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;
183 struct vring_desc *desc;
185 /* A counter to avoid desc dead loop chain */
186 uint16_t nr_desc = 1;
187 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
188 uint16_t copy_nb = vq->batch_copy_nb_elems;
191 desc = &descs[desc_idx];
192 desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
193 desc->len, VHOST_ACCESS_RW);
195 * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
196 * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
197 * otherwise stores offset on the stack instead of in a register.
199 if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr) {
204 rte_prefetch0((void *)(uintptr_t)desc_addr);
206 virtio_enqueue_offload(m, (struct virtio_net_hdr *)(uintptr_t)desc_addr);
207 vhost_log_write(dev, desc->addr, dev->vhost_hlen);
208 PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
210 desc_offset = dev->vhost_hlen;
211 desc_avail = desc->len - dev->vhost_hlen;
213 mbuf_avail = rte_pktmbuf_data_len(m);
215 while (mbuf_avail != 0 || m->next != NULL) {
216 /* done with current mbuf, fetch next */
217 if (mbuf_avail == 0) {
221 mbuf_avail = rte_pktmbuf_data_len(m);
224 /* done with current desc buf, fetch next */
225 if (desc_avail == 0) {
226 if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
227 /* Room in vring buffer is not enough */
231 if (unlikely(desc->next >= size || ++nr_desc > size)) {
236 desc = &descs[desc->next];
237 desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
240 if (unlikely(!desc_addr)) {
246 desc_avail = desc->len;
249 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
250 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
251 rte_memcpy((void *)((uintptr_t)(desc_addr +
253 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
255 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
256 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
259 batch_copy[copy_nb].dst =
260 (void *)((uintptr_t)(desc_addr + desc_offset));
261 batch_copy[copy_nb].src =
262 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
263 batch_copy[copy_nb].log_addr = desc->addr + desc_offset;
264 batch_copy[copy_nb].len = cpy_len;
268 mbuf_avail -= cpy_len;
269 mbuf_offset += cpy_len;
270 desc_avail -= cpy_len;
271 desc_offset += cpy_len;
275 vq->batch_copy_nb_elems = copy_nb;
281 * This function adds buffers to the virtio devices RX virtqueue. Buffers can
282 * be received from the physical port or from another virtio device. A packet
283 * count is returned to indicate the number of packets that are successfully
284 * added to the RX queue. This function works when the mbuf is scattered, but
285 * it doesn't support the mergeable feature.
287 static __rte_always_inline uint32_t
288 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
289 struct rte_mbuf **pkts, uint32_t count)
291 struct vhost_virtqueue *vq;
292 uint16_t avail_idx, free_entries, start_idx;
293 uint16_t desc_indexes[MAX_PKT_BURST];
294 struct vring_desc *descs;
298 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
299 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
300 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
301 dev->vid, __func__, queue_id);
305 vq = dev->virtqueue[queue_id];
307 rte_spinlock_lock(&vq->access_lock);
309 if (unlikely(vq->enabled == 0))
310 goto out_access_unlock;
312 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
313 vhost_user_iotlb_rd_lock(vq);
315 if (unlikely(vq->access_ok == 0)) {
316 if (unlikely(vring_translate(dev, vq) < 0)) {
322 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
323 start_idx = vq->last_used_idx;
324 free_entries = avail_idx - start_idx;
325 count = RTE_MIN(count, free_entries);
326 count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
330 LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
331 dev->vid, start_idx, start_idx + count);
333 vq->batch_copy_nb_elems = 0;
335 /* Retrieve all of the desc indexes first to avoid caching issues. */
336 rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
337 for (i = 0; i < count; i++) {
338 used_idx = (start_idx + i) & (vq->size - 1);
339 desc_indexes[i] = vq->avail->ring[used_idx];
340 vq->used->ring[used_idx].id = desc_indexes[i];
341 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
343 vhost_log_used_vring(dev, vq,
344 offsetof(struct vring_used, ring[used_idx]),
345 sizeof(vq->used->ring[used_idx]));
348 rte_prefetch0(&vq->desc[desc_indexes[0]]);
349 for (i = 0; i < count; i++) {
350 uint16_t desc_idx = desc_indexes[i];
353 if (vq->desc[desc_idx].flags & VRING_DESC_F_INDIRECT) {
354 descs = (struct vring_desc *)(uintptr_t)
355 vhost_iova_to_vva(dev,
356 vq, vq->desc[desc_idx].addr,
357 vq->desc[desc_idx].len,
359 if (unlikely(!descs)) {
365 sz = vq->desc[desc_idx].len / sizeof(*descs);
371 err = copy_mbuf_to_desc(dev, vq, descs, pkts[i], desc_idx, sz);
378 rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
381 do_data_copy_enqueue(dev, vq);
385 *(volatile uint16_t *)&vq->used->idx += count;
386 vq->last_used_idx += count;
387 vhost_log_used_vring(dev, vq,
388 offsetof(struct vring_used, idx),
389 sizeof(vq->used->idx));
391 vhost_vring_call(dev, vq);
393 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
394 vhost_user_iotlb_rd_unlock(vq);
397 rte_spinlock_unlock(&vq->access_lock);
402 static __rte_always_inline int
403 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
404 uint32_t avail_idx, uint32_t *vec_idx,
405 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
406 uint16_t *desc_chain_len)
408 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
409 uint32_t vec_id = *vec_idx;
411 struct vring_desc *descs = vq->desc;
413 *desc_chain_head = idx;
415 if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
416 descs = (struct vring_desc *)(uintptr_t)
417 vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
420 if (unlikely(!descs))
427 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
430 len += descs[idx].len;
431 buf_vec[vec_id].buf_addr = descs[idx].addr;
432 buf_vec[vec_id].buf_len = descs[idx].len;
433 buf_vec[vec_id].desc_idx = idx;
436 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
439 idx = descs[idx].next;
442 *desc_chain_len = len;
449 * Returns -1 on fail, 0 on success
452 reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
453 uint32_t size, struct buf_vector *buf_vec,
454 uint16_t *num_buffers, uint16_t avail_head)
457 uint32_t vec_idx = 0;
460 uint16_t head_idx = 0;
464 cur_idx = vq->last_avail_idx;
467 if (unlikely(cur_idx == avail_head))
470 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
471 &head_idx, &len) < 0))
473 len = RTE_MIN(len, size);
474 update_shadow_used_ring(vq, head_idx, len);
482 * if we tried all available ring items, and still
483 * can't get enough buf, it means something abnormal
486 if (unlikely(tries >= vq->size))
493 static __rte_always_inline int
494 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
495 struct rte_mbuf *m, struct buf_vector *buf_vec,
496 uint16_t num_buffers)
498 uint32_t vec_idx = 0;
500 uint32_t mbuf_offset, mbuf_avail;
501 uint32_t desc_offset, desc_avail;
503 uint64_t hdr_addr, hdr_phys_addr;
504 struct rte_mbuf *hdr_mbuf;
505 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
506 uint16_t copy_nb = vq->batch_copy_nb_elems;
509 if (unlikely(m == NULL)) {
514 desc_addr = vhost_iova_to_vva(dev, vq, buf_vec[vec_idx].buf_addr,
515 buf_vec[vec_idx].buf_len,
517 if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr) {
523 hdr_addr = desc_addr;
524 hdr_phys_addr = buf_vec[vec_idx].buf_addr;
525 rte_prefetch0((void *)(uintptr_t)hdr_addr);
527 LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
528 dev->vid, num_buffers);
530 desc_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
531 desc_offset = dev->vhost_hlen;
533 mbuf_avail = rte_pktmbuf_data_len(m);
535 while (mbuf_avail != 0 || m->next != NULL) {
536 /* done with current desc buf, get the next one */
537 if (desc_avail == 0) {
540 vhost_iova_to_vva(dev, vq,
541 buf_vec[vec_idx].buf_addr,
542 buf_vec[vec_idx].buf_len,
544 if (unlikely(!desc_addr)) {
549 /* Prefetch buffer address. */
550 rte_prefetch0((void *)(uintptr_t)desc_addr);
552 desc_avail = buf_vec[vec_idx].buf_len;
555 /* done with current mbuf, get the next one */
556 if (mbuf_avail == 0) {
560 mbuf_avail = rte_pktmbuf_data_len(m);
564 struct virtio_net_hdr_mrg_rxbuf *hdr;
566 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)
568 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
569 ASSIGN_UNLESS_EQUAL(hdr->num_buffers, num_buffers);
571 vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
572 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
578 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
580 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
581 rte_memcpy((void *)((uintptr_t)(desc_addr +
583 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
586 buf_vec[vec_idx].buf_addr + desc_offset,
588 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
591 batch_copy[copy_nb].dst =
592 (void *)((uintptr_t)(desc_addr + desc_offset));
593 batch_copy[copy_nb].src =
594 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
595 batch_copy[copy_nb].log_addr =
596 buf_vec[vec_idx].buf_addr + desc_offset;
597 batch_copy[copy_nb].len = cpy_len;
601 mbuf_avail -= cpy_len;
602 mbuf_offset += cpy_len;
603 desc_avail -= cpy_len;
604 desc_offset += cpy_len;
608 vq->batch_copy_nb_elems = copy_nb;
613 static __rte_always_inline uint32_t
614 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
615 struct rte_mbuf **pkts, uint32_t count)
617 struct vhost_virtqueue *vq;
618 uint32_t pkt_idx = 0;
619 uint16_t num_buffers;
620 struct buf_vector buf_vec[BUF_VECTOR_MAX];
623 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
624 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
625 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
626 dev->vid, __func__, queue_id);
630 vq = dev->virtqueue[queue_id];
632 rte_spinlock_lock(&vq->access_lock);
634 if (unlikely(vq->enabled == 0))
635 goto out_access_unlock;
637 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
638 vhost_user_iotlb_rd_lock(vq);
640 if (unlikely(vq->access_ok == 0))
641 if (unlikely(vring_translate(dev, vq) < 0))
644 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
648 vq->batch_copy_nb_elems = 0;
650 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
652 vq->shadow_used_idx = 0;
653 avail_head = *((volatile uint16_t *)&vq->avail->idx);
654 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
655 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
657 if (unlikely(reserve_avail_buf_mergeable(dev, vq,
658 pkt_len, buf_vec, &num_buffers,
660 LOG_DEBUG(VHOST_DATA,
661 "(%d) failed to get enough desc from vring\n",
663 vq->shadow_used_idx -= num_buffers;
667 LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
668 dev->vid, vq->last_avail_idx,
669 vq->last_avail_idx + num_buffers);
671 if (copy_mbuf_to_desc_mergeable(dev, vq, pkts[pkt_idx],
672 buf_vec, num_buffers) < 0) {
673 vq->shadow_used_idx -= num_buffers;
677 vq->last_avail_idx += num_buffers;
680 do_data_copy_enqueue(dev, vq);
682 if (likely(vq->shadow_used_idx)) {
683 flush_shadow_used_ring(dev, vq);
684 vhost_vring_call(dev, vq);
688 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
689 vhost_user_iotlb_rd_unlock(vq);
692 rte_spinlock_unlock(&vq->access_lock);
698 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
699 struct rte_mbuf **pkts, uint16_t count)
701 struct virtio_net *dev = get_device(vid);
706 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
707 RTE_LOG(ERR, VHOST_DATA,
708 "(%d) %s: built-in vhost net backend is disabled.\n",
713 if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
714 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
716 return virtio_dev_rx(dev, queue_id, pkts, count);
720 virtio_net_with_host_offload(struct virtio_net *dev)
723 ((1ULL << VIRTIO_NET_F_CSUM) |
724 (1ULL << VIRTIO_NET_F_HOST_ECN) |
725 (1ULL << VIRTIO_NET_F_HOST_TSO4) |
726 (1ULL << VIRTIO_NET_F_HOST_TSO6) |
727 (1ULL << VIRTIO_NET_F_HOST_UFO)))
734 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
736 struct ipv4_hdr *ipv4_hdr;
737 struct ipv6_hdr *ipv6_hdr;
739 struct ether_hdr *eth_hdr;
742 eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
744 m->l2_len = sizeof(struct ether_hdr);
745 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
747 if (ethertype == ETHER_TYPE_VLAN) {
748 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
750 m->l2_len += sizeof(struct vlan_hdr);
751 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
754 l3_hdr = (char *)eth_hdr + m->l2_len;
757 case ETHER_TYPE_IPv4:
759 *l4_proto = ipv4_hdr->next_proto_id;
760 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
761 *l4_hdr = (char *)l3_hdr + m->l3_len;
762 m->ol_flags |= PKT_TX_IPV4;
764 case ETHER_TYPE_IPv6:
766 *l4_proto = ipv6_hdr->proto;
767 m->l3_len = sizeof(struct ipv6_hdr);
768 *l4_hdr = (char *)l3_hdr + m->l3_len;
769 m->ol_flags |= PKT_TX_IPV6;
779 static __rte_always_inline void
780 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
782 uint16_t l4_proto = 0;
784 struct tcp_hdr *tcp_hdr = NULL;
786 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
789 parse_ethernet(m, &l4_proto, &l4_hdr);
790 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
791 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
792 switch (hdr->csum_offset) {
793 case (offsetof(struct tcp_hdr, cksum)):
794 if (l4_proto == IPPROTO_TCP)
795 m->ol_flags |= PKT_TX_TCP_CKSUM;
797 case (offsetof(struct udp_hdr, dgram_cksum)):
798 if (l4_proto == IPPROTO_UDP)
799 m->ol_flags |= PKT_TX_UDP_CKSUM;
801 case (offsetof(struct sctp_hdr, cksum)):
802 if (l4_proto == IPPROTO_SCTP)
803 m->ol_flags |= PKT_TX_SCTP_CKSUM;
811 if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
812 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
813 case VIRTIO_NET_HDR_GSO_TCPV4:
814 case VIRTIO_NET_HDR_GSO_TCPV6:
816 m->ol_flags |= PKT_TX_TCP_SEG;
817 m->tso_segsz = hdr->gso_size;
818 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
820 case VIRTIO_NET_HDR_GSO_UDP:
821 m->ol_flags |= PKT_TX_UDP_SEG;
822 m->tso_segsz = hdr->gso_size;
823 m->l4_len = sizeof(struct udp_hdr);
826 RTE_LOG(WARNING, VHOST_DATA,
827 "unsupported gso type %u.\n", hdr->gso_type);
833 static __rte_always_inline void
834 put_zmbuf(struct zcopy_mbuf *zmbuf)
839 static __rte_always_inline int
840 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
841 struct vring_desc *descs, uint16_t max_desc,
842 struct rte_mbuf *m, uint16_t desc_idx,
843 struct rte_mempool *mbuf_pool)
845 struct vring_desc *desc;
847 uint32_t desc_avail, desc_offset;
848 uint32_t mbuf_avail, mbuf_offset;
850 struct rte_mbuf *cur = m, *prev = m;
851 struct virtio_net_hdr *hdr = NULL;
852 /* A counter to avoid desc dead loop chain */
853 uint32_t nr_desc = 1;
854 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
855 uint16_t copy_nb = vq->batch_copy_nb_elems;
858 desc = &descs[desc_idx];
859 if (unlikely((desc->len < dev->vhost_hlen)) ||
860 (desc->flags & VRING_DESC_F_INDIRECT)) {
865 desc_addr = vhost_iova_to_vva(dev,
869 if (unlikely(!desc_addr)) {
874 if (virtio_net_with_host_offload(dev)) {
875 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
880 * A virtio driver normally uses at least 2 desc buffers
881 * for Tx: the first for storing the header, and others
882 * for storing the data.
884 if (likely((desc->len == dev->vhost_hlen) &&
885 (desc->flags & VRING_DESC_F_NEXT) != 0)) {
886 desc = &descs[desc->next];
887 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
892 desc_addr = vhost_iova_to_vva(dev,
896 if (unlikely(!desc_addr)) {
902 desc_avail = desc->len;
905 desc_avail = desc->len - dev->vhost_hlen;
906 desc_offset = dev->vhost_hlen;
909 rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
911 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
914 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
918 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
921 * A desc buf might across two host physical pages that are
922 * not continuous. In such case (gpa_to_hpa returns 0), data
923 * will be copied even though zero copy is enabled.
925 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
926 desc->addr + desc_offset, cpy_len)))) {
927 cur->data_len = cpy_len;
929 cur->buf_addr = (void *)(uintptr_t)(desc_addr
934 * In zero copy mode, one mbuf can only reference data
935 * for one or partial of one desc buff.
937 mbuf_avail = cpy_len;
939 if (likely(cpy_len > MAX_BATCH_LEN ||
940 copy_nb >= vq->size ||
941 (hdr && cur == m))) {
942 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
944 (void *)((uintptr_t)(desc_addr +
948 batch_copy[copy_nb].dst =
949 rte_pktmbuf_mtod_offset(cur, void *,
951 batch_copy[copy_nb].src =
952 (void *)((uintptr_t)(desc_addr +
954 batch_copy[copy_nb].len = cpy_len;
959 mbuf_avail -= cpy_len;
960 mbuf_offset += cpy_len;
961 desc_avail -= cpy_len;
962 desc_offset += cpy_len;
964 /* This desc reaches to its end, get the next one */
965 if (desc_avail == 0) {
966 if ((desc->flags & VRING_DESC_F_NEXT) == 0)
969 if (unlikely(desc->next >= max_desc ||
970 ++nr_desc > max_desc)) {
974 desc = &descs[desc->next];
975 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
980 desc_addr = vhost_iova_to_vva(dev,
984 if (unlikely(!desc_addr)) {
989 rte_prefetch0((void *)(uintptr_t)desc_addr);
992 desc_avail = desc->len;
994 PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
998 * This mbuf reaches to its end, get a new one
1001 if (mbuf_avail == 0) {
1002 cur = rte_pktmbuf_alloc(mbuf_pool);
1003 if (unlikely(cur == NULL)) {
1004 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1005 "allocate memory for mbuf.\n");
1009 if (unlikely(dev->dequeue_zero_copy))
1010 rte_mbuf_refcnt_update(cur, 1);
1013 prev->data_len = mbuf_offset;
1015 m->pkt_len += mbuf_offset;
1019 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1023 prev->data_len = mbuf_offset;
1024 m->pkt_len += mbuf_offset;
1027 vhost_dequeue_offload(hdr, m);
1030 vq->batch_copy_nb_elems = copy_nb;
1035 static __rte_always_inline void
1036 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
1037 uint32_t used_idx, uint32_t desc_idx)
1039 vq->used->ring[used_idx].id = desc_idx;
1040 vq->used->ring[used_idx].len = 0;
1041 vhost_log_used_vring(dev, vq,
1042 offsetof(struct vring_used, ring[used_idx]),
1043 sizeof(vq->used->ring[used_idx]));
1046 static __rte_always_inline void
1047 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
1050 if (unlikely(count == 0))
1056 vq->used->idx += count;
1057 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
1058 sizeof(vq->used->idx));
1059 vhost_vring_call(dev, vq);
1062 static __rte_always_inline struct zcopy_mbuf *
1063 get_zmbuf(struct vhost_virtqueue *vq)
1069 /* search [last_zmbuf_idx, zmbuf_size) */
1070 i = vq->last_zmbuf_idx;
1071 last = vq->zmbuf_size;
1074 for (; i < last; i++) {
1075 if (vq->zmbufs[i].in_use == 0) {
1076 vq->last_zmbuf_idx = i + 1;
1077 vq->zmbufs[i].in_use = 1;
1078 return &vq->zmbufs[i];
1084 /* search [0, last_zmbuf_idx) */
1086 last = vq->last_zmbuf_idx;
1093 static __rte_always_inline bool
1094 mbuf_is_consumed(struct rte_mbuf *m)
1097 if (rte_mbuf_refcnt_read(m) > 1)
1105 static __rte_always_inline void
1106 restore_mbuf(struct rte_mbuf *m)
1108 uint32_t mbuf_size, priv_size;
1111 priv_size = rte_pktmbuf_priv_size(m->pool);
1112 mbuf_size = sizeof(struct rte_mbuf) + priv_size;
1113 /* start of buffer is after mbuf structure and priv data */
1115 m->buf_addr = (char *)m + mbuf_size;
1116 m->buf_iova = rte_mempool_virt2iova(m) + mbuf_size;
1122 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1123 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1125 struct virtio_net *dev;
1126 struct rte_mbuf *rarp_mbuf = NULL;
1127 struct vhost_virtqueue *vq;
1128 uint32_t desc_indexes[MAX_PKT_BURST];
1131 uint16_t free_entries;
1134 dev = get_device(vid);
1138 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1139 RTE_LOG(ERR, VHOST_DATA,
1140 "(%d) %s: built-in vhost net backend is disabled.\n",
1141 dev->vid, __func__);
1145 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1146 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1147 dev->vid, __func__, queue_id);
1151 vq = dev->virtqueue[queue_id];
1153 if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
1156 if (unlikely(vq->enabled == 0))
1157 goto out_access_unlock;
1159 vq->batch_copy_nb_elems = 0;
1161 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1162 vhost_user_iotlb_rd_lock(vq);
1164 if (unlikely(vq->access_ok == 0))
1165 if (unlikely(vring_translate(dev, vq) < 0))
1168 if (unlikely(dev->dequeue_zero_copy)) {
1169 struct zcopy_mbuf *zmbuf, *next;
1172 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1173 zmbuf != NULL; zmbuf = next) {
1174 next = TAILQ_NEXT(zmbuf, next);
1176 if (mbuf_is_consumed(zmbuf->mbuf)) {
1177 used_idx = vq->last_used_idx++ & (vq->size - 1);
1178 update_used_ring(dev, vq, used_idx,
1182 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1183 restore_mbuf(zmbuf->mbuf);
1184 rte_pktmbuf_free(zmbuf->mbuf);
1190 update_used_idx(dev, vq, nr_updated);
1194 * Construct a RARP broadcast packet, and inject it to the "pkts"
1195 * array, to looks like that guest actually send such packet.
1197 * Check user_send_rarp() for more information.
1199 * broadcast_rarp shares a cacheline in the virtio_net structure
1200 * with some fields that are accessed during enqueue and
1201 * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1202 * result in false sharing between enqueue and dequeue.
1204 * Prevent unnecessary false sharing by reading broadcast_rarp first
1205 * and only performing cmpset if the read indicates it is likely to
1209 if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1210 rte_atomic16_cmpset((volatile uint16_t *)
1211 &dev->broadcast_rarp.cnt, 1, 0))) {
1213 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
1214 if (rarp_mbuf == NULL) {
1215 RTE_LOG(ERR, VHOST_DATA,
1216 "Failed to make RARP packet.\n");
1222 free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1224 if (free_entries == 0)
1227 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1229 /* Prefetch available and used ring */
1230 avail_idx = vq->last_avail_idx & (vq->size - 1);
1231 used_idx = vq->last_used_idx & (vq->size - 1);
1232 rte_prefetch0(&vq->avail->ring[avail_idx]);
1233 rte_prefetch0(&vq->used->ring[used_idx]);
1235 count = RTE_MIN(count, MAX_PKT_BURST);
1236 count = RTE_MIN(count, free_entries);
1237 LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1240 /* Retrieve all of the head indexes first to avoid caching issues. */
1241 for (i = 0; i < count; i++) {
1242 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1243 used_idx = (vq->last_used_idx + i) & (vq->size - 1);
1244 desc_indexes[i] = vq->avail->ring[avail_idx];
1246 if (likely(dev->dequeue_zero_copy == 0))
1247 update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1250 /* Prefetch descriptor index. */
1251 rte_prefetch0(&vq->desc[desc_indexes[0]]);
1252 for (i = 0; i < count; i++) {
1253 struct vring_desc *desc;
1257 if (likely(i + 1 < count))
1258 rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1260 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1261 desc = (struct vring_desc *)(uintptr_t)
1262 vhost_iova_to_vva(dev, vq,
1263 vq->desc[desc_indexes[i]].addr,
1266 if (unlikely(!desc))
1269 rte_prefetch0(desc);
1270 sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1275 idx = desc_indexes[i];
1278 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1279 if (unlikely(pkts[i] == NULL)) {
1280 RTE_LOG(ERR, VHOST_DATA,
1281 "Failed to allocate memory for mbuf.\n");
1285 err = copy_desc_to_mbuf(dev, vq, desc, sz, pkts[i], idx,
1287 if (unlikely(err)) {
1288 rte_pktmbuf_free(pkts[i]);
1292 if (unlikely(dev->dequeue_zero_copy)) {
1293 struct zcopy_mbuf *zmbuf;
1295 zmbuf = get_zmbuf(vq);
1297 rte_pktmbuf_free(pkts[i]);
1300 zmbuf->mbuf = pkts[i];
1301 zmbuf->desc_idx = desc_indexes[i];
1304 * Pin lock the mbuf; we will check later to see
1305 * whether the mbuf is freed (when we are the last
1306 * user) or not. If that's the case, we then could
1307 * update the used ring safely.
1309 rte_mbuf_refcnt_update(pkts[i], 1);
1312 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1315 vq->last_avail_idx += i;
1317 if (likely(dev->dequeue_zero_copy == 0)) {
1318 do_data_copy_dequeue(vq);
1319 vq->last_used_idx += i;
1320 update_used_idx(dev, vq, i);
1324 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1325 vhost_user_iotlb_rd_unlock(vq);
1328 rte_spinlock_unlock(&vq->access_lock);
1330 if (unlikely(rarp_mbuf != NULL)) {
1332 * Inject it to the head of "pkts" array, so that switch's mac
1333 * learning table will get updated first.
1335 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1336 pkts[0] = rarp_mbuf;