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         for (i = 0; i < vq->shadow_used_idx; i++) {
 114                 uint16_t flags;
 115
 116                 if (vq->shadow_used_packed[i].len)
 117                         flags = VRING_DESC_F_WRITE;
 118                 else
 119                         flags = 0;
 120
 121                 if (vq->used_wrap_counter) {
 122                         flags |= VRING_DESC_F_USED;
 123                         flags |= VRING_DESC_F_AVAIL;
 124                 } else {
 125                         flags &= ~VRING_DESC_F_USED;
 126                         flags &= ~VRING_DESC_F_AVAIL;
 127                 }
 128
 129                 if (i > 0) {
 130                         vq->desc_packed[vq->last_used_idx].flags = flags;
 131
 132                         vhost_log_cache_used_vring(dev, vq,
 133                                         vq->last_used_idx *
 134                                         sizeof(struct vring_packed_desc),
 135                                         sizeof(struct vring_packed_desc));
 136                 } else {
 137                         head_idx = vq->last_used_idx;
 138                         head_flags = flags;
 139                 }
 140
 141                 vq_inc_last_used_packed(vq, vq->shadow_used_packed[i].count);
 142         }
 143
 144         __atomic_store_n(&vq->desc_packed[head_idx].flags, head_flags,
 145                          __ATOMIC_RELEASE);
 146
 147         vhost_log_cache_used_vring(dev, vq,
 148                                 head_idx *
 149                                 sizeof(struct vring_packed_desc),
 150                                 sizeof(struct vring_packed_desc));
 151
 152         vq->shadow_used_idx = 0;
 153         vhost_log_cache_sync(dev, vq);
 154 }
 155
 156 static __rte_always_inline void
 157 update_shadow_used_ring_packed(struct vhost_virtqueue *vq,
 158                          uint16_t desc_idx, uint32_t len, uint16_t count)
 159 {
 160         uint16_t i = vq->shadow_used_idx++;
 161
 162         vq->shadow_used_packed[i].id  = desc_idx;
 163         vq->shadow_used_packed[i].len = len;
 164         vq->shadow_used_packed[i].count = count;
 165 }
 166
 167 static inline void
 168 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
 169 {
 170         struct batch_copy_elem *elem = vq->batch_copy_elems;
 171         uint16_t count = vq->batch_copy_nb_elems;
 172         int i;
 173
 174         for (i = 0; i < count; i++) {
 175                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 176                 vhost_log_cache_write_iova(dev, vq, elem[i].log_addr,
 177                                            elem[i].len);
 178                 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
 179         }
 180
 181         vq->batch_copy_nb_elems = 0;
 182 }
 183
 184 static inline void
 185 do_data_copy_dequeue(struct vhost_virtqueue *vq)
 186 {
 187         struct batch_copy_elem *elem = vq->batch_copy_elems;
 188         uint16_t count = vq->batch_copy_nb_elems;
 189         int i;
 190
 191         for (i = 0; i < count; i++)
 192                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 193
 194         vq->batch_copy_nb_elems = 0;
 195 }
 196
 197 /* avoid write operation when necessary, to lessen cache issues */
 198 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 199         if ((var) != (val))                     \
 200                 (var) = (val);                  \
 201 } while (0)
 202
 203 static __rte_always_inline void
 204 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 205 {
 206         uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
 207
 208         if (m_buf->ol_flags & PKT_TX_TCP_SEG)
 209                 csum_l4 |= PKT_TX_TCP_CKSUM;
 210
 211         if (csum_l4) {
 212                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 213                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 214
 215                 switch (csum_l4) {
 216                 case PKT_TX_TCP_CKSUM:
 217                         net_hdr->csum_offset = (offsetof(struct rte_tcp_hdr,
 218                                                 cksum));
 219                         break;
 220                 case PKT_TX_UDP_CKSUM:
 221                         net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
 222                                                 dgram_cksum));
 223                         break;
 224                 case PKT_TX_SCTP_CKSUM:
 225                         net_hdr->csum_offset = (offsetof(struct rte_sctp_hdr,
 226                                                 cksum));
 227                         break;
 228                 }
 229         } else {
 230                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 231                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 232                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 233         }
 234
 235         /* IP cksum verification cannot be bypassed, then calculate here */
 236         if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
 237                 struct rte_ipv4_hdr *ipv4_hdr;
 238
 239                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_ipv4_hdr *,
 240                                                    m_buf->l2_len);
 241                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 242         }
 243
 244         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 245                 if (m_buf->ol_flags & PKT_TX_IPV4)
 246                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 247                 else
 248                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 249                 net_hdr->gso_size = m_buf->tso_segsz;
 250                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 251                                         + m_buf->l4_len;
 252         } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
 253                 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 254                 net_hdr->gso_size = m_buf->tso_segsz;
 255                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
 256                         m_buf->l4_len;
 257         } else {
 258                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 259                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 260                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 261         }
 262 }
 263
 264 static __rte_always_inline int
 265 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 266                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 267                 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
 268 {
 269         uint16_t vec_id = *vec_idx;
 270
 271         while (desc_len) {
 272                 uint64_t desc_addr;
 273                 uint64_t desc_chunck_len = desc_len;
 274
 275                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 276                         return -1;
 277
 278                 desc_addr = vhost_iova_to_vva(dev, vq,
 279                                 desc_iova,
 280                                 &desc_chunck_len,
 281                                 perm);
 282                 if (unlikely(!desc_addr))
 283                         return -1;
 284
 285                 rte_prefetch0((void *)(uintptr_t)desc_addr);
 286
 287                 buf_vec[vec_id].buf_iova = desc_iova;
 288                 buf_vec[vec_id].buf_addr = desc_addr;
 289                 buf_vec[vec_id].buf_len  = desc_chunck_len;
 290
 291                 desc_len -= desc_chunck_len;
 292                 desc_iova += desc_chunck_len;
 293                 vec_id++;
 294         }
 295         *vec_idx = vec_id;
 296
 297         return 0;
 298 }
 299
 300 static __rte_always_inline int
 301 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 302                          uint32_t avail_idx, uint16_t *vec_idx,
 303                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 304                          uint32_t *desc_chain_len, uint8_t perm)
 305 {
 306         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 307         uint16_t vec_id = *vec_idx;
 308         uint32_t len    = 0;
 309         uint64_t dlen;
 310         uint32_t nr_descs = vq->size;
 311         uint32_t cnt    = 0;
 312         struct vring_desc *descs = vq->desc;
 313         struct vring_desc *idesc = NULL;
 314
 315         if (unlikely(idx >= vq->size))
 316                 return -1;
 317
 318         *desc_chain_head = idx;
 319
 320         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 321                 dlen = vq->desc[idx].len;
 322                 nr_descs = dlen / sizeof(struct vring_desc);
 323                 if (unlikely(nr_descs > vq->size))
 324                         return -1;
 325
 326                 descs = (struct vring_desc *)(uintptr_t)
 327                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 328                                                 &dlen,
 329                                                 VHOST_ACCESS_RO);
 330                 if (unlikely(!descs))
 331                         return -1;
 332
 333                 if (unlikely(dlen < vq->desc[idx].len)) {
 334                         /*
 335                          * The indirect desc table is not contiguous
 336                          * in process VA space, we have to copy it.
 337                          */
 338                         idesc = vhost_alloc_copy_ind_table(dev, vq,
 339                                         vq->desc[idx].addr, vq->desc[idx].len);
 340                         if (unlikely(!idesc))
 341                                 return -1;
 342
 343                         descs = idesc;
 344                 }
 345
 346                 idx = 0;
 347         }
 348
 349         while (1) {
 350                 if (unlikely(idx >= nr_descs || cnt++ >= nr_descs)) {
 351                         free_ind_table(idesc);
 352                         return -1;
 353                 }
 354
 355                 len += descs[idx].len;
 356
 357                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 358                                                 descs[idx].addr, descs[idx].len,
 359                                                 perm))) {
 360                         free_ind_table(idesc);
 361                         return -1;
 362                 }
 363
 364                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 365                         break;
 366
 367                 idx = descs[idx].next;
 368         }
 369
 370         *desc_chain_len = len;
 371         *vec_idx = vec_id;
 372
 373         if (unlikely(!!idesc))
 374                 free_ind_table(idesc);
 375
 376         return 0;
 377 }
 378
 379 /*
 380  * Returns -1 on fail, 0 on success
 381  */
 382 static inline int
 383 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 384                                 uint32_t size, struct buf_vector *buf_vec,
 385                                 uint16_t *num_buffers, uint16_t avail_head,
 386                                 uint16_t *nr_vec)
 387 {
 388         uint16_t cur_idx;
 389         uint16_t vec_idx = 0;
 390         uint16_t max_tries, tries = 0;
 391
 392         uint16_t head_idx = 0;
 393         uint32_t len = 0;
 394
 395         *num_buffers = 0;
 396         cur_idx  = vq->last_avail_idx;
 397
 398         if (rxvq_is_mergeable(dev))
 399                 max_tries = vq->size - 1;
 400         else
 401                 max_tries = 1;
 402
 403         while (size > 0) {
 404                 if (unlikely(cur_idx == avail_head))
 405                         return -1;
 406                 /*
 407                  * if we tried all available ring items, and still
 408                  * can't get enough buf, it means something abnormal
 409                  * happened.
 410                  */
 411                 if (unlikely(++tries > max_tries))
 412                         return -1;
 413
 414                 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
 415                                                 &vec_idx, buf_vec,
 416                                                 &head_idx, &len,
 417                                                 VHOST_ACCESS_RW) < 0))
 418                         return -1;
 419                 len = RTE_MIN(len, size);
 420                 update_shadow_used_ring_split(vq, head_idx, len);
 421                 size -= len;
 422
 423                 cur_idx++;
 424                 *num_buffers += 1;
 425         }
 426
 427         *nr_vec = vec_idx;
 428
 429         return 0;
 430 }
 431
 432 static __rte_always_inline int
 433 fill_vec_buf_packed_indirect(struct virtio_net *dev,
 434                         struct vhost_virtqueue *vq,
 435                         struct vring_packed_desc *desc, uint16_t *vec_idx,
 436                         struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
 437 {
 438         uint16_t i;
 439         uint32_t nr_descs;
 440         uint16_t vec_id = *vec_idx;
 441         uint64_t dlen;
 442         struct vring_packed_desc *descs, *idescs = NULL;
 443
 444         dlen = desc->len;
 445         descs = (struct vring_packed_desc *)(uintptr_t)
 446                 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
 447         if (unlikely(!descs))
 448                 return -1;
 449
 450         if (unlikely(dlen < desc->len)) {
 451                 /*
 452                  * The indirect desc table is not contiguous
 453                  * in process VA space, we have to copy it.
 454                  */
 455                 idescs = vhost_alloc_copy_ind_table(dev,
 456                                 vq, desc->addr, desc->len);
 457                 if (unlikely(!idescs))
 458                         return -1;
 459
 460                 descs = idescs;
 461         }
 462
 463         nr_descs =  desc->len / sizeof(struct vring_packed_desc);
 464         if (unlikely(nr_descs >= vq->size)) {
 465                 free_ind_table(idescs);
 466                 return -1;
 467         }
 468
 469         for (i = 0; i < nr_descs; i++) {
 470                 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
 471                         free_ind_table(idescs);
 472                         return -1;
 473                 }
 474
 475                 *len += descs[i].len;
 476                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 477                                                 descs[i].addr, descs[i].len,
 478                                                 perm)))
 479                         return -1;
 480         }
 481         *vec_idx = vec_id;
 482
 483         if (unlikely(!!idescs))
 484                 free_ind_table(idescs);
 485
 486         return 0;
 487 }
 488
 489 static __rte_always_inline int
 490 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 491                                 uint16_t avail_idx, uint16_t *desc_count,
 492                                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 493                                 uint16_t *buf_id, uint32_t *len, uint8_t perm)
 494 {
 495         bool wrap_counter = vq->avail_wrap_counter;
 496         struct vring_packed_desc *descs = vq->desc_packed;
 497         uint16_t vec_id = *vec_idx;
 498
 499         if (avail_idx < vq->last_avail_idx)
 500                 wrap_counter ^= 1;
 501
 502         /*
 503          * Perform a load-acquire barrier in desc_is_avail to
 504          * enforce the ordering between desc flags and desc
 505          * content.
 506          */
 507         if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
 508                 return -1;
 509
 510         *desc_count = 0;
 511         *len = 0;
 512
 513         while (1) {
 514                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 515                         return -1;
 516
 517                 if (unlikely(*desc_count >= vq->size))
 518                         return -1;
 519
 520                 *desc_count += 1;
 521                 *buf_id = descs[avail_idx].id;
 522
 523                 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
 524                         if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
 525                                                         &descs[avail_idx],
 526                                                         &vec_id, buf_vec,
 527                                                         len, perm) < 0))
 528                                 return -1;
 529                 } else {
 530                         *len += descs[avail_idx].len;
 531
 532                         if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 533                                                         descs[avail_idx].addr,
 534                                                         descs[avail_idx].len,
 535                                                         perm)))
 536                                 return -1;
 537                 }
 538
 539                 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
 540                         break;
 541
 542                 if (++avail_idx >= vq->size) {
 543                         avail_idx -= vq->size;
 544                         wrap_counter ^= 1;
 545                 }
 546         }
 547
 548         *vec_idx = vec_id;
 549
 550         return 0;
 551 }
 552
 553 /*
 554  * Returns -1 on fail, 0 on success
 555  */
 556 static inline int
 557 reserve_avail_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 558                                 uint32_t size, struct buf_vector *buf_vec,
 559                                 uint16_t *nr_vec, uint16_t *num_buffers,
 560                                 uint16_t *nr_descs)
 561 {
 562         uint16_t avail_idx;
 563         uint16_t vec_idx = 0;
 564         uint16_t max_tries, tries = 0;
 565
 566         uint16_t buf_id = 0;
 567         uint32_t len = 0;
 568         uint16_t desc_count;
 569
 570         *num_buffers = 0;
 571         avail_idx = vq->last_avail_idx;
 572
 573         if (rxvq_is_mergeable(dev))
 574                 max_tries = vq->size - 1;
 575         else
 576                 max_tries = 1;
 577
 578         while (size > 0) {
 579                 /*
 580                  * if we tried all available ring items, and still
 581                  * can't get enough buf, it means something abnormal
 582                  * happened.
 583                  */
 584                 if (unlikely(++tries > max_tries))
 585                         return -1;
 586
 587                 if (unlikely(fill_vec_buf_packed(dev, vq,
 588                                                 avail_idx, &desc_count,
 589                                                 buf_vec, &vec_idx,
 590                                                 &buf_id, &len,
 591                                                 VHOST_ACCESS_RW) < 0))
 592                         return -1;
 593
 594                 len = RTE_MIN(len, size);
 595                 update_shadow_used_ring_packed(vq, buf_id, len, desc_count);
 596                 size -= len;
 597
 598                 avail_idx += desc_count;
 599                 if (avail_idx >= vq->size)
 600                         avail_idx -= vq->size;
 601
 602                 *nr_descs += desc_count;
 603                 *num_buffers += 1;
 604         }
 605
 606         *nr_vec = vec_idx;
 607
 608         return 0;
 609 }
 610
 611 static __rte_noinline void
 612 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 613                 struct buf_vector *buf_vec,
 614                 struct virtio_net_hdr_mrg_rxbuf *hdr)
 615 {
 616         uint64_t len;
 617         uint64_t remain = dev->vhost_hlen;
 618         uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 619         uint64_t iova = buf_vec->buf_iova;
 620
 621         while (remain) {
 622                 len = RTE_MIN(remain,
 623                                 buf_vec->buf_len);
 624                 dst = buf_vec->buf_addr;
 625                 rte_memcpy((void *)(uintptr_t)dst,
 626                                 (void *)(uintptr_t)src,
 627                                 len);
 628
 629                 PRINT_PACKET(dev, (uintptr_t)dst,
 630                                 (uint32_t)len, 0);
 631                 vhost_log_cache_write_iova(dev, vq,
 632                                 iova, len);
 633
 634                 remain -= len;
 635                 iova += len;
 636                 src += len;
 637                 buf_vec++;
 638         }
 639 }
 640
 641 static __rte_always_inline int
 642 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 643                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 644                             uint16_t nr_vec, uint16_t num_buffers)
 645 {
 646         uint32_t vec_idx = 0;
 647         uint32_t mbuf_offset, mbuf_avail;
 648         uint32_t buf_offset, buf_avail;
 649         uint64_t buf_addr, buf_iova, buf_len;
 650         uint32_t cpy_len;
 651         uint64_t hdr_addr;
 652         struct rte_mbuf *hdr_mbuf;
 653         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 654         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 655         int error = 0;
 656
 657         if (unlikely(m == NULL)) {
 658                 error = -1;
 659                 goto out;
 660         }
 661
 662         buf_addr = buf_vec[vec_idx].buf_addr;
 663         buf_iova = buf_vec[vec_idx].buf_iova;
 664         buf_len = buf_vec[vec_idx].buf_len;
 665
 666         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 667                 error = -1;
 668                 goto out;
 669         }
 670
 671         hdr_mbuf = m;
 672         hdr_addr = buf_addr;
 673         if (unlikely(buf_len < dev->vhost_hlen))
 674                 hdr = &tmp_hdr;
 675         else
 676                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
 677
 678         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 679                 dev->vid, num_buffers);
 680
 681         if (unlikely(buf_len < dev->vhost_hlen)) {
 682                 buf_offset = dev->vhost_hlen - buf_len;
 683                 vec_idx++;
 684                 buf_addr = buf_vec[vec_idx].buf_addr;
 685                 buf_iova = buf_vec[vec_idx].buf_iova;
 686                 buf_len = buf_vec[vec_idx].buf_len;
 687                 buf_avail = buf_len - buf_offset;
 688         } else {
 689                 buf_offset = dev->vhost_hlen;
 690                 buf_avail = buf_len - dev->vhost_hlen;
 691         }
 692
 693         mbuf_avail  = rte_pktmbuf_data_len(m);
 694         mbuf_offset = 0;
 695         while (mbuf_avail != 0 || m->next != NULL) {
 696                 /* done with current buf, get the next one */
 697                 if (buf_avail == 0) {
 698                         vec_idx++;
 699                         if (unlikely(vec_idx >= nr_vec)) {
 700                                 error = -1;
 701                                 goto out;
 702                         }
 703
 704                         buf_addr = buf_vec[vec_idx].buf_addr;
 705                         buf_iova = buf_vec[vec_idx].buf_iova;
 706                         buf_len = buf_vec[vec_idx].buf_len;
 707
 708                         buf_offset = 0;
 709                         buf_avail  = buf_len;
 710                 }
 711
 712                 /* done with current mbuf, get the next one */
 713                 if (mbuf_avail == 0) {
 714                         m = m->next;
 715
 716                         mbuf_offset = 0;
 717                         mbuf_avail  = rte_pktmbuf_data_len(m);
 718                 }
 719
 720                 if (hdr_addr) {
 721                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 722                         if (rxvq_is_mergeable(dev))
 723                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
 724                                                 num_buffers);
 725
 726                         if (unlikely(hdr == &tmp_hdr)) {
 727                                 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
 728                         } else {
 729                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 730                                                 dev->vhost_hlen, 0);
 731                                 vhost_log_cache_write_iova(dev, vq,
 732                                                 buf_vec[0].buf_iova,
 733                                                 dev->vhost_hlen);
 734                         }
 735
 736                         hdr_addr = 0;
 737                 }
 738
 739                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 740
 741                 if (likely(cpy_len > MAX_BATCH_LEN ||
 742                                         vq->batch_copy_nb_elems >= vq->size)) {
 743                         rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
 744                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 745                                 cpy_len);
 746                         vhost_log_cache_write_iova(dev, vq,
 747                                                    buf_iova + buf_offset,
 748                                                    cpy_len);
 749                         PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
 750                                 cpy_len, 0);
 751                 } else {
 752                         batch_copy[vq->batch_copy_nb_elems].dst =
 753                                 (void *)((uintptr_t)(buf_addr + buf_offset));
 754                         batch_copy[vq->batch_copy_nb_elems].src =
 755                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 756                         batch_copy[vq->batch_copy_nb_elems].log_addr =
 757                                 buf_iova + buf_offset;
 758                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
 759                         vq->batch_copy_nb_elems++;
 760                 }
 761
 762                 mbuf_avail  -= cpy_len;
 763                 mbuf_offset += cpy_len;
 764                 buf_avail  -= cpy_len;
 765                 buf_offset += cpy_len;
 766         }
 767
 768 out:
 769
 770         return error;
 771 }
 772
 773 static __rte_always_inline int
 774 vhost_enqueue_single_packed(struct virtio_net *dev,
 775                             struct vhost_virtqueue *vq,
 776                             struct rte_mbuf *pkt,
 777                             struct buf_vector *buf_vec,
 778                             uint16_t *nr_descs)
 779 {
 780         uint16_t nr_vec = 0;
 781         uint16_t avail_idx = vq->last_avail_idx;
 782         uint16_t max_tries, tries = 0;
 783         uint16_t buf_id = 0;
 784         uint32_t len = 0;
 785         uint16_t desc_count;
 786         uint32_t size = pkt->pkt_len + dev->vhost_hlen;
 787         uint16_t num_buffers = 0;
 788
 789         if (rxvq_is_mergeable(dev))
 790                 max_tries = vq->size - 1;
 791         else
 792                 max_tries = 1;
 793
 794         while (size > 0) {
 795                 /*
 796                  * if we tried all available ring items, and still
 797                  * can't get enough buf, it means something abnormal
 798                  * happened.
 799                  */
 800                 if (unlikely(++tries > max_tries))
 801                         return -1;
 802
 803                 if (unlikely(fill_vec_buf_packed(dev, vq,
 804                                                 avail_idx, &desc_count,
 805                                                 buf_vec, &nr_vec,
 806                                                 &buf_id, &len,
 807                                                 VHOST_ACCESS_RW) < 0))
 808                         return -1;
 809
 810                 len = RTE_MIN(len, size);
 811                 size -= len;
 812
 813                 num_buffers += 1;
 814
 815                 *nr_descs += desc_count;
 816                 avail_idx += desc_count;
 817                 if (avail_idx >= vq->size)
 818                         avail_idx -= vq->size;
 819         }
 820
 821         if (copy_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, num_buffers) < 0)
 822                 return -1;
 823
 824         return 0;
 825 }
 826
 827 static __rte_noinline uint32_t
 828 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 829         struct rte_mbuf **pkts, uint32_t count)
 830 {
 831         uint32_t pkt_idx = 0;
 832         uint16_t num_buffers;
 833         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 834         uint16_t avail_head;
 835
 836         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 837
 838         /*
 839          * The ordering between avail index and
 840          * desc reads needs to be enforced.
 841          */
 842         rte_smp_rmb();
 843
 844         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 845
 846         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 847                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 848                 uint16_t nr_vec = 0;
 849
 850                 if (unlikely(reserve_avail_buf_split(dev, vq,
 851                                                 pkt_len, buf_vec, &num_buffers,
 852                                                 avail_head, &nr_vec) < 0)) {
 853                         VHOST_LOG_DEBUG(VHOST_DATA,
 854                                 "(%d) failed to get enough desc from vring\n",
 855                                 dev->vid);
 856                         vq->shadow_used_idx -= num_buffers;
 857                         break;
 858                 }
 859
 860                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 861                         dev->vid, vq->last_avail_idx,
 862                         vq->last_avail_idx + num_buffers);
 863
 864                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
 865                                                 buf_vec, nr_vec,
 866                                                 num_buffers) < 0) {
 867                         vq->shadow_used_idx -= num_buffers;
 868                         break;
 869                 }
 870
 871                 vq->last_avail_idx += num_buffers;
 872         }
 873
 874         do_data_copy_enqueue(dev, vq);
 875
 876         if (likely(vq->shadow_used_idx)) {
 877                 flush_shadow_used_ring_split(dev, vq);
 878                 vhost_vring_call_split(dev, vq);
 879         }
 880
 881         return pkt_idx;
 882 }
 883
 884 static __rte_unused int16_t
 885 virtio_dev_rx_single_packed(struct virtio_net *dev,
 886                             struct vhost_virtqueue *vq,
 887                             struct rte_mbuf *pkt)
 888 {
 889         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 890         uint16_t nr_descs = 0;
 891
 892         rte_smp_rmb();
 893         if (unlikely(vhost_enqueue_single_packed(dev, vq, pkt, buf_vec,
 894                                                  &nr_descs) < 0)) {
 895                 VHOST_LOG_DEBUG(VHOST_DATA,
 896                                 "(%d) failed to get enough desc from vring\n",
 897                                 dev->vid);
 898                 return -1;
 899         }
 900
 901         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 902                         dev->vid, vq->last_avail_idx,
 903                         vq->last_avail_idx + nr_descs);
 904
 905         vq_inc_last_avail_packed(vq, nr_descs);
 906
 907         return 0;
 908 }
 909
 910 static __rte_noinline uint32_t
 911 virtio_dev_rx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 912         struct rte_mbuf **pkts, uint32_t count)
 913 {
 914         uint32_t pkt_idx = 0;
 915         uint16_t num_buffers;
 916         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 917
 918         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 919                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 920                 uint16_t nr_vec = 0;
 921                 uint16_t nr_descs = 0;
 922
 923                 if (unlikely(reserve_avail_buf_packed(dev, vq,
 924                                                 pkt_len, buf_vec, &nr_vec,
 925                                                 &num_buffers, &nr_descs) < 0)) {
 926                         VHOST_LOG_DEBUG(VHOST_DATA,
 927                                 "(%d) failed to get enough desc from vring\n",
 928                                 dev->vid);
 929                         vq->shadow_used_idx -= num_buffers;
 930                         break;
 931                 }
 932
 933                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 934                         dev->vid, vq->last_avail_idx,
 935                         vq->last_avail_idx + num_buffers);
 936
 937                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
 938                                                 buf_vec, nr_vec,
 939                                                 num_buffers) < 0) {
 940                         vq->shadow_used_idx -= num_buffers;
 941                         break;
 942                 }
 943
 944                 vq_inc_last_avail_packed(vq, nr_descs);
 945         }
 946
 947         do_data_copy_enqueue(dev, vq);
 948
 949         if (likely(vq->shadow_used_idx)) {
 950                 flush_shadow_used_ring_packed(dev, vq);
 951                 vhost_vring_call_packed(dev, vq);
 952         }
 953
 954         return pkt_idx;
 955 }
 956
 957 static __rte_always_inline uint32_t
 958 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 959         struct rte_mbuf **pkts, uint32_t count)
 960 {
 961         struct vhost_virtqueue *vq;
 962         uint32_t nb_tx = 0;
 963
 964         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 965         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 966                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 967                         dev->vid, __func__, queue_id);
 968                 return 0;
 969         }
 970
 971         vq = dev->virtqueue[queue_id];
 972
 973         rte_spinlock_lock(&vq->access_lock);
 974
 975         if (unlikely(vq->enabled == 0))
 976                 goto out_access_unlock;
 977
 978         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 979                 vhost_user_iotlb_rd_lock(vq);
 980
 981         if (unlikely(vq->access_ok == 0))
 982                 if (unlikely(vring_translate(dev, vq) < 0))
 983                         goto out;
 984
 985         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 986         if (count == 0)
 987                 goto out;
 988
 989         if (vq_is_packed(dev))
 990                 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
 991         else
 992                 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
 993
 994 out:
 995         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 996                 vhost_user_iotlb_rd_unlock(vq);
 997
 998 out_access_unlock:
 999         rte_spinlock_unlock(&vq->access_lock);
1000
1001         return nb_tx;
1002 }
1003
1004 uint16_t
1005 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
1006         struct rte_mbuf **pkts, uint16_t count)
1007 {
1008         struct virtio_net *dev = get_device(vid);
1009
1010         if (!dev)
1011                 return 0;
1012
1013         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1014                 RTE_LOG(ERR, VHOST_DATA,
1015                         "(%d) %s: built-in vhost net backend is disabled.\n",
1016                         dev->vid, __func__);
1017                 return 0;
1018         }
1019
1020         return virtio_dev_rx(dev, queue_id, pkts, count);
1021 }
1022
1023 static inline bool
1024 virtio_net_with_host_offload(struct virtio_net *dev)
1025 {
1026         if (dev->features &
1027                         ((1ULL << VIRTIO_NET_F_CSUM) |
1028                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
1029                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
1030                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
1031                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
1032                 return true;
1033
1034         return false;
1035 }
1036
1037 static void
1038 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
1039 {
1040         struct rte_ipv4_hdr *ipv4_hdr;
1041         struct rte_ipv6_hdr *ipv6_hdr;
1042         void *l3_hdr = NULL;
1043         struct rte_ether_hdr *eth_hdr;
1044         uint16_t ethertype;
1045
1046         eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
1047
1048         m->l2_len = sizeof(struct rte_ether_hdr);
1049         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
1050
1051         if (ethertype == RTE_ETHER_TYPE_VLAN) {
1052                 struct rte_vlan_hdr *vlan_hdr =
1053                         (struct rte_vlan_hdr *)(eth_hdr + 1);
1054
1055                 m->l2_len += sizeof(struct rte_vlan_hdr);
1056                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
1057         }
1058
1059         l3_hdr = (char *)eth_hdr + m->l2_len;
1060
1061         switch (ethertype) {
1062         case RTE_ETHER_TYPE_IPV4:
1063                 ipv4_hdr = l3_hdr;
1064                 *l4_proto = ipv4_hdr->next_proto_id;
1065                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
1066                 *l4_hdr = (char *)l3_hdr + m->l3_len;
1067                 m->ol_flags |= PKT_TX_IPV4;
1068                 break;
1069         case RTE_ETHER_TYPE_IPV6:
1070                 ipv6_hdr = l3_hdr;
1071                 *l4_proto = ipv6_hdr->proto;
1072                 m->l3_len = sizeof(struct rte_ipv6_hdr);
1073                 *l4_hdr = (char *)l3_hdr + m->l3_len;
1074                 m->ol_flags |= PKT_TX_IPV6;
1075                 break;
1076         default:
1077                 m->l3_len = 0;
1078                 *l4_proto = 0;
1079                 *l4_hdr = NULL;
1080                 break;
1081         }
1082 }
1083
1084 static __rte_always_inline void
1085 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
1086 {
1087         uint16_t l4_proto = 0;
1088         void *l4_hdr = NULL;
1089         struct rte_tcp_hdr *tcp_hdr = NULL;
1090
1091         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1092                 return;
1093
1094         parse_ethernet(m, &l4_proto, &l4_hdr);
1095         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1096                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
1097                         switch (hdr->csum_offset) {
1098                         case (offsetof(struct rte_tcp_hdr, cksum)):
1099                                 if (l4_proto == IPPROTO_TCP)
1100                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
1101                                 break;
1102                         case (offsetof(struct rte_udp_hdr, dgram_cksum)):
1103                                 if (l4_proto == IPPROTO_UDP)
1104                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
1105                                 break;
1106                         case (offsetof(struct rte_sctp_hdr, cksum)):
1107                                 if (l4_proto == IPPROTO_SCTP)
1108                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
1109                                 break;
1110                         default:
1111                                 break;
1112                         }
1113                 }
1114         }
1115
1116         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1117                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1118                 case VIRTIO_NET_HDR_GSO_TCPV4:
1119                 case VIRTIO_NET_HDR_GSO_TCPV6:
1120                         tcp_hdr = l4_hdr;
1121                         m->ol_flags |= PKT_TX_TCP_SEG;
1122                         m->tso_segsz = hdr->gso_size;
1123                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
1124                         break;
1125                 case VIRTIO_NET_HDR_GSO_UDP:
1126                         m->ol_flags |= PKT_TX_UDP_SEG;
1127                         m->tso_segsz = hdr->gso_size;
1128                         m->l4_len = sizeof(struct rte_udp_hdr);
1129                         break;
1130                 default:
1131                         RTE_LOG(WARNING, VHOST_DATA,
1132                                 "unsupported gso type %u.\n", hdr->gso_type);
1133                         break;
1134                 }
1135         }
1136 }
1137
1138 static __rte_noinline void
1139 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
1140                 struct buf_vector *buf_vec)
1141 {
1142         uint64_t len;
1143         uint64_t remain = sizeof(struct virtio_net_hdr);
1144         uint64_t src;
1145         uint64_t dst = (uint64_t)(uintptr_t)hdr;
1146
1147         while (remain) {
1148                 len = RTE_MIN(remain, buf_vec->buf_len);
1149                 src = buf_vec->buf_addr;
1150                 rte_memcpy((void *)(uintptr_t)dst,
1151                                 (void *)(uintptr_t)src, len);
1152
1153                 remain -= len;
1154                 dst += len;
1155                 buf_vec++;
1156         }
1157 }
1158
1159 static __rte_always_inline int
1160 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
1161                   struct buf_vector *buf_vec, uint16_t nr_vec,
1162                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
1163 {
1164         uint32_t buf_avail, buf_offset;
1165         uint64_t buf_addr, buf_iova, buf_len;
1166         uint32_t mbuf_avail, mbuf_offset;
1167         uint32_t cpy_len;
1168         struct rte_mbuf *cur = m, *prev = m;
1169         struct virtio_net_hdr tmp_hdr;
1170         struct virtio_net_hdr *hdr = NULL;
1171         /* A counter to avoid desc dead loop chain */
1172         uint16_t vec_idx = 0;
1173         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
1174         int error = 0;
1175
1176         buf_addr = buf_vec[vec_idx].buf_addr;
1177         buf_iova = buf_vec[vec_idx].buf_iova;
1178         buf_len = buf_vec[vec_idx].buf_len;
1179
1180         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
1181                 error = -1;
1182                 goto out;
1183         }
1184
1185         if (virtio_net_with_host_offload(dev)) {
1186                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
1187                         /*
1188                          * No luck, the virtio-net header doesn't fit
1189                          * in a contiguous virtual area.
1190                          */
1191                         copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
1192                         hdr = &tmp_hdr;
1193                 } else {
1194                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
1195                 }
1196         }
1197
1198         /*
1199          * A virtio driver normally uses at least 2 desc buffers
1200          * for Tx: the first for storing the header, and others
1201          * for storing the data.
1202          */
1203         if (unlikely(buf_len < dev->vhost_hlen)) {
1204                 buf_offset = dev->vhost_hlen - buf_len;
1205                 vec_idx++;
1206                 buf_addr = buf_vec[vec_idx].buf_addr;
1207                 buf_iova = buf_vec[vec_idx].buf_iova;
1208                 buf_len = buf_vec[vec_idx].buf_len;
1209                 buf_avail  = buf_len - buf_offset;
1210         } else if (buf_len == dev->vhost_hlen) {
1211                 if (unlikely(++vec_idx >= nr_vec))
1212                         goto out;
1213                 buf_addr = buf_vec[vec_idx].buf_addr;
1214                 buf_iova = buf_vec[vec_idx].buf_iova;
1215                 buf_len = buf_vec[vec_idx].buf_len;
1216
1217                 buf_offset = 0;
1218                 buf_avail = buf_len;
1219         } else {
1220                 buf_offset = dev->vhost_hlen;
1221                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
1222         }
1223
1224         PRINT_PACKET(dev,
1225                         (uintptr_t)(buf_addr + buf_offset),
1226                         (uint32_t)buf_avail, 0);
1227
1228         mbuf_offset = 0;
1229         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
1230         while (1) {
1231                 uint64_t hpa;
1232
1233                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1234
1235                 /*
1236                  * A desc buf might across two host physical pages that are
1237                  * not continuous. In such case (gpa_to_hpa returns 0), data
1238                  * will be copied even though zero copy is enabled.
1239                  */
1240                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
1241                                         buf_iova + buf_offset, cpy_len)))) {
1242                         cur->data_len = cpy_len;
1243                         cur->data_off = 0;
1244                         cur->buf_addr =
1245                                 (void *)(uintptr_t)(buf_addr + buf_offset);
1246                         cur->buf_iova = hpa;
1247
1248                         /*
1249                          * In zero copy mode, one mbuf can only reference data
1250                          * for one or partial of one desc buff.
1251                          */
1252                         mbuf_avail = cpy_len;
1253                 } else {
1254                         if (likely(cpy_len > MAX_BATCH_LEN ||
1255                                    vq->batch_copy_nb_elems >= vq->size ||
1256                                    (hdr && cur == m))) {
1257                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
1258                                                                    mbuf_offset),
1259                                            (void *)((uintptr_t)(buf_addr +
1260                                                            buf_offset)),
1261                                            cpy_len);
1262                         } else {
1263                                 batch_copy[vq->batch_copy_nb_elems].dst =
1264                                         rte_pktmbuf_mtod_offset(cur, void *,
1265                                                                 mbuf_offset);
1266                                 batch_copy[vq->batch_copy_nb_elems].src =
1267                                         (void *)((uintptr_t)(buf_addr +
1268                                                                 buf_offset));
1269                                 batch_copy[vq->batch_copy_nb_elems].len =
1270                                         cpy_len;
1271                                 vq->batch_copy_nb_elems++;
1272                         }
1273                 }
1274
1275                 mbuf_avail  -= cpy_len;
1276                 mbuf_offset += cpy_len;
1277                 buf_avail -= cpy_len;
1278                 buf_offset += cpy_len;
1279
1280                 /* This buf reaches to its end, get the next one */
1281                 if (buf_avail == 0) {
1282                         if (++vec_idx >= nr_vec)
1283                                 break;
1284
1285                         buf_addr = buf_vec[vec_idx].buf_addr;
1286                         buf_iova = buf_vec[vec_idx].buf_iova;
1287                         buf_len = buf_vec[vec_idx].buf_len;
1288
1289                         buf_offset = 0;
1290                         buf_avail  = buf_len;
1291
1292                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
1293                                         (uint32_t)buf_avail, 0);
1294                 }
1295
1296                 /*
1297                  * This mbuf reaches to its end, get a new one
1298                  * to hold more data.
1299                  */
1300                 if (mbuf_avail == 0) {
1301                         cur = rte_pktmbuf_alloc(mbuf_pool);
1302                         if (unlikely(cur == NULL)) {
1303                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1304                                         "allocate memory for mbuf.\n");
1305                                 error = -1;
1306                                 goto out;
1307                         }
1308                         if (unlikely(dev->dequeue_zero_copy))
1309                                 rte_mbuf_refcnt_update(cur, 1);
1310
1311                         prev->next = cur;
1312                         prev->data_len = mbuf_offset;
1313                         m->nb_segs += 1;
1314                         m->pkt_len += mbuf_offset;
1315                         prev = cur;
1316
1317                         mbuf_offset = 0;
1318                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1319                 }
1320         }
1321
1322         prev->data_len = mbuf_offset;
1323         m->pkt_len    += mbuf_offset;
1324
1325         if (hdr)
1326                 vhost_dequeue_offload(hdr, m);
1327
1328 out:
1329
1330         return error;
1331 }
1332
1333 static __rte_always_inline struct zcopy_mbuf *
1334 get_zmbuf(struct vhost_virtqueue *vq)
1335 {
1336         uint16_t i;
1337         uint16_t last;
1338         int tries = 0;
1339
1340         /* search [last_zmbuf_idx, zmbuf_size) */
1341         i = vq->last_zmbuf_idx;
1342         last = vq->zmbuf_size;
1343
1344 again:
1345         for (; i < last; i++) {
1346                 if (vq->zmbufs[i].in_use == 0) {
1347                         vq->last_zmbuf_idx = i + 1;
1348                         vq->zmbufs[i].in_use = 1;
1349                         return &vq->zmbufs[i];
1350                 }
1351         }
1352
1353         tries++;
1354         if (tries == 1) {
1355                 /* search [0, last_zmbuf_idx) */
1356                 i = 0;
1357                 last = vq->last_zmbuf_idx;
1358                 goto again;
1359         }
1360
1361         return NULL;
1362 }
1363
1364 static void
1365 virtio_dev_extbuf_free(void *addr __rte_unused, void *opaque)
1366 {
1367         rte_free(opaque);
1368 }
1369
1370 static int
1371 virtio_dev_extbuf_alloc(struct rte_mbuf *pkt, uint32_t size)
1372 {
1373         struct rte_mbuf_ext_shared_info *shinfo = NULL;
1374         uint32_t total_len = RTE_PKTMBUF_HEADROOM + size;
1375         uint16_t buf_len;
1376         rte_iova_t iova;
1377         void *buf;
1378
1379         /* Try to use pkt buffer to store shinfo to reduce the amount of memory
1380          * required, otherwise store shinfo in the new buffer.
1381          */
1382         if (rte_pktmbuf_tailroom(pkt) >= sizeof(*shinfo))
1383                 shinfo = rte_pktmbuf_mtod(pkt,
1384                                           struct rte_mbuf_ext_shared_info *);
1385         else {
1386                 total_len += sizeof(*shinfo) + sizeof(uintptr_t);
1387                 total_len = RTE_ALIGN_CEIL(total_len, sizeof(uintptr_t));
1388         }
1389
1390         if (unlikely(total_len > UINT16_MAX))
1391                 return -ENOSPC;
1392
1393         buf_len = total_len;
1394         buf = rte_malloc(NULL, buf_len, RTE_CACHE_LINE_SIZE);
1395         if (unlikely(buf == NULL))
1396                 return -ENOMEM;
1397
1398         /* Initialize shinfo */
1399         if (shinfo) {
1400                 shinfo->free_cb = virtio_dev_extbuf_free;
1401                 shinfo->fcb_opaque = buf;
1402                 rte_mbuf_ext_refcnt_set(shinfo, 1);
1403         } else {
1404                 shinfo = rte_pktmbuf_ext_shinfo_init_helper(buf, &buf_len,
1405                                               virtio_dev_extbuf_free, buf);
1406                 if (unlikely(shinfo == NULL)) {
1407                         rte_free(buf);
1408                         RTE_LOG(ERR, VHOST_DATA, "Failed to init shinfo\n");
1409                         return -1;
1410                 }
1411         }
1412
1413         iova = rte_malloc_virt2iova(buf);
1414         rte_pktmbuf_attach_extbuf(pkt, buf, iova, buf_len, shinfo);
1415         rte_pktmbuf_reset_headroom(pkt);
1416
1417         return 0;
1418 }
1419
1420 /*
1421  * Allocate a host supported pktmbuf.
1422  */
1423 static __rte_always_inline struct rte_mbuf *
1424 virtio_dev_pktmbuf_alloc(struct virtio_net *dev, struct rte_mempool *mp,
1425                          uint32_t data_len)
1426 {
1427         struct rte_mbuf *pkt = rte_pktmbuf_alloc(mp);
1428
1429         if (unlikely(pkt == NULL))
1430                 return NULL;
1431
1432         if (rte_pktmbuf_tailroom(pkt) >= data_len)
1433                 return pkt;
1434
1435         /* attach an external buffer if supported */
1436         if (dev->extbuf && !virtio_dev_extbuf_alloc(pkt, data_len))
1437                 return pkt;
1438
1439         /* check if chained buffers are allowed */
1440         if (!dev->linearbuf)
1441                 return pkt;
1442
1443         /* Data doesn't fit into the buffer and the host supports
1444          * only linear buffers
1445          */
1446         rte_pktmbuf_free(pkt);
1447
1448         return NULL;
1449 }
1450
1451 static __rte_noinline uint16_t
1452 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1453         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1454 {
1455         uint16_t i;
1456         uint16_t free_entries;
1457
1458         if (unlikely(dev->dequeue_zero_copy)) {
1459                 struct zcopy_mbuf *zmbuf, *next;
1460
1461                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1462                      zmbuf != NULL; zmbuf = next) {
1463                         next = TAILQ_NEXT(zmbuf, next);
1464
1465                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1466                                 update_shadow_used_ring_split(vq,
1467                                                 zmbuf->desc_idx, 0);
1468                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1469                                 restore_mbuf(zmbuf->mbuf);
1470                                 rte_pktmbuf_free(zmbuf->mbuf);
1471                                 put_zmbuf(zmbuf);
1472                                 vq->nr_zmbuf -= 1;
1473                         }
1474                 }
1475
1476                 if (likely(vq->shadow_used_idx)) {
1477                         flush_shadow_used_ring_split(dev, vq);
1478                         vhost_vring_call_split(dev, vq);
1479                 }
1480         }
1481
1482         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1483                         vq->last_avail_idx;
1484         if (free_entries == 0)
1485                 return 0;
1486
1487         /*
1488          * The ordering between avail index and
1489          * desc reads needs to be enforced.
1490          */
1491         rte_smp_rmb();
1492
1493         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1494
1495         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1496
1497         count = RTE_MIN(count, MAX_PKT_BURST);
1498         count = RTE_MIN(count, free_entries);
1499         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1500                         dev->vid, count);
1501
1502         for (i = 0; i < count; i++) {
1503                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1504                 uint16_t head_idx;
1505                 uint32_t buf_len;
1506                 uint16_t nr_vec = 0;
1507                 int err;
1508
1509                 if (unlikely(fill_vec_buf_split(dev, vq,
1510                                                 vq->last_avail_idx + i,
1511                                                 &nr_vec, buf_vec,
1512                                                 &head_idx, &buf_len,
1513                                                 VHOST_ACCESS_RO) < 0))
1514                         break;
1515
1516                 if (likely(dev->dequeue_zero_copy == 0))
1517                         update_shadow_used_ring_split(vq, head_idx, 0);
1518
1519                 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
1520                 if (unlikely(pkts[i] == NULL))
1521                         break;
1522
1523                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1524                                 mbuf_pool);
1525                 if (unlikely(err)) {
1526                         rte_pktmbuf_free(pkts[i]);
1527                         break;
1528                 }
1529
1530                 if (unlikely(dev->dequeue_zero_copy)) {
1531                         struct zcopy_mbuf *zmbuf;
1532
1533                         zmbuf = get_zmbuf(vq);
1534                         if (!zmbuf) {
1535                                 rte_pktmbuf_free(pkts[i]);
1536                                 break;
1537                         }
1538                         zmbuf->mbuf = pkts[i];
1539                         zmbuf->desc_idx = head_idx;
1540
1541                         /*
1542                          * Pin lock the mbuf; we will check later to see
1543                          * whether the mbuf is freed (when we are the last
1544                          * user) or not. If that's the case, we then could
1545                          * update the used ring safely.
1546                          */
1547                         rte_mbuf_refcnt_update(pkts[i], 1);
1548
1549                         vq->nr_zmbuf += 1;
1550                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1551                 }
1552         }
1553         vq->last_avail_idx += i;
1554
1555         if (likely(dev->dequeue_zero_copy == 0)) {
1556                 do_data_copy_dequeue(vq);
1557                 if (unlikely(i < count))
1558                         vq->shadow_used_idx = i;
1559                 if (likely(vq->shadow_used_idx)) {
1560                         flush_shadow_used_ring_split(dev, vq);
1561                         vhost_vring_call_split(dev, vq);
1562                 }
1563         }
1564
1565         return i;
1566 }
1567
1568 static __rte_noinline uint16_t
1569 virtio_dev_tx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
1570         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1571 {
1572         uint16_t i;
1573
1574         if (unlikely(dev->dequeue_zero_copy)) {
1575                 struct zcopy_mbuf *zmbuf, *next;
1576
1577                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1578                      zmbuf != NULL; zmbuf = next) {
1579                         next = TAILQ_NEXT(zmbuf, next);
1580
1581                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1582                                 update_shadow_used_ring_packed(vq,
1583                                                 zmbuf->desc_idx,
1584                                                 0,
1585                                                 zmbuf->desc_count);
1586
1587                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1588                                 restore_mbuf(zmbuf->mbuf);
1589                                 rte_pktmbuf_free(zmbuf->mbuf);
1590                                 put_zmbuf(zmbuf);
1591                                 vq->nr_zmbuf -= 1;
1592                         }
1593                 }
1594
1595                 if (likely(vq->shadow_used_idx)) {
1596                         flush_shadow_used_ring_packed(dev, vq);
1597                         vhost_vring_call_packed(dev, vq);
1598                 }
1599         }
1600
1601         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1602
1603         count = RTE_MIN(count, MAX_PKT_BURST);
1604         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1605                         dev->vid, count);
1606
1607         for (i = 0; i < count; i++) {
1608                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1609                 uint16_t buf_id;
1610                 uint32_t buf_len;
1611                 uint16_t desc_count, nr_vec = 0;
1612                 int err;
1613
1614                 if (unlikely(fill_vec_buf_packed(dev, vq,
1615                                                 vq->last_avail_idx, &desc_count,
1616                                                 buf_vec, &nr_vec,
1617                                                 &buf_id, &buf_len,
1618                                                 VHOST_ACCESS_RO) < 0))
1619                         break;
1620
1621                 if (likely(dev->dequeue_zero_copy == 0))
1622                         update_shadow_used_ring_packed(vq, buf_id, 0,
1623                                         desc_count);
1624
1625                 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
1626                 if (unlikely(pkts[i] == NULL))
1627                         break;
1628
1629                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1630                                 mbuf_pool);
1631                 if (unlikely(err)) {
1632                         rte_pktmbuf_free(pkts[i]);
1633                         break;
1634                 }
1635
1636                 if (unlikely(dev->dequeue_zero_copy)) {
1637                         struct zcopy_mbuf *zmbuf;
1638
1639                         zmbuf = get_zmbuf(vq);
1640                         if (!zmbuf) {
1641                                 rte_pktmbuf_free(pkts[i]);
1642                                 break;
1643                         }
1644                         zmbuf->mbuf = pkts[i];
1645                         zmbuf->desc_idx = buf_id;
1646                         zmbuf->desc_count = desc_count;
1647
1648                         /*
1649                          * Pin lock the mbuf; we will check later to see
1650                          * whether the mbuf is freed (when we are the last
1651                          * user) or not. If that's the case, we then could
1652                          * update the used ring safely.
1653                          */
1654                         rte_mbuf_refcnt_update(pkts[i], 1);
1655
1656                         vq->nr_zmbuf += 1;
1657                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1658                 }
1659
1660                 vq_inc_last_avail_packed(vq, desc_count);
1661         }
1662
1663         if (likely(dev->dequeue_zero_copy == 0)) {
1664                 do_data_copy_dequeue(vq);
1665                 if (unlikely(i < count))
1666                         vq->shadow_used_idx = i;
1667                 if (likely(vq->shadow_used_idx)) {
1668                         flush_shadow_used_ring_packed(dev, vq);
1669                         vhost_vring_call_packed(dev, vq);
1670                 }
1671         }
1672
1673         return i;
1674 }
1675
1676 uint16_t
1677 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1678         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1679 {
1680         struct virtio_net *dev;
1681         struct rte_mbuf *rarp_mbuf = NULL;
1682         struct vhost_virtqueue *vq;
1683
1684         dev = get_device(vid);
1685         if (!dev)
1686                 return 0;
1687
1688         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1689                 RTE_LOG(ERR, VHOST_DATA,
1690                         "(%d) %s: built-in vhost net backend is disabled.\n",
1691                         dev->vid, __func__);
1692                 return 0;
1693         }
1694
1695         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1696                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1697                         dev->vid, __func__, queue_id);
1698                 return 0;
1699         }
1700
1701         vq = dev->virtqueue[queue_id];
1702
1703         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
1704                 return 0;
1705
1706         if (unlikely(vq->enabled == 0)) {
1707                 count = 0;
1708                 goto out_access_unlock;
1709         }
1710
1711         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1712                 vhost_user_iotlb_rd_lock(vq);
1713
1714         if (unlikely(vq->access_ok == 0))
1715                 if (unlikely(vring_translate(dev, vq) < 0)) {
1716                         count = 0;
1717                         goto out;
1718                 }
1719
1720         /*
1721          * Construct a RARP broadcast packet, and inject it to the "pkts"
1722          * array, to looks like that guest actually send such packet.
1723          *
1724          * Check user_send_rarp() for more information.
1725          *
1726          * broadcast_rarp shares a cacheline in the virtio_net structure
1727          * with some fields that are accessed during enqueue and
1728          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1729          * result in false sharing between enqueue and dequeue.
1730          *
1731          * Prevent unnecessary false sharing by reading broadcast_rarp first
1732          * and only performing cmpset if the read indicates it is likely to
1733          * be set.
1734          */
1735         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1736                         rte_atomic16_cmpset((volatile uint16_t *)
1737                                 &dev->broadcast_rarp.cnt, 1, 0))) {
1738
1739                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
1740                 if (rarp_mbuf == NULL) {
1741                         RTE_LOG(ERR, VHOST_DATA,
1742                                 "Failed to make RARP packet.\n");
1743                         count = 0;
1744                         goto out;
1745                 }
1746                 count -= 1;
1747         }
1748
1749         if (vq_is_packed(dev))
1750                 count = virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count);
1751         else
1752                 count = virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count);
1753
1754 out:
1755         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1756                 vhost_user_iotlb_rd_unlock(vq);
1757
1758 out_access_unlock:
1759         rte_spinlock_unlock(&vq->access_lock);
1760
1761         if (unlikely(rarp_mbuf != NULL)) {
1762                 /*
1763                  * Inject it to the head of "pkts" array, so that switch's mac
1764                  * learning table will get updated first.
1765                  */
1766                 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
1767                 pkts[0] = rarp_mbuf;
1768                 count += 1;
1769         }
1770
1771         return count;
1772 }