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 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
707 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
709 return virtio_dev_rx(dev, queue_id, pkts, count);
713 virtio_net_with_host_offload(struct virtio_net *dev)
716 ((1ULL << VIRTIO_NET_F_CSUM) |
717 (1ULL << VIRTIO_NET_F_HOST_ECN) |
718 (1ULL << VIRTIO_NET_F_HOST_TSO4) |
719 (1ULL << VIRTIO_NET_F_HOST_TSO6) |
720 (1ULL << VIRTIO_NET_F_HOST_UFO)))
727 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
729 struct ipv4_hdr *ipv4_hdr;
730 struct ipv6_hdr *ipv6_hdr;
732 struct ether_hdr *eth_hdr;
735 eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
737 m->l2_len = sizeof(struct ether_hdr);
738 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
740 if (ethertype == ETHER_TYPE_VLAN) {
741 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
743 m->l2_len += sizeof(struct vlan_hdr);
744 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
747 l3_hdr = (char *)eth_hdr + m->l2_len;
750 case ETHER_TYPE_IPv4:
752 *l4_proto = ipv4_hdr->next_proto_id;
753 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
754 *l4_hdr = (char *)l3_hdr + m->l3_len;
755 m->ol_flags |= PKT_TX_IPV4;
757 case ETHER_TYPE_IPv6:
759 *l4_proto = ipv6_hdr->proto;
760 m->l3_len = sizeof(struct ipv6_hdr);
761 *l4_hdr = (char *)l3_hdr + m->l3_len;
762 m->ol_flags |= PKT_TX_IPV6;
772 static __rte_always_inline void
773 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
775 uint16_t l4_proto = 0;
777 struct tcp_hdr *tcp_hdr = NULL;
779 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
782 parse_ethernet(m, &l4_proto, &l4_hdr);
783 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
784 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
785 switch (hdr->csum_offset) {
786 case (offsetof(struct tcp_hdr, cksum)):
787 if (l4_proto == IPPROTO_TCP)
788 m->ol_flags |= PKT_TX_TCP_CKSUM;
790 case (offsetof(struct udp_hdr, dgram_cksum)):
791 if (l4_proto == IPPROTO_UDP)
792 m->ol_flags |= PKT_TX_UDP_CKSUM;
794 case (offsetof(struct sctp_hdr, cksum)):
795 if (l4_proto == IPPROTO_SCTP)
796 m->ol_flags |= PKT_TX_SCTP_CKSUM;
804 if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
805 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
806 case VIRTIO_NET_HDR_GSO_TCPV4:
807 case VIRTIO_NET_HDR_GSO_TCPV6:
809 m->ol_flags |= PKT_TX_TCP_SEG;
810 m->tso_segsz = hdr->gso_size;
811 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
813 case VIRTIO_NET_HDR_GSO_UDP:
814 m->ol_flags |= PKT_TX_UDP_SEG;
815 m->tso_segsz = hdr->gso_size;
816 m->l4_len = sizeof(struct udp_hdr);
819 RTE_LOG(WARNING, VHOST_DATA,
820 "unsupported gso type %u.\n", hdr->gso_type);
826 static __rte_always_inline void
827 put_zmbuf(struct zcopy_mbuf *zmbuf)
832 static __rte_always_inline int
833 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
834 struct vring_desc *descs, uint16_t max_desc,
835 struct rte_mbuf *m, uint16_t desc_idx,
836 struct rte_mempool *mbuf_pool)
838 struct vring_desc *desc;
840 uint32_t desc_avail, desc_offset;
841 uint32_t mbuf_avail, mbuf_offset;
843 struct rte_mbuf *cur = m, *prev = m;
844 struct virtio_net_hdr *hdr = NULL;
845 /* A counter to avoid desc dead loop chain */
846 uint32_t nr_desc = 1;
847 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
848 uint16_t copy_nb = vq->batch_copy_nb_elems;
851 desc = &descs[desc_idx];
852 if (unlikely((desc->len < dev->vhost_hlen)) ||
853 (desc->flags & VRING_DESC_F_INDIRECT)) {
858 desc_addr = vhost_iova_to_vva(dev,
862 if (unlikely(!desc_addr)) {
867 if (virtio_net_with_host_offload(dev)) {
868 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
873 * A virtio driver normally uses at least 2 desc buffers
874 * for Tx: the first for storing the header, and others
875 * for storing the data.
877 if (likely((desc->len == dev->vhost_hlen) &&
878 (desc->flags & VRING_DESC_F_NEXT) != 0)) {
879 desc = &descs[desc->next];
880 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
885 desc_addr = vhost_iova_to_vva(dev,
889 if (unlikely(!desc_addr)) {
895 desc_avail = desc->len;
898 desc_avail = desc->len - dev->vhost_hlen;
899 desc_offset = dev->vhost_hlen;
902 rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
904 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
907 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
911 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
914 * A desc buf might across two host physical pages that are
915 * not continuous. In such case (gpa_to_hpa returns 0), data
916 * will be copied even though zero copy is enabled.
918 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
919 desc->addr + desc_offset, cpy_len)))) {
920 cur->data_len = cpy_len;
922 cur->buf_addr = (void *)(uintptr_t)(desc_addr
927 * In zero copy mode, one mbuf can only reference data
928 * for one or partial of one desc buff.
930 mbuf_avail = cpy_len;
932 if (likely(cpy_len > MAX_BATCH_LEN ||
933 copy_nb >= vq->size ||
934 (hdr && cur == m))) {
935 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
937 (void *)((uintptr_t)(desc_addr +
941 batch_copy[copy_nb].dst =
942 rte_pktmbuf_mtod_offset(cur, void *,
944 batch_copy[copy_nb].src =
945 (void *)((uintptr_t)(desc_addr +
947 batch_copy[copy_nb].len = cpy_len;
952 mbuf_avail -= cpy_len;
953 mbuf_offset += cpy_len;
954 desc_avail -= cpy_len;
955 desc_offset += cpy_len;
957 /* This desc reaches to its end, get the next one */
958 if (desc_avail == 0) {
959 if ((desc->flags & VRING_DESC_F_NEXT) == 0)
962 if (unlikely(desc->next >= max_desc ||
963 ++nr_desc > max_desc)) {
967 desc = &descs[desc->next];
968 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
973 desc_addr = vhost_iova_to_vva(dev,
977 if (unlikely(!desc_addr)) {
982 rte_prefetch0((void *)(uintptr_t)desc_addr);
985 desc_avail = desc->len;
987 PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
991 * This mbuf reaches to its end, get a new one
994 if (mbuf_avail == 0) {
995 cur = rte_pktmbuf_alloc(mbuf_pool);
996 if (unlikely(cur == NULL)) {
997 RTE_LOG(ERR, VHOST_DATA, "Failed to "
998 "allocate memory for mbuf.\n");
1002 if (unlikely(dev->dequeue_zero_copy))
1003 rte_mbuf_refcnt_update(cur, 1);
1006 prev->data_len = mbuf_offset;
1008 m->pkt_len += mbuf_offset;
1012 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1016 prev->data_len = mbuf_offset;
1017 m->pkt_len += mbuf_offset;
1020 vhost_dequeue_offload(hdr, m);
1023 vq->batch_copy_nb_elems = copy_nb;
1028 static __rte_always_inline void
1029 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
1030 uint32_t used_idx, uint32_t desc_idx)
1032 vq->used->ring[used_idx].id = desc_idx;
1033 vq->used->ring[used_idx].len = 0;
1034 vhost_log_used_vring(dev, vq,
1035 offsetof(struct vring_used, ring[used_idx]),
1036 sizeof(vq->used->ring[used_idx]));
1039 static __rte_always_inline void
1040 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
1043 if (unlikely(count == 0))
1049 vq->used->idx += count;
1050 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
1051 sizeof(vq->used->idx));
1052 vhost_vring_call(dev, vq);
1055 static __rte_always_inline struct zcopy_mbuf *
1056 get_zmbuf(struct vhost_virtqueue *vq)
1062 /* search [last_zmbuf_idx, zmbuf_size) */
1063 i = vq->last_zmbuf_idx;
1064 last = vq->zmbuf_size;
1067 for (; i < last; i++) {
1068 if (vq->zmbufs[i].in_use == 0) {
1069 vq->last_zmbuf_idx = i + 1;
1070 vq->zmbufs[i].in_use = 1;
1071 return &vq->zmbufs[i];
1077 /* search [0, last_zmbuf_idx) */
1079 last = vq->last_zmbuf_idx;
1086 static __rte_always_inline bool
1087 mbuf_is_consumed(struct rte_mbuf *m)
1090 if (rte_mbuf_refcnt_read(m) > 1)
1098 static __rte_always_inline void
1099 restore_mbuf(struct rte_mbuf *m)
1101 uint32_t mbuf_size, priv_size;
1104 priv_size = rte_pktmbuf_priv_size(m->pool);
1105 mbuf_size = sizeof(struct rte_mbuf) + priv_size;
1106 /* start of buffer is after mbuf structure and priv data */
1108 m->buf_addr = (char *)m + mbuf_size;
1109 m->buf_iova = rte_mempool_virt2iova(m) + mbuf_size;
1115 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1116 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1118 struct virtio_net *dev;
1119 struct rte_mbuf *rarp_mbuf = NULL;
1120 struct vhost_virtqueue *vq;
1121 uint32_t desc_indexes[MAX_PKT_BURST];
1124 uint16_t free_entries;
1127 dev = get_device(vid);
1131 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1132 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1133 dev->vid, __func__, queue_id);
1137 vq = dev->virtqueue[queue_id];
1139 if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
1142 if (unlikely(vq->enabled == 0))
1143 goto out_access_unlock;
1145 vq->batch_copy_nb_elems = 0;
1147 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1148 vhost_user_iotlb_rd_lock(vq);
1150 if (unlikely(vq->access_ok == 0))
1151 if (unlikely(vring_translate(dev, vq) < 0))
1154 if (unlikely(dev->dequeue_zero_copy)) {
1155 struct zcopy_mbuf *zmbuf, *next;
1158 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1159 zmbuf != NULL; zmbuf = next) {
1160 next = TAILQ_NEXT(zmbuf, next);
1162 if (mbuf_is_consumed(zmbuf->mbuf)) {
1163 used_idx = vq->last_used_idx++ & (vq->size - 1);
1164 update_used_ring(dev, vq, used_idx,
1168 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1169 restore_mbuf(zmbuf->mbuf);
1170 rte_pktmbuf_free(zmbuf->mbuf);
1176 update_used_idx(dev, vq, nr_updated);
1180 * Construct a RARP broadcast packet, and inject it to the "pkts"
1181 * array, to looks like that guest actually send such packet.
1183 * Check user_send_rarp() for more information.
1185 * broadcast_rarp shares a cacheline in the virtio_net structure
1186 * with some fields that are accessed during enqueue and
1187 * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1188 * result in false sharing between enqueue and dequeue.
1190 * Prevent unnecessary false sharing by reading broadcast_rarp first
1191 * and only performing cmpset if the read indicates it is likely to
1195 if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1196 rte_atomic16_cmpset((volatile uint16_t *)
1197 &dev->broadcast_rarp.cnt, 1, 0))) {
1199 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
1200 if (rarp_mbuf == NULL) {
1201 RTE_LOG(ERR, VHOST_DATA,
1202 "Failed to make RARP packet.\n");
1208 free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1210 if (free_entries == 0)
1213 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1215 /* Prefetch available and used ring */
1216 avail_idx = vq->last_avail_idx & (vq->size - 1);
1217 used_idx = vq->last_used_idx & (vq->size - 1);
1218 rte_prefetch0(&vq->avail->ring[avail_idx]);
1219 rte_prefetch0(&vq->used->ring[used_idx]);
1221 count = RTE_MIN(count, MAX_PKT_BURST);
1222 count = RTE_MIN(count, free_entries);
1223 LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1226 /* Retrieve all of the head indexes first to avoid caching issues. */
1227 for (i = 0; i < count; i++) {
1228 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1229 used_idx = (vq->last_used_idx + i) & (vq->size - 1);
1230 desc_indexes[i] = vq->avail->ring[avail_idx];
1232 if (likely(dev->dequeue_zero_copy == 0))
1233 update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1236 /* Prefetch descriptor index. */
1237 rte_prefetch0(&vq->desc[desc_indexes[0]]);
1238 for (i = 0; i < count; i++) {
1239 struct vring_desc *desc;
1243 if (likely(i + 1 < count))
1244 rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1246 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1247 desc = (struct vring_desc *)(uintptr_t)
1248 vhost_iova_to_vva(dev, vq,
1249 vq->desc[desc_indexes[i]].addr,
1252 if (unlikely(!desc))
1255 rte_prefetch0(desc);
1256 sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1261 idx = desc_indexes[i];
1264 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1265 if (unlikely(pkts[i] == NULL)) {
1266 RTE_LOG(ERR, VHOST_DATA,
1267 "Failed to allocate memory for mbuf.\n");
1271 err = copy_desc_to_mbuf(dev, vq, desc, sz, pkts[i], idx,
1273 if (unlikely(err)) {
1274 rte_pktmbuf_free(pkts[i]);
1278 if (unlikely(dev->dequeue_zero_copy)) {
1279 struct zcopy_mbuf *zmbuf;
1281 zmbuf = get_zmbuf(vq);
1283 rte_pktmbuf_free(pkts[i]);
1286 zmbuf->mbuf = pkts[i];
1287 zmbuf->desc_idx = desc_indexes[i];
1290 * Pin lock the mbuf; we will check later to see
1291 * whether the mbuf is freed (when we are the last
1292 * user) or not. If that's the case, we then could
1293 * update the used ring safely.
1295 rte_mbuf_refcnt_update(pkts[i], 1);
1298 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1301 vq->last_avail_idx += i;
1303 if (likely(dev->dequeue_zero_copy == 0)) {
1304 do_data_copy_dequeue(vq);
1305 vq->last_used_idx += i;
1306 update_used_idx(dev, vq, i);
1310 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1311 vhost_user_iotlb_rd_unlock(vq);
1314 rte_spinlock_unlock(&vq->access_lock);
1316 if (unlikely(rarp_mbuf != NULL)) {
1318 * Inject it to the head of "pkts" array, so that switch's mac
1319 * learning table will get updated first.
1321 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1322 pkts[0] = rarp_mbuf;