4 * Copyright(c) 2010-2016 Intel Corporation. All rights reserved.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * * Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * * Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * * Neither the name of Intel Corporation nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
36 #include <linux/virtio_net.h>
39 #include <rte_memcpy.h>
40 #include <rte_ether.h>
42 #include <rte_vhost.h>
51 #define MAX_PKT_BURST 32
53 #define MAX_BATCH_LEN 256
56 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
58 return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
61 static __rte_always_inline void
62 do_flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
63 uint16_t to, uint16_t from, uint16_t size)
65 rte_memcpy(&vq->used->ring[to],
66 &vq->shadow_used_ring[from],
67 size * sizeof(struct vring_used_elem));
68 vhost_log_used_vring(dev, vq,
69 offsetof(struct vring_used, ring[to]),
70 size * sizeof(struct vring_used_elem));
73 static __rte_always_inline void
74 flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq)
76 uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
78 if (used_idx + vq->shadow_used_idx <= vq->size) {
79 do_flush_shadow_used_ring(dev, vq, used_idx, 0,
84 /* update used ring interval [used_idx, vq->size] */
85 size = vq->size - used_idx;
86 do_flush_shadow_used_ring(dev, vq, used_idx, 0, size);
88 /* update the left half used ring interval [0, left_size] */
89 do_flush_shadow_used_ring(dev, vq, 0, size,
90 vq->shadow_used_idx - size);
92 vq->last_used_idx += vq->shadow_used_idx;
96 *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
97 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
98 sizeof(vq->used->idx));
101 static __rte_always_inline void
102 update_shadow_used_ring(struct vhost_virtqueue *vq,
103 uint16_t desc_idx, uint16_t len)
105 uint16_t i = vq->shadow_used_idx++;
107 vq->shadow_used_ring[i].id = desc_idx;
108 vq->shadow_used_ring[i].len = len;
112 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
114 struct batch_copy_elem *elem = vq->batch_copy_elems;
115 uint16_t count = vq->batch_copy_nb_elems;
118 for (i = 0; i < count; i++) {
119 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
120 vhost_log_write(dev, elem[i].log_addr, elem[i].len);
121 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
126 do_data_copy_dequeue(struct vhost_virtqueue *vq)
128 struct batch_copy_elem *elem = vq->batch_copy_elems;
129 uint16_t count = vq->batch_copy_nb_elems;
132 for (i = 0; i < count; i++)
133 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
136 /* avoid write operation when necessary, to lessen cache issues */
137 #define ASSIGN_UNLESS_EQUAL(var, val) do { \
138 if ((var) != (val)) \
143 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
145 uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
147 if (m_buf->ol_flags & PKT_TX_TCP_SEG)
148 csum_l4 |= PKT_TX_TCP_CKSUM;
151 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
152 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
155 case PKT_TX_TCP_CKSUM:
156 net_hdr->csum_offset = (offsetof(struct tcp_hdr,
159 case PKT_TX_UDP_CKSUM:
160 net_hdr->csum_offset = (offsetof(struct udp_hdr,
163 case PKT_TX_SCTP_CKSUM:
164 net_hdr->csum_offset = (offsetof(struct sctp_hdr,
169 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
170 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
171 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
174 /* IP cksum verification cannot be bypassed, then calculate here */
175 if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
176 struct ipv4_hdr *ipv4_hdr;
178 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct ipv4_hdr *,
180 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
183 if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
184 if (m_buf->ol_flags & PKT_TX_IPV4)
185 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
187 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
188 net_hdr->gso_size = m_buf->tso_segsz;
189 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
192 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
193 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
194 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
198 static __rte_always_inline int
199 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
200 struct vring_desc *descs, struct rte_mbuf *m,
201 uint16_t desc_idx, uint32_t size)
203 uint32_t desc_avail, desc_offset;
204 uint32_t mbuf_avail, mbuf_offset;
206 struct vring_desc *desc;
208 /* A counter to avoid desc dead loop chain */
209 uint16_t nr_desc = 1;
210 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
211 uint16_t copy_nb = vq->batch_copy_nb_elems;
214 desc = &descs[desc_idx];
215 desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
216 desc->len, VHOST_ACCESS_RW);
218 * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
219 * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
220 * otherwise stores offset on the stack instead of in a register.
222 if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr) {
227 rte_prefetch0((void *)(uintptr_t)desc_addr);
229 virtio_enqueue_offload(m, (struct virtio_net_hdr *)(uintptr_t)desc_addr);
230 vhost_log_write(dev, desc->addr, dev->vhost_hlen);
231 PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
233 desc_offset = dev->vhost_hlen;
234 desc_avail = desc->len - dev->vhost_hlen;
236 mbuf_avail = rte_pktmbuf_data_len(m);
238 while (mbuf_avail != 0 || m->next != NULL) {
239 /* done with current mbuf, fetch next */
240 if (mbuf_avail == 0) {
244 mbuf_avail = rte_pktmbuf_data_len(m);
247 /* done with current desc buf, fetch next */
248 if (desc_avail == 0) {
249 if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
250 /* Room in vring buffer is not enough */
254 if (unlikely(desc->next >= size || ++nr_desc > size)) {
259 desc = &descs[desc->next];
260 desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
263 if (unlikely(!desc_addr)) {
269 desc_avail = desc->len;
272 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
273 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
274 rte_memcpy((void *)((uintptr_t)(desc_addr +
276 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
278 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
279 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
282 batch_copy[copy_nb].dst =
283 (void *)((uintptr_t)(desc_addr + desc_offset));
284 batch_copy[copy_nb].src =
285 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
286 batch_copy[copy_nb].log_addr = desc->addr + desc_offset;
287 batch_copy[copy_nb].len = cpy_len;
291 mbuf_avail -= cpy_len;
292 mbuf_offset += cpy_len;
293 desc_avail -= cpy_len;
294 desc_offset += cpy_len;
298 vq->batch_copy_nb_elems = copy_nb;
304 * This function adds buffers to the virtio devices RX virtqueue. Buffers can
305 * be received from the physical port or from another virtio device. A packet
306 * count is returned to indicate the number of packets that are successfully
307 * added to the RX queue. This function works when the mbuf is scattered, but
308 * it doesn't support the mergeable feature.
310 static __rte_always_inline uint32_t
311 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
312 struct rte_mbuf **pkts, uint32_t count)
314 struct vhost_virtqueue *vq;
315 uint16_t avail_idx, free_entries, start_idx;
316 uint16_t desc_indexes[MAX_PKT_BURST];
317 struct vring_desc *descs;
321 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
322 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
323 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
324 dev->vid, __func__, queue_id);
328 vq = dev->virtqueue[queue_id];
329 if (unlikely(vq->enabled == 0))
332 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
333 start_idx = vq->last_used_idx;
334 free_entries = avail_idx - start_idx;
335 count = RTE_MIN(count, free_entries);
336 count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
340 LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
341 dev->vid, start_idx, start_idx + count);
343 vq->batch_copy_nb_elems = 0;
345 /* Retrieve all of the desc indexes first to avoid caching issues. */
346 rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
347 for (i = 0; i < count; i++) {
348 used_idx = (start_idx + i) & (vq->size - 1);
349 desc_indexes[i] = vq->avail->ring[used_idx];
350 vq->used->ring[used_idx].id = desc_indexes[i];
351 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
353 vhost_log_used_vring(dev, vq,
354 offsetof(struct vring_used, ring[used_idx]),
355 sizeof(vq->used->ring[used_idx]));
358 rte_prefetch0(&vq->desc[desc_indexes[0]]);
360 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
361 vhost_user_iotlb_rd_lock(vq);
363 for (i = 0; i < count; i++) {
364 uint16_t desc_idx = desc_indexes[i];
367 if (vq->desc[desc_idx].flags & VRING_DESC_F_INDIRECT) {
368 descs = (struct vring_desc *)(uintptr_t)
369 vhost_iova_to_vva(dev,
370 vq, vq->desc[desc_idx].addr,
371 vq->desc[desc_idx].len,
373 if (unlikely(!descs)) {
379 sz = vq->desc[desc_idx].len / sizeof(*descs);
385 err = copy_mbuf_to_desc(dev, vq, descs, pkts[i], desc_idx, sz);
392 rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
395 do_data_copy_enqueue(dev, vq);
397 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
398 vhost_user_iotlb_rd_unlock(vq);
402 *(volatile uint16_t *)&vq->used->idx += count;
403 vq->last_used_idx += count;
404 vhost_log_used_vring(dev, vq,
405 offsetof(struct vring_used, idx),
406 sizeof(vq->used->idx));
408 /* flush used->idx update before we read avail->flags. */
411 /* Kick the guest if necessary. */
412 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
413 && (vq->callfd >= 0))
414 eventfd_write(vq->callfd, (eventfd_t)1);
418 static __rte_always_inline int
419 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
420 uint32_t avail_idx, uint32_t *vec_idx,
421 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
422 uint16_t *desc_chain_len)
424 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
425 uint32_t vec_id = *vec_idx;
427 struct vring_desc *descs = vq->desc;
429 *desc_chain_head = idx;
431 if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
432 descs = (struct vring_desc *)(uintptr_t)
433 vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
436 if (unlikely(!descs))
443 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
446 len += descs[idx].len;
447 buf_vec[vec_id].buf_addr = descs[idx].addr;
448 buf_vec[vec_id].buf_len = descs[idx].len;
449 buf_vec[vec_id].desc_idx = idx;
452 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
455 idx = descs[idx].next;
458 *desc_chain_len = len;
465 * Returns -1 on fail, 0 on success
468 reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
469 uint32_t size, struct buf_vector *buf_vec,
470 uint16_t *num_buffers, uint16_t avail_head)
473 uint32_t vec_idx = 0;
476 uint16_t head_idx = 0;
480 cur_idx = vq->last_avail_idx;
483 if (unlikely(cur_idx == avail_head))
486 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
487 &head_idx, &len) < 0))
489 len = RTE_MIN(len, size);
490 update_shadow_used_ring(vq, head_idx, len);
498 * if we tried all available ring items, and still
499 * can't get enough buf, it means something abnormal
502 if (unlikely(tries >= vq->size))
509 static __rte_always_inline int
510 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
511 struct rte_mbuf *m, struct buf_vector *buf_vec,
512 uint16_t num_buffers)
514 uint32_t vec_idx = 0;
516 uint32_t mbuf_offset, mbuf_avail;
517 uint32_t desc_offset, desc_avail;
519 uint64_t hdr_addr, hdr_phys_addr;
520 struct rte_mbuf *hdr_mbuf;
521 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
522 uint16_t copy_nb = vq->batch_copy_nb_elems;
525 if (unlikely(m == NULL)) {
530 desc_addr = vhost_iova_to_vva(dev, vq, buf_vec[vec_idx].buf_addr,
531 buf_vec[vec_idx].buf_len,
533 if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr) {
539 hdr_addr = desc_addr;
540 hdr_phys_addr = buf_vec[vec_idx].buf_addr;
541 rte_prefetch0((void *)(uintptr_t)hdr_addr);
543 LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
544 dev->vid, num_buffers);
546 desc_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
547 desc_offset = dev->vhost_hlen;
549 mbuf_avail = rte_pktmbuf_data_len(m);
551 while (mbuf_avail != 0 || m->next != NULL) {
552 /* done with current desc buf, get the next one */
553 if (desc_avail == 0) {
556 vhost_iova_to_vva(dev, vq,
557 buf_vec[vec_idx].buf_addr,
558 buf_vec[vec_idx].buf_len,
560 if (unlikely(!desc_addr)) {
565 /* Prefetch buffer address. */
566 rte_prefetch0((void *)(uintptr_t)desc_addr);
568 desc_avail = buf_vec[vec_idx].buf_len;
571 /* done with current mbuf, get the next one */
572 if (mbuf_avail == 0) {
576 mbuf_avail = rte_pktmbuf_data_len(m);
580 struct virtio_net_hdr_mrg_rxbuf *hdr;
582 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)
584 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
585 ASSIGN_UNLESS_EQUAL(hdr->num_buffers, num_buffers);
587 vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
588 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
594 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
596 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
597 rte_memcpy((void *)((uintptr_t)(desc_addr +
599 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
602 buf_vec[vec_idx].buf_addr + desc_offset,
604 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
607 batch_copy[copy_nb].dst =
608 (void *)((uintptr_t)(desc_addr + desc_offset));
609 batch_copy[copy_nb].src =
610 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
611 batch_copy[copy_nb].log_addr =
612 buf_vec[vec_idx].buf_addr + desc_offset;
613 batch_copy[copy_nb].len = cpy_len;
617 mbuf_avail -= cpy_len;
618 mbuf_offset += cpy_len;
619 desc_avail -= cpy_len;
620 desc_offset += cpy_len;
624 vq->batch_copy_nb_elems = copy_nb;
629 static __rte_always_inline uint32_t
630 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
631 struct rte_mbuf **pkts, uint32_t count)
633 struct vhost_virtqueue *vq;
634 uint32_t pkt_idx = 0;
635 uint16_t num_buffers;
636 struct buf_vector buf_vec[BUF_VECTOR_MAX];
639 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
640 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
641 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
642 dev->vid, __func__, queue_id);
646 vq = dev->virtqueue[queue_id];
647 if (unlikely(vq->enabled == 0))
650 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
654 vq->batch_copy_nb_elems = 0;
656 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
658 vq->shadow_used_idx = 0;
659 avail_head = *((volatile uint16_t *)&vq->avail->idx);
661 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
662 vhost_user_iotlb_rd_lock(vq);
664 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
665 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
667 if (unlikely(reserve_avail_buf_mergeable(dev, vq,
668 pkt_len, buf_vec, &num_buffers,
670 LOG_DEBUG(VHOST_DATA,
671 "(%d) failed to get enough desc from vring\n",
673 vq->shadow_used_idx -= num_buffers;
677 LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
678 dev->vid, vq->last_avail_idx,
679 vq->last_avail_idx + num_buffers);
681 if (copy_mbuf_to_desc_mergeable(dev, vq, pkts[pkt_idx],
682 buf_vec, num_buffers) < 0) {
683 vq->shadow_used_idx -= num_buffers;
687 vq->last_avail_idx += num_buffers;
690 do_data_copy_enqueue(dev, vq);
692 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
693 vhost_user_iotlb_rd_unlock(vq);
695 if (likely(vq->shadow_used_idx)) {
696 flush_shadow_used_ring(dev, vq);
698 /* flush used->idx update before we read avail->flags. */
701 /* Kick the guest if necessary. */
702 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
703 && (vq->callfd >= 0))
704 eventfd_write(vq->callfd, (eventfd_t)1);
711 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
712 struct rte_mbuf **pkts, uint16_t count)
714 struct virtio_net *dev = get_device(vid);
719 if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
720 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
722 return virtio_dev_rx(dev, queue_id, pkts, count);
726 virtio_net_with_host_offload(struct virtio_net *dev)
729 ((1ULL << VIRTIO_NET_F_CSUM) |
730 (1ULL << VIRTIO_NET_F_HOST_ECN) |
731 (1ULL << VIRTIO_NET_F_HOST_TSO4) |
732 (1ULL << VIRTIO_NET_F_HOST_TSO6) |
733 (1ULL << VIRTIO_NET_F_HOST_UFO)))
740 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
742 struct ipv4_hdr *ipv4_hdr;
743 struct ipv6_hdr *ipv6_hdr;
745 struct ether_hdr *eth_hdr;
748 eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
750 m->l2_len = sizeof(struct ether_hdr);
751 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
753 if (ethertype == ETHER_TYPE_VLAN) {
754 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
756 m->l2_len += sizeof(struct vlan_hdr);
757 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
760 l3_hdr = (char *)eth_hdr + m->l2_len;
763 case ETHER_TYPE_IPv4:
765 *l4_proto = ipv4_hdr->next_proto_id;
766 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
767 *l4_hdr = (char *)l3_hdr + m->l3_len;
768 m->ol_flags |= PKT_TX_IPV4;
770 case ETHER_TYPE_IPv6:
772 *l4_proto = ipv6_hdr->proto;
773 m->l3_len = sizeof(struct ipv6_hdr);
774 *l4_hdr = (char *)l3_hdr + m->l3_len;
775 m->ol_flags |= PKT_TX_IPV6;
785 static __rte_always_inline void
786 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
788 uint16_t l4_proto = 0;
790 struct tcp_hdr *tcp_hdr = NULL;
792 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
795 parse_ethernet(m, &l4_proto, &l4_hdr);
796 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
797 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
798 switch (hdr->csum_offset) {
799 case (offsetof(struct tcp_hdr, cksum)):
800 if (l4_proto == IPPROTO_TCP)
801 m->ol_flags |= PKT_TX_TCP_CKSUM;
803 case (offsetof(struct udp_hdr, dgram_cksum)):
804 if (l4_proto == IPPROTO_UDP)
805 m->ol_flags |= PKT_TX_UDP_CKSUM;
807 case (offsetof(struct sctp_hdr, cksum)):
808 if (l4_proto == IPPROTO_SCTP)
809 m->ol_flags |= PKT_TX_SCTP_CKSUM;
817 if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
818 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
819 case VIRTIO_NET_HDR_GSO_TCPV4:
820 case VIRTIO_NET_HDR_GSO_TCPV6:
822 m->ol_flags |= PKT_TX_TCP_SEG;
823 m->tso_segsz = hdr->gso_size;
824 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
827 RTE_LOG(WARNING, VHOST_DATA,
828 "unsupported gso type %u.\n", hdr->gso_type);
834 #define RARP_PKT_SIZE 64
837 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
839 struct ether_hdr *eth_hdr;
840 struct arp_hdr *rarp;
842 if (rarp_mbuf->buf_len < 64) {
843 RTE_LOG(WARNING, VHOST_DATA,
844 "failed to make RARP; mbuf size too small %u (< %d)\n",
845 rarp_mbuf->buf_len, RARP_PKT_SIZE);
849 /* Ethernet header. */
850 eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
851 memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
852 ether_addr_copy(mac, ð_hdr->s_addr);
853 eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
856 rarp = (struct arp_hdr *)(eth_hdr + 1);
857 rarp->arp_hrd = htons(ARP_HRD_ETHER);
858 rarp->arp_pro = htons(ETHER_TYPE_IPv4);
859 rarp->arp_hln = ETHER_ADDR_LEN;
861 rarp->arp_op = htons(ARP_OP_REVREQUEST);
863 ether_addr_copy(mac, &rarp->arp_data.arp_sha);
864 ether_addr_copy(mac, &rarp->arp_data.arp_tha);
865 memset(&rarp->arp_data.arp_sip, 0x00, 4);
866 memset(&rarp->arp_data.arp_tip, 0x00, 4);
868 rarp_mbuf->pkt_len = rarp_mbuf->data_len = RARP_PKT_SIZE;
873 static __rte_always_inline void
874 put_zmbuf(struct zcopy_mbuf *zmbuf)
879 static __rte_always_inline int
880 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
881 struct vring_desc *descs, uint16_t max_desc,
882 struct rte_mbuf *m, uint16_t desc_idx,
883 struct rte_mempool *mbuf_pool)
885 struct vring_desc *desc;
887 uint32_t desc_avail, desc_offset;
888 uint32_t mbuf_avail, mbuf_offset;
890 struct rte_mbuf *cur = m, *prev = m;
891 struct virtio_net_hdr *hdr = NULL;
892 /* A counter to avoid desc dead loop chain */
893 uint32_t nr_desc = 1;
894 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
895 uint16_t copy_nb = vq->batch_copy_nb_elems;
898 desc = &descs[desc_idx];
899 if (unlikely((desc->len < dev->vhost_hlen)) ||
900 (desc->flags & VRING_DESC_F_INDIRECT)) {
905 desc_addr = vhost_iova_to_vva(dev,
909 if (unlikely(!desc_addr)) {
914 if (virtio_net_with_host_offload(dev)) {
915 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
920 * A virtio driver normally uses at least 2 desc buffers
921 * for Tx: the first for storing the header, and others
922 * for storing the data.
924 if (likely((desc->len == dev->vhost_hlen) &&
925 (desc->flags & VRING_DESC_F_NEXT) != 0)) {
926 desc = &descs[desc->next];
927 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
932 desc_addr = vhost_iova_to_vva(dev,
936 if (unlikely(!desc_addr)) {
942 desc_avail = desc->len;
945 desc_avail = desc->len - dev->vhost_hlen;
946 desc_offset = dev->vhost_hlen;
949 rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
951 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
954 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
958 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
961 * A desc buf might across two host physical pages that are
962 * not continuous. In such case (gpa_to_hpa returns 0), data
963 * will be copied even though zero copy is enabled.
965 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
966 desc->addr + desc_offset, cpy_len)))) {
967 cur->data_len = cpy_len;
969 cur->buf_addr = (void *)(uintptr_t)desc_addr;
970 cur->buf_physaddr = hpa;
973 * In zero copy mode, one mbuf can only reference data
974 * for one or partial of one desc buff.
976 mbuf_avail = cpy_len;
978 if (likely(cpy_len > MAX_BATCH_LEN ||
979 copy_nb >= vq->size)) {
980 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
982 (void *)((uintptr_t)(desc_addr +
986 batch_copy[copy_nb].dst =
987 rte_pktmbuf_mtod_offset(cur, void *,
989 batch_copy[copy_nb].src =
990 (void *)((uintptr_t)(desc_addr +
992 batch_copy[copy_nb].len = cpy_len;
997 mbuf_avail -= cpy_len;
998 mbuf_offset += cpy_len;
999 desc_avail -= cpy_len;
1000 desc_offset += cpy_len;
1002 /* This desc reaches to its end, get the next one */
1003 if (desc_avail == 0) {
1004 if ((desc->flags & VRING_DESC_F_NEXT) == 0)
1007 if (unlikely(desc->next >= max_desc ||
1008 ++nr_desc > max_desc)) {
1012 desc = &descs[desc->next];
1013 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
1018 desc_addr = vhost_iova_to_vva(dev,
1022 if (unlikely(!desc_addr)) {
1027 rte_prefetch0((void *)(uintptr_t)desc_addr);
1030 desc_avail = desc->len;
1032 PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
1036 * This mbuf reaches to its end, get a new one
1037 * to hold more data.
1039 if (mbuf_avail == 0) {
1040 cur = rte_pktmbuf_alloc(mbuf_pool);
1041 if (unlikely(cur == NULL)) {
1042 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1043 "allocate memory for mbuf.\n");
1047 if (unlikely(dev->dequeue_zero_copy))
1048 rte_mbuf_refcnt_update(cur, 1);
1051 prev->data_len = mbuf_offset;
1053 m->pkt_len += mbuf_offset;
1057 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1061 prev->data_len = mbuf_offset;
1062 m->pkt_len += mbuf_offset;
1065 vhost_dequeue_offload(hdr, m);
1068 vq->batch_copy_nb_elems = copy_nb;
1073 static __rte_always_inline void
1074 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
1075 uint32_t used_idx, uint32_t desc_idx)
1077 vq->used->ring[used_idx].id = desc_idx;
1078 vq->used->ring[used_idx].len = 0;
1079 vhost_log_used_vring(dev, vq,
1080 offsetof(struct vring_used, ring[used_idx]),
1081 sizeof(vq->used->ring[used_idx]));
1084 static __rte_always_inline void
1085 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
1088 if (unlikely(count == 0))
1094 vq->used->idx += count;
1095 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
1096 sizeof(vq->used->idx));
1098 /* Kick guest if required. */
1099 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
1100 && (vq->callfd >= 0))
1101 eventfd_write(vq->callfd, (eventfd_t)1);
1104 static __rte_always_inline struct zcopy_mbuf *
1105 get_zmbuf(struct vhost_virtqueue *vq)
1111 /* search [last_zmbuf_idx, zmbuf_size) */
1112 i = vq->last_zmbuf_idx;
1113 last = vq->zmbuf_size;
1116 for (; i < last; i++) {
1117 if (vq->zmbufs[i].in_use == 0) {
1118 vq->last_zmbuf_idx = i + 1;
1119 vq->zmbufs[i].in_use = 1;
1120 return &vq->zmbufs[i];
1126 /* search [0, last_zmbuf_idx) */
1128 last = vq->last_zmbuf_idx;
1135 static __rte_always_inline bool
1136 mbuf_is_consumed(struct rte_mbuf *m)
1139 if (rte_mbuf_refcnt_read(m) > 1)
1148 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1149 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1151 struct virtio_net *dev;
1152 struct rte_mbuf *rarp_mbuf = NULL;
1153 struct vhost_virtqueue *vq;
1154 uint32_t desc_indexes[MAX_PKT_BURST];
1157 uint16_t free_entries;
1160 dev = get_device(vid);
1164 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1165 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1166 dev->vid, __func__, queue_id);
1170 vq = dev->virtqueue[queue_id];
1171 if (unlikely(vq->enabled == 0))
1174 vq->batch_copy_nb_elems = 0;
1176 if (unlikely(dev->dequeue_zero_copy)) {
1177 struct zcopy_mbuf *zmbuf, *next;
1180 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1181 zmbuf != NULL; zmbuf = next) {
1182 next = TAILQ_NEXT(zmbuf, next);
1184 if (mbuf_is_consumed(zmbuf->mbuf)) {
1185 used_idx = vq->last_used_idx++ & (vq->size - 1);
1186 update_used_ring(dev, vq, used_idx,
1190 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1191 rte_pktmbuf_free(zmbuf->mbuf);
1197 update_used_idx(dev, vq, nr_updated);
1201 * Construct a RARP broadcast packet, and inject it to the "pkts"
1202 * array, to looks like that guest actually send such packet.
1204 * Check user_send_rarp() for more information.
1206 * broadcast_rarp shares a cacheline in the virtio_net structure
1207 * with some fields that are accessed during enqueue and
1208 * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1209 * result in false sharing between enqueue and dequeue.
1211 * Prevent unnecessary false sharing by reading broadcast_rarp first
1212 * and only performing cmpset if the read indicates it is likely to
1216 if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1217 rte_atomic16_cmpset((volatile uint16_t *)
1218 &dev->broadcast_rarp.cnt, 1, 0))) {
1220 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
1221 if (rarp_mbuf == NULL) {
1222 RTE_LOG(ERR, VHOST_DATA,
1223 "Failed to allocate memory for mbuf.\n");
1227 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
1228 rte_pktmbuf_free(rarp_mbuf);
1235 free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1237 if (free_entries == 0)
1240 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1242 /* Prefetch available and used ring */
1243 avail_idx = vq->last_avail_idx & (vq->size - 1);
1244 used_idx = vq->last_used_idx & (vq->size - 1);
1245 rte_prefetch0(&vq->avail->ring[avail_idx]);
1246 rte_prefetch0(&vq->used->ring[used_idx]);
1248 count = RTE_MIN(count, MAX_PKT_BURST);
1249 count = RTE_MIN(count, free_entries);
1250 LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1253 /* Retrieve all of the head indexes first to avoid caching issues. */
1254 for (i = 0; i < count; i++) {
1255 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1256 used_idx = (vq->last_used_idx + i) & (vq->size - 1);
1257 desc_indexes[i] = vq->avail->ring[avail_idx];
1259 if (likely(dev->dequeue_zero_copy == 0))
1260 update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1263 /* Prefetch descriptor index. */
1264 rte_prefetch0(&vq->desc[desc_indexes[0]]);
1266 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1267 vhost_user_iotlb_rd_lock(vq);
1269 for (i = 0; i < count; i++) {
1270 struct vring_desc *desc;
1274 if (likely(i + 1 < count))
1275 rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1277 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1278 desc = (struct vring_desc *)(uintptr_t)
1279 vhost_iova_to_vva(dev, vq,
1280 vq->desc[desc_indexes[i]].addr,
1283 if (unlikely(!desc))
1286 rte_prefetch0(desc);
1287 sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1292 idx = desc_indexes[i];
1295 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1296 if (unlikely(pkts[i] == NULL)) {
1297 RTE_LOG(ERR, VHOST_DATA,
1298 "Failed to allocate memory for mbuf.\n");
1302 err = copy_desc_to_mbuf(dev, vq, desc, sz, pkts[i], idx,
1304 if (unlikely(err)) {
1305 rte_pktmbuf_free(pkts[i]);
1309 if (unlikely(dev->dequeue_zero_copy)) {
1310 struct zcopy_mbuf *zmbuf;
1312 zmbuf = get_zmbuf(vq);
1314 rte_pktmbuf_free(pkts[i]);
1317 zmbuf->mbuf = pkts[i];
1318 zmbuf->desc_idx = desc_indexes[i];
1321 * Pin lock the mbuf; we will check later to see
1322 * whether the mbuf is freed (when we are the last
1323 * user) or not. If that's the case, we then could
1324 * update the used ring safely.
1326 rte_mbuf_refcnt_update(pkts[i], 1);
1329 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1332 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1333 vhost_user_iotlb_rd_unlock(vq);
1335 vq->last_avail_idx += i;
1337 if (likely(dev->dequeue_zero_copy == 0)) {
1338 do_data_copy_dequeue(vq);
1339 vq->last_used_idx += i;
1340 update_used_idx(dev, vq, i);
1344 if (unlikely(rarp_mbuf != NULL)) {
1346 * Inject it to the head of "pkts" array, so that switch's mac
1347 * learning table will get updated first.
1349 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1350 pkts[0] = rarp_mbuf;