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)
230 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
231 uint32_t vec_id = *vec_idx;
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(vec_id >= BUF_VECTOR_MAX || idx >= vq->size)) {
265 free_ind_table(idesc);
269 len += descs[idx].len;
270 buf_vec[vec_id].buf_addr = descs[idx].addr;
271 buf_vec[vec_id].buf_len = descs[idx].len;
272 buf_vec[vec_id].desc_idx = idx;
275 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
278 idx = descs[idx].next;
281 *desc_chain_len = len;
284 if (unlikely(!!idesc))
285 free_ind_table(idesc);
291 * Returns -1 on fail, 0 on success
294 reserve_avail_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
295 uint32_t size, struct buf_vector *buf_vec,
296 uint16_t *num_buffers, uint16_t avail_head)
299 uint32_t vec_idx = 0;
300 uint16_t max_tries, tries = 0;
302 uint16_t head_idx = 0;
306 cur_idx = vq->last_avail_idx;
308 if (rxvq_is_mergeable(dev))
309 max_tries = vq->size;
314 if (unlikely(cur_idx == avail_head))
317 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
318 &head_idx, &len) < 0))
320 len = RTE_MIN(len, size);
321 update_shadow_used_ring(vq, head_idx, len);
329 * if we tried all available ring items, and still
330 * can't get enough buf, it means something abnormal
333 if (unlikely(tries > max_tries))
340 static __rte_always_inline int
341 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
342 struct rte_mbuf *m, struct buf_vector *buf_vec,
343 uint16_t num_buffers)
345 uint32_t vec_idx = 0;
346 uint64_t desc_addr, desc_gaddr;
347 uint32_t mbuf_offset, mbuf_avail;
348 uint32_t desc_offset, desc_avail;
350 uint64_t desc_chunck_len;
351 uint64_t hdr_addr, hdr_phys_addr;
352 struct rte_mbuf *hdr_mbuf;
353 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
354 struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
355 uint16_t copy_nb = vq->batch_copy_nb_elems;
358 if (unlikely(m == NULL)) {
363 desc_chunck_len = buf_vec[vec_idx].buf_len;
364 desc_gaddr = buf_vec[vec_idx].buf_addr;
365 desc_addr = vhost_iova_to_vva(dev, vq,
369 if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr) {
375 hdr_addr = desc_addr;
376 if (unlikely(desc_chunck_len < dev->vhost_hlen))
379 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
380 hdr_phys_addr = desc_gaddr;
381 rte_prefetch0((void *)(uintptr_t)hdr_addr);
383 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
384 dev->vid, num_buffers);
386 desc_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
387 if (unlikely(desc_chunck_len < dev->vhost_hlen)) {
388 desc_chunck_len = desc_avail;
389 desc_gaddr += dev->vhost_hlen;
390 desc_addr = vhost_iova_to_vva(dev, vq,
394 if (unlikely(!desc_addr)) {
401 desc_offset = dev->vhost_hlen;
402 desc_chunck_len -= dev->vhost_hlen;
406 mbuf_avail = rte_pktmbuf_data_len(m);
408 while (mbuf_avail != 0 || m->next != NULL) {
409 /* done with current desc buf, get the next one */
410 if (desc_avail == 0) {
412 desc_chunck_len = buf_vec[vec_idx].buf_len;
413 desc_gaddr = buf_vec[vec_idx].buf_addr;
415 vhost_iova_to_vva(dev, vq,
419 if (unlikely(!desc_addr)) {
424 /* Prefetch buffer address. */
425 rte_prefetch0((void *)(uintptr_t)desc_addr);
427 desc_avail = buf_vec[vec_idx].buf_len;
428 } else if (unlikely(desc_chunck_len == 0)) {
429 desc_chunck_len = desc_avail;
430 desc_gaddr += desc_offset;
431 desc_addr = vhost_iova_to_vva(dev, vq,
433 &desc_chunck_len, VHOST_ACCESS_RW);
434 if (unlikely(!desc_addr)) {
441 /* done with current mbuf, get the next one */
442 if (mbuf_avail == 0) {
446 mbuf_avail = rte_pktmbuf_data_len(m);
450 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
451 if (rxvq_is_mergeable(dev))
452 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
455 if (unlikely(hdr == &tmp_hdr)) {
457 uint64_t remain = dev->vhost_hlen;
458 uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
459 uint64_t guest_addr = hdr_phys_addr;
463 dst = vhost_iova_to_vva(dev, vq,
466 if (unlikely(!dst || !len)) {
471 rte_memcpy((void *)(uintptr_t)dst,
472 (void *)(uintptr_t)src,
475 PRINT_PACKET(dev, (uintptr_t)dst,
477 vhost_log_cache_write(dev, vq,
485 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
487 vhost_log_cache_write(dev, vq, hdr_phys_addr,
494 cpy_len = RTE_MIN(desc_chunck_len, mbuf_avail);
496 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
497 rte_memcpy((void *)((uintptr_t)(desc_addr +
499 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
501 vhost_log_cache_write(dev, vq, desc_gaddr + desc_offset,
503 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
506 batch_copy[copy_nb].dst =
507 (void *)((uintptr_t)(desc_addr + desc_offset));
508 batch_copy[copy_nb].src =
509 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
510 batch_copy[copy_nb].log_addr = desc_gaddr + desc_offset;
511 batch_copy[copy_nb].len = cpy_len;
515 mbuf_avail -= cpy_len;
516 mbuf_offset += cpy_len;
517 desc_avail -= cpy_len;
518 desc_offset += cpy_len;
519 desc_chunck_len -= cpy_len;
523 vq->batch_copy_nb_elems = copy_nb;
528 static __rte_always_inline uint32_t
529 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
530 struct rte_mbuf **pkts, uint32_t count)
532 struct vhost_virtqueue *vq;
533 uint32_t pkt_idx = 0;
534 uint16_t num_buffers;
535 struct buf_vector buf_vec[BUF_VECTOR_MAX];
538 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
539 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
540 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
541 dev->vid, __func__, queue_id);
545 vq = dev->virtqueue[queue_id];
547 rte_spinlock_lock(&vq->access_lock);
549 if (unlikely(vq->enabled == 0))
550 goto out_access_unlock;
552 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
553 vhost_user_iotlb_rd_lock(vq);
555 if (unlikely(vq->access_ok == 0))
556 if (unlikely(vring_translate(dev, vq) < 0))
559 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
563 vq->batch_copy_nb_elems = 0;
565 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
567 vq->shadow_used_idx = 0;
568 avail_head = *((volatile uint16_t *)&vq->avail->idx);
569 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
570 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,
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],
587 buf_vec, num_buffers) < 0) {
588 vq->shadow_used_idx -= num_buffers;
592 vq->last_avail_idx += num_buffers;
595 do_data_copy_enqueue(dev, vq);
597 if (likely(vq->shadow_used_idx)) {
598 flush_shadow_used_ring(dev, vq);
599 vhost_vring_call(dev, vq);
603 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
604 vhost_user_iotlb_rd_unlock(vq);
607 rte_spinlock_unlock(&vq->access_lock);
613 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
614 struct rte_mbuf **pkts, uint16_t count)
616 struct virtio_net *dev = get_device(vid);
621 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
622 RTE_LOG(ERR, VHOST_DATA,
623 "(%d) %s: built-in vhost net backend is disabled.\n",
628 return virtio_dev_rx(dev, queue_id, pkts, count);
632 virtio_net_with_host_offload(struct virtio_net *dev)
635 ((1ULL << VIRTIO_NET_F_CSUM) |
636 (1ULL << VIRTIO_NET_F_HOST_ECN) |
637 (1ULL << VIRTIO_NET_F_HOST_TSO4) |
638 (1ULL << VIRTIO_NET_F_HOST_TSO6) |
639 (1ULL << VIRTIO_NET_F_HOST_UFO)))
646 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
648 struct ipv4_hdr *ipv4_hdr;
649 struct ipv6_hdr *ipv6_hdr;
651 struct ether_hdr *eth_hdr;
654 eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
656 m->l2_len = sizeof(struct ether_hdr);
657 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
659 if (ethertype == ETHER_TYPE_VLAN) {
660 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
662 m->l2_len += sizeof(struct vlan_hdr);
663 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
666 l3_hdr = (char *)eth_hdr + m->l2_len;
669 case ETHER_TYPE_IPv4:
671 *l4_proto = ipv4_hdr->next_proto_id;
672 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
673 *l4_hdr = (char *)l3_hdr + m->l3_len;
674 m->ol_flags |= PKT_TX_IPV4;
676 case ETHER_TYPE_IPv6:
678 *l4_proto = ipv6_hdr->proto;
679 m->l3_len = sizeof(struct ipv6_hdr);
680 *l4_hdr = (char *)l3_hdr + m->l3_len;
681 m->ol_flags |= PKT_TX_IPV6;
691 static __rte_always_inline void
692 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
694 uint16_t l4_proto = 0;
696 struct tcp_hdr *tcp_hdr = NULL;
698 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
701 parse_ethernet(m, &l4_proto, &l4_hdr);
702 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
703 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
704 switch (hdr->csum_offset) {
705 case (offsetof(struct tcp_hdr, cksum)):
706 if (l4_proto == IPPROTO_TCP)
707 m->ol_flags |= PKT_TX_TCP_CKSUM;
709 case (offsetof(struct udp_hdr, dgram_cksum)):
710 if (l4_proto == IPPROTO_UDP)
711 m->ol_flags |= PKT_TX_UDP_CKSUM;
713 case (offsetof(struct sctp_hdr, cksum)):
714 if (l4_proto == IPPROTO_SCTP)
715 m->ol_flags |= PKT_TX_SCTP_CKSUM;
723 if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
724 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
725 case VIRTIO_NET_HDR_GSO_TCPV4:
726 case VIRTIO_NET_HDR_GSO_TCPV6:
728 m->ol_flags |= PKT_TX_TCP_SEG;
729 m->tso_segsz = hdr->gso_size;
730 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
732 case VIRTIO_NET_HDR_GSO_UDP:
733 m->ol_flags |= PKT_TX_UDP_SEG;
734 m->tso_segsz = hdr->gso_size;
735 m->l4_len = sizeof(struct udp_hdr);
738 RTE_LOG(WARNING, VHOST_DATA,
739 "unsupported gso type %u.\n", hdr->gso_type);
745 static __rte_always_inline void
746 put_zmbuf(struct zcopy_mbuf *zmbuf)
751 static __rte_always_inline int
752 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
753 struct vring_desc *descs, uint16_t max_desc,
754 struct rte_mbuf *m, uint16_t desc_idx,
755 struct rte_mempool *mbuf_pool)
757 struct vring_desc *desc;
758 uint64_t desc_addr, desc_gaddr;
759 uint32_t desc_avail, desc_offset;
760 uint32_t mbuf_avail, mbuf_offset;
762 uint64_t desc_chunck_len;
763 struct rte_mbuf *cur = m, *prev = m;
764 struct virtio_net_hdr tmp_hdr;
765 struct virtio_net_hdr *hdr = NULL;
766 /* A counter to avoid desc dead loop chain */
767 uint32_t nr_desc = 1;
768 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
769 uint16_t copy_nb = vq->batch_copy_nb_elems;
772 desc = &descs[desc_idx];
773 if (unlikely((desc->len < dev->vhost_hlen)) ||
774 (desc->flags & VRING_DESC_F_INDIRECT)) {
779 desc_chunck_len = desc->len;
780 desc_gaddr = desc->addr;
781 desc_addr = vhost_iova_to_vva(dev,
785 if (unlikely(!desc_addr)) {
790 if (virtio_net_with_host_offload(dev)) {
791 if (unlikely(desc_chunck_len < sizeof(struct virtio_net_hdr))) {
792 uint64_t len = desc_chunck_len;
793 uint64_t remain = sizeof(struct virtio_net_hdr);
794 uint64_t src = desc_addr;
795 uint64_t dst = (uint64_t)(uintptr_t)&tmp_hdr;
796 uint64_t guest_addr = desc_gaddr;
799 * No luck, the virtio-net header doesn't fit
800 * in a contiguous virtual area.
804 src = vhost_iova_to_vva(dev, vq,
807 if (unlikely(!src || !len)) {
812 rte_memcpy((void *)(uintptr_t)dst,
813 (void *)(uintptr_t)src, len);
822 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
828 * A virtio driver normally uses at least 2 desc buffers
829 * for Tx: the first for storing the header, and others
830 * for storing the data.
832 if (likely((desc->len == dev->vhost_hlen) &&
833 (desc->flags & VRING_DESC_F_NEXT) != 0)) {
834 desc = &descs[desc->next];
835 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
840 desc_chunck_len = desc->len;
841 desc_gaddr = desc->addr;
842 desc_addr = vhost_iova_to_vva(dev,
846 if (unlikely(!desc_addr)) {
852 desc_avail = desc->len;
855 desc_avail = desc->len - dev->vhost_hlen;
857 if (unlikely(desc_chunck_len < dev->vhost_hlen)) {
858 desc_chunck_len = desc_avail;
859 desc_gaddr += dev->vhost_hlen;
860 desc_addr = vhost_iova_to_vva(dev,
864 if (unlikely(!desc_addr)) {
871 desc_offset = dev->vhost_hlen;
872 desc_chunck_len -= dev->vhost_hlen;
876 rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
878 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
879 (uint32_t)desc_chunck_len, 0);
882 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
886 cpy_len = RTE_MIN(desc_chunck_len, mbuf_avail);
889 * A desc buf might across two host physical pages that are
890 * not continuous. In such case (gpa_to_hpa returns 0), data
891 * will be copied even though zero copy is enabled.
893 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
894 desc_gaddr + desc_offset, cpy_len)))) {
895 cur->data_len = cpy_len;
897 cur->buf_addr = (void *)(uintptr_t)(desc_addr
902 * In zero copy mode, one mbuf can only reference data
903 * for one or partial of one desc buff.
905 mbuf_avail = cpy_len;
907 if (likely(cpy_len > MAX_BATCH_LEN ||
908 copy_nb >= vq->size ||
910 desc->len != desc_chunck_len)) {
911 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
913 (void *)((uintptr_t)(desc_addr +
917 batch_copy[copy_nb].dst =
918 rte_pktmbuf_mtod_offset(cur, void *,
920 batch_copy[copy_nb].src =
921 (void *)((uintptr_t)(desc_addr +
923 batch_copy[copy_nb].len = cpy_len;
928 mbuf_avail -= cpy_len;
929 mbuf_offset += cpy_len;
930 desc_avail -= cpy_len;
931 desc_chunck_len -= cpy_len;
932 desc_offset += cpy_len;
934 /* This desc reaches to its end, get the next one */
935 if (desc_avail == 0) {
936 if ((desc->flags & VRING_DESC_F_NEXT) == 0)
939 if (unlikely(desc->next >= max_desc ||
940 ++nr_desc > max_desc)) {
944 desc = &descs[desc->next];
945 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
950 desc_chunck_len = desc->len;
951 desc_gaddr = desc->addr;
952 desc_addr = vhost_iova_to_vva(dev,
956 if (unlikely(!desc_addr)) {
961 rte_prefetch0((void *)(uintptr_t)desc_addr);
964 desc_avail = desc->len;
966 PRINT_PACKET(dev, (uintptr_t)desc_addr,
967 (uint32_t)desc_chunck_len, 0);
968 } else if (unlikely(desc_chunck_len == 0)) {
969 desc_chunck_len = desc_avail;
970 desc_gaddr += desc_offset;
971 desc_addr = vhost_iova_to_vva(dev, vq,
975 if (unlikely(!desc_addr)) {
981 PRINT_PACKET(dev, (uintptr_t)desc_addr,
982 (uint32_t)desc_chunck_len, 0);
986 * This mbuf reaches to its end, get a new one
989 if (mbuf_avail == 0) {
990 cur = rte_pktmbuf_alloc(mbuf_pool);
991 if (unlikely(cur == NULL)) {
992 RTE_LOG(ERR, VHOST_DATA, "Failed to "
993 "allocate memory for mbuf.\n");
997 if (unlikely(dev->dequeue_zero_copy))
998 rte_mbuf_refcnt_update(cur, 1);
1001 prev->data_len = mbuf_offset;
1003 m->pkt_len += mbuf_offset;
1007 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1011 prev->data_len = mbuf_offset;
1012 m->pkt_len += mbuf_offset;
1015 vhost_dequeue_offload(hdr, m);
1018 vq->batch_copy_nb_elems = copy_nb;
1023 static __rte_always_inline void
1024 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
1025 uint32_t used_idx, uint32_t desc_idx)
1027 vq->used->ring[used_idx].id = desc_idx;
1028 vq->used->ring[used_idx].len = 0;
1029 vhost_log_cache_used_vring(dev, vq,
1030 offsetof(struct vring_used, ring[used_idx]),
1031 sizeof(vq->used->ring[used_idx]));
1034 static __rte_always_inline void
1035 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
1038 if (unlikely(count == 0))
1044 vhost_log_cache_sync(dev, vq);
1046 vq->used->idx += count;
1047 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
1048 sizeof(vq->used->idx));
1049 vhost_vring_call(dev, vq);
1052 static __rte_always_inline struct zcopy_mbuf *
1053 get_zmbuf(struct vhost_virtqueue *vq)
1059 /* search [last_zmbuf_idx, zmbuf_size) */
1060 i = vq->last_zmbuf_idx;
1061 last = vq->zmbuf_size;
1064 for (; i < last; i++) {
1065 if (vq->zmbufs[i].in_use == 0) {
1066 vq->last_zmbuf_idx = i + 1;
1067 vq->zmbufs[i].in_use = 1;
1068 return &vq->zmbufs[i];
1074 /* search [0, last_zmbuf_idx) */
1076 last = vq->last_zmbuf_idx;
1083 static __rte_always_inline bool
1084 mbuf_is_consumed(struct rte_mbuf *m)
1087 if (rte_mbuf_refcnt_read(m) > 1)
1095 static __rte_always_inline void
1096 restore_mbuf(struct rte_mbuf *m)
1098 uint32_t mbuf_size, priv_size;
1101 priv_size = rte_pktmbuf_priv_size(m->pool);
1102 mbuf_size = sizeof(struct rte_mbuf) + priv_size;
1103 /* start of buffer is after mbuf structure and priv data */
1105 m->buf_addr = (char *)m + mbuf_size;
1106 m->buf_iova = rte_mempool_virt2iova(m) + mbuf_size;
1112 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1113 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1115 struct virtio_net *dev;
1116 struct rte_mbuf *rarp_mbuf = NULL;
1117 struct vhost_virtqueue *vq;
1118 uint32_t desc_indexes[MAX_PKT_BURST];
1121 uint16_t free_entries;
1124 dev = get_device(vid);
1128 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1129 RTE_LOG(ERR, VHOST_DATA,
1130 "(%d) %s: built-in vhost net backend is disabled.\n",
1131 dev->vid, __func__);
1135 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1136 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1137 dev->vid, __func__, queue_id);
1141 vq = dev->virtqueue[queue_id];
1143 if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
1146 if (unlikely(vq->enabled == 0))
1147 goto out_access_unlock;
1149 vq->batch_copy_nb_elems = 0;
1151 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1152 vhost_user_iotlb_rd_lock(vq);
1154 if (unlikely(vq->access_ok == 0))
1155 if (unlikely(vring_translate(dev, vq) < 0))
1158 if (unlikely(dev->dequeue_zero_copy)) {
1159 struct zcopy_mbuf *zmbuf, *next;
1162 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1163 zmbuf != NULL; zmbuf = next) {
1164 next = TAILQ_NEXT(zmbuf, next);
1166 if (mbuf_is_consumed(zmbuf->mbuf)) {
1167 used_idx = vq->last_used_idx++ & (vq->size - 1);
1168 update_used_ring(dev, vq, used_idx,
1172 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1173 restore_mbuf(zmbuf->mbuf);
1174 rte_pktmbuf_free(zmbuf->mbuf);
1180 update_used_idx(dev, vq, nr_updated);
1184 * Construct a RARP broadcast packet, and inject it to the "pkts"
1185 * array, to looks like that guest actually send such packet.
1187 * Check user_send_rarp() for more information.
1189 * broadcast_rarp shares a cacheline in the virtio_net structure
1190 * with some fields that are accessed during enqueue and
1191 * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1192 * result in false sharing between enqueue and dequeue.
1194 * Prevent unnecessary false sharing by reading broadcast_rarp first
1195 * and only performing cmpset if the read indicates it is likely to
1199 if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1200 rte_atomic16_cmpset((volatile uint16_t *)
1201 &dev->broadcast_rarp.cnt, 1, 0))) {
1203 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
1204 if (rarp_mbuf == NULL) {
1205 RTE_LOG(ERR, VHOST_DATA,
1206 "Failed to make RARP packet.\n");
1212 free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1214 if (free_entries == 0)
1217 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1219 /* Prefetch available and used ring */
1220 avail_idx = vq->last_avail_idx & (vq->size - 1);
1221 used_idx = vq->last_used_idx & (vq->size - 1);
1222 rte_prefetch0(&vq->avail->ring[avail_idx]);
1223 rte_prefetch0(&vq->used->ring[used_idx]);
1225 count = RTE_MIN(count, MAX_PKT_BURST);
1226 count = RTE_MIN(count, free_entries);
1227 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1230 /* Retrieve all of the head indexes first to avoid caching issues. */
1231 for (i = 0; i < count; i++) {
1232 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1233 used_idx = (vq->last_used_idx + i) & (vq->size - 1);
1234 desc_indexes[i] = vq->avail->ring[avail_idx];
1236 if (likely(dev->dequeue_zero_copy == 0))
1237 update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1240 /* Prefetch descriptor index. */
1241 rte_prefetch0(&vq->desc[desc_indexes[0]]);
1242 for (i = 0; i < count; i++) {
1243 struct vring_desc *desc, *idesc = NULL;
1248 if (likely(i + 1 < count))
1249 rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1251 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1252 dlen = vq->desc[desc_indexes[i]].len;
1253 desc = (struct vring_desc *)(uintptr_t)
1254 vhost_iova_to_vva(dev, vq,
1255 vq->desc[desc_indexes[i]].addr,
1258 if (unlikely(!desc))
1261 if (unlikely(dlen < vq->desc[desc_indexes[i]].len)) {
1263 * The indirect desc table is not contiguous
1264 * in process VA space, we have to copy it.
1266 idesc = alloc_copy_ind_table(dev, vq,
1267 &vq->desc[desc_indexes[i]]);
1268 if (unlikely(!idesc))
1274 rte_prefetch0(desc);
1275 sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1280 idx = desc_indexes[i];
1283 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1284 if (unlikely(pkts[i] == NULL)) {
1285 RTE_LOG(ERR, VHOST_DATA,
1286 "Failed to allocate memory for mbuf.\n");
1287 free_ind_table(idesc);
1291 err = copy_desc_to_mbuf(dev, vq, desc, sz, pkts[i], idx,
1293 if (unlikely(err)) {
1294 rte_pktmbuf_free(pkts[i]);
1295 free_ind_table(idesc);
1299 if (unlikely(dev->dequeue_zero_copy)) {
1300 struct zcopy_mbuf *zmbuf;
1302 zmbuf = get_zmbuf(vq);
1304 rte_pktmbuf_free(pkts[i]);
1305 free_ind_table(idesc);
1308 zmbuf->mbuf = pkts[i];
1309 zmbuf->desc_idx = desc_indexes[i];
1312 * Pin lock the mbuf; we will check later to see
1313 * whether the mbuf is freed (when we are the last
1314 * user) or not. If that's the case, we then could
1315 * update the used ring safely.
1317 rte_mbuf_refcnt_update(pkts[i], 1);
1320 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1323 if (unlikely(!!idesc))
1324 free_ind_table(idesc);
1326 vq->last_avail_idx += i;
1328 if (likely(dev->dequeue_zero_copy == 0)) {
1329 do_data_copy_dequeue(vq);
1330 vq->last_used_idx += i;
1331 update_used_idx(dev, vq, i);
1335 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1336 vhost_user_iotlb_rd_unlock(vq);
1339 rte_spinlock_unlock(&vq->access_lock);
1341 if (unlikely(rarp_mbuf != NULL)) {
1343 * Inject it to the head of "pkts" array, so that switch's mac
1344 * learning table will get updated first.
1346 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1347 pkts[0] = rarp_mbuf;