1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2016 Intel Corporation
7 #include <linux/virtio_net.h>
10 #include <rte_memcpy.h>
11 #include <rte_ether.h>
13 #include <rte_vhost.h>
18 #include <rte_spinlock.h>
23 #define MAX_PKT_BURST 32
25 #define MAX_BATCH_LEN 256
28 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
30 return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
33 static __rte_always_inline void
34 do_flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
35 uint16_t to, uint16_t from, uint16_t size)
37 rte_memcpy(&vq->used->ring[to],
38 &vq->shadow_used_ring[from],
39 size * sizeof(struct vring_used_elem));
40 vhost_log_used_vring(dev, vq,
41 offsetof(struct vring_used, ring[to]),
42 size * sizeof(struct vring_used_elem));
45 static __rte_always_inline void
46 flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq)
48 uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
50 if (used_idx + vq->shadow_used_idx <= vq->size) {
51 do_flush_shadow_used_ring(dev, vq, used_idx, 0,
56 /* update used ring interval [used_idx, vq->size] */
57 size = vq->size - used_idx;
58 do_flush_shadow_used_ring(dev, vq, used_idx, 0, size);
60 /* update the left half used ring interval [0, left_size] */
61 do_flush_shadow_used_ring(dev, vq, 0, size,
62 vq->shadow_used_idx - size);
64 vq->last_used_idx += vq->shadow_used_idx;
68 *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
69 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
70 sizeof(vq->used->idx));
73 static __rte_always_inline void
74 update_shadow_used_ring(struct vhost_virtqueue *vq,
75 uint16_t desc_idx, uint16_t len)
77 uint16_t i = vq->shadow_used_idx++;
79 vq->shadow_used_ring[i].id = desc_idx;
80 vq->shadow_used_ring[i].len = len;
84 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
86 struct batch_copy_elem *elem = vq->batch_copy_elems;
87 uint16_t count = vq->batch_copy_nb_elems;
90 for (i = 0; i < count; i++) {
91 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
92 vhost_log_write(dev, elem[i].log_addr, elem[i].len);
93 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
98 do_data_copy_dequeue(struct vhost_virtqueue *vq)
100 struct batch_copy_elem *elem = vq->batch_copy_elems;
101 uint16_t count = vq->batch_copy_nb_elems;
104 for (i = 0; i < count; i++)
105 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
108 /* avoid write operation when necessary, to lessen cache issues */
109 #define ASSIGN_UNLESS_EQUAL(var, val) do { \
110 if ((var) != (val)) \
115 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
117 uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
119 if (m_buf->ol_flags & PKT_TX_TCP_SEG)
120 csum_l4 |= PKT_TX_TCP_CKSUM;
123 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
124 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
127 case PKT_TX_TCP_CKSUM:
128 net_hdr->csum_offset = (offsetof(struct tcp_hdr,
131 case PKT_TX_UDP_CKSUM:
132 net_hdr->csum_offset = (offsetof(struct udp_hdr,
135 case PKT_TX_SCTP_CKSUM:
136 net_hdr->csum_offset = (offsetof(struct sctp_hdr,
141 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
142 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
143 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
146 /* IP cksum verification cannot be bypassed, then calculate here */
147 if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
148 struct ipv4_hdr *ipv4_hdr;
150 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct ipv4_hdr *,
152 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
155 if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
156 if (m_buf->ol_flags & PKT_TX_IPV4)
157 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
159 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
160 net_hdr->gso_size = m_buf->tso_segsz;
161 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
163 } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
164 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
165 net_hdr->gso_size = m_buf->tso_segsz;
166 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
169 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
170 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
171 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
175 static __rte_always_inline int
176 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
177 struct vring_desc *descs, struct rte_mbuf *m,
178 uint16_t desc_idx, uint32_t size)
180 uint32_t desc_avail, desc_offset;
181 uint32_t mbuf_avail, mbuf_offset;
184 struct vring_desc *desc;
186 /* A counter to avoid desc dead loop chain */
187 uint16_t nr_desc = 1;
188 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
189 uint16_t copy_nb = vq->batch_copy_nb_elems;
192 desc = &descs[desc_idx];
194 desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
195 &dlen, VHOST_ACCESS_RW);
197 * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
198 * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
199 * otherwise stores offset on the stack instead of in a register.
201 if (unlikely(dlen != desc->len || desc->len < dev->vhost_hlen) ||
207 rte_prefetch0((void *)(uintptr_t)desc_addr);
209 virtio_enqueue_offload(m, (struct virtio_net_hdr *)(uintptr_t)desc_addr);
210 vhost_log_write(dev, desc->addr, dev->vhost_hlen);
211 PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
213 desc_offset = dev->vhost_hlen;
214 desc_avail = desc->len - dev->vhost_hlen;
216 mbuf_avail = rte_pktmbuf_data_len(m);
218 while (mbuf_avail != 0 || m->next != NULL) {
219 /* done with current mbuf, fetch next */
220 if (mbuf_avail == 0) {
224 mbuf_avail = rte_pktmbuf_data_len(m);
227 /* done with current desc buf, fetch next */
228 if (desc_avail == 0) {
229 if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
230 /* Room in vring buffer is not enough */
234 if (unlikely(desc->next >= size || ++nr_desc > size)) {
239 desc = &descs[desc->next];
241 desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
244 if (unlikely(!desc_addr || dlen != desc->len)) {
250 desc_avail = desc->len;
253 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
254 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
255 rte_memcpy((void *)((uintptr_t)(desc_addr +
257 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
259 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
260 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
263 batch_copy[copy_nb].dst =
264 (void *)((uintptr_t)(desc_addr + desc_offset));
265 batch_copy[copy_nb].src =
266 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
267 batch_copy[copy_nb].log_addr = desc->addr + desc_offset;
268 batch_copy[copy_nb].len = cpy_len;
272 mbuf_avail -= cpy_len;
273 mbuf_offset += cpy_len;
274 desc_avail -= cpy_len;
275 desc_offset += cpy_len;
279 vq->batch_copy_nb_elems = copy_nb;
285 * This function adds buffers to the virtio devices RX virtqueue. Buffers can
286 * be received from the physical port or from another virtio device. A packet
287 * count is returned to indicate the number of packets that are successfully
288 * added to the RX queue. This function works when the mbuf is scattered, but
289 * it doesn't support the mergeable feature.
291 static __rte_always_inline uint32_t
292 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
293 struct rte_mbuf **pkts, uint32_t count)
295 struct vhost_virtqueue *vq;
296 uint16_t avail_idx, free_entries, start_idx;
297 uint16_t desc_indexes[MAX_PKT_BURST];
298 struct vring_desc *descs;
302 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
303 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
304 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
305 dev->vid, __func__, queue_id);
309 vq = dev->virtqueue[queue_id];
311 rte_spinlock_lock(&vq->access_lock);
313 if (unlikely(vq->enabled == 0))
314 goto out_access_unlock;
316 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
317 vhost_user_iotlb_rd_lock(vq);
319 if (unlikely(vq->access_ok == 0)) {
320 if (unlikely(vring_translate(dev, vq) < 0)) {
326 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
327 start_idx = vq->last_used_idx;
328 free_entries = avail_idx - start_idx;
329 count = RTE_MIN(count, free_entries);
330 count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
334 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
335 dev->vid, start_idx, start_idx + count);
337 vq->batch_copy_nb_elems = 0;
339 /* Retrieve all of the desc indexes first to avoid caching issues. */
340 rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
341 for (i = 0; i < count; i++) {
342 used_idx = (start_idx + i) & (vq->size - 1);
343 desc_indexes[i] = vq->avail->ring[used_idx];
344 vq->used->ring[used_idx].id = desc_indexes[i];
345 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
347 vhost_log_used_vring(dev, vq,
348 offsetof(struct vring_used, ring[used_idx]),
349 sizeof(vq->used->ring[used_idx]));
352 rte_prefetch0(&vq->desc[desc_indexes[0]]);
353 for (i = 0; i < count; i++) {
354 uint16_t desc_idx = desc_indexes[i];
357 if (vq->desc[desc_idx].flags & VRING_DESC_F_INDIRECT) {
358 uint64_t dlen = vq->desc[desc_idx].len;
359 descs = (struct vring_desc *)(uintptr_t)
360 vhost_iova_to_vva(dev,
361 vq, vq->desc[desc_idx].addr,
362 &dlen, VHOST_ACCESS_RO);
363 if (unlikely(!descs ||
364 dlen != vq->desc[desc_idx].len)) {
370 sz = vq->desc[desc_idx].len / sizeof(*descs);
376 err = copy_mbuf_to_desc(dev, vq, descs, pkts[i], desc_idx, sz);
383 rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
386 do_data_copy_enqueue(dev, vq);
390 *(volatile uint16_t *)&vq->used->idx += count;
391 vq->last_used_idx += count;
392 vhost_log_used_vring(dev, vq,
393 offsetof(struct vring_used, idx),
394 sizeof(vq->used->idx));
396 vhost_vring_call(dev, vq);
398 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
399 vhost_user_iotlb_rd_unlock(vq);
402 rte_spinlock_unlock(&vq->access_lock);
407 static __rte_always_inline int
408 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
409 uint32_t avail_idx, uint32_t *vec_idx,
410 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
411 uint16_t *desc_chain_len)
413 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
414 uint32_t vec_id = *vec_idx;
417 struct vring_desc *descs = vq->desc;
419 *desc_chain_head = idx;
421 if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
422 dlen = vq->desc[idx].len;
423 descs = (struct vring_desc *)(uintptr_t)
424 vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
427 if (unlikely(!descs || dlen != vq->desc[idx].len))
434 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
437 len += descs[idx].len;
438 buf_vec[vec_id].buf_addr = descs[idx].addr;
439 buf_vec[vec_id].buf_len = descs[idx].len;
440 buf_vec[vec_id].desc_idx = idx;
443 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
446 idx = descs[idx].next;
449 *desc_chain_len = len;
456 * Returns -1 on fail, 0 on success
459 reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
460 uint32_t size, struct buf_vector *buf_vec,
461 uint16_t *num_buffers, uint16_t avail_head)
464 uint32_t vec_idx = 0;
467 uint16_t head_idx = 0;
471 cur_idx = vq->last_avail_idx;
474 if (unlikely(cur_idx == avail_head))
477 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
478 &head_idx, &len) < 0))
480 len = RTE_MIN(len, size);
481 update_shadow_used_ring(vq, head_idx, len);
489 * if we tried all available ring items, and still
490 * can't get enough buf, it means something abnormal
493 if (unlikely(tries >= vq->size))
500 static __rte_always_inline int
501 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
502 struct rte_mbuf *m, struct buf_vector *buf_vec,
503 uint16_t num_buffers)
505 uint32_t vec_idx = 0;
507 uint32_t mbuf_offset, mbuf_avail;
508 uint32_t desc_offset, desc_avail;
511 uint64_t hdr_addr, hdr_phys_addr;
512 struct rte_mbuf *hdr_mbuf;
513 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
514 uint16_t copy_nb = vq->batch_copy_nb_elems;
517 if (unlikely(m == NULL)) {
522 dlen = buf_vec[vec_idx].buf_len;
523 desc_addr = vhost_iova_to_vva(dev, vq, buf_vec[vec_idx].buf_addr,
524 &dlen, VHOST_ACCESS_RW);
525 if (dlen != buf_vec[vec_idx].buf_len ||
526 buf_vec[vec_idx].buf_len < dev->vhost_hlen ||
533 hdr_addr = desc_addr;
534 hdr_phys_addr = buf_vec[vec_idx].buf_addr;
535 rte_prefetch0((void *)(uintptr_t)hdr_addr);
537 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
538 dev->vid, num_buffers);
540 desc_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
541 desc_offset = dev->vhost_hlen;
543 mbuf_avail = rte_pktmbuf_data_len(m);
545 while (mbuf_avail != 0 || m->next != NULL) {
546 /* done with current desc buf, get the next one */
547 if (desc_avail == 0) {
549 dlen = buf_vec[vec_idx].buf_len;
551 vhost_iova_to_vva(dev, vq,
552 buf_vec[vec_idx].buf_addr,
555 if (unlikely(!desc_addr ||
556 dlen != buf_vec[vec_idx].buf_len)) {
561 /* Prefetch buffer address. */
562 rte_prefetch0((void *)(uintptr_t)desc_addr);
564 desc_avail = buf_vec[vec_idx].buf_len;
567 /* done with current mbuf, get the next one */
568 if (mbuf_avail == 0) {
572 mbuf_avail = rte_pktmbuf_data_len(m);
576 struct virtio_net_hdr_mrg_rxbuf *hdr;
578 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)
580 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
581 ASSIGN_UNLESS_EQUAL(hdr->num_buffers, num_buffers);
583 vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
584 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
590 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
592 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
593 rte_memcpy((void *)((uintptr_t)(desc_addr +
595 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
598 buf_vec[vec_idx].buf_addr + desc_offset,
600 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
603 batch_copy[copy_nb].dst =
604 (void *)((uintptr_t)(desc_addr + desc_offset));
605 batch_copy[copy_nb].src =
606 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
607 batch_copy[copy_nb].log_addr =
608 buf_vec[vec_idx].buf_addr + desc_offset;
609 batch_copy[copy_nb].len = cpy_len;
613 mbuf_avail -= cpy_len;
614 mbuf_offset += cpy_len;
615 desc_avail -= cpy_len;
616 desc_offset += cpy_len;
620 vq->batch_copy_nb_elems = copy_nb;
625 static __rte_always_inline uint32_t
626 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
627 struct rte_mbuf **pkts, uint32_t count)
629 struct vhost_virtqueue *vq;
630 uint32_t pkt_idx = 0;
631 uint16_t num_buffers;
632 struct buf_vector buf_vec[BUF_VECTOR_MAX];
635 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
636 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
637 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
638 dev->vid, __func__, queue_id);
642 vq = dev->virtqueue[queue_id];
644 rte_spinlock_lock(&vq->access_lock);
646 if (unlikely(vq->enabled == 0))
647 goto out_access_unlock;
649 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
650 vhost_user_iotlb_rd_lock(vq);
652 if (unlikely(vq->access_ok == 0))
653 if (unlikely(vring_translate(dev, vq) < 0))
656 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
660 vq->batch_copy_nb_elems = 0;
662 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
664 vq->shadow_used_idx = 0;
665 avail_head = *((volatile uint16_t *)&vq->avail->idx);
666 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
667 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
669 if (unlikely(reserve_avail_buf_mergeable(dev, vq,
670 pkt_len, buf_vec, &num_buffers,
672 VHOST_LOG_DEBUG(VHOST_DATA,
673 "(%d) failed to get enough desc from vring\n",
675 vq->shadow_used_idx -= num_buffers;
679 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
680 dev->vid, vq->last_avail_idx,
681 vq->last_avail_idx + num_buffers);
683 if (copy_mbuf_to_desc_mergeable(dev, vq, pkts[pkt_idx],
684 buf_vec, num_buffers) < 0) {
685 vq->shadow_used_idx -= num_buffers;
689 vq->last_avail_idx += num_buffers;
692 do_data_copy_enqueue(dev, vq);
694 if (likely(vq->shadow_used_idx)) {
695 flush_shadow_used_ring(dev, vq);
696 vhost_vring_call(dev, vq);
700 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
701 vhost_user_iotlb_rd_unlock(vq);
704 rte_spinlock_unlock(&vq->access_lock);
710 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
711 struct rte_mbuf **pkts, uint16_t count)
713 struct virtio_net *dev = get_device(vid);
718 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
719 RTE_LOG(ERR, VHOST_DATA,
720 "(%d) %s: built-in vhost net backend is disabled.\n",
725 if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
726 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
728 return virtio_dev_rx(dev, queue_id, pkts, count);
732 virtio_net_with_host_offload(struct virtio_net *dev)
735 ((1ULL << VIRTIO_NET_F_CSUM) |
736 (1ULL << VIRTIO_NET_F_HOST_ECN) |
737 (1ULL << VIRTIO_NET_F_HOST_TSO4) |
738 (1ULL << VIRTIO_NET_F_HOST_TSO6) |
739 (1ULL << VIRTIO_NET_F_HOST_UFO)))
746 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
748 struct ipv4_hdr *ipv4_hdr;
749 struct ipv6_hdr *ipv6_hdr;
751 struct ether_hdr *eth_hdr;
754 eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
756 m->l2_len = sizeof(struct ether_hdr);
757 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
759 if (ethertype == ETHER_TYPE_VLAN) {
760 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
762 m->l2_len += sizeof(struct vlan_hdr);
763 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
766 l3_hdr = (char *)eth_hdr + m->l2_len;
769 case ETHER_TYPE_IPv4:
771 *l4_proto = ipv4_hdr->next_proto_id;
772 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
773 *l4_hdr = (char *)l3_hdr + m->l3_len;
774 m->ol_flags |= PKT_TX_IPV4;
776 case ETHER_TYPE_IPv6:
778 *l4_proto = ipv6_hdr->proto;
779 m->l3_len = sizeof(struct ipv6_hdr);
780 *l4_hdr = (char *)l3_hdr + m->l3_len;
781 m->ol_flags |= PKT_TX_IPV6;
791 static __rte_always_inline void
792 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
794 uint16_t l4_proto = 0;
796 struct tcp_hdr *tcp_hdr = NULL;
798 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
801 parse_ethernet(m, &l4_proto, &l4_hdr);
802 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
803 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
804 switch (hdr->csum_offset) {
805 case (offsetof(struct tcp_hdr, cksum)):
806 if (l4_proto == IPPROTO_TCP)
807 m->ol_flags |= PKT_TX_TCP_CKSUM;
809 case (offsetof(struct udp_hdr, dgram_cksum)):
810 if (l4_proto == IPPROTO_UDP)
811 m->ol_flags |= PKT_TX_UDP_CKSUM;
813 case (offsetof(struct sctp_hdr, cksum)):
814 if (l4_proto == IPPROTO_SCTP)
815 m->ol_flags |= PKT_TX_SCTP_CKSUM;
823 if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
824 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
825 case VIRTIO_NET_HDR_GSO_TCPV4:
826 case VIRTIO_NET_HDR_GSO_TCPV6:
828 m->ol_flags |= PKT_TX_TCP_SEG;
829 m->tso_segsz = hdr->gso_size;
830 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
832 case VIRTIO_NET_HDR_GSO_UDP:
833 m->ol_flags |= PKT_TX_UDP_SEG;
834 m->tso_segsz = hdr->gso_size;
835 m->l4_len = sizeof(struct udp_hdr);
838 RTE_LOG(WARNING, VHOST_DATA,
839 "unsupported gso type %u.\n", hdr->gso_type);
845 static __rte_always_inline void
846 put_zmbuf(struct zcopy_mbuf *zmbuf)
851 static __rte_always_inline int
852 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
853 struct vring_desc *descs, uint16_t max_desc,
854 struct rte_mbuf *m, uint16_t desc_idx,
855 struct rte_mempool *mbuf_pool)
857 struct vring_desc *desc;
859 uint32_t desc_avail, desc_offset;
860 uint32_t mbuf_avail, mbuf_offset;
863 struct rte_mbuf *cur = m, *prev = m;
864 struct virtio_net_hdr *hdr = NULL;
865 /* A counter to avoid desc dead loop chain */
866 uint32_t nr_desc = 1;
867 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
868 uint16_t copy_nb = vq->batch_copy_nb_elems;
871 desc = &descs[desc_idx];
872 if (unlikely((desc->len < dev->vhost_hlen)) ||
873 (desc->flags & VRING_DESC_F_INDIRECT)) {
879 desc_addr = vhost_iova_to_vva(dev,
883 if (unlikely(!desc_addr || dlen != desc->len)) {
888 if (virtio_net_with_host_offload(dev)) {
889 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
894 * A virtio driver normally uses at least 2 desc buffers
895 * for Tx: the first for storing the header, and others
896 * for storing the data.
898 if (likely((desc->len == dev->vhost_hlen) &&
899 (desc->flags & VRING_DESC_F_NEXT) != 0)) {
900 desc = &descs[desc->next];
901 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
907 desc_addr = vhost_iova_to_vva(dev,
911 if (unlikely(!desc_addr || dlen != desc->len)) {
917 desc_avail = desc->len;
920 desc_avail = desc->len - dev->vhost_hlen;
921 desc_offset = dev->vhost_hlen;
924 rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
926 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
929 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
933 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
936 * A desc buf might across two host physical pages that are
937 * not continuous. In such case (gpa_to_hpa returns 0), data
938 * will be copied even though zero copy is enabled.
940 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
941 desc->addr + desc_offset, cpy_len)))) {
942 cur->data_len = cpy_len;
944 cur->buf_addr = (void *)(uintptr_t)(desc_addr
949 * In zero copy mode, one mbuf can only reference data
950 * for one or partial of one desc buff.
952 mbuf_avail = cpy_len;
954 if (likely(cpy_len > MAX_BATCH_LEN ||
955 copy_nb >= vq->size ||
956 (hdr && cur == m))) {
957 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
959 (void *)((uintptr_t)(desc_addr +
963 batch_copy[copy_nb].dst =
964 rte_pktmbuf_mtod_offset(cur, void *,
966 batch_copy[copy_nb].src =
967 (void *)((uintptr_t)(desc_addr +
969 batch_copy[copy_nb].len = cpy_len;
974 mbuf_avail -= cpy_len;
975 mbuf_offset += cpy_len;
976 desc_avail -= cpy_len;
977 desc_offset += cpy_len;
979 /* This desc reaches to its end, get the next one */
980 if (desc_avail == 0) {
981 if ((desc->flags & VRING_DESC_F_NEXT) == 0)
984 if (unlikely(desc->next >= max_desc ||
985 ++nr_desc > max_desc)) {
989 desc = &descs[desc->next];
990 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
996 desc_addr = vhost_iova_to_vva(dev,
998 &dlen, VHOST_ACCESS_RO);
999 if (unlikely(!desc_addr || dlen != desc->len)) {
1004 rte_prefetch0((void *)(uintptr_t)desc_addr);
1007 desc_avail = desc->len;
1009 PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
1013 * This mbuf reaches to its end, get a new one
1014 * to hold more data.
1016 if (mbuf_avail == 0) {
1017 cur = rte_pktmbuf_alloc(mbuf_pool);
1018 if (unlikely(cur == NULL)) {
1019 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1020 "allocate memory for mbuf.\n");
1024 if (unlikely(dev->dequeue_zero_copy))
1025 rte_mbuf_refcnt_update(cur, 1);
1028 prev->data_len = mbuf_offset;
1030 m->pkt_len += mbuf_offset;
1034 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1038 prev->data_len = mbuf_offset;
1039 m->pkt_len += mbuf_offset;
1042 vhost_dequeue_offload(hdr, m);
1045 vq->batch_copy_nb_elems = copy_nb;
1050 static __rte_always_inline void
1051 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
1052 uint32_t used_idx, uint32_t desc_idx)
1054 vq->used->ring[used_idx].id = desc_idx;
1055 vq->used->ring[used_idx].len = 0;
1056 vhost_log_used_vring(dev, vq,
1057 offsetof(struct vring_used, ring[used_idx]),
1058 sizeof(vq->used->ring[used_idx]));
1061 static __rte_always_inline void
1062 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
1065 if (unlikely(count == 0))
1071 vq->used->idx += count;
1072 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
1073 sizeof(vq->used->idx));
1074 vhost_vring_call(dev, vq);
1077 static __rte_always_inline struct zcopy_mbuf *
1078 get_zmbuf(struct vhost_virtqueue *vq)
1084 /* search [last_zmbuf_idx, zmbuf_size) */
1085 i = vq->last_zmbuf_idx;
1086 last = vq->zmbuf_size;
1089 for (; i < last; i++) {
1090 if (vq->zmbufs[i].in_use == 0) {
1091 vq->last_zmbuf_idx = i + 1;
1092 vq->zmbufs[i].in_use = 1;
1093 return &vq->zmbufs[i];
1099 /* search [0, last_zmbuf_idx) */
1101 last = vq->last_zmbuf_idx;
1108 static __rte_always_inline bool
1109 mbuf_is_consumed(struct rte_mbuf *m)
1112 if (rte_mbuf_refcnt_read(m) > 1)
1120 static __rte_always_inline void
1121 restore_mbuf(struct rte_mbuf *m)
1123 uint32_t mbuf_size, priv_size;
1126 priv_size = rte_pktmbuf_priv_size(m->pool);
1127 mbuf_size = sizeof(struct rte_mbuf) + priv_size;
1128 /* start of buffer is after mbuf structure and priv data */
1130 m->buf_addr = (char *)m + mbuf_size;
1131 m->buf_iova = rte_mempool_virt2iova(m) + mbuf_size;
1137 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1138 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1140 struct virtio_net *dev;
1141 struct rte_mbuf *rarp_mbuf = NULL;
1142 struct vhost_virtqueue *vq;
1143 uint32_t desc_indexes[MAX_PKT_BURST];
1146 uint16_t free_entries;
1149 dev = get_device(vid);
1153 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1154 RTE_LOG(ERR, VHOST_DATA,
1155 "(%d) %s: built-in vhost net backend is disabled.\n",
1156 dev->vid, __func__);
1160 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1161 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1162 dev->vid, __func__, queue_id);
1166 vq = dev->virtqueue[queue_id];
1168 if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
1171 if (unlikely(vq->enabled == 0))
1172 goto out_access_unlock;
1174 vq->batch_copy_nb_elems = 0;
1176 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1177 vhost_user_iotlb_rd_lock(vq);
1179 if (unlikely(vq->access_ok == 0))
1180 if (unlikely(vring_translate(dev, vq) < 0))
1183 if (unlikely(dev->dequeue_zero_copy)) {
1184 struct zcopy_mbuf *zmbuf, *next;
1187 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1188 zmbuf != NULL; zmbuf = next) {
1189 next = TAILQ_NEXT(zmbuf, next);
1191 if (mbuf_is_consumed(zmbuf->mbuf)) {
1192 used_idx = vq->last_used_idx++ & (vq->size - 1);
1193 update_used_ring(dev, vq, used_idx,
1197 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1198 restore_mbuf(zmbuf->mbuf);
1199 rte_pktmbuf_free(zmbuf->mbuf);
1205 update_used_idx(dev, vq, nr_updated);
1209 * Construct a RARP broadcast packet, and inject it to the "pkts"
1210 * array, to looks like that guest actually send such packet.
1212 * Check user_send_rarp() for more information.
1214 * broadcast_rarp shares a cacheline in the virtio_net structure
1215 * with some fields that are accessed during enqueue and
1216 * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1217 * result in false sharing between enqueue and dequeue.
1219 * Prevent unnecessary false sharing by reading broadcast_rarp first
1220 * and only performing cmpset if the read indicates it is likely to
1224 if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1225 rte_atomic16_cmpset((volatile uint16_t *)
1226 &dev->broadcast_rarp.cnt, 1, 0))) {
1228 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
1229 if (rarp_mbuf == NULL) {
1230 RTE_LOG(ERR, VHOST_DATA,
1231 "Failed to make RARP packet.\n");
1237 free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1239 if (free_entries == 0)
1242 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1244 /* Prefetch available and used ring */
1245 avail_idx = vq->last_avail_idx & (vq->size - 1);
1246 used_idx = vq->last_used_idx & (vq->size - 1);
1247 rte_prefetch0(&vq->avail->ring[avail_idx]);
1248 rte_prefetch0(&vq->used->ring[used_idx]);
1250 count = RTE_MIN(count, MAX_PKT_BURST);
1251 count = RTE_MIN(count, free_entries);
1252 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1255 /* Retrieve all of the head indexes first to avoid caching issues. */
1256 for (i = 0; i < count; i++) {
1257 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1258 used_idx = (vq->last_used_idx + i) & (vq->size - 1);
1259 desc_indexes[i] = vq->avail->ring[avail_idx];
1261 if (likely(dev->dequeue_zero_copy == 0))
1262 update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1265 /* Prefetch descriptor index. */
1266 rte_prefetch0(&vq->desc[desc_indexes[0]]);
1267 for (i = 0; i < count; i++) {
1268 struct vring_desc *desc;
1273 if (likely(i + 1 < count))
1274 rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1276 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1277 dlen = vq->desc[desc_indexes[i]].len;
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 ||
1284 dlen != vq->desc[desc_indexes[i]].len))
1287 rte_prefetch0(desc);
1288 sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1293 idx = desc_indexes[i];
1296 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1297 if (unlikely(pkts[i] == NULL)) {
1298 RTE_LOG(ERR, VHOST_DATA,
1299 "Failed to allocate memory for mbuf.\n");
1303 err = copy_desc_to_mbuf(dev, vq, desc, sz, pkts[i], idx,
1305 if (unlikely(err)) {
1306 rte_pktmbuf_free(pkts[i]);
1310 if (unlikely(dev->dequeue_zero_copy)) {
1311 struct zcopy_mbuf *zmbuf;
1313 zmbuf = get_zmbuf(vq);
1315 rte_pktmbuf_free(pkts[i]);
1318 zmbuf->mbuf = pkts[i];
1319 zmbuf->desc_idx = desc_indexes[i];
1322 * Pin lock the mbuf; we will check later to see
1323 * whether the mbuf is freed (when we are the last
1324 * user) or not. If that's the case, we then could
1325 * update the used ring safely.
1327 rte_mbuf_refcnt_update(pkts[i], 1);
1330 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1333 vq->last_avail_idx += i;
1335 if (likely(dev->dequeue_zero_copy == 0)) {
1336 do_data_copy_dequeue(vq);
1337 vq->last_used_idx += i;
1338 update_used_idx(dev, vq, i);
1342 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1343 vhost_user_iotlb_rd_unlock(vq);
1346 rte_spinlock_unlock(&vq->access_lock);
1348 if (unlikely(rarp_mbuf != NULL)) {
1350 * Inject it to the head of "pkts" array, so that switch's mac
1351 * learning table will get updated first.
1353 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1354 pkts[0] = rarp_mbuf;