lib/librte_vhost/virtio_net.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2016 Intel Corporation
   3  */
   4
   5 #include <stdint.h>
   6 #include <stdbool.h>
   7 #include <linux/virtio_net.h>
   8
   9 #include <rte_mbuf.h>
  10 #include <rte_memcpy.h>
  11 #include <rte_ether.h>
  12 #include <rte_ip.h>
  13 #include <rte_vhost.h>
  14 #include <rte_tcp.h>
  15 #include <rte_udp.h>
  16 #include <rte_sctp.h>
  17 #include <rte_arp.h>
  18 #include <rte_spinlock.h>
  19 #include <rte_malloc.h>
  20
  21 #include "iotlb.h"
  22 #include "vhost.h"
  23
  24 #define MAX_PKT_BURST 32
  25
  26 #define MAX_BATCH_LEN 256
  27
  28 static  __rte_always_inline bool
  29 rxvq_is_mergeable(struct virtio_net *dev)
  30 {
  31         return dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF);
  32 }
  33
  34 static bool
  35 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
  36 {
  37         return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
  38 }
  39
  40 static __rte_always_inline void
  41 do_flush_shadow_used_ring_split(struct virtio_net *dev,
  42                         struct vhost_virtqueue *vq,
  43                         uint16_t to, uint16_t from, uint16_t size)
  44 {
  45         rte_memcpy(&vq->used->ring[to],
  46                         &vq->shadow_used_split[from],
  47                         size * sizeof(struct vring_used_elem));
  48         vhost_log_cache_used_vring(dev, vq,
  49                         offsetof(struct vring_used, ring[to]),
  50                         size * sizeof(struct vring_used_elem));
  51 }
  52
  53 static __rte_always_inline void
  54 flush_shadow_used_ring_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
  55 {
  56         uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
  57
  58         if (used_idx + vq->shadow_used_idx <= vq->size) {
  59                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0,
  60                                           vq->shadow_used_idx);
  61         } else {
  62                 uint16_t size;
  63
  64                 /* update used ring interval [used_idx, vq->size] */
  65                 size = vq->size - used_idx;
  66                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0, size);
  67
  68                 /* update the left half used ring interval [0, left_size] */
  69                 do_flush_shadow_used_ring_split(dev, vq, 0, size,
  70                                           vq->shadow_used_idx - size);
  71         }
  72         vq->last_used_idx += vq->shadow_used_idx;
  73
  74         rte_smp_wmb();
  75
  76         vhost_log_cache_sync(dev, vq);
  77
  78         *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
  79         vq->shadow_used_idx = 0;
  80         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
  81                 sizeof(vq->used->idx));
  82 }
  83
  84 static __rte_always_inline void
  85 update_shadow_used_ring_split(struct vhost_virtqueue *vq,
  86                          uint16_t desc_idx, uint32_t len)
  87 {
  88         uint16_t i = vq->shadow_used_idx++;
  89
  90         vq->shadow_used_split[i].id  = desc_idx;
  91         vq->shadow_used_split[i].len = len;
  92 }
  93
  94 static __rte_always_inline void
  95 flush_shadow_used_ring_packed(struct virtio_net *dev,
  96                         struct vhost_virtqueue *vq)
  97 {
  98         int i;
  99         uint16_t used_idx = vq->last_used_idx;
 100         uint16_t head_idx = vq->last_used_idx;
 101         uint16_t head_flags = 0;
 102
 103         /* Split loop in two to save memory barriers */
 104         for (i = 0; i < vq->shadow_used_idx; i++) {
 105                 vq->desc_packed[used_idx].id = vq->shadow_used_packed[i].id;
 106                 vq->desc_packed[used_idx].len = vq->shadow_used_packed[i].len;
 107
 108                 used_idx += vq->shadow_used_packed[i].count;
 109                 if (used_idx >= vq->size)
 110                         used_idx -= vq->size;
 111         }
 112
 113         rte_smp_wmb();
 114
 115         for (i = 0; i < vq->shadow_used_idx; i++) {
 116                 uint16_t flags;
 117
 118                 if (vq->shadow_used_packed[i].len)
 119                         flags = VRING_DESC_F_WRITE;
 120                 else
 121                         flags = 0;
 122
 123                 if (vq->used_wrap_counter) {
 124                         flags |= VRING_DESC_F_USED;
 125                         flags |= VRING_DESC_F_AVAIL;
 126                 } else {
 127                         flags &= ~VRING_DESC_F_USED;
 128                         flags &= ~VRING_DESC_F_AVAIL;
 129                 }
 130
 131                 if (i > 0) {
 132                         vq->desc_packed[vq->last_used_idx].flags = flags;
 133
 134                         vhost_log_cache_used_vring(dev, vq,
 135                                         vq->last_used_idx *
 136                                         sizeof(struct vring_packed_desc),
 137                                         sizeof(struct vring_packed_desc));
 138                 } else {
 139                         head_idx = vq->last_used_idx;
 140                         head_flags = flags;
 141                 }
 142
 143                 vq->last_used_idx += vq->shadow_used_packed[i].count;
 144                 if (vq->last_used_idx >= vq->size) {
 145                         vq->used_wrap_counter ^= 1;
 146                         vq->last_used_idx -= vq->size;
 147                 }
 148         }
 149
 150         vq->desc_packed[head_idx].flags = head_flags;
 151
 152         vhost_log_cache_used_vring(dev, vq,
 153                                 head_idx *
 154                                 sizeof(struct vring_packed_desc),
 155                                 sizeof(struct vring_packed_desc));
 156
 157         vq->shadow_used_idx = 0;
 158         vhost_log_cache_sync(dev, vq);
 159 }
 160
 161 static __rte_always_inline void
 162 update_shadow_used_ring_packed(struct vhost_virtqueue *vq,
 163                          uint16_t desc_idx, uint32_t len, uint16_t count)
 164 {
 165         uint16_t i = vq->shadow_used_idx++;
 166
 167         vq->shadow_used_packed[i].id  = desc_idx;
 168         vq->shadow_used_packed[i].len = len;
 169         vq->shadow_used_packed[i].count = count;
 170 }
 171
 172 static inline void
 173 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
 174 {
 175         struct batch_copy_elem *elem = vq->batch_copy_elems;
 176         uint16_t count = vq->batch_copy_nb_elems;
 177         int i;
 178
 179         for (i = 0; i < count; i++) {
 180                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 181                 vhost_log_cache_write(dev, vq, elem[i].log_addr, elem[i].len);
 182                 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
 183         }
 184
 185         vq->batch_copy_nb_elems = 0;
 186 }
 187
 188 static inline void
 189 do_data_copy_dequeue(struct vhost_virtqueue *vq)
 190 {
 191         struct batch_copy_elem *elem = vq->batch_copy_elems;
 192         uint16_t count = vq->batch_copy_nb_elems;
 193         int i;
 194
 195         for (i = 0; i < count; i++)
 196                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 197
 198         vq->batch_copy_nb_elems = 0;
 199 }
 200
 201 /* avoid write operation when necessary, to lessen cache issues */
 202 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 203         if ((var) != (val))                     \
 204                 (var) = (val);                  \
 205 } while (0)
 206
 207 static __rte_always_inline void
 208 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 209 {
 210         uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
 211
 212         if (m_buf->ol_flags & PKT_TX_TCP_SEG)
 213                 csum_l4 |= PKT_TX_TCP_CKSUM;
 214
 215         if (csum_l4) {
 216                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 217                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 218
 219                 switch (csum_l4) {
 220                 case PKT_TX_TCP_CKSUM:
 221                         net_hdr->csum_offset = (offsetof(struct tcp_hdr,
 222                                                 cksum));
 223                         break;
 224                 case PKT_TX_UDP_CKSUM:
 225                         net_hdr->csum_offset = (offsetof(struct udp_hdr,
 226                                                 dgram_cksum));
 227                         break;
 228                 case PKT_TX_SCTP_CKSUM:
 229                         net_hdr->csum_offset = (offsetof(struct sctp_hdr,
 230                                                 cksum));
 231                         break;
 232                 }
 233         } else {
 234                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 235                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 236                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 237         }
 238
 239         /* IP cksum verification cannot be bypassed, then calculate here */
 240         if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
 241                 struct ipv4_hdr *ipv4_hdr;
 242
 243                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct ipv4_hdr *,
 244                                                    m_buf->l2_len);
 245                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 246         }
 247
 248         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 249                 if (m_buf->ol_flags & PKT_TX_IPV4)
 250                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 251                 else
 252                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 253                 net_hdr->gso_size = m_buf->tso_segsz;
 254                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 255                                         + m_buf->l4_len;
 256         } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
 257                 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 258                 net_hdr->gso_size = m_buf->tso_segsz;
 259                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
 260                         m_buf->l4_len;
 261         } else {
 262                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 263                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 264                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 265         }
 266 }
 267
 268 static __rte_always_inline int
 269 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 270                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 271                 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
 272 {
 273         uint16_t vec_id = *vec_idx;
 274
 275         while (desc_len) {
 276                 uint64_t desc_addr;
 277                 uint64_t desc_chunck_len = desc_len;
 278
 279                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 280                         return -1;
 281
 282                 desc_addr = vhost_iova_to_vva(dev, vq,
 283                                 desc_iova,
 284                                 &desc_chunck_len,
 285                                 perm);
 286                 if (unlikely(!desc_addr))
 287                         return -1;
 288
 289                 buf_vec[vec_id].buf_iova = desc_iova;
 290                 buf_vec[vec_id].buf_addr = desc_addr;
 291                 buf_vec[vec_id].buf_len  = desc_chunck_len;
 292
 293                 desc_len -= desc_chunck_len;
 294                 desc_iova += desc_chunck_len;
 295                 vec_id++;
 296         }
 297         *vec_idx = vec_id;
 298
 299         return 0;
 300 }
 301
 302 static __rte_always_inline int
 303 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 304                          uint32_t avail_idx, uint16_t *vec_idx,
 305                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 306                          uint32_t *desc_chain_len, uint8_t perm)
 307 {
 308         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 309         uint16_t vec_id = *vec_idx;
 310         uint32_t len    = 0;
 311         uint64_t dlen;
 312         uint32_t nr_descs = vq->size;
 313         uint32_t cnt    = 0;
 314         struct vring_desc *descs = vq->desc;
 315         struct vring_desc *idesc = NULL;
 316
 317         if (unlikely(idx >= vq->size))
 318                 return -1;
 319
 320         *desc_chain_head = idx;
 321
 322         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 323                 dlen = vq->desc[idx].len;
 324                 nr_descs = dlen / sizeof(struct vring_desc);
 325                 if (unlikely(nr_descs > vq->size))
 326                         return -1;
 327
 328                 descs = (struct vring_desc *)(uintptr_t)
 329                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 330                                                 &dlen,
 331                                                 VHOST_ACCESS_RO);
 332                 if (unlikely(!descs))
 333                         return -1;
 334
 335                 if (unlikely(dlen < vq->desc[idx].len)) {
 336                         /*
 337                          * The indirect desc table is not contiguous
 338                          * in process VA space, we have to copy it.
 339                          */
 340                         idesc = alloc_copy_ind_table(dev, vq,
 341                                         vq->desc[idx].addr, vq->desc[idx].len);
 342                         if (unlikely(!idesc))
 343                                 return -1;
 344
 345                         descs = idesc;
 346                 }
 347
 348                 idx = 0;
 349         }
 350
 351         while (1) {
 352                 if (unlikely(idx >= nr_descs || cnt++ >= nr_descs)) {
 353                         free_ind_table(idesc);
 354                         return -1;
 355                 }
 356
 357                 len += descs[idx].len;
 358
 359                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 360                                                 descs[idx].addr, descs[idx].len,
 361                                                 perm))) {
 362                         free_ind_table(idesc);
 363                         return -1;
 364                 }
 365
 366                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 367                         break;
 368
 369                 idx = descs[idx].next;
 370         }
 371
 372         *desc_chain_len = len;
 373         *vec_idx = vec_id;
 374
 375         if (unlikely(!!idesc))
 376                 free_ind_table(idesc);
 377
 378         return 0;
 379 }
 380
 381 /*
 382  * Returns -1 on fail, 0 on success
 383  */
 384 static inline int
 385 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 386                                 uint32_t size, struct buf_vector *buf_vec,
 387                                 uint16_t *num_buffers, uint16_t avail_head,
 388                                 uint16_t *nr_vec)
 389 {
 390         uint16_t cur_idx;
 391         uint16_t vec_idx = 0;
 392         uint16_t max_tries, tries = 0;
 393
 394         uint16_t head_idx = 0;
 395         uint32_t len = 0;
 396
 397         *num_buffers = 0;
 398         cur_idx  = vq->last_avail_idx;
 399
 400         if (rxvq_is_mergeable(dev))
 401                 max_tries = vq->size - 1;
 402         else
 403                 max_tries = 1;
 404
 405         while (size > 0) {
 406                 if (unlikely(cur_idx == avail_head))
 407                         return -1;
 408                 /*
 409                  * if we tried all available ring items, and still
 410                  * can't get enough buf, it means something abnormal
 411                  * happened.
 412                  */
 413                 if (unlikely(++tries > max_tries))
 414                         return -1;
 415
 416                 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
 417                                                 &vec_idx, buf_vec,
 418                                                 &head_idx, &len,
 419                                                 VHOST_ACCESS_RW) < 0))
 420                         return -1;
 421                 len = RTE_MIN(len, size);
 422                 update_shadow_used_ring_split(vq, head_idx, len);
 423                 size -= len;
 424
 425                 cur_idx++;
 426                 *num_buffers += 1;
 427         }
 428
 429         *nr_vec = vec_idx;
 430
 431         return 0;
 432 }
 433
 434 static __rte_always_inline int
 435 fill_vec_buf_packed_indirect(struct virtio_net *dev,
 436                         struct vhost_virtqueue *vq,
 437                         struct vring_packed_desc *desc, uint16_t *vec_idx,
 438                         struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
 439 {
 440         uint16_t i;
 441         uint32_t nr_descs;
 442         uint16_t vec_id = *vec_idx;
 443         uint64_t dlen;
 444         struct vring_packed_desc *descs, *idescs = NULL;
 445
 446         dlen = desc->len;
 447         descs = (struct vring_packed_desc *)(uintptr_t)
 448                 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
 449         if (unlikely(!descs))
 450                 return -1;
 451
 452         if (unlikely(dlen < desc->len)) {
 453                 /*
 454                  * The indirect desc table is not contiguous
 455                  * in process VA space, we have to copy it.
 456                  */
 457                 idescs = alloc_copy_ind_table(dev, vq, desc->addr, desc->len);
 458                 if (unlikely(!idescs))
 459                         return -1;
 460
 461                 descs = idescs;
 462         }
 463
 464         nr_descs =  desc->len / sizeof(struct vring_packed_desc);
 465         if (unlikely(nr_descs >= vq->size)) {
 466                 free_ind_table(idescs);
 467                 return -1;
 468         }
 469
 470         for (i = 0; i < nr_descs; i++) {
 471                 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
 472                         free_ind_table(idescs);
 473                         return -1;
 474                 }
 475
 476                 *len += descs[i].len;
 477                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 478                                                 descs[i].addr, descs[i].len,
 479                                                 perm)))
 480                         return -1;
 481         }
 482         *vec_idx = vec_id;
 483
 484         if (unlikely(!!idescs))
 485                 free_ind_table(idescs);
 486
 487         return 0;
 488 }
 489
 490 static __rte_always_inline int
 491 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 492                                 uint16_t avail_idx, uint16_t *desc_count,
 493                                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 494                                 uint16_t *buf_id, uint32_t *len, uint8_t perm)
 495 {
 496         bool wrap_counter = vq->avail_wrap_counter;
 497         struct vring_packed_desc *descs = vq->desc_packed;
 498         uint16_t vec_id = *vec_idx;
 499
 500         if (avail_idx < vq->last_avail_idx)
 501                 wrap_counter ^= 1;
 502
 503         if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
 504                 return -1;
 505
 506         /*
 507          * The ordering between desc flags and desc
 508          * content reads need to be enforced.
 509          */
 510         rte_smp_rmb();
 511
 512         *desc_count = 0;
 513         *len = 0;
 514
 515         while (1) {
 516                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 517                         return -1;
 518
 519                 if (unlikely(*desc_count >= vq->size))
 520                         return -1;
 521
 522                 *desc_count += 1;
 523                 *buf_id = descs[avail_idx].id;
 524
 525                 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
 526                         if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
 527                                                         &descs[avail_idx],
 528                                                         &vec_id, buf_vec,
 529                                                         len, perm) < 0))
 530                                 return -1;
 531                 } else {
 532                         *len += descs[avail_idx].len;
 533
 534                         if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 535                                                         descs[avail_idx].addr,
 536                                                         descs[avail_idx].len,
 537                                                         perm)))
 538                                 return -1;
 539                 }
 540
 541                 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
 542                         break;
 543
 544                 if (++avail_idx >= vq->size) {
 545                         avail_idx -= vq->size;
 546                         wrap_counter ^= 1;
 547                 }
 548         }
 549
 550         *vec_idx = vec_id;
 551
 552         return 0;
 553 }
 554
 555 /*
 556  * Returns -1 on fail, 0 on success
 557  */
 558 static inline int
 559 reserve_avail_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 560                                 uint32_t size, struct buf_vector *buf_vec,
 561                                 uint16_t *nr_vec, uint16_t *num_buffers,
 562                                 uint16_t *nr_descs)
 563 {
 564         uint16_t avail_idx;
 565         uint16_t vec_idx = 0;
 566         uint16_t max_tries, tries = 0;
 567
 568         uint16_t buf_id = 0;
 569         uint32_t len = 0;
 570         uint16_t desc_count;
 571
 572         *num_buffers = 0;
 573         avail_idx = vq->last_avail_idx;
 574
 575         if (rxvq_is_mergeable(dev))
 576                 max_tries = vq->size - 1;
 577         else
 578                 max_tries = 1;
 579
 580         while (size > 0) {
 581                 /*
 582                  * if we tried all available ring items, and still
 583                  * can't get enough buf, it means something abnormal
 584                  * happened.
 585                  */
 586                 if (unlikely(++tries > max_tries))
 587                         return -1;
 588
 589                 if (unlikely(fill_vec_buf_packed(dev, vq,
 590                                                 avail_idx, &desc_count,
 591                                                 buf_vec, &vec_idx,
 592                                                 &buf_id, &len,
 593                                                 VHOST_ACCESS_RW) < 0))
 594                         return -1;
 595
 596                 len = RTE_MIN(len, size);
 597                 update_shadow_used_ring_packed(vq, buf_id, len, desc_count);
 598                 size -= len;
 599
 600                 avail_idx += desc_count;
 601                 if (avail_idx >= vq->size)
 602                         avail_idx -= vq->size;
 603
 604                 *nr_descs += desc_count;
 605                 *num_buffers += 1;
 606         }
 607
 608         *nr_vec = vec_idx;
 609
 610         return 0;
 611 }
 612
 613 static __rte_always_inline int
 614 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 615                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 616                             uint16_t nr_vec, uint16_t num_buffers)
 617 {
 618         uint32_t vec_idx = 0;
 619         uint32_t mbuf_offset, mbuf_avail;
 620         uint32_t buf_offset, buf_avail;
 621         uint64_t buf_addr, buf_iova, buf_len;
 622         uint32_t cpy_len;
 623         uint64_t hdr_addr;
 624         struct rte_mbuf *hdr_mbuf;
 625         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 626         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 627         int error = 0;
 628
 629         if (unlikely(m == NULL)) {
 630                 error = -1;
 631                 goto out;
 632         }
 633
 634         buf_addr = buf_vec[vec_idx].buf_addr;
 635         buf_iova = buf_vec[vec_idx].buf_iova;
 636         buf_len = buf_vec[vec_idx].buf_len;
 637
 638         if (nr_vec > 1)
 639                 rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
 640
 641         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 642                 error = -1;
 643                 goto out;
 644         }
 645
 646         hdr_mbuf = m;
 647         hdr_addr = buf_addr;
 648         if (unlikely(buf_len < dev->vhost_hlen))
 649                 hdr = &tmp_hdr;
 650         else
 651                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
 652
 653         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 654                 dev->vid, num_buffers);
 655
 656         if (unlikely(buf_len < dev->vhost_hlen)) {
 657                 buf_offset = dev->vhost_hlen - buf_len;
 658                 vec_idx++;
 659                 buf_addr = buf_vec[vec_idx].buf_addr;
 660                 buf_iova = buf_vec[vec_idx].buf_iova;
 661                 buf_len = buf_vec[vec_idx].buf_len;
 662                 buf_avail = buf_len - buf_offset;
 663         } else {
 664                 buf_offset = dev->vhost_hlen;
 665                 buf_avail = buf_len - dev->vhost_hlen;
 666         }
 667
 668         mbuf_avail  = rte_pktmbuf_data_len(m);
 669         mbuf_offset = 0;
 670         while (mbuf_avail != 0 || m->next != NULL) {
 671                 /* done with current buf, get the next one */
 672                 if (buf_avail == 0) {
 673                         vec_idx++;
 674                         if (unlikely(vec_idx >= nr_vec)) {
 675                                 error = -1;
 676                                 goto out;
 677                         }
 678
 679                         buf_addr = buf_vec[vec_idx].buf_addr;
 680                         buf_iova = buf_vec[vec_idx].buf_iova;
 681                         buf_len = buf_vec[vec_idx].buf_len;
 682
 683                         /* Prefetch next buffer address. */
 684                         if (vec_idx + 1 < nr_vec)
 685                                 rte_prefetch0((void *)(uintptr_t)
 686                                                 buf_vec[vec_idx + 1].buf_addr);
 687                         buf_offset = 0;
 688                         buf_avail  = buf_len;
 689                 }
 690
 691                 /* done with current mbuf, get the next one */
 692                 if (mbuf_avail == 0) {
 693                         m = m->next;
 694
 695                         mbuf_offset = 0;
 696                         mbuf_avail  = rte_pktmbuf_data_len(m);
 697                 }
 698
 699                 if (hdr_addr) {
 700                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 701                         if (rxvq_is_mergeable(dev))
 702                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
 703                                                 num_buffers);
 704
 705                         if (unlikely(hdr == &tmp_hdr)) {
 706                                 uint64_t len;
 707                                 uint64_t remain = dev->vhost_hlen;
 708                                 uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 709                                 uint64_t iova = buf_vec[0].buf_iova;
 710                                 uint16_t hdr_vec_idx = 0;
 711
 712                                 while (remain) {
 713                                         len = RTE_MIN(remain,
 714                                                 buf_vec[hdr_vec_idx].buf_len);
 715                                         dst = buf_vec[hdr_vec_idx].buf_addr;
 716                                         rte_memcpy((void *)(uintptr_t)dst,
 717                                                         (void *)(uintptr_t)src,
 718                                                         len);
 719
 720                                         PRINT_PACKET(dev, (uintptr_t)dst,
 721                                                         (uint32_t)len, 0);
 722                                         vhost_log_cache_write(dev, vq,
 723                                                         iova, len);
 724
 725                                         remain -= len;
 726                                         iova += len;
 727                                         src += len;
 728                                         hdr_vec_idx++;
 729                                 }
 730                         } else {
 731                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 732                                                 dev->vhost_hlen, 0);
 733                                 vhost_log_cache_write(dev, vq,
 734                                                 buf_vec[0].buf_iova,
 735                                                 dev->vhost_hlen);
 736                         }
 737
 738                         hdr_addr = 0;
 739                 }
 740
 741                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 742
 743                 if (likely(cpy_len > MAX_BATCH_LEN ||
 744                                         vq->batch_copy_nb_elems >= vq->size)) {
 745                         rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
 746                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 747                                 cpy_len);
 748                         vhost_log_cache_write(dev, vq, buf_iova + buf_offset,
 749                                         cpy_len);
 750                         PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
 751                                 cpy_len, 0);
 752                 } else {
 753                         batch_copy[vq->batch_copy_nb_elems].dst =
 754                                 (void *)((uintptr_t)(buf_addr + buf_offset));
 755                         batch_copy[vq->batch_copy_nb_elems].src =
 756                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 757                         batch_copy[vq->batch_copy_nb_elems].log_addr =
 758                                 buf_iova + buf_offset;
 759                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
 760                         vq->batch_copy_nb_elems++;
 761                 }
 762
 763                 mbuf_avail  -= cpy_len;
 764                 mbuf_offset += cpy_len;
 765                 buf_avail  -= cpy_len;
 766                 buf_offset += cpy_len;
 767         }
 768
 769 out:
 770
 771         return error;
 772 }
 773
 774 static __rte_always_inline uint32_t
 775 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 776         struct rte_mbuf **pkts, uint32_t count)
 777 {
 778         uint32_t pkt_idx = 0;
 779         uint16_t num_buffers;
 780         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 781         uint16_t avail_head;
 782
 783         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 784
 785         /*
 786          * The ordering between avail index and
 787          * desc reads needs to be enforced.
 788          */
 789         rte_smp_rmb();
 790
 791         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 792
 793         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 794                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 795                 uint16_t nr_vec = 0;
 796
 797                 if (unlikely(reserve_avail_buf_split(dev, vq,
 798                                                 pkt_len, buf_vec, &num_buffers,
 799                                                 avail_head, &nr_vec) < 0)) {
 800                         VHOST_LOG_DEBUG(VHOST_DATA,
 801                                 "(%d) failed to get enough desc from vring\n",
 802                                 dev->vid);
 803                         vq->shadow_used_idx -= num_buffers;
 804                         break;
 805                 }
 806
 807                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
 808
 809                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 810                         dev->vid, vq->last_avail_idx,
 811                         vq->last_avail_idx + num_buffers);
 812
 813                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
 814                                                 buf_vec, nr_vec,
 815                                                 num_buffers) < 0) {
 816                         vq->shadow_used_idx -= num_buffers;
 817                         break;
 818                 }
 819
 820                 vq->last_avail_idx += num_buffers;
 821         }
 822
 823         do_data_copy_enqueue(dev, vq);
 824
 825         if (likely(vq->shadow_used_idx)) {
 826                 flush_shadow_used_ring_split(dev, vq);
 827                 vhost_vring_call_split(dev, vq);
 828         }
 829
 830         return pkt_idx;
 831 }
 832
 833 static __rte_always_inline uint32_t
 834 virtio_dev_rx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 835         struct rte_mbuf **pkts, uint32_t count)
 836 {
 837         uint32_t pkt_idx = 0;
 838         uint16_t num_buffers;
 839         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 840
 841         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 842                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 843                 uint16_t nr_vec = 0;
 844                 uint16_t nr_descs = 0;
 845
 846                 if (unlikely(reserve_avail_buf_packed(dev, vq,
 847                                                 pkt_len, buf_vec, &nr_vec,
 848                                                 &num_buffers, &nr_descs) < 0)) {
 849                         VHOST_LOG_DEBUG(VHOST_DATA,
 850                                 "(%d) failed to get enough desc from vring\n",
 851                                 dev->vid);
 852                         vq->shadow_used_idx -= num_buffers;
 853                         break;
 854                 }
 855
 856                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
 857
 858                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 859                         dev->vid, vq->last_avail_idx,
 860                         vq->last_avail_idx + num_buffers);
 861
 862                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
 863                                                 buf_vec, nr_vec,
 864                                                 num_buffers) < 0) {
 865                         vq->shadow_used_idx -= num_buffers;
 866                         break;
 867                 }
 868
 869                 vq->last_avail_idx += nr_descs;
 870                 if (vq->last_avail_idx >= vq->size) {
 871                         vq->last_avail_idx -= vq->size;
 872                         vq->avail_wrap_counter ^= 1;
 873                 }
 874         }
 875
 876         do_data_copy_enqueue(dev, vq);
 877
 878         if (likely(vq->shadow_used_idx)) {
 879                 flush_shadow_used_ring_packed(dev, vq);
 880                 vhost_vring_call_packed(dev, vq);
 881         }
 882
 883         return pkt_idx;
 884 }
 885
 886 static __rte_always_inline uint32_t
 887 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 888         struct rte_mbuf **pkts, uint32_t count)
 889 {
 890         struct vhost_virtqueue *vq;
 891         uint32_t nb_tx = 0;
 892
 893         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 894         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 895                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 896                         dev->vid, __func__, queue_id);
 897                 return 0;
 898         }
 899
 900         vq = dev->virtqueue[queue_id];
 901
 902         rte_spinlock_lock(&vq->access_lock);
 903
 904         if (unlikely(vq->enabled == 0))
 905                 goto out_access_unlock;
 906
 907         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 908                 vhost_user_iotlb_rd_lock(vq);
 909
 910         if (unlikely(vq->access_ok == 0))
 911                 if (unlikely(vring_translate(dev, vq) < 0))
 912                         goto out;
 913
 914         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 915         if (count == 0)
 916                 goto out;
 917
 918         if (vq_is_packed(dev))
 919                 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
 920         else
 921                 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
 922
 923 out:
 924         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 925                 vhost_user_iotlb_rd_unlock(vq);
 926
 927 out_access_unlock:
 928         rte_spinlock_unlock(&vq->access_lock);
 929
 930         return nb_tx;
 931 }
 932
 933 uint16_t
 934 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 935         struct rte_mbuf **pkts, uint16_t count)
 936 {
 937         struct virtio_net *dev = get_device(vid);
 938
 939         if (!dev)
 940                 return 0;
 941
 942         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
 943                 RTE_LOG(ERR, VHOST_DATA,
 944                         "(%d) %s: built-in vhost net backend is disabled.\n",
 945                         dev->vid, __func__);
 946                 return 0;
 947         }
 948
 949         return virtio_dev_rx(dev, queue_id, pkts, count);
 950 }
 951
 952 static inline bool
 953 virtio_net_with_host_offload(struct virtio_net *dev)
 954 {
 955         if (dev->features &
 956                         ((1ULL << VIRTIO_NET_F_CSUM) |
 957                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
 958                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
 959                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
 960                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
 961                 return true;
 962
 963         return false;
 964 }
 965
 966 static void
 967 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 968 {
 969         struct ipv4_hdr *ipv4_hdr;
 970         struct ipv6_hdr *ipv6_hdr;
 971         void *l3_hdr = NULL;
 972         struct rte_ether_hdr *eth_hdr;
 973         uint16_t ethertype;
 974
 975         eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
 976
 977         m->l2_len = sizeof(struct rte_ether_hdr);
 978         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 979
 980         if (ethertype == ETHER_TYPE_VLAN) {
 981                 struct rte_vlan_hdr *vlan_hdr =
 982                         (struct rte_vlan_hdr *)(eth_hdr + 1);
 983
 984                 m->l2_len += sizeof(struct rte_vlan_hdr);
 985                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 986         }
 987
 988         l3_hdr = (char *)eth_hdr + m->l2_len;
 989
 990         switch (ethertype) {
 991         case ETHER_TYPE_IPv4:
 992                 ipv4_hdr = l3_hdr;
 993                 *l4_proto = ipv4_hdr->next_proto_id;
 994                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 995                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 996                 m->ol_flags |= PKT_TX_IPV4;
 997                 break;
 998         case ETHER_TYPE_IPv6:
 999                 ipv6_hdr = l3_hdr;
1000                 *l4_proto = ipv6_hdr->proto;
1001                 m->l3_len = sizeof(struct ipv6_hdr);
1002                 *l4_hdr = (char *)l3_hdr + m->l3_len;
1003                 m->ol_flags |= PKT_TX_IPV6;
1004                 break;
1005         default:
1006                 m->l3_len = 0;
1007                 *l4_proto = 0;
1008                 *l4_hdr = NULL;
1009                 break;
1010         }
1011 }
1012
1013 static __rte_always_inline void
1014 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
1015 {
1016         uint16_t l4_proto = 0;
1017         void *l4_hdr = NULL;
1018         struct tcp_hdr *tcp_hdr = NULL;
1019
1020         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1021                 return;
1022
1023         parse_ethernet(m, &l4_proto, &l4_hdr);
1024         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1025                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
1026                         switch (hdr->csum_offset) {
1027                         case (offsetof(struct tcp_hdr, cksum)):
1028                                 if (l4_proto == IPPROTO_TCP)
1029                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
1030                                 break;
1031                         case (offsetof(struct udp_hdr, dgram_cksum)):
1032                                 if (l4_proto == IPPROTO_UDP)
1033                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
1034                                 break;
1035                         case (offsetof(struct sctp_hdr, cksum)):
1036                                 if (l4_proto == IPPROTO_SCTP)
1037                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
1038                                 break;
1039                         default:
1040                                 break;
1041                         }
1042                 }
1043         }
1044
1045         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1046                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1047                 case VIRTIO_NET_HDR_GSO_TCPV4:
1048                 case VIRTIO_NET_HDR_GSO_TCPV6:
1049                         tcp_hdr = l4_hdr;
1050                         m->ol_flags |= PKT_TX_TCP_SEG;
1051                         m->tso_segsz = hdr->gso_size;
1052                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
1053                         break;
1054                 case VIRTIO_NET_HDR_GSO_UDP:
1055                         m->ol_flags |= PKT_TX_UDP_SEG;
1056                         m->tso_segsz = hdr->gso_size;
1057                         m->l4_len = sizeof(struct udp_hdr);
1058                         break;
1059                 default:
1060                         RTE_LOG(WARNING, VHOST_DATA,
1061                                 "unsupported gso type %u.\n", hdr->gso_type);
1062                         break;
1063                 }
1064         }
1065 }
1066
1067 static __rte_always_inline int
1068 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
1069                   struct buf_vector *buf_vec, uint16_t nr_vec,
1070                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
1071 {
1072         uint32_t buf_avail, buf_offset;
1073         uint64_t buf_addr, buf_iova, buf_len;
1074         uint32_t mbuf_avail, mbuf_offset;
1075         uint32_t cpy_len;
1076         struct rte_mbuf *cur = m, *prev = m;
1077         struct virtio_net_hdr tmp_hdr;
1078         struct virtio_net_hdr *hdr = NULL;
1079         /* A counter to avoid desc dead loop chain */
1080         uint16_t vec_idx = 0;
1081         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
1082         int error = 0;
1083
1084         buf_addr = buf_vec[vec_idx].buf_addr;
1085         buf_iova = buf_vec[vec_idx].buf_iova;
1086         buf_len = buf_vec[vec_idx].buf_len;
1087
1088         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
1089                 error = -1;
1090                 goto out;
1091         }
1092
1093         if (likely(nr_vec > 1))
1094                 rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
1095
1096         if (virtio_net_with_host_offload(dev)) {
1097                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
1098                         uint64_t len;
1099                         uint64_t remain = sizeof(struct virtio_net_hdr);
1100                         uint64_t src;
1101                         uint64_t dst = (uint64_t)(uintptr_t)&tmp_hdr;
1102                         uint16_t hdr_vec_idx = 0;
1103
1104                         /*
1105                          * No luck, the virtio-net header doesn't fit
1106                          * in a contiguous virtual area.
1107                          */
1108                         while (remain) {
1109                                 len = RTE_MIN(remain,
1110                                         buf_vec[hdr_vec_idx].buf_len);
1111                                 src = buf_vec[hdr_vec_idx].buf_addr;
1112                                 rte_memcpy((void *)(uintptr_t)dst,
1113                                                    (void *)(uintptr_t)src, len);
1114
1115                                 remain -= len;
1116                                 dst += len;
1117                                 hdr_vec_idx++;
1118                         }
1119
1120                         hdr = &tmp_hdr;
1121                 } else {
1122                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
1123                         rte_prefetch0(hdr);
1124                 }
1125         }
1126
1127         /*
1128          * A virtio driver normally uses at least 2 desc buffers
1129          * for Tx: the first for storing the header, and others
1130          * for storing the data.
1131          */
1132         if (unlikely(buf_len < dev->vhost_hlen)) {
1133                 buf_offset = dev->vhost_hlen - buf_len;
1134                 vec_idx++;
1135                 buf_addr = buf_vec[vec_idx].buf_addr;
1136                 buf_iova = buf_vec[vec_idx].buf_iova;
1137                 buf_len = buf_vec[vec_idx].buf_len;
1138                 buf_avail  = buf_len - buf_offset;
1139         } else if (buf_len == dev->vhost_hlen) {
1140                 if (unlikely(++vec_idx >= nr_vec))
1141                         goto out;
1142                 buf_addr = buf_vec[vec_idx].buf_addr;
1143                 buf_iova = buf_vec[vec_idx].buf_iova;
1144                 buf_len = buf_vec[vec_idx].buf_len;
1145
1146                 buf_offset = 0;
1147                 buf_avail = buf_len;
1148         } else {
1149                 buf_offset = dev->vhost_hlen;
1150                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
1151         }
1152
1153         rte_prefetch0((void *)(uintptr_t)
1154                         (buf_addr + buf_offset));
1155
1156         PRINT_PACKET(dev,
1157                         (uintptr_t)(buf_addr + buf_offset),
1158                         (uint32_t)buf_avail, 0);
1159
1160         mbuf_offset = 0;
1161         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
1162         while (1) {
1163                 uint64_t hpa;
1164
1165                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1166
1167                 /*
1168                  * A desc buf might across two host physical pages that are
1169                  * not continuous. In such case (gpa_to_hpa returns 0), data
1170                  * will be copied even though zero copy is enabled.
1171                  */
1172                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
1173                                         buf_iova + buf_offset, cpy_len)))) {
1174                         cur->data_len = cpy_len;
1175                         cur->data_off = 0;
1176                         cur->buf_addr =
1177                                 (void *)(uintptr_t)(buf_addr + buf_offset);
1178                         cur->buf_iova = hpa;
1179
1180                         /*
1181                          * In zero copy mode, one mbuf can only reference data
1182                          * for one or partial of one desc buff.
1183                          */
1184                         mbuf_avail = cpy_len;
1185                 } else {
1186                         if (likely(cpy_len > MAX_BATCH_LEN ||
1187                                    vq->batch_copy_nb_elems >= vq->size ||
1188                                    (hdr && cur == m))) {
1189                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
1190                                                                    mbuf_offset),
1191                                            (void *)((uintptr_t)(buf_addr +
1192                                                            buf_offset)),
1193                                            cpy_len);
1194                         } else {
1195                                 batch_copy[vq->batch_copy_nb_elems].dst =
1196                                         rte_pktmbuf_mtod_offset(cur, void *,
1197                                                                 mbuf_offset);
1198                                 batch_copy[vq->batch_copy_nb_elems].src =
1199                                         (void *)((uintptr_t)(buf_addr +
1200                                                                 buf_offset));
1201                                 batch_copy[vq->batch_copy_nb_elems].len =
1202                                         cpy_len;
1203                                 vq->batch_copy_nb_elems++;
1204                         }
1205                 }
1206
1207                 mbuf_avail  -= cpy_len;
1208                 mbuf_offset += cpy_len;
1209                 buf_avail -= cpy_len;
1210                 buf_offset += cpy_len;
1211
1212                 /* This buf reaches to its end, get the next one */
1213                 if (buf_avail == 0) {
1214                         if (++vec_idx >= nr_vec)
1215                                 break;
1216
1217                         buf_addr = buf_vec[vec_idx].buf_addr;
1218                         buf_iova = buf_vec[vec_idx].buf_iova;
1219                         buf_len = buf_vec[vec_idx].buf_len;
1220
1221                         /*
1222                          * Prefecth desc n + 1 buffer while
1223                          * desc n buffer is processed.
1224                          */
1225                         if (vec_idx + 1 < nr_vec)
1226                                 rte_prefetch0((void *)(uintptr_t)
1227                                                 buf_vec[vec_idx + 1].buf_addr);
1228
1229                         buf_offset = 0;
1230                         buf_avail  = buf_len;
1231
1232                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
1233                                         (uint32_t)buf_avail, 0);
1234                 }
1235
1236                 /*
1237                  * This mbuf reaches to its end, get a new one
1238                  * to hold more data.
1239                  */
1240                 if (mbuf_avail == 0) {
1241                         cur = rte_pktmbuf_alloc(mbuf_pool);
1242                         if (unlikely(cur == NULL)) {
1243                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1244                                         "allocate memory for mbuf.\n");
1245                                 error = -1;
1246                                 goto out;
1247                         }
1248                         if (unlikely(dev->dequeue_zero_copy))
1249                                 rte_mbuf_refcnt_update(cur, 1);
1250
1251                         prev->next = cur;
1252                         prev->data_len = mbuf_offset;
1253                         m->nb_segs += 1;
1254                         m->pkt_len += mbuf_offset;
1255                         prev = cur;
1256
1257                         mbuf_offset = 0;
1258                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1259                 }
1260         }
1261
1262         prev->data_len = mbuf_offset;
1263         m->pkt_len    += mbuf_offset;
1264
1265         if (hdr)
1266                 vhost_dequeue_offload(hdr, m);
1267
1268 out:
1269
1270         return error;
1271 }
1272
1273 static __rte_always_inline struct zcopy_mbuf *
1274 get_zmbuf(struct vhost_virtqueue *vq)
1275 {
1276         uint16_t i;
1277         uint16_t last;
1278         int tries = 0;
1279
1280         /* search [last_zmbuf_idx, zmbuf_size) */
1281         i = vq->last_zmbuf_idx;
1282         last = vq->zmbuf_size;
1283
1284 again:
1285         for (; i < last; i++) {
1286                 if (vq->zmbufs[i].in_use == 0) {
1287                         vq->last_zmbuf_idx = i + 1;
1288                         vq->zmbufs[i].in_use = 1;
1289                         return &vq->zmbufs[i];
1290                 }
1291         }
1292
1293         tries++;
1294         if (tries == 1) {
1295                 /* search [0, last_zmbuf_idx) */
1296                 i = 0;
1297                 last = vq->last_zmbuf_idx;
1298                 goto again;
1299         }
1300
1301         return NULL;
1302 }
1303
1304 static __rte_always_inline uint16_t
1305 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1306         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1307 {
1308         uint16_t i;
1309         uint16_t free_entries;
1310
1311         if (unlikely(dev->dequeue_zero_copy)) {
1312                 struct zcopy_mbuf *zmbuf, *next;
1313
1314                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1315                      zmbuf != NULL; zmbuf = next) {
1316                         next = TAILQ_NEXT(zmbuf, next);
1317
1318                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1319                                 update_shadow_used_ring_split(vq,
1320                                                 zmbuf->desc_idx, 0);
1321                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1322                                 restore_mbuf(zmbuf->mbuf);
1323                                 rte_pktmbuf_free(zmbuf->mbuf);
1324                                 put_zmbuf(zmbuf);
1325                                 vq->nr_zmbuf -= 1;
1326                         }
1327                 }
1328
1329                 if (likely(vq->shadow_used_idx)) {
1330                         flush_shadow_used_ring_split(dev, vq);
1331                         vhost_vring_call_split(dev, vq);
1332                 }
1333         }
1334
1335         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1336                         vq->last_avail_idx;
1337         if (free_entries == 0)
1338                 return 0;
1339
1340         /*
1341          * The ordering between avail index and
1342          * desc reads needs to be enforced.
1343          */
1344         rte_smp_rmb();
1345
1346         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1347
1348         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1349
1350         count = RTE_MIN(count, MAX_PKT_BURST);
1351         count = RTE_MIN(count, free_entries);
1352         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1353                         dev->vid, count);
1354
1355         for (i = 0; i < count; i++) {
1356                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1357                 uint16_t head_idx;
1358                 uint32_t dummy_len;
1359                 uint16_t nr_vec = 0;
1360                 int err;
1361
1362                 if (unlikely(fill_vec_buf_split(dev, vq,
1363                                                 vq->last_avail_idx + i,
1364                                                 &nr_vec, buf_vec,
1365                                                 &head_idx, &dummy_len,
1366                                                 VHOST_ACCESS_RO) < 0))
1367                         break;
1368
1369                 if (likely(dev->dequeue_zero_copy == 0))
1370                         update_shadow_used_ring_split(vq, head_idx, 0);
1371
1372                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
1373
1374                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1375                 if (unlikely(pkts[i] == NULL)) {
1376                         RTE_LOG(ERR, VHOST_DATA,
1377                                 "Failed to allocate memory for mbuf.\n");
1378                         break;
1379                 }
1380
1381                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1382                                 mbuf_pool);
1383                 if (unlikely(err)) {
1384                         rte_pktmbuf_free(pkts[i]);
1385                         break;
1386                 }
1387
1388                 if (unlikely(dev->dequeue_zero_copy)) {
1389                         struct zcopy_mbuf *zmbuf;
1390
1391                         zmbuf = get_zmbuf(vq);
1392                         if (!zmbuf) {
1393                                 rte_pktmbuf_free(pkts[i]);
1394                                 break;
1395                         }
1396                         zmbuf->mbuf = pkts[i];
1397                         zmbuf->desc_idx = head_idx;
1398
1399                         /*
1400                          * Pin lock the mbuf; we will check later to see
1401                          * whether the mbuf is freed (when we are the last
1402                          * user) or not. If that's the case, we then could
1403                          * update the used ring safely.
1404                          */
1405                         rte_mbuf_refcnt_update(pkts[i], 1);
1406
1407                         vq->nr_zmbuf += 1;
1408                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1409                 }
1410         }
1411         vq->last_avail_idx += i;
1412
1413         if (likely(dev->dequeue_zero_copy == 0)) {
1414                 do_data_copy_dequeue(vq);
1415                 if (unlikely(i < count))
1416                         vq->shadow_used_idx = i;
1417                 if (likely(vq->shadow_used_idx)) {
1418                         flush_shadow_used_ring_split(dev, vq);
1419                         vhost_vring_call_split(dev, vq);
1420                 }
1421         }
1422
1423         return i;
1424 }
1425
1426 static __rte_always_inline uint16_t
1427 virtio_dev_tx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
1428         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1429 {
1430         uint16_t i;
1431
1432         if (unlikely(dev->dequeue_zero_copy)) {
1433                 struct zcopy_mbuf *zmbuf, *next;
1434
1435                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1436                      zmbuf != NULL; zmbuf = next) {
1437                         next = TAILQ_NEXT(zmbuf, next);
1438
1439                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1440                                 update_shadow_used_ring_packed(vq,
1441                                                 zmbuf->desc_idx,
1442                                                 0,
1443                                                 zmbuf->desc_count);
1444
1445                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1446                                 restore_mbuf(zmbuf->mbuf);
1447                                 rte_pktmbuf_free(zmbuf->mbuf);
1448                                 put_zmbuf(zmbuf);
1449                                 vq->nr_zmbuf -= 1;
1450                         }
1451                 }
1452
1453                 if (likely(vq->shadow_used_idx)) {
1454                         flush_shadow_used_ring_packed(dev, vq);
1455                         vhost_vring_call_packed(dev, vq);
1456                 }
1457         }
1458
1459         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1460
1461         count = RTE_MIN(count, MAX_PKT_BURST);
1462         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1463                         dev->vid, count);
1464
1465         for (i = 0; i < count; i++) {
1466                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1467                 uint16_t buf_id;
1468                 uint32_t dummy_len;
1469                 uint16_t desc_count, nr_vec = 0;
1470                 int err;
1471
1472                 if (unlikely(fill_vec_buf_packed(dev, vq,
1473                                                 vq->last_avail_idx, &desc_count,
1474                                                 buf_vec, &nr_vec,
1475                                                 &buf_id, &dummy_len,
1476                                                 VHOST_ACCESS_RO) < 0))
1477                         break;
1478
1479                 if (likely(dev->dequeue_zero_copy == 0))
1480                         update_shadow_used_ring_packed(vq, buf_id, 0,
1481                                         desc_count);
1482
1483                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
1484
1485                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1486                 if (unlikely(pkts[i] == NULL)) {
1487                         RTE_LOG(ERR, VHOST_DATA,
1488                                 "Failed to allocate memory for mbuf.\n");
1489                         break;
1490                 }
1491
1492                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1493                                 mbuf_pool);
1494                 if (unlikely(err)) {
1495                         rte_pktmbuf_free(pkts[i]);
1496                         break;
1497                 }
1498
1499                 if (unlikely(dev->dequeue_zero_copy)) {
1500                         struct zcopy_mbuf *zmbuf;
1501
1502                         zmbuf = get_zmbuf(vq);
1503                         if (!zmbuf) {
1504                                 rte_pktmbuf_free(pkts[i]);
1505                                 break;
1506                         }
1507                         zmbuf->mbuf = pkts[i];
1508                         zmbuf->desc_idx = buf_id;
1509                         zmbuf->desc_count = desc_count;
1510
1511                         /*
1512                          * Pin lock the mbuf; we will check later to see
1513                          * whether the mbuf is freed (when we are the last
1514                          * user) or not. If that's the case, we then could
1515                          * update the used ring safely.
1516                          */
1517                         rte_mbuf_refcnt_update(pkts[i], 1);
1518
1519                         vq->nr_zmbuf += 1;
1520                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1521                 }
1522
1523                 vq->last_avail_idx += desc_count;
1524                 if (vq->last_avail_idx >= vq->size) {
1525                         vq->last_avail_idx -= vq->size;
1526                         vq->avail_wrap_counter ^= 1;
1527                 }
1528         }
1529
1530         if (likely(dev->dequeue_zero_copy == 0)) {
1531                 do_data_copy_dequeue(vq);
1532                 if (unlikely(i < count))
1533                         vq->shadow_used_idx = i;
1534                 if (likely(vq->shadow_used_idx)) {
1535                         flush_shadow_used_ring_packed(dev, vq);
1536                         vhost_vring_call_packed(dev, vq);
1537                 }
1538         }
1539
1540         return i;
1541 }
1542
1543 uint16_t
1544 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1545         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1546 {
1547         struct virtio_net *dev;
1548         struct rte_mbuf *rarp_mbuf = NULL;
1549         struct vhost_virtqueue *vq;
1550
1551         dev = get_device(vid);
1552         if (!dev)
1553                 return 0;
1554
1555         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1556                 RTE_LOG(ERR, VHOST_DATA,
1557                         "(%d) %s: built-in vhost net backend is disabled.\n",
1558                         dev->vid, __func__);
1559                 return 0;
1560         }
1561
1562         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1563                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1564                         dev->vid, __func__, queue_id);
1565                 return 0;
1566         }
1567
1568         vq = dev->virtqueue[queue_id];
1569
1570         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
1571                 return 0;
1572
1573         if (unlikely(vq->enabled == 0)) {
1574                 count = 0;
1575                 goto out_access_unlock;
1576         }
1577
1578         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1579                 vhost_user_iotlb_rd_lock(vq);
1580
1581         if (unlikely(vq->access_ok == 0))
1582                 if (unlikely(vring_translate(dev, vq) < 0)) {
1583                         count = 0;
1584                         goto out;
1585                 }
1586
1587         /*
1588          * Construct a RARP broadcast packet, and inject it to the "pkts"
1589          * array, to looks like that guest actually send such packet.
1590          *
1591          * Check user_send_rarp() for more information.
1592          *
1593          * broadcast_rarp shares a cacheline in the virtio_net structure
1594          * with some fields that are accessed during enqueue and
1595          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1596          * result in false sharing between enqueue and dequeue.
1597          *
1598          * Prevent unnecessary false sharing by reading broadcast_rarp first
1599          * and only performing cmpset if the read indicates it is likely to
1600          * be set.
1601          */
1602         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1603                         rte_atomic16_cmpset((volatile uint16_t *)
1604                                 &dev->broadcast_rarp.cnt, 1, 0))) {
1605
1606                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
1607                 if (rarp_mbuf == NULL) {
1608                         RTE_LOG(ERR, VHOST_DATA,
1609                                 "Failed to make RARP packet.\n");
1610                         count = 0;
1611                         goto out;
1612                 }
1613                 count -= 1;
1614         }
1615
1616         if (vq_is_packed(dev))
1617                 count = virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count);
1618         else
1619                 count = virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count);
1620
1621 out:
1622         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1623                 vhost_user_iotlb_rd_unlock(vq);
1624
1625 out_access_unlock:
1626         rte_spinlock_unlock(&vq->access_lock);
1627
1628         if (unlikely(rarp_mbuf != NULL)) {
1629                 /*
1630                  * Inject it to the head of "pkts" array, so that switch's mac
1631                  * learning table will get updated first.
1632                  */
1633                 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
1634                 pkts[0] = rarp_mbuf;
1635                 count += 1;
1636         }
1637
1638         return count;
1639 }