4 * Copyright(c) 2010-2016 Intel Corporation. All rights reserved.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * * Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * * Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * * Neither the name of Intel Corporation nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
36 #include <linux/virtio_net.h>
39 #include <rte_memcpy.h>
40 #include <rte_ether.h>
42 #include <rte_vhost.h>
50 #define MAX_PKT_BURST 32
53 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
55 return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
58 static __rte_always_inline void
59 do_flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
60 uint16_t to, uint16_t from, uint16_t size)
62 rte_memcpy(&vq->used->ring[to],
63 &vq->shadow_used_ring[from],
64 size * sizeof(struct vring_used_elem));
65 vhost_log_used_vring(dev, vq,
66 offsetof(struct vring_used, ring[to]),
67 size * sizeof(struct vring_used_elem));
70 static __rte_always_inline void
71 flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq)
73 uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
75 if (used_idx + vq->shadow_used_idx <= vq->size) {
76 do_flush_shadow_used_ring(dev, vq, used_idx, 0,
81 /* update used ring interval [used_idx, vq->size] */
82 size = vq->size - used_idx;
83 do_flush_shadow_used_ring(dev, vq, used_idx, 0, size);
85 /* update the left half used ring interval [0, left_size] */
86 do_flush_shadow_used_ring(dev, vq, 0, size,
87 vq->shadow_used_idx - size);
89 vq->last_used_idx += vq->shadow_used_idx;
93 *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
94 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
95 sizeof(vq->used->idx));
98 static __rte_always_inline void
99 update_shadow_used_ring(struct vhost_virtqueue *vq,
100 uint16_t desc_idx, uint16_t len)
102 uint16_t i = vq->shadow_used_idx++;
104 vq->shadow_used_ring[i].id = desc_idx;
105 vq->shadow_used_ring[i].len = 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
164 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
165 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
166 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
170 static __rte_always_inline int
171 copy_mbuf_to_desc(struct virtio_net *dev, struct vring_desc *descs,
172 struct rte_mbuf *m, 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;
182 desc = &descs[desc_idx];
183 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
185 * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
186 * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
187 * otherwise stores offset on the stack instead of in a register.
189 if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr)
192 rte_prefetch0((void *)(uintptr_t)desc_addr);
194 virtio_enqueue_offload(m, (struct virtio_net_hdr *)(uintptr_t)desc_addr);
195 vhost_log_write(dev, desc->addr, dev->vhost_hlen);
196 PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
198 desc_offset = dev->vhost_hlen;
199 desc_avail = desc->len - dev->vhost_hlen;
201 mbuf_avail = rte_pktmbuf_data_len(m);
203 while (mbuf_avail != 0 || m->next != NULL) {
204 /* done with current mbuf, fetch next */
205 if (mbuf_avail == 0) {
209 mbuf_avail = rte_pktmbuf_data_len(m);
212 /* done with current desc buf, fetch next */
213 if (desc_avail == 0) {
214 if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
215 /* Room in vring buffer is not enough */
218 if (unlikely(desc->next >= size || ++nr_desc > size))
221 desc = &descs[desc->next];
222 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
223 if (unlikely(!desc_addr))
227 desc_avail = desc->len;
230 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
231 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
232 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
234 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
235 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
238 mbuf_avail -= cpy_len;
239 mbuf_offset += cpy_len;
240 desc_avail -= cpy_len;
241 desc_offset += cpy_len;
248 * This function adds buffers to the virtio devices RX virtqueue. Buffers can
249 * be received from the physical port or from another virtio device. A packet
250 * count is returned to indicate the number of packets that are successfully
251 * added to the RX queue. This function works when the mbuf is scattered, but
252 * it doesn't support the mergeable feature.
254 static __rte_always_inline uint32_t
255 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
256 struct rte_mbuf **pkts, uint32_t count)
258 struct vhost_virtqueue *vq;
259 uint16_t avail_idx, free_entries, start_idx;
260 uint16_t desc_indexes[MAX_PKT_BURST];
261 struct vring_desc *descs;
265 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
266 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
267 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
268 dev->vid, __func__, queue_id);
272 vq = dev->virtqueue[queue_id];
273 if (unlikely(vq->enabled == 0))
276 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
277 start_idx = vq->last_used_idx;
278 free_entries = avail_idx - start_idx;
279 count = RTE_MIN(count, free_entries);
280 count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
284 LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
285 dev->vid, start_idx, start_idx + count);
287 /* Retrieve all of the desc indexes first to avoid caching issues. */
288 rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
289 for (i = 0; i < count; i++) {
290 used_idx = (start_idx + i) & (vq->size - 1);
291 desc_indexes[i] = vq->avail->ring[used_idx];
292 vq->used->ring[used_idx].id = desc_indexes[i];
293 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
295 vhost_log_used_vring(dev, vq,
296 offsetof(struct vring_used, ring[used_idx]),
297 sizeof(vq->used->ring[used_idx]));
300 rte_prefetch0(&vq->desc[desc_indexes[0]]);
301 for (i = 0; i < count; i++) {
302 uint16_t desc_idx = desc_indexes[i];
305 if (vq->desc[desc_idx].flags & VRING_DESC_F_INDIRECT) {
306 descs = (struct vring_desc *)(uintptr_t)
307 rte_vhost_gpa_to_vva(dev->mem,
308 vq->desc[desc_idx].addr);
309 if (unlikely(!descs)) {
315 sz = vq->desc[desc_idx].len / sizeof(*descs);
321 err = copy_mbuf_to_desc(dev, descs, pkts[i], desc_idx, sz);
323 used_idx = (start_idx + i) & (vq->size - 1);
324 vq->used->ring[used_idx].len = dev->vhost_hlen;
325 vhost_log_used_vring(dev, vq,
326 offsetof(struct vring_used, ring[used_idx]),
327 sizeof(vq->used->ring[used_idx]));
331 rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
336 *(volatile uint16_t *)&vq->used->idx += count;
337 vq->last_used_idx += count;
338 vhost_log_used_vring(dev, vq,
339 offsetof(struct vring_used, idx),
340 sizeof(vq->used->idx));
342 /* flush used->idx update before we read avail->flags. */
345 /* Kick the guest if necessary. */
346 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
347 && (vq->callfd >= 0))
348 eventfd_write(vq->callfd, (eventfd_t)1);
352 static __rte_always_inline int
353 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
354 uint32_t avail_idx, uint32_t *vec_idx,
355 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
356 uint16_t *desc_chain_len)
358 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
359 uint32_t vec_id = *vec_idx;
361 struct vring_desc *descs = vq->desc;
363 *desc_chain_head = idx;
365 if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
366 descs = (struct vring_desc *)(uintptr_t)
367 rte_vhost_gpa_to_vva(dev->mem, vq->desc[idx].addr);
368 if (unlikely(!descs))
375 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
378 len += descs[idx].len;
379 buf_vec[vec_id].buf_addr = descs[idx].addr;
380 buf_vec[vec_id].buf_len = descs[idx].len;
381 buf_vec[vec_id].desc_idx = idx;
384 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
387 idx = descs[idx].next;
390 *desc_chain_len = len;
397 * Returns -1 on fail, 0 on success
400 reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
401 uint32_t size, struct buf_vector *buf_vec,
402 uint16_t *num_buffers, uint16_t avail_head)
405 uint32_t vec_idx = 0;
408 uint16_t head_idx = 0;
412 cur_idx = vq->last_avail_idx;
415 if (unlikely(cur_idx == avail_head))
418 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
419 &head_idx, &len) < 0))
421 len = RTE_MIN(len, size);
422 update_shadow_used_ring(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 >= vq->size))
441 static __rte_always_inline int
442 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct rte_mbuf *m,
443 struct buf_vector *buf_vec, uint16_t num_buffers)
445 uint32_t vec_idx = 0;
447 uint32_t mbuf_offset, mbuf_avail;
448 uint32_t desc_offset, desc_avail;
450 uint64_t hdr_addr, hdr_phys_addr;
451 struct rte_mbuf *hdr_mbuf;
453 if (unlikely(m == NULL))
456 desc_addr = rte_vhost_gpa_to_vva(dev->mem, buf_vec[vec_idx].buf_addr);
457 if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr)
461 hdr_addr = desc_addr;
462 hdr_phys_addr = buf_vec[vec_idx].buf_addr;
463 rte_prefetch0((void *)(uintptr_t)hdr_addr);
465 LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
466 dev->vid, num_buffers);
468 desc_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
469 desc_offset = dev->vhost_hlen;
471 mbuf_avail = rte_pktmbuf_data_len(m);
473 while (mbuf_avail != 0 || m->next != NULL) {
474 /* done with current desc buf, get the next one */
475 if (desc_avail == 0) {
477 desc_addr = rte_vhost_gpa_to_vva(dev->mem,
478 buf_vec[vec_idx].buf_addr);
479 if (unlikely(!desc_addr))
482 /* Prefetch buffer address. */
483 rte_prefetch0((void *)(uintptr_t)desc_addr);
485 desc_avail = buf_vec[vec_idx].buf_len;
488 /* done with current mbuf, get the next one */
489 if (mbuf_avail == 0) {
493 mbuf_avail = rte_pktmbuf_data_len(m);
497 struct virtio_net_hdr_mrg_rxbuf *hdr;
499 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)
501 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
502 ASSIGN_UNLESS_EQUAL(hdr->num_buffers, num_buffers);
504 vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
505 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
511 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
512 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
513 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
515 vhost_log_write(dev, buf_vec[vec_idx].buf_addr + desc_offset,
517 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
520 mbuf_avail -= cpy_len;
521 mbuf_offset += cpy_len;
522 desc_avail -= cpy_len;
523 desc_offset += cpy_len;
529 static __rte_always_inline uint32_t
530 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
531 struct rte_mbuf **pkts, uint32_t count)
533 struct vhost_virtqueue *vq;
534 uint32_t pkt_idx = 0;
535 uint16_t num_buffers;
536 struct buf_vector buf_vec[BUF_VECTOR_MAX];
539 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
540 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
541 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
542 dev->vid, __func__, queue_id);
546 vq = dev->virtqueue[queue_id];
547 if (unlikely(vq->enabled == 0))
550 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
554 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
556 vq->shadow_used_idx = 0;
557 avail_head = *((volatile uint16_t *)&vq->avail->idx);
558 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
559 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
561 if (unlikely(reserve_avail_buf_mergeable(dev, vq,
562 pkt_len, buf_vec, &num_buffers,
564 LOG_DEBUG(VHOST_DATA,
565 "(%d) failed to get enough desc from vring\n",
567 vq->shadow_used_idx -= num_buffers;
571 LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
572 dev->vid, vq->last_avail_idx,
573 vq->last_avail_idx + num_buffers);
575 if (copy_mbuf_to_desc_mergeable(dev, pkts[pkt_idx],
576 buf_vec, num_buffers) < 0) {
577 vq->shadow_used_idx -= num_buffers;
581 vq->last_avail_idx += num_buffers;
584 if (likely(vq->shadow_used_idx)) {
585 flush_shadow_used_ring(dev, vq);
587 /* flush used->idx update before we read avail->flags. */
590 /* Kick the guest if necessary. */
591 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
592 && (vq->callfd >= 0))
593 eventfd_write(vq->callfd, (eventfd_t)1);
600 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
601 struct rte_mbuf **pkts, uint16_t count)
603 struct virtio_net *dev = get_device(vid);
608 if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
609 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
611 return virtio_dev_rx(dev, queue_id, pkts, count);
615 virtio_net_with_host_offload(struct virtio_net *dev)
618 ((1ULL << VIRTIO_NET_F_CSUM) |
619 (1ULL << VIRTIO_NET_F_HOST_ECN) |
620 (1ULL << VIRTIO_NET_F_HOST_TSO4) |
621 (1ULL << VIRTIO_NET_F_HOST_TSO6) |
622 (1ULL << VIRTIO_NET_F_HOST_UFO)))
629 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
631 struct ipv4_hdr *ipv4_hdr;
632 struct ipv6_hdr *ipv6_hdr;
634 struct ether_hdr *eth_hdr;
637 eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
639 m->l2_len = sizeof(struct ether_hdr);
640 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
642 if (ethertype == ETHER_TYPE_VLAN) {
643 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
645 m->l2_len += sizeof(struct vlan_hdr);
646 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
649 l3_hdr = (char *)eth_hdr + m->l2_len;
652 case ETHER_TYPE_IPv4:
654 *l4_proto = ipv4_hdr->next_proto_id;
655 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
656 *l4_hdr = (char *)l3_hdr + m->l3_len;
657 m->ol_flags |= PKT_TX_IPV4;
659 case ETHER_TYPE_IPv6:
661 *l4_proto = ipv6_hdr->proto;
662 m->l3_len = sizeof(struct ipv6_hdr);
663 *l4_hdr = (char *)l3_hdr + m->l3_len;
664 m->ol_flags |= PKT_TX_IPV6;
674 static __rte_always_inline void
675 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
677 uint16_t l4_proto = 0;
679 struct tcp_hdr *tcp_hdr = NULL;
681 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
684 parse_ethernet(m, &l4_proto, &l4_hdr);
685 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
686 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
687 switch (hdr->csum_offset) {
688 case (offsetof(struct tcp_hdr, cksum)):
689 if (l4_proto == IPPROTO_TCP)
690 m->ol_flags |= PKT_TX_TCP_CKSUM;
692 case (offsetof(struct udp_hdr, dgram_cksum)):
693 if (l4_proto == IPPROTO_UDP)
694 m->ol_flags |= PKT_TX_UDP_CKSUM;
696 case (offsetof(struct sctp_hdr, cksum)):
697 if (l4_proto == IPPROTO_SCTP)
698 m->ol_flags |= PKT_TX_SCTP_CKSUM;
706 if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
707 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
708 case VIRTIO_NET_HDR_GSO_TCPV4:
709 case VIRTIO_NET_HDR_GSO_TCPV6:
711 m->ol_flags |= PKT_TX_TCP_SEG;
712 m->tso_segsz = hdr->gso_size;
713 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
716 RTE_LOG(WARNING, VHOST_DATA,
717 "unsupported gso type %u.\n", hdr->gso_type);
723 #define RARP_PKT_SIZE 64
726 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
728 struct ether_hdr *eth_hdr;
729 struct arp_hdr *rarp;
731 if (rarp_mbuf->buf_len < 64) {
732 RTE_LOG(WARNING, VHOST_DATA,
733 "failed to make RARP; mbuf size too small %u (< %d)\n",
734 rarp_mbuf->buf_len, RARP_PKT_SIZE);
738 /* Ethernet header. */
739 eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
740 memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
741 ether_addr_copy(mac, ð_hdr->s_addr);
742 eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
745 rarp = (struct arp_hdr *)(eth_hdr + 1);
746 rarp->arp_hrd = htons(ARP_HRD_ETHER);
747 rarp->arp_pro = htons(ETHER_TYPE_IPv4);
748 rarp->arp_hln = ETHER_ADDR_LEN;
750 rarp->arp_op = htons(ARP_OP_REVREQUEST);
752 ether_addr_copy(mac, &rarp->arp_data.arp_sha);
753 ether_addr_copy(mac, &rarp->arp_data.arp_tha);
754 memset(&rarp->arp_data.arp_sip, 0x00, 4);
755 memset(&rarp->arp_data.arp_tip, 0x00, 4);
757 rarp_mbuf->pkt_len = rarp_mbuf->data_len = RARP_PKT_SIZE;
762 static __rte_always_inline void
763 put_zmbuf(struct zcopy_mbuf *zmbuf)
768 static __rte_always_inline int
769 copy_desc_to_mbuf(struct virtio_net *dev, struct vring_desc *descs,
770 uint16_t max_desc, struct rte_mbuf *m, uint16_t desc_idx,
771 struct rte_mempool *mbuf_pool)
773 struct vring_desc *desc;
775 uint32_t desc_avail, desc_offset;
776 uint32_t mbuf_avail, mbuf_offset;
778 struct rte_mbuf *cur = m, *prev = m;
779 struct virtio_net_hdr *hdr = NULL;
780 /* A counter to avoid desc dead loop chain */
781 uint32_t nr_desc = 1;
783 desc = &descs[desc_idx];
784 if (unlikely((desc->len < dev->vhost_hlen)) ||
785 (desc->flags & VRING_DESC_F_INDIRECT))
788 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
789 if (unlikely(!desc_addr))
792 if (virtio_net_with_host_offload(dev)) {
793 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
798 * A virtio driver normally uses at least 2 desc buffers
799 * for Tx: the first for storing the header, and others
800 * for storing the data.
802 if (likely((desc->len == dev->vhost_hlen) &&
803 (desc->flags & VRING_DESC_F_NEXT) != 0)) {
804 desc = &descs[desc->next];
805 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
808 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
809 if (unlikely(!desc_addr))
813 desc_avail = desc->len;
816 desc_avail = desc->len - dev->vhost_hlen;
817 desc_offset = dev->vhost_hlen;
820 rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
822 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
825 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
829 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
832 * A desc buf might across two host physical pages that are
833 * not continuous. In such case (gpa_to_hpa returns 0), data
834 * will be copied even though zero copy is enabled.
836 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
837 desc->addr + desc_offset, cpy_len)))) {
838 cur->data_len = cpy_len;
840 cur->buf_addr = (void *)(uintptr_t)desc_addr;
841 cur->buf_physaddr = hpa;
844 * In zero copy mode, one mbuf can only reference data
845 * for one or partial of one desc buff.
847 mbuf_avail = cpy_len;
849 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
851 (void *)((uintptr_t)(desc_addr + desc_offset)),
855 mbuf_avail -= cpy_len;
856 mbuf_offset += cpy_len;
857 desc_avail -= cpy_len;
858 desc_offset += cpy_len;
860 /* This desc reaches to its end, get the next one */
861 if (desc_avail == 0) {
862 if ((desc->flags & VRING_DESC_F_NEXT) == 0)
865 if (unlikely(desc->next >= max_desc ||
866 ++nr_desc > max_desc))
868 desc = &descs[desc->next];
869 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
872 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
873 if (unlikely(!desc_addr))
876 rte_prefetch0((void *)(uintptr_t)desc_addr);
879 desc_avail = desc->len;
881 PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
885 * This mbuf reaches to its end, get a new one
888 if (mbuf_avail == 0) {
889 cur = rte_pktmbuf_alloc(mbuf_pool);
890 if (unlikely(cur == NULL)) {
891 RTE_LOG(ERR, VHOST_DATA, "Failed to "
892 "allocate memory for mbuf.\n");
895 if (unlikely(dev->dequeue_zero_copy))
896 rte_mbuf_refcnt_update(cur, 1);
899 prev->data_len = mbuf_offset;
901 m->pkt_len += mbuf_offset;
905 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
909 prev->data_len = mbuf_offset;
910 m->pkt_len += mbuf_offset;
913 vhost_dequeue_offload(hdr, m);
918 static __rte_always_inline void
919 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
920 uint32_t used_idx, uint32_t desc_idx)
922 vq->used->ring[used_idx].id = desc_idx;
923 vq->used->ring[used_idx].len = 0;
924 vhost_log_used_vring(dev, vq,
925 offsetof(struct vring_used, ring[used_idx]),
926 sizeof(vq->used->ring[used_idx]));
929 static __rte_always_inline void
930 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
933 if (unlikely(count == 0))
939 vq->used->idx += count;
940 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
941 sizeof(vq->used->idx));
943 /* Kick guest if required. */
944 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
945 && (vq->callfd >= 0))
946 eventfd_write(vq->callfd, (eventfd_t)1);
949 static __rte_always_inline struct zcopy_mbuf *
950 get_zmbuf(struct vhost_virtqueue *vq)
956 /* search [last_zmbuf_idx, zmbuf_size) */
957 i = vq->last_zmbuf_idx;
958 last = vq->zmbuf_size;
961 for (; i < last; i++) {
962 if (vq->zmbufs[i].in_use == 0) {
963 vq->last_zmbuf_idx = i + 1;
964 vq->zmbufs[i].in_use = 1;
965 return &vq->zmbufs[i];
971 /* search [0, last_zmbuf_idx) */
973 last = vq->last_zmbuf_idx;
980 static __rte_always_inline bool
981 mbuf_is_consumed(struct rte_mbuf *m)
984 if (rte_mbuf_refcnt_read(m) > 1)
993 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
994 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
996 struct virtio_net *dev;
997 struct rte_mbuf *rarp_mbuf = NULL;
998 struct vhost_virtqueue *vq;
999 uint32_t desc_indexes[MAX_PKT_BURST];
1002 uint16_t free_entries;
1005 dev = get_device(vid);
1009 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1010 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1011 dev->vid, __func__, queue_id);
1015 vq = dev->virtqueue[queue_id];
1016 if (unlikely(vq->enabled == 0))
1019 if (unlikely(dev->dequeue_zero_copy)) {
1020 struct zcopy_mbuf *zmbuf, *next;
1023 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1024 zmbuf != NULL; zmbuf = next) {
1025 next = TAILQ_NEXT(zmbuf, next);
1027 if (mbuf_is_consumed(zmbuf->mbuf)) {
1028 used_idx = vq->last_used_idx++ & (vq->size - 1);
1029 update_used_ring(dev, vq, used_idx,
1033 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1034 rte_pktmbuf_free(zmbuf->mbuf);
1040 update_used_idx(dev, vq, nr_updated);
1044 * Construct a RARP broadcast packet, and inject it to the "pkts"
1045 * array, to looks like that guest actually send such packet.
1047 * Check user_send_rarp() for more information.
1049 * broadcast_rarp shares a cacheline in the virtio_net structure
1050 * with some fields that are accessed during enqueue and
1051 * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1052 * result in false sharing between enqueue and dequeue.
1054 * Prevent unnecessary false sharing by reading broadcast_rarp first
1055 * and only performing cmpset if the read indicates it is likely to
1059 if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1060 rte_atomic16_cmpset((volatile uint16_t *)
1061 &dev->broadcast_rarp.cnt, 1, 0))) {
1063 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
1064 if (rarp_mbuf == NULL) {
1065 RTE_LOG(ERR, VHOST_DATA,
1066 "Failed to allocate memory for mbuf.\n");
1070 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
1071 rte_pktmbuf_free(rarp_mbuf);
1078 free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1080 if (free_entries == 0)
1083 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1085 /* Prefetch available and used ring */
1086 avail_idx = vq->last_avail_idx & (vq->size - 1);
1087 used_idx = vq->last_used_idx & (vq->size - 1);
1088 rte_prefetch0(&vq->avail->ring[avail_idx]);
1089 rte_prefetch0(&vq->used->ring[used_idx]);
1091 count = RTE_MIN(count, MAX_PKT_BURST);
1092 count = RTE_MIN(count, free_entries);
1093 LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1096 /* Retrieve all of the head indexes first to avoid caching issues. */
1097 for (i = 0; i < count; i++) {
1098 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1099 used_idx = (vq->last_used_idx + i) & (vq->size - 1);
1100 desc_indexes[i] = vq->avail->ring[avail_idx];
1102 if (likely(dev->dequeue_zero_copy == 0))
1103 update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1106 /* Prefetch descriptor index. */
1107 rte_prefetch0(&vq->desc[desc_indexes[0]]);
1108 for (i = 0; i < count; i++) {
1109 struct vring_desc *desc;
1113 if (likely(i + 1 < count))
1114 rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1116 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1117 desc = (struct vring_desc *)(uintptr_t)
1118 rte_vhost_gpa_to_vva(dev->mem,
1119 vq->desc[desc_indexes[i]].addr);
1120 if (unlikely(!desc))
1123 rte_prefetch0(desc);
1124 sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1129 idx = desc_indexes[i];
1132 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1133 if (unlikely(pkts[i] == NULL)) {
1134 RTE_LOG(ERR, VHOST_DATA,
1135 "Failed to allocate memory for mbuf.\n");
1139 err = copy_desc_to_mbuf(dev, desc, sz, pkts[i], idx, mbuf_pool);
1140 if (unlikely(err)) {
1141 rte_pktmbuf_free(pkts[i]);
1145 if (unlikely(dev->dequeue_zero_copy)) {
1146 struct zcopy_mbuf *zmbuf;
1148 zmbuf = get_zmbuf(vq);
1150 rte_pktmbuf_free(pkts[i]);
1153 zmbuf->mbuf = pkts[i];
1154 zmbuf->desc_idx = desc_indexes[i];
1157 * Pin lock the mbuf; we will check later to see
1158 * whether the mbuf is freed (when we are the last
1159 * user) or not. If that's the case, we then could
1160 * update the used ring safely.
1162 rte_mbuf_refcnt_update(pkts[i], 1);
1165 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1168 vq->last_avail_idx += i;
1170 if (likely(dev->dequeue_zero_copy == 0)) {
1171 vq->last_used_idx += i;
1172 update_used_idx(dev, vq, i);
1176 if (unlikely(rarp_mbuf != NULL)) {
1178 * Inject it to the head of "pkts" array, so that switch's mac
1179 * learning table will get updated first.
1181 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1182 pkts[0] = rarp_mbuf;