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>
19 #include <rte_malloc.h>
24 #define MAX_PKT_BURST 32
26 #define MAX_BATCH_LEN 256
28 static __rte_always_inline bool
29 rxvq_is_mergeable(struct virtio_net *dev)
31 return dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF);
35 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
37 return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
40 static __rte_always_inline struct vring_desc *
41 alloc_copy_ind_table(struct virtio_net *dev, struct vhost_virtqueue *vq,
42 struct vring_desc *desc)
44 struct vring_desc *idesc;
46 uint64_t len, remain = desc->len;
47 uint64_t desc_addr = desc->addr;
49 idesc = rte_malloc(__func__, desc->len, 0);
53 dst = (uint64_t)(uintptr_t)idesc;
57 src = vhost_iova_to_vva(dev, vq, desc_addr, &len,
59 if (unlikely(!src || !len)) {
64 rte_memcpy((void *)(uintptr_t)dst, (void *)(uintptr_t)src, len);
74 static __rte_always_inline void
75 free_ind_table(struct vring_desc *idesc)
80 static __rte_always_inline void
81 do_flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
82 uint16_t to, uint16_t from, uint16_t size)
84 rte_memcpy(&vq->used->ring[to],
85 &vq->shadow_used_ring[from],
86 size * sizeof(struct vring_used_elem));
87 vhost_log_cache_used_vring(dev, vq,
88 offsetof(struct vring_used, ring[to]),
89 size * sizeof(struct vring_used_elem));
92 static __rte_always_inline void
93 flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq)
95 uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
97 if (used_idx + vq->shadow_used_idx <= vq->size) {
98 do_flush_shadow_used_ring(dev, vq, used_idx, 0,
103 /* update used ring interval [used_idx, vq->size] */
104 size = vq->size - used_idx;
105 do_flush_shadow_used_ring(dev, vq, used_idx, 0, size);
107 /* update the left half used ring interval [0, left_size] */
108 do_flush_shadow_used_ring(dev, vq, 0, size,
109 vq->shadow_used_idx - size);
111 vq->last_used_idx += vq->shadow_used_idx;
115 vhost_log_cache_sync(dev, vq);
117 *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
118 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
119 sizeof(vq->used->idx));
122 static __rte_always_inline void
123 update_shadow_used_ring(struct vhost_virtqueue *vq,
124 uint16_t desc_idx, uint16_t len)
126 uint16_t i = vq->shadow_used_idx++;
128 vq->shadow_used_ring[i].id = desc_idx;
129 vq->shadow_used_ring[i].len = len;
133 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
135 struct batch_copy_elem *elem = vq->batch_copy_elems;
136 uint16_t count = vq->batch_copy_nb_elems;
139 for (i = 0; i < count; i++) {
140 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
141 vhost_log_cache_write(dev, vq, elem[i].log_addr, elem[i].len);
142 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
147 do_data_copy_dequeue(struct vhost_virtqueue *vq)
149 struct batch_copy_elem *elem = vq->batch_copy_elems;
150 uint16_t count = vq->batch_copy_nb_elems;
153 for (i = 0; i < count; i++)
154 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
157 /* avoid write operation when necessary, to lessen cache issues */
158 #define ASSIGN_UNLESS_EQUAL(var, val) do { \
159 if ((var) != (val)) \
163 static __rte_always_inline void
164 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
166 uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
168 if (m_buf->ol_flags & PKT_TX_TCP_SEG)
169 csum_l4 |= PKT_TX_TCP_CKSUM;
172 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
173 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
176 case PKT_TX_TCP_CKSUM:
177 net_hdr->csum_offset = (offsetof(struct tcp_hdr,
180 case PKT_TX_UDP_CKSUM:
181 net_hdr->csum_offset = (offsetof(struct udp_hdr,
184 case PKT_TX_SCTP_CKSUM:
185 net_hdr->csum_offset = (offsetof(struct sctp_hdr,
190 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
191 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
192 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
195 /* IP cksum verification cannot be bypassed, then calculate here */
196 if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
197 struct ipv4_hdr *ipv4_hdr;
199 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct ipv4_hdr *,
201 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
204 if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
205 if (m_buf->ol_flags & PKT_TX_IPV4)
206 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
208 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
209 net_hdr->gso_size = m_buf->tso_segsz;
210 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
212 } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
213 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
214 net_hdr->gso_size = m_buf->tso_segsz;
215 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
218 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
219 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
220 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
224 static __rte_always_inline int
225 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
226 uint32_t avail_idx, uint32_t *vec_idx,
227 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
228 uint16_t *desc_chain_len, uint8_t perm)
230 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
231 uint32_t vec_id = *vec_idx;
233 uint64_t dlen, desc_avail, desc_iova;
234 struct vring_desc *descs = vq->desc;
235 struct vring_desc *idesc = NULL;
237 *desc_chain_head = idx;
239 if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
240 dlen = vq->desc[idx].len;
241 descs = (struct vring_desc *)(uintptr_t)
242 vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
245 if (unlikely(!descs))
248 if (unlikely(dlen < vq->desc[idx].len)) {
250 * The indirect desc table is not contiguous
251 * in process VA space, we have to copy it.
253 idesc = alloc_copy_ind_table(dev, vq, &vq->desc[idx]);
254 if (unlikely(!idesc))
264 if (unlikely(idx >= vq->size)) {
265 free_ind_table(idesc);
270 len += descs[idx].len;
271 desc_avail = descs[idx].len;
272 desc_iova = descs[idx].addr;
276 uint64_t desc_chunck_len = desc_avail;
278 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
279 free_ind_table(idesc);
283 desc_addr = vhost_iova_to_vva(dev, vq,
287 if (unlikely(!desc_addr)) {
288 free_ind_table(idesc);
292 buf_vec[vec_id].buf_iova = desc_iova;
293 buf_vec[vec_id].buf_addr = desc_addr;
294 buf_vec[vec_id].buf_len = desc_chunck_len;
295 buf_vec[vec_id].desc_idx = idx;
297 desc_avail -= desc_chunck_len;
298 desc_iova += desc_chunck_len;
302 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
305 idx = descs[idx].next;
308 *desc_chain_len = len;
311 if (unlikely(!!idesc))
312 free_ind_table(idesc);
318 * Returns -1 on fail, 0 on success
321 reserve_avail_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
322 uint32_t size, struct buf_vector *buf_vec,
323 uint16_t *num_buffers, uint16_t avail_head,
327 uint32_t vec_idx = 0;
328 uint16_t max_tries, tries = 0;
330 uint16_t head_idx = 0;
334 cur_idx = vq->last_avail_idx;
336 if (rxvq_is_mergeable(dev))
337 max_tries = vq->size;
342 if (unlikely(cur_idx == avail_head))
345 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
347 VHOST_ACCESS_RW) < 0))
349 len = RTE_MIN(len, size);
350 update_shadow_used_ring(vq, head_idx, len);
358 * if we tried all available ring items, and still
359 * can't get enough buf, it means something abnormal
362 if (unlikely(tries > max_tries))
371 static __rte_always_inline int
372 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
373 struct rte_mbuf *m, struct buf_vector *buf_vec,
374 uint16_t nr_vec, uint16_t num_buffers)
376 uint32_t vec_idx = 0;
377 uint32_t mbuf_offset, mbuf_avail;
378 uint32_t buf_offset, buf_avail;
379 uint64_t buf_addr, buf_iova, buf_len;
382 struct rte_mbuf *hdr_mbuf;
383 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
384 struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
387 if (unlikely(m == NULL)) {
392 buf_addr = buf_vec[vec_idx].buf_addr;
393 buf_iova = buf_vec[vec_idx].buf_iova;
394 buf_len = buf_vec[vec_idx].buf_len;
396 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
403 if (unlikely(buf_len < dev->vhost_hlen))
406 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
407 rte_prefetch0((void *)(uintptr_t)hdr_addr);
409 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
410 dev->vid, num_buffers);
412 if (unlikely(buf_len < dev->vhost_hlen)) {
413 buf_offset = dev->vhost_hlen - buf_len;
415 buf_addr = buf_vec[vec_idx].buf_addr;
416 buf_iova = buf_vec[vec_idx].buf_iova;
417 buf_len = buf_vec[vec_idx].buf_len;
418 buf_avail = buf_len - buf_offset;
420 buf_offset = dev->vhost_hlen;
421 buf_avail = buf_len - dev->vhost_hlen;
424 mbuf_avail = rte_pktmbuf_data_len(m);
426 while (mbuf_avail != 0 || m->next != NULL) {
427 /* done with current buf, get the next one */
428 if (buf_avail == 0) {
430 if (unlikely(vec_idx >= nr_vec)) {
435 buf_addr = buf_vec[vec_idx].buf_addr;
436 buf_iova = buf_vec[vec_idx].buf_iova;
437 buf_len = buf_vec[vec_idx].buf_len;
439 /* Prefetch buffer address. */
440 rte_prefetch0((void *)(uintptr_t)buf_addr);
445 /* done with current mbuf, get the next one */
446 if (mbuf_avail == 0) {
450 mbuf_avail = rte_pktmbuf_data_len(m);
454 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
455 if (rxvq_is_mergeable(dev))
456 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
459 if (unlikely(hdr == &tmp_hdr)) {
461 uint64_t remain = dev->vhost_hlen;
462 uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
463 uint64_t iova = buf_vec[0].buf_iova;
464 uint16_t hdr_vec_idx = 0;
468 dst = buf_vec[hdr_vec_idx].buf_addr;
469 rte_memcpy((void *)(uintptr_t)dst,
470 (void *)(uintptr_t)src,
473 PRINT_PACKET(dev, (uintptr_t)dst,
475 vhost_log_cache_write(dev, vq,
484 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
486 vhost_log_cache_write(dev, vq,
494 cpy_len = RTE_MIN(buf_len, mbuf_avail);
496 if (likely(cpy_len > MAX_BATCH_LEN ||
497 vq->batch_copy_nb_elems >= vq->size)) {
498 rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
499 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
501 vhost_log_cache_write(dev, vq, buf_iova + buf_offset,
503 PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
506 batch_copy[vq->batch_copy_nb_elems].dst =
507 (void *)((uintptr_t)(buf_addr + buf_offset));
508 batch_copy[vq->batch_copy_nb_elems].src =
509 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
510 batch_copy[vq->batch_copy_nb_elems].log_addr =
511 buf_iova + buf_offset;
512 batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
513 vq->batch_copy_nb_elems++;
516 mbuf_avail -= cpy_len;
517 mbuf_offset += cpy_len;
518 buf_avail -= cpy_len;
519 buf_offset += cpy_len;
527 static __rte_always_inline uint32_t
528 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
529 struct rte_mbuf **pkts, uint32_t count)
531 struct vhost_virtqueue *vq;
532 uint32_t pkt_idx = 0;
533 uint16_t num_buffers;
534 struct buf_vector buf_vec[BUF_VECTOR_MAX];
537 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
538 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
539 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
540 dev->vid, __func__, queue_id);
544 vq = dev->virtqueue[queue_id];
546 rte_spinlock_lock(&vq->access_lock);
548 if (unlikely(vq->enabled == 0))
549 goto out_access_unlock;
551 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
552 vhost_user_iotlb_rd_lock(vq);
554 if (unlikely(vq->access_ok == 0))
555 if (unlikely(vring_translate(dev, vq) < 0))
558 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
562 vq->batch_copy_nb_elems = 0;
564 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
566 vq->shadow_used_idx = 0;
567 avail_head = *((volatile uint16_t *)&vq->avail->idx);
568 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
569 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
572 if (unlikely(reserve_avail_buf(dev, vq,
573 pkt_len, buf_vec, &num_buffers,
574 avail_head, &nr_vec) < 0)) {
575 VHOST_LOG_DEBUG(VHOST_DATA,
576 "(%d) failed to get enough desc from vring\n",
578 vq->shadow_used_idx -= num_buffers;
582 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
583 dev->vid, vq->last_avail_idx,
584 vq->last_avail_idx + num_buffers);
586 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
589 vq->shadow_used_idx -= num_buffers;
593 vq->last_avail_idx += num_buffers;
596 do_data_copy_enqueue(dev, vq);
598 if (likely(vq->shadow_used_idx)) {
599 flush_shadow_used_ring(dev, vq);
600 vhost_vring_call(dev, vq);
604 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
605 vhost_user_iotlb_rd_unlock(vq);
608 rte_spinlock_unlock(&vq->access_lock);
614 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
615 struct rte_mbuf **pkts, uint16_t count)
617 struct virtio_net *dev = get_device(vid);
622 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
623 RTE_LOG(ERR, VHOST_DATA,
624 "(%d) %s: built-in vhost net backend is disabled.\n",
629 return virtio_dev_rx(dev, queue_id, pkts, count);
633 virtio_net_with_host_offload(struct virtio_net *dev)
636 ((1ULL << VIRTIO_NET_F_CSUM) |
637 (1ULL << VIRTIO_NET_F_HOST_ECN) |
638 (1ULL << VIRTIO_NET_F_HOST_TSO4) |
639 (1ULL << VIRTIO_NET_F_HOST_TSO6) |
640 (1ULL << VIRTIO_NET_F_HOST_UFO)))
647 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
649 struct ipv4_hdr *ipv4_hdr;
650 struct ipv6_hdr *ipv6_hdr;
652 struct ether_hdr *eth_hdr;
655 eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
657 m->l2_len = sizeof(struct ether_hdr);
658 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
660 if (ethertype == ETHER_TYPE_VLAN) {
661 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
663 m->l2_len += sizeof(struct vlan_hdr);
664 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
667 l3_hdr = (char *)eth_hdr + m->l2_len;
670 case ETHER_TYPE_IPv4:
672 *l4_proto = ipv4_hdr->next_proto_id;
673 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
674 *l4_hdr = (char *)l3_hdr + m->l3_len;
675 m->ol_flags |= PKT_TX_IPV4;
677 case ETHER_TYPE_IPv6:
679 *l4_proto = ipv6_hdr->proto;
680 m->l3_len = sizeof(struct ipv6_hdr);
681 *l4_hdr = (char *)l3_hdr + m->l3_len;
682 m->ol_flags |= PKT_TX_IPV6;
692 static __rte_always_inline void
693 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
695 uint16_t l4_proto = 0;
697 struct tcp_hdr *tcp_hdr = NULL;
699 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
702 parse_ethernet(m, &l4_proto, &l4_hdr);
703 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
704 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
705 switch (hdr->csum_offset) {
706 case (offsetof(struct tcp_hdr, cksum)):
707 if (l4_proto == IPPROTO_TCP)
708 m->ol_flags |= PKT_TX_TCP_CKSUM;
710 case (offsetof(struct udp_hdr, dgram_cksum)):
711 if (l4_proto == IPPROTO_UDP)
712 m->ol_flags |= PKT_TX_UDP_CKSUM;
714 case (offsetof(struct sctp_hdr, cksum)):
715 if (l4_proto == IPPROTO_SCTP)
716 m->ol_flags |= PKT_TX_SCTP_CKSUM;
724 if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
725 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
726 case VIRTIO_NET_HDR_GSO_TCPV4:
727 case VIRTIO_NET_HDR_GSO_TCPV6:
729 m->ol_flags |= PKT_TX_TCP_SEG;
730 m->tso_segsz = hdr->gso_size;
731 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
733 case VIRTIO_NET_HDR_GSO_UDP:
734 m->ol_flags |= PKT_TX_UDP_SEG;
735 m->tso_segsz = hdr->gso_size;
736 m->l4_len = sizeof(struct udp_hdr);
739 RTE_LOG(WARNING, VHOST_DATA,
740 "unsupported gso type %u.\n", hdr->gso_type);
746 static __rte_always_inline void
747 put_zmbuf(struct zcopy_mbuf *zmbuf)
752 static __rte_always_inline int
753 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
754 struct buf_vector *buf_vec, uint16_t nr_vec,
755 struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
757 uint32_t buf_avail, buf_offset;
758 uint64_t buf_addr, buf_iova, buf_len;
759 uint32_t mbuf_avail, mbuf_offset;
761 struct rte_mbuf *cur = m, *prev = m;
762 struct virtio_net_hdr tmp_hdr;
763 struct virtio_net_hdr *hdr = NULL;
764 /* A counter to avoid desc dead loop chain */
765 uint16_t vec_idx = 0;
766 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
769 buf_addr = buf_vec[vec_idx].buf_addr;
770 buf_iova = buf_vec[vec_idx].buf_iova;
771 buf_len = buf_vec[vec_idx].buf_len;
773 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
778 if (likely(nr_vec > 1))
779 rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
781 if (virtio_net_with_host_offload(dev)) {
782 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
784 uint64_t remain = sizeof(struct virtio_net_hdr);
786 uint64_t dst = (uint64_t)(uintptr_t)&tmp_hdr;
787 uint16_t hdr_vec_idx = 0;
790 * No luck, the virtio-net header doesn't fit
791 * in a contiguous virtual area.
795 src = buf_vec[hdr_vec_idx].buf_addr;
796 rte_memcpy((void *)(uintptr_t)dst,
797 (void *)(uintptr_t)src, len);
806 hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
812 * A virtio driver normally uses at least 2 desc buffers
813 * for Tx: the first for storing the header, and others
814 * for storing the data.
816 if (unlikely(buf_len < dev->vhost_hlen)) {
817 buf_offset = dev->vhost_hlen - buf_len;
819 buf_addr = buf_vec[vec_idx].buf_addr;
820 buf_iova = buf_vec[vec_idx].buf_iova;
821 buf_len = buf_vec[vec_idx].buf_len;
822 buf_avail = buf_len - buf_offset;
823 } else if (buf_len == dev->vhost_hlen) {
824 if (unlikely(++vec_idx >= nr_vec))
826 buf_addr = buf_vec[vec_idx].buf_addr;
827 buf_iova = buf_vec[vec_idx].buf_iova;
828 buf_len = buf_vec[vec_idx].buf_len;
833 buf_offset = dev->vhost_hlen;
834 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
837 rte_prefetch0((void *)(uintptr_t)
838 (buf_addr + buf_offset));
841 (uintptr_t)(buf_addr + buf_offset),
842 (uint32_t)buf_avail, 0);
845 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
849 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
852 * A desc buf might across two host physical pages that are
853 * not continuous. In such case (gpa_to_hpa returns 0), data
854 * will be copied even though zero copy is enabled.
856 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
857 buf_iova + buf_offset, cpy_len)))) {
858 cur->data_len = cpy_len;
861 (void *)(uintptr_t)(buf_addr + buf_offset);
865 * In zero copy mode, one mbuf can only reference data
866 * for one or partial of one desc buff.
868 mbuf_avail = cpy_len;
870 if (likely(cpy_len > MAX_BATCH_LEN ||
871 vq->batch_copy_nb_elems >= vq->size ||
872 (hdr && cur == m))) {
873 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
875 (void *)((uintptr_t)(buf_addr +
879 batch_copy[vq->batch_copy_nb_elems].dst =
880 rte_pktmbuf_mtod_offset(cur, void *,
882 batch_copy[vq->batch_copy_nb_elems].src =
883 (void *)((uintptr_t)(buf_addr +
885 batch_copy[vq->batch_copy_nb_elems].len =
887 vq->batch_copy_nb_elems++;
891 mbuf_avail -= cpy_len;
892 mbuf_offset += cpy_len;
893 buf_avail -= cpy_len;
894 buf_offset += cpy_len;
896 /* This buf reaches to its end, get the next one */
897 if (buf_avail == 0) {
898 if (++vec_idx >= nr_vec)
901 buf_addr = buf_vec[vec_idx].buf_addr;
902 buf_iova = buf_vec[vec_idx].buf_iova;
903 buf_len = buf_vec[vec_idx].buf_len;
906 * Prefecth desc n + 1 buffer while
907 * desc n buffer is processed.
909 if (vec_idx + 1 < nr_vec)
910 rte_prefetch0((void *)(uintptr_t)
911 buf_vec[vec_idx + 1].buf_addr);
916 PRINT_PACKET(dev, (uintptr_t)buf_addr,
917 (uint32_t)buf_avail, 0);
921 * This mbuf reaches to its end, get a new one
924 if (mbuf_avail == 0) {
925 cur = rte_pktmbuf_alloc(mbuf_pool);
926 if (unlikely(cur == NULL)) {
927 RTE_LOG(ERR, VHOST_DATA, "Failed to "
928 "allocate memory for mbuf.\n");
932 if (unlikely(dev->dequeue_zero_copy))
933 rte_mbuf_refcnt_update(cur, 1);
936 prev->data_len = mbuf_offset;
938 m->pkt_len += mbuf_offset;
942 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
946 prev->data_len = mbuf_offset;
947 m->pkt_len += mbuf_offset;
950 vhost_dequeue_offload(hdr, m);
957 static __rte_always_inline struct zcopy_mbuf *
958 get_zmbuf(struct vhost_virtqueue *vq)
964 /* search [last_zmbuf_idx, zmbuf_size) */
965 i = vq->last_zmbuf_idx;
966 last = vq->zmbuf_size;
969 for (; i < last; i++) {
970 if (vq->zmbufs[i].in_use == 0) {
971 vq->last_zmbuf_idx = i + 1;
972 vq->zmbufs[i].in_use = 1;
973 return &vq->zmbufs[i];
979 /* search [0, last_zmbuf_idx) */
981 last = vq->last_zmbuf_idx;
988 static __rte_always_inline bool
989 mbuf_is_consumed(struct rte_mbuf *m)
992 if (rte_mbuf_refcnt_read(m) > 1)
1000 static __rte_always_inline void
1001 restore_mbuf(struct rte_mbuf *m)
1003 uint32_t mbuf_size, priv_size;
1006 priv_size = rte_pktmbuf_priv_size(m->pool);
1007 mbuf_size = sizeof(struct rte_mbuf) + priv_size;
1008 /* start of buffer is after mbuf structure and priv data */
1010 m->buf_addr = (char *)m + mbuf_size;
1011 m->buf_iova = rte_mempool_virt2iova(m) + mbuf_size;
1017 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1018 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1020 struct virtio_net *dev;
1021 struct rte_mbuf *rarp_mbuf = NULL;
1022 struct vhost_virtqueue *vq;
1024 uint16_t free_entries;
1026 dev = get_device(vid);
1030 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1031 RTE_LOG(ERR, VHOST_DATA,
1032 "(%d) %s: built-in vhost net backend is disabled.\n",
1033 dev->vid, __func__);
1037 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1038 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1039 dev->vid, __func__, queue_id);
1043 vq = dev->virtqueue[queue_id];
1045 if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
1048 if (unlikely(vq->enabled == 0))
1049 goto out_access_unlock;
1051 vq->batch_copy_nb_elems = 0;
1052 vq->shadow_used_idx = 0;
1054 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1055 vhost_user_iotlb_rd_lock(vq);
1057 if (unlikely(vq->access_ok == 0))
1058 if (unlikely(vring_translate(dev, vq) < 0))
1061 if (unlikely(dev->dequeue_zero_copy)) {
1062 struct zcopy_mbuf *zmbuf, *next;
1065 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1066 zmbuf != NULL; zmbuf = next) {
1067 next = TAILQ_NEXT(zmbuf, next);
1069 if (mbuf_is_consumed(zmbuf->mbuf)) {
1070 update_shadow_used_ring(vq, zmbuf->desc_idx, 0);
1073 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1074 restore_mbuf(zmbuf->mbuf);
1075 rte_pktmbuf_free(zmbuf->mbuf);
1081 flush_shadow_used_ring(dev, vq);
1082 vhost_vring_call(dev, vq);
1083 vq->shadow_used_idx = 0;
1086 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1089 * Construct a RARP broadcast packet, and inject it to the "pkts"
1090 * array, to looks like that guest actually send such packet.
1092 * Check user_send_rarp() for more information.
1094 * broadcast_rarp shares a cacheline in the virtio_net structure
1095 * with some fields that are accessed during enqueue and
1096 * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1097 * result in false sharing between enqueue and dequeue.
1099 * Prevent unnecessary false sharing by reading broadcast_rarp first
1100 * and only performing cmpset if the read indicates it is likely to
1104 if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1105 rte_atomic16_cmpset((volatile uint16_t *)
1106 &dev->broadcast_rarp.cnt, 1, 0))) {
1108 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
1109 if (rarp_mbuf == NULL) {
1110 RTE_LOG(ERR, VHOST_DATA,
1111 "Failed to make RARP packet.\n");
1117 free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1119 if (free_entries == 0)
1122 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1124 count = RTE_MIN(count, MAX_PKT_BURST);
1125 count = RTE_MIN(count, free_entries);
1126 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1129 for (i = 0; i < count; i++) {
1130 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1131 uint16_t head_idx, dummy_len;
1132 uint32_t nr_vec = 0;
1135 if (unlikely(fill_vec_buf(dev, vq,
1136 vq->last_avail_idx + i,
1138 &head_idx, &dummy_len,
1139 VHOST_ACCESS_RO) < 0))
1142 if (likely(dev->dequeue_zero_copy == 0))
1143 update_shadow_used_ring(vq, head_idx, 0);
1145 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
1147 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1148 if (unlikely(pkts[i] == NULL)) {
1149 RTE_LOG(ERR, VHOST_DATA,
1150 "Failed to allocate memory for mbuf.\n");
1154 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1156 if (unlikely(err)) {
1157 rte_pktmbuf_free(pkts[i]);
1161 if (unlikely(dev->dequeue_zero_copy)) {
1162 struct zcopy_mbuf *zmbuf;
1164 zmbuf = get_zmbuf(vq);
1166 rte_pktmbuf_free(pkts[i]);
1169 zmbuf->mbuf = pkts[i];
1170 zmbuf->desc_idx = head_idx;
1173 * Pin lock the mbuf; we will check later to see
1174 * whether the mbuf is freed (when we are the last
1175 * user) or not. If that's the case, we then could
1176 * update the used ring safely.
1178 rte_mbuf_refcnt_update(pkts[i], 1);
1181 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1184 vq->last_avail_idx += i;
1186 if (likely(dev->dequeue_zero_copy == 0)) {
1187 do_data_copy_dequeue(vq);
1188 if (unlikely(i < count))
1189 vq->shadow_used_idx = i;
1190 flush_shadow_used_ring(dev, vq);
1191 vhost_vring_call(dev, vq);
1195 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1196 vhost_user_iotlb_rd_unlock(vq);
1199 rte_spinlock_unlock(&vq->access_lock);
1201 if (unlikely(rarp_mbuf != NULL)) {
1203 * Inject it to the head of "pkts" array, so that switch's mac
1204 * learning table will get updated first.
1206 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1207 pkts[0] = rarp_mbuf;