
 254                 rte_prefetch0(cookie);
 255                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 256                 rx_pkts[i]  = cookie;
 257                 vq->vq_used_cons_idx++;
 258                 vq->vq_descx[used_idx].cookie = NULL;
 259         }
 260
 261         vq_ring_free_inorder(vq, used_idx, i);
 262         return i;
 263 }
 264
 265 #ifndef DEFAULT_TX_FREE_THRESH
 266 #define DEFAULT_TX_FREE_THRESH 32
 267 #endif
 268
 269 static void
 270 virtio_xmit_cleanup_inorder_packed(struct virtqueue *vq, int num)
 271 {
 272         uint16_t used_idx, id, curr_id, free_cnt = 0;
 273         uint16_t size = vq->vq_nentries;
 274         struct vring_packed_desc *desc = vq->vq_packed.ring.desc;
 275         struct vq_desc_extra *dxp;
 276
 277         used_idx = vq->vq_used_cons_idx;
 278         while (num > 0 && desc_is_used(&desc[used_idx], vq)) {
 279                 virtio_rmb(vq->hw->weak_barriers);
 280                 id = desc[used_idx].id;
 281                 do {
 282                         curr_id = used_idx;
 283                         dxp = &vq->vq_descx[used_idx];
 284                         used_idx += dxp->ndescs;
 285                         free_cnt += dxp->ndescs;
 286                         num -= dxp->ndescs;
 287                         if (used_idx >= size) {
 288                                 used_idx -= size;
 289                                 vq->vq_packed.used_wrap_counter ^= 1;
 290                         }
 291                         if (dxp->cookie != NULL) {
 292                                 rte_pktmbuf_free(dxp->cookie);
 293                                 dxp->cookie = NULL;
 294                         }
 295                 } while (curr_id != id);
 296         }
 297         vq->vq_used_cons_idx = used_idx;
 298         vq->vq_free_cnt += free_cnt;
 299 }
 300
 301 static void
 302 virtio_xmit_cleanup_normal_packed(struct virtqueue *vq, int num)
 303 {
 304         uint16_t used_idx, id;
 305         uint16_t size = vq->vq_nentries;
 306         struct vring_packed_desc *desc = vq->vq_packed.ring.desc;
 307         struct vq_desc_extra *dxp;
 308
 309         used_idx = vq->vq_used_cons_idx;
 310         while (num-- && desc_is_used(&desc[used_idx], vq)) {
 311                 virtio_rmb(vq->hw->weak_barriers);
 312                 id = desc[used_idx].id;
 313                 dxp = &vq->vq_descx[id];
 314                 vq->vq_used_cons_idx += dxp->ndescs;
 315                 if (vq->vq_used_cons_idx >= size) {
 316                         vq->vq_used_cons_idx -= size;
 317                         vq->vq_packed.used_wrap_counter ^= 1;
 318                 }
 319                 vq_ring_free_id_packed(vq, id);
 320                 if (dxp->cookie != NULL) {
 321                         rte_pktmbuf_free(dxp->cookie);
 322                         dxp->cookie = NULL;
 323                 }
 324                 used_idx = vq->vq_used_cons_idx;
 325         }
 326 }
 327
 328 /* Cleanup from completed transmits. */
 329 static inline void
 330 virtio_xmit_cleanup_packed(struct virtqueue *vq, int num, int in_order)
 331 {
 332         if (in_order)
 333                 virtio_xmit_cleanup_inorder_packed(vq, num);
 334         else
 335                 virtio_xmit_cleanup_normal_packed(vq, num);
 336 }
 337
 338 static void
 339 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
 340 {
 341         uint16_t i, used_idx, desc_idx;
 342         for (i = 0; i < num; i++) {
 343                 struct vring_used_elem *uep;
 344                 struct vq_desc_extra *dxp;
 345
 346                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 347                 uep = &vq->vq_split.ring.used->ring[used_idx];
 348
 349                 desc_idx = (uint16_t) uep->id;
 350                 dxp = &vq->vq_descx[desc_idx];
 351                 vq->vq_used_cons_idx++;
 352                 vq_ring_free_chain(vq, desc_idx);
 353
 354                 if (dxp->cookie != NULL) {
 355                         rte_pktmbuf_free(dxp->cookie);
 356                         dxp->cookie = NULL;
 357                 }
 358         }
 359 }
 360
 361 /* Cleanup from completed inorder transmits. */
 362 static __rte_always_inline void
 363 virtio_xmit_cleanup_inorder(struct virtqueue *vq, uint16_t num)
 364 {
 365         uint16_t i, idx = vq->vq_used_cons_idx;
 366         int16_t free_cnt = 0;
 367         struct vq_desc_extra *dxp = NULL;
 368
 369         if (unlikely(num == 0))
 370                 return;
 371
 372         for (i = 0; i < num; i++) {
 373                 dxp = &vq->vq_descx[idx++ & (vq->vq_nentries - 1)];
 374                 free_cnt += dxp->ndescs;
 375                 if (dxp->cookie != NULL) {
 376                         rte_pktmbuf_free(dxp->cookie);
 377                         dxp->cookie = NULL;
 378                 }
 379         }
 380
 381         vq->vq_free_cnt += free_cnt;
 382         vq->vq_used_cons_idx = idx;
 383 }
 384
 385 static inline int
 386 virtqueue_enqueue_refill_inorder(struct virtqueue *vq,
 387                         struct rte_mbuf **cookies,
 388                         uint16_t num)
 389 {
 390         struct vq_desc_extra *dxp;
 391         struct virtio_hw *hw = vq->hw;
 392         struct vring_desc *start_dp;
 393         uint16_t head_idx, idx, i = 0;
 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         head_idx = vq->vq_desc_head_idx & (vq->vq_nentries - 1);
 401         start_dp = vq->vq_split.ring.desc;
 402
 403         while (i < num) {
 404                 idx = head_idx & (vq->vq_nentries - 1);
 405                 dxp = &vq->vq_descx[idx];
 406                 dxp->cookie = (void *)cookies[i];
 407                 dxp->ndescs = 1;
 408
 409                 start_dp[idx].addr =
 410                                 VIRTIO_MBUF_ADDR(cookies[i], vq) +
 411                                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 412                 start_dp[idx].len =
 413                                 cookies[i]->buf_len -
 414                                 RTE_PKTMBUF_HEADROOM +
 415                                 hw->vtnet_hdr_size;
 416                 start_dp[idx].flags =  VRING_DESC_F_WRITE;
 417
 418                 vq_update_avail_ring(vq, idx);
 419                 head_idx++;
 420                 i++;
 421         }
 422
 423         vq->vq_desc_head_idx += num;
 424         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 425         return 0;
 426 }
 427
 428 static inline int
 429 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf **cookie,
 430                                 uint16_t num)
 431 {
 432         struct vq_desc_extra *dxp;
 433         struct virtio_hw *hw = vq->hw;
 434         struct vring_desc *start_dp = vq->vq_split.ring.desc;
 435         uint16_t idx, i;
 436
 437         if (unlikely(vq->vq_free_cnt == 0))
 438                 return -ENOSPC;
 439         if (unlikely(vq->vq_free_cnt < num))
 440                 return -EMSGSIZE;
 441
 442         if (unlikely(vq->vq_desc_head_idx >= vq->vq_nentries))
 443                 return -EFAULT;
 444
 445         for (i = 0; i < num; i++) {
 446                 idx = vq->vq_desc_head_idx;
 447                 dxp = &vq->vq_descx[idx];
 448                 dxp->cookie = (void *)cookie[i];
 449                 dxp->ndescs = 1;
 450
 451                 start_dp[idx].addr =
 452                         VIRTIO_MBUF_ADDR(cookie[i], vq) +
 453                         RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 454                 start_dp[idx].len =
 455                         cookie[i]->buf_len - RTE_PKTMBUF_HEADROOM +
 456                         hw->vtnet_hdr_size;
 457                 start_dp[idx].flags = VRING_DESC_F_WRITE;
 458                 vq->vq_desc_head_idx = start_dp[idx].next;
 459                 vq_update_avail_ring(vq, idx);
 460                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END) {
 461                         vq->vq_desc_tail_idx = vq->vq_desc_head_idx;
 462                         break;
 463                 }
 464         }
 465
 466         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 467
 468         return 0;
 469 }
 470
 471 static inline int
 472 virtqueue_enqueue_recv_refill_packed(struct virtqueue *vq,
 473                                      struct rte_mbuf **cookie, uint16_t num)
 474 {
 475         struct vring_packed_desc *start_dp = vq->vq_packed.ring.desc;
 476         uint16_t flags = vq->vq_packed.cached_flags;
 477         struct virtio_hw *hw = vq->hw;
 478         struct vq_desc_extra *dxp;
 479         uint16_t idx;
 480         int i;
 481
 482         if (unlikely(vq->vq_free_cnt == 0))
 483                 return -ENOSPC;
 484         if (unlikely(vq->vq_free_cnt < num))
 485                 return -EMSGSIZE;
 486
 487         for (i = 0; i < num; i++) {
 488                 idx = vq->vq_avail_idx;
 489                 dxp = &vq->vq_descx[idx];
 490                 dxp->cookie = (void *)cookie[i];
 491                 dxp->ndescs = 1;
 492
 493                 start_dp[idx].addr = VIRTIO_MBUF_ADDR(cookie[i], vq) +
 494                                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 495                 start_dp[idx].len = cookie[i]->buf_len - RTE_PKTMBUF_HEADROOM
 496                                         + hw->vtnet_hdr_size;
 497
 498                 vq->vq_desc_head_idx = dxp->next;
 499                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 500                         vq->vq_desc_tail_idx = vq->vq_desc_head_idx;
 501                 virtio_wmb(hw->weak_barriers);
 502                 start_dp[idx].flags = flags;
 503                 if (++vq->vq_avail_idx >= vq->vq_nentries) {
 504                         vq->vq_avail_idx -= vq->vq_nentries;
 505                         vq->vq_packed.cached_flags ^=
 506                                 VRING_PACKED_DESC_F_AVAIL_USED;
 507                         flags = vq->vq_packed.cached_flags;
 508                 }
 509         }
 510         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 511         return 0;
 512 }
 513
 514 /* When doing TSO, the IP length is not included in the pseudo header
 515  * checksum of the packet given to the PMD, but for virtio it is
 516  * expected.
 517  */
 518 static void
 519 virtio_tso_fix_cksum(struct rte_mbuf *m)
 520 {
 521         /* common case: header is not fragmented */
 522         if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
 523                         m->l4_len)) {
 524                 struct rte_ipv4_hdr *iph;
 525                 struct rte_ipv6_hdr *ip6h;
 526                 struct rte_tcp_hdr *th;
 527                 uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
 528                 uint32_t tmp;
 529
 530                 iph = rte_pktmbuf_mtod_offset(m,
 531                                         struct rte_ipv4_hdr *, m->l2_len);
 532                 th = RTE_PTR_ADD(iph, m->l3_len);
 533                 if ((iph->version_ihl >> 4) == 4) {
 534                         iph->hdr_checksum = 0;
 535                         iph->hdr_checksum = rte_ipv4_cksum(iph);
 536                         ip_len = iph->total_length;
 537                         ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
 538                                 m->l3_len);
 539                 } else {
 540                         ip6h = (struct rte_ipv6_hdr *)iph;
 541                         ip_paylen = ip6h->payload_len;
 542                 }
 543
 544                 /* calculate the new phdr checksum not including ip_paylen */
 545                 prev_cksum = th->cksum;
 546                 tmp = prev_cksum;
 547                 tmp += ip_paylen;
 548                 tmp = (tmp & 0xffff) + (tmp >> 16);
 549                 new_cksum = tmp;
 550
 551                 /* replace it in the packet */
 552                 th->cksum = new_cksum;
 553         }
 554 }
 555
 556
 557 /* avoid write operation when necessary, to lessen cache issues */
 558 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 559         if ((var) != (val))                     \
 560                 (var) = (val);                  \
 561 } while (0)
 562
 563 #define virtqueue_clear_net_hdr(_hdr) do {              \
 564         ASSIGN_UNLESS_EQUAL((_hdr)->csum_start, 0);     \
 565         ASSIGN_UNLESS_EQUAL((_hdr)->csum_offset, 0);    \
 566         ASSIGN_UNLESS_EQUAL((_hdr)->flags, 0);          \
 567         ASSIGN_UNLESS_EQUAL((_hdr)->gso_type, 0);       \
 568         ASSIGN_UNLESS_EQUAL((_hdr)->gso_size, 0);       \
 569         ASSIGN_UNLESS_EQUAL((_hdr)->hdr_len, 0);        \
 570 } while (0)
 571
 572 static inline void
 573 virtqueue_xmit_offload(struct virtio_net_hdr *hdr,
 574                         struct rte_mbuf *cookie,
 575                         bool offload)
 576 {
 577         if (offload) {
 578                 if (cookie->ol_flags & PKT_TX_TCP_SEG)
 579                         cookie->ol_flags |= PKT_TX_TCP_CKSUM;
 580
 581                 switch (cookie->ol_flags & PKT_TX_L4_MASK) {
 582                 case PKT_TX_UDP_CKSUM:
 583                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 584                         hdr->csum_offset = offsetof(struct rte_udp_hdr,
 585                                 dgram_cksum);
 586                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 587                         break;
 588
 589                 case PKT_TX_TCP_CKSUM:
 590                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 591                         hdr->csum_offset = offsetof(struct rte_tcp_hdr, cksum);
 592                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 593                         break;
 594
 595                 default:
 596                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 597                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 598                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 599                         break;
 600                 }
 601
 602                 /* TCP Segmentation Offload */
 603                 if (cookie->ol_flags & PKT_TX_TCP_SEG) {
 604                         hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
 605                                 VIRTIO_NET_HDR_GSO_TCPV6 :
 606                                 VIRTIO_NET_HDR_GSO_TCPV4;
 607                         hdr->gso_size = cookie->tso_segsz;
 608                         hdr->hdr_len =
 609                                 cookie->l2_len +
 610                                 cookie->l3_len +
 611                                 cookie->l4_len;
 612                 } else {
 613                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 614                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 615                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 616                 }
 617         }
 618 }
 619
 620 static inline void
 621 virtqueue_enqueue_xmit_inorder(struct virtnet_tx *txvq,
 622                         struct rte_mbuf **cookies,
 623                         uint16_t num)
 624 {
 625         struct vq_desc_extra *dxp;
 626         struct virtqueue *vq = txvq->vq;
 627         struct vring_desc *start_dp;
 628         struct virtio_net_hdr *hdr;
 629         uint16_t idx;
 630         uint16_t head_size = vq->hw->vtnet_hdr_size;
 631         uint16_t i = 0;
 632
 633         idx = vq->vq_desc_head_idx;
 634         start_dp = vq->vq_split.ring.desc;
 635
 636         while (i < num) {
 637                 idx = idx & (vq->vq_nentries - 1);
 638                 dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
 639                 dxp->cookie = (void *)cookies[i];
 640                 dxp->ndescs = 1;
 641                 virtio_update_packet_stats(&txvq->stats, cookies[i]);
 642
 643                 hdr = (struct virtio_net_hdr *)(char *)cookies[i]->buf_addr +
 644                         cookies[i]->data_off - head_size;
 645
 646                 /* if offload disabled, hdr is not zeroed yet, do it now */
 647                 if (!vq->hw->has_tx_offload)
 648                         virtqueue_clear_net_hdr(hdr);
 649                 else
 650                         virtqueue_xmit_offload(hdr, cookies[i], true);
 651
 652                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookies[i], vq);
 653                 start_dp[idx].len   = cookies[i]->data_len + head_size;
 654                 start_dp[idx].flags = 0;
 655
 656
 657                 vq_update_avail_ring(vq, idx);
 658
 659                 idx++;
 660                 i++;
 661         };
 662
 663         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 664         vq->vq_desc_head_idx = idx & (vq->vq_nentries - 1);
 665 }
 666
 667 static inline void
 668 virtqueue_enqueue_xmit_packed_fast(struct virtnet_tx *txvq,
 669                                    struct rte_mbuf *cookie,
 670                                    int in_order)
 671 {
 672         struct virtqueue *vq = txvq->vq;
 673         struct vring_packed_desc *dp;
 674         struct vq_desc_extra *dxp;
 675         uint16_t idx, id, flags;
 676         uint16_t head_size = vq->hw->vtnet_hdr_size;
 677         struct virtio_net_hdr *hdr;
 678
 679         id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx;
 680         idx = vq->vq_avail_idx;
 681         dp = &vq->vq_packed.ring.desc[idx];
 682
 683         dxp = &vq->vq_descx[id];
 684         dxp->ndescs = 1;
 685         dxp->cookie = cookie;
 686
 687         flags = vq->vq_packed.cached_flags;
 688
 689         /* prepend cannot fail, checked by caller */
 690         hdr = (struct virtio_net_hdr *)(char *)cookie->buf_addr +
 691                 cookie->data_off - head_size;
 692
 693         /* if offload disabled, hdr is not zeroed yet, do it now */
 694         if (!vq->hw->has_tx_offload)
 695                 virtqueue_clear_net_hdr(hdr);
 696         else
 697                 virtqueue_xmit_offload(hdr, cookie, true);
 698
 699         dp->addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
 700         dp->len  = cookie->data_len + head_size;
 701         dp->id   = id;
 702
 703         if (++vq->vq_avail_idx >= vq->vq_nentries) {
 704                 vq->vq_avail_idx -= vq->vq_nentries;
 705                 vq->vq_packed.cached_flags ^= VRING_PACKED_DESC_F_AVAIL_USED;
 706         }
 707
 708         vq->vq_free_cnt--;
 709
 710         if (!in_order) {
 711                 vq->vq_desc_head_idx = dxp->next;
 712                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 713                         vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
 714         }
 715
 716         virtio_wmb(vq->hw->weak_barriers);
 717         dp->flags = flags;
 718 }
 719
 720 static inline void
 721 virtqueue_enqueue_xmit_packed(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
 722                               uint16_t needed, int can_push, int in_order)
 723 {
 724         struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 725         struct vq_desc_extra *dxp;
 726         struct virtqueue *vq = txvq->vq;
 727         struct vring_packed_desc *start_dp, *head_dp;
 728         uint16_t idx, id, head_idx, head_flags;
 729         uint16_t head_size = vq->hw->vtnet_hdr_size;
 730         struct virtio_net_hdr *hdr;
 731         uint16_t prev;
 732         bool prepend_header = false;
 733
 734         id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx;
 735
 736         dxp = &vq->vq_descx[id];
 737         dxp->ndescs = needed;
 738         dxp->cookie = cookie;
 739
 740         head_idx = vq->vq_avail_idx;
 741         idx = head_idx;
 742         prev = head_idx;
 743         start_dp = vq->vq_packed.ring.desc;
 744
 745         head_dp = &vq->vq_packed.ring.desc[idx];
 746         head_flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 747         head_flags |= vq->vq_packed.cached_flags;
 748
 749         if (can_push) {
 750                 /* prepend cannot fail, checked by caller */
 751                 hdr = (struct virtio_net_hdr *)(char *)cookie->buf_addr +
 752                         cookie->data_off - head_size;
 753                 prepend_header = true;
 754
 755                 /* if offload disabled, it is not zeroed below, do it now */
 756                 if (!vq->hw->has_tx_offload)
 757                         virtqueue_clear_net_hdr(hdr);
 758         } else {
 759                 /* setup first tx ring slot to point to header
 760                  * stored in reserved region.
 761                  */
 762                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 763                         RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
 764                 start_dp[idx].len   = vq->hw->vtnet_hdr_size;
 765                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 766                 idx++;
 767                 if (idx >= vq->vq_nentries) {
 768                         idx -= vq->vq_nentries;
 769                         vq->vq_packed.cached_flags ^=
 770                                 VRING_PACKED_DESC_F_AVAIL_USED;
 771                 }
 772         }
 773
 774         virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload);
 775
 776         do {
 777                 uint16_t flags;
 778
 779                 start_dp[idx].addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
 780                 start_dp[idx].len  = cookie->data_len;
 781                 if (prepend_header) {
 782                         start_dp[idx].len += head_size;
 783                         prepend_header = false;
 784                 }
 785
 786                 if (likely(idx != head_idx)) {
 787                         flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 788                         flags |= vq->vq_packed.cached_flags;
 789                         start_dp[idx].flags = flags;
 790                 }
 791                 prev = idx;
 792                 idx++;
 793                 if (idx >= vq->vq_nentries) {
 794                         idx -= vq->vq_nentries;
 795                         vq->vq_packed.cached_flags ^=
 796                                 VRING_PACKED_DESC_F_AVAIL_USED;
 797                 }
 798         } while ((cookie = cookie->next) != NULL);
 799
 800         start_dp[prev].id = id;
 801
 802         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 803         vq->vq_avail_idx = idx;
 804
 805         if (!in_order) {
 806                 vq->vq_desc_head_idx = dxp->next;
 807                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 808                         vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
 809         }
 810
 811         virtio_wmb(vq->hw->weak_barriers);
 812         head_dp->flags = head_flags;
 813 }
 814
 815 static inline void
 816 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
 817                         uint16_t needed, int use_indirect, int can_push,
 818                         int in_order)
 819 {
 820         struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 821         struct vq_desc_extra *dxp;
 822         struct virtqueue *vq = txvq->vq;
 823         struct vring_desc *start_dp;
 824         uint16_t seg_num = cookie->nb_segs;
 825         uint16_t head_idx, idx;
 826         uint16_t head_size = vq->hw->vtnet_hdr_size;
 827         bool prepend_header = false;
 828         struct virtio_net_hdr *hdr;
 829
 830         head_idx = vq->vq_desc_head_idx;
 831         idx = head_idx;
 832         if (in_order)
 833                 dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
 834         else
 835                 dxp = &vq->vq_descx[idx];
 836         dxp->cookie = (void *)cookie;
 837         dxp->ndescs = needed;
 838
 839         start_dp = vq->vq_split.ring.desc;
 840
 841         if (can_push) {
 842                 /* prepend cannot fail, checked by caller */
 843                 hdr = (struct virtio_net_hdr *)(char *)cookie->buf_addr +
 844                         cookie->data_off - head_size;
 845                 prepend_header = true;
 846
 847                 /* if offload disabled, it is not zeroed below, do it now */
 848                 if (!vq->hw->has_tx_offload)
 849                         virtqueue_clear_net_hdr(hdr);
 850         } else if (use_indirect) {
 851                 /* setup tx ring slot to point to indirect
 852                  * descriptor list stored in reserved region.
 853                  *
 854                  * the first slot in indirect ring is already preset
 855                  * to point to the header in reserved region
 856                  */
 857                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 858                         RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
 859                 start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
 860                 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
 861                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 862
 863                 /* loop below will fill in rest of the indirect elements */
 864                 start_dp = txr[idx].tx_indir;
 865                 idx = 1;
 866         } else {
 867                 /* setup first tx ring slot to point to header
 868                  * stored in reserved region.
 869                  */
 870                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 871                         RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
 872                 start_dp[idx].len   = vq->hw->vtnet_hdr_size;
 873                 start_dp[idx].flags = VRING_DESC_F_NEXT;
 874                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 875
 876                 idx = start_dp[idx].next;
 877         }
 878
 879         virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload);
 880
 881         do {
 882                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
 883                 start_dp[idx].len   = cookie->data_len;
 884                 if (prepend_header) {
 885                         start_dp[idx].len += head_size;
 886                         prepend_header = false;
 887                 }
 888                 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 889                 idx = start_dp[idx].next;
 890         } while ((cookie = cookie->next) != NULL);
 891
 892         if (use_indirect)
 893                 idx = vq->vq_split.ring.desc[head_idx].next;
 894
 895         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 896
 897         vq->vq_desc_head_idx = idx;
 898         vq_update_avail_ring(vq, head_idx);
 899
 900         if (!in_order) {
 901                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 902                         vq->vq_desc_tail_idx = idx;
 903         }
 904 }
 905
 906 void
 907 virtio_dev_cq_start(struct rte_eth_dev *dev)
 908 {
 909         struct virtio_hw *hw = dev->data->dev_private;
 910
 911         if (hw->cvq && hw->cvq->vq) {
 912                 rte_spinlock_init(&hw->cvq->lock);
 913                 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
 914         }
 915 }
 916
 917 int
 918 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
 919                         uint16_t queue_idx,
 920                         uint16_t nb_desc,
 921                         unsigned int socket_id __rte_unused,
 922                         const struct rte_eth_rxconf *rx_conf __rte_unused,
 923                         struct rte_mempool *mp)
 924 {
 925         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 926         struct virtio_hw *hw = dev->data->dev_private;
 927         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 928         struct virtnet_rx *rxvq;
 929
 930         PMD_INIT_FUNC_TRACE();
 931
 932         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 933                 nb_desc = vq->vq_nentries;
 934         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 935
 936         rxvq = &vq->rxq;
 937         rxvq->queue_id = queue_idx;
 938         rxvq->mpool = mp;
 939         dev->data->rx_queues[queue_idx] = rxvq;
 940
 941         return 0;
 942 }
 943
 944 int
 945 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
 946 {
 947         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 948         struct virtio_hw *hw = dev->data->dev_private;
 949         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 950         struct virtnet_rx *rxvq = &vq->rxq;
 951         struct rte_mbuf *m;
 952         uint16_t desc_idx;
 953         int error, nbufs, i;
 954
 955         PMD_INIT_FUNC_TRACE();
 956
 957         /* Allocate blank mbufs for the each rx descriptor */
 958         nbufs = 0;
 959
 960         if (hw->use_simple_rx) {
 961                 for (desc_idx = 0; desc_idx < vq->vq_nentries;
 962                      desc_idx++) {
 963                         vq->vq_split.ring.avail->ring[desc_idx] = desc_idx;
 964                         vq->vq_split.ring.desc[desc_idx].flags =
 965                                 VRING_DESC_F_WRITE;
 966                 }
 967
 968                 virtio_rxq_vec_setup(rxvq);
 969         }
 970
 971         memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
 972         for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
 973              desc_idx++) {
 974                 vq->sw_ring[vq->vq_nentries + desc_idx] =
 975                         &rxvq->fake_mbuf;
 976         }
 977
 978         if (hw->use_simple_rx) {
 979                 while (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
 980                         virtio_rxq_rearm_vec(rxvq);
 981                         nbufs += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
 982                 }
 983         } else if (hw->use_inorder_rx) {
 984                 if ((!virtqueue_full(vq))) {
 985                         uint16_t free_cnt = vq->vq_free_cnt;
 986                         struct rte_mbuf *pkts[free_cnt];
 987
 988                         if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, pkts,
 989                                 free_cnt)) {
 990                                 error = virtqueue_enqueue_refill_inorder(vq,
 991                                                 pkts,
 992                                                 free_cnt);
 993                                 if (unlikely(error)) {
 994                                         for (i = 0; i < free_cnt; i++)
 995                                                 rte_pktmbuf_free(pkts[i]);
 996                                 }
 997                         }
 998
 999                         nbufs += free_cnt;
1000                         vq_update_avail_idx(vq);
1001                 }
1002         } else {
1003                 while (!virtqueue_full(vq)) {
1004                         m = rte_mbuf_raw_alloc(rxvq->mpool);
1005                         if (m == NULL)
1006                                 break;
1007
1008                         /* Enqueue allocated buffers */
1009                         if (vtpci_packed_queue(vq->hw))
1010                                 error = virtqueue_enqueue_recv_refill_packed(vq,
1011                                                 &m, 1);
1012                         else
1013                                 error = virtqueue_enqueue_recv_refill(vq,
1014                                                 &m, 1);
1015                         if (error) {
1016                                 rte_pktmbuf_free(m);
1017                                 break;
1018                         }
1019                         nbufs++;
1020                 }
1021
1022                 if (!vtpci_packed_queue(vq->hw))
1023                         vq_update_avail_idx(vq);
1024         }
1025
1026         PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
1027
1028         VIRTQUEUE_DUMP(vq);
1029
1030         return 0;
1031 }
1032
1033 /*
1034  * struct rte_eth_dev *dev: Used to update dev
1035  * uint16_t nb_desc: Defaults to values read from config space
1036  * unsigned int socket_id: Used to allocate memzone
1037  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
1038  * uint16_t queue_idx: Just used as an index in dev txq list
1039  */
1040 int
1041 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
1042                         uint16_t queue_idx,
1043                         uint16_t nb_desc,
1044                         unsigned int socket_id __rte_unused,
1045                         const struct rte_eth_txconf *tx_conf)
1046 {
1047         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
1048         struct virtio_hw *hw = dev->data->dev_private;
1049         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
1050         struct virtnet_tx *txvq;
1051         uint16_t tx_free_thresh;
1052
1053         PMD_INIT_FUNC_TRACE();
1054
1055         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
1056                 nb_desc = vq->vq_nentries;
1057         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
1058
1059         txvq = &vq->txq;
1060         txvq->queue_id = queue_idx;
1061
1062         tx_free_thresh = tx_conf->tx_free_thresh;
1063         if (tx_free_thresh == 0)
1064                 tx_free_thresh =
1065                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
1066
1067         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
1068                 RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
1069                         "number of TX entries minus 3 (%u)."
1070                         " (tx_free_thresh=%u port=%u queue=%u)\n",
1071                         vq->vq_nentries - 3,
1072                         tx_free_thresh, dev->data->port_id, queue_idx);
1073                 return -EINVAL;
1074         }
1075
1076         vq->vq_free_thresh = tx_free_thresh;
1077
1078         dev->data->tx_queues[queue_idx] = txvq;
1079         return 0;
1080 }
1081
1082 int
1083 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
1084                                 uint16_t queue_idx)
1085 {
1086         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
1087         struct virtio_hw *hw = dev->data->dev_private;
1088         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
1089
1090         PMD_INIT_FUNC_TRACE();
1091
1092         if (!vtpci_packed_queue(hw)) {
1093                 if (hw->use_inorder_tx)
1094                         vq->vq_split.ring.desc[vq->vq_nentries - 1].next = 0;
1095         }
1096
1097         VIRTQUEUE_DUMP(vq);
1098
1099         return 0;
1100 }
1101
1102 static inline void
1103 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
1104 {
1105         int error;
1106         /*
1107          * Requeue the discarded mbuf. This should always be
1108          * successful since it was just dequeued.
1109          */
1110         if (vtpci_packed_queue(vq->hw))
1111                 error = virtqueue_enqueue_recv_refill_packed(vq, &m, 1);
1112         else
1113                 error = virtqueue_enqueue_recv_refill(vq, &m, 1);
1114
1115         if (unlikely(error)) {
1116                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
1117                 rte_pktmbuf_free(m);
1118         }
1119 }
1120
1121 static inline void
1122 virtio_discard_rxbuf_inorder(struct virtqueue *vq, struct rte_mbuf *m)
1123 {
1124         int error;
1125
1126         error = virtqueue_enqueue_refill_inorder(vq, &m, 1);
1127         if (unlikely(error)) {
1128                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
1129                 rte_pktmbuf_free(m);
1130         }
1131 }
1132
1133 /* Optionally fill offload information in structure */
1134 static inline int
1135 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
1136 {
1137         struct rte_net_hdr_lens hdr_lens;
1138         uint32_t hdrlen, ptype;
1139         int l4_supported = 0;
1140
1141         /* nothing to do */
1142         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1143                 return 0;
1144
1145         m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
1146
1147         ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
1148         m->packet_type = ptype;
1149         if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
1150             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
1151             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
1152                 l4_supported = 1;
1153
1154         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1155                 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
1156                 if (hdr->csum_start <= hdrlen && l4_supported) {
1157                         m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
1158                 } else {
1159                         /* Unknown proto or tunnel, do sw cksum. We can assume
1160                          * the cksum field is in the first segment since the
1161                          * buffers we provided to the host are large enough.
1162                          * In case of SCTP, this will be wrong since it's a CRC
1163                          * but there's nothing we can do.
1164                          */
1165                         uint16_t csum = 0, off;
1166
1167                         rte_raw_cksum_mbuf(m, hdr->csum_start,
1168                                 rte_pktmbuf_pkt_len(m) - hdr->csum_start,
1169                                 &csum);
1170                         if (likely(csum != 0xffff))
1171                                 csum = ~csum;
1172                         off = hdr->csum_offset + hdr->csum_start;
1173                         if (rte_pktmbuf_data_len(m) >= off + 1)
1174                                 *rte_pktmbuf_mtod_offset(m, uint16_t *,
1175                                         off) = csum;
1176                 }
1177         } else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
1178                 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
1179         }
1180
1181         /* GSO request, save required information in mbuf */
1182         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1183                 /* Check unsupported modes */
1184                 if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
1185                     (hdr->gso_size == 0)) {
1186                         return -EINVAL;
1187                 }
1188
1189                 /* Update mss lengthes in mbuf */
1190                 m->tso_segsz = hdr->gso_size;
1191                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1192                         case VIRTIO_NET_HDR_GSO_TCPV4:
1193                         case VIRTIO_NET_HDR_GSO_TCPV6:
1194                                 m->ol_flags |= PKT_RX_LRO | \
1195                                         PKT_RX_L4_CKSUM_NONE;
1196                                 break;
1197                         default:
1198                                 return -EINVAL;
1199                 }
1200         }
1201
1202         return 0;
1203 }
1204
1205 #define VIRTIO_MBUF_BURST_SZ 64
1206 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
1207 uint16_t
1208 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1209 {
1210         struct virtnet_rx *rxvq = rx_queue;
1211         struct virtqueue *vq = rxvq->vq;
1212         struct virtio_hw *hw = vq->hw;
1213         struct rte_mbuf *rxm;
1214         uint16_t nb_used, num, nb_rx;
1215         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1216         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1217         int error;
1218         uint32_t i, nb_enqueued;
1219         uint32_t hdr_size;
1220         struct virtio_net_hdr *hdr;
1221
1222         nb_rx = 0;
1223         if (unlikely(hw->started == 0))
1224                 return nb_rx;
1225
1226         nb_used = VIRTQUEUE_NUSED(vq);
1227
1228         virtio_rmb(hw->weak_barriers);
1229
1230         num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1231         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1232                 num = VIRTIO_MBUF_BURST_SZ;
1233         if (likely(num > DESC_PER_CACHELINE))
1234                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1235
1236         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1237         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
1238
1239         nb_enqueued = 0;
1240         hdr_size = hw->vtnet_hdr_size;
1241
1242         for (i = 0; i < num ; i++) {
1243                 rxm = rcv_pkts[i];
1244
1245                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1246
1247                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1248                         PMD_RX_LOG(ERR, "Packet drop");
1249                         nb_enqueued++;
1250                         virtio_discard_rxbuf(vq, rxm);
1251                         rxvq->stats.errors++;
1252                         continue;
1253                 }
1254
1255                 rxm->port = rxvq->port_id;
1256                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1257                 rxm->ol_flags = 0;
1258                 rxm->vlan_tci = 0;
1259
1260                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1261                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1262
1263                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1264                         RTE_PKTMBUF_HEADROOM - hdr_size);
1265
1266                 if (hw->vlan_strip)
1267                         rte_vlan_strip(rxm);
1268
1269                 if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1270                         virtio_discard_rxbuf(vq, rxm);
1271                         rxvq->stats.errors++;
1272                         continue;
1273                 }
1274
1275                 virtio_rx_stats_updated(rxvq, rxm);
1276
1277                 rx_pkts[nb_rx++] = rxm;
1278         }
1279
1280         rxvq->stats.packets += nb_rx;
1281
1282         /* Allocate new mbuf for the used descriptor */
1283         if (likely(!virtqueue_full(vq))) {
1284                 uint16_t free_cnt = vq->vq_free_cnt;
1285                 struct rte_mbuf *new_pkts[free_cnt];
1286
1287                 if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1288                                                 free_cnt) == 0)) {
1289                         error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1290                                         free_cnt);
1291                         if (unlikely(error)) {
1292                                 for (i = 0; i < free_cnt; i++)
1293                                         rte_pktmbuf_free(new_pkts[i]);
1294                         }
1295                         nb_enqueued += free_cnt;
1296                 } else {
1297                         struct rte_eth_dev *dev =
1298                                 &rte_eth_devices[rxvq->port_id];
1299                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1300                 }
1301         }
1302
1303         if (likely(nb_enqueued)) {
1304                 vq_update_avail_idx(vq);
1305
1306                 if (unlikely(virtqueue_kick_prepare(vq))) {
1307                         virtqueue_notify(vq);
1308                         PMD_RX_LOG(DEBUG, "Notified");
1309                 }
1310         }
1311
1312         return nb_rx;
1313 }
1314
1315 uint16_t
1316 virtio_recv_pkts_packed(void *rx_queue, struct rte_mbuf **rx_pkts,
1317                         uint16_t nb_pkts)
1318 {
1319         struct virtnet_rx *rxvq = rx_queue;
1320         struct virtqueue *vq = rxvq->vq;
1321         struct virtio_hw *hw = vq->hw;
1322         struct rte_mbuf *rxm;
1323         uint16_t num, nb_rx;
1324         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1325         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1326         int error;
1327         uint32_t i, nb_enqueued;
1328         uint32_t hdr_size;
1329         struct virtio_net_hdr *hdr;
1330
1331         nb_rx = 0;
1332         if (unlikely(hw->started == 0))
1333                 return nb_rx;
1334
1335         num = RTE_MIN(VIRTIO_MBUF_BURST_SZ, nb_pkts);
1336         if (likely(num > DESC_PER_CACHELINE))
1337                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1338
1339         num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1340         PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1341
1342         nb_enqueued = 0;
1343         hdr_size = hw->vtnet_hdr_size;
1344
1345         for (i = 0; i < num; i++) {
1346                 rxm = rcv_pkts[i];
1347
1348                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1349
1350                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1351                         PMD_RX_LOG(ERR, "Packet drop");
1352                         nb_enqueued++;
1353                         virtio_discard_rxbuf(vq, rxm);
1354                         rxvq->stats.errors++;
1355                         continue;
1356                 }
1357
1358                 rxm->port = rxvq->port_id;
1359                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1360                 rxm->ol_flags = 0;
1361                 rxm->vlan_tci = 0;
1362
1363                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1364                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1365
1366                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1367                         RTE_PKTMBUF_HEADROOM - hdr_size);
1368
1369                 if (hw->vlan_strip)
1370                         rte_vlan_strip(rxm);
1371
1372                 if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1373                         virtio_discard_rxbuf(vq, rxm);
1374                         rxvq->stats.errors++;
1375                         continue;
1376                 }
1377
1378                 virtio_rx_stats_updated(rxvq, rxm);
1379
1380                 rx_pkts[nb_rx++] = rxm;
1381         }
1382
1383         rxvq->stats.packets += nb_rx;
1384
1385         /* Allocate new mbuf for the used descriptor */
1386         if (likely(!virtqueue_full(vq))) {
1387                 uint16_t free_cnt = vq->vq_free_cnt;
1388                 struct rte_mbuf *new_pkts[free_cnt];
1389
1390                 if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1391                                                 free_cnt) == 0)) {
1392                         error = virtqueue_enqueue_recv_refill_packed(vq,
1393                                         new_pkts, free_cnt);
1394                         if (unlikely(error)) {
1395                                 for (i = 0; i < free_cnt; i++)
1396                                         rte_pktmbuf_free(new_pkts[i]);
1397                         }
1398                         nb_enqueued += free_cnt;
1399                 } else {
1400                         struct rte_eth_dev *dev =
1401                                 &rte_eth_devices[rxvq->port_id];
1402                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1403                 }
1404         }
1405
1406         if (likely(nb_enqueued)) {
1407                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1408                         virtqueue_notify(vq);
1409                         PMD_RX_LOG(DEBUG, "Notified");
1410                 }
1411         }
1412
1413         return nb_rx;
1414 }
1415
1416
1417 uint16_t
1418 virtio_recv_pkts_inorder(void *rx_queue,
1419                         struct rte_mbuf **rx_pkts,
1420                         uint16_t nb_pkts)
1421 {
1422         struct virtnet_rx *rxvq = rx_queue;
1423         struct virtqueue *vq = rxvq->vq;
1424         struct virtio_hw *hw = vq->hw;
1425         struct rte_mbuf *rxm;
1426         struct rte_mbuf *prev = NULL;
1427         uint16_t nb_used, num, nb_rx;
1428         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1429         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1430         int error;
1431         uint32_t nb_enqueued;
1432         uint32_t seg_num;
1433         uint32_t seg_res;
1434         uint32_t hdr_size;
1435         int32_t i;
1436
1437         nb_rx = 0;
1438         if (unlikely(hw->started == 0))
1439                 return nb_rx;
1440
1441         nb_used = VIRTQUEUE_NUSED(vq);
1442         nb_used = RTE_MIN(nb_used, nb_pkts);
1443         nb_used = RTE_MIN(nb_used, VIRTIO_MBUF_BURST_SZ);
1444
1445         virtio_rmb(hw->weak_barriers);
1446
1447         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1448
1449         nb_enqueued = 0;
1450         seg_num = 1;
1451         seg_res = 0;
1452         hdr_size = hw->vtnet_hdr_size;
1453
1454         num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len, nb_used);
1455
1456         for (i = 0; i < num; i++) {
1457                 struct virtio_net_hdr_mrg_rxbuf *header;
1458
1459                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1460                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1461
1462                 rxm = rcv_pkts[i];
1463
1464                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1465                         PMD_RX_LOG(ERR, "Packet drop");
1466                         nb_enqueued++;
1467                         virtio_discard_rxbuf_inorder(vq, rxm);
1468                         rxvq->stats.errors++;
1469                         continue;
1470                 }
1471
1472                 header = (struct virtio_net_hdr_mrg_rxbuf *)
1473                          ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1474                          - hdr_size);
1475
1476                 if (vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
1477                         seg_num = header->num_buffers;
1478                         if (seg_num == 0)
1479                                 seg_num = 1;
1480                 } else {
1481                         seg_num = 1;
1482                 }
1483
1484                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1485                 rxm->nb_segs = seg_num;
1486                 rxm->ol_flags = 0;
1487                 rxm->vlan_tci = 0;
1488                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1489                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1490
1491                 rxm->port = rxvq->port_id;
1492
1493                 rx_pkts[nb_rx] = rxm;
1494                 prev = rxm;
1495
1496                 if (vq->hw->has_rx_offload &&
1497                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1498                         virtio_discard_rxbuf_inorder(vq, rxm);
1499                         rxvq->stats.errors++;
1500                         continue;
1501                 }
1502
1503                 if (hw->vlan_strip)
1504                         rte_vlan_strip(rx_pkts[nb_rx]);
1505
1506                 seg_res = seg_num - 1;
1507
1508                 /* Merge remaining segments */
1509                 while (seg_res != 0 && i < (num - 1)) {
1510                         i++;
1511
1512                         rxm = rcv_pkts[i];
1513                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1514                         rxm->pkt_len = (uint32_t)(len[i]);
1515                         rxm->data_len = (uint16_t)(len[i]);
1516
1517                         rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1518
1519                         prev->next = rxm;
1520                         prev = rxm;
1521                         seg_res -= 1;
1522                 }
1523
1524                 if (!seg_res) {
1525                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1526                         nb_rx++;
1527                 }
1528         }
1529
1530         /* Last packet still need merge segments */
1531         while (seg_res != 0) {
1532                 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1533                                         VIRTIO_MBUF_BURST_SZ);
1534
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                                 extra_idx += 1;
1552                         };
1553                         seg_res -= rcv_cnt;
1554
1555                         if (!seg_res) {
1556                                 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1557                                 nb_rx++;
1558                         }
1559                 } else {
1560                         PMD_RX_LOG(ERR,
1561                                         "No enough segments for packet.");
1562                         rte_pktmbuf_free(rx_pkts[nb_rx]);
1563                         rxvq->stats.errors++;
1564                         break;
1565                 }
1566         }
1567
1568         rxvq->stats.packets += nb_rx;
1569
1570         /* Allocate new mbuf for the used descriptor */
1571
1572         if (likely(!virtqueue_full(vq))) {
1573                 /* free_cnt may include mrg descs */
1574                 uint16_t free_cnt = vq->vq_free_cnt;
1575                 struct rte_mbuf *new_pkts[free_cnt];
1576
1577                 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1578                         error = virtqueue_enqueue_refill_inorder(vq, new_pkts,
1579                                         free_cnt);
1580                         if (unlikely(error)) {
1581                                 for (i = 0; i < free_cnt; i++)
1582                                         rte_pktmbuf_free(new_pkts[i]);
1583                         }
1584                         nb_enqueued += free_cnt;
1585                 } else {
1586                         struct rte_eth_dev *dev =
1587                                 &rte_eth_devices[rxvq->port_id];
1588                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1589                 }
1590         }
1591
1592         if (likely(nb_enqueued)) {
1593                 vq_update_avail_idx(vq);
1594
1595                 if (unlikely(virtqueue_kick_prepare(vq))) {
1596                         virtqueue_notify(vq);
1597                         PMD_RX_LOG(DEBUG, "Notified");
1598                 }
1599         }
1600
1601         return nb_rx;
1602 }
1603
1604 uint16_t
1605 virtio_recv_mergeable_pkts(void *rx_queue,
1606                         struct rte_mbuf **rx_pkts,
1607                         uint16_t nb_pkts)
1608 {
1609         struct virtnet_rx *rxvq = rx_queue;
1610         struct virtqueue *vq = rxvq->vq;
1611         struct virtio_hw *hw = vq->hw;
1612         struct rte_mbuf *rxm;
1613         struct rte_mbuf *prev = NULL;
1614         uint16_t nb_used, num, nb_rx = 0;
1615         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1616         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1617         int error;
1618         uint32_t nb_enqueued = 0;
1619         uint32_t seg_num = 0;
1620         uint32_t seg_res = 0;
1621         uint32_t hdr_size = hw->vtnet_hdr_size;
1622         int32_t i;
1623
1624         if (unlikely(hw->started == 0))
1625                 return nb_rx;
1626
1627         nb_used = VIRTQUEUE_NUSED(vq);
1628
1629         virtio_rmb(hw->weak_barriers);
1630
1631         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1632
1633         num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1634         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1635                 num = VIRTIO_MBUF_BURST_SZ;
1636         if (likely(num > DESC_PER_CACHELINE))
1637                 num = num - ((vq->vq_used_cons_idx + num) %
1638                                 DESC_PER_CACHELINE);
1639
1640
1641         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1642
1643         for (i = 0; i < num; i++) {
1644                 struct virtio_net_hdr_mrg_rxbuf *header;
1645
1646                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1647                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1648
1649                 rxm = rcv_pkts[i];
1650
1651                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1652                         PMD_RX_LOG(ERR, "Packet drop");
1653                         nb_enqueued++;
1654                         virtio_discard_rxbuf(vq, rxm);
1655                         rxvq->stats.errors++;
1656                         continue;
1657                 }
1658
1659                 header = (struct virtio_net_hdr_mrg_rxbuf *)
1660                          ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1661                          - hdr_size);
1662                 seg_num = header->num_buffers;
1663                 if (seg_num == 0)
1664                         seg_num = 1;
1665
1666                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1667                 rxm->nb_segs = seg_num;
1668                 rxm->ol_flags = 0;
1669                 rxm->vlan_tci = 0;
1670                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1671                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1672
1673                 rxm->port = rxvq->port_id;
1674
1675                 rx_pkts[nb_rx] = rxm;
1676                 prev = rxm;
1677
1678                 if (hw->has_rx_offload &&
1679                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1680                         virtio_discard_rxbuf(vq, rxm);
1681                         rxvq->stats.errors++;
1682                         continue;
1683                 }
1684
1685                 if (hw->vlan_strip)
1686                         rte_vlan_strip(rx_pkts[nb_rx]);
1687
1688                 seg_res = seg_num - 1;
1689
1690                 /* Merge remaining segments */
1691                 while (seg_res != 0 && i < (num - 1)) {
1692                         i++;
1693
1694                         rxm = rcv_pkts[i];
1695                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1696                         rxm->pkt_len = (uint32_t)(len[i]);
1697                         rxm->data_len = (uint16_t)(len[i]);
1698
1699                         rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1700
1701                         prev->next = rxm;
1702                         prev = rxm;
1703                         seg_res -= 1;
1704                 }
1705
1706                 if (!seg_res) {
1707                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1708                         nb_rx++;
1709                 }
1710         }
1711
1712         /* Last packet still need merge segments */
1713         while (seg_res != 0) {
1714                 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1715                                         VIRTIO_MBUF_BURST_SZ);
1716
1717                 if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
1718                         virtio_rmb(hw->weak_barriers);
1719                         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len,
1720                                                            rcv_cnt);
1721                         uint16_t extra_idx = 0;
1722
1723                         rcv_cnt = num;
1724                         while (extra_idx < rcv_cnt) {
1725                                 rxm = rcv_pkts[extra_idx];
1726                                 rxm->data_off =
1727                                         RTE_PKTMBUF_HEADROOM - hdr_size;
1728                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1729                                 rxm->data_len = (uint16_t)(len[extra_idx]);
1730                                 prev->next = rxm;
1731                                 prev = rxm;
1732                                 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1733                                 extra_idx += 1;
1734                         };
1735                         seg_res -= rcv_cnt;
1736
1737                         if (!seg_res) {
1738                                 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1739                                 nb_rx++;
1740                         }
1741                 } else {
1742                         PMD_RX_LOG(ERR,
1743                                         "No enough segments for packet.");
1744                         rte_pktmbuf_free(rx_pkts[nb_rx]);
1745                         rxvq->stats.errors++;
1746                         break;
1747                 }
1748         }
1749
1750         rxvq->stats.packets += nb_rx;
1751
1752         /* Allocate new mbuf for the used descriptor */
1753         if (likely(!virtqueue_full(vq))) {
1754                 /* free_cnt may include mrg descs */
1755                 uint16_t free_cnt = vq->vq_free_cnt;
1756                 struct rte_mbuf *new_pkts[free_cnt];
1757
1758                 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1759                         error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1760                                         free_cnt);
1761                         if (unlikely(error)) {
1762                                 for (i = 0; i < free_cnt; i++)
1763                                         rte_pktmbuf_free(new_pkts[i]);
1764                         }
1765                         nb_enqueued += free_cnt;
1766                 } else {
1767                         struct rte_eth_dev *dev =
1768                                 &rte_eth_devices[rxvq->port_id];
1769                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1770                 }
1771         }
1772
1773         if (likely(nb_enqueued)) {
1774                 vq_update_avail_idx(vq);
1775
1776                 if (unlikely(virtqueue_kick_prepare(vq))) {
1777                         virtqueue_notify(vq);
1778                         PMD_RX_LOG(DEBUG, "Notified");
1779                 }
1780         }
1781
1782         return nb_rx;
1783 }
1784
1785 uint16_t
1786 virtio_recv_mergeable_pkts_packed(void *rx_queue,
1787                         struct rte_mbuf **rx_pkts,
1788                         uint16_t nb_pkts)
1789 {
1790         struct virtnet_rx *rxvq = rx_queue;
1791         struct virtqueue *vq = rxvq->vq;
1792         struct virtio_hw *hw = vq->hw;
1793         struct rte_mbuf *rxm;
1794         struct rte_mbuf *prev = NULL;
1795         uint16_t num, nb_rx = 0;
1796         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1797         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1798         uint32_t nb_enqueued = 0;
1799         uint32_t seg_num = 0;
1800         uint32_t seg_res = 0;
1801         uint32_t hdr_size = hw->vtnet_hdr_size;
1802         int32_t i;
1803         int error;
1804
1805         if (unlikely(hw->started == 0))
1806                 return nb_rx;
1807
1808
1809         num = nb_pkts;
1810         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1811                 num = VIRTIO_MBUF_BURST_SZ;
1812         if (likely(num > DESC_PER_CACHELINE))
1813                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1814
1815         num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1816
1817         for (i = 0; i < num; i++) {
1818                 struct virtio_net_hdr_mrg_rxbuf *header;
1819
1820                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1821                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1822
1823                 rxm = rcv_pkts[i];
1824
1825                 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1826                         PMD_RX_LOG(ERR, "Packet drop");
1827                         nb_enqueued++;
1828                         virtio_discard_rxbuf(vq, rxm);
1829                         rxvq->stats.errors++;
1830                         continue;
1831                 }
1832
1833                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)
1834                           rxm->buf_addr + RTE_PKTMBUF_HEADROOM - hdr_size);
1835                 seg_num = header->num_buffers;
1836
1837                 if (seg_num == 0)
1838                         seg_num = 1;
1839
1840                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1841                 rxm->nb_segs = seg_num;
1842                 rxm->ol_flags = 0;
1843                 rxm->vlan_tci = 0;
1844                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1845                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1846
1847                 rxm->port = rxvq->port_id;
1848                 rx_pkts[nb_rx] = rxm;
1849                 prev = rxm;
1850
1851                 if (hw->has_rx_offload &&
1852                                 virtio_rx_offload(rxm, &header->hdr) < 0) {
1853                         virtio_discard_rxbuf(vq, rxm);
1854                         rxvq->stats.errors++;
1855                         continue;
1856                 }
1857
1858                 if (hw->vlan_strip)
1859                         rte_vlan_strip(rx_pkts[nb_rx]);
1860
1861                 seg_res = seg_num - 1;
1862
1863                 /* Merge remaining segments */
1864                 while (seg_res != 0 && i < (num - 1)) {
1865                         i++;
1866
1867                         rxm = rcv_pkts[i];
1868                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1869                         rxm->pkt_len = (uint32_t)(len[i]);
1870                         rxm->data_len = (uint16_t)(len[i]);
1871
1872                         rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1873
1874                         prev->next = rxm;
1875                         prev = rxm;
1876                         seg_res -= 1;
1877                 }
1878
1879                 if (!seg_res) {
1880                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1881                         nb_rx++;
1882                 }
1883         }
1884
1885         /* Last packet still need merge segments */
1886         while (seg_res != 0) {
1887                 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1888                                         VIRTIO_MBUF_BURST_SZ);
1889                 uint16_t extra_idx = 0;
1890
1891                 rcv_cnt = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts,
1892                                 len, rcv_cnt);
1893                 if (unlikely(rcv_cnt == 0)) {
1894                         PMD_RX_LOG(ERR, "No enough segments for packet.");
1895                         rte_pktmbuf_free(rx_pkts[nb_rx]);
1896                         rxvq->stats.errors++;
1897                         break;
1898                 }
1899
1900                 while (extra_idx < rcv_cnt) {
1901                         rxm = rcv_pkts[extra_idx];
1902
1903                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1904                         rxm->pkt_len = (uint32_t)(len[extra_idx]);
1905                         rxm->data_len = (uint16_t)(len[extra_idx]);
1906
1907                         prev->next = rxm;
1908                         prev = rxm;
1909                         rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1910                         extra_idx += 1;
1911                 }
1912                 seg_res -= rcv_cnt;
1913                 if (!seg_res) {
1914                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1915                         nb_rx++;
1916                 }
1917         }
1918
1919         rxvq->stats.packets += nb_rx;
1920
1921         /* Allocate new mbuf for the used descriptor */
1922         if (likely(!virtqueue_full(vq))) {
1923                 /* free_cnt may include mrg descs */
1924                 uint16_t free_cnt = vq->vq_free_cnt;
1925                 struct rte_mbuf *new_pkts[free_cnt];
1926
1927                 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1928                         error = virtqueue_enqueue_recv_refill_packed(vq,
1929                                         new_pkts, free_cnt);
1930                         if (unlikely(error)) {
1931                                 for (i = 0; i < free_cnt; i++)
1932                                         rte_pktmbuf_free(new_pkts[i]);
1933                         }
1934                         nb_enqueued += free_cnt;
1935                 } else {
1936                         struct rte_eth_dev *dev =
1937                                 &rte_eth_devices[rxvq->port_id];
1938                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1939                 }
1940         }
1941
1942         if (likely(nb_enqueued)) {
1943                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1944                         virtqueue_notify(vq);
1945                         PMD_RX_LOG(DEBUG, "Notified");
1946                 }
1947         }
1948
1949         return nb_rx;
1950 }
1951
1952 uint16_t
1953 virtio_xmit_pkts_prepare(void *tx_queue __rte_unused, struct rte_mbuf **tx_pkts,
1954                         uint16_t nb_pkts)
1955 {
1956         uint16_t nb_tx;
1957         int error;
1958
1959         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1960                 struct rte_mbuf *m = tx_pkts[nb_tx];
1961
1962 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
1963                 error = rte_validate_tx_offload(m);
1964                 if (unlikely(error)) {
1965                         rte_errno = -error;
1966                         break;
1967                 }
1968 #endif
1969
1970                 /* Do VLAN tag insertion */
1971                 if (unlikely(m->ol_flags & PKT_TX_VLAN_PKT)) {
1972                         error = rte_vlan_insert(&m);
1973                         /* rte_vlan_insert() may change pointer
1974                          * even in the case of failure
1975                          */
1976                         tx_pkts[nb_tx] = m;
1977
1978                         if (unlikely(error)) {
1979                                 rte_errno = -error;
1980                                 break;
1981                         }
1982                 }
1983
1984                 error = rte_net_intel_cksum_prepare(m);
1985                 if (unlikely(error)) {
1986                         rte_errno = -error;
1987                         break;
1988                 }
1989
1990                 if (m->ol_flags & PKT_TX_TCP_SEG)
1991                         virtio_tso_fix_cksum(m);
1992         }
1993
1994         return nb_tx;
1995 }
1996
1997 uint16_t
1998 virtio_xmit_pkts_packed(void *tx_queue, struct rte_mbuf **tx_pkts,
1999                         uint16_t nb_pkts)
2000 {
2001         struct virtnet_tx *txvq = tx_queue;
2002         struct virtqueue *vq = txvq->vq;
2003         struct virtio_hw *hw = vq->hw;
2004         uint16_t hdr_size = hw->vtnet_hdr_size;
2005         uint16_t nb_tx = 0;
2006         bool in_order = hw->use_inorder_tx;
2007
2008         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
2009                 return nb_tx;
2010
2011         if (unlikely(nb_pkts < 1))
2012                 return nb_pkts;
2013
2014         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
2015
2016         if (nb_pkts > vq->vq_free_cnt)
2017                 virtio_xmit_cleanup_packed(vq, nb_pkts - vq->vq_free_cnt,
2018                                            in_order);
2019
2020         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
2021                 struct rte_mbuf *txm = tx_pkts[nb_tx];
2022                 int can_push = 0, slots, need;
2023
2024                 /* optimize ring usage */
2025                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
2026                       vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
2027                     rte_mbuf_refcnt_read(txm) == 1 &&
2028                     RTE_MBUF_DIRECT(txm) &&
2029                     txm->nb_segs == 1 &&
2030                     rte_pktmbuf_headroom(txm) >= hdr_size &&
2031                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
2032                            __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
2033                         can_push = 1;
2034
2035                 /* How many main ring entries are needed to this Tx?
2036                  * any_layout => number of segments
2037                  * default    => number of segments + 1
2038                  */
2039                 slots = txm->nb_segs + !can_push;
2040                 need = slots - vq->vq_free_cnt;
2041
2042                 /* Positive value indicates it need free vring descriptors */
2043                 if (unlikely(need > 0)) {
2044                         virtio_xmit_cleanup_packed(vq, need, in_order);
2045                         need = slots - vq->vq_free_cnt;
2046                         if (unlikely(need > 0)) {
2047                                 PMD_TX_LOG(ERR,
2048                                            "No free tx descriptors to transmit");
2049                                 break;
2050                         }
2051                 }
2052
2053                 /* Enqueue Packet buffers */
2054                 if (can_push)
2055                         virtqueue_enqueue_xmit_packed_fast(txvq, txm, in_order);
2056                 else
2057                         virtqueue_enqueue_xmit_packed(txvq, txm, slots, 0,
2058                                                       in_order);
2059
2060                 virtio_update_packet_stats(&txvq->stats, txm);
2061         }
2062
2063         txvq->stats.packets += nb_tx;
2064
2065         if (likely(nb_tx)) {
2066                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
2067                         virtqueue_notify(vq);
2068                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2069                 }
2070         }
2071
2072         return nb_tx;
2073 }
2074
2075 uint16_t
2076 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
2077 {
2078         struct virtnet_tx *txvq = tx_queue;
2079         struct virtqueue *vq = txvq->vq;
2080         struct virtio_hw *hw = vq->hw;
2081         uint16_t hdr_size = hw->vtnet_hdr_size;
2082         uint16_t nb_used, nb_tx = 0;
2083
2084         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
2085                 return nb_tx;
2086
2087         if (unlikely(nb_pkts < 1))
2088                 return nb_pkts;
2089
2090         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
2091         nb_used = VIRTQUEUE_NUSED(vq);
2092
2093         virtio_rmb(hw->weak_barriers);
2094         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
2095                 virtio_xmit_cleanup(vq, nb_used);
2096
2097         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
2098                 struct rte_mbuf *txm = tx_pkts[nb_tx];
2099                 int can_push = 0, use_indirect = 0, slots, need;
2100
2101                 /* optimize ring usage */
2102                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
2103                       vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
2104                     rte_mbuf_refcnt_read(txm) == 1 &&
2105                     RTE_MBUF_DIRECT(txm) &&
2106                     txm->nb_segs == 1 &&
2107                     rte_pktmbuf_headroom(txm) >= hdr_size &&
2108                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
2109                                    __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
2110                         can_push = 1;
2111                 else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
2112                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
2113                         use_indirect = 1;
2114
2115                 /* How many main ring entries are needed to this Tx?
2116                  * any_layout => number of segments
2117                  * indirect   => 1
2118                  * default    => number of segments + 1
2119                  */
2120                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
2121                 need = slots - vq->vq_free_cnt;
2122
2123                 /* Positive value indicates it need free vring descriptors */
2124                 if (unlikely(need > 0)) {
2125                         nb_used = VIRTQUEUE_NUSED(vq);
2126                         virtio_rmb(hw->weak_barriers);
2127                         need = RTE_MIN(need, (int)nb_used);
2128
2129                         virtio_xmit_cleanup(vq, need);
2130                         need = slots - vq->vq_free_cnt;
2131                         if (unlikely(need > 0)) {
2132                                 PMD_TX_LOG(ERR,
2133                                            "No free tx descriptors to transmit");
2134                                 break;
2135                         }
2136                 }
2137
2138                 /* Enqueue Packet buffers */
2139                 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
2140                         can_push, 0);
2141
2142                 virtio_update_packet_stats(&txvq->stats, txm);
2143         }
2144
2145         txvq->stats.packets += nb_tx;
2146
2147         if (likely(nb_tx)) {
2148                 vq_update_avail_idx(vq);
2149
2150                 if (unlikely(virtqueue_kick_prepare(vq))) {
2151                         virtqueue_notify(vq);
2152                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2153                 }
2154         }
2155
2156         return nb_tx;
2157 }
2158
2159 static __rte_always_inline int
2160 virtio_xmit_try_cleanup_inorder(struct virtqueue *vq, uint16_t need)
2161 {
2162         uint16_t nb_used, nb_clean, nb_descs;
2163         struct virtio_hw *hw = vq->hw;
2164
2165         nb_descs = vq->vq_free_cnt + need;
2166         nb_used = VIRTQUEUE_NUSED(vq);
2167         virtio_rmb(hw->weak_barriers);
2168         nb_clean = RTE_MIN(need, (int)nb_used);
2169
2170         virtio_xmit_cleanup_inorder(vq, nb_clean);
2171
2172         return nb_descs - vq->vq_free_cnt;
2173 }
2174
2175 uint16_t
2176 virtio_xmit_pkts_inorder(void *tx_queue,
2177                         struct rte_mbuf **tx_pkts,
2178                         uint16_t nb_pkts)
2179 {
2180         struct virtnet_tx *txvq = tx_queue;
2181         struct virtqueue *vq = txvq->vq;
2182         struct virtio_hw *hw = vq->hw;
2183         uint16_t hdr_size = hw->vtnet_hdr_size;
2184         uint16_t nb_used, nb_tx = 0, nb_inorder_pkts = 0;
2185         struct rte_mbuf *inorder_pkts[nb_pkts];
2186         int need;
2187
2188         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
2189                 return nb_tx;
2190
2191         if (unlikely(nb_pkts < 1))
2192                 return nb_pkts;
2193
2194         VIRTQUEUE_DUMP(vq);
2195         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
2196         nb_used = VIRTQUEUE_NUSED(vq);
2197
2198         virtio_rmb(hw->weak_barriers);
2199         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
2200                 virtio_xmit_cleanup_inorder(vq, nb_used);
2201
2202         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
2203                 struct rte_mbuf *txm = tx_pkts[nb_tx];
2204                 int slots;
2205
2206                 /* optimize ring usage */
2207                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
2208                      vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
2209                      rte_mbuf_refcnt_read(txm) == 1 &&
2210                      RTE_MBUF_DIRECT(txm) &&
2211                      txm->nb_segs == 1 &&
2212                      rte_pktmbuf_headroom(txm) >= hdr_size &&
2213                      rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
2214                                 __alignof__(struct virtio_net_hdr_mrg_rxbuf))) {
2215                         inorder_pkts[nb_inorder_pkts] = txm;
2216                         nb_inorder_pkts++;
2217
2218                         continue;
2219                 }
2220
2221                 if (nb_inorder_pkts) {
2222                         need = nb_inorder_pkts - vq->vq_free_cnt;
2223                         if (unlikely(need > 0)) {
2224                                 need = virtio_xmit_try_cleanup_inorder(vq,
2225                                                                        need);
2226                                 if (unlikely(need > 0)) {
2227                                         PMD_TX_LOG(ERR,
2228                                                 "No free tx descriptors to "
2229                                                 "transmit");
2230                                         break;
2231                                 }
2232                         }
2233                         virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2234                                                         nb_inorder_pkts);
2235                         nb_inorder_pkts = 0;
2236                 }
2237
2238                 slots = txm->nb_segs + 1;
2239                 need = slots - vq->vq_free_cnt;
2240                 if (unlikely(need > 0)) {
2241                         need = virtio_xmit_try_cleanup_inorder(vq, slots);
2242
2243                         if (unlikely(need > 0)) {
2244                                 PMD_TX_LOG(ERR,
2245                                         "No free tx descriptors to transmit");
2246                                 break;
2247                         }
2248                 }
2249                 /* Enqueue Packet buffers */
2250                 virtqueue_enqueue_xmit(txvq, txm, slots, 0, 0, 1);
2251
2252                 virtio_update_packet_stats(&txvq->stats, txm);
2253         }
2254
2255         /* Transmit all inorder packets */
2256         if (nb_inorder_pkts) {
2257                 need = nb_inorder_pkts - vq->vq_free_cnt;
2258                 if (unlikely(need > 0)) {
2259                         need = virtio_xmit_try_cleanup_inorder(vq,
2260                                                                   need);
2261                         if (unlikely(need > 0)) {
2262                                 PMD_TX_LOG(ERR,
2263                                         "No free tx descriptors to transmit");
2264                                 nb_inorder_pkts = vq->vq_free_cnt;
2265                                 nb_tx -= need;
2266                         }
2267                 }
2268
2269                 virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2270                                                 nb_inorder_pkts);
2271         }
2272
2273         txvq->stats.packets += nb_tx;
2274
2275         if (likely(nb_tx)) {
2276                 vq_update_avail_idx(vq);
2277
2278                 if (unlikely(virtqueue_kick_prepare(vq))) {
2279                         virtqueue_notify(vq);
2280                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2281                 }
2282         }
2283
2284         VIRTQUEUE_DUMP(vq);
2285
2286         return nb_tx;
2287 }