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 struct vring_desc *
  41 alloc_copy_ind_table(struct virtio_net *dev, struct vhost_virtqueue *vq,
  42                                          struct vring_desc *desc)
  43 {
  44         struct vring_desc *idesc;
  45         uint64_t src, dst;
  46         uint64_t len, remain = desc->len;
  47         uint64_t desc_addr = desc->addr;
  48
  49         idesc = rte_malloc(__func__, desc->len, 0);
  50         if (unlikely(!idesc))
  51                 return 0;
  52
  53         dst = (uint64_t)(uintptr_t)idesc;
  54
  55         while (remain) {
  56                 len = remain;
  57                 src = vhost_iova_to_vva(dev, vq, desc_addr, &len,
  58                                 VHOST_ACCESS_RO);
  59                 if (unlikely(!src || !len)) {
  60                         rte_free(idesc);
  61                         return 0;
  62                 }
  63
  64                 rte_memcpy((void *)(uintptr_t)dst, (void *)(uintptr_t)src, len);
  65
  66                 remain -= len;
  67                 dst += len;
  68                 desc_addr += len;
  69         }
  70
  71         return idesc;
  72 }
  73
  74 static __rte_always_inline void
  75 free_ind_table(struct vring_desc *idesc)
  76 {
  77         rte_free(idesc);
  78 }
  79
  80 static __rte_always_inline void
  81 do_flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
  82                           uint16_t to, uint16_t from, uint16_t size)
  83 {
  84         rte_memcpy(&vq->used->ring[to],
  85                         &vq->shadow_used_ring[from],
  86                         size * sizeof(struct vring_used_elem));
  87         vhost_log_cache_used_vring(dev, vq,
  88                         offsetof(struct vring_used, ring[to]),
  89                         size * sizeof(struct vring_used_elem));
  90 }
  91
  92 static __rte_always_inline void
  93 flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq)
  94 {
  95         uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
  96
  97         if (used_idx + vq->shadow_used_idx <= vq->size) {
  98                 do_flush_shadow_used_ring(dev, vq, used_idx, 0,
  99                                           vq->shadow_used_idx);
 100         } else {
 101                 uint16_t size;
 102
 103                 /* update used ring interval [used_idx, vq->size] */
 104                 size = vq->size - used_idx;
 105                 do_flush_shadow_used_ring(dev, vq, used_idx, 0, size);
 106
 107                 /* update the left half used ring interval [0, left_size] */
 108                 do_flush_shadow_used_ring(dev, vq, 0, size,
 109                                           vq->shadow_used_idx - size);
 110         }
 111         vq->last_used_idx += vq->shadow_used_idx;
 112
 113         rte_smp_wmb();
 114
 115         vhost_log_cache_sync(dev, vq);
 116
 117         *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
 118         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 119                 sizeof(vq->used->idx));
 120 }
 121
 122 static __rte_always_inline void
 123 update_shadow_used_ring(struct vhost_virtqueue *vq,
 124                          uint16_t desc_idx, uint16_t len)
 125 {
 126         uint16_t i = vq->shadow_used_idx++;
 127
 128         vq->shadow_used_ring[i].id  = desc_idx;
 129         vq->shadow_used_ring[i].len = len;
 130 }
 131
 132 static inline void
 133 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
 134 {
 135         struct batch_copy_elem *elem = vq->batch_copy_elems;
 136         uint16_t count = vq->batch_copy_nb_elems;
 137         int i;
 138
 139         for (i = 0; i < count; i++) {
 140                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 141                 vhost_log_cache_write(dev, vq, elem[i].log_addr, elem[i].len);
 142                 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
 143         }
 144 }
 145
 146 static inline void
 147 do_data_copy_dequeue(struct vhost_virtqueue *vq)
 148 {
 149         struct batch_copy_elem *elem = vq->batch_copy_elems;
 150         uint16_t count = vq->batch_copy_nb_elems;
 151         int i;
 152
 153         for (i = 0; i < count; i++)
 154                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 155 }
 156
 157 /* avoid write operation when necessary, to lessen cache issues */
 158 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 159         if ((var) != (val))                     \
 160                 (var) = (val);                  \
 161 } while (0)
 162
 163 static __rte_always_inline void
 164 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 165 {
 166         uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
 167
 168         if (m_buf->ol_flags & PKT_TX_TCP_SEG)
 169                 csum_l4 |= PKT_TX_TCP_CKSUM;
 170
 171         if (csum_l4) {
 172                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 173                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 174
 175                 switch (csum_l4) {
 176                 case PKT_TX_TCP_CKSUM:
 177                         net_hdr->csum_offset = (offsetof(struct tcp_hdr,
 178                                                 cksum));
 179                         break;
 180                 case PKT_TX_UDP_CKSUM:
 181                         net_hdr->csum_offset = (offsetof(struct udp_hdr,
 182                                                 dgram_cksum));
 183                         break;
 184                 case PKT_TX_SCTP_CKSUM:
 185                         net_hdr->csum_offset = (offsetof(struct sctp_hdr,
 186                                                 cksum));
 187                         break;
 188                 }
 189         } else {
 190                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 191                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 192                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 193         }
 194
 195         /* IP cksum verification cannot be bypassed, then calculate here */
 196         if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
 197                 struct ipv4_hdr *ipv4_hdr;
 198
 199                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct ipv4_hdr *,
 200                                                    m_buf->l2_len);
 201                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 202         }
 203
 204         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 205                 if (m_buf->ol_flags & PKT_TX_IPV4)
 206                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 207                 else
 208                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 209                 net_hdr->gso_size = m_buf->tso_segsz;
 210                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 211                                         + m_buf->l4_len;
 212         } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
 213                 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 214                 net_hdr->gso_size = m_buf->tso_segsz;
 215                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
 216                         m_buf->l4_len;
 217         } else {
 218                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 219                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 220                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 221         }
 222 }
 223
 224 static __rte_always_inline int
 225 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 226                          uint32_t avail_idx, uint32_t *vec_idx,
 227                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 228                          uint16_t *desc_chain_len, uint8_t perm)
 229 {
 230         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 231         uint32_t vec_id = *vec_idx;
 232         uint32_t len    = 0;
 233         uint64_t dlen, desc_avail, desc_iova;
 234         struct vring_desc *descs = vq->desc;
 235         struct vring_desc *idesc = NULL;
 236
 237         *desc_chain_head = idx;
 238
 239         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 240                 dlen = vq->desc[idx].len;
 241                 descs = (struct vring_desc *)(uintptr_t)
 242                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 243                                                 &dlen,
 244                                                 VHOST_ACCESS_RO);
 245                 if (unlikely(!descs))
 246                         return -1;
 247
 248                 if (unlikely(dlen < vq->desc[idx].len)) {
 249                         /*
 250                          * The indirect desc table is not contiguous
 251                          * in process VA space, we have to copy it.
 252                          */
 253                         idesc = alloc_copy_ind_table(dev, vq, &vq->desc[idx]);
 254                         if (unlikely(!idesc))
 255                                 return -1;
 256
 257                         descs = idesc;
 258                 }
 259
 260                 idx = 0;
 261         }
 262
 263         while (1) {
 264                 if (unlikely(idx >= vq->size)) {
 265                         free_ind_table(idesc);
 266                         return -1;
 267                 }
 268
 269
 270                 len += descs[idx].len;
 271                 desc_avail = descs[idx].len;
 272                 desc_iova = descs[idx].addr;
 273
 274                 while (desc_avail) {
 275                         uint64_t desc_addr;
 276                         uint64_t desc_chunck_len = desc_avail;
 277
 278                         if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
 279                                 free_ind_table(idesc);
 280                                 return -1;
 281                         }
 282
 283                         desc_addr = vhost_iova_to_vva(dev, vq,
 284                                         desc_iova,
 285                                         &desc_chunck_len,
 286                                         perm);
 287                         if (unlikely(!desc_addr)) {
 288                                 free_ind_table(idesc);
 289                                 return -1;
 290                         }
 291
 292                         buf_vec[vec_id].buf_iova = desc_iova;
 293                         buf_vec[vec_id].buf_addr = desc_addr;
 294                         buf_vec[vec_id].buf_len  = desc_chunck_len;
 295                         buf_vec[vec_id].desc_idx = idx;
 296
 297                         desc_avail -= desc_chunck_len;
 298                         desc_iova += desc_chunck_len;
 299                         vec_id++;
 300                 }
 301
 302                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 303                         break;
 304
 305                 idx = descs[idx].next;
 306         }
 307
 308         *desc_chain_len = len;
 309         *vec_idx = vec_id;
 310
 311         if (unlikely(!!idesc))
 312                 free_ind_table(idesc);
 313
 314         return 0;
 315 }
 316
 317 /*
 318  * Returns -1 on fail, 0 on success
 319  */
 320 static inline int
 321 reserve_avail_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 322                                 uint32_t size, struct buf_vector *buf_vec,
 323                                 uint16_t *num_buffers, uint16_t avail_head,
 324                                 uint16_t *nr_vec)
 325 {
 326         uint16_t cur_idx;
 327         uint32_t vec_idx = 0;
 328         uint16_t max_tries, tries = 0;
 329
 330         uint16_t head_idx = 0;
 331         uint16_t len = 0;
 332
 333         *num_buffers = 0;
 334         cur_idx  = vq->last_avail_idx;
 335
 336         if (rxvq_is_mergeable(dev))
 337                 max_tries = vq->size;
 338         else
 339                 max_tries = 1;
 340
 341         while (size > 0) {
 342                 if (unlikely(cur_idx == avail_head))
 343                         return -1;
 344
 345                 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
 346                                                 &head_idx, &len,
 347                                                 VHOST_ACCESS_RW) < 0))
 348                         return -1;
 349                 len = RTE_MIN(len, size);
 350                 update_shadow_used_ring(vq, head_idx, len);
 351                 size -= len;
 352
 353                 cur_idx++;
 354                 tries++;
 355                 *num_buffers += 1;
 356
 357                 /*
 358                  * if we tried all available ring items, and still
 359                  * can't get enough buf, it means something abnormal
 360                  * happened.
 361                  */
 362                 if (unlikely(tries > max_tries))
 363                         return -1;
 364         }
 365
 366         *nr_vec = vec_idx;
 367
 368         return 0;
 369 }
 370
 371 static __rte_always_inline int
 372 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 373                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 374                             uint16_t nr_vec, uint16_t num_buffers)
 375 {
 376         uint32_t vec_idx = 0;
 377         uint32_t mbuf_offset, mbuf_avail;
 378         uint32_t buf_offset, buf_avail;
 379         uint64_t buf_addr, buf_iova, buf_len;
 380         uint32_t cpy_len;
 381         uint64_t hdr_addr;
 382         struct rte_mbuf *hdr_mbuf;
 383         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 384         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 385         int error = 0;
 386
 387         if (unlikely(m == NULL)) {
 388                 error = -1;
 389                 goto out;
 390         }
 391
 392         buf_addr = buf_vec[vec_idx].buf_addr;
 393         buf_iova = buf_vec[vec_idx].buf_iova;
 394         buf_len = buf_vec[vec_idx].buf_len;
 395
 396         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 397                 error = -1;
 398                 goto out;
 399         }
 400
 401         hdr_mbuf = m;
 402         hdr_addr = buf_addr;
 403         if (unlikely(buf_len < dev->vhost_hlen))
 404                 hdr = &tmp_hdr;
 405         else
 406                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
 407         rte_prefetch0((void *)(uintptr_t)hdr_addr);
 408
 409         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 410                 dev->vid, num_buffers);
 411
 412         if (unlikely(buf_len < dev->vhost_hlen)) {
 413                 buf_offset = dev->vhost_hlen - buf_len;
 414                 vec_idx++;
 415                 buf_addr = buf_vec[vec_idx].buf_addr;
 416                 buf_iova = buf_vec[vec_idx].buf_iova;
 417                 buf_len = buf_vec[vec_idx].buf_len;
 418                 buf_avail = buf_len - buf_offset;
 419         } else {
 420                 buf_offset = dev->vhost_hlen;
 421                 buf_avail = buf_len - dev->vhost_hlen;
 422         }
 423
 424         mbuf_avail  = rte_pktmbuf_data_len(m);
 425         mbuf_offset = 0;
 426         while (mbuf_avail != 0 || m->next != NULL) {
 427                 /* done with current buf, get the next one */
 428                 if (buf_avail == 0) {
 429                         vec_idx++;
 430                         if (unlikely(vec_idx >= nr_vec)) {
 431                                 error = -1;
 432                                 goto out;
 433                         }
 434
 435                         buf_addr = buf_vec[vec_idx].buf_addr;
 436                         buf_iova = buf_vec[vec_idx].buf_iova;
 437                         buf_len = buf_vec[vec_idx].buf_len;
 438
 439                         /* Prefetch buffer address. */
 440                         rte_prefetch0((void *)(uintptr_t)buf_addr);
 441                         buf_offset = 0;
 442                         buf_avail  = buf_len;
 443                 }
 444
 445                 /* done with current mbuf, get the next one */
 446                 if (mbuf_avail == 0) {
 447                         m = m->next;
 448
 449                         mbuf_offset = 0;
 450                         mbuf_avail  = rte_pktmbuf_data_len(m);
 451                 }
 452
 453                 if (hdr_addr) {
 454                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 455                         if (rxvq_is_mergeable(dev))
 456                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
 457                                                 num_buffers);
 458
 459                         if (unlikely(hdr == &tmp_hdr)) {
 460                                 uint64_t len;
 461                                 uint64_t remain = dev->vhost_hlen;
 462                                 uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 463                                 uint64_t iova = buf_vec[0].buf_iova;
 464                                 uint16_t hdr_vec_idx = 0;
 465
 466                                 while (remain) {
 467                                         len = remain;
 468                                         dst = buf_vec[hdr_vec_idx].buf_addr;
 469                                         rte_memcpy((void *)(uintptr_t)dst,
 470                                                         (void *)(uintptr_t)src,
 471                                                         len);
 472
 473                                         PRINT_PACKET(dev, (uintptr_t)dst,
 474                                                         (uint32_t)len, 0);
 475                                         vhost_log_cache_write(dev, vq,
 476                                                         iova, len);
 477
 478                                         remain -= len;
 479                                         iova += len;
 480                                         src += len;
 481                                         hdr_vec_idx++;
 482                                 }
 483                         } else {
 484                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 485                                                 dev->vhost_hlen, 0);
 486                                 vhost_log_cache_write(dev, vq,
 487                                                 buf_vec[0].buf_iova,
 488                                                 dev->vhost_hlen);
 489                         }
 490
 491                         hdr_addr = 0;
 492                 }
 493
 494                 cpy_len = RTE_MIN(buf_len, mbuf_avail);
 495
 496                 if (likely(cpy_len > MAX_BATCH_LEN ||
 497                                         vq->batch_copy_nb_elems >= vq->size)) {
 498                         rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
 499                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 500                                 cpy_len);
 501                         vhost_log_cache_write(dev, vq, buf_iova + buf_offset,
 502                                         cpy_len);
 503                         PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
 504                                 cpy_len, 0);
 505                 } else {
 506                         batch_copy[vq->batch_copy_nb_elems].dst =
 507                                 (void *)((uintptr_t)(buf_addr + buf_offset));
 508                         batch_copy[vq->batch_copy_nb_elems].src =
 509                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 510                         batch_copy[vq->batch_copy_nb_elems].log_addr =
 511                                 buf_iova + buf_offset;
 512                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
 513                         vq->batch_copy_nb_elems++;
 514                 }
 515
 516                 mbuf_avail  -= cpy_len;
 517                 mbuf_offset += cpy_len;
 518                 buf_avail  -= cpy_len;
 519                 buf_offset += cpy_len;
 520         }
 521
 522 out:
 523
 524         return error;
 525 }
 526
 527 static __rte_always_inline uint32_t
 528 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 529         struct rte_mbuf **pkts, uint32_t count)
 530 {
 531         struct vhost_virtqueue *vq;
 532         uint32_t pkt_idx = 0;
 533         uint16_t num_buffers;
 534         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 535         uint16_t avail_head;
 536
 537         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 538         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 539                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 540                         dev->vid, __func__, queue_id);
 541                 return 0;
 542         }
 543
 544         vq = dev->virtqueue[queue_id];
 545
 546         rte_spinlock_lock(&vq->access_lock);
 547
 548         if (unlikely(vq->enabled == 0))
 549                 goto out_access_unlock;
 550
 551         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 552                 vhost_user_iotlb_rd_lock(vq);
 553
 554         if (unlikely(vq->access_ok == 0))
 555                 if (unlikely(vring_translate(dev, vq) < 0))
 556                         goto out;
 557
 558         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 559         if (count == 0)
 560                 goto out;
 561
 562         vq->batch_copy_nb_elems = 0;
 563
 564         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 565
 566         vq->shadow_used_idx = 0;
 567         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 568         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 569                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 570                 uint16_t nr_vec = 0;
 571
 572                 if (unlikely(reserve_avail_buf(dev, vq,
 573                                                 pkt_len, buf_vec, &num_buffers,
 574                                                 avail_head, &nr_vec) < 0)) {
 575                         VHOST_LOG_DEBUG(VHOST_DATA,
 576                                 "(%d) failed to get enough desc from vring\n",
 577                                 dev->vid);
 578                         vq->shadow_used_idx -= num_buffers;
 579                         break;
 580                 }
 581
 582                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 583                         dev->vid, vq->last_avail_idx,
 584                         vq->last_avail_idx + num_buffers);
 585
 586                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
 587                                                 buf_vec, nr_vec,
 588                                                 num_buffers) < 0) {
 589                         vq->shadow_used_idx -= num_buffers;
 590                         break;
 591                 }
 592
 593                 vq->last_avail_idx += num_buffers;
 594         }
 595
 596         do_data_copy_enqueue(dev, vq);
 597
 598         if (likely(vq->shadow_used_idx)) {
 599                 flush_shadow_used_ring(dev, vq);
 600                 vhost_vring_call(dev, vq);
 601         }
 602
 603 out:
 604         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 605                 vhost_user_iotlb_rd_unlock(vq);
 606
 607 out_access_unlock:
 608         rte_spinlock_unlock(&vq->access_lock);
 609
 610         return pkt_idx;
 611 }
 612
 613 uint16_t
 614 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 615         struct rte_mbuf **pkts, uint16_t count)
 616 {
 617         struct virtio_net *dev = get_device(vid);
 618
 619         if (!dev)
 620                 return 0;
 621
 622         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
 623                 RTE_LOG(ERR, VHOST_DATA,
 624                         "(%d) %s: built-in vhost net backend is disabled.\n",
 625                         dev->vid, __func__);
 626                 return 0;
 627         }
 628
 629         return virtio_dev_rx(dev, queue_id, pkts, count);
 630 }
 631
 632 static inline bool
 633 virtio_net_with_host_offload(struct virtio_net *dev)
 634 {
 635         if (dev->features &
 636                         ((1ULL << VIRTIO_NET_F_CSUM) |
 637                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
 638                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
 639                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
 640                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
 641                 return true;
 642
 643         return false;
 644 }
 645
 646 static void
 647 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 648 {
 649         struct ipv4_hdr *ipv4_hdr;
 650         struct ipv6_hdr *ipv6_hdr;
 651         void *l3_hdr = NULL;
 652         struct ether_hdr *eth_hdr;
 653         uint16_t ethertype;
 654
 655         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 656
 657         m->l2_len = sizeof(struct ether_hdr);
 658         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 659
 660         if (ethertype == ETHER_TYPE_VLAN) {
 661                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 662
 663                 m->l2_len += sizeof(struct vlan_hdr);
 664                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 665         }
 666
 667         l3_hdr = (char *)eth_hdr + m->l2_len;
 668
 669         switch (ethertype) {
 670         case ETHER_TYPE_IPv4:
 671                 ipv4_hdr = l3_hdr;
 672                 *l4_proto = ipv4_hdr->next_proto_id;
 673                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 674                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 675                 m->ol_flags |= PKT_TX_IPV4;
 676                 break;
 677         case ETHER_TYPE_IPv6:
 678                 ipv6_hdr = l3_hdr;
 679                 *l4_proto = ipv6_hdr->proto;
 680                 m->l3_len = sizeof(struct ipv6_hdr);
 681                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 682                 m->ol_flags |= PKT_TX_IPV6;
 683                 break;
 684         default:
 685                 m->l3_len = 0;
 686                 *l4_proto = 0;
 687                 *l4_hdr = NULL;
 688                 break;
 689         }
 690 }
 691
 692 static __rte_always_inline void
 693 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 694 {
 695         uint16_t l4_proto = 0;
 696         void *l4_hdr = NULL;
 697         struct tcp_hdr *tcp_hdr = NULL;
 698
 699         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 700                 return;
 701
 702         parse_ethernet(m, &l4_proto, &l4_hdr);
 703         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 704                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 705                         switch (hdr->csum_offset) {
 706                         case (offsetof(struct tcp_hdr, cksum)):
 707                                 if (l4_proto == IPPROTO_TCP)
 708                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 709                                 break;
 710                         case (offsetof(struct udp_hdr, dgram_cksum)):
 711                                 if (l4_proto == IPPROTO_UDP)
 712                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 713                                 break;
 714                         case (offsetof(struct sctp_hdr, cksum)):
 715                                 if (l4_proto == IPPROTO_SCTP)
 716                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 717                                 break;
 718                         default:
 719                                 break;
 720                         }
 721                 }
 722         }
 723
 724         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 725                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 726                 case VIRTIO_NET_HDR_GSO_TCPV4:
 727                 case VIRTIO_NET_HDR_GSO_TCPV6:
 728                         tcp_hdr = l4_hdr;
 729                         m->ol_flags |= PKT_TX_TCP_SEG;
 730                         m->tso_segsz = hdr->gso_size;
 731                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 732                         break;
 733                 case VIRTIO_NET_HDR_GSO_UDP:
 734                         m->ol_flags |= PKT_TX_UDP_SEG;
 735                         m->tso_segsz = hdr->gso_size;
 736                         m->l4_len = sizeof(struct udp_hdr);
 737                         break;
 738                 default:
 739                         RTE_LOG(WARNING, VHOST_DATA,
 740                                 "unsupported gso type %u.\n", hdr->gso_type);
 741                         break;
 742                 }
 743         }
 744 }
 745
 746 static __rte_always_inline void
 747 put_zmbuf(struct zcopy_mbuf *zmbuf)
 748 {
 749         zmbuf->in_use = 0;
 750 }
 751
 752 static __rte_always_inline int
 753 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 754                   struct buf_vector *buf_vec, uint16_t nr_vec,
 755                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
 756 {
 757         uint32_t buf_avail, buf_offset;
 758         uint64_t buf_addr, buf_iova, buf_len;
 759         uint32_t mbuf_avail, mbuf_offset;
 760         uint32_t cpy_len;
 761         struct rte_mbuf *cur = m, *prev = m;
 762         struct virtio_net_hdr tmp_hdr;
 763         struct virtio_net_hdr *hdr = NULL;
 764         /* A counter to avoid desc dead loop chain */
 765         uint16_t vec_idx = 0;
 766         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 767         int error = 0;
 768
 769         buf_addr = buf_vec[vec_idx].buf_addr;
 770         buf_iova = buf_vec[vec_idx].buf_iova;
 771         buf_len = buf_vec[vec_idx].buf_len;
 772
 773         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 774                 error = -1;
 775                 goto out;
 776         }
 777
 778         if (likely(nr_vec > 1))
 779                 rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
 780
 781         if (virtio_net_with_host_offload(dev)) {
 782                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
 783                         uint64_t len;
 784                         uint64_t remain = sizeof(struct virtio_net_hdr);
 785                         uint64_t src;
 786                         uint64_t dst = (uint64_t)(uintptr_t)&tmp_hdr;
 787                         uint16_t hdr_vec_idx = 0;
 788
 789                         /*
 790                          * No luck, the virtio-net header doesn't fit
 791                          * in a contiguous virtual area.
 792                          */
 793                         while (remain) {
 794                                 len = remain;
 795                                 src = buf_vec[hdr_vec_idx].buf_addr;
 796                                 rte_memcpy((void *)(uintptr_t)dst,
 797                                                    (void *)(uintptr_t)src, len);
 798
 799                                 remain -= len;
 800                                 dst += len;
 801                                 hdr_vec_idx++;
 802                         }
 803
 804                         hdr = &tmp_hdr;
 805                 } else {
 806                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
 807                         rte_prefetch0(hdr);
 808                 }
 809         }
 810
 811         /*
 812          * A virtio driver normally uses at least 2 desc buffers
 813          * for Tx: the first for storing the header, and others
 814          * for storing the data.
 815          */
 816         if (unlikely(buf_len < dev->vhost_hlen)) {
 817                 buf_offset = dev->vhost_hlen - buf_len;
 818                 vec_idx++;
 819                 buf_addr = buf_vec[vec_idx].buf_addr;
 820                 buf_iova = buf_vec[vec_idx].buf_iova;
 821                 buf_len = buf_vec[vec_idx].buf_len;
 822                 buf_avail  = buf_len - buf_offset;
 823         } else if (buf_len == dev->vhost_hlen) {
 824                 if (unlikely(++vec_idx >= nr_vec))
 825                         goto out;
 826                 buf_addr = buf_vec[vec_idx].buf_addr;
 827                 buf_iova = buf_vec[vec_idx].buf_iova;
 828                 buf_len = buf_vec[vec_idx].buf_len;
 829
 830                 buf_offset = 0;
 831                 buf_avail = buf_len;
 832         } else {
 833                 buf_offset = dev->vhost_hlen;
 834                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
 835         }
 836
 837         rte_prefetch0((void *)(uintptr_t)
 838                         (buf_addr + buf_offset));
 839
 840         PRINT_PACKET(dev,
 841                         (uintptr_t)(buf_addr + buf_offset),
 842                         (uint32_t)buf_avail, 0);
 843
 844         mbuf_offset = 0;
 845         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 846         while (1) {
 847                 uint64_t hpa;
 848
 849                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 850
 851                 /*
 852                  * A desc buf might across two host physical pages that are
 853                  * not continuous. In such case (gpa_to_hpa returns 0), data
 854                  * will be copied even though zero copy is enabled.
 855                  */
 856                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
 857                                         buf_iova + buf_offset, cpy_len)))) {
 858                         cur->data_len = cpy_len;
 859                         cur->data_off = 0;
 860                         cur->buf_addr =
 861                                 (void *)(uintptr_t)(buf_addr + buf_offset);
 862                         cur->buf_iova = hpa;
 863
 864                         /*
 865                          * In zero copy mode, one mbuf can only reference data
 866                          * for one or partial of one desc buff.
 867                          */
 868                         mbuf_avail = cpy_len;
 869                 } else {
 870                         if (likely(cpy_len > MAX_BATCH_LEN ||
 871                                    vq->batch_copy_nb_elems >= vq->size ||
 872                                    (hdr && cur == m))) {
 873                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
 874                                                                    mbuf_offset),
 875                                            (void *)((uintptr_t)(buf_addr +
 876                                                            buf_offset)),
 877                                            cpy_len);
 878                         } else {
 879                                 batch_copy[vq->batch_copy_nb_elems].dst =
 880                                         rte_pktmbuf_mtod_offset(cur, void *,
 881                                                                 mbuf_offset);
 882                                 batch_copy[vq->batch_copy_nb_elems].src =
 883                                         (void *)((uintptr_t)(buf_addr +
 884                                                                 buf_offset));
 885                                 batch_copy[vq->batch_copy_nb_elems].len =
 886                                         cpy_len;
 887                                 vq->batch_copy_nb_elems++;
 888                         }
 889                 }
 890
 891                 mbuf_avail  -= cpy_len;
 892                 mbuf_offset += cpy_len;
 893                 buf_avail -= cpy_len;
 894                 buf_offset += cpy_len;
 895
 896                 /* This buf reaches to its end, get the next one */
 897                 if (buf_avail == 0) {
 898                         if (++vec_idx >= nr_vec)
 899                                 break;
 900
 901                         buf_addr = buf_vec[vec_idx].buf_addr;
 902                         buf_iova = buf_vec[vec_idx].buf_iova;
 903                         buf_len = buf_vec[vec_idx].buf_len;
 904
 905                         rte_prefetch0((void *)(uintptr_t)buf_addr);
 906
 907                         buf_offset = 0;
 908                         buf_avail  = buf_len;
 909
 910                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
 911                                         (uint32_t)buf_avail, 0);
 912                 }
 913
 914                 /*
 915                  * This mbuf reaches to its end, get a new one
 916                  * to hold more data.
 917                  */
 918                 if (mbuf_avail == 0) {
 919                         cur = rte_pktmbuf_alloc(mbuf_pool);
 920                         if (unlikely(cur == NULL)) {
 921                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
 922                                         "allocate memory for mbuf.\n");
 923                                 error = -1;
 924                                 goto out;
 925                         }
 926                         if (unlikely(dev->dequeue_zero_copy))
 927                                 rte_mbuf_refcnt_update(cur, 1);
 928
 929                         prev->next = cur;
 930                         prev->data_len = mbuf_offset;
 931                         m->nb_segs += 1;
 932                         m->pkt_len += mbuf_offset;
 933                         prev = cur;
 934
 935                         mbuf_offset = 0;
 936                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
 937                 }
 938         }
 939
 940         prev->data_len = mbuf_offset;
 941         m->pkt_len    += mbuf_offset;
 942
 943         if (hdr)
 944                 vhost_dequeue_offload(hdr, m);
 945
 946 out:
 947
 948         return error;
 949 }
 950
 951 static __rte_always_inline struct zcopy_mbuf *
 952 get_zmbuf(struct vhost_virtqueue *vq)
 953 {
 954         uint16_t i;
 955         uint16_t last;
 956         int tries = 0;
 957
 958         /* search [last_zmbuf_idx, zmbuf_size) */
 959         i = vq->last_zmbuf_idx;
 960         last = vq->zmbuf_size;
 961
 962 again:
 963         for (; i < last; i++) {
 964                 if (vq->zmbufs[i].in_use == 0) {
 965                         vq->last_zmbuf_idx = i + 1;
 966                         vq->zmbufs[i].in_use = 1;
 967                         return &vq->zmbufs[i];
 968                 }
 969         }
 970
 971         tries++;
 972         if (tries == 1) {
 973                 /* search [0, last_zmbuf_idx) */
 974                 i = 0;
 975                 last = vq->last_zmbuf_idx;
 976                 goto again;
 977         }
 978
 979         return NULL;
 980 }
 981
 982 static __rte_always_inline bool
 983 mbuf_is_consumed(struct rte_mbuf *m)
 984 {
 985         while (m) {
 986                 if (rte_mbuf_refcnt_read(m) > 1)
 987                         return false;
 988                 m = m->next;
 989         }
 990
 991         return true;
 992 }
 993
 994 static __rte_always_inline void
 995 restore_mbuf(struct rte_mbuf *m)
 996 {
 997         uint32_t mbuf_size, priv_size;
 998
 999         while (m) {
1000                 priv_size = rte_pktmbuf_priv_size(m->pool);
1001                 mbuf_size = sizeof(struct rte_mbuf) + priv_size;
1002                 /* start of buffer is after mbuf structure and priv data */
1003
1004                 m->buf_addr = (char *)m + mbuf_size;
1005                 m->buf_iova = rte_mempool_virt2iova(m) + mbuf_size;
1006                 m = m->next;
1007         }
1008 }
1009
1010 uint16_t
1011 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1012         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1013 {
1014         struct virtio_net *dev;
1015         struct rte_mbuf *rarp_mbuf = NULL;
1016         struct vhost_virtqueue *vq;
1017         uint32_t i = 0;
1018         uint16_t free_entries;
1019
1020         dev = get_device(vid);
1021         if (!dev)
1022                 return 0;
1023
1024         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1025                 RTE_LOG(ERR, VHOST_DATA,
1026                         "(%d) %s: built-in vhost net backend is disabled.\n",
1027                         dev->vid, __func__);
1028                 return 0;
1029         }
1030
1031         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1032                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1033                         dev->vid, __func__, queue_id);
1034                 return 0;
1035         }
1036
1037         vq = dev->virtqueue[queue_id];
1038
1039         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
1040                 return 0;
1041
1042         if (unlikely(vq->enabled == 0))
1043                 goto out_access_unlock;
1044
1045         vq->batch_copy_nb_elems = 0;
1046         vq->shadow_used_idx = 0;
1047
1048         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1049                 vhost_user_iotlb_rd_lock(vq);
1050
1051         if (unlikely(vq->access_ok == 0))
1052                 if (unlikely(vring_translate(dev, vq) < 0))
1053                         goto out;
1054
1055         if (unlikely(dev->dequeue_zero_copy)) {
1056                 struct zcopy_mbuf *zmbuf, *next;
1057                 int nr_updated = 0;
1058
1059                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1060                      zmbuf != NULL; zmbuf = next) {
1061                         next = TAILQ_NEXT(zmbuf, next);
1062
1063                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1064                                 update_shadow_used_ring(vq, zmbuf->desc_idx, 0);
1065                                 nr_updated += 1;
1066
1067                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1068                                 restore_mbuf(zmbuf->mbuf);
1069                                 rte_pktmbuf_free(zmbuf->mbuf);
1070                                 put_zmbuf(zmbuf);
1071                                 vq->nr_zmbuf -= 1;
1072                         }
1073                 }
1074
1075                 flush_shadow_used_ring(dev, vq);
1076                 vhost_vring_call(dev, vq);
1077                 vq->shadow_used_idx = 0;
1078         }
1079
1080         /*
1081          * Construct a RARP broadcast packet, and inject it to the "pkts"
1082          * array, to looks like that guest actually send such packet.
1083          *
1084          * Check user_send_rarp() for more information.
1085          *
1086          * broadcast_rarp shares a cacheline in the virtio_net structure
1087          * with some fields that are accessed during enqueue and
1088          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1089          * result in false sharing between enqueue and dequeue.
1090          *
1091          * Prevent unnecessary false sharing by reading broadcast_rarp first
1092          * and only performing cmpset if the read indicates it is likely to
1093          * be set.
1094          */
1095
1096         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1097                         rte_atomic16_cmpset((volatile uint16_t *)
1098                                 &dev->broadcast_rarp.cnt, 1, 0))) {
1099
1100                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
1101                 if (rarp_mbuf == NULL) {
1102                         RTE_LOG(ERR, VHOST_DATA,
1103                                 "Failed to make RARP packet.\n");
1104                         return 0;
1105                 }
1106                 count -= 1;
1107         }
1108
1109         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1110                         vq->last_avail_idx;
1111         if (free_entries == 0)
1112                 goto out;
1113
1114         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1115
1116         count = RTE_MIN(count, MAX_PKT_BURST);
1117         count = RTE_MIN(count, free_entries);
1118         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1119                         dev->vid, count);
1120
1121         for (i = 0; i < count; i++) {
1122                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1123                 uint16_t head_idx, dummy_len;
1124                 uint32_t nr_vec = 0;
1125                 int err;
1126
1127                 if (unlikely(fill_vec_buf(dev, vq,
1128                                                 vq->last_avail_idx + i,
1129                                                 &nr_vec, buf_vec,
1130                                                 &head_idx, &dummy_len,
1131                                                 VHOST_ACCESS_RO) < 0))
1132                         break;
1133
1134                 if (likely(dev->dequeue_zero_copy == 0))
1135                         update_shadow_used_ring(vq, head_idx, 0);
1136
1137                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1138                 if (unlikely(pkts[i] == NULL)) {
1139                         RTE_LOG(ERR, VHOST_DATA,
1140                                 "Failed to allocate memory for mbuf.\n");
1141                         break;
1142                 }
1143
1144                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1145                                 mbuf_pool);
1146                 if (unlikely(err)) {
1147                         rte_pktmbuf_free(pkts[i]);
1148                         break;
1149                 }
1150
1151                 if (unlikely(dev->dequeue_zero_copy)) {
1152                         struct zcopy_mbuf *zmbuf;
1153
1154                         zmbuf = get_zmbuf(vq);
1155                         if (!zmbuf) {
1156                                 rte_pktmbuf_free(pkts[i]);
1157                                 break;
1158                         }
1159                         zmbuf->mbuf = pkts[i];
1160                         zmbuf->desc_idx = head_idx;
1161
1162                         /*
1163                          * Pin lock the mbuf; we will check later to see
1164                          * whether the mbuf is freed (when we are the last
1165                          * user) or not. If that's the case, we then could
1166                          * update the used ring safely.
1167                          */
1168                         rte_mbuf_refcnt_update(pkts[i], 1);
1169
1170                         vq->nr_zmbuf += 1;
1171                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1172                 }
1173         }
1174         vq->last_avail_idx += i;
1175
1176         if (likely(dev->dequeue_zero_copy == 0)) {
1177                 do_data_copy_dequeue(vq);
1178                 if (unlikely(i < count))
1179                         vq->shadow_used_idx = i;
1180                 flush_shadow_used_ring(dev, vq);
1181                 vhost_vring_call(dev, vq);
1182         }
1183
1184 out:
1185         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1186                 vhost_user_iotlb_rd_unlock(vq);
1187
1188 out_access_unlock:
1189         rte_spinlock_unlock(&vq->access_lock);
1190
1191         if (unlikely(rarp_mbuf != NULL)) {
1192                 /*
1193                  * Inject it to the head of "pkts" array, so that switch's mac
1194                  * learning table will get updated first.
1195                  */
1196                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1197                 pkts[0] = rarp_mbuf;
1198                 i += 1;
1199         }
1200
1201         return i;
1202 }