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>
22 #define MAX_PKT_BURST 32
24 #define MAX_BATCH_LEN 256
27 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
29 return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
32 static __rte_always_inline void
33 do_flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
34 uint16_t to, uint16_t from, uint16_t size)
36 rte_memcpy(&vq->used->ring[to],
37 &vq->shadow_used_ring[from],
38 size * sizeof(struct vring_used_elem));
39 vhost_log_used_vring(dev, vq,
40 offsetof(struct vring_used, ring[to]),
41 size * sizeof(struct vring_used_elem));
44 static __rte_always_inline void
45 flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq)
47 uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
49 if (used_idx + vq->shadow_used_idx <= vq->size) {
50 do_flush_shadow_used_ring(dev, vq, used_idx, 0,
55 /* update used ring interval [used_idx, vq->size] */
56 size = vq->size - used_idx;
57 do_flush_shadow_used_ring(dev, vq, used_idx, 0, size);
59 /* update the left half used ring interval [0, left_size] */
60 do_flush_shadow_used_ring(dev, vq, 0, size,
61 vq->shadow_used_idx - size);
63 vq->last_used_idx += vq->shadow_used_idx;
67 *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
68 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
69 sizeof(vq->used->idx));
72 static __rte_always_inline void
73 update_shadow_used_ring(struct vhost_virtqueue *vq,
74 uint16_t desc_idx, uint16_t len)
76 uint16_t i = vq->shadow_used_idx++;
78 vq->shadow_used_ring[i].id = desc_idx;
79 vq->shadow_used_ring[i].len = len;
83 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
85 struct batch_copy_elem *elem = vq->batch_copy_elems;
86 uint16_t count = vq->batch_copy_nb_elems;
89 for (i = 0; i < count; i++) {
90 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
91 vhost_log_write(dev, elem[i].log_addr, elem[i].len);
92 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
97 do_data_copy_dequeue(struct vhost_virtqueue *vq)
99 struct batch_copy_elem *elem = vq->batch_copy_elems;
100 uint16_t count = vq->batch_copy_nb_elems;
103 for (i = 0; i < count; i++)
104 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
107 /* avoid write operation when necessary, to lessen cache issues */
108 #define ASSIGN_UNLESS_EQUAL(var, val) do { \
109 if ((var) != (val)) \
114 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
116 uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
118 if (m_buf->ol_flags & PKT_TX_TCP_SEG)
119 csum_l4 |= PKT_TX_TCP_CKSUM;
122 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
123 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
126 case PKT_TX_TCP_CKSUM:
127 net_hdr->csum_offset = (offsetof(struct tcp_hdr,
130 case PKT_TX_UDP_CKSUM:
131 net_hdr->csum_offset = (offsetof(struct udp_hdr,
134 case PKT_TX_SCTP_CKSUM:
135 net_hdr->csum_offset = (offsetof(struct sctp_hdr,
140 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
141 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
142 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
145 /* IP cksum verification cannot be bypassed, then calculate here */
146 if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
147 struct ipv4_hdr *ipv4_hdr;
149 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct ipv4_hdr *,
151 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
154 if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
155 if (m_buf->ol_flags & PKT_TX_IPV4)
156 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
158 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
159 net_hdr->gso_size = m_buf->tso_segsz;
160 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
163 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
164 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
165 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
169 static __rte_always_inline int
170 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
171 struct vring_desc *descs, struct rte_mbuf *m,
172 uint16_t desc_idx, uint32_t size)
174 uint32_t desc_avail, desc_offset;
175 uint32_t mbuf_avail, mbuf_offset;
177 struct vring_desc *desc;
179 /* A counter to avoid desc dead loop chain */
180 uint16_t nr_desc = 1;
181 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
182 uint16_t copy_nb = vq->batch_copy_nb_elems;
185 desc = &descs[desc_idx];
186 desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
187 desc->len, VHOST_ACCESS_RW);
189 * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
190 * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
191 * otherwise stores offset on the stack instead of in a register.
193 if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr) {
198 rte_prefetch0((void *)(uintptr_t)desc_addr);
200 virtio_enqueue_offload(m, (struct virtio_net_hdr *)(uintptr_t)desc_addr);
201 vhost_log_write(dev, desc->addr, dev->vhost_hlen);
202 PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
204 desc_offset = dev->vhost_hlen;
205 desc_avail = desc->len - dev->vhost_hlen;
207 mbuf_avail = rte_pktmbuf_data_len(m);
209 while (mbuf_avail != 0 || m->next != NULL) {
210 /* done with current mbuf, fetch next */
211 if (mbuf_avail == 0) {
215 mbuf_avail = rte_pktmbuf_data_len(m);
218 /* done with current desc buf, fetch next */
219 if (desc_avail == 0) {
220 if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
221 /* Room in vring buffer is not enough */
225 if (unlikely(desc->next >= size || ++nr_desc > size)) {
230 desc = &descs[desc->next];
231 desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
234 if (unlikely(!desc_addr)) {
240 desc_avail = desc->len;
243 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
244 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
245 rte_memcpy((void *)((uintptr_t)(desc_addr +
247 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
249 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
250 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
253 batch_copy[copy_nb].dst =
254 (void *)((uintptr_t)(desc_addr + desc_offset));
255 batch_copy[copy_nb].src =
256 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
257 batch_copy[copy_nb].log_addr = desc->addr + desc_offset;
258 batch_copy[copy_nb].len = cpy_len;
262 mbuf_avail -= cpy_len;
263 mbuf_offset += cpy_len;
264 desc_avail -= cpy_len;
265 desc_offset += cpy_len;
269 vq->batch_copy_nb_elems = copy_nb;
275 * This function adds buffers to the virtio devices RX virtqueue. Buffers can
276 * be received from the physical port or from another virtio device. A packet
277 * count is returned to indicate the number of packets that are successfully
278 * added to the RX queue. This function works when the mbuf is scattered, but
279 * it doesn't support the mergeable feature.
281 static __rte_always_inline uint32_t
282 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
283 struct rte_mbuf **pkts, uint32_t count)
285 struct vhost_virtqueue *vq;
286 uint16_t avail_idx, free_entries, start_idx;
287 uint16_t desc_indexes[MAX_PKT_BURST];
288 struct vring_desc *descs;
292 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
293 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
294 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
295 dev->vid, __func__, queue_id);
299 vq = dev->virtqueue[queue_id];
300 if (unlikely(vq->enabled == 0))
303 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
304 vhost_user_iotlb_rd_lock(vq);
306 if (unlikely(vq->access_ok == 0)) {
307 if (unlikely(vring_translate(dev, vq) < 0)) {
313 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
314 start_idx = vq->last_used_idx;
315 free_entries = avail_idx - start_idx;
316 count = RTE_MIN(count, free_entries);
317 count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
321 LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
322 dev->vid, start_idx, start_idx + count);
324 vq->batch_copy_nb_elems = 0;
326 /* Retrieve all of the desc indexes first to avoid caching issues. */
327 rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
328 for (i = 0; i < count; i++) {
329 used_idx = (start_idx + i) & (vq->size - 1);
330 desc_indexes[i] = vq->avail->ring[used_idx];
331 vq->used->ring[used_idx].id = desc_indexes[i];
332 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
334 vhost_log_used_vring(dev, vq,
335 offsetof(struct vring_used, ring[used_idx]),
336 sizeof(vq->used->ring[used_idx]));
339 rte_prefetch0(&vq->desc[desc_indexes[0]]);
340 for (i = 0; i < count; i++) {
341 uint16_t desc_idx = desc_indexes[i];
344 if (vq->desc[desc_idx].flags & VRING_DESC_F_INDIRECT) {
345 descs = (struct vring_desc *)(uintptr_t)
346 vhost_iova_to_vva(dev,
347 vq, vq->desc[desc_idx].addr,
348 vq->desc[desc_idx].len,
350 if (unlikely(!descs)) {
356 sz = vq->desc[desc_idx].len / sizeof(*descs);
362 err = copy_mbuf_to_desc(dev, vq, descs, pkts[i], desc_idx, sz);
369 rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
372 do_data_copy_enqueue(dev, vq);
376 *(volatile uint16_t *)&vq->used->idx += count;
377 vq->last_used_idx += count;
378 vhost_log_used_vring(dev, vq,
379 offsetof(struct vring_used, idx),
380 sizeof(vq->used->idx));
382 /* flush used->idx update before we read avail->flags. */
385 /* Kick the guest if necessary. */
386 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
387 && (vq->callfd >= 0))
388 eventfd_write(vq->callfd, (eventfd_t)1);
390 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
391 vhost_user_iotlb_rd_unlock(vq);
396 static __rte_always_inline int
397 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
398 uint32_t avail_idx, uint32_t *vec_idx,
399 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
400 uint16_t *desc_chain_len)
402 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
403 uint32_t vec_id = *vec_idx;
405 struct vring_desc *descs = vq->desc;
407 *desc_chain_head = idx;
409 if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
410 descs = (struct vring_desc *)(uintptr_t)
411 vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
414 if (unlikely(!descs))
421 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
424 len += descs[idx].len;
425 buf_vec[vec_id].buf_addr = descs[idx].addr;
426 buf_vec[vec_id].buf_len = descs[idx].len;
427 buf_vec[vec_id].desc_idx = idx;
430 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
433 idx = descs[idx].next;
436 *desc_chain_len = len;
443 * Returns -1 on fail, 0 on success
446 reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
447 uint32_t size, struct buf_vector *buf_vec,
448 uint16_t *num_buffers, uint16_t avail_head)
451 uint32_t vec_idx = 0;
454 uint16_t head_idx = 0;
458 cur_idx = vq->last_avail_idx;
461 if (unlikely(cur_idx == avail_head))
464 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
465 &head_idx, &len) < 0))
467 len = RTE_MIN(len, size);
468 update_shadow_used_ring(vq, head_idx, len);
476 * if we tried all available ring items, and still
477 * can't get enough buf, it means something abnormal
480 if (unlikely(tries >= vq->size))
487 static __rte_always_inline int
488 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
489 struct rte_mbuf *m, struct buf_vector *buf_vec,
490 uint16_t num_buffers)
492 uint32_t vec_idx = 0;
494 uint32_t mbuf_offset, mbuf_avail;
495 uint32_t desc_offset, desc_avail;
497 uint64_t hdr_addr, hdr_phys_addr;
498 struct rte_mbuf *hdr_mbuf;
499 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
500 uint16_t copy_nb = vq->batch_copy_nb_elems;
503 if (unlikely(m == NULL)) {
508 desc_addr = vhost_iova_to_vva(dev, vq, buf_vec[vec_idx].buf_addr,
509 buf_vec[vec_idx].buf_len,
511 if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr) {
517 hdr_addr = desc_addr;
518 hdr_phys_addr = buf_vec[vec_idx].buf_addr;
519 rte_prefetch0((void *)(uintptr_t)hdr_addr);
521 LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
522 dev->vid, num_buffers);
524 desc_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
525 desc_offset = dev->vhost_hlen;
527 mbuf_avail = rte_pktmbuf_data_len(m);
529 while (mbuf_avail != 0 || m->next != NULL) {
530 /* done with current desc buf, get the next one */
531 if (desc_avail == 0) {
534 vhost_iova_to_vva(dev, vq,
535 buf_vec[vec_idx].buf_addr,
536 buf_vec[vec_idx].buf_len,
538 if (unlikely(!desc_addr)) {
543 /* Prefetch buffer address. */
544 rte_prefetch0((void *)(uintptr_t)desc_addr);
546 desc_avail = buf_vec[vec_idx].buf_len;
549 /* done with current mbuf, get the next one */
550 if (mbuf_avail == 0) {
554 mbuf_avail = rte_pktmbuf_data_len(m);
558 struct virtio_net_hdr_mrg_rxbuf *hdr;
560 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)
562 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
563 ASSIGN_UNLESS_EQUAL(hdr->num_buffers, num_buffers);
565 vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
566 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
572 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
574 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
575 rte_memcpy((void *)((uintptr_t)(desc_addr +
577 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
580 buf_vec[vec_idx].buf_addr + desc_offset,
582 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
585 batch_copy[copy_nb].dst =
586 (void *)((uintptr_t)(desc_addr + desc_offset));
587 batch_copy[copy_nb].src =
588 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
589 batch_copy[copy_nb].log_addr =
590 buf_vec[vec_idx].buf_addr + desc_offset;
591 batch_copy[copy_nb].len = cpy_len;
595 mbuf_avail -= cpy_len;
596 mbuf_offset += cpy_len;
597 desc_avail -= cpy_len;
598 desc_offset += cpy_len;
602 vq->batch_copy_nb_elems = copy_nb;
607 static __rte_always_inline uint32_t
608 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
609 struct rte_mbuf **pkts, uint32_t count)
611 struct vhost_virtqueue *vq;
612 uint32_t pkt_idx = 0;
613 uint16_t num_buffers;
614 struct buf_vector buf_vec[BUF_VECTOR_MAX];
617 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
618 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
619 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
620 dev->vid, __func__, queue_id);
624 vq = dev->virtqueue[queue_id];
625 if (unlikely(vq->enabled == 0))
628 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
629 vhost_user_iotlb_rd_lock(vq);
631 if (unlikely(vq->access_ok == 0))
632 if (unlikely(vring_translate(dev, vq) < 0))
635 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
639 vq->batch_copy_nb_elems = 0;
641 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
643 vq->shadow_used_idx = 0;
644 avail_head = *((volatile uint16_t *)&vq->avail->idx);
645 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
646 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
648 if (unlikely(reserve_avail_buf_mergeable(dev, vq,
649 pkt_len, buf_vec, &num_buffers,
651 LOG_DEBUG(VHOST_DATA,
652 "(%d) failed to get enough desc from vring\n",
654 vq->shadow_used_idx -= num_buffers;
658 LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
659 dev->vid, vq->last_avail_idx,
660 vq->last_avail_idx + num_buffers);
662 if (copy_mbuf_to_desc_mergeable(dev, vq, pkts[pkt_idx],
663 buf_vec, num_buffers) < 0) {
664 vq->shadow_used_idx -= num_buffers;
668 vq->last_avail_idx += num_buffers;
671 do_data_copy_enqueue(dev, vq);
673 if (likely(vq->shadow_used_idx)) {
674 flush_shadow_used_ring(dev, vq);
676 /* flush used->idx update before we read avail->flags. */
679 /* Kick the guest if necessary. */
680 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
681 && (vq->callfd >= 0))
682 eventfd_write(vq->callfd, (eventfd_t)1);
686 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
687 vhost_user_iotlb_rd_unlock(vq);
693 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
694 struct rte_mbuf **pkts, uint16_t count)
696 struct virtio_net *dev = get_device(vid);
701 if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
702 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
704 return virtio_dev_rx(dev, queue_id, pkts, count);
708 virtio_net_with_host_offload(struct virtio_net *dev)
711 ((1ULL << VIRTIO_NET_F_CSUM) |
712 (1ULL << VIRTIO_NET_F_HOST_ECN) |
713 (1ULL << VIRTIO_NET_F_HOST_TSO4) |
714 (1ULL << VIRTIO_NET_F_HOST_TSO6) |
715 (1ULL << VIRTIO_NET_F_HOST_UFO)))
722 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
724 struct ipv4_hdr *ipv4_hdr;
725 struct ipv6_hdr *ipv6_hdr;
727 struct ether_hdr *eth_hdr;
730 eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
732 m->l2_len = sizeof(struct ether_hdr);
733 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
735 if (ethertype == ETHER_TYPE_VLAN) {
736 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
738 m->l2_len += sizeof(struct vlan_hdr);
739 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
742 l3_hdr = (char *)eth_hdr + m->l2_len;
745 case ETHER_TYPE_IPv4:
747 *l4_proto = ipv4_hdr->next_proto_id;
748 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
749 *l4_hdr = (char *)l3_hdr + m->l3_len;
750 m->ol_flags |= PKT_TX_IPV4;
752 case ETHER_TYPE_IPv6:
754 *l4_proto = ipv6_hdr->proto;
755 m->l3_len = sizeof(struct ipv6_hdr);
756 *l4_hdr = (char *)l3_hdr + m->l3_len;
757 m->ol_flags |= PKT_TX_IPV6;
767 static __rte_always_inline void
768 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
770 uint16_t l4_proto = 0;
772 struct tcp_hdr *tcp_hdr = NULL;
774 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
777 parse_ethernet(m, &l4_proto, &l4_hdr);
778 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
779 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
780 switch (hdr->csum_offset) {
781 case (offsetof(struct tcp_hdr, cksum)):
782 if (l4_proto == IPPROTO_TCP)
783 m->ol_flags |= PKT_TX_TCP_CKSUM;
785 case (offsetof(struct udp_hdr, dgram_cksum)):
786 if (l4_proto == IPPROTO_UDP)
787 m->ol_flags |= PKT_TX_UDP_CKSUM;
789 case (offsetof(struct sctp_hdr, cksum)):
790 if (l4_proto == IPPROTO_SCTP)
791 m->ol_flags |= PKT_TX_SCTP_CKSUM;
799 if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
800 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
801 case VIRTIO_NET_HDR_GSO_TCPV4:
802 case VIRTIO_NET_HDR_GSO_TCPV6:
804 m->ol_flags |= PKT_TX_TCP_SEG;
805 m->tso_segsz = hdr->gso_size;
806 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
809 RTE_LOG(WARNING, VHOST_DATA,
810 "unsupported gso type %u.\n", hdr->gso_type);
816 #define RARP_PKT_SIZE 64
819 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
821 struct ether_hdr *eth_hdr;
822 struct arp_hdr *rarp;
824 if (rarp_mbuf->buf_len < 64) {
825 RTE_LOG(WARNING, VHOST_DATA,
826 "failed to make RARP; mbuf size too small %u (< %d)\n",
827 rarp_mbuf->buf_len, RARP_PKT_SIZE);
831 /* Ethernet header. */
832 eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
833 memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
834 ether_addr_copy(mac, ð_hdr->s_addr);
835 eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
838 rarp = (struct arp_hdr *)(eth_hdr + 1);
839 rarp->arp_hrd = htons(ARP_HRD_ETHER);
840 rarp->arp_pro = htons(ETHER_TYPE_IPv4);
841 rarp->arp_hln = ETHER_ADDR_LEN;
843 rarp->arp_op = htons(ARP_OP_REVREQUEST);
845 ether_addr_copy(mac, &rarp->arp_data.arp_sha);
846 ether_addr_copy(mac, &rarp->arp_data.arp_tha);
847 memset(&rarp->arp_data.arp_sip, 0x00, 4);
848 memset(&rarp->arp_data.arp_tip, 0x00, 4);
850 rarp_mbuf->pkt_len = rarp_mbuf->data_len = RARP_PKT_SIZE;
855 static __rte_always_inline void
856 put_zmbuf(struct zcopy_mbuf *zmbuf)
861 static __rte_always_inline int
862 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
863 struct vring_desc *descs, uint16_t max_desc,
864 struct rte_mbuf *m, uint16_t desc_idx,
865 struct rte_mempool *mbuf_pool)
867 struct vring_desc *desc;
869 uint32_t desc_avail, desc_offset;
870 uint32_t mbuf_avail, mbuf_offset;
872 struct rte_mbuf *cur = m, *prev = m;
873 struct virtio_net_hdr *hdr = NULL;
874 /* A counter to avoid desc dead loop chain */
875 uint32_t nr_desc = 1;
876 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
877 uint16_t copy_nb = vq->batch_copy_nb_elems;
880 desc = &descs[desc_idx];
881 if (unlikely((desc->len < dev->vhost_hlen)) ||
882 (desc->flags & VRING_DESC_F_INDIRECT)) {
887 desc_addr = vhost_iova_to_vva(dev,
891 if (unlikely(!desc_addr)) {
896 if (virtio_net_with_host_offload(dev)) {
897 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
902 * A virtio driver normally uses at least 2 desc buffers
903 * for Tx: the first for storing the header, and others
904 * for storing the data.
906 if (likely((desc->len == dev->vhost_hlen) &&
907 (desc->flags & VRING_DESC_F_NEXT) != 0)) {
908 desc = &descs[desc->next];
909 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
914 desc_addr = vhost_iova_to_vva(dev,
918 if (unlikely(!desc_addr)) {
924 desc_avail = desc->len;
927 desc_avail = desc->len - dev->vhost_hlen;
928 desc_offset = dev->vhost_hlen;
931 rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
933 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
936 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
940 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
943 * A desc buf might across two host physical pages that are
944 * not continuous. In such case (gpa_to_hpa returns 0), data
945 * will be copied even though zero copy is enabled.
947 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
948 desc->addr + desc_offset, cpy_len)))) {
949 cur->data_len = cpy_len;
951 cur->buf_addr = (void *)(uintptr_t)desc_addr;
955 * In zero copy mode, one mbuf can only reference data
956 * for one or partial of one desc buff.
958 mbuf_avail = cpy_len;
960 if (likely(cpy_len > MAX_BATCH_LEN ||
961 copy_nb >= vq->size ||
962 (hdr && cur == m))) {
963 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
965 (void *)((uintptr_t)(desc_addr +
969 batch_copy[copy_nb].dst =
970 rte_pktmbuf_mtod_offset(cur, void *,
972 batch_copy[copy_nb].src =
973 (void *)((uintptr_t)(desc_addr +
975 batch_copy[copy_nb].len = cpy_len;
980 mbuf_avail -= cpy_len;
981 mbuf_offset += cpy_len;
982 desc_avail -= cpy_len;
983 desc_offset += cpy_len;
985 /* This desc reaches to its end, get the next one */
986 if (desc_avail == 0) {
987 if ((desc->flags & VRING_DESC_F_NEXT) == 0)
990 if (unlikely(desc->next >= max_desc ||
991 ++nr_desc > max_desc)) {
995 desc = &descs[desc->next];
996 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
1001 desc_addr = vhost_iova_to_vva(dev,
1005 if (unlikely(!desc_addr)) {
1010 rte_prefetch0((void *)(uintptr_t)desc_addr);
1013 desc_avail = desc->len;
1015 PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
1019 * This mbuf reaches to its end, get a new one
1020 * to hold more data.
1022 if (mbuf_avail == 0) {
1023 cur = rte_pktmbuf_alloc(mbuf_pool);
1024 if (unlikely(cur == NULL)) {
1025 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1026 "allocate memory for mbuf.\n");
1030 if (unlikely(dev->dequeue_zero_copy))
1031 rte_mbuf_refcnt_update(cur, 1);
1034 prev->data_len = mbuf_offset;
1036 m->pkt_len += mbuf_offset;
1040 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1044 prev->data_len = mbuf_offset;
1045 m->pkt_len += mbuf_offset;
1048 vhost_dequeue_offload(hdr, m);
1051 vq->batch_copy_nb_elems = copy_nb;
1056 static __rte_always_inline void
1057 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
1058 uint32_t used_idx, uint32_t desc_idx)
1060 vq->used->ring[used_idx].id = desc_idx;
1061 vq->used->ring[used_idx].len = 0;
1062 vhost_log_used_vring(dev, vq,
1063 offsetof(struct vring_used, ring[used_idx]),
1064 sizeof(vq->used->ring[used_idx]));
1067 static __rte_always_inline void
1068 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
1071 if (unlikely(count == 0))
1077 vq->used->idx += count;
1078 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
1079 sizeof(vq->used->idx));
1081 /* Kick guest if required. */
1082 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
1083 && (vq->callfd >= 0))
1084 eventfd_write(vq->callfd, (eventfd_t)1);
1087 static __rte_always_inline struct zcopy_mbuf *
1088 get_zmbuf(struct vhost_virtqueue *vq)
1094 /* search [last_zmbuf_idx, zmbuf_size) */
1095 i = vq->last_zmbuf_idx;
1096 last = vq->zmbuf_size;
1099 for (; i < last; i++) {
1100 if (vq->zmbufs[i].in_use == 0) {
1101 vq->last_zmbuf_idx = i + 1;
1102 vq->zmbufs[i].in_use = 1;
1103 return &vq->zmbufs[i];
1109 /* search [0, last_zmbuf_idx) */
1111 last = vq->last_zmbuf_idx;
1118 static __rte_always_inline bool
1119 mbuf_is_consumed(struct rte_mbuf *m)
1122 if (rte_mbuf_refcnt_read(m) > 1)
1131 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1132 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1134 struct virtio_net *dev;
1135 struct rte_mbuf *rarp_mbuf = NULL;
1136 struct vhost_virtqueue *vq;
1137 uint32_t desc_indexes[MAX_PKT_BURST];
1140 uint16_t free_entries;
1143 dev = get_device(vid);
1147 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1148 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1149 dev->vid, __func__, queue_id);
1153 vq = dev->virtqueue[queue_id];
1154 if (unlikely(vq->enabled == 0))
1157 vq->batch_copy_nb_elems = 0;
1159 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1160 vhost_user_iotlb_rd_lock(vq);
1162 if (unlikely(vq->access_ok == 0))
1163 if (unlikely(vring_translate(dev, vq) < 0))
1166 if (unlikely(dev->dequeue_zero_copy)) {
1167 struct zcopy_mbuf *zmbuf, *next;
1170 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1171 zmbuf != NULL; zmbuf = next) {
1172 next = TAILQ_NEXT(zmbuf, next);
1174 if (mbuf_is_consumed(zmbuf->mbuf)) {
1175 used_idx = vq->last_used_idx++ & (vq->size - 1);
1176 update_used_ring(dev, vq, used_idx,
1180 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1181 rte_pktmbuf_free(zmbuf->mbuf);
1187 update_used_idx(dev, vq, nr_updated);
1191 * Construct a RARP broadcast packet, and inject it to the "pkts"
1192 * array, to looks like that guest actually send such packet.
1194 * Check user_send_rarp() for more information.
1196 * broadcast_rarp shares a cacheline in the virtio_net structure
1197 * with some fields that are accessed during enqueue and
1198 * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1199 * result in false sharing between enqueue and dequeue.
1201 * Prevent unnecessary false sharing by reading broadcast_rarp first
1202 * and only performing cmpset if the read indicates it is likely to
1206 if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1207 rte_atomic16_cmpset((volatile uint16_t *)
1208 &dev->broadcast_rarp.cnt, 1, 0))) {
1210 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
1211 if (rarp_mbuf == NULL) {
1212 RTE_LOG(ERR, VHOST_DATA,
1213 "Failed to allocate memory for mbuf.\n");
1217 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
1218 rte_pktmbuf_free(rarp_mbuf);
1225 free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1227 if (free_entries == 0)
1230 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1232 /* Prefetch available and used ring */
1233 avail_idx = vq->last_avail_idx & (vq->size - 1);
1234 used_idx = vq->last_used_idx & (vq->size - 1);
1235 rte_prefetch0(&vq->avail->ring[avail_idx]);
1236 rte_prefetch0(&vq->used->ring[used_idx]);
1238 count = RTE_MIN(count, MAX_PKT_BURST);
1239 count = RTE_MIN(count, free_entries);
1240 LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1243 /* Retrieve all of the head indexes first to avoid caching issues. */
1244 for (i = 0; i < count; i++) {
1245 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1246 used_idx = (vq->last_used_idx + i) & (vq->size - 1);
1247 desc_indexes[i] = vq->avail->ring[avail_idx];
1249 if (likely(dev->dequeue_zero_copy == 0))
1250 update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1253 /* Prefetch descriptor index. */
1254 rte_prefetch0(&vq->desc[desc_indexes[0]]);
1255 for (i = 0; i < count; i++) {
1256 struct vring_desc *desc;
1260 if (likely(i + 1 < count))
1261 rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1263 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1264 desc = (struct vring_desc *)(uintptr_t)
1265 vhost_iova_to_vva(dev, vq,
1266 vq->desc[desc_indexes[i]].addr,
1269 if (unlikely(!desc))
1272 rte_prefetch0(desc);
1273 sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1278 idx = desc_indexes[i];
1281 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1282 if (unlikely(pkts[i] == NULL)) {
1283 RTE_LOG(ERR, VHOST_DATA,
1284 "Failed to allocate memory for mbuf.\n");
1288 err = copy_desc_to_mbuf(dev, vq, desc, sz, pkts[i], idx,
1290 if (unlikely(err)) {
1291 rte_pktmbuf_free(pkts[i]);
1295 if (unlikely(dev->dequeue_zero_copy)) {
1296 struct zcopy_mbuf *zmbuf;
1298 zmbuf = get_zmbuf(vq);
1300 rte_pktmbuf_free(pkts[i]);
1303 zmbuf->mbuf = pkts[i];
1304 zmbuf->desc_idx = desc_indexes[i];
1307 * Pin lock the mbuf; we will check later to see
1308 * whether the mbuf is freed (when we are the last
1309 * user) or not. If that's the case, we then could
1310 * update the used ring safely.
1312 rte_mbuf_refcnt_update(pkts[i], 1);
1315 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1318 vq->last_avail_idx += i;
1320 if (likely(dev->dequeue_zero_copy == 0)) {
1321 do_data_copy_dequeue(vq);
1322 vq->last_used_idx += i;
1323 update_used_idx(dev, vq, i);
1327 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1328 vhost_user_iotlb_rd_unlock(vq);
1330 if (unlikely(rarp_mbuf != NULL)) {
1332 * Inject it to the head of "pkts" array, so that switch's mac
1333 * learning table will get updated first.
1335 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1336 pkts[0] = rarp_mbuf;