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 memset(net_hdr, 0, sizeof(struct virtio_net_hdr));
99 if (m_buf->ol_flags & PKT_TX_L4_MASK) {
100 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
101 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
103 switch (m_buf->ol_flags & PKT_TX_L4_MASK) {
104 case PKT_TX_TCP_CKSUM:
105 net_hdr->csum_offset = (offsetof(struct tcp_hdr,
108 case PKT_TX_UDP_CKSUM:
109 net_hdr->csum_offset = (offsetof(struct udp_hdr,
112 case PKT_TX_SCTP_CKSUM:
113 net_hdr->csum_offset = (offsetof(struct sctp_hdr,
119 if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
120 if (m_buf->ol_flags & PKT_TX_IPV4)
121 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
123 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
124 net_hdr->gso_size = m_buf->tso_segsz;
125 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
133 copy_virtio_net_hdr(struct vhost_virtqueue *vq, uint64_t desc_addr,
134 struct virtio_net_hdr_mrg_rxbuf hdr)
136 if (vq->vhost_hlen == sizeof(struct virtio_net_hdr_mrg_rxbuf))
137 *(struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr = hdr;
139 *(struct virtio_net_hdr *)(uintptr_t)desc_addr = hdr.hdr;
142 static inline int __attribute__((always_inline))
143 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
144 struct rte_mbuf *m, uint16_t desc_idx, uint32_t *copied)
146 uint32_t desc_avail, desc_offset;
147 uint32_t mbuf_avail, mbuf_offset;
149 struct vring_desc *desc;
151 struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
153 desc = &vq->desc[desc_idx];
154 if (unlikely(desc->len < vq->vhost_hlen))
157 desc_addr = gpa_to_vva(dev, desc->addr);
158 rte_prefetch0((void *)(uintptr_t)desc_addr);
160 virtio_enqueue_offload(m, &virtio_hdr.hdr);
161 copy_virtio_net_hdr(vq, desc_addr, virtio_hdr);
162 vhost_log_write(dev, desc->addr, vq->vhost_hlen);
163 PRINT_PACKET(dev, (uintptr_t)desc_addr, vq->vhost_hlen, 0);
165 desc_offset = vq->vhost_hlen;
166 desc_avail = desc->len - vq->vhost_hlen;
168 *copied = rte_pktmbuf_pkt_len(m);
169 mbuf_avail = rte_pktmbuf_data_len(m);
171 while (mbuf_avail != 0 || m->next != NULL) {
172 /* done with current mbuf, fetch next */
173 if (mbuf_avail == 0) {
177 mbuf_avail = rte_pktmbuf_data_len(m);
180 /* done with current desc buf, fetch next */
181 if (desc_avail == 0) {
182 if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
183 /* Room in vring buffer is not enough */
186 if (unlikely(desc->next >= vq->size))
189 desc = &vq->desc[desc->next];
190 desc_addr = gpa_to_vva(dev, desc->addr);
192 desc_avail = desc->len;
195 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
196 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
197 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
199 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
200 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
203 mbuf_avail -= cpy_len;
204 mbuf_offset += cpy_len;
205 desc_avail -= cpy_len;
206 desc_offset += cpy_len;
213 * As many data cores may want to access available buffers
214 * they need to be reserved.
216 static inline uint32_t
217 reserve_avail_buf(struct vhost_virtqueue *vq, uint32_t count,
218 uint16_t *start, uint16_t *end)
220 uint16_t res_start_idx;
221 uint16_t res_end_idx;
223 uint16_t free_entries;
226 count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
229 res_start_idx = vq->last_used_idx_res;
230 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
232 free_entries = avail_idx - res_start_idx;
233 count = RTE_MIN(count, free_entries);
237 res_end_idx = res_start_idx + count;
240 * update vq->last_used_idx_res atomically; try again if failed.
242 * TODO: Allow to disable cmpset if no concurrency in application.
244 success = rte_atomic16_cmpset(&vq->last_used_idx_res,
245 res_start_idx, res_end_idx);
246 if (unlikely(!success))
249 *start = res_start_idx;
256 * This function adds buffers to the virtio devices RX virtqueue. Buffers can
257 * be received from the physical port or from another virtio device. A packet
258 * count is returned to indicate the number of packets that are succesfully
259 * added to the RX queue. This function works when the mbuf is scattered, but
260 * it doesn't support the mergeable feature.
262 static inline uint32_t __attribute__((always_inline))
263 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
264 struct rte_mbuf **pkts, uint32_t count)
266 struct vhost_virtqueue *vq;
267 uint16_t res_start_idx, res_end_idx;
268 uint16_t desc_indexes[MAX_PKT_BURST];
271 LOG_DEBUG(VHOST_DATA, "(%"PRIu64") virtio_dev_rx()\n", dev->device_fh);
272 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
273 RTE_LOG(ERR, VHOST_DATA,
274 "%s (%"PRIu64"): virtqueue idx:%d invalid.\n",
275 __func__, dev->device_fh, queue_id);
279 vq = dev->virtqueue[queue_id];
280 if (unlikely(vq->enabled == 0))
283 count = reserve_avail_buf(vq, count, &res_start_idx, &res_end_idx);
287 LOG_DEBUG(VHOST_DATA,
288 "(%"PRIu64") res_start_idx %d| res_end_idx Index %d\n",
289 dev->device_fh, res_start_idx, res_end_idx);
291 /* Retrieve all of the desc indexes first to avoid caching issues. */
292 rte_prefetch0(&vq->avail->ring[res_start_idx & (vq->size - 1)]);
293 for (i = 0; i < count; i++) {
294 desc_indexes[i] = vq->avail->ring[(res_start_idx + i) &
298 rte_prefetch0(&vq->desc[desc_indexes[0]]);
299 for (i = 0; i < count; i++) {
300 uint16_t desc_idx = desc_indexes[i];
301 uint16_t used_idx = (res_start_idx + i) & (vq->size - 1);
305 err = copy_mbuf_to_desc(dev, vq, pkts[i], desc_idx, &copied);
307 vq->used->ring[used_idx].id = desc_idx;
309 vq->used->ring[used_idx].len = vq->vhost_hlen;
311 vq->used->ring[used_idx].len = copied + vq->vhost_hlen;
312 vhost_log_used_vring(dev, vq,
313 offsetof(struct vring_used, ring[used_idx]),
314 sizeof(vq->used->ring[used_idx]));
317 rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
320 rte_compiler_barrier();
322 /* Wait until it's our turn to add our buffer to the used ring. */
323 while (unlikely(vq->last_used_idx != res_start_idx))
326 *(volatile uint16_t *)&vq->used->idx += count;
327 vq->last_used_idx = res_end_idx;
328 vhost_log_used_vring(dev, vq,
329 offsetof(struct vring_used, idx),
330 sizeof(vq->used->idx));
332 /* flush used->idx update before we read avail->flags. */
335 /* Kick the guest if necessary. */
336 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT))
337 eventfd_write(vq->callfd, (eventfd_t)1);
342 fill_vec_buf(struct vhost_virtqueue *vq, uint32_t avail_idx,
343 uint32_t *allocated, uint32_t *vec_idx)
345 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
346 uint32_t vec_id = *vec_idx;
347 uint32_t len = *allocated;
350 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
353 len += vq->desc[idx].len;
354 vq->buf_vec[vec_id].buf_addr = vq->desc[idx].addr;
355 vq->buf_vec[vec_id].buf_len = vq->desc[idx].len;
356 vq->buf_vec[vec_id].desc_idx = idx;
359 if ((vq->desc[idx].flags & VRING_DESC_F_NEXT) == 0)
362 idx = vq->desc[idx].next;
372 * As many data cores may want to access available buffers concurrently,
373 * they need to be reserved.
375 * Returns -1 on fail, 0 on success
378 reserve_avail_buf_mergeable(struct vhost_virtqueue *vq, uint32_t size,
379 uint16_t *start, uint16_t *end)
381 uint16_t res_start_idx;
382 uint16_t res_cur_idx;
389 res_start_idx = vq->last_used_idx_res;
390 res_cur_idx = res_start_idx;
396 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
397 if (unlikely(res_cur_idx == avail_idx))
400 if (unlikely(fill_vec_buf(vq, res_cur_idx, &allocated,
407 if (allocated >= size)
411 * if we tried all available ring items, and still
412 * can't get enough buf, it means something abnormal
415 if (unlikely(tries >= vq->size))
420 * update vq->last_used_idx_res atomically.
421 * retry again if failed.
423 if (rte_atomic16_cmpset(&vq->last_used_idx_res,
424 res_start_idx, res_cur_idx) == 0)
427 *start = res_start_idx;
432 static inline uint32_t __attribute__((always_inline))
433 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
434 uint16_t res_start_idx, uint16_t res_end_idx,
437 struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
438 uint32_t vec_idx = 0;
439 uint16_t cur_idx = res_start_idx;
441 uint32_t mbuf_offset, mbuf_avail;
442 uint32_t desc_offset, desc_avail;
444 uint16_t desc_idx, used_idx;
446 if (unlikely(m == NULL))
449 LOG_DEBUG(VHOST_DATA,
450 "(%"PRIu64") Current Index %d| End Index %d\n",
451 dev->device_fh, cur_idx, res_end_idx);
453 if (vq->buf_vec[vec_idx].buf_len < vq->vhost_hlen)
456 desc_addr = gpa_to_vva(dev, vq->buf_vec[vec_idx].buf_addr);
457 rte_prefetch0((void *)(uintptr_t)desc_addr);
459 virtio_hdr.num_buffers = res_end_idx - res_start_idx;
460 LOG_DEBUG(VHOST_DATA, "(%"PRIu64") RX: Num merge buffers %d\n",
461 dev->device_fh, virtio_hdr.num_buffers);
463 virtio_enqueue_offload(m, &virtio_hdr.hdr);
464 copy_virtio_net_hdr(vq, desc_addr, virtio_hdr);
465 vhost_log_write(dev, vq->buf_vec[vec_idx].buf_addr, vq->vhost_hlen);
466 PRINT_PACKET(dev, (uintptr_t)desc_addr, vq->vhost_hlen, 0);
468 desc_avail = vq->buf_vec[vec_idx].buf_len - vq->vhost_hlen;
469 desc_offset = vq->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) {
476 desc_idx = vq->buf_vec[vec_idx].desc_idx;
478 if (!(vq->desc[desc_idx].flags & VRING_DESC_F_NEXT)) {
479 /* Update used ring with desc information */
480 used_idx = cur_idx++ & (vq->size - 1);
481 vq->used->ring[used_idx].id = desc_idx;
482 vq->used->ring[used_idx].len = desc_offset;
483 vhost_log_used_vring(dev, vq,
484 offsetof(struct vring_used,
486 sizeof(vq->used->ring[used_idx]));
490 desc_addr = gpa_to_vva(dev, vq->buf_vec[vec_idx].buf_addr);
492 /* Prefetch buffer address. */
493 rte_prefetch0((void *)(uintptr_t)desc_addr);
495 desc_avail = vq->buf_vec[vec_idx].buf_len;
498 /* done with current mbuf, get the next one */
499 if (mbuf_avail == 0) {
503 mbuf_avail = rte_pktmbuf_data_len(m);
506 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
507 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
508 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
510 vhost_log_write(dev, vq->buf_vec[vec_idx].buf_addr + desc_offset,
512 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
515 mbuf_avail -= cpy_len;
516 mbuf_offset += cpy_len;
517 desc_avail -= cpy_len;
518 desc_offset += cpy_len;
521 used_idx = cur_idx & (vq->size - 1);
522 vq->used->ring[used_idx].id = vq->buf_vec[vec_idx].desc_idx;
523 vq->used->ring[used_idx].len = desc_offset;
524 vhost_log_used_vring(dev, vq,
525 offsetof(struct vring_used, ring[used_idx]),
526 sizeof(vq->used->ring[used_idx]));
528 return res_end_idx - res_start_idx;
531 static inline uint32_t __attribute__((always_inline))
532 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
533 struct rte_mbuf **pkts, uint32_t count)
535 struct vhost_virtqueue *vq;
536 uint32_t pkt_idx = 0, nr_used = 0;
539 LOG_DEBUG(VHOST_DATA, "(%"PRIu64") virtio_dev_merge_rx()\n",
541 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
542 RTE_LOG(ERR, VHOST_DATA,
543 "%s (%"PRIu64"): virtqueue idx:%d invalid.\n",
544 __func__, dev->device_fh, queue_id);
548 vq = dev->virtqueue[queue_id];
549 if (unlikely(vq->enabled == 0))
552 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
556 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
557 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + vq->vhost_hlen;
559 if (unlikely(reserve_avail_buf_mergeable(vq, pkt_len,
560 &start, &end) < 0)) {
561 LOG_DEBUG(VHOST_DATA,
562 "(%" PRIu64 ") Failed to get enough desc from vring\n",
567 nr_used = copy_mbuf_to_desc_mergeable(dev, vq, start, end,
569 rte_compiler_barrier();
572 * Wait until it's our turn to add our buffer
575 while (unlikely(vq->last_used_idx != start))
578 *(volatile uint16_t *)&vq->used->idx += nr_used;
579 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
580 sizeof(vq->used->idx));
581 vq->last_used_idx = end;
584 if (likely(pkt_idx)) {
585 /* flush used->idx update before we read avail->flags. */
588 /* Kick the guest if necessary. */
589 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT))
590 eventfd_write(vq->callfd, (eventfd_t)1);
597 rte_vhost_enqueue_burst(struct virtio_net *dev, uint16_t queue_id,
598 struct rte_mbuf **pkts, uint16_t count)
600 if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
601 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
603 return virtio_dev_rx(dev, queue_id, pkts, count);
607 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
609 struct ipv4_hdr *ipv4_hdr;
610 struct ipv6_hdr *ipv6_hdr;
612 struct ether_hdr *eth_hdr;
615 eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
617 m->l2_len = sizeof(struct ether_hdr);
618 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
620 if (ethertype == ETHER_TYPE_VLAN) {
621 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
623 m->l2_len += sizeof(struct vlan_hdr);
624 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
627 l3_hdr = (char *)eth_hdr + m->l2_len;
630 case ETHER_TYPE_IPv4:
631 ipv4_hdr = (struct ipv4_hdr *)l3_hdr;
632 *l4_proto = ipv4_hdr->next_proto_id;
633 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
634 *l4_hdr = (char *)l3_hdr + m->l3_len;
635 m->ol_flags |= PKT_TX_IPV4;
637 case ETHER_TYPE_IPv6:
638 ipv6_hdr = (struct ipv6_hdr *)l3_hdr;
639 *l4_proto = ipv6_hdr->proto;
640 m->l3_len = sizeof(struct ipv6_hdr);
641 *l4_hdr = (char *)l3_hdr + m->l3_len;
642 m->ol_flags |= PKT_TX_IPV6;
651 static inline void __attribute__((always_inline))
652 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
654 uint16_t l4_proto = 0;
656 struct tcp_hdr *tcp_hdr = NULL;
658 parse_ethernet(m, &l4_proto, &l4_hdr);
659 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
660 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
661 switch (hdr->csum_offset) {
662 case (offsetof(struct tcp_hdr, cksum)):
663 if (l4_proto == IPPROTO_TCP)
664 m->ol_flags |= PKT_TX_TCP_CKSUM;
666 case (offsetof(struct udp_hdr, dgram_cksum)):
667 if (l4_proto == IPPROTO_UDP)
668 m->ol_flags |= PKT_TX_UDP_CKSUM;
670 case (offsetof(struct sctp_hdr, cksum)):
671 if (l4_proto == IPPROTO_SCTP)
672 m->ol_flags |= PKT_TX_SCTP_CKSUM;
680 if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
681 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
682 case VIRTIO_NET_HDR_GSO_TCPV4:
683 case VIRTIO_NET_HDR_GSO_TCPV6:
684 tcp_hdr = (struct tcp_hdr *)l4_hdr;
685 m->ol_flags |= PKT_TX_TCP_SEG;
686 m->tso_segsz = hdr->gso_size;
687 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
690 RTE_LOG(WARNING, VHOST_DATA,
691 "unsupported gso type %u.\n", hdr->gso_type);
697 #define RARP_PKT_SIZE 64
700 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
702 struct ether_hdr *eth_hdr;
703 struct arp_hdr *rarp;
705 if (rarp_mbuf->buf_len < 64) {
706 RTE_LOG(WARNING, VHOST_DATA,
707 "failed to make RARP; mbuf size too small %u (< %d)\n",
708 rarp_mbuf->buf_len, RARP_PKT_SIZE);
712 /* Ethernet header. */
713 eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
714 memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
715 ether_addr_copy(mac, ð_hdr->s_addr);
716 eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
719 rarp = (struct arp_hdr *)(eth_hdr + 1);
720 rarp->arp_hrd = htons(ARP_HRD_ETHER);
721 rarp->arp_pro = htons(ETHER_TYPE_IPv4);
722 rarp->arp_hln = ETHER_ADDR_LEN;
724 rarp->arp_op = htons(ARP_OP_REVREQUEST);
726 ether_addr_copy(mac, &rarp->arp_data.arp_sha);
727 ether_addr_copy(mac, &rarp->arp_data.arp_tha);
728 memset(&rarp->arp_data.arp_sip, 0x00, 4);
729 memset(&rarp->arp_data.arp_tip, 0x00, 4);
731 rarp_mbuf->pkt_len = rarp_mbuf->data_len = RARP_PKT_SIZE;
736 static inline int __attribute__((always_inline))
737 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
738 struct rte_mbuf *m, uint16_t desc_idx,
739 struct rte_mempool *mbuf_pool)
741 struct vring_desc *desc;
743 uint32_t desc_avail, desc_offset;
744 uint32_t mbuf_avail, mbuf_offset;
746 struct rte_mbuf *cur = m, *prev = m;
747 struct virtio_net_hdr *hdr;
748 /* A counter to avoid desc dead loop chain */
749 uint32_t nr_desc = 1;
751 desc = &vq->desc[desc_idx];
752 if (unlikely(desc->len < vq->vhost_hlen))
755 desc_addr = gpa_to_vva(dev, desc->addr);
756 rte_prefetch0((void *)(uintptr_t)desc_addr);
758 /* Retrieve virtio net header */
759 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
760 desc_avail = desc->len - vq->vhost_hlen;
761 desc_offset = vq->vhost_hlen;
764 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
765 while (desc_avail != 0 || (desc->flags & VRING_DESC_F_NEXT) != 0) {
766 /* This desc reaches to its end, get the next one */
767 if (desc_avail == 0) {
768 if (unlikely(desc->next >= vq->size ||
769 ++nr_desc >= vq->size))
771 desc = &vq->desc[desc->next];
773 desc_addr = gpa_to_vva(dev, desc->addr);
774 rte_prefetch0((void *)(uintptr_t)desc_addr);
777 desc_avail = desc->len;
779 PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
783 * This mbuf reaches to its end, get a new one
786 if (mbuf_avail == 0) {
787 cur = rte_pktmbuf_alloc(mbuf_pool);
788 if (unlikely(cur == NULL)) {
789 RTE_LOG(ERR, VHOST_DATA, "Failed to "
790 "allocate memory for mbuf.\n");
795 prev->data_len = mbuf_offset;
797 m->pkt_len += mbuf_offset;
801 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
804 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
805 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *, mbuf_offset),
806 (void *)((uintptr_t)(desc_addr + desc_offset)),
809 mbuf_avail -= cpy_len;
810 mbuf_offset += cpy_len;
811 desc_avail -= cpy_len;
812 desc_offset += cpy_len;
815 prev->data_len = mbuf_offset;
816 m->pkt_len += mbuf_offset;
818 if (hdr->flags != 0 || hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE)
819 vhost_dequeue_offload(hdr, m);
825 rte_vhost_dequeue_burst(struct virtio_net *dev, uint16_t queue_id,
826 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
828 struct rte_mbuf *rarp_mbuf = NULL;
829 struct vhost_virtqueue *vq;
830 uint32_t desc_indexes[MAX_PKT_BURST];
833 uint16_t free_entries;
836 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->virt_qp_nb))) {
837 RTE_LOG(ERR, VHOST_DATA,
838 "%s (%"PRIu64"): virtqueue idx:%d invalid.\n",
839 __func__, dev->device_fh, queue_id);
843 vq = dev->virtqueue[queue_id];
844 if (unlikely(vq->enabled == 0))
848 * Construct a RARP broadcast packet, and inject it to the "pkts"
849 * array, to looks like that guest actually send such packet.
851 * Check user_send_rarp() for more information.
853 if (unlikely(rte_atomic16_cmpset((volatile uint16_t *)
854 &dev->broadcast_rarp.cnt, 1, 0))) {
855 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
856 if (rarp_mbuf == NULL) {
857 RTE_LOG(ERR, VHOST_DATA,
858 "Failed to allocate memory for mbuf.\n");
862 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
863 rte_pktmbuf_free(rarp_mbuf);
870 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
871 free_entries = avail_idx - vq->last_used_idx;
872 if (free_entries == 0)
875 LOG_DEBUG(VHOST_DATA, "%s (%"PRIu64")\n", __func__, dev->device_fh);
877 /* Prefetch available ring to retrieve head indexes. */
878 used_idx = vq->last_used_idx & (vq->size - 1);
879 rte_prefetch0(&vq->avail->ring[used_idx]);
881 count = RTE_MIN(count, MAX_PKT_BURST);
882 count = RTE_MIN(count, free_entries);
883 LOG_DEBUG(VHOST_DATA, "(%"PRIu64") about to dequeue %u buffers\n",
884 dev->device_fh, count);
886 /* Retrieve all of the head indexes first to avoid caching issues. */
887 for (i = 0; i < count; i++) {
888 desc_indexes[i] = vq->avail->ring[(vq->last_used_idx + i) &
892 /* Prefetch descriptor index. */
893 rte_prefetch0(&vq->desc[desc_indexes[0]]);
894 rte_prefetch0(&vq->used->ring[vq->last_used_idx & (vq->size - 1)]);
896 for (i = 0; i < count; i++) {
899 if (likely(i + 1 < count)) {
900 rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
901 rte_prefetch0(&vq->used->ring[(used_idx + 1) &
905 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
906 if (unlikely(pkts[i] == NULL)) {
907 RTE_LOG(ERR, VHOST_DATA,
908 "Failed to allocate memory for mbuf.\n");
911 err = copy_desc_to_mbuf(dev, vq, pkts[i], desc_indexes[i],
914 rte_pktmbuf_free(pkts[i]);
918 used_idx = vq->last_used_idx++ & (vq->size - 1);
919 vq->used->ring[used_idx].id = desc_indexes[i];
920 vq->used->ring[used_idx].len = 0;
921 vhost_log_used_vring(dev, vq,
922 offsetof(struct vring_used, ring[used_idx]),
923 sizeof(vq->used->ring[used_idx]));
926 rte_compiler_barrier();
928 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
929 sizeof(vq->used->idx));
931 /* Kick guest if required. */
932 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT))
933 eventfd_write(vq->callfd, (eventfd_t)1);
936 if (unlikely(rarp_mbuf != NULL)) {
938 * Inject it to the head of "pkts" array, so that switch's mac
939 * learning table will get updated first.
941 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));