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_virtio_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 desc_addr = gpa_to_vva(dev, desc->addr);
152 * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
153 * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
154 * otherwise stores offset on the stack instead of in a register.
156 if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr)
159 rte_prefetch0((void *)(uintptr_t)desc_addr);
161 virtio_enqueue_offload(m, &virtio_hdr.hdr);
162 copy_virtio_net_hdr(dev, desc_addr, virtio_hdr);
163 vhost_log_write(dev, desc->addr, dev->vhost_hlen);
164 PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
166 desc_offset = dev->vhost_hlen;
167 desc_avail = desc->len - dev->vhost_hlen;
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);
191 if (unlikely(!desc_addr))
195 desc_avail = desc->len;
198 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
199 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
200 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
202 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
203 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
206 mbuf_avail -= cpy_len;
207 mbuf_offset += cpy_len;
208 desc_avail -= cpy_len;
209 desc_offset += cpy_len;
216 * This function adds buffers to the virtio devices RX virtqueue. Buffers can
217 * be received from the physical port or from another virtio device. A packet
218 * count is returned to indicate the number of packets that are succesfully
219 * added to the RX queue. This function works when the mbuf is scattered, but
220 * it doesn't support the mergeable feature.
222 static inline uint32_t __attribute__((always_inline))
223 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
224 struct rte_mbuf **pkts, uint32_t count)
226 struct vhost_virtqueue *vq;
227 uint16_t avail_idx, free_entries, start_idx;
228 uint16_t desc_indexes[MAX_PKT_BURST];
232 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
233 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
234 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
235 dev->vid, __func__, queue_id);
239 vq = dev->virtqueue[queue_id];
240 if (unlikely(vq->enabled == 0))
243 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
244 start_idx = vq->last_used_idx;
245 free_entries = avail_idx - start_idx;
246 count = RTE_MIN(count, free_entries);
247 count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
251 LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
252 dev->vid, start_idx, start_idx + count);
254 /* Retrieve all of the desc indexes first to avoid caching issues. */
255 rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
256 for (i = 0; i < count; i++) {
257 used_idx = (start_idx + i) & (vq->size - 1);
258 desc_indexes[i] = vq->avail->ring[used_idx];
259 vq->used->ring[used_idx].id = desc_indexes[i];
260 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
262 vhost_log_used_vring(dev, vq,
263 offsetof(struct vring_used, ring[used_idx]),
264 sizeof(vq->used->ring[used_idx]));
267 rte_prefetch0(&vq->desc[desc_indexes[0]]);
268 for (i = 0; i < count; i++) {
269 uint16_t desc_idx = desc_indexes[i];
272 err = copy_mbuf_to_desc(dev, vq, pkts[i], desc_idx);
274 used_idx = (start_idx + i) & (vq->size - 1);
275 vq->used->ring[used_idx].len = dev->vhost_hlen;
276 vhost_log_used_vring(dev, vq,
277 offsetof(struct vring_used, ring[used_idx]),
278 sizeof(vq->used->ring[used_idx]));
282 rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
287 *(volatile uint16_t *)&vq->used->idx += count;
288 vq->last_used_idx += count;
289 vhost_log_used_vring(dev, vq,
290 offsetof(struct vring_used, idx),
291 sizeof(vq->used->idx));
293 /* flush used->idx update before we read avail->flags. */
296 /* Kick the guest if necessary. */
297 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
298 && (vq->callfd >= 0))
299 eventfd_write(vq->callfd, (eventfd_t)1);
304 fill_vec_buf(struct vhost_virtqueue *vq, uint32_t avail_idx,
305 uint32_t *allocated, uint32_t *vec_idx,
306 struct buf_vector *buf_vec)
308 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
309 uint32_t vec_id = *vec_idx;
310 uint32_t len = *allocated;
313 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
316 len += vq->desc[idx].len;
317 buf_vec[vec_id].buf_addr = vq->desc[idx].addr;
318 buf_vec[vec_id].buf_len = vq->desc[idx].len;
319 buf_vec[vec_id].desc_idx = idx;
322 if ((vq->desc[idx].flags & VRING_DESC_F_NEXT) == 0)
325 idx = vq->desc[idx].next;
335 * Returns -1 on fail, 0 on success
338 reserve_avail_buf_mergeable(struct vhost_virtqueue *vq, uint32_t size,
339 struct buf_vector *buf_vec, uint16_t *num_buffers)
343 uint32_t allocated = 0;
344 uint32_t vec_idx = 0;
347 cur_idx = vq->last_avail_idx;
350 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
351 if (unlikely(cur_idx == avail_idx))
354 if (unlikely(fill_vec_buf(vq, cur_idx, &allocated,
355 &vec_idx, buf_vec) < 0))
361 if (allocated >= size)
365 * if we tried all available ring items, and still
366 * can't get enough buf, it means something abnormal
369 if (unlikely(tries >= vq->size))
373 *num_buffers = cur_idx - vq->last_avail_idx;
377 static inline int __attribute__((always_inline))
378 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
379 struct rte_mbuf *m, struct buf_vector *buf_vec,
380 uint16_t num_buffers)
382 struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
383 uint32_t vec_idx = 0;
384 uint16_t cur_idx = vq->last_used_idx;
386 uint32_t desc_chain_head;
387 uint32_t desc_chain_len;
388 uint32_t mbuf_offset, mbuf_avail;
389 uint32_t desc_offset, desc_avail;
391 uint16_t desc_idx, used_idx;
392 uint64_t hdr_addr, hdr_phys_addr;
393 struct rte_mbuf *hdr_mbuf;
395 if (unlikely(m == NULL))
398 LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
399 dev->vid, cur_idx, cur_idx + num_buffers);
401 desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
402 if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr)
406 hdr_addr = desc_addr;
407 hdr_phys_addr = buf_vec[vec_idx].buf_addr;
408 rte_prefetch0((void *)(uintptr_t)hdr_addr);
410 virtio_hdr.num_buffers = num_buffers;
411 LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
412 dev->vid, virtio_hdr.num_buffers);
414 desc_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
415 desc_offset = dev->vhost_hlen;
416 desc_chain_head = buf_vec[vec_idx].desc_idx;
417 desc_chain_len = desc_offset;
419 mbuf_avail = rte_pktmbuf_data_len(m);
421 while (mbuf_avail != 0 || m->next != NULL) {
422 /* done with current desc buf, get the next one */
423 if (desc_avail == 0) {
424 desc_idx = buf_vec[vec_idx].desc_idx;
427 if (!(vq->desc[desc_idx].flags & VRING_DESC_F_NEXT)) {
428 /* Update used ring with desc information */
429 used_idx = cur_idx++ & (vq->size - 1);
430 vq->used->ring[used_idx].id = desc_chain_head;
431 vq->used->ring[used_idx].len = desc_chain_len;
432 vhost_log_used_vring(dev, vq,
433 offsetof(struct vring_used,
435 sizeof(vq->used->ring[used_idx]));
436 desc_chain_head = buf_vec[vec_idx].desc_idx;
440 desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
441 if (unlikely(!desc_addr))
444 /* Prefetch buffer address. */
445 rte_prefetch0((void *)(uintptr_t)desc_addr);
447 desc_avail = buf_vec[vec_idx].buf_len;
450 /* done with current mbuf, get the next one */
451 if (mbuf_avail == 0) {
455 mbuf_avail = rte_pktmbuf_data_len(m);
459 virtio_enqueue_offload(hdr_mbuf, &virtio_hdr.hdr);
460 copy_virtio_net_hdr(dev, hdr_addr, virtio_hdr);
461 vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
462 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
468 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
469 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
470 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
472 vhost_log_write(dev, buf_vec[vec_idx].buf_addr + desc_offset,
474 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
477 mbuf_avail -= cpy_len;
478 mbuf_offset += cpy_len;
479 desc_avail -= cpy_len;
480 desc_offset += cpy_len;
481 desc_chain_len += cpy_len;
484 used_idx = cur_idx & (vq->size - 1);
485 vq->used->ring[used_idx].id = desc_chain_head;
486 vq->used->ring[used_idx].len = desc_chain_len;
487 vhost_log_used_vring(dev, vq,
488 offsetof(struct vring_used, ring[used_idx]),
489 sizeof(vq->used->ring[used_idx]));
494 static inline uint32_t __attribute__((always_inline))
495 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
496 struct rte_mbuf **pkts, uint32_t count)
498 struct vhost_virtqueue *vq;
499 uint32_t pkt_idx = 0;
500 uint16_t num_buffers;
501 struct buf_vector buf_vec[BUF_VECTOR_MAX];
503 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
504 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
505 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
506 dev->vid, __func__, queue_id);
510 vq = dev->virtqueue[queue_id];
511 if (unlikely(vq->enabled == 0))
514 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
518 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
519 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
521 if (unlikely(reserve_avail_buf_mergeable(vq, pkt_len, buf_vec,
522 &num_buffers) < 0)) {
523 LOG_DEBUG(VHOST_DATA,
524 "(%d) failed to get enough desc from vring\n",
529 if (copy_mbuf_to_desc_mergeable(dev, vq, pkts[pkt_idx],
530 buf_vec, num_buffers) < 0)
535 *(volatile uint16_t *)&vq->used->idx += num_buffers;
536 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
537 sizeof(vq->used->idx));
538 vq->last_used_idx += num_buffers;
539 vq->last_avail_idx += num_buffers;
542 if (likely(pkt_idx)) {
543 /* flush used->idx update before we read avail->flags. */
546 /* Kick the guest if necessary. */
547 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
548 && (vq->callfd >= 0))
549 eventfd_write(vq->callfd, (eventfd_t)1);
556 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
557 struct rte_mbuf **pkts, uint16_t count)
559 struct virtio_net *dev = get_device(vid);
564 if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
565 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
567 return virtio_dev_rx(dev, queue_id, pkts, count);
571 virtio_net_with_host_offload(struct virtio_net *dev)
574 (VIRTIO_NET_F_CSUM | VIRTIO_NET_F_HOST_ECN |
575 VIRTIO_NET_F_HOST_TSO4 | VIRTIO_NET_F_HOST_TSO6 |
576 VIRTIO_NET_F_HOST_UFO))
583 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
585 struct ipv4_hdr *ipv4_hdr;
586 struct ipv6_hdr *ipv6_hdr;
588 struct ether_hdr *eth_hdr;
591 eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
593 m->l2_len = sizeof(struct ether_hdr);
594 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
596 if (ethertype == ETHER_TYPE_VLAN) {
597 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
599 m->l2_len += sizeof(struct vlan_hdr);
600 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
603 l3_hdr = (char *)eth_hdr + m->l2_len;
606 case ETHER_TYPE_IPv4:
607 ipv4_hdr = (struct ipv4_hdr *)l3_hdr;
608 *l4_proto = ipv4_hdr->next_proto_id;
609 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
610 *l4_hdr = (char *)l3_hdr + m->l3_len;
611 m->ol_flags |= PKT_TX_IPV4;
613 case ETHER_TYPE_IPv6:
614 ipv6_hdr = (struct ipv6_hdr *)l3_hdr;
615 *l4_proto = ipv6_hdr->proto;
616 m->l3_len = sizeof(struct ipv6_hdr);
617 *l4_hdr = (char *)l3_hdr + m->l3_len;
618 m->ol_flags |= PKT_TX_IPV6;
627 static inline void __attribute__((always_inline))
628 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
630 uint16_t l4_proto = 0;
632 struct tcp_hdr *tcp_hdr = NULL;
634 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
637 parse_ethernet(m, &l4_proto, &l4_hdr);
638 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
639 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
640 switch (hdr->csum_offset) {
641 case (offsetof(struct tcp_hdr, cksum)):
642 if (l4_proto == IPPROTO_TCP)
643 m->ol_flags |= PKT_TX_TCP_CKSUM;
645 case (offsetof(struct udp_hdr, dgram_cksum)):
646 if (l4_proto == IPPROTO_UDP)
647 m->ol_flags |= PKT_TX_UDP_CKSUM;
649 case (offsetof(struct sctp_hdr, cksum)):
650 if (l4_proto == IPPROTO_SCTP)
651 m->ol_flags |= PKT_TX_SCTP_CKSUM;
659 if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
660 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
661 case VIRTIO_NET_HDR_GSO_TCPV4:
662 case VIRTIO_NET_HDR_GSO_TCPV6:
663 tcp_hdr = (struct tcp_hdr *)l4_hdr;
664 m->ol_flags |= PKT_TX_TCP_SEG;
665 m->tso_segsz = hdr->gso_size;
666 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
669 RTE_LOG(WARNING, VHOST_DATA,
670 "unsupported gso type %u.\n", hdr->gso_type);
676 #define RARP_PKT_SIZE 64
679 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
681 struct ether_hdr *eth_hdr;
682 struct arp_hdr *rarp;
684 if (rarp_mbuf->buf_len < 64) {
685 RTE_LOG(WARNING, VHOST_DATA,
686 "failed to make RARP; mbuf size too small %u (< %d)\n",
687 rarp_mbuf->buf_len, RARP_PKT_SIZE);
691 /* Ethernet header. */
692 eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
693 memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
694 ether_addr_copy(mac, ð_hdr->s_addr);
695 eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
698 rarp = (struct arp_hdr *)(eth_hdr + 1);
699 rarp->arp_hrd = htons(ARP_HRD_ETHER);
700 rarp->arp_pro = htons(ETHER_TYPE_IPv4);
701 rarp->arp_hln = ETHER_ADDR_LEN;
703 rarp->arp_op = htons(ARP_OP_REVREQUEST);
705 ether_addr_copy(mac, &rarp->arp_data.arp_sha);
706 ether_addr_copy(mac, &rarp->arp_data.arp_tha);
707 memset(&rarp->arp_data.arp_sip, 0x00, 4);
708 memset(&rarp->arp_data.arp_tip, 0x00, 4);
710 rarp_mbuf->pkt_len = rarp_mbuf->data_len = RARP_PKT_SIZE;
715 static inline void __attribute__((always_inline))
716 put_zmbuf(struct zcopy_mbuf *zmbuf)
721 static inline int __attribute__((always_inline))
722 copy_desc_to_mbuf(struct virtio_net *dev, struct vring_desc *descs,
723 uint16_t max_desc, struct rte_mbuf *m, uint16_t desc_idx,
724 struct rte_mempool *mbuf_pool)
726 struct vring_desc *desc;
728 uint32_t desc_avail, desc_offset;
729 uint32_t mbuf_avail, mbuf_offset;
731 struct rte_mbuf *cur = m, *prev = m;
732 struct virtio_net_hdr *hdr = NULL;
733 /* A counter to avoid desc dead loop chain */
734 uint32_t nr_desc = 1;
736 desc = &descs[desc_idx];
737 if (unlikely((desc->len < dev->vhost_hlen)) ||
738 (desc->flags & VRING_DESC_F_INDIRECT))
741 desc_addr = gpa_to_vva(dev, desc->addr);
742 if (unlikely(!desc_addr))
745 if (virtio_net_with_host_offload(dev)) {
746 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
751 * A virtio driver normally uses at least 2 desc buffers
752 * for Tx: the first for storing the header, and others
753 * for storing the data.
755 if (likely((desc->len == dev->vhost_hlen) &&
756 (desc->flags & VRING_DESC_F_NEXT) != 0)) {
757 desc = &descs[desc->next];
758 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
761 desc_addr = gpa_to_vva(dev, desc->addr);
762 if (unlikely(!desc_addr))
766 desc_avail = desc->len;
769 desc_avail = desc->len - dev->vhost_hlen;
770 desc_offset = dev->vhost_hlen;
773 rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
775 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
778 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
782 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
785 * A desc buf might across two host physical pages that are
786 * not continuous. In such case (gpa_to_hpa returns 0), data
787 * will be copied even though zero copy is enabled.
789 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
790 desc->addr + desc_offset, cpy_len)))) {
791 cur->data_len = cpy_len;
793 cur->buf_addr = (void *)(uintptr_t)desc_addr;
794 cur->buf_physaddr = hpa;
797 * In zero copy mode, one mbuf can only reference data
798 * for one or partial of one desc buff.
800 mbuf_avail = cpy_len;
802 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
804 (void *)((uintptr_t)(desc_addr + desc_offset)),
808 mbuf_avail -= cpy_len;
809 mbuf_offset += cpy_len;
810 desc_avail -= cpy_len;
811 desc_offset += cpy_len;
813 /* This desc reaches to its end, get the next one */
814 if (desc_avail == 0) {
815 if ((desc->flags & VRING_DESC_F_NEXT) == 0)
818 if (unlikely(desc->next >= max_desc ||
819 ++nr_desc > max_desc))
821 desc = &descs[desc->next];
822 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
825 desc_addr = gpa_to_vva(dev, desc->addr);
826 if (unlikely(!desc_addr))
829 rte_prefetch0((void *)(uintptr_t)desc_addr);
832 desc_avail = desc->len;
834 PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
838 * This mbuf reaches to its end, get a new one
841 if (mbuf_avail == 0) {
842 cur = rte_pktmbuf_alloc(mbuf_pool);
843 if (unlikely(cur == NULL)) {
844 RTE_LOG(ERR, VHOST_DATA, "Failed to "
845 "allocate memory for mbuf.\n");
850 prev->data_len = mbuf_offset;
852 m->pkt_len += mbuf_offset;
856 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
860 prev->data_len = mbuf_offset;
861 m->pkt_len += mbuf_offset;
864 vhost_dequeue_offload(hdr, m);
869 static inline void __attribute__((always_inline))
870 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
871 uint32_t used_idx, uint32_t desc_idx)
873 vq->used->ring[used_idx].id = desc_idx;
874 vq->used->ring[used_idx].len = 0;
875 vhost_log_used_vring(dev, vq,
876 offsetof(struct vring_used, ring[used_idx]),
877 sizeof(vq->used->ring[used_idx]));
880 static inline void __attribute__((always_inline))
881 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
884 if (unlikely(count == 0))
890 vq->used->idx += count;
891 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
892 sizeof(vq->used->idx));
894 /* Kick guest if required. */
895 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
896 && (vq->callfd >= 0))
897 eventfd_write(vq->callfd, (eventfd_t)1);
900 static inline struct zcopy_mbuf *__attribute__((always_inline))
901 get_zmbuf(struct vhost_virtqueue *vq)
907 /* search [last_zmbuf_idx, zmbuf_size) */
908 i = vq->last_zmbuf_idx;
909 last = vq->zmbuf_size;
912 for (; i < last; i++) {
913 if (vq->zmbufs[i].in_use == 0) {
914 vq->last_zmbuf_idx = i + 1;
915 vq->zmbufs[i].in_use = 1;
916 return &vq->zmbufs[i];
922 /* search [0, last_zmbuf_idx) */
924 last = vq->last_zmbuf_idx;
931 static inline bool __attribute__((always_inline))
932 mbuf_is_consumed(struct rte_mbuf *m)
935 if (rte_mbuf_refcnt_read(m) > 1)
944 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
945 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
947 struct virtio_net *dev;
948 struct rte_mbuf *rarp_mbuf = NULL;
949 struct vhost_virtqueue *vq;
950 uint32_t desc_indexes[MAX_PKT_BURST];
953 uint16_t free_entries;
956 dev = get_device(vid);
960 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->virt_qp_nb))) {
961 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
962 dev->vid, __func__, queue_id);
966 vq = dev->virtqueue[queue_id];
967 if (unlikely(vq->enabled == 0))
970 if (unlikely(dev->dequeue_zero_copy)) {
971 struct zcopy_mbuf *zmbuf, *next;
974 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
975 zmbuf != NULL; zmbuf = next) {
976 next = TAILQ_NEXT(zmbuf, next);
978 if (mbuf_is_consumed(zmbuf->mbuf)) {
979 used_idx = vq->last_used_idx++ & (vq->size - 1);
980 update_used_ring(dev, vq, used_idx,
984 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
985 rte_pktmbuf_free(zmbuf->mbuf);
991 update_used_idx(dev, vq, nr_updated);
995 * Construct a RARP broadcast packet, and inject it to the "pkts"
996 * array, to looks like that guest actually send such packet.
998 * Check user_send_rarp() for more information.
1000 if (unlikely(rte_atomic16_cmpset((volatile uint16_t *)
1001 &dev->broadcast_rarp.cnt, 1, 0))) {
1002 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
1003 if (rarp_mbuf == NULL) {
1004 RTE_LOG(ERR, VHOST_DATA,
1005 "Failed to allocate memory for mbuf.\n");
1009 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
1010 rte_pktmbuf_free(rarp_mbuf);
1017 free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1019 if (free_entries == 0)
1022 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1024 /* Prefetch available and used ring */
1025 avail_idx = vq->last_avail_idx & (vq->size - 1);
1026 used_idx = vq->last_used_idx & (vq->size - 1);
1027 rte_prefetch0(&vq->avail->ring[avail_idx]);
1028 rte_prefetch0(&vq->used->ring[used_idx]);
1030 count = RTE_MIN(count, MAX_PKT_BURST);
1031 count = RTE_MIN(count, free_entries);
1032 LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1035 /* Retrieve all of the head indexes first to avoid caching issues. */
1036 for (i = 0; i < count; i++) {
1037 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1038 used_idx = (vq->last_used_idx + i) & (vq->size - 1);
1039 desc_indexes[i] = vq->avail->ring[avail_idx];
1041 if (likely(dev->dequeue_zero_copy == 0))
1042 update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1045 /* Prefetch descriptor index. */
1046 rte_prefetch0(&vq->desc[desc_indexes[0]]);
1047 for (i = 0; i < count; i++) {
1048 struct vring_desc *desc;
1052 if (likely(i + 1 < count))
1053 rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1055 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1056 desc = (struct vring_desc *)(uintptr_t)gpa_to_vva(dev,
1057 vq->desc[desc_indexes[i]].addr);
1058 if (unlikely(!desc))
1061 rte_prefetch0(desc);
1062 sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1067 idx = desc_indexes[i];
1070 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1071 if (unlikely(pkts[i] == NULL)) {
1072 RTE_LOG(ERR, VHOST_DATA,
1073 "Failed to allocate memory for mbuf.\n");
1077 err = copy_desc_to_mbuf(dev, desc, sz, pkts[i], idx, mbuf_pool);
1078 if (unlikely(err)) {
1079 rte_pktmbuf_free(pkts[i]);
1083 if (unlikely(dev->dequeue_zero_copy)) {
1084 struct zcopy_mbuf *zmbuf;
1086 zmbuf = get_zmbuf(vq);
1088 rte_pktmbuf_free(pkts[i]);
1091 zmbuf->mbuf = pkts[i];
1092 zmbuf->desc_idx = desc_indexes[i];
1095 * Pin lock the mbuf; we will check later to see
1096 * whether the mbuf is freed (when we are the last
1097 * user) or not. If that's the case, we then could
1098 * update the used ring safely.
1100 rte_mbuf_refcnt_update(pkts[i], 1);
1103 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1106 vq->last_avail_idx += i;
1108 if (likely(dev->dequeue_zero_copy == 0)) {
1109 vq->last_used_idx += i;
1110 update_used_idx(dev, vq, i);
1114 if (unlikely(rarp_mbuf != NULL)) {
1116 * Inject it to the head of "pkts" array, so that switch's mac
1117 * learning table will get updated first.
1119 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1120 pkts[0] = rarp_mbuf;