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>
22 #define MAX_PKT_BURST 32
24 #define MAX_BATCH_LEN 256
27 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
29 return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
32 static __rte_always_inline void
33 do_flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
34 uint16_t to, uint16_t from, uint16_t size)
36 rte_memcpy(&vq->used->ring[to],
37 &vq->shadow_used_ring[from],
38 size * sizeof(struct vring_used_elem));
39 vhost_log_used_vring(dev, vq,
40 offsetof(struct vring_used, ring[to]),
41 size * sizeof(struct vring_used_elem));
44 static __rte_always_inline void
45 flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq)
47 uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
49 if (used_idx + vq->shadow_used_idx <= vq->size) {
50 do_flush_shadow_used_ring(dev, vq, used_idx, 0,
55 /* update used ring interval [used_idx, vq->size] */
56 size = vq->size - used_idx;
57 do_flush_shadow_used_ring(dev, vq, used_idx, 0, size);
59 /* update the left half used ring interval [0, left_size] */
60 do_flush_shadow_used_ring(dev, vq, 0, size,
61 vq->shadow_used_idx - size);
63 vq->last_used_idx += vq->shadow_used_idx;
67 *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
68 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
69 sizeof(vq->used->idx));
72 static __rte_always_inline void
73 update_shadow_used_ring(struct vhost_virtqueue *vq,
74 uint16_t desc_idx, uint16_t len)
76 uint16_t i = vq->shadow_used_idx++;
78 vq->shadow_used_ring[i].id = desc_idx;
79 vq->shadow_used_ring[i].len = len;
83 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
85 struct batch_copy_elem *elem = vq->batch_copy_elems;
86 uint16_t count = vq->batch_copy_nb_elems;
89 for (i = 0; i < count; i++) {
90 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
91 vhost_log_write(dev, elem[i].log_addr, elem[i].len);
92 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
97 do_data_copy_dequeue(struct vhost_virtqueue *vq)
99 struct batch_copy_elem *elem = vq->batch_copy_elems;
100 uint16_t count = vq->batch_copy_nb_elems;
103 for (i = 0; i < count; i++)
104 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
107 /* avoid write operation when necessary, to lessen cache issues */
108 #define ASSIGN_UNLESS_EQUAL(var, val) do { \
109 if ((var) != (val)) \
114 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
116 uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
118 if (m_buf->ol_flags & PKT_TX_TCP_SEG)
119 csum_l4 |= PKT_TX_TCP_CKSUM;
122 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
123 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
126 case PKT_TX_TCP_CKSUM:
127 net_hdr->csum_offset = (offsetof(struct tcp_hdr,
130 case PKT_TX_UDP_CKSUM:
131 net_hdr->csum_offset = (offsetof(struct udp_hdr,
134 case PKT_TX_SCTP_CKSUM:
135 net_hdr->csum_offset = (offsetof(struct sctp_hdr,
140 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
141 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
142 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
145 /* IP cksum verification cannot be bypassed, then calculate here */
146 if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
147 struct ipv4_hdr *ipv4_hdr;
149 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct ipv4_hdr *,
151 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
154 if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
155 if (m_buf->ol_flags & PKT_TX_IPV4)
156 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
158 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
159 net_hdr->gso_size = m_buf->tso_segsz;
160 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
163 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
164 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
165 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
169 static __rte_always_inline int
170 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
171 struct vring_desc *descs, struct rte_mbuf *m,
172 uint16_t desc_idx, uint32_t size)
174 uint32_t desc_avail, desc_offset;
175 uint32_t mbuf_avail, mbuf_offset;
177 struct vring_desc *desc;
179 /* A counter to avoid desc dead loop chain */
180 uint16_t nr_desc = 1;
181 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
182 uint16_t copy_nb = vq->batch_copy_nb_elems;
185 desc = &descs[desc_idx];
186 desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
187 desc->len, VHOST_ACCESS_RW);
189 * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
190 * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
191 * otherwise stores offset on the stack instead of in a register.
193 if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr) {
198 rte_prefetch0((void *)(uintptr_t)desc_addr);
200 virtio_enqueue_offload(m, (struct virtio_net_hdr *)(uintptr_t)desc_addr);
201 vhost_log_write(dev, desc->addr, dev->vhost_hlen);
202 PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
204 desc_offset = dev->vhost_hlen;
205 desc_avail = desc->len - dev->vhost_hlen;
207 mbuf_avail = rte_pktmbuf_data_len(m);
209 while (mbuf_avail != 0 || m->next != NULL) {
210 /* done with current mbuf, fetch next */
211 if (mbuf_avail == 0) {
215 mbuf_avail = rte_pktmbuf_data_len(m);
218 /* done with current desc buf, fetch next */
219 if (desc_avail == 0) {
220 if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
221 /* Room in vring buffer is not enough */
225 if (unlikely(desc->next >= size || ++nr_desc > size)) {
230 desc = &descs[desc->next];
231 desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
234 if (unlikely(!desc_addr)) {
240 desc_avail = desc->len;
243 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
244 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
245 rte_memcpy((void *)((uintptr_t)(desc_addr +
247 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
249 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
250 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
253 batch_copy[copy_nb].dst =
254 (void *)((uintptr_t)(desc_addr + desc_offset));
255 batch_copy[copy_nb].src =
256 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
257 batch_copy[copy_nb].log_addr = desc->addr + desc_offset;
258 batch_copy[copy_nb].len = cpy_len;
262 mbuf_avail -= cpy_len;
263 mbuf_offset += cpy_len;
264 desc_avail -= cpy_len;
265 desc_offset += cpy_len;
269 vq->batch_copy_nb_elems = copy_nb;
275 * This function adds buffers to the virtio devices RX virtqueue. Buffers can
276 * be received from the physical port or from another virtio device. A packet
277 * count is returned to indicate the number of packets that are successfully
278 * added to the RX queue. This function works when the mbuf is scattered, but
279 * it doesn't support the mergeable feature.
281 static __rte_always_inline uint32_t
282 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
283 struct rte_mbuf **pkts, uint32_t count)
285 struct vhost_virtqueue *vq;
286 uint16_t avail_idx, free_entries, start_idx;
287 uint16_t desc_indexes[MAX_PKT_BURST];
288 struct vring_desc *descs;
292 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
293 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
294 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
295 dev->vid, __func__, queue_id);
299 vq = dev->virtqueue[queue_id];
300 if (unlikely(vq->enabled == 0))
303 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
304 vhost_user_iotlb_rd_lock(vq);
306 if (unlikely(vq->access_ok == 0)) {
307 if (unlikely(vring_translate(dev, vq) < 0)) {
313 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
314 start_idx = vq->last_used_idx;
315 free_entries = avail_idx - start_idx;
316 count = RTE_MIN(count, free_entries);
317 count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
321 LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
322 dev->vid, start_idx, start_idx + count);
324 vq->batch_copy_nb_elems = 0;
326 /* Retrieve all of the desc indexes first to avoid caching issues. */
327 rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
328 for (i = 0; i < count; i++) {
329 used_idx = (start_idx + i) & (vq->size - 1);
330 desc_indexes[i] = vq->avail->ring[used_idx];
331 vq->used->ring[used_idx].id = desc_indexes[i];
332 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
334 vhost_log_used_vring(dev, vq,
335 offsetof(struct vring_used, ring[used_idx]),
336 sizeof(vq->used->ring[used_idx]));
339 rte_prefetch0(&vq->desc[desc_indexes[0]]);
340 for (i = 0; i < count; i++) {
341 uint16_t desc_idx = desc_indexes[i];
344 if (vq->desc[desc_idx].flags & VRING_DESC_F_INDIRECT) {
345 descs = (struct vring_desc *)(uintptr_t)
346 vhost_iova_to_vva(dev,
347 vq, vq->desc[desc_idx].addr,
348 vq->desc[desc_idx].len,
350 if (unlikely(!descs)) {
356 sz = vq->desc[desc_idx].len / sizeof(*descs);
362 err = copy_mbuf_to_desc(dev, vq, descs, pkts[i], desc_idx, sz);
369 rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
372 do_data_copy_enqueue(dev, vq);
376 *(volatile uint16_t *)&vq->used->idx += count;
377 vq->last_used_idx += count;
378 vhost_log_used_vring(dev, vq,
379 offsetof(struct vring_used, idx),
380 sizeof(vq->used->idx));
382 vhost_vring_call(vq);
384 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
385 vhost_user_iotlb_rd_unlock(vq);
390 static __rte_always_inline int
391 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
392 uint32_t avail_idx, uint32_t *vec_idx,
393 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
394 uint16_t *desc_chain_len)
396 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
397 uint32_t vec_id = *vec_idx;
399 struct vring_desc *descs = vq->desc;
401 *desc_chain_head = idx;
403 if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
404 descs = (struct vring_desc *)(uintptr_t)
405 vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
408 if (unlikely(!descs))
415 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
418 len += descs[idx].len;
419 buf_vec[vec_id].buf_addr = descs[idx].addr;
420 buf_vec[vec_id].buf_len = descs[idx].len;
421 buf_vec[vec_id].desc_idx = idx;
424 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
427 idx = descs[idx].next;
430 *desc_chain_len = len;
437 * Returns -1 on fail, 0 on success
440 reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
441 uint32_t size, struct buf_vector *buf_vec,
442 uint16_t *num_buffers, uint16_t avail_head)
445 uint32_t vec_idx = 0;
448 uint16_t head_idx = 0;
452 cur_idx = vq->last_avail_idx;
455 if (unlikely(cur_idx == avail_head))
458 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
459 &head_idx, &len) < 0))
461 len = RTE_MIN(len, size);
462 update_shadow_used_ring(vq, head_idx, len);
470 * if we tried all available ring items, and still
471 * can't get enough buf, it means something abnormal
474 if (unlikely(tries >= vq->size))
481 static __rte_always_inline int
482 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
483 struct rte_mbuf *m, struct buf_vector *buf_vec,
484 uint16_t num_buffers)
486 uint32_t vec_idx = 0;
488 uint32_t mbuf_offset, mbuf_avail;
489 uint32_t desc_offset, desc_avail;
491 uint64_t hdr_addr, hdr_phys_addr;
492 struct rte_mbuf *hdr_mbuf;
493 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
494 uint16_t copy_nb = vq->batch_copy_nb_elems;
497 if (unlikely(m == NULL)) {
502 desc_addr = vhost_iova_to_vva(dev, vq, buf_vec[vec_idx].buf_addr,
503 buf_vec[vec_idx].buf_len,
505 if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr) {
511 hdr_addr = desc_addr;
512 hdr_phys_addr = buf_vec[vec_idx].buf_addr;
513 rte_prefetch0((void *)(uintptr_t)hdr_addr);
515 LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
516 dev->vid, num_buffers);
518 desc_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
519 desc_offset = dev->vhost_hlen;
521 mbuf_avail = rte_pktmbuf_data_len(m);
523 while (mbuf_avail != 0 || m->next != NULL) {
524 /* done with current desc buf, get the next one */
525 if (desc_avail == 0) {
528 vhost_iova_to_vva(dev, vq,
529 buf_vec[vec_idx].buf_addr,
530 buf_vec[vec_idx].buf_len,
532 if (unlikely(!desc_addr)) {
537 /* Prefetch buffer address. */
538 rte_prefetch0((void *)(uintptr_t)desc_addr);
540 desc_avail = buf_vec[vec_idx].buf_len;
543 /* done with current mbuf, get the next one */
544 if (mbuf_avail == 0) {
548 mbuf_avail = rte_pktmbuf_data_len(m);
552 struct virtio_net_hdr_mrg_rxbuf *hdr;
554 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)
556 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
557 ASSIGN_UNLESS_EQUAL(hdr->num_buffers, num_buffers);
559 vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
560 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
566 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
568 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
569 rte_memcpy((void *)((uintptr_t)(desc_addr +
571 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
574 buf_vec[vec_idx].buf_addr + desc_offset,
576 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
579 batch_copy[copy_nb].dst =
580 (void *)((uintptr_t)(desc_addr + desc_offset));
581 batch_copy[copy_nb].src =
582 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
583 batch_copy[copy_nb].log_addr =
584 buf_vec[vec_idx].buf_addr + desc_offset;
585 batch_copy[copy_nb].len = cpy_len;
589 mbuf_avail -= cpy_len;
590 mbuf_offset += cpy_len;
591 desc_avail -= cpy_len;
592 desc_offset += cpy_len;
596 vq->batch_copy_nb_elems = copy_nb;
601 static __rte_always_inline uint32_t
602 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
603 struct rte_mbuf **pkts, uint32_t count)
605 struct vhost_virtqueue *vq;
606 uint32_t pkt_idx = 0;
607 uint16_t num_buffers;
608 struct buf_vector buf_vec[BUF_VECTOR_MAX];
611 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
612 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
613 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
614 dev->vid, __func__, queue_id);
618 vq = dev->virtqueue[queue_id];
619 if (unlikely(vq->enabled == 0))
622 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
623 vhost_user_iotlb_rd_lock(vq);
625 if (unlikely(vq->access_ok == 0))
626 if (unlikely(vring_translate(dev, vq) < 0))
629 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
633 vq->batch_copy_nb_elems = 0;
635 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
637 vq->shadow_used_idx = 0;
638 avail_head = *((volatile uint16_t *)&vq->avail->idx);
639 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
640 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
642 if (unlikely(reserve_avail_buf_mergeable(dev, vq,
643 pkt_len, buf_vec, &num_buffers,
645 LOG_DEBUG(VHOST_DATA,
646 "(%d) failed to get enough desc from vring\n",
648 vq->shadow_used_idx -= num_buffers;
652 LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
653 dev->vid, vq->last_avail_idx,
654 vq->last_avail_idx + num_buffers);
656 if (copy_mbuf_to_desc_mergeable(dev, vq, pkts[pkt_idx],
657 buf_vec, num_buffers) < 0) {
658 vq->shadow_used_idx -= num_buffers;
662 vq->last_avail_idx += num_buffers;
665 do_data_copy_enqueue(dev, vq);
667 if (likely(vq->shadow_used_idx)) {
668 flush_shadow_used_ring(dev, vq);
669 vhost_vring_call(vq);
673 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
674 vhost_user_iotlb_rd_unlock(vq);
680 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
681 struct rte_mbuf **pkts, uint16_t count)
683 struct virtio_net *dev = get_device(vid);
688 if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
689 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
691 return virtio_dev_rx(dev, queue_id, pkts, count);
695 virtio_net_with_host_offload(struct virtio_net *dev)
698 ((1ULL << VIRTIO_NET_F_CSUM) |
699 (1ULL << VIRTIO_NET_F_HOST_ECN) |
700 (1ULL << VIRTIO_NET_F_HOST_TSO4) |
701 (1ULL << VIRTIO_NET_F_HOST_TSO6) |
702 (1ULL << VIRTIO_NET_F_HOST_UFO)))
709 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
711 struct ipv4_hdr *ipv4_hdr;
712 struct ipv6_hdr *ipv6_hdr;
714 struct ether_hdr *eth_hdr;
717 eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
719 m->l2_len = sizeof(struct ether_hdr);
720 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
722 if (ethertype == ETHER_TYPE_VLAN) {
723 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
725 m->l2_len += sizeof(struct vlan_hdr);
726 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
729 l3_hdr = (char *)eth_hdr + m->l2_len;
732 case ETHER_TYPE_IPv4:
734 *l4_proto = ipv4_hdr->next_proto_id;
735 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
736 *l4_hdr = (char *)l3_hdr + m->l3_len;
737 m->ol_flags |= PKT_TX_IPV4;
739 case ETHER_TYPE_IPv6:
741 *l4_proto = ipv6_hdr->proto;
742 m->l3_len = sizeof(struct ipv6_hdr);
743 *l4_hdr = (char *)l3_hdr + m->l3_len;
744 m->ol_flags |= PKT_TX_IPV6;
754 static __rte_always_inline void
755 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
757 uint16_t l4_proto = 0;
759 struct tcp_hdr *tcp_hdr = NULL;
761 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
764 parse_ethernet(m, &l4_proto, &l4_hdr);
765 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
766 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
767 switch (hdr->csum_offset) {
768 case (offsetof(struct tcp_hdr, cksum)):
769 if (l4_proto == IPPROTO_TCP)
770 m->ol_flags |= PKT_TX_TCP_CKSUM;
772 case (offsetof(struct udp_hdr, dgram_cksum)):
773 if (l4_proto == IPPROTO_UDP)
774 m->ol_flags |= PKT_TX_UDP_CKSUM;
776 case (offsetof(struct sctp_hdr, cksum)):
777 if (l4_proto == IPPROTO_SCTP)
778 m->ol_flags |= PKT_TX_SCTP_CKSUM;
786 if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
787 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
788 case VIRTIO_NET_HDR_GSO_TCPV4:
789 case VIRTIO_NET_HDR_GSO_TCPV6:
791 m->ol_flags |= PKT_TX_TCP_SEG;
792 m->tso_segsz = hdr->gso_size;
793 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
796 RTE_LOG(WARNING, VHOST_DATA,
797 "unsupported gso type %u.\n", hdr->gso_type);
803 #define RARP_PKT_SIZE 64
806 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
808 struct ether_hdr *eth_hdr;
809 struct arp_hdr *rarp;
811 if (rarp_mbuf->buf_len < 64) {
812 RTE_LOG(WARNING, VHOST_DATA,
813 "failed to make RARP; mbuf size too small %u (< %d)\n",
814 rarp_mbuf->buf_len, RARP_PKT_SIZE);
818 /* Ethernet header. */
819 eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
820 memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
821 ether_addr_copy(mac, ð_hdr->s_addr);
822 eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
825 rarp = (struct arp_hdr *)(eth_hdr + 1);
826 rarp->arp_hrd = htons(ARP_HRD_ETHER);
827 rarp->arp_pro = htons(ETHER_TYPE_IPv4);
828 rarp->arp_hln = ETHER_ADDR_LEN;
830 rarp->arp_op = htons(ARP_OP_REVREQUEST);
832 ether_addr_copy(mac, &rarp->arp_data.arp_sha);
833 ether_addr_copy(mac, &rarp->arp_data.arp_tha);
834 memset(&rarp->arp_data.arp_sip, 0x00, 4);
835 memset(&rarp->arp_data.arp_tip, 0x00, 4);
837 rarp_mbuf->pkt_len = rarp_mbuf->data_len = RARP_PKT_SIZE;
842 static __rte_always_inline void
843 put_zmbuf(struct zcopy_mbuf *zmbuf)
848 static __rte_always_inline int
849 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
850 struct vring_desc *descs, uint16_t max_desc,
851 struct rte_mbuf *m, uint16_t desc_idx,
852 struct rte_mempool *mbuf_pool)
854 struct vring_desc *desc;
856 uint32_t desc_avail, desc_offset;
857 uint32_t mbuf_avail, mbuf_offset;
859 struct rte_mbuf *cur = m, *prev = m;
860 struct virtio_net_hdr *hdr = NULL;
861 /* A counter to avoid desc dead loop chain */
862 uint32_t nr_desc = 1;
863 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
864 uint16_t copy_nb = vq->batch_copy_nb_elems;
867 desc = &descs[desc_idx];
868 if (unlikely((desc->len < dev->vhost_hlen)) ||
869 (desc->flags & VRING_DESC_F_INDIRECT)) {
874 desc_addr = vhost_iova_to_vva(dev,
878 if (unlikely(!desc_addr)) {
883 if (virtio_net_with_host_offload(dev)) {
884 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
889 * A virtio driver normally uses at least 2 desc buffers
890 * for Tx: the first for storing the header, and others
891 * for storing the data.
893 if (likely((desc->len == dev->vhost_hlen) &&
894 (desc->flags & VRING_DESC_F_NEXT) != 0)) {
895 desc = &descs[desc->next];
896 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
901 desc_addr = vhost_iova_to_vva(dev,
905 if (unlikely(!desc_addr)) {
911 desc_avail = desc->len;
914 desc_avail = desc->len - dev->vhost_hlen;
915 desc_offset = dev->vhost_hlen;
918 rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
920 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
923 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
927 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
930 * A desc buf might across two host physical pages that are
931 * not continuous. In such case (gpa_to_hpa returns 0), data
932 * will be copied even though zero copy is enabled.
934 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
935 desc->addr + desc_offset, cpy_len)))) {
936 cur->data_len = cpy_len;
938 cur->buf_addr = (void *)(uintptr_t)(desc_addr
943 * In zero copy mode, one mbuf can only reference data
944 * for one or partial of one desc buff.
946 mbuf_avail = cpy_len;
948 if (likely(cpy_len > MAX_BATCH_LEN ||
949 copy_nb >= vq->size ||
950 (hdr && cur == m))) {
951 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
953 (void *)((uintptr_t)(desc_addr +
957 batch_copy[copy_nb].dst =
958 rte_pktmbuf_mtod_offset(cur, void *,
960 batch_copy[copy_nb].src =
961 (void *)((uintptr_t)(desc_addr +
963 batch_copy[copy_nb].len = cpy_len;
968 mbuf_avail -= cpy_len;
969 mbuf_offset += cpy_len;
970 desc_avail -= cpy_len;
971 desc_offset += cpy_len;
973 /* This desc reaches to its end, get the next one */
974 if (desc_avail == 0) {
975 if ((desc->flags & VRING_DESC_F_NEXT) == 0)
978 if (unlikely(desc->next >= max_desc ||
979 ++nr_desc > max_desc)) {
983 desc = &descs[desc->next];
984 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
989 desc_addr = vhost_iova_to_vva(dev,
993 if (unlikely(!desc_addr)) {
998 rte_prefetch0((void *)(uintptr_t)desc_addr);
1001 desc_avail = desc->len;
1003 PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
1007 * This mbuf reaches to its end, get a new one
1008 * to hold more data.
1010 if (mbuf_avail == 0) {
1011 cur = rte_pktmbuf_alloc(mbuf_pool);
1012 if (unlikely(cur == NULL)) {
1013 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1014 "allocate memory for mbuf.\n");
1018 if (unlikely(dev->dequeue_zero_copy))
1019 rte_mbuf_refcnt_update(cur, 1);
1022 prev->data_len = mbuf_offset;
1024 m->pkt_len += mbuf_offset;
1028 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1032 prev->data_len = mbuf_offset;
1033 m->pkt_len += mbuf_offset;
1036 vhost_dequeue_offload(hdr, m);
1039 vq->batch_copy_nb_elems = copy_nb;
1044 static __rte_always_inline void
1045 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
1046 uint32_t used_idx, uint32_t desc_idx)
1048 vq->used->ring[used_idx].id = desc_idx;
1049 vq->used->ring[used_idx].len = 0;
1050 vhost_log_used_vring(dev, vq,
1051 offsetof(struct vring_used, ring[used_idx]),
1052 sizeof(vq->used->ring[used_idx]));
1055 static __rte_always_inline void
1056 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
1059 if (unlikely(count == 0))
1065 vq->used->idx += count;
1066 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
1067 sizeof(vq->used->idx));
1068 vhost_vring_call(vq);
1071 static __rte_always_inline struct zcopy_mbuf *
1072 get_zmbuf(struct vhost_virtqueue *vq)
1078 /* search [last_zmbuf_idx, zmbuf_size) */
1079 i = vq->last_zmbuf_idx;
1080 last = vq->zmbuf_size;
1083 for (; i < last; i++) {
1084 if (vq->zmbufs[i].in_use == 0) {
1085 vq->last_zmbuf_idx = i + 1;
1086 vq->zmbufs[i].in_use = 1;
1087 return &vq->zmbufs[i];
1093 /* search [0, last_zmbuf_idx) */
1095 last = vq->last_zmbuf_idx;
1102 static __rte_always_inline bool
1103 mbuf_is_consumed(struct rte_mbuf *m)
1106 if (rte_mbuf_refcnt_read(m) > 1)
1115 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1116 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1118 struct virtio_net *dev;
1119 struct rte_mbuf *rarp_mbuf = NULL;
1120 struct vhost_virtqueue *vq;
1121 uint32_t desc_indexes[MAX_PKT_BURST];
1124 uint16_t free_entries;
1127 dev = get_device(vid);
1131 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1132 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1133 dev->vid, __func__, queue_id);
1137 vq = dev->virtqueue[queue_id];
1138 if (unlikely(vq->enabled == 0))
1141 vq->batch_copy_nb_elems = 0;
1143 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1144 vhost_user_iotlb_rd_lock(vq);
1146 if (unlikely(vq->access_ok == 0))
1147 if (unlikely(vring_translate(dev, vq) < 0))
1150 if (unlikely(dev->dequeue_zero_copy)) {
1151 struct zcopy_mbuf *zmbuf, *next;
1154 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1155 zmbuf != NULL; zmbuf = next) {
1156 next = TAILQ_NEXT(zmbuf, next);
1158 if (mbuf_is_consumed(zmbuf->mbuf)) {
1159 used_idx = vq->last_used_idx++ & (vq->size - 1);
1160 update_used_ring(dev, vq, used_idx,
1164 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1165 rte_pktmbuf_free(zmbuf->mbuf);
1171 update_used_idx(dev, vq, nr_updated);
1175 * Construct a RARP broadcast packet, and inject it to the "pkts"
1176 * array, to looks like that guest actually send such packet.
1178 * Check user_send_rarp() for more information.
1180 * broadcast_rarp shares a cacheline in the virtio_net structure
1181 * with some fields that are accessed during enqueue and
1182 * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1183 * result in false sharing between enqueue and dequeue.
1185 * Prevent unnecessary false sharing by reading broadcast_rarp first
1186 * and only performing cmpset if the read indicates it is likely to
1190 if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1191 rte_atomic16_cmpset((volatile uint16_t *)
1192 &dev->broadcast_rarp.cnt, 1, 0))) {
1194 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
1195 if (rarp_mbuf == NULL) {
1196 RTE_LOG(ERR, VHOST_DATA,
1197 "Failed to allocate memory for mbuf.\n");
1201 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
1202 rte_pktmbuf_free(rarp_mbuf);
1209 free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1211 if (free_entries == 0)
1214 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1216 /* Prefetch available and used ring */
1217 avail_idx = vq->last_avail_idx & (vq->size - 1);
1218 used_idx = vq->last_used_idx & (vq->size - 1);
1219 rte_prefetch0(&vq->avail->ring[avail_idx]);
1220 rte_prefetch0(&vq->used->ring[used_idx]);
1222 count = RTE_MIN(count, MAX_PKT_BURST);
1223 count = RTE_MIN(count, free_entries);
1224 LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1227 /* Retrieve all of the head indexes first to avoid caching issues. */
1228 for (i = 0; i < count; i++) {
1229 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1230 used_idx = (vq->last_used_idx + i) & (vq->size - 1);
1231 desc_indexes[i] = vq->avail->ring[avail_idx];
1233 if (likely(dev->dequeue_zero_copy == 0))
1234 update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1237 /* Prefetch descriptor index. */
1238 rte_prefetch0(&vq->desc[desc_indexes[0]]);
1239 for (i = 0; i < count; i++) {
1240 struct vring_desc *desc;
1244 if (likely(i + 1 < count))
1245 rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1247 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1248 desc = (struct vring_desc *)(uintptr_t)
1249 vhost_iova_to_vva(dev, vq,
1250 vq->desc[desc_indexes[i]].addr,
1253 if (unlikely(!desc))
1256 rte_prefetch0(desc);
1257 sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1262 idx = desc_indexes[i];
1265 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1266 if (unlikely(pkts[i] == NULL)) {
1267 RTE_LOG(ERR, VHOST_DATA,
1268 "Failed to allocate memory for mbuf.\n");
1272 err = copy_desc_to_mbuf(dev, vq, desc, sz, pkts[i], idx,
1274 if (unlikely(err)) {
1275 rte_pktmbuf_free(pkts[i]);
1279 if (unlikely(dev->dequeue_zero_copy)) {
1280 struct zcopy_mbuf *zmbuf;
1282 zmbuf = get_zmbuf(vq);
1284 rte_pktmbuf_free(pkts[i]);
1287 zmbuf->mbuf = pkts[i];
1288 zmbuf->desc_idx = desc_indexes[i];
1291 * Pin lock the mbuf; we will check later to see
1292 * whether the mbuf is freed (when we are the last
1293 * user) or not. If that's the case, we then could
1294 * update the used ring safely.
1296 rte_mbuf_refcnt_update(pkts[i], 1);
1299 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1302 vq->last_avail_idx += i;
1304 if (likely(dev->dequeue_zero_copy == 0)) {
1305 do_data_copy_dequeue(vq);
1306 vq->last_used_idx += i;
1307 update_used_idx(dev, vq, i);
1311 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1312 vhost_user_iotlb_rd_unlock(vq);
1314 if (unlikely(rarp_mbuf != NULL)) {
1316 * Inject it to the head of "pkts" array, so that switch's mac
1317 * learning table will get updated first.
1319 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1320 pkts[0] = rarp_mbuf;