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 if (m_buf->ol_flags & PKT_TX_L4_MASK) {
118 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
119 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
121 switch (m_buf->ol_flags & PKT_TX_L4_MASK) {
122 case PKT_TX_TCP_CKSUM:
123 net_hdr->csum_offset = (offsetof(struct tcp_hdr,
126 case PKT_TX_UDP_CKSUM:
127 net_hdr->csum_offset = (offsetof(struct udp_hdr,
130 case PKT_TX_SCTP_CKSUM:
131 net_hdr->csum_offset = (offsetof(struct sctp_hdr,
136 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
137 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
138 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
141 if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
142 if (m_buf->ol_flags & PKT_TX_IPV4)
143 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
145 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
146 net_hdr->gso_size = m_buf->tso_segsz;
147 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
150 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
151 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
152 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
156 static __rte_always_inline int
157 copy_mbuf_to_desc(struct virtio_net *dev, struct vring_desc *descs,
158 struct rte_mbuf *m, uint16_t desc_idx, uint32_t size)
160 uint32_t desc_avail, desc_offset;
161 uint32_t mbuf_avail, mbuf_offset;
163 struct vring_desc *desc;
165 /* A counter to avoid desc dead loop chain */
166 uint16_t nr_desc = 1;
168 desc = &descs[desc_idx];
169 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
171 * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
172 * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
173 * otherwise stores offset on the stack instead of in a register.
175 if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr)
178 rte_prefetch0((void *)(uintptr_t)desc_addr);
180 virtio_enqueue_offload(m, (struct virtio_net_hdr *)(uintptr_t)desc_addr);
181 vhost_log_write(dev, desc->addr, dev->vhost_hlen);
182 PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
184 desc_offset = dev->vhost_hlen;
185 desc_avail = desc->len - dev->vhost_hlen;
187 mbuf_avail = rte_pktmbuf_data_len(m);
189 while (mbuf_avail != 0 || m->next != NULL) {
190 /* done with current mbuf, fetch next */
191 if (mbuf_avail == 0) {
195 mbuf_avail = rte_pktmbuf_data_len(m);
198 /* done with current desc buf, fetch next */
199 if (desc_avail == 0) {
200 if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
201 /* Room in vring buffer is not enough */
204 if (unlikely(desc->next >= size || ++nr_desc > size))
207 desc = &descs[desc->next];
208 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
209 if (unlikely(!desc_addr))
213 desc_avail = desc->len;
216 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
217 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
218 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
220 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
221 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
224 mbuf_avail -= cpy_len;
225 mbuf_offset += cpy_len;
226 desc_avail -= cpy_len;
227 desc_offset += cpy_len;
234 * This function adds buffers to the virtio devices RX virtqueue. Buffers can
235 * be received from the physical port or from another virtio device. A packet
236 * count is returned to indicate the number of packets that are successfully
237 * added to the RX queue. This function works when the mbuf is scattered, but
238 * it doesn't support the mergeable feature.
240 static __rte_always_inline uint32_t
241 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
242 struct rte_mbuf **pkts, uint32_t count)
244 struct vhost_virtqueue *vq;
245 uint16_t avail_idx, free_entries, start_idx;
246 uint16_t desc_indexes[MAX_PKT_BURST];
247 struct vring_desc *descs;
251 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
252 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
253 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
254 dev->vid, __func__, queue_id);
258 vq = dev->virtqueue[queue_id];
259 if (unlikely(vq->enabled == 0))
262 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
263 start_idx = vq->last_used_idx;
264 free_entries = avail_idx - start_idx;
265 count = RTE_MIN(count, free_entries);
266 count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
270 LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
271 dev->vid, start_idx, start_idx + count);
273 /* Retrieve all of the desc indexes first to avoid caching issues. */
274 rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
275 for (i = 0; i < count; i++) {
276 used_idx = (start_idx + i) & (vq->size - 1);
277 desc_indexes[i] = vq->avail->ring[used_idx];
278 vq->used->ring[used_idx].id = desc_indexes[i];
279 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
281 vhost_log_used_vring(dev, vq,
282 offsetof(struct vring_used, ring[used_idx]),
283 sizeof(vq->used->ring[used_idx]));
286 rte_prefetch0(&vq->desc[desc_indexes[0]]);
287 for (i = 0; i < count; i++) {
288 uint16_t desc_idx = desc_indexes[i];
291 if (vq->desc[desc_idx].flags & VRING_DESC_F_INDIRECT) {
292 descs = (struct vring_desc *)(uintptr_t)
293 rte_vhost_gpa_to_vva(dev->mem,
294 vq->desc[desc_idx].addr);
295 if (unlikely(!descs)) {
301 sz = vq->desc[desc_idx].len / sizeof(*descs);
307 err = copy_mbuf_to_desc(dev, descs, pkts[i], desc_idx, sz);
309 used_idx = (start_idx + i) & (vq->size - 1);
310 vq->used->ring[used_idx].len = dev->vhost_hlen;
311 vhost_log_used_vring(dev, vq,
312 offsetof(struct vring_used, ring[used_idx]),
313 sizeof(vq->used->ring[used_idx]));
317 rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
322 *(volatile uint16_t *)&vq->used->idx += count;
323 vq->last_used_idx += count;
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);
338 static __rte_always_inline int
339 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
340 uint32_t avail_idx, uint32_t *vec_idx,
341 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
342 uint16_t *desc_chain_len)
344 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
345 uint32_t vec_id = *vec_idx;
347 struct vring_desc *descs = vq->desc;
349 *desc_chain_head = idx;
351 if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
352 descs = (struct vring_desc *)(uintptr_t)
353 rte_vhost_gpa_to_vva(dev->mem, vq->desc[idx].addr);
354 if (unlikely(!descs))
361 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
364 len += descs[idx].len;
365 buf_vec[vec_id].buf_addr = descs[idx].addr;
366 buf_vec[vec_id].buf_len = descs[idx].len;
367 buf_vec[vec_id].desc_idx = idx;
370 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
373 idx = descs[idx].next;
376 *desc_chain_len = len;
383 * Returns -1 on fail, 0 on success
386 reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
387 uint32_t size, struct buf_vector *buf_vec,
388 uint16_t *num_buffers, uint16_t avail_head)
391 uint32_t vec_idx = 0;
394 uint16_t head_idx = 0;
398 cur_idx = vq->last_avail_idx;
401 if (unlikely(cur_idx == avail_head))
404 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
405 &head_idx, &len) < 0))
407 len = RTE_MIN(len, size);
408 update_shadow_used_ring(vq, head_idx, len);
416 * if we tried all available ring items, and still
417 * can't get enough buf, it means something abnormal
420 if (unlikely(tries >= vq->size))
427 static __rte_always_inline int
428 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct rte_mbuf *m,
429 struct buf_vector *buf_vec, uint16_t num_buffers)
431 uint32_t vec_idx = 0;
433 uint32_t mbuf_offset, mbuf_avail;
434 uint32_t desc_offset, desc_avail;
436 uint64_t hdr_addr, hdr_phys_addr;
437 struct rte_mbuf *hdr_mbuf;
439 if (unlikely(m == NULL))
442 desc_addr = rte_vhost_gpa_to_vva(dev->mem, buf_vec[vec_idx].buf_addr);
443 if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr)
447 hdr_addr = desc_addr;
448 hdr_phys_addr = buf_vec[vec_idx].buf_addr;
449 rte_prefetch0((void *)(uintptr_t)hdr_addr);
451 LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
452 dev->vid, num_buffers);
454 desc_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
455 desc_offset = dev->vhost_hlen;
457 mbuf_avail = rte_pktmbuf_data_len(m);
459 while (mbuf_avail != 0 || m->next != NULL) {
460 /* done with current desc buf, get the next one */
461 if (desc_avail == 0) {
463 desc_addr = rte_vhost_gpa_to_vva(dev->mem,
464 buf_vec[vec_idx].buf_addr);
465 if (unlikely(!desc_addr))
468 /* Prefetch buffer address. */
469 rte_prefetch0((void *)(uintptr_t)desc_addr);
471 desc_avail = buf_vec[vec_idx].buf_len;
474 /* done with current mbuf, get the next one */
475 if (mbuf_avail == 0) {
479 mbuf_avail = rte_pktmbuf_data_len(m);
483 struct virtio_net_hdr_mrg_rxbuf *hdr;
485 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)
487 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
488 ASSIGN_UNLESS_EQUAL(hdr->num_buffers, num_buffers);
490 vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
491 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
497 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
498 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
499 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
501 vhost_log_write(dev, buf_vec[vec_idx].buf_addr + desc_offset,
503 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
506 mbuf_avail -= cpy_len;
507 mbuf_offset += cpy_len;
508 desc_avail -= cpy_len;
509 desc_offset += cpy_len;
515 static __rte_always_inline uint32_t
516 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
517 struct rte_mbuf **pkts, uint32_t count)
519 struct vhost_virtqueue *vq;
520 uint32_t pkt_idx = 0;
521 uint16_t num_buffers;
522 struct buf_vector buf_vec[BUF_VECTOR_MAX];
525 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
526 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
527 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
528 dev->vid, __func__, queue_id);
532 vq = dev->virtqueue[queue_id];
533 if (unlikely(vq->enabled == 0))
536 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
540 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
542 vq->shadow_used_idx = 0;
543 avail_head = *((volatile uint16_t *)&vq->avail->idx);
544 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
545 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
547 if (unlikely(reserve_avail_buf_mergeable(dev, vq,
548 pkt_len, buf_vec, &num_buffers,
550 LOG_DEBUG(VHOST_DATA,
551 "(%d) failed to get enough desc from vring\n",
553 vq->shadow_used_idx -= num_buffers;
557 LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
558 dev->vid, vq->last_avail_idx,
559 vq->last_avail_idx + num_buffers);
561 if (copy_mbuf_to_desc_mergeable(dev, pkts[pkt_idx],
562 buf_vec, num_buffers) < 0) {
563 vq->shadow_used_idx -= num_buffers;
567 vq->last_avail_idx += num_buffers;
570 if (likely(vq->shadow_used_idx)) {
571 flush_shadow_used_ring(dev, vq);
573 /* flush used->idx update before we read avail->flags. */
576 /* Kick the guest if necessary. */
577 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
578 && (vq->callfd >= 0))
579 eventfd_write(vq->callfd, (eventfd_t)1);
586 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
587 struct rte_mbuf **pkts, uint16_t count)
589 struct virtio_net *dev = get_device(vid);
594 if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
595 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
597 return virtio_dev_rx(dev, queue_id, pkts, count);
601 virtio_net_with_host_offload(struct virtio_net *dev)
604 (VIRTIO_NET_F_CSUM | VIRTIO_NET_F_HOST_ECN |
605 VIRTIO_NET_F_HOST_TSO4 | VIRTIO_NET_F_HOST_TSO6 |
606 VIRTIO_NET_F_HOST_UFO))
613 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
615 struct ipv4_hdr *ipv4_hdr;
616 struct ipv6_hdr *ipv6_hdr;
618 struct ether_hdr *eth_hdr;
621 eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
623 m->l2_len = sizeof(struct ether_hdr);
624 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
626 if (ethertype == ETHER_TYPE_VLAN) {
627 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
629 m->l2_len += sizeof(struct vlan_hdr);
630 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
633 l3_hdr = (char *)eth_hdr + m->l2_len;
636 case ETHER_TYPE_IPv4:
638 *l4_proto = ipv4_hdr->next_proto_id;
639 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
640 *l4_hdr = (char *)l3_hdr + m->l3_len;
641 m->ol_flags |= PKT_TX_IPV4;
643 case ETHER_TYPE_IPv6:
645 *l4_proto = ipv6_hdr->proto;
646 m->l3_len = sizeof(struct ipv6_hdr);
647 *l4_hdr = (char *)l3_hdr + m->l3_len;
648 m->ol_flags |= PKT_TX_IPV6;
658 static __rte_always_inline void
659 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
661 uint16_t l4_proto = 0;
663 struct tcp_hdr *tcp_hdr = NULL;
665 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
668 parse_ethernet(m, &l4_proto, &l4_hdr);
669 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
670 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
671 switch (hdr->csum_offset) {
672 case (offsetof(struct tcp_hdr, cksum)):
673 if (l4_proto == IPPROTO_TCP)
674 m->ol_flags |= PKT_TX_TCP_CKSUM;
676 case (offsetof(struct udp_hdr, dgram_cksum)):
677 if (l4_proto == IPPROTO_UDP)
678 m->ol_flags |= PKT_TX_UDP_CKSUM;
680 case (offsetof(struct sctp_hdr, cksum)):
681 if (l4_proto == IPPROTO_SCTP)
682 m->ol_flags |= PKT_TX_SCTP_CKSUM;
690 if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
691 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
692 case VIRTIO_NET_HDR_GSO_TCPV4:
693 case VIRTIO_NET_HDR_GSO_TCPV6:
695 m->ol_flags |= PKT_TX_TCP_SEG;
696 m->tso_segsz = hdr->gso_size;
697 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
700 RTE_LOG(WARNING, VHOST_DATA,
701 "unsupported gso type %u.\n", hdr->gso_type);
707 #define RARP_PKT_SIZE 64
710 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
712 struct ether_hdr *eth_hdr;
713 struct arp_hdr *rarp;
715 if (rarp_mbuf->buf_len < 64) {
716 RTE_LOG(WARNING, VHOST_DATA,
717 "failed to make RARP; mbuf size too small %u (< %d)\n",
718 rarp_mbuf->buf_len, RARP_PKT_SIZE);
722 /* Ethernet header. */
723 eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
724 memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
725 ether_addr_copy(mac, ð_hdr->s_addr);
726 eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
729 rarp = (struct arp_hdr *)(eth_hdr + 1);
730 rarp->arp_hrd = htons(ARP_HRD_ETHER);
731 rarp->arp_pro = htons(ETHER_TYPE_IPv4);
732 rarp->arp_hln = ETHER_ADDR_LEN;
734 rarp->arp_op = htons(ARP_OP_REVREQUEST);
736 ether_addr_copy(mac, &rarp->arp_data.arp_sha);
737 ether_addr_copy(mac, &rarp->arp_data.arp_tha);
738 memset(&rarp->arp_data.arp_sip, 0x00, 4);
739 memset(&rarp->arp_data.arp_tip, 0x00, 4);
741 rarp_mbuf->pkt_len = rarp_mbuf->data_len = RARP_PKT_SIZE;
746 static __rte_always_inline void
747 put_zmbuf(struct zcopy_mbuf *zmbuf)
752 static __rte_always_inline int
753 copy_desc_to_mbuf(struct virtio_net *dev, struct vring_desc *descs,
754 uint16_t max_desc, struct rte_mbuf *m, uint16_t desc_idx,
755 struct rte_mempool *mbuf_pool)
757 struct vring_desc *desc;
759 uint32_t desc_avail, desc_offset;
760 uint32_t mbuf_avail, mbuf_offset;
762 struct rte_mbuf *cur = m, *prev = m;
763 struct virtio_net_hdr *hdr = NULL;
764 /* A counter to avoid desc dead loop chain */
765 uint32_t nr_desc = 1;
767 desc = &descs[desc_idx];
768 if (unlikely((desc->len < dev->vhost_hlen)) ||
769 (desc->flags & VRING_DESC_F_INDIRECT))
772 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
773 if (unlikely(!desc_addr))
776 if (virtio_net_with_host_offload(dev)) {
777 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
782 * A virtio driver normally uses at least 2 desc buffers
783 * for Tx: the first for storing the header, and others
784 * for storing the data.
786 if (likely((desc->len == dev->vhost_hlen) &&
787 (desc->flags & VRING_DESC_F_NEXT) != 0)) {
788 desc = &descs[desc->next];
789 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
792 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
793 if (unlikely(!desc_addr))
797 desc_avail = desc->len;
800 desc_avail = desc->len - dev->vhost_hlen;
801 desc_offset = dev->vhost_hlen;
804 rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
806 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
809 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
813 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
816 * A desc buf might across two host physical pages that are
817 * not continuous. In such case (gpa_to_hpa returns 0), data
818 * will be copied even though zero copy is enabled.
820 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
821 desc->addr + desc_offset, cpy_len)))) {
822 cur->data_len = cpy_len;
824 cur->buf_addr = (void *)(uintptr_t)desc_addr;
825 cur->buf_physaddr = hpa;
828 * In zero copy mode, one mbuf can only reference data
829 * for one or partial of one desc buff.
831 mbuf_avail = cpy_len;
833 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
835 (void *)((uintptr_t)(desc_addr + desc_offset)),
839 mbuf_avail -= cpy_len;
840 mbuf_offset += cpy_len;
841 desc_avail -= cpy_len;
842 desc_offset += cpy_len;
844 /* This desc reaches to its end, get the next one */
845 if (desc_avail == 0) {
846 if ((desc->flags & VRING_DESC_F_NEXT) == 0)
849 if (unlikely(desc->next >= max_desc ||
850 ++nr_desc > max_desc))
852 desc = &descs[desc->next];
853 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
856 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
857 if (unlikely(!desc_addr))
860 rte_prefetch0((void *)(uintptr_t)desc_addr);
863 desc_avail = desc->len;
865 PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
869 * This mbuf reaches to its end, get a new one
872 if (mbuf_avail == 0) {
873 cur = rte_pktmbuf_alloc(mbuf_pool);
874 if (unlikely(cur == NULL)) {
875 RTE_LOG(ERR, VHOST_DATA, "Failed to "
876 "allocate memory for mbuf.\n");
879 if (unlikely(dev->dequeue_zero_copy))
880 rte_mbuf_refcnt_update(cur, 1);
883 prev->data_len = mbuf_offset;
885 m->pkt_len += mbuf_offset;
889 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
893 prev->data_len = mbuf_offset;
894 m->pkt_len += mbuf_offset;
897 vhost_dequeue_offload(hdr, m);
902 static __rte_always_inline void
903 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
904 uint32_t used_idx, uint32_t desc_idx)
906 vq->used->ring[used_idx].id = desc_idx;
907 vq->used->ring[used_idx].len = 0;
908 vhost_log_used_vring(dev, vq,
909 offsetof(struct vring_used, ring[used_idx]),
910 sizeof(vq->used->ring[used_idx]));
913 static __rte_always_inline void
914 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
917 if (unlikely(count == 0))
923 vq->used->idx += count;
924 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
925 sizeof(vq->used->idx));
927 /* Kick guest if required. */
928 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
929 && (vq->callfd >= 0))
930 eventfd_write(vq->callfd, (eventfd_t)1);
933 static __rte_always_inline struct zcopy_mbuf *
934 get_zmbuf(struct vhost_virtqueue *vq)
940 /* search [last_zmbuf_idx, zmbuf_size) */
941 i = vq->last_zmbuf_idx;
942 last = vq->zmbuf_size;
945 for (; i < last; i++) {
946 if (vq->zmbufs[i].in_use == 0) {
947 vq->last_zmbuf_idx = i + 1;
948 vq->zmbufs[i].in_use = 1;
949 return &vq->zmbufs[i];
955 /* search [0, last_zmbuf_idx) */
957 last = vq->last_zmbuf_idx;
964 static __rte_always_inline bool
965 mbuf_is_consumed(struct rte_mbuf *m)
968 if (rte_mbuf_refcnt_read(m) > 1)
977 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
978 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
980 struct virtio_net *dev;
981 struct rte_mbuf *rarp_mbuf = NULL;
982 struct vhost_virtqueue *vq;
983 uint32_t desc_indexes[MAX_PKT_BURST];
986 uint16_t free_entries;
989 dev = get_device(vid);
993 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
994 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
995 dev->vid, __func__, queue_id);
999 vq = dev->virtqueue[queue_id];
1000 if (unlikely(vq->enabled == 0))
1003 if (unlikely(dev->dequeue_zero_copy)) {
1004 struct zcopy_mbuf *zmbuf, *next;
1007 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1008 zmbuf != NULL; zmbuf = next) {
1009 next = TAILQ_NEXT(zmbuf, next);
1011 if (mbuf_is_consumed(zmbuf->mbuf)) {
1012 used_idx = vq->last_used_idx++ & (vq->size - 1);
1013 update_used_ring(dev, vq, used_idx,
1017 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1018 rte_pktmbuf_free(zmbuf->mbuf);
1024 update_used_idx(dev, vq, nr_updated);
1028 * Construct a RARP broadcast packet, and inject it to the "pkts"
1029 * array, to looks like that guest actually send such packet.
1031 * Check user_send_rarp() for more information.
1033 * broadcast_rarp shares a cacheline in the virtio_net structure
1034 * with some fields that are accessed during enqueue and
1035 * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1036 * result in false sharing between enqueue and dequeue.
1038 * Prevent unnecessary false sharing by reading broadcast_rarp first
1039 * and only performing cmpset if the read indicates it is likely to
1043 if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1044 rte_atomic16_cmpset((volatile uint16_t *)
1045 &dev->broadcast_rarp.cnt, 1, 0))) {
1047 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
1048 if (rarp_mbuf == NULL) {
1049 RTE_LOG(ERR, VHOST_DATA,
1050 "Failed to allocate memory for mbuf.\n");
1054 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
1055 rte_pktmbuf_free(rarp_mbuf);
1062 free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1064 if (free_entries == 0)
1067 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1069 /* Prefetch available and used ring */
1070 avail_idx = vq->last_avail_idx & (vq->size - 1);
1071 used_idx = vq->last_used_idx & (vq->size - 1);
1072 rte_prefetch0(&vq->avail->ring[avail_idx]);
1073 rte_prefetch0(&vq->used->ring[used_idx]);
1075 count = RTE_MIN(count, MAX_PKT_BURST);
1076 count = RTE_MIN(count, free_entries);
1077 LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1080 /* Retrieve all of the head indexes first to avoid caching issues. */
1081 for (i = 0; i < count; i++) {
1082 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1083 used_idx = (vq->last_used_idx + i) & (vq->size - 1);
1084 desc_indexes[i] = vq->avail->ring[avail_idx];
1086 if (likely(dev->dequeue_zero_copy == 0))
1087 update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1090 /* Prefetch descriptor index. */
1091 rte_prefetch0(&vq->desc[desc_indexes[0]]);
1092 for (i = 0; i < count; i++) {
1093 struct vring_desc *desc;
1097 if (likely(i + 1 < count))
1098 rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1100 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1101 desc = (struct vring_desc *)(uintptr_t)
1102 rte_vhost_gpa_to_vva(dev->mem,
1103 vq->desc[desc_indexes[i]].addr);
1104 if (unlikely(!desc))
1107 rte_prefetch0(desc);
1108 sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1113 idx = desc_indexes[i];
1116 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1117 if (unlikely(pkts[i] == NULL)) {
1118 RTE_LOG(ERR, VHOST_DATA,
1119 "Failed to allocate memory for mbuf.\n");
1123 err = copy_desc_to_mbuf(dev, desc, sz, pkts[i], idx, mbuf_pool);
1124 if (unlikely(err)) {
1125 rte_pktmbuf_free(pkts[i]);
1129 if (unlikely(dev->dequeue_zero_copy)) {
1130 struct zcopy_mbuf *zmbuf;
1132 zmbuf = get_zmbuf(vq);
1134 rte_pktmbuf_free(pkts[i]);
1137 zmbuf->mbuf = pkts[i];
1138 zmbuf->desc_idx = desc_indexes[i];
1141 * Pin lock the mbuf; we will check later to see
1142 * whether the mbuf is freed (when we are the last
1143 * user) or not. If that's the case, we then could
1144 * update the used ring safely.
1146 rte_mbuf_refcnt_update(pkts[i], 1);
1149 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1152 vq->last_avail_idx += i;
1154 if (likely(dev->dequeue_zero_copy == 0)) {
1155 vq->last_used_idx += i;
1156 update_used_idx(dev, vq, i);
1160 if (unlikely(rarp_mbuf != NULL)) {
1162 * Inject it to the head of "pkts" array, so that switch's mac
1163 * learning table will get updated first.
1165 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1166 pkts[0] = rarp_mbuf;