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 rte_tcp_hdr,
 222                                                 cksum));
 223                         break;
 224                 case PKT_TX_UDP_CKSUM:
 225                         net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
 226                                                 dgram_cksum));
 227                         break;
 228                 case PKT_TX_SCTP_CKSUM:
 229                         net_hdr->csum_offset = (offsetof(struct rte_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 rte_ipv4_hdr *ipv4_hdr;
 242
 243                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_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 = vhost_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 = vhost_alloc_copy_ind_table(dev,
 458                                 vq, desc->addr, desc->len);
 459                 if (unlikely(!idescs))
 460                         return -1;
 461
 462                 descs = idescs;
 463         }
 464
 465         nr_descs =  desc->len / sizeof(struct vring_packed_desc);
 466         if (unlikely(nr_descs >= vq->size)) {
 467                 free_ind_table(idescs);
 468                 return -1;
 469         }
 470
 471         for (i = 0; i < nr_descs; i++) {
 472                 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
 473                         free_ind_table(idescs);
 474                         return -1;
 475                 }
 476
 477                 *len += descs[i].len;
 478                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 479                                                 descs[i].addr, descs[i].len,
 480                                                 perm)))
 481                         return -1;
 482         }
 483         *vec_idx = vec_id;
 484
 485         if (unlikely(!!idescs))
 486                 free_ind_table(idescs);
 487
 488         return 0;
 489 }
 490
 491 static __rte_always_inline int
 492 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 493                                 uint16_t avail_idx, uint16_t *desc_count,
 494                                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 495                                 uint16_t *buf_id, uint32_t *len, uint8_t perm)
 496 {
 497         bool wrap_counter = vq->avail_wrap_counter;
 498         struct vring_packed_desc *descs = vq->desc_packed;
 499         uint16_t vec_id = *vec_idx;
 500
 501         if (avail_idx < vq->last_avail_idx)
 502                 wrap_counter ^= 1;
 503
 504         if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
 505                 return -1;
 506
 507         /*
 508          * The ordering between desc flags and desc
 509          * content reads need to be enforced.
 510          */
 511         rte_smp_rmb();
 512
 513         *desc_count = 0;
 514         *len = 0;
 515
 516         while (1) {
 517                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 518                         return -1;
 519
 520                 if (unlikely(*desc_count >= vq->size))
 521                         return -1;
 522
 523                 *desc_count += 1;
 524                 *buf_id = descs[avail_idx].id;
 525
 526                 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
 527                         if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
 528                                                         &descs[avail_idx],
 529                                                         &vec_id, buf_vec,
 530                                                         len, perm) < 0))
 531                                 return -1;
 532                 } else {
 533                         *len += descs[avail_idx].len;
 534
 535                         if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 536                                                         descs[avail_idx].addr,
 537                                                         descs[avail_idx].len,
 538                                                         perm)))
 539                                 return -1;
 540                 }
 541
 542                 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
 543                         break;
 544
 545                 if (++avail_idx >= vq->size) {
 546                         avail_idx -= vq->size;
 547                         wrap_counter ^= 1;
 548                 }
 549         }
 550
 551         *vec_idx = vec_id;
 552
 553         return 0;
 554 }
 555
 556 /*
 557  * Returns -1 on fail, 0 on success
 558  */
 559 static inline int
 560 reserve_avail_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 561                                 uint32_t size, struct buf_vector *buf_vec,
 562                                 uint16_t *nr_vec, uint16_t *num_buffers,
 563                                 uint16_t *nr_descs)
 564 {
 565         uint16_t avail_idx;
 566         uint16_t vec_idx = 0;
 567         uint16_t max_tries, tries = 0;
 568
 569         uint16_t buf_id = 0;
 570         uint32_t len = 0;
 571         uint16_t desc_count;
 572
 573         *num_buffers = 0;
 574         avail_idx = vq->last_avail_idx;
 575
 576         if (rxvq_is_mergeable(dev))
 577                 max_tries = vq->size - 1;
 578         else
 579                 max_tries = 1;
 580
 581         while (size > 0) {
 582                 /*
 583                  * if we tried all available ring items, and still
 584                  * can't get enough buf, it means something abnormal
 585                  * happened.
 586                  */
 587                 if (unlikely(++tries > max_tries))
 588                         return -1;
 589
 590                 if (unlikely(fill_vec_buf_packed(dev, vq,
 591                                                 avail_idx, &desc_count,
 592                                                 buf_vec, &vec_idx,
 593                                                 &buf_id, &len,
 594                                                 VHOST_ACCESS_RW) < 0))
 595                         return -1;
 596
 597                 len = RTE_MIN(len, size);
 598                 update_shadow_used_ring_packed(vq, buf_id, len, desc_count);
 599                 size -= len;
 600
 601                 avail_idx += desc_count;
 602                 if (avail_idx >= vq->size)
 603                         avail_idx -= vq->size;
 604
 605                 *nr_descs += desc_count;
 606                 *num_buffers += 1;
 607         }
 608
 609         *nr_vec = vec_idx;
 610
 611         return 0;
 612 }
 613
 614 static __rte_noinline void
 615 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 616                 struct buf_vector *buf_vec,
 617                 struct virtio_net_hdr_mrg_rxbuf *hdr)
 618 {
 619         uint64_t len;
 620         uint64_t remain = dev->vhost_hlen;
 621         uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 622         uint64_t iova = buf_vec->buf_iova;
 623
 624         while (remain) {
 625                 len = RTE_MIN(remain,
 626                                 buf_vec->buf_len);
 627                 dst = buf_vec->buf_addr;
 628                 rte_memcpy((void *)(uintptr_t)dst,
 629                                 (void *)(uintptr_t)src,
 630                                 len);
 631
 632                 PRINT_PACKET(dev, (uintptr_t)dst,
 633                                 (uint32_t)len, 0);
 634                 vhost_log_cache_write(dev, vq,
 635                                 iova, len);
 636
 637                 remain -= len;
 638                 iova += len;
 639                 src += len;
 640                 buf_vec++;
 641         }
 642 }
 643
 644 static __rte_always_inline int
 645 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 646                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 647                             uint16_t nr_vec, uint16_t num_buffers)
 648 {
 649         uint32_t vec_idx = 0;
 650         uint32_t mbuf_offset, mbuf_avail;
 651         uint32_t buf_offset, buf_avail;
 652         uint64_t buf_addr, buf_iova, buf_len;
 653         uint32_t cpy_len;
 654         uint64_t hdr_addr;
 655         struct rte_mbuf *hdr_mbuf;
 656         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 657         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 658         int error = 0;
 659
 660         if (unlikely(m == NULL)) {
 661                 error = -1;
 662                 goto out;
 663         }
 664
 665         buf_addr = buf_vec[vec_idx].buf_addr;
 666         buf_iova = buf_vec[vec_idx].buf_iova;
 667         buf_len = buf_vec[vec_idx].buf_len;
 668
 669         if (nr_vec > 1)
 670                 rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
 671
 672         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 673                 error = -1;
 674                 goto out;
 675         }
 676
 677         hdr_mbuf = m;
 678         hdr_addr = buf_addr;
 679         if (unlikely(buf_len < dev->vhost_hlen))
 680                 hdr = &tmp_hdr;
 681         else
 682                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
 683
 684         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 685                 dev->vid, num_buffers);
 686
 687         if (unlikely(buf_len < dev->vhost_hlen)) {
 688                 buf_offset = dev->vhost_hlen - buf_len;
 689                 vec_idx++;
 690                 buf_addr = buf_vec[vec_idx].buf_addr;
 691                 buf_iova = buf_vec[vec_idx].buf_iova;
 692                 buf_len = buf_vec[vec_idx].buf_len;
 693                 buf_avail = buf_len - buf_offset;
 694         } else {
 695                 buf_offset = dev->vhost_hlen;
 696                 buf_avail = buf_len - dev->vhost_hlen;
 697         }
 698
 699         mbuf_avail  = rte_pktmbuf_data_len(m);
 700         mbuf_offset = 0;
 701         while (mbuf_avail != 0 || m->next != NULL) {
 702                 /* done with current buf, get the next one */
 703                 if (buf_avail == 0) {
 704                         vec_idx++;
 705                         if (unlikely(vec_idx >= nr_vec)) {
 706                                 error = -1;
 707                                 goto out;
 708                         }
 709
 710                         buf_addr = buf_vec[vec_idx].buf_addr;
 711                         buf_iova = buf_vec[vec_idx].buf_iova;
 712                         buf_len = buf_vec[vec_idx].buf_len;
 713
 714                         /* Prefetch next buffer address. */
 715                         if (vec_idx + 1 < nr_vec)
 716                                 rte_prefetch0((void *)(uintptr_t)
 717                                                 buf_vec[vec_idx + 1].buf_addr);
 718                         buf_offset = 0;
 719                         buf_avail  = buf_len;
 720                 }
 721
 722                 /* done with current mbuf, get the next one */
 723                 if (mbuf_avail == 0) {
 724                         m = m->next;
 725
 726                         mbuf_offset = 0;
 727                         mbuf_avail  = rte_pktmbuf_data_len(m);
 728                 }
 729
 730                 if (hdr_addr) {
 731                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 732                         if (rxvq_is_mergeable(dev))
 733                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
 734                                                 num_buffers);
 735
 736                         if (unlikely(hdr == &tmp_hdr)) {
 737                                 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
 738                         } else {
 739                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 740                                                 dev->vhost_hlen, 0);
 741                                 vhost_log_cache_write(dev, vq,
 742                                                 buf_vec[0].buf_iova,
 743                                                 dev->vhost_hlen);
 744                         }
 745
 746                         hdr_addr = 0;
 747                 }
 748
 749                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 750
 751                 if (likely(cpy_len > MAX_BATCH_LEN ||
 752                                         vq->batch_copy_nb_elems >= vq->size)) {
 753                         rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
 754                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 755                                 cpy_len);
 756                         vhost_log_cache_write(dev, vq, buf_iova + buf_offset,
 757                                         cpy_len);
 758                         PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
 759                                 cpy_len, 0);
 760                 } else {
 761                         batch_copy[vq->batch_copy_nb_elems].dst =
 762                                 (void *)((uintptr_t)(buf_addr + buf_offset));
 763                         batch_copy[vq->batch_copy_nb_elems].src =
 764                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 765                         batch_copy[vq->batch_copy_nb_elems].log_addr =
 766                                 buf_iova + buf_offset;
 767                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
 768                         vq->batch_copy_nb_elems++;
 769                 }
 770
 771                 mbuf_avail  -= cpy_len;
 772                 mbuf_offset += cpy_len;
 773                 buf_avail  -= cpy_len;
 774                 buf_offset += cpy_len;
 775         }
 776
 777 out:
 778
 779         return error;
 780 }
 781
 782 static __rte_noinline uint32_t
 783 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 784         struct rte_mbuf **pkts, uint32_t count)
 785 {
 786         uint32_t pkt_idx = 0;
 787         uint16_t num_buffers;
 788         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 789         uint16_t avail_head;
 790
 791         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 792
 793         /*
 794          * The ordering between avail index and
 795          * desc reads needs to be enforced.
 796          */
 797         rte_smp_rmb();
 798
 799         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 800
 801         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 802                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 803                 uint16_t nr_vec = 0;
 804
 805                 if (unlikely(reserve_avail_buf_split(dev, vq,
 806                                                 pkt_len, buf_vec, &num_buffers,
 807                                                 avail_head, &nr_vec) < 0)) {
 808                         VHOST_LOG_DEBUG(VHOST_DATA,
 809                                 "(%d) failed to get enough desc from vring\n",
 810                                 dev->vid);
 811                         vq->shadow_used_idx -= num_buffers;
 812                         break;
 813                 }
 814
 815                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
 816
 817                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 818                         dev->vid, vq->last_avail_idx,
 819                         vq->last_avail_idx + num_buffers);
 820
 821                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
 822                                                 buf_vec, nr_vec,
 823                                                 num_buffers) < 0) {
 824                         vq->shadow_used_idx -= num_buffers;
 825                         break;
 826                 }
 827
 828                 vq->last_avail_idx += num_buffers;
 829         }
 830
 831         do_data_copy_enqueue(dev, vq);
 832
 833         if (likely(vq->shadow_used_idx)) {
 834                 flush_shadow_used_ring_split(dev, vq);
 835                 vhost_vring_call_split(dev, vq);
 836         }
 837
 838         return pkt_idx;
 839 }
 840
 841 static __rte_noinline uint32_t
 842 virtio_dev_rx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 843         struct rte_mbuf **pkts, uint32_t count)
 844 {
 845         uint32_t pkt_idx = 0;
 846         uint16_t num_buffers;
 847         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 848
 849         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 850                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 851                 uint16_t nr_vec = 0;
 852                 uint16_t nr_descs = 0;
 853
 854                 if (unlikely(reserve_avail_buf_packed(dev, vq,
 855                                                 pkt_len, buf_vec, &nr_vec,
 856                                                 &num_buffers, &nr_descs) < 0)) {
 857                         VHOST_LOG_DEBUG(VHOST_DATA,
 858                                 "(%d) failed to get enough desc from vring\n",
 859                                 dev->vid);
 860                         vq->shadow_used_idx -= num_buffers;
 861                         break;
 862                 }
 863
 864                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
 865
 866                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 867                         dev->vid, vq->last_avail_idx,
 868                         vq->last_avail_idx + num_buffers);
 869
 870                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
 871                                                 buf_vec, nr_vec,
 872                                                 num_buffers) < 0) {
 873                         vq->shadow_used_idx -= num_buffers;
 874                         break;
 875                 }
 876
 877                 vq->last_avail_idx += nr_descs;
 878                 if (vq->last_avail_idx >= vq->size) {
 879                         vq->last_avail_idx -= vq->size;
 880                         vq->avail_wrap_counter ^= 1;
 881                 }
 882         }
 883
 884         do_data_copy_enqueue(dev, vq);
 885
 886         if (likely(vq->shadow_used_idx)) {
 887                 flush_shadow_used_ring_packed(dev, vq);
 888                 vhost_vring_call_packed(dev, vq);
 889         }
 890
 891         return pkt_idx;
 892 }
 893
 894 static __rte_always_inline uint32_t
 895 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 896         struct rte_mbuf **pkts, uint32_t count)
 897 {
 898         struct vhost_virtqueue *vq;
 899         uint32_t nb_tx = 0;
 900
 901         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 902         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 903                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 904                         dev->vid, __func__, queue_id);
 905                 return 0;
 906         }
 907
 908         vq = dev->virtqueue[queue_id];
 909
 910         rte_spinlock_lock(&vq->access_lock);
 911
 912         if (unlikely(vq->enabled == 0))
 913                 goto out_access_unlock;
 914
 915         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 916                 vhost_user_iotlb_rd_lock(vq);
 917
 918         if (unlikely(vq->access_ok == 0))
 919                 if (unlikely(vring_translate(dev, vq) < 0))
 920                         goto out;
 921
 922         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 923         if (count == 0)
 924                 goto out;
 925
 926         if (vq_is_packed(dev))
 927                 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
 928         else
 929                 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
 930
 931 out:
 932         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 933                 vhost_user_iotlb_rd_unlock(vq);
 934
 935 out_access_unlock:
 936         rte_spinlock_unlock(&vq->access_lock);
 937
 938         return nb_tx;
 939 }
 940
 941 uint16_t
 942 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 943         struct rte_mbuf **pkts, uint16_t count)
 944 {
 945         struct virtio_net *dev = get_device(vid);
 946
 947         if (!dev)
 948                 return 0;
 949
 950         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
 951                 RTE_LOG(ERR, VHOST_DATA,
 952                         "(%d) %s: built-in vhost net backend is disabled.\n",
 953                         dev->vid, __func__);
 954                 return 0;
 955         }
 956
 957         return virtio_dev_rx(dev, queue_id, pkts, count);
 958 }
 959
 960 static inline bool
 961 virtio_net_with_host_offload(struct virtio_net *dev)
 962 {
 963         if (dev->features &
 964                         ((1ULL << VIRTIO_NET_F_CSUM) |
 965                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
 966                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
 967                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
 968                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
 969                 return true;
 970
 971         return false;
 972 }
 973
 974 static void
 975 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 976 {
 977         struct rte_ipv4_hdr *ipv4_hdr;
 978         struct rte_ipv6_hdr *ipv6_hdr;
 979         void *l3_hdr = NULL;
 980         struct rte_ether_hdr *eth_hdr;
 981         uint16_t ethertype;
 982
 983         eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
 984
 985         m->l2_len = sizeof(struct rte_ether_hdr);
 986         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 987
 988         if (ethertype == RTE_ETHER_TYPE_VLAN) {
 989                 struct rte_vlan_hdr *vlan_hdr =
 990                         (struct rte_vlan_hdr *)(eth_hdr + 1);
 991
 992                 m->l2_len += sizeof(struct rte_vlan_hdr);
 993                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 994         }
 995
 996         l3_hdr = (char *)eth_hdr + m->l2_len;
 997
 998         switch (ethertype) {
 999         case RTE_ETHER_TYPE_IPV4:
1000                 ipv4_hdr = l3_hdr;
1001                 *l4_proto = ipv4_hdr->next_proto_id;
1002                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
1003                 *l4_hdr = (char *)l3_hdr + m->l3_len;
1004                 m->ol_flags |= PKT_TX_IPV4;
1005                 break;
1006         case RTE_ETHER_TYPE_IPV6:
1007                 ipv6_hdr = l3_hdr;
1008                 *l4_proto = ipv6_hdr->proto;
1009                 m->l3_len = sizeof(struct rte_ipv6_hdr);
1010                 *l4_hdr = (char *)l3_hdr + m->l3_len;
1011                 m->ol_flags |= PKT_TX_IPV6;
1012                 break;
1013         default:
1014                 m->l3_len = 0;
1015                 *l4_proto = 0;
1016                 *l4_hdr = NULL;
1017                 break;
1018         }
1019 }
1020
1021 static __rte_always_inline void
1022 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
1023 {
1024         uint16_t l4_proto = 0;
1025         void *l4_hdr = NULL;
1026         struct rte_tcp_hdr *tcp_hdr = NULL;
1027
1028         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1029                 return;
1030
1031         parse_ethernet(m, &l4_proto, &l4_hdr);
1032         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1033                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
1034                         switch (hdr->csum_offset) {
1035                         case (offsetof(struct rte_tcp_hdr, cksum)):
1036                                 if (l4_proto == IPPROTO_TCP)
1037                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
1038                                 break;
1039                         case (offsetof(struct rte_udp_hdr, dgram_cksum)):
1040                                 if (l4_proto == IPPROTO_UDP)
1041                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
1042                                 break;
1043                         case (offsetof(struct rte_sctp_hdr, cksum)):
1044                                 if (l4_proto == IPPROTO_SCTP)
1045                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
1046                                 break;
1047                         default:
1048                                 break;
1049                         }
1050                 }
1051         }
1052
1053         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1054                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1055                 case VIRTIO_NET_HDR_GSO_TCPV4:
1056                 case VIRTIO_NET_HDR_GSO_TCPV6:
1057                         tcp_hdr = l4_hdr;
1058                         m->ol_flags |= PKT_TX_TCP_SEG;
1059                         m->tso_segsz = hdr->gso_size;
1060                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
1061                         break;
1062                 case VIRTIO_NET_HDR_GSO_UDP:
1063                         m->ol_flags |= PKT_TX_UDP_SEG;
1064                         m->tso_segsz = hdr->gso_size;
1065                         m->l4_len = sizeof(struct rte_udp_hdr);
1066                         break;
1067                 default:
1068                         RTE_LOG(WARNING, VHOST_DATA,
1069                                 "unsupported gso type %u.\n", hdr->gso_type);
1070                         break;
1071                 }
1072         }
1073 }
1074
1075 static __rte_noinline void
1076 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
1077                 struct buf_vector *buf_vec)
1078 {
1079         uint64_t len;
1080         uint64_t remain = sizeof(struct virtio_net_hdr);
1081         uint64_t src;
1082         uint64_t dst = (uint64_t)(uintptr_t)hdr;
1083
1084         while (remain) {
1085                 len = RTE_MIN(remain, buf_vec->buf_len);
1086                 src = buf_vec->buf_addr;
1087                 rte_memcpy((void *)(uintptr_t)dst,
1088                                 (void *)(uintptr_t)src, len);
1089
1090                 remain -= len;
1091                 dst += len;
1092                 buf_vec++;
1093         }
1094 }
1095
1096 static __rte_always_inline int
1097 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
1098                   struct buf_vector *buf_vec, uint16_t nr_vec,
1099                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
1100 {
1101         uint32_t buf_avail, buf_offset;
1102         uint64_t buf_addr, buf_iova, buf_len;
1103         uint32_t mbuf_avail, mbuf_offset;
1104         uint32_t cpy_len;
1105         struct rte_mbuf *cur = m, *prev = m;
1106         struct virtio_net_hdr tmp_hdr;
1107         struct virtio_net_hdr *hdr = NULL;
1108         /* A counter to avoid desc dead loop chain */
1109         uint16_t vec_idx = 0;
1110         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
1111         int error = 0;
1112
1113         buf_addr = buf_vec[vec_idx].buf_addr;
1114         buf_iova = buf_vec[vec_idx].buf_iova;
1115         buf_len = buf_vec[vec_idx].buf_len;
1116
1117         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
1118                 error = -1;
1119                 goto out;
1120         }
1121
1122         if (likely(nr_vec > 1))
1123                 rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
1124
1125         if (virtio_net_with_host_offload(dev)) {
1126                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
1127                         /*
1128                          * No luck, the virtio-net header doesn't fit
1129                          * in a contiguous virtual area.
1130                          */
1131                         copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
1132                         hdr = &tmp_hdr;
1133                 } else {
1134                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
1135                         rte_prefetch0(hdr);
1136                 }
1137         }
1138
1139         /*
1140          * A virtio driver normally uses at least 2 desc buffers
1141          * for Tx: the first for storing the header, and others
1142          * for storing the data.
1143          */
1144         if (unlikely(buf_len < dev->vhost_hlen)) {
1145                 buf_offset = dev->vhost_hlen - buf_len;
1146                 vec_idx++;
1147                 buf_addr = buf_vec[vec_idx].buf_addr;
1148                 buf_iova = buf_vec[vec_idx].buf_iova;
1149                 buf_len = buf_vec[vec_idx].buf_len;
1150                 buf_avail  = buf_len - buf_offset;
1151         } else if (buf_len == dev->vhost_hlen) {
1152                 if (unlikely(++vec_idx >= nr_vec))
1153                         goto out;
1154                 buf_addr = buf_vec[vec_idx].buf_addr;
1155                 buf_iova = buf_vec[vec_idx].buf_iova;
1156                 buf_len = buf_vec[vec_idx].buf_len;
1157
1158                 buf_offset = 0;
1159                 buf_avail = buf_len;
1160         } else {
1161                 buf_offset = dev->vhost_hlen;
1162                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
1163         }
1164
1165         rte_prefetch0((void *)(uintptr_t)
1166                         (buf_addr + buf_offset));
1167
1168         PRINT_PACKET(dev,
1169                         (uintptr_t)(buf_addr + buf_offset),
1170                         (uint32_t)buf_avail, 0);
1171
1172         mbuf_offset = 0;
1173         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
1174         while (1) {
1175                 uint64_t hpa;
1176
1177                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1178
1179                 /*
1180                  * A desc buf might across two host physical pages that are
1181                  * not continuous. In such case (gpa_to_hpa returns 0), data
1182                  * will be copied even though zero copy is enabled.
1183                  */
1184                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
1185                                         buf_iova + buf_offset, cpy_len)))) {
1186                         cur->data_len = cpy_len;
1187                         cur->data_off = 0;
1188                         cur->buf_addr =
1189                                 (void *)(uintptr_t)(buf_addr + buf_offset);
1190                         cur->buf_iova = hpa;
1191
1192                         /*
1193                          * In zero copy mode, one mbuf can only reference data
1194                          * for one or partial of one desc buff.
1195                          */
1196                         mbuf_avail = cpy_len;
1197                 } else {
1198                         if (likely(cpy_len > MAX_BATCH_LEN ||
1199                                    vq->batch_copy_nb_elems >= vq->size ||
1200                                    (hdr && cur == m))) {
1201                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
1202                                                                    mbuf_offset),
1203                                            (void *)((uintptr_t)(buf_addr +
1204                                                            buf_offset)),
1205                                            cpy_len);
1206                         } else {
1207                                 batch_copy[vq->batch_copy_nb_elems].dst =
1208                                         rte_pktmbuf_mtod_offset(cur, void *,
1209                                                                 mbuf_offset);
1210                                 batch_copy[vq->batch_copy_nb_elems].src =
1211                                         (void *)((uintptr_t)(buf_addr +
1212                                                                 buf_offset));
1213                                 batch_copy[vq->batch_copy_nb_elems].len =
1214                                         cpy_len;
1215                                 vq->batch_copy_nb_elems++;
1216                         }
1217                 }
1218
1219                 mbuf_avail  -= cpy_len;
1220                 mbuf_offset += cpy_len;
1221                 buf_avail -= cpy_len;
1222                 buf_offset += cpy_len;
1223
1224                 /* This buf reaches to its end, get the next one */
1225                 if (buf_avail == 0) {
1226                         if (++vec_idx >= nr_vec)
1227                                 break;
1228
1229                         buf_addr = buf_vec[vec_idx].buf_addr;
1230                         buf_iova = buf_vec[vec_idx].buf_iova;
1231                         buf_len = buf_vec[vec_idx].buf_len;
1232
1233                         /*
1234                          * Prefecth desc n + 1 buffer while
1235                          * desc n buffer is processed.
1236                          */
1237                         if (vec_idx + 1 < nr_vec)
1238                                 rte_prefetch0((void *)(uintptr_t)
1239                                                 buf_vec[vec_idx + 1].buf_addr);
1240
1241                         buf_offset = 0;
1242                         buf_avail  = buf_len;
1243
1244                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
1245                                         (uint32_t)buf_avail, 0);
1246                 }
1247
1248                 /*
1249                  * This mbuf reaches to its end, get a new one
1250                  * to hold more data.
1251                  */
1252                 if (mbuf_avail == 0) {
1253                         cur = rte_pktmbuf_alloc(mbuf_pool);
1254                         if (unlikely(cur == NULL)) {
1255                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1256                                         "allocate memory for mbuf.\n");
1257                                 error = -1;
1258                                 goto out;
1259                         }
1260                         if (unlikely(dev->dequeue_zero_copy))
1261                                 rte_mbuf_refcnt_update(cur, 1);
1262
1263                         prev->next = cur;
1264                         prev->data_len = mbuf_offset;
1265                         m->nb_segs += 1;
1266                         m->pkt_len += mbuf_offset;
1267                         prev = cur;
1268
1269                         mbuf_offset = 0;
1270                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1271                 }
1272         }
1273
1274         prev->data_len = mbuf_offset;
1275         m->pkt_len    += mbuf_offset;
1276
1277         if (hdr)
1278                 vhost_dequeue_offload(hdr, m);
1279
1280 out:
1281
1282         return error;
1283 }
1284
1285 static __rte_always_inline struct zcopy_mbuf *
1286 get_zmbuf(struct vhost_virtqueue *vq)
1287 {
1288         uint16_t i;
1289         uint16_t last;
1290         int tries = 0;
1291
1292         /* search [last_zmbuf_idx, zmbuf_size) */
1293         i = vq->last_zmbuf_idx;
1294         last = vq->zmbuf_size;
1295
1296 again:
1297         for (; i < last; i++) {
1298                 if (vq->zmbufs[i].in_use == 0) {
1299                         vq->last_zmbuf_idx = i + 1;
1300                         vq->zmbufs[i].in_use = 1;
1301                         return &vq->zmbufs[i];
1302                 }
1303         }
1304
1305         tries++;
1306         if (tries == 1) {
1307                 /* search [0, last_zmbuf_idx) */
1308                 i = 0;
1309                 last = vq->last_zmbuf_idx;
1310                 goto again;
1311         }
1312
1313         return NULL;
1314 }
1315
1316 static __rte_noinline uint16_t
1317 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1318         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1319 {
1320         uint16_t i;
1321         uint16_t free_entries;
1322
1323         if (unlikely(dev->dequeue_zero_copy)) {
1324                 struct zcopy_mbuf *zmbuf, *next;
1325
1326                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1327                      zmbuf != NULL; zmbuf = next) {
1328                         next = TAILQ_NEXT(zmbuf, next);
1329
1330                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1331                                 update_shadow_used_ring_split(vq,
1332                                                 zmbuf->desc_idx, 0);
1333                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1334                                 restore_mbuf(zmbuf->mbuf);
1335                                 rte_pktmbuf_free(zmbuf->mbuf);
1336                                 put_zmbuf(zmbuf);
1337                                 vq->nr_zmbuf -= 1;
1338                         }
1339                 }
1340
1341                 if (likely(vq->shadow_used_idx)) {
1342                         flush_shadow_used_ring_split(dev, vq);
1343                         vhost_vring_call_split(dev, vq);
1344                 }
1345         }
1346
1347         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1348                         vq->last_avail_idx;
1349         if (free_entries == 0)
1350                 return 0;
1351
1352         /*
1353          * The ordering between avail index and
1354          * desc reads needs to be enforced.
1355          */
1356         rte_smp_rmb();
1357
1358         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1359
1360         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1361
1362         count = RTE_MIN(count, MAX_PKT_BURST);
1363         count = RTE_MIN(count, free_entries);
1364         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1365                         dev->vid, count);
1366
1367         for (i = 0; i < count; i++) {
1368                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1369                 uint16_t head_idx;
1370                 uint32_t dummy_len;
1371                 uint16_t nr_vec = 0;
1372                 int err;
1373
1374                 if (unlikely(fill_vec_buf_split(dev, vq,
1375                                                 vq->last_avail_idx + i,
1376                                                 &nr_vec, buf_vec,
1377                                                 &head_idx, &dummy_len,
1378                                                 VHOST_ACCESS_RO) < 0))
1379                         break;
1380
1381                 if (likely(dev->dequeue_zero_copy == 0))
1382                         update_shadow_used_ring_split(vq, head_idx, 0);
1383
1384                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
1385
1386                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1387                 if (unlikely(pkts[i] == NULL)) {
1388                         RTE_LOG(ERR, VHOST_DATA,
1389                                 "Failed to allocate memory for mbuf.\n");
1390                         break;
1391                 }
1392
1393                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1394                                 mbuf_pool);
1395                 if (unlikely(err)) {
1396                         rte_pktmbuf_free(pkts[i]);
1397                         break;
1398                 }
1399
1400                 if (unlikely(dev->dequeue_zero_copy)) {
1401                         struct zcopy_mbuf *zmbuf;
1402
1403                         zmbuf = get_zmbuf(vq);
1404                         if (!zmbuf) {
1405                                 rte_pktmbuf_free(pkts[i]);
1406                                 break;
1407                         }
1408                         zmbuf->mbuf = pkts[i];
1409                         zmbuf->desc_idx = head_idx;
1410
1411                         /*
1412                          * Pin lock the mbuf; we will check later to see
1413                          * whether the mbuf is freed (when we are the last
1414                          * user) or not. If that's the case, we then could
1415                          * update the used ring safely.
1416                          */
1417                         rte_mbuf_refcnt_update(pkts[i], 1);
1418
1419                         vq->nr_zmbuf += 1;
1420                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1421                 }
1422         }
1423         vq->last_avail_idx += i;
1424
1425         if (likely(dev->dequeue_zero_copy == 0)) {
1426                 do_data_copy_dequeue(vq);
1427                 if (unlikely(i < count))
1428                         vq->shadow_used_idx = i;
1429                 if (likely(vq->shadow_used_idx)) {
1430                         flush_shadow_used_ring_split(dev, vq);
1431                         vhost_vring_call_split(dev, vq);
1432                 }
1433         }
1434
1435         return i;
1436 }
1437
1438 static __rte_noinline uint16_t
1439 virtio_dev_tx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
1440         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1441 {
1442         uint16_t i;
1443
1444         if (unlikely(dev->dequeue_zero_copy)) {
1445                 struct zcopy_mbuf *zmbuf, *next;
1446
1447                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1448                      zmbuf != NULL; zmbuf = next) {
1449                         next = TAILQ_NEXT(zmbuf, next);
1450
1451                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1452                                 update_shadow_used_ring_packed(vq,
1453                                                 zmbuf->desc_idx,
1454                                                 0,
1455                                                 zmbuf->desc_count);
1456
1457                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1458                                 restore_mbuf(zmbuf->mbuf);
1459                                 rte_pktmbuf_free(zmbuf->mbuf);
1460                                 put_zmbuf(zmbuf);
1461                                 vq->nr_zmbuf -= 1;
1462                         }
1463                 }
1464
1465                 if (likely(vq->shadow_used_idx)) {
1466                         flush_shadow_used_ring_packed(dev, vq);
1467                         vhost_vring_call_packed(dev, vq);
1468                 }
1469         }
1470
1471         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1472
1473         count = RTE_MIN(count, MAX_PKT_BURST);
1474         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1475                         dev->vid, count);
1476
1477         for (i = 0; i < count; i++) {
1478                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1479                 uint16_t buf_id;
1480                 uint32_t dummy_len;
1481                 uint16_t desc_count, nr_vec = 0;
1482                 int err;
1483
1484                 if (unlikely(fill_vec_buf_packed(dev, vq,
1485                                                 vq->last_avail_idx, &desc_count,
1486                                                 buf_vec, &nr_vec,
1487                                                 &buf_id, &dummy_len,
1488                                                 VHOST_ACCESS_RO) < 0))
1489                         break;
1490
1491                 if (likely(dev->dequeue_zero_copy == 0))
1492                         update_shadow_used_ring_packed(vq, buf_id, 0,
1493                                         desc_count);
1494
1495                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
1496
1497                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1498                 if (unlikely(pkts[i] == NULL)) {
1499                         RTE_LOG(ERR, VHOST_DATA,
1500                                 "Failed to allocate memory for mbuf.\n");
1501                         break;
1502                 }
1503
1504                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1505                                 mbuf_pool);
1506                 if (unlikely(err)) {
1507                         rte_pktmbuf_free(pkts[i]);
1508                         break;
1509                 }
1510
1511                 if (unlikely(dev->dequeue_zero_copy)) {
1512                         struct zcopy_mbuf *zmbuf;
1513
1514                         zmbuf = get_zmbuf(vq);
1515                         if (!zmbuf) {
1516                                 rte_pktmbuf_free(pkts[i]);
1517                                 break;
1518                         }
1519                         zmbuf->mbuf = pkts[i];
1520                         zmbuf->desc_idx = buf_id;
1521                         zmbuf->desc_count = desc_count;
1522
1523                         /*
1524                          * Pin lock the mbuf; we will check later to see
1525                          * whether the mbuf is freed (when we are the last
1526                          * user) or not. If that's the case, we then could
1527                          * update the used ring safely.
1528                          */
1529                         rte_mbuf_refcnt_update(pkts[i], 1);
1530
1531                         vq->nr_zmbuf += 1;
1532                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1533                 }
1534
1535                 vq->last_avail_idx += desc_count;
1536                 if (vq->last_avail_idx >= vq->size) {
1537                         vq->last_avail_idx -= vq->size;
1538                         vq->avail_wrap_counter ^= 1;
1539                 }
1540         }
1541
1542         if (likely(dev->dequeue_zero_copy == 0)) {
1543                 do_data_copy_dequeue(vq);
1544                 if (unlikely(i < count))
1545                         vq->shadow_used_idx = i;
1546                 if (likely(vq->shadow_used_idx)) {
1547                         flush_shadow_used_ring_packed(dev, vq);
1548                         vhost_vring_call_packed(dev, vq);
1549                 }
1550         }
1551
1552         return i;
1553 }
1554
1555 uint16_t
1556 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1557         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1558 {
1559         struct virtio_net *dev;
1560         struct rte_mbuf *rarp_mbuf = NULL;
1561         struct vhost_virtqueue *vq;
1562
1563         dev = get_device(vid);
1564         if (!dev)
1565                 return 0;
1566
1567         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1568                 RTE_LOG(ERR, VHOST_DATA,
1569                         "(%d) %s: built-in vhost net backend is disabled.\n",
1570                         dev->vid, __func__);
1571                 return 0;
1572         }
1573
1574         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1575                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1576                         dev->vid, __func__, queue_id);
1577                 return 0;
1578         }
1579
1580         vq = dev->virtqueue[queue_id];
1581
1582         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
1583                 return 0;
1584
1585         if (unlikely(vq->enabled == 0)) {
1586                 count = 0;
1587                 goto out_access_unlock;
1588         }
1589
1590         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1591                 vhost_user_iotlb_rd_lock(vq);
1592
1593         if (unlikely(vq->access_ok == 0))
1594                 if (unlikely(vring_translate(dev, vq) < 0)) {
1595                         count = 0;
1596                         goto out;
1597                 }
1598
1599         /*
1600          * Construct a RARP broadcast packet, and inject it to the "pkts"
1601          * array, to looks like that guest actually send such packet.
1602          *
1603          * Check user_send_rarp() for more information.
1604          *
1605          * broadcast_rarp shares a cacheline in the virtio_net structure
1606          * with some fields that are accessed during enqueue and
1607          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1608          * result in false sharing between enqueue and dequeue.
1609          *
1610          * Prevent unnecessary false sharing by reading broadcast_rarp first
1611          * and only performing cmpset if the read indicates it is likely to
1612          * be set.
1613          */
1614         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1615                         rte_atomic16_cmpset((volatile uint16_t *)
1616                                 &dev->broadcast_rarp.cnt, 1, 0))) {
1617
1618                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
1619                 if (rarp_mbuf == NULL) {
1620                         RTE_LOG(ERR, VHOST_DATA,
1621                                 "Failed to make RARP packet.\n");
1622                         count = 0;
1623                         goto out;
1624                 }
1625                 count -= 1;
1626         }
1627
1628         if (vq_is_packed(dev))
1629                 count = virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count);
1630         else
1631                 count = virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count);
1632
1633 out:
1634         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1635                 vhost_user_iotlb_rd_unlock(vq);
1636
1637 out_access_unlock:
1638         rte_spinlock_unlock(&vq->access_lock);
1639
1640         if (unlikely(rarp_mbuf != NULL)) {
1641                 /*
1642                  * Inject it to the head of "pkts" array, so that switch's mac
1643                  * learning table will get updated first.
1644                  */
1645                 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
1646                 pkts[0] = rarp_mbuf;
1647                 count += 1;
1648         }
1649
1650         return count;
1651 }