4 * Copyright(c) 2010-2014 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_virtio_net.h>
48 #include "vhost-net.h"
50 #define MAX_PKT_BURST 32
51 #define VHOST_LOG_PAGE 4096
53 static inline void __attribute__((always_inline))
54 vhost_log_page(uint8_t *log_base, uint64_t page)
56 log_base[page / 8] |= 1 << (page % 8);
59 static inline void __attribute__((always_inline))
60 vhost_log_write(struct virtio_net *dev, uint64_t addr, uint64_t len)
64 if (likely(((dev->features & (1ULL << VHOST_F_LOG_ALL)) == 0) ||
65 !dev->log_base || !len))
68 if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
71 /* To make sure guest memory updates are committed before logging */
74 page = addr / VHOST_LOG_PAGE;
75 while (page * VHOST_LOG_PAGE < addr + len) {
76 vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
81 static inline void __attribute__((always_inline))
82 vhost_log_used_vring(struct virtio_net *dev, struct vhost_virtqueue *vq,
83 uint64_t offset, uint64_t len)
85 vhost_log_write(dev, vq->log_guest_addr + offset, len);
89 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t qp_nb)
91 return (is_tx ^ (idx & 1)) == 0 && idx < qp_nb * VIRTIO_QNUM;
95 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
97 if (m_buf->ol_flags & PKT_TX_L4_MASK) {
98 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
99 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
101 switch (m_buf->ol_flags & PKT_TX_L4_MASK) {
102 case PKT_TX_TCP_CKSUM:
103 net_hdr->csum_offset = (offsetof(struct tcp_hdr,
106 case PKT_TX_UDP_CKSUM:
107 net_hdr->csum_offset = (offsetof(struct udp_hdr,
110 case PKT_TX_SCTP_CKSUM:
111 net_hdr->csum_offset = (offsetof(struct sctp_hdr,
117 if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
118 if (m_buf->ol_flags & PKT_TX_IPV4)
119 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
121 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
122 net_hdr->gso_size = m_buf->tso_segsz;
123 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
129 copy_virtio_net_hdr(struct virtio_net *dev, uint64_t desc_addr,
130 struct virtio_net_hdr_mrg_rxbuf hdr)
132 if (dev->vhost_hlen == sizeof(struct virtio_net_hdr_mrg_rxbuf))
133 *(struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr = hdr;
135 *(struct virtio_net_hdr *)(uintptr_t)desc_addr = hdr.hdr;
138 static inline int __attribute__((always_inline))
139 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
140 struct rte_mbuf *m, uint16_t desc_idx)
142 uint32_t desc_avail, desc_offset;
143 uint32_t mbuf_avail, mbuf_offset;
145 struct vring_desc *desc;
147 struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
149 desc = &vq->desc[desc_idx];
150 if (unlikely(desc->len < dev->vhost_hlen))
153 desc_addr = gpa_to_vva(dev, desc->addr);
154 rte_prefetch0((void *)(uintptr_t)desc_addr);
156 virtio_enqueue_offload(m, &virtio_hdr.hdr);
157 copy_virtio_net_hdr(dev, desc_addr, virtio_hdr);
158 vhost_log_write(dev, desc->addr, dev->vhost_hlen);
159 PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
161 desc_offset = dev->vhost_hlen;
162 desc_avail = desc->len - dev->vhost_hlen;
164 mbuf_avail = rte_pktmbuf_data_len(m);
166 while (mbuf_avail != 0 || m->next != NULL) {
167 /* done with current mbuf, fetch next */
168 if (mbuf_avail == 0) {
172 mbuf_avail = rte_pktmbuf_data_len(m);
175 /* done with current desc buf, fetch next */
176 if (desc_avail == 0) {
177 if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
178 /* Room in vring buffer is not enough */
181 if (unlikely(desc->next >= vq->size))
184 desc = &vq->desc[desc->next];
185 desc_addr = gpa_to_vva(dev, desc->addr);
187 desc_avail = desc->len;
190 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
191 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
192 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
194 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
195 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
198 mbuf_avail -= cpy_len;
199 mbuf_offset += cpy_len;
200 desc_avail -= cpy_len;
201 desc_offset += cpy_len;
208 * As many data cores may want to access available buffers
209 * they need to be reserved.
211 static inline uint32_t
212 reserve_avail_buf(struct vhost_virtqueue *vq, uint32_t count,
213 uint16_t *start, uint16_t *end)
215 uint16_t res_start_idx;
216 uint16_t res_end_idx;
218 uint16_t free_entries;
221 count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
224 res_start_idx = vq->last_used_idx_res;
225 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
227 free_entries = avail_idx - res_start_idx;
228 count = RTE_MIN(count, free_entries);
232 res_end_idx = res_start_idx + count;
235 * update vq->last_used_idx_res atomically; try again if failed.
237 * TODO: Allow to disable cmpset if no concurrency in application.
239 success = rte_atomic16_cmpset(&vq->last_used_idx_res,
240 res_start_idx, res_end_idx);
241 if (unlikely(!success))
244 *start = res_start_idx;
251 * This function adds buffers to the virtio devices RX virtqueue. Buffers can
252 * be received from the physical port or from another virtio device. A packet
253 * count is returned to indicate the number of packets that are succesfully
254 * added to the RX queue. This function works when the mbuf is scattered, but
255 * it doesn't support the mergeable feature.
257 static inline uint32_t __attribute__((always_inline))
258 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
259 struct rte_mbuf **pkts, uint32_t count)
261 struct vhost_virtqueue *vq;
262 uint16_t res_start_idx, res_end_idx;
263 uint16_t desc_indexes[MAX_PKT_BURST];
267 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
268 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
269 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
270 dev->vid, __func__, queue_id);
274 vq = dev->virtqueue[queue_id];
275 if (unlikely(vq->enabled == 0))
278 count = reserve_avail_buf(vq, count, &res_start_idx, &res_end_idx);
282 LOG_DEBUG(VHOST_DATA, "(%d) res_start_idx %d | res_end_idx Index %d\n",
283 dev->vid, res_start_idx, res_end_idx);
285 /* Retrieve all of the desc indexes first to avoid caching issues. */
286 rte_prefetch0(&vq->avail->ring[res_start_idx & (vq->size - 1)]);
287 for (i = 0; i < count; i++) {
288 used_idx = (res_start_idx + i) & (vq->size - 1);
289 desc_indexes[i] = vq->avail->ring[used_idx];
290 vq->used->ring[used_idx].id = desc_indexes[i];
291 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
293 vhost_log_used_vring(dev, vq,
294 offsetof(struct vring_used, ring[used_idx]),
295 sizeof(vq->used->ring[used_idx]));
298 rte_prefetch0(&vq->desc[desc_indexes[0]]);
299 for (i = 0; i < count; i++) {
300 uint16_t desc_idx = desc_indexes[i];
303 err = copy_mbuf_to_desc(dev, vq, pkts[i], desc_idx);
305 used_idx = (res_start_idx + i) & (vq->size - 1);
306 vq->used->ring[used_idx].len = dev->vhost_hlen;
307 vhost_log_used_vring(dev, vq,
308 offsetof(struct vring_used, ring[used_idx]),
309 sizeof(vq->used->ring[used_idx]));
313 rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
318 /* Wait until it's our turn to add our buffer to the used ring. */
319 while (unlikely(vq->last_used_idx != res_start_idx))
322 *(volatile uint16_t *)&vq->used->idx += count;
323 vq->last_used_idx = res_end_idx;
324 vhost_log_used_vring(dev, vq,
325 offsetof(struct vring_used, idx),
326 sizeof(vq->used->idx));
328 /* flush used->idx update before we read avail->flags. */
331 /* Kick the guest if necessary. */
332 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
333 && (vq->callfd >= 0))
334 eventfd_write(vq->callfd, (eventfd_t)1);
339 fill_vec_buf(struct vhost_virtqueue *vq, uint32_t avail_idx,
340 uint32_t *allocated, uint32_t *vec_idx,
341 struct buf_vector *buf_vec)
343 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
344 uint32_t vec_id = *vec_idx;
345 uint32_t len = *allocated;
348 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
351 len += vq->desc[idx].len;
352 buf_vec[vec_id].buf_addr = vq->desc[idx].addr;
353 buf_vec[vec_id].buf_len = vq->desc[idx].len;
354 buf_vec[vec_id].desc_idx = idx;
357 if ((vq->desc[idx].flags & VRING_DESC_F_NEXT) == 0)
360 idx = vq->desc[idx].next;
370 * As many data cores may want to access available buffers concurrently,
371 * they need to be reserved.
373 * Returns -1 on fail, 0 on success
376 reserve_avail_buf_mergeable(struct vhost_virtqueue *vq, uint32_t size,
377 uint16_t *start, uint16_t *end,
378 struct buf_vector *buf_vec)
380 uint16_t res_start_idx;
381 uint16_t res_cur_idx;
388 res_start_idx = vq->last_used_idx_res;
389 res_cur_idx = res_start_idx;
395 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
396 if (unlikely(res_cur_idx == avail_idx))
399 if (unlikely(fill_vec_buf(vq, res_cur_idx, &allocated,
400 &vec_idx, buf_vec) < 0))
406 if (allocated >= size)
410 * if we tried all available ring items, and still
411 * can't get enough buf, it means something abnormal
414 if (unlikely(tries >= vq->size))
419 * update vq->last_used_idx_res atomically.
420 * retry again if failed.
422 if (rte_atomic16_cmpset(&vq->last_used_idx_res,
423 res_start_idx, res_cur_idx) == 0)
426 *start = res_start_idx;
431 static inline uint32_t __attribute__((always_inline))
432 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
433 uint16_t res_start_idx, uint16_t res_end_idx,
434 struct rte_mbuf *m, struct buf_vector *buf_vec)
436 struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
437 uint32_t vec_idx = 0;
438 uint16_t cur_idx = res_start_idx;
440 uint32_t mbuf_offset, mbuf_avail;
441 uint32_t desc_offset, desc_avail;
443 uint16_t desc_idx, used_idx;
445 if (unlikely(m == NULL))
448 LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
449 dev->vid, cur_idx, res_end_idx);
451 if (buf_vec[vec_idx].buf_len < dev->vhost_hlen)
454 desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
455 rte_prefetch0((void *)(uintptr_t)desc_addr);
457 virtio_hdr.num_buffers = res_end_idx - res_start_idx;
458 LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
459 dev->vid, virtio_hdr.num_buffers);
461 virtio_enqueue_offload(m, &virtio_hdr.hdr);
462 copy_virtio_net_hdr(dev, desc_addr, virtio_hdr);
463 vhost_log_write(dev, buf_vec[vec_idx].buf_addr, dev->vhost_hlen);
464 PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
466 desc_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
467 desc_offset = dev->vhost_hlen;
469 mbuf_avail = rte_pktmbuf_data_len(m);
471 while (mbuf_avail != 0 || m->next != NULL) {
472 /* done with current desc buf, get the next one */
473 if (desc_avail == 0) {
474 desc_idx = buf_vec[vec_idx].desc_idx;
476 if (!(vq->desc[desc_idx].flags & VRING_DESC_F_NEXT)) {
477 /* Update used ring with desc information */
478 used_idx = cur_idx++ & (vq->size - 1);
479 vq->used->ring[used_idx].id = desc_idx;
480 vq->used->ring[used_idx].len = desc_offset;
481 vhost_log_used_vring(dev, vq,
482 offsetof(struct vring_used,
484 sizeof(vq->used->ring[used_idx]));
488 desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
490 /* Prefetch buffer address. */
491 rte_prefetch0((void *)(uintptr_t)desc_addr);
493 desc_avail = buf_vec[vec_idx].buf_len;
496 /* done with current mbuf, get the next one */
497 if (mbuf_avail == 0) {
501 mbuf_avail = rte_pktmbuf_data_len(m);
504 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
505 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
506 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
508 vhost_log_write(dev, buf_vec[vec_idx].buf_addr + desc_offset,
510 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
513 mbuf_avail -= cpy_len;
514 mbuf_offset += cpy_len;
515 desc_avail -= cpy_len;
516 desc_offset += cpy_len;
519 used_idx = cur_idx & (vq->size - 1);
520 vq->used->ring[used_idx].id = buf_vec[vec_idx].desc_idx;
521 vq->used->ring[used_idx].len = desc_offset;
522 vhost_log_used_vring(dev, vq,
523 offsetof(struct vring_used, ring[used_idx]),
524 sizeof(vq->used->ring[used_idx]));
526 return res_end_idx - res_start_idx;
529 static inline uint32_t __attribute__((always_inline))
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, nr_used = 0;
536 struct buf_vector buf_vec[BUF_VECTOR_MAX];
538 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
539 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
540 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
541 dev->vid, __func__, queue_id);
545 vq = dev->virtqueue[queue_id];
546 if (unlikely(vq->enabled == 0))
549 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
553 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
554 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
556 if (unlikely(reserve_avail_buf_mergeable(vq, pkt_len, &start,
557 &end, buf_vec) < 0)) {
558 LOG_DEBUG(VHOST_DATA,
559 "(%d) failed to get enough desc from vring\n",
564 nr_used = copy_mbuf_to_desc_mergeable(dev, vq, start, end,
565 pkts[pkt_idx], buf_vec);
569 * Wait until it's our turn to add our buffer
572 while (unlikely(vq->last_used_idx != start))
575 *(volatile uint16_t *)&vq->used->idx += nr_used;
576 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
577 sizeof(vq->used->idx));
578 vq->last_used_idx = end;
581 if (likely(pkt_idx)) {
582 /* flush used->idx update before we read avail->flags. */
585 /* Kick the guest if necessary. */
586 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
587 && (vq->callfd >= 0))
588 eventfd_write(vq->callfd, (eventfd_t)1);
595 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
596 struct rte_mbuf **pkts, uint16_t count)
598 struct virtio_net *dev = get_device(vid);
603 if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
604 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
606 return virtio_dev_rx(dev, queue_id, pkts, count);
610 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
612 struct ipv4_hdr *ipv4_hdr;
613 struct ipv6_hdr *ipv6_hdr;
615 struct ether_hdr *eth_hdr;
618 eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
620 m->l2_len = sizeof(struct ether_hdr);
621 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
623 if (ethertype == ETHER_TYPE_VLAN) {
624 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
626 m->l2_len += sizeof(struct vlan_hdr);
627 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
630 l3_hdr = (char *)eth_hdr + m->l2_len;
633 case ETHER_TYPE_IPv4:
634 ipv4_hdr = (struct ipv4_hdr *)l3_hdr;
635 *l4_proto = ipv4_hdr->next_proto_id;
636 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
637 *l4_hdr = (char *)l3_hdr + m->l3_len;
638 m->ol_flags |= PKT_TX_IPV4;
640 case ETHER_TYPE_IPv6:
641 ipv6_hdr = (struct ipv6_hdr *)l3_hdr;
642 *l4_proto = ipv6_hdr->proto;
643 m->l3_len = sizeof(struct ipv6_hdr);
644 *l4_hdr = (char *)l3_hdr + m->l3_len;
645 m->ol_flags |= PKT_TX_IPV6;
654 static inline void __attribute__((always_inline))
655 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
657 uint16_t l4_proto = 0;
659 struct tcp_hdr *tcp_hdr = NULL;
661 parse_ethernet(m, &l4_proto, &l4_hdr);
662 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
663 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
664 switch (hdr->csum_offset) {
665 case (offsetof(struct tcp_hdr, cksum)):
666 if (l4_proto == IPPROTO_TCP)
667 m->ol_flags |= PKT_TX_TCP_CKSUM;
669 case (offsetof(struct udp_hdr, dgram_cksum)):
670 if (l4_proto == IPPROTO_UDP)
671 m->ol_flags |= PKT_TX_UDP_CKSUM;
673 case (offsetof(struct sctp_hdr, cksum)):
674 if (l4_proto == IPPROTO_SCTP)
675 m->ol_flags |= PKT_TX_SCTP_CKSUM;
683 if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
684 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
685 case VIRTIO_NET_HDR_GSO_TCPV4:
686 case VIRTIO_NET_HDR_GSO_TCPV6:
687 tcp_hdr = (struct tcp_hdr *)l4_hdr;
688 m->ol_flags |= PKT_TX_TCP_SEG;
689 m->tso_segsz = hdr->gso_size;
690 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
693 RTE_LOG(WARNING, VHOST_DATA,
694 "unsupported gso type %u.\n", hdr->gso_type);
700 #define RARP_PKT_SIZE 64
703 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
705 struct ether_hdr *eth_hdr;
706 struct arp_hdr *rarp;
708 if (rarp_mbuf->buf_len < 64) {
709 RTE_LOG(WARNING, VHOST_DATA,
710 "failed to make RARP; mbuf size too small %u (< %d)\n",
711 rarp_mbuf->buf_len, RARP_PKT_SIZE);
715 /* Ethernet header. */
716 eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
717 memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
718 ether_addr_copy(mac, ð_hdr->s_addr);
719 eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
722 rarp = (struct arp_hdr *)(eth_hdr + 1);
723 rarp->arp_hrd = htons(ARP_HRD_ETHER);
724 rarp->arp_pro = htons(ETHER_TYPE_IPv4);
725 rarp->arp_hln = ETHER_ADDR_LEN;
727 rarp->arp_op = htons(ARP_OP_REVREQUEST);
729 ether_addr_copy(mac, &rarp->arp_data.arp_sha);
730 ether_addr_copy(mac, &rarp->arp_data.arp_tha);
731 memset(&rarp->arp_data.arp_sip, 0x00, 4);
732 memset(&rarp->arp_data.arp_tip, 0x00, 4);
734 rarp_mbuf->pkt_len = rarp_mbuf->data_len = RARP_PKT_SIZE;
739 static inline int __attribute__((always_inline))
740 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
741 struct rte_mbuf *m, uint16_t desc_idx,
742 struct rte_mempool *mbuf_pool)
744 struct vring_desc *desc;
746 uint32_t desc_avail, desc_offset;
747 uint32_t mbuf_avail, mbuf_offset;
749 struct rte_mbuf *cur = m, *prev = m;
750 struct virtio_net_hdr *hdr;
751 /* A counter to avoid desc dead loop chain */
752 uint32_t nr_desc = 1;
754 desc = &vq->desc[desc_idx];
755 if (unlikely(desc->len < dev->vhost_hlen))
758 desc_addr = gpa_to_vva(dev, desc->addr);
759 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
763 * A virtio driver normally uses at least 2 desc buffers
764 * for Tx: the first for storing the header, and others
765 * for storing the data.
767 if (likely((desc->len == dev->vhost_hlen) &&
768 (desc->flags & VRING_DESC_F_NEXT) != 0)) {
769 desc = &vq->desc[desc->next];
771 desc_addr = gpa_to_vva(dev, desc->addr);
772 rte_prefetch0((void *)(uintptr_t)desc_addr);
775 desc_avail = desc->len;
778 PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
780 desc_avail = desc->len - dev->vhost_hlen;
781 desc_offset = dev->vhost_hlen;
785 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
787 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
788 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *, mbuf_offset),
789 (void *)((uintptr_t)(desc_addr + desc_offset)),
792 mbuf_avail -= cpy_len;
793 mbuf_offset += cpy_len;
794 desc_avail -= cpy_len;
795 desc_offset += cpy_len;
797 /* This desc reaches to its end, get the next one */
798 if (desc_avail == 0) {
799 if ((desc->flags & VRING_DESC_F_NEXT) == 0)
802 if (unlikely(desc->next >= vq->size ||
803 ++nr_desc >= vq->size))
805 desc = &vq->desc[desc->next];
807 desc_addr = gpa_to_vva(dev, desc->addr);
808 rte_prefetch0((void *)(uintptr_t)desc_addr);
811 desc_avail = desc->len;
813 PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
817 * This mbuf reaches to its end, get a new one
820 if (mbuf_avail == 0) {
821 cur = rte_pktmbuf_alloc(mbuf_pool);
822 if (unlikely(cur == NULL)) {
823 RTE_LOG(ERR, VHOST_DATA, "Failed to "
824 "allocate memory for mbuf.\n");
829 prev->data_len = mbuf_offset;
831 m->pkt_len += mbuf_offset;
835 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
839 prev->data_len = mbuf_offset;
840 m->pkt_len += mbuf_offset;
842 if (hdr->flags != 0 || hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE)
843 vhost_dequeue_offload(hdr, m);
849 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
850 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
852 struct virtio_net *dev;
853 struct rte_mbuf *rarp_mbuf = NULL;
854 struct vhost_virtqueue *vq;
855 uint32_t desc_indexes[MAX_PKT_BURST];
858 uint16_t free_entries;
861 dev = get_device(vid);
865 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->virt_qp_nb))) {
866 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
867 dev->vid, __func__, queue_id);
871 vq = dev->virtqueue[queue_id];
872 if (unlikely(vq->enabled == 0))
876 * Construct a RARP broadcast packet, and inject it to the "pkts"
877 * array, to looks like that guest actually send such packet.
879 * Check user_send_rarp() for more information.
881 if (unlikely(rte_atomic16_cmpset((volatile uint16_t *)
882 &dev->broadcast_rarp.cnt, 1, 0))) {
883 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
884 if (rarp_mbuf == NULL) {
885 RTE_LOG(ERR, VHOST_DATA,
886 "Failed to allocate memory for mbuf.\n");
890 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
891 rte_pktmbuf_free(rarp_mbuf);
898 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
899 free_entries = avail_idx - vq->last_used_idx;
900 if (free_entries == 0)
903 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
905 /* Prefetch available ring to retrieve head indexes. */
906 used_idx = vq->last_used_idx & (vq->size - 1);
907 rte_prefetch0(&vq->avail->ring[used_idx]);
908 rte_prefetch0(&vq->used->ring[used_idx]);
910 count = RTE_MIN(count, MAX_PKT_BURST);
911 count = RTE_MIN(count, free_entries);
912 LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
915 /* Retrieve all of the head indexes first to avoid caching issues. */
916 for (i = 0; i < count; i++) {
917 used_idx = (vq->last_used_idx + i) & (vq->size - 1);
918 desc_indexes[i] = vq->avail->ring[used_idx];
920 vq->used->ring[used_idx].id = desc_indexes[i];
921 vq->used->ring[used_idx].len = 0;
922 vhost_log_used_vring(dev, vq,
923 offsetof(struct vring_used, ring[used_idx]),
924 sizeof(vq->used->ring[used_idx]));
927 /* Prefetch descriptor index. */
928 rte_prefetch0(&vq->desc[desc_indexes[0]]);
929 for (i = 0; i < count; i++) {
932 if (likely(i + 1 < count))
933 rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
935 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
936 if (unlikely(pkts[i] == NULL)) {
937 RTE_LOG(ERR, VHOST_DATA,
938 "Failed to allocate memory for mbuf.\n");
941 err = copy_desc_to_mbuf(dev, vq, pkts[i], desc_indexes[i],
944 rte_pktmbuf_free(pkts[i]);
952 vq->last_used_idx += i;
953 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
954 sizeof(vq->used->idx));
956 /* Kick guest if required. */
957 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
958 && (vq->callfd >= 0))
959 eventfd_write(vq->callfd, (eventfd_t)1);
962 if (unlikely(rarp_mbuf != NULL)) {
964 * Inject it to the head of "pkts" array, so that switch's mac
965 * learning table will get updated first.
967 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));