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
162 } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
163 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
164 net_hdr->gso_size = m_buf->tso_segsz;
165 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
168 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
169 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
170 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
174 static __rte_always_inline int
175 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
176 struct vring_desc *descs, struct rte_mbuf *m,
177 uint16_t desc_idx, uint32_t size)
179 uint32_t desc_avail, desc_offset;
180 uint32_t mbuf_avail, mbuf_offset;
182 struct vring_desc *desc;
184 /* A counter to avoid desc dead loop chain */
185 uint16_t nr_desc = 1;
186 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
187 uint16_t copy_nb = vq->batch_copy_nb_elems;
190 desc = &descs[desc_idx];
191 desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
192 desc->len, VHOST_ACCESS_RW);
194 * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
195 * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
196 * otherwise stores offset on the stack instead of in a register.
198 if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr) {
203 rte_prefetch0((void *)(uintptr_t)desc_addr);
205 virtio_enqueue_offload(m, (struct virtio_net_hdr *)(uintptr_t)desc_addr);
206 vhost_log_write(dev, desc->addr, dev->vhost_hlen);
207 PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
209 desc_offset = dev->vhost_hlen;
210 desc_avail = desc->len - dev->vhost_hlen;
212 mbuf_avail = rte_pktmbuf_data_len(m);
214 while (mbuf_avail != 0 || m->next != NULL) {
215 /* done with current mbuf, fetch next */
216 if (mbuf_avail == 0) {
220 mbuf_avail = rte_pktmbuf_data_len(m);
223 /* done with current desc buf, fetch next */
224 if (desc_avail == 0) {
225 if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
226 /* Room in vring buffer is not enough */
230 if (unlikely(desc->next >= size || ++nr_desc > size)) {
235 desc = &descs[desc->next];
236 desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
239 if (unlikely(!desc_addr)) {
245 desc_avail = desc->len;
248 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
249 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
250 rte_memcpy((void *)((uintptr_t)(desc_addr +
252 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
254 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
255 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
258 batch_copy[copy_nb].dst =
259 (void *)((uintptr_t)(desc_addr + desc_offset));
260 batch_copy[copy_nb].src =
261 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
262 batch_copy[copy_nb].log_addr = desc->addr + desc_offset;
263 batch_copy[copy_nb].len = cpy_len;
267 mbuf_avail -= cpy_len;
268 mbuf_offset += cpy_len;
269 desc_avail -= cpy_len;
270 desc_offset += cpy_len;
274 vq->batch_copy_nb_elems = copy_nb;
280 * This function adds buffers to the virtio devices RX virtqueue. Buffers can
281 * be received from the physical port or from another virtio device. A packet
282 * count is returned to indicate the number of packets that are successfully
283 * added to the RX queue. This function works when the mbuf is scattered, but
284 * it doesn't support the mergeable feature.
286 static __rte_always_inline uint32_t
287 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
288 struct rte_mbuf **pkts, uint32_t count)
290 struct vhost_virtqueue *vq;
291 uint16_t avail_idx, free_entries, start_idx;
292 uint16_t desc_indexes[MAX_PKT_BURST];
293 struct vring_desc *descs;
297 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
298 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
299 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
300 dev->vid, __func__, queue_id);
304 vq = dev->virtqueue[queue_id];
305 if (unlikely(vq->enabled == 0))
308 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
309 vhost_user_iotlb_rd_lock(vq);
311 if (unlikely(vq->access_ok == 0)) {
312 if (unlikely(vring_translate(dev, vq) < 0)) {
318 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
319 start_idx = vq->last_used_idx;
320 free_entries = avail_idx - start_idx;
321 count = RTE_MIN(count, free_entries);
322 count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
326 LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
327 dev->vid, start_idx, start_idx + count);
329 vq->batch_copy_nb_elems = 0;
331 /* Retrieve all of the desc indexes first to avoid caching issues. */
332 rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
333 for (i = 0; i < count; i++) {
334 used_idx = (start_idx + i) & (vq->size - 1);
335 desc_indexes[i] = vq->avail->ring[used_idx];
336 vq->used->ring[used_idx].id = desc_indexes[i];
337 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
339 vhost_log_used_vring(dev, vq,
340 offsetof(struct vring_used, ring[used_idx]),
341 sizeof(vq->used->ring[used_idx]));
344 rte_prefetch0(&vq->desc[desc_indexes[0]]);
345 for (i = 0; i < count; i++) {
346 uint16_t desc_idx = desc_indexes[i];
349 if (vq->desc[desc_idx].flags & VRING_DESC_F_INDIRECT) {
350 descs = (struct vring_desc *)(uintptr_t)
351 vhost_iova_to_vva(dev,
352 vq, vq->desc[desc_idx].addr,
353 vq->desc[desc_idx].len,
355 if (unlikely(!descs)) {
361 sz = vq->desc[desc_idx].len / sizeof(*descs);
367 err = copy_mbuf_to_desc(dev, vq, descs, pkts[i], desc_idx, sz);
374 rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
377 do_data_copy_enqueue(dev, vq);
381 *(volatile uint16_t *)&vq->used->idx += count;
382 vq->last_used_idx += count;
383 vhost_log_used_vring(dev, vq,
384 offsetof(struct vring_used, idx),
385 sizeof(vq->used->idx));
387 vhost_vring_call(vq);
389 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
390 vhost_user_iotlb_rd_unlock(vq);
395 static __rte_always_inline int
396 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
397 uint32_t avail_idx, uint32_t *vec_idx,
398 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
399 uint16_t *desc_chain_len)
401 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
402 uint32_t vec_id = *vec_idx;
404 struct vring_desc *descs = vq->desc;
406 *desc_chain_head = idx;
408 if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
409 descs = (struct vring_desc *)(uintptr_t)
410 vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
413 if (unlikely(!descs))
420 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
423 len += descs[idx].len;
424 buf_vec[vec_id].buf_addr = descs[idx].addr;
425 buf_vec[vec_id].buf_len = descs[idx].len;
426 buf_vec[vec_id].desc_idx = idx;
429 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
432 idx = descs[idx].next;
435 *desc_chain_len = len;
442 * Returns -1 on fail, 0 on success
445 reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
446 uint32_t size, struct buf_vector *buf_vec,
447 uint16_t *num_buffers, uint16_t avail_head)
450 uint32_t vec_idx = 0;
453 uint16_t head_idx = 0;
457 cur_idx = vq->last_avail_idx;
460 if (unlikely(cur_idx == avail_head))
463 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
464 &head_idx, &len) < 0))
466 len = RTE_MIN(len, size);
467 update_shadow_used_ring(vq, head_idx, len);
475 * if we tried all available ring items, and still
476 * can't get enough buf, it means something abnormal
479 if (unlikely(tries >= vq->size))
486 static __rte_always_inline int
487 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
488 struct rte_mbuf *m, struct buf_vector *buf_vec,
489 uint16_t num_buffers)
491 uint32_t vec_idx = 0;
493 uint32_t mbuf_offset, mbuf_avail;
494 uint32_t desc_offset, desc_avail;
496 uint64_t hdr_addr, hdr_phys_addr;
497 struct rte_mbuf *hdr_mbuf;
498 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
499 uint16_t copy_nb = vq->batch_copy_nb_elems;
502 if (unlikely(m == NULL)) {
507 desc_addr = vhost_iova_to_vva(dev, vq, buf_vec[vec_idx].buf_addr,
508 buf_vec[vec_idx].buf_len,
510 if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr) {
516 hdr_addr = desc_addr;
517 hdr_phys_addr = buf_vec[vec_idx].buf_addr;
518 rte_prefetch0((void *)(uintptr_t)hdr_addr);
520 LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
521 dev->vid, num_buffers);
523 desc_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
524 desc_offset = dev->vhost_hlen;
526 mbuf_avail = rte_pktmbuf_data_len(m);
528 while (mbuf_avail != 0 || m->next != NULL) {
529 /* done with current desc buf, get the next one */
530 if (desc_avail == 0) {
533 vhost_iova_to_vva(dev, vq,
534 buf_vec[vec_idx].buf_addr,
535 buf_vec[vec_idx].buf_len,
537 if (unlikely(!desc_addr)) {
542 /* Prefetch buffer address. */
543 rte_prefetch0((void *)(uintptr_t)desc_addr);
545 desc_avail = buf_vec[vec_idx].buf_len;
548 /* done with current mbuf, get the next one */
549 if (mbuf_avail == 0) {
553 mbuf_avail = rte_pktmbuf_data_len(m);
557 struct virtio_net_hdr_mrg_rxbuf *hdr;
559 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)
561 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
562 ASSIGN_UNLESS_EQUAL(hdr->num_buffers, num_buffers);
564 vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
565 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
571 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
573 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
574 rte_memcpy((void *)((uintptr_t)(desc_addr +
576 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
579 buf_vec[vec_idx].buf_addr + desc_offset,
581 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
584 batch_copy[copy_nb].dst =
585 (void *)((uintptr_t)(desc_addr + desc_offset));
586 batch_copy[copy_nb].src =
587 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
588 batch_copy[copy_nb].log_addr =
589 buf_vec[vec_idx].buf_addr + desc_offset;
590 batch_copy[copy_nb].len = cpy_len;
594 mbuf_avail -= cpy_len;
595 mbuf_offset += cpy_len;
596 desc_avail -= cpy_len;
597 desc_offset += cpy_len;
601 vq->batch_copy_nb_elems = copy_nb;
606 static __rte_always_inline uint32_t
607 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
608 struct rte_mbuf **pkts, uint32_t count)
610 struct vhost_virtqueue *vq;
611 uint32_t pkt_idx = 0;
612 uint16_t num_buffers;
613 struct buf_vector buf_vec[BUF_VECTOR_MAX];
616 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
617 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
618 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
619 dev->vid, __func__, queue_id);
623 vq = dev->virtqueue[queue_id];
624 if (unlikely(vq->enabled == 0))
627 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
628 vhost_user_iotlb_rd_lock(vq);
630 if (unlikely(vq->access_ok == 0))
631 if (unlikely(vring_translate(dev, vq) < 0))
634 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
638 vq->batch_copy_nb_elems = 0;
640 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
642 vq->shadow_used_idx = 0;
643 avail_head = *((volatile uint16_t *)&vq->avail->idx);
644 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
645 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
647 if (unlikely(reserve_avail_buf_mergeable(dev, vq,
648 pkt_len, buf_vec, &num_buffers,
650 LOG_DEBUG(VHOST_DATA,
651 "(%d) failed to get enough desc from vring\n",
653 vq->shadow_used_idx -= num_buffers;
657 LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
658 dev->vid, vq->last_avail_idx,
659 vq->last_avail_idx + num_buffers);
661 if (copy_mbuf_to_desc_mergeable(dev, vq, pkts[pkt_idx],
662 buf_vec, num_buffers) < 0) {
663 vq->shadow_used_idx -= num_buffers;
667 vq->last_avail_idx += num_buffers;
670 do_data_copy_enqueue(dev, vq);
672 if (likely(vq->shadow_used_idx)) {
673 flush_shadow_used_ring(dev, vq);
674 vhost_vring_call(vq);
678 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
679 vhost_user_iotlb_rd_unlock(vq);
685 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
686 struct rte_mbuf **pkts, uint16_t count)
688 struct virtio_net *dev = get_device(vid);
693 if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
694 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
696 return virtio_dev_rx(dev, queue_id, pkts, count);
700 virtio_net_with_host_offload(struct virtio_net *dev)
703 ((1ULL << VIRTIO_NET_F_CSUM) |
704 (1ULL << VIRTIO_NET_F_HOST_ECN) |
705 (1ULL << VIRTIO_NET_F_HOST_TSO4) |
706 (1ULL << VIRTIO_NET_F_HOST_TSO6) |
707 (1ULL << VIRTIO_NET_F_HOST_UFO)))
714 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
716 struct ipv4_hdr *ipv4_hdr;
717 struct ipv6_hdr *ipv6_hdr;
719 struct ether_hdr *eth_hdr;
722 eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
724 m->l2_len = sizeof(struct ether_hdr);
725 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
727 if (ethertype == ETHER_TYPE_VLAN) {
728 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
730 m->l2_len += sizeof(struct vlan_hdr);
731 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
734 l3_hdr = (char *)eth_hdr + m->l2_len;
737 case ETHER_TYPE_IPv4:
739 *l4_proto = ipv4_hdr->next_proto_id;
740 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
741 *l4_hdr = (char *)l3_hdr + m->l3_len;
742 m->ol_flags |= PKT_TX_IPV4;
744 case ETHER_TYPE_IPv6:
746 *l4_proto = ipv6_hdr->proto;
747 m->l3_len = sizeof(struct ipv6_hdr);
748 *l4_hdr = (char *)l3_hdr + m->l3_len;
749 m->ol_flags |= PKT_TX_IPV6;
759 static __rte_always_inline void
760 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
762 uint16_t l4_proto = 0;
764 struct tcp_hdr *tcp_hdr = NULL;
766 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
769 parse_ethernet(m, &l4_proto, &l4_hdr);
770 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
771 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
772 switch (hdr->csum_offset) {
773 case (offsetof(struct tcp_hdr, cksum)):
774 if (l4_proto == IPPROTO_TCP)
775 m->ol_flags |= PKT_TX_TCP_CKSUM;
777 case (offsetof(struct udp_hdr, dgram_cksum)):
778 if (l4_proto == IPPROTO_UDP)
779 m->ol_flags |= PKT_TX_UDP_CKSUM;
781 case (offsetof(struct sctp_hdr, cksum)):
782 if (l4_proto == IPPROTO_SCTP)
783 m->ol_flags |= PKT_TX_SCTP_CKSUM;
791 if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
792 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
793 case VIRTIO_NET_HDR_GSO_TCPV4:
794 case VIRTIO_NET_HDR_GSO_TCPV6:
796 m->ol_flags |= PKT_TX_TCP_SEG;
797 m->tso_segsz = hdr->gso_size;
798 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
800 case VIRTIO_NET_HDR_GSO_UDP:
801 m->ol_flags |= PKT_TX_UDP_SEG;
802 m->tso_segsz = hdr->gso_size;
803 m->l4_len = sizeof(struct udp_hdr);
806 RTE_LOG(WARNING, VHOST_DATA,
807 "unsupported gso type %u.\n", hdr->gso_type);
813 #define RARP_PKT_SIZE 64
816 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
818 struct ether_hdr *eth_hdr;
819 struct arp_hdr *rarp;
821 if (rarp_mbuf->buf_len < 64) {
822 RTE_LOG(WARNING, VHOST_DATA,
823 "failed to make RARP; mbuf size too small %u (< %d)\n",
824 rarp_mbuf->buf_len, RARP_PKT_SIZE);
828 /* Ethernet header. */
829 eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
830 memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
831 ether_addr_copy(mac, ð_hdr->s_addr);
832 eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
835 rarp = (struct arp_hdr *)(eth_hdr + 1);
836 rarp->arp_hrd = htons(ARP_HRD_ETHER);
837 rarp->arp_pro = htons(ETHER_TYPE_IPv4);
838 rarp->arp_hln = ETHER_ADDR_LEN;
840 rarp->arp_op = htons(ARP_OP_REVREQUEST);
842 ether_addr_copy(mac, &rarp->arp_data.arp_sha);
843 ether_addr_copy(mac, &rarp->arp_data.arp_tha);
844 memset(&rarp->arp_data.arp_sip, 0x00, 4);
845 memset(&rarp->arp_data.arp_tip, 0x00, 4);
847 rarp_mbuf->pkt_len = rarp_mbuf->data_len = RARP_PKT_SIZE;
852 static __rte_always_inline void
853 put_zmbuf(struct zcopy_mbuf *zmbuf)
858 static __rte_always_inline int
859 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
860 struct vring_desc *descs, uint16_t max_desc,
861 struct rte_mbuf *m, uint16_t desc_idx,
862 struct rte_mempool *mbuf_pool)
864 struct vring_desc *desc;
866 uint32_t desc_avail, desc_offset;
867 uint32_t mbuf_avail, mbuf_offset;
869 struct rte_mbuf *cur = m, *prev = m;
870 struct virtio_net_hdr *hdr = NULL;
871 /* A counter to avoid desc dead loop chain */
872 uint32_t nr_desc = 1;
873 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
874 uint16_t copy_nb = vq->batch_copy_nb_elems;
877 desc = &descs[desc_idx];
878 if (unlikely((desc->len < dev->vhost_hlen)) ||
879 (desc->flags & VRING_DESC_F_INDIRECT)) {
884 desc_addr = vhost_iova_to_vva(dev,
888 if (unlikely(!desc_addr)) {
893 if (virtio_net_with_host_offload(dev)) {
894 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
899 * A virtio driver normally uses at least 2 desc buffers
900 * for Tx: the first for storing the header, and others
901 * for storing the data.
903 if (likely((desc->len == dev->vhost_hlen) &&
904 (desc->flags & VRING_DESC_F_NEXT) != 0)) {
905 desc = &descs[desc->next];
906 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
911 desc_addr = vhost_iova_to_vva(dev,
915 if (unlikely(!desc_addr)) {
921 desc_avail = desc->len;
924 desc_avail = desc->len - dev->vhost_hlen;
925 desc_offset = dev->vhost_hlen;
928 rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
930 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
933 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
937 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
940 * A desc buf might across two host physical pages that are
941 * not continuous. In such case (gpa_to_hpa returns 0), data
942 * will be copied even though zero copy is enabled.
944 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
945 desc->addr + desc_offset, cpy_len)))) {
946 cur->data_len = cpy_len;
948 cur->buf_addr = (void *)(uintptr_t)(desc_addr
953 * In zero copy mode, one mbuf can only reference data
954 * for one or partial of one desc buff.
956 mbuf_avail = cpy_len;
958 if (likely(cpy_len > MAX_BATCH_LEN ||
959 copy_nb >= vq->size ||
960 (hdr && cur == m))) {
961 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
963 (void *)((uintptr_t)(desc_addr +
967 batch_copy[copy_nb].dst =
968 rte_pktmbuf_mtod_offset(cur, void *,
970 batch_copy[copy_nb].src =
971 (void *)((uintptr_t)(desc_addr +
973 batch_copy[copy_nb].len = cpy_len;
978 mbuf_avail -= cpy_len;
979 mbuf_offset += cpy_len;
980 desc_avail -= cpy_len;
981 desc_offset += cpy_len;
983 /* This desc reaches to its end, get the next one */
984 if (desc_avail == 0) {
985 if ((desc->flags & VRING_DESC_F_NEXT) == 0)
988 if (unlikely(desc->next >= max_desc ||
989 ++nr_desc > max_desc)) {
993 desc = &descs[desc->next];
994 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
999 desc_addr = vhost_iova_to_vva(dev,
1003 if (unlikely(!desc_addr)) {
1008 rte_prefetch0((void *)(uintptr_t)desc_addr);
1011 desc_avail = desc->len;
1013 PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
1017 * This mbuf reaches to its end, get a new one
1018 * to hold more data.
1020 if (mbuf_avail == 0) {
1021 cur = rte_pktmbuf_alloc(mbuf_pool);
1022 if (unlikely(cur == NULL)) {
1023 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1024 "allocate memory for mbuf.\n");
1028 if (unlikely(dev->dequeue_zero_copy))
1029 rte_mbuf_refcnt_update(cur, 1);
1032 prev->data_len = mbuf_offset;
1034 m->pkt_len += mbuf_offset;
1038 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1042 prev->data_len = mbuf_offset;
1043 m->pkt_len += mbuf_offset;
1046 vhost_dequeue_offload(hdr, m);
1049 vq->batch_copy_nb_elems = copy_nb;
1054 static __rte_always_inline void
1055 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
1056 uint32_t used_idx, uint32_t desc_idx)
1058 vq->used->ring[used_idx].id = desc_idx;
1059 vq->used->ring[used_idx].len = 0;
1060 vhost_log_used_vring(dev, vq,
1061 offsetof(struct vring_used, ring[used_idx]),
1062 sizeof(vq->used->ring[used_idx]));
1065 static __rte_always_inline void
1066 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
1069 if (unlikely(count == 0))
1075 vq->used->idx += count;
1076 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
1077 sizeof(vq->used->idx));
1078 vhost_vring_call(vq);
1081 static __rte_always_inline struct zcopy_mbuf *
1082 get_zmbuf(struct vhost_virtqueue *vq)
1088 /* search [last_zmbuf_idx, zmbuf_size) */
1089 i = vq->last_zmbuf_idx;
1090 last = vq->zmbuf_size;
1093 for (; i < last; i++) {
1094 if (vq->zmbufs[i].in_use == 0) {
1095 vq->last_zmbuf_idx = i + 1;
1096 vq->zmbufs[i].in_use = 1;
1097 return &vq->zmbufs[i];
1103 /* search [0, last_zmbuf_idx) */
1105 last = vq->last_zmbuf_idx;
1112 static __rte_always_inline bool
1113 mbuf_is_consumed(struct rte_mbuf *m)
1116 if (rte_mbuf_refcnt_read(m) > 1)
1125 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1126 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1128 struct virtio_net *dev;
1129 struct rte_mbuf *rarp_mbuf = NULL;
1130 struct vhost_virtqueue *vq;
1131 uint32_t desc_indexes[MAX_PKT_BURST];
1134 uint16_t free_entries;
1137 dev = get_device(vid);
1141 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1142 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1143 dev->vid, __func__, queue_id);
1147 vq = dev->virtqueue[queue_id];
1148 if (unlikely(vq->enabled == 0))
1151 vq->batch_copy_nb_elems = 0;
1153 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1154 vhost_user_iotlb_rd_lock(vq);
1156 if (unlikely(vq->access_ok == 0))
1157 if (unlikely(vring_translate(dev, vq) < 0))
1160 if (unlikely(dev->dequeue_zero_copy)) {
1161 struct zcopy_mbuf *zmbuf, *next;
1164 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1165 zmbuf != NULL; zmbuf = next) {
1166 next = TAILQ_NEXT(zmbuf, next);
1168 if (mbuf_is_consumed(zmbuf->mbuf)) {
1169 used_idx = vq->last_used_idx++ & (vq->size - 1);
1170 update_used_ring(dev, vq, used_idx,
1174 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1175 rte_pktmbuf_free(zmbuf->mbuf);
1181 update_used_idx(dev, vq, nr_updated);
1185 * Construct a RARP broadcast packet, and inject it to the "pkts"
1186 * array, to looks like that guest actually send such packet.
1188 * Check user_send_rarp() for more information.
1190 * broadcast_rarp shares a cacheline in the virtio_net structure
1191 * with some fields that are accessed during enqueue and
1192 * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1193 * result in false sharing between enqueue and dequeue.
1195 * Prevent unnecessary false sharing by reading broadcast_rarp first
1196 * and only performing cmpset if the read indicates it is likely to
1200 if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1201 rte_atomic16_cmpset((volatile uint16_t *)
1202 &dev->broadcast_rarp.cnt, 1, 0))) {
1204 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
1205 if (rarp_mbuf == NULL) {
1206 RTE_LOG(ERR, VHOST_DATA,
1207 "Failed to allocate memory for mbuf.\n");
1211 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
1212 rte_pktmbuf_free(rarp_mbuf);
1219 free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1221 if (free_entries == 0)
1224 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1226 /* Prefetch available and used ring */
1227 avail_idx = vq->last_avail_idx & (vq->size - 1);
1228 used_idx = vq->last_used_idx & (vq->size - 1);
1229 rte_prefetch0(&vq->avail->ring[avail_idx]);
1230 rte_prefetch0(&vq->used->ring[used_idx]);
1232 count = RTE_MIN(count, MAX_PKT_BURST);
1233 count = RTE_MIN(count, free_entries);
1234 LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1237 /* Retrieve all of the head indexes first to avoid caching issues. */
1238 for (i = 0; i < count; i++) {
1239 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1240 used_idx = (vq->last_used_idx + i) & (vq->size - 1);
1241 desc_indexes[i] = vq->avail->ring[avail_idx];
1243 if (likely(dev->dequeue_zero_copy == 0))
1244 update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1247 /* Prefetch descriptor index. */
1248 rte_prefetch0(&vq->desc[desc_indexes[0]]);
1249 for (i = 0; i < count; i++) {
1250 struct vring_desc *desc;
1254 if (likely(i + 1 < count))
1255 rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1257 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1258 desc = (struct vring_desc *)(uintptr_t)
1259 vhost_iova_to_vva(dev, vq,
1260 vq->desc[desc_indexes[i]].addr,
1263 if (unlikely(!desc))
1266 rte_prefetch0(desc);
1267 sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1272 idx = desc_indexes[i];
1275 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1276 if (unlikely(pkts[i] == NULL)) {
1277 RTE_LOG(ERR, VHOST_DATA,
1278 "Failed to allocate memory for mbuf.\n");
1282 err = copy_desc_to_mbuf(dev, vq, desc, sz, pkts[i], idx,
1284 if (unlikely(err)) {
1285 rte_pktmbuf_free(pkts[i]);
1289 if (unlikely(dev->dequeue_zero_copy)) {
1290 struct zcopy_mbuf *zmbuf;
1292 zmbuf = get_zmbuf(vq);
1294 rte_pktmbuf_free(pkts[i]);
1297 zmbuf->mbuf = pkts[i];
1298 zmbuf->desc_idx = desc_indexes[i];
1301 * Pin lock the mbuf; we will check later to see
1302 * whether the mbuf is freed (when we are the last
1303 * user) or not. If that's the case, we then could
1304 * update the used ring safely.
1306 rte_mbuf_refcnt_update(pkts[i], 1);
1309 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1312 vq->last_avail_idx += i;
1314 if (likely(dev->dequeue_zero_copy == 0)) {
1315 do_data_copy_dequeue(vq);
1316 vq->last_used_idx += i;
1317 update_used_idx(dev, vq, i);
1321 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1322 vhost_user_iotlb_rd_unlock(vq);
1324 if (unlikely(rarp_mbuf != NULL)) {
1326 * Inject it to the head of "pkts" array, so that switch's mac
1327 * learning table will get updated first.
1329 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1330 pkts[0] = rarp_mbuf;