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 void *
41 alloc_copy_ind_table(struct virtio_net *dev, struct vhost_virtqueue *vq,
42 uint64_t desc_addr, uint64_t desc_len)
46 uint64_t len, remain = desc_len;
48 idesc = rte_malloc(__func__, desc_len, 0);
52 dst = (uint64_t)(uintptr_t)idesc;
56 src = vhost_iova_to_vva(dev, vq, desc_addr, &len,
58 if (unlikely(!src || !len)) {
63 rte_memcpy((void *)(uintptr_t)dst, (void *)(uintptr_t)src, len);
73 static __rte_always_inline void
74 free_ind_table(void *idesc)
79 static __rte_always_inline void
80 do_flush_shadow_used_ring_split(struct virtio_net *dev,
81 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_split[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_split(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_split(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_split(dev, vq, used_idx, 0, size);
107 /* update the left half used ring interval [0, left_size] */
108 do_flush_shadow_used_ring_split(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 vq->shadow_used_idx = 0;
119 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
120 sizeof(vq->used->idx));
123 static __rte_always_inline void
124 update_shadow_used_ring_split(struct vhost_virtqueue *vq,
125 uint16_t desc_idx, uint16_t len)
127 uint16_t i = vq->shadow_used_idx++;
129 vq->shadow_used_split[i].id = desc_idx;
130 vq->shadow_used_split[i].len = len;
133 static __rte_unused __rte_always_inline void
134 flush_shadow_used_ring_packed(struct virtio_net *dev,
135 struct vhost_virtqueue *vq)
138 uint16_t used_idx = vq->last_used_idx;
140 /* Split loop in two to save memory barriers */
141 for (i = 0; i < vq->shadow_used_idx; i++) {
142 vq->desc_packed[used_idx].id = vq->shadow_used_packed[i].id;
143 vq->desc_packed[used_idx].len = vq->shadow_used_packed[i].len;
145 used_idx += vq->shadow_used_packed[i].count;
146 if (used_idx >= vq->size)
147 used_idx -= vq->size;
152 for (i = 0; i < vq->shadow_used_idx; i++) {
155 if (vq->shadow_used_packed[i].len)
156 flags = VRING_DESC_F_WRITE;
160 if (vq->used_wrap_counter) {
161 flags |= VRING_DESC_F_USED;
162 flags |= VRING_DESC_F_AVAIL;
164 flags &= ~VRING_DESC_F_USED;
165 flags &= ~VRING_DESC_F_AVAIL;
168 vq->desc_packed[vq->last_used_idx].flags = flags;
170 vhost_log_cache_used_vring(dev, vq,
172 sizeof(struct vring_packed_desc),
173 sizeof(struct vring_packed_desc));
175 vq->last_used_idx += vq->shadow_used_packed[i].count;
176 if (vq->last_used_idx >= vq->size) {
177 vq->used_wrap_counter ^= 1;
178 vq->last_used_idx -= vq->size;
183 vq->shadow_used_idx = 0;
184 vhost_log_cache_sync(dev, vq);
187 static __rte_unused __rte_always_inline void
188 update_shadow_used_ring_packed(struct vhost_virtqueue *vq,
189 uint16_t desc_idx, uint16_t len, uint16_t count)
191 uint16_t i = vq->shadow_used_idx++;
193 vq->shadow_used_packed[i].id = desc_idx;
194 vq->shadow_used_packed[i].len = len;
195 vq->shadow_used_packed[i].count = count;
199 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
201 struct batch_copy_elem *elem = vq->batch_copy_elems;
202 uint16_t count = vq->batch_copy_nb_elems;
205 for (i = 0; i < count; i++) {
206 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
207 vhost_log_cache_write(dev, vq, elem[i].log_addr, elem[i].len);
208 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
211 vq->batch_copy_nb_elems = 0;
215 do_data_copy_dequeue(struct vhost_virtqueue *vq)
217 struct batch_copy_elem *elem = vq->batch_copy_elems;
218 uint16_t count = vq->batch_copy_nb_elems;
221 for (i = 0; i < count; i++)
222 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
224 vq->batch_copy_nb_elems = 0;
227 /* avoid write operation when necessary, to lessen cache issues */
228 #define ASSIGN_UNLESS_EQUAL(var, val) do { \
229 if ((var) != (val)) \
233 static __rte_always_inline void
234 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
236 uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
238 if (m_buf->ol_flags & PKT_TX_TCP_SEG)
239 csum_l4 |= PKT_TX_TCP_CKSUM;
242 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
243 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
246 case PKT_TX_TCP_CKSUM:
247 net_hdr->csum_offset = (offsetof(struct tcp_hdr,
250 case PKT_TX_UDP_CKSUM:
251 net_hdr->csum_offset = (offsetof(struct udp_hdr,
254 case PKT_TX_SCTP_CKSUM:
255 net_hdr->csum_offset = (offsetof(struct sctp_hdr,
260 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
261 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
262 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
265 /* IP cksum verification cannot be bypassed, then calculate here */
266 if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
267 struct ipv4_hdr *ipv4_hdr;
269 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct ipv4_hdr *,
271 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
274 if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
275 if (m_buf->ol_flags & PKT_TX_IPV4)
276 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
278 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
279 net_hdr->gso_size = m_buf->tso_segsz;
280 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
282 } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
283 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
284 net_hdr->gso_size = m_buf->tso_segsz;
285 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
288 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
289 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
290 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
294 static __rte_always_inline int
295 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
296 uint32_t avail_idx, uint16_t *vec_idx,
297 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
298 uint16_t *desc_chain_len, uint8_t perm)
300 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
301 uint16_t vec_id = *vec_idx;
303 uint64_t dlen, desc_avail, desc_iova;
304 struct vring_desc *descs = vq->desc;
305 struct vring_desc *idesc = NULL;
307 *desc_chain_head = idx;
309 if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
310 dlen = vq->desc[idx].len;
311 descs = (struct vring_desc *)(uintptr_t)
312 vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
315 if (unlikely(!descs))
318 if (unlikely(dlen < vq->desc[idx].len)) {
320 * The indirect desc table is not contiguous
321 * in process VA space, we have to copy it.
323 idesc = alloc_copy_ind_table(dev, vq,
324 vq->desc[idx].addr, vq->desc[idx].len);
325 if (unlikely(!idesc))
335 if (unlikely(idx >= vq->size)) {
336 free_ind_table(idesc);
341 len += descs[idx].len;
342 desc_avail = descs[idx].len;
343 desc_iova = descs[idx].addr;
347 uint64_t desc_chunck_len = desc_avail;
349 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
350 free_ind_table(idesc);
354 desc_addr = vhost_iova_to_vva(dev, vq,
358 if (unlikely(!desc_addr)) {
359 free_ind_table(idesc);
363 buf_vec[vec_id].buf_iova = desc_iova;
364 buf_vec[vec_id].buf_addr = desc_addr;
365 buf_vec[vec_id].buf_len = desc_chunck_len;
366 buf_vec[vec_id].desc_idx = idx;
368 desc_avail -= desc_chunck_len;
369 desc_iova += desc_chunck_len;
373 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
376 idx = descs[idx].next;
379 *desc_chain_len = len;
382 if (unlikely(!!idesc))
383 free_ind_table(idesc);
389 * Returns -1 on fail, 0 on success
392 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
393 uint32_t size, struct buf_vector *buf_vec,
394 uint16_t *num_buffers, uint16_t avail_head,
398 uint16_t vec_idx = 0;
399 uint16_t max_tries, tries = 0;
401 uint16_t head_idx = 0;
405 cur_idx = vq->last_avail_idx;
407 if (rxvq_is_mergeable(dev))
408 max_tries = vq->size;
413 if (unlikely(cur_idx == avail_head))
416 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
419 VHOST_ACCESS_RW) < 0))
421 len = RTE_MIN(len, size);
422 update_shadow_used_ring_split(vq, head_idx, len);
430 * if we tried all available ring items, and still
431 * can't get enough buf, it means something abnormal
434 if (unlikely(tries > max_tries))
443 static __rte_always_inline int
444 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
445 struct rte_mbuf *m, struct buf_vector *buf_vec,
446 uint16_t nr_vec, uint16_t num_buffers)
448 uint32_t vec_idx = 0;
449 uint32_t mbuf_offset, mbuf_avail;
450 uint32_t buf_offset, buf_avail;
451 uint64_t buf_addr, buf_iova, buf_len;
454 struct rte_mbuf *hdr_mbuf;
455 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
456 struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
459 if (unlikely(m == NULL)) {
464 buf_addr = buf_vec[vec_idx].buf_addr;
465 buf_iova = buf_vec[vec_idx].buf_iova;
466 buf_len = buf_vec[vec_idx].buf_len;
469 rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
471 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
478 if (unlikely(buf_len < dev->vhost_hlen))
481 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
483 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
484 dev->vid, num_buffers);
486 if (unlikely(buf_len < dev->vhost_hlen)) {
487 buf_offset = dev->vhost_hlen - buf_len;
489 buf_addr = buf_vec[vec_idx].buf_addr;
490 buf_iova = buf_vec[vec_idx].buf_iova;
491 buf_len = buf_vec[vec_idx].buf_len;
492 buf_avail = buf_len - buf_offset;
494 buf_offset = dev->vhost_hlen;
495 buf_avail = buf_len - dev->vhost_hlen;
498 mbuf_avail = rte_pktmbuf_data_len(m);
500 while (mbuf_avail != 0 || m->next != NULL) {
501 /* done with current buf, get the next one */
502 if (buf_avail == 0) {
504 if (unlikely(vec_idx >= nr_vec)) {
509 buf_addr = buf_vec[vec_idx].buf_addr;
510 buf_iova = buf_vec[vec_idx].buf_iova;
511 buf_len = buf_vec[vec_idx].buf_len;
513 /* Prefetch next buffer address. */
514 if (vec_idx + 1 < nr_vec)
515 rte_prefetch0((void *)(uintptr_t)
516 buf_vec[vec_idx + 1].buf_addr);
521 /* done with current mbuf, get the next one */
522 if (mbuf_avail == 0) {
526 mbuf_avail = rte_pktmbuf_data_len(m);
530 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
531 if (rxvq_is_mergeable(dev))
532 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
535 if (unlikely(hdr == &tmp_hdr)) {
537 uint64_t remain = dev->vhost_hlen;
538 uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
539 uint64_t iova = buf_vec[0].buf_iova;
540 uint16_t hdr_vec_idx = 0;
544 dst = buf_vec[hdr_vec_idx].buf_addr;
545 rte_memcpy((void *)(uintptr_t)dst,
546 (void *)(uintptr_t)src,
549 PRINT_PACKET(dev, (uintptr_t)dst,
551 vhost_log_cache_write(dev, vq,
560 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
562 vhost_log_cache_write(dev, vq,
570 cpy_len = RTE_MIN(buf_len, mbuf_avail);
572 if (likely(cpy_len > MAX_BATCH_LEN ||
573 vq->batch_copy_nb_elems >= vq->size)) {
574 rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
575 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
577 vhost_log_cache_write(dev, vq, buf_iova + buf_offset,
579 PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
582 batch_copy[vq->batch_copy_nb_elems].dst =
583 (void *)((uintptr_t)(buf_addr + buf_offset));
584 batch_copy[vq->batch_copy_nb_elems].src =
585 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
586 batch_copy[vq->batch_copy_nb_elems].log_addr =
587 buf_iova + buf_offset;
588 batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
589 vq->batch_copy_nb_elems++;
592 mbuf_avail -= cpy_len;
593 mbuf_offset += cpy_len;
594 buf_avail -= cpy_len;
595 buf_offset += cpy_len;
603 static __rte_always_inline uint32_t
604 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
605 struct rte_mbuf **pkts, uint32_t count)
607 uint32_t pkt_idx = 0;
608 uint16_t num_buffers;
609 struct buf_vector buf_vec[BUF_VECTOR_MAX];
612 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
613 avail_head = *((volatile uint16_t *)&vq->avail->idx);
615 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
616 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
619 if (unlikely(reserve_avail_buf_split(dev, vq,
620 pkt_len, buf_vec, &num_buffers,
621 avail_head, &nr_vec) < 0)) {
622 VHOST_LOG_DEBUG(VHOST_DATA,
623 "(%d) failed to get enough desc from vring\n",
625 vq->shadow_used_idx -= num_buffers;
629 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
631 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
632 dev->vid, vq->last_avail_idx,
633 vq->last_avail_idx + num_buffers);
635 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
638 vq->shadow_used_idx -= num_buffers;
642 vq->last_avail_idx += num_buffers;
645 do_data_copy_enqueue(dev, vq);
647 if (likely(vq->shadow_used_idx)) {
648 flush_shadow_used_ring_split(dev, vq);
649 vhost_vring_call(dev, vq);
655 static __rte_always_inline uint32_t
656 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
657 struct rte_mbuf **pkts, uint32_t count)
659 struct vhost_virtqueue *vq;
661 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
662 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
663 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
664 dev->vid, __func__, queue_id);
668 vq = dev->virtqueue[queue_id];
670 rte_spinlock_lock(&vq->access_lock);
672 if (unlikely(vq->enabled == 0))
673 goto out_access_unlock;
675 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
676 vhost_user_iotlb_rd_lock(vq);
678 if (unlikely(vq->access_ok == 0))
679 if (unlikely(vring_translate(dev, vq) < 0))
682 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
686 count = virtio_dev_rx_split(dev, vq, pkts, count);
689 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
690 vhost_user_iotlb_rd_unlock(vq);
693 rte_spinlock_unlock(&vq->access_lock);
699 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
700 struct rte_mbuf **pkts, uint16_t count)
702 struct virtio_net *dev = get_device(vid);
707 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
708 RTE_LOG(ERR, VHOST_DATA,
709 "(%d) %s: built-in vhost net backend is disabled.\n",
714 return virtio_dev_rx(dev, queue_id, pkts, count);
718 virtio_net_with_host_offload(struct virtio_net *dev)
721 ((1ULL << VIRTIO_NET_F_CSUM) |
722 (1ULL << VIRTIO_NET_F_HOST_ECN) |
723 (1ULL << VIRTIO_NET_F_HOST_TSO4) |
724 (1ULL << VIRTIO_NET_F_HOST_TSO6) |
725 (1ULL << VIRTIO_NET_F_HOST_UFO)))
732 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
734 struct ipv4_hdr *ipv4_hdr;
735 struct ipv6_hdr *ipv6_hdr;
737 struct ether_hdr *eth_hdr;
740 eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
742 m->l2_len = sizeof(struct ether_hdr);
743 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
745 if (ethertype == ETHER_TYPE_VLAN) {
746 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
748 m->l2_len += sizeof(struct vlan_hdr);
749 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
752 l3_hdr = (char *)eth_hdr + m->l2_len;
755 case ETHER_TYPE_IPv4:
757 *l4_proto = ipv4_hdr->next_proto_id;
758 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
759 *l4_hdr = (char *)l3_hdr + m->l3_len;
760 m->ol_flags |= PKT_TX_IPV4;
762 case ETHER_TYPE_IPv6:
764 *l4_proto = ipv6_hdr->proto;
765 m->l3_len = sizeof(struct ipv6_hdr);
766 *l4_hdr = (char *)l3_hdr + m->l3_len;
767 m->ol_flags |= PKT_TX_IPV6;
777 static __rte_always_inline void
778 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
780 uint16_t l4_proto = 0;
782 struct tcp_hdr *tcp_hdr = NULL;
784 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
787 parse_ethernet(m, &l4_proto, &l4_hdr);
788 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
789 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
790 switch (hdr->csum_offset) {
791 case (offsetof(struct tcp_hdr, cksum)):
792 if (l4_proto == IPPROTO_TCP)
793 m->ol_flags |= PKT_TX_TCP_CKSUM;
795 case (offsetof(struct udp_hdr, dgram_cksum)):
796 if (l4_proto == IPPROTO_UDP)
797 m->ol_flags |= PKT_TX_UDP_CKSUM;
799 case (offsetof(struct sctp_hdr, cksum)):
800 if (l4_proto == IPPROTO_SCTP)
801 m->ol_flags |= PKT_TX_SCTP_CKSUM;
809 if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
810 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
811 case VIRTIO_NET_HDR_GSO_TCPV4:
812 case VIRTIO_NET_HDR_GSO_TCPV6:
814 m->ol_flags |= PKT_TX_TCP_SEG;
815 m->tso_segsz = hdr->gso_size;
816 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
818 case VIRTIO_NET_HDR_GSO_UDP:
819 m->ol_flags |= PKT_TX_UDP_SEG;
820 m->tso_segsz = hdr->gso_size;
821 m->l4_len = sizeof(struct udp_hdr);
824 RTE_LOG(WARNING, VHOST_DATA,
825 "unsupported gso type %u.\n", hdr->gso_type);
831 static __rte_always_inline void
832 put_zmbuf(struct zcopy_mbuf *zmbuf)
837 static __rte_always_inline int
838 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
839 struct buf_vector *buf_vec, uint16_t nr_vec,
840 struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
842 uint32_t buf_avail, buf_offset;
843 uint64_t buf_addr, buf_iova, buf_len;
844 uint32_t mbuf_avail, mbuf_offset;
846 struct rte_mbuf *cur = m, *prev = m;
847 struct virtio_net_hdr tmp_hdr;
848 struct virtio_net_hdr *hdr = NULL;
849 /* A counter to avoid desc dead loop chain */
850 uint16_t vec_idx = 0;
851 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
854 buf_addr = buf_vec[vec_idx].buf_addr;
855 buf_iova = buf_vec[vec_idx].buf_iova;
856 buf_len = buf_vec[vec_idx].buf_len;
858 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
863 if (likely(nr_vec > 1))
864 rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
866 if (virtio_net_with_host_offload(dev)) {
867 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
869 uint64_t remain = sizeof(struct virtio_net_hdr);
871 uint64_t dst = (uint64_t)(uintptr_t)&tmp_hdr;
872 uint16_t hdr_vec_idx = 0;
875 * No luck, the virtio-net header doesn't fit
876 * in a contiguous virtual area.
880 src = buf_vec[hdr_vec_idx].buf_addr;
881 rte_memcpy((void *)(uintptr_t)dst,
882 (void *)(uintptr_t)src, len);
891 hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
897 * A virtio driver normally uses at least 2 desc buffers
898 * for Tx: the first for storing the header, and others
899 * for storing the data.
901 if (unlikely(buf_len < dev->vhost_hlen)) {
902 buf_offset = dev->vhost_hlen - buf_len;
904 buf_addr = buf_vec[vec_idx].buf_addr;
905 buf_iova = buf_vec[vec_idx].buf_iova;
906 buf_len = buf_vec[vec_idx].buf_len;
907 buf_avail = buf_len - buf_offset;
908 } else if (buf_len == dev->vhost_hlen) {
909 if (unlikely(++vec_idx >= nr_vec))
911 buf_addr = buf_vec[vec_idx].buf_addr;
912 buf_iova = buf_vec[vec_idx].buf_iova;
913 buf_len = buf_vec[vec_idx].buf_len;
918 buf_offset = dev->vhost_hlen;
919 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
922 rte_prefetch0((void *)(uintptr_t)
923 (buf_addr + buf_offset));
926 (uintptr_t)(buf_addr + buf_offset),
927 (uint32_t)buf_avail, 0);
930 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
934 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
937 * A desc buf might across two host physical pages that are
938 * not continuous. In such case (gpa_to_hpa returns 0), data
939 * will be copied even though zero copy is enabled.
941 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
942 buf_iova + buf_offset, cpy_len)))) {
943 cur->data_len = cpy_len;
946 (void *)(uintptr_t)(buf_addr + buf_offset);
950 * In zero copy mode, one mbuf can only reference data
951 * for one or partial of one desc buff.
953 mbuf_avail = cpy_len;
955 if (likely(cpy_len > MAX_BATCH_LEN ||
956 vq->batch_copy_nb_elems >= vq->size ||
957 (hdr && cur == m))) {
958 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
960 (void *)((uintptr_t)(buf_addr +
964 batch_copy[vq->batch_copy_nb_elems].dst =
965 rte_pktmbuf_mtod_offset(cur, void *,
967 batch_copy[vq->batch_copy_nb_elems].src =
968 (void *)((uintptr_t)(buf_addr +
970 batch_copy[vq->batch_copy_nb_elems].len =
972 vq->batch_copy_nb_elems++;
976 mbuf_avail -= cpy_len;
977 mbuf_offset += cpy_len;
978 buf_avail -= cpy_len;
979 buf_offset += cpy_len;
981 /* This buf reaches to its end, get the next one */
982 if (buf_avail == 0) {
983 if (++vec_idx >= nr_vec)
986 buf_addr = buf_vec[vec_idx].buf_addr;
987 buf_iova = buf_vec[vec_idx].buf_iova;
988 buf_len = buf_vec[vec_idx].buf_len;
991 * Prefecth desc n + 1 buffer while
992 * desc n buffer is processed.
994 if (vec_idx + 1 < nr_vec)
995 rte_prefetch0((void *)(uintptr_t)
996 buf_vec[vec_idx + 1].buf_addr);
1001 PRINT_PACKET(dev, (uintptr_t)buf_addr,
1002 (uint32_t)buf_avail, 0);
1006 * This mbuf reaches to its end, get a new one
1007 * to hold more data.
1009 if (mbuf_avail == 0) {
1010 cur = rte_pktmbuf_alloc(mbuf_pool);
1011 if (unlikely(cur == NULL)) {
1012 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1013 "allocate memory for mbuf.\n");
1017 if (unlikely(dev->dequeue_zero_copy))
1018 rte_mbuf_refcnt_update(cur, 1);
1021 prev->data_len = mbuf_offset;
1023 m->pkt_len += mbuf_offset;
1027 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1031 prev->data_len = mbuf_offset;
1032 m->pkt_len += mbuf_offset;
1035 vhost_dequeue_offload(hdr, m);
1042 static __rte_always_inline struct zcopy_mbuf *
1043 get_zmbuf(struct vhost_virtqueue *vq)
1049 /* search [last_zmbuf_idx, zmbuf_size) */
1050 i = vq->last_zmbuf_idx;
1051 last = vq->zmbuf_size;
1054 for (; i < last; i++) {
1055 if (vq->zmbufs[i].in_use == 0) {
1056 vq->last_zmbuf_idx = i + 1;
1057 vq->zmbufs[i].in_use = 1;
1058 return &vq->zmbufs[i];
1064 /* search [0, last_zmbuf_idx) */
1066 last = vq->last_zmbuf_idx;
1073 static __rte_always_inline bool
1074 mbuf_is_consumed(struct rte_mbuf *m)
1077 if (rte_mbuf_refcnt_read(m) > 1)
1085 static __rte_always_inline void
1086 restore_mbuf(struct rte_mbuf *m)
1088 uint32_t mbuf_size, priv_size;
1091 priv_size = rte_pktmbuf_priv_size(m->pool);
1092 mbuf_size = sizeof(struct rte_mbuf) + priv_size;
1093 /* start of buffer is after mbuf structure and priv data */
1095 m->buf_addr = (char *)m + mbuf_size;
1096 m->buf_iova = rte_mempool_virt2iova(m) + mbuf_size;
1101 static __rte_always_inline uint16_t
1102 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1103 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1106 uint16_t free_entries;
1108 if (unlikely(dev->dequeue_zero_copy)) {
1109 struct zcopy_mbuf *zmbuf, *next;
1112 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1113 zmbuf != NULL; zmbuf = next) {
1114 next = TAILQ_NEXT(zmbuf, next);
1116 if (mbuf_is_consumed(zmbuf->mbuf)) {
1117 update_shadow_used_ring_split(vq,
1118 zmbuf->desc_idx, 0);
1121 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1122 restore_mbuf(zmbuf->mbuf);
1123 rte_pktmbuf_free(zmbuf->mbuf);
1129 flush_shadow_used_ring_split(dev, vq);
1130 vhost_vring_call(dev, vq);
1133 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1135 free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1137 if (free_entries == 0)
1140 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1142 count = RTE_MIN(count, MAX_PKT_BURST);
1143 count = RTE_MIN(count, free_entries);
1144 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1147 for (i = 0; i < count; i++) {
1148 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1149 uint16_t head_idx, dummy_len;
1150 uint16_t nr_vec = 0;
1153 if (unlikely(fill_vec_buf_split(dev, vq,
1154 vq->last_avail_idx + i,
1156 &head_idx, &dummy_len,
1157 VHOST_ACCESS_RO) < 0))
1160 if (likely(dev->dequeue_zero_copy == 0))
1161 update_shadow_used_ring_split(vq, head_idx, 0);
1163 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
1165 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1166 if (unlikely(pkts[i] == NULL)) {
1167 RTE_LOG(ERR, VHOST_DATA,
1168 "Failed to allocate memory for mbuf.\n");
1172 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1174 if (unlikely(err)) {
1175 rte_pktmbuf_free(pkts[i]);
1179 if (unlikely(dev->dequeue_zero_copy)) {
1180 struct zcopy_mbuf *zmbuf;
1182 zmbuf = get_zmbuf(vq);
1184 rte_pktmbuf_free(pkts[i]);
1187 zmbuf->mbuf = pkts[i];
1188 zmbuf->desc_idx = head_idx;
1191 * Pin lock the mbuf; we will check later to see
1192 * whether the mbuf is freed (when we are the last
1193 * user) or not. If that's the case, we then could
1194 * update the used ring safely.
1196 rte_mbuf_refcnt_update(pkts[i], 1);
1199 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1202 vq->last_avail_idx += i;
1204 if (likely(dev->dequeue_zero_copy == 0)) {
1205 do_data_copy_dequeue(vq);
1206 if (unlikely(i < count))
1207 vq->shadow_used_idx = i;
1208 flush_shadow_used_ring_split(dev, vq);
1209 vhost_vring_call(dev, vq);
1216 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1217 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1219 struct virtio_net *dev;
1220 struct rte_mbuf *rarp_mbuf = NULL;
1221 struct vhost_virtqueue *vq;
1223 dev = get_device(vid);
1227 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1228 RTE_LOG(ERR, VHOST_DATA,
1229 "(%d) %s: built-in vhost net backend is disabled.\n",
1230 dev->vid, __func__);
1234 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1235 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1236 dev->vid, __func__, queue_id);
1240 vq = dev->virtqueue[queue_id];
1242 if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
1245 if (unlikely(vq->enabled == 0))
1246 goto out_access_unlock;
1248 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1249 vhost_user_iotlb_rd_lock(vq);
1251 if (unlikely(vq->access_ok == 0))
1252 if (unlikely(vring_translate(dev, vq) < 0))
1256 * Construct a RARP broadcast packet, and inject it to the "pkts"
1257 * array, to looks like that guest actually send such packet.
1259 * Check user_send_rarp() for more information.
1261 * broadcast_rarp shares a cacheline in the virtio_net structure
1262 * with some fields that are accessed during enqueue and
1263 * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1264 * result in false sharing between enqueue and dequeue.
1266 * Prevent unnecessary false sharing by reading broadcast_rarp first
1267 * and only performing cmpset if the read indicates it is likely to
1270 if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1271 rte_atomic16_cmpset((volatile uint16_t *)
1272 &dev->broadcast_rarp.cnt, 1, 0))) {
1274 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
1275 if (rarp_mbuf == NULL) {
1276 RTE_LOG(ERR, VHOST_DATA,
1277 "Failed to make RARP packet.\n");
1283 count = virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count);
1286 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1287 vhost_user_iotlb_rd_unlock(vq);
1290 rte_spinlock_unlock(&vq->access_lock);
1292 if (unlikely(rarp_mbuf != NULL)) {
1294 * Inject it to the head of "pkts" array, so that switch's mac
1295 * learning table will get updated first.
1297 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
1298 pkts[0] = rarp_mbuf;