4 * Copyright(c) 2010-2016 Intel Corporation. All rights reserved.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * * Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * * Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * * Neither the name of Intel Corporation nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
36 #include <linux/virtio_net.h>
39 #include <rte_memcpy.h>
40 #include <rte_ether.h>
42 #include <rte_vhost.h>
50 #define MAX_PKT_BURST 32
53 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
55 return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
58 static __rte_always_inline void
59 do_flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
60 uint16_t to, uint16_t from, uint16_t size)
62 rte_memcpy(&vq->used->ring[to],
63 &vq->shadow_used_ring[from],
64 size * sizeof(struct vring_used_elem));
65 vhost_log_used_vring(dev, vq,
66 offsetof(struct vring_used, ring[to]),
67 size * sizeof(struct vring_used_elem));
70 static __rte_always_inline void
71 flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq)
73 uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
75 if (used_idx + vq->shadow_used_idx <= vq->size) {
76 do_flush_shadow_used_ring(dev, vq, used_idx, 0,
81 /* update used ring interval [used_idx, vq->size] */
82 size = vq->size - used_idx;
83 do_flush_shadow_used_ring(dev, vq, used_idx, 0, size);
85 /* update the left half used ring interval [0, left_size] */
86 do_flush_shadow_used_ring(dev, vq, 0, size,
87 vq->shadow_used_idx - size);
89 vq->last_used_idx += vq->shadow_used_idx;
93 *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
94 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
95 sizeof(vq->used->idx));
98 static __rte_always_inline void
99 update_shadow_used_ring(struct vhost_virtqueue *vq,
100 uint16_t desc_idx, uint16_t len)
102 uint16_t i = vq->shadow_used_idx++;
104 vq->shadow_used_ring[i].id = desc_idx;
105 vq->shadow_used_ring[i].len = len;
108 /* avoid write operation when necessary, to lessen cache issues */
109 #define ASSIGN_UNLESS_EQUAL(var, val) do { \
110 if ((var) != (val)) \
115 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
117 uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
119 if (m_buf->ol_flags & PKT_TX_TCP_SEG)
120 csum_l4 |= PKT_TX_TCP_CKSUM;
123 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
124 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
127 case PKT_TX_TCP_CKSUM:
128 net_hdr->csum_offset = (offsetof(struct tcp_hdr,
131 case PKT_TX_UDP_CKSUM:
132 net_hdr->csum_offset = (offsetof(struct udp_hdr,
135 case PKT_TX_SCTP_CKSUM:
136 net_hdr->csum_offset = (offsetof(struct sctp_hdr,
141 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
142 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
143 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
146 if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
147 if (m_buf->ol_flags & PKT_TX_IPV4)
148 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
150 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
151 net_hdr->gso_size = m_buf->tso_segsz;
152 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
155 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
156 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
157 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
161 static __rte_always_inline int
162 copy_mbuf_to_desc(struct virtio_net *dev, struct vring_desc *descs,
163 struct rte_mbuf *m, uint16_t desc_idx, uint32_t size)
165 uint32_t desc_avail, desc_offset;
166 uint32_t mbuf_avail, mbuf_offset;
168 struct vring_desc *desc;
170 /* A counter to avoid desc dead loop chain */
171 uint16_t nr_desc = 1;
173 desc = &descs[desc_idx];
174 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
176 * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
177 * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
178 * otherwise stores offset on the stack instead of in a register.
180 if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr)
183 rte_prefetch0((void *)(uintptr_t)desc_addr);
185 virtio_enqueue_offload(m, (struct virtio_net_hdr *)(uintptr_t)desc_addr);
186 vhost_log_write(dev, desc->addr, dev->vhost_hlen);
187 PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
189 desc_offset = dev->vhost_hlen;
190 desc_avail = desc->len - dev->vhost_hlen;
192 mbuf_avail = rte_pktmbuf_data_len(m);
194 while (mbuf_avail != 0 || m->next != NULL) {
195 /* done with current mbuf, fetch next */
196 if (mbuf_avail == 0) {
200 mbuf_avail = rte_pktmbuf_data_len(m);
203 /* done with current desc buf, fetch next */
204 if (desc_avail == 0) {
205 if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
206 /* Room in vring buffer is not enough */
209 if (unlikely(desc->next >= size || ++nr_desc > size))
212 desc = &descs[desc->next];
213 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
214 if (unlikely(!desc_addr))
218 desc_avail = desc->len;
221 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
222 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
223 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
225 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
226 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
229 mbuf_avail -= cpy_len;
230 mbuf_offset += cpy_len;
231 desc_avail -= cpy_len;
232 desc_offset += cpy_len;
239 * This function adds buffers to the virtio devices RX virtqueue. Buffers can
240 * be received from the physical port or from another virtio device. A packet
241 * count is returned to indicate the number of packets that are successfully
242 * added to the RX queue. This function works when the mbuf is scattered, but
243 * it doesn't support the mergeable feature.
245 static __rte_always_inline uint32_t
246 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
247 struct rte_mbuf **pkts, uint32_t count)
249 struct vhost_virtqueue *vq;
250 uint16_t avail_idx, free_entries, start_idx;
251 uint16_t desc_indexes[MAX_PKT_BURST];
252 struct vring_desc *descs;
256 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
257 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
258 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
259 dev->vid, __func__, queue_id);
263 vq = dev->virtqueue[queue_id];
264 if (unlikely(vq->enabled == 0))
267 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
268 start_idx = vq->last_used_idx;
269 free_entries = avail_idx - start_idx;
270 count = RTE_MIN(count, free_entries);
271 count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
275 LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
276 dev->vid, start_idx, start_idx + count);
278 /* Retrieve all of the desc indexes first to avoid caching issues. */
279 rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
280 for (i = 0; i < count; i++) {
281 used_idx = (start_idx + i) & (vq->size - 1);
282 desc_indexes[i] = vq->avail->ring[used_idx];
283 vq->used->ring[used_idx].id = desc_indexes[i];
284 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
286 vhost_log_used_vring(dev, vq,
287 offsetof(struct vring_used, ring[used_idx]),
288 sizeof(vq->used->ring[used_idx]));
291 rte_prefetch0(&vq->desc[desc_indexes[0]]);
292 for (i = 0; i < count; i++) {
293 uint16_t desc_idx = desc_indexes[i];
296 if (vq->desc[desc_idx].flags & VRING_DESC_F_INDIRECT) {
297 descs = (struct vring_desc *)(uintptr_t)
298 rte_vhost_gpa_to_vva(dev->mem,
299 vq->desc[desc_idx].addr);
300 if (unlikely(!descs)) {
306 sz = vq->desc[desc_idx].len / sizeof(*descs);
312 err = copy_mbuf_to_desc(dev, descs, pkts[i], desc_idx, sz);
314 used_idx = (start_idx + i) & (vq->size - 1);
315 vq->used->ring[used_idx].len = dev->vhost_hlen;
316 vhost_log_used_vring(dev, vq,
317 offsetof(struct vring_used, ring[used_idx]),
318 sizeof(vq->used->ring[used_idx]));
322 rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
327 *(volatile uint16_t *)&vq->used->idx += count;
328 vq->last_used_idx += count;
329 vhost_log_used_vring(dev, vq,
330 offsetof(struct vring_used, idx),
331 sizeof(vq->used->idx));
333 /* flush used->idx update before we read avail->flags. */
336 /* Kick the guest if necessary. */
337 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
338 && (vq->callfd >= 0))
339 eventfd_write(vq->callfd, (eventfd_t)1);
343 static __rte_always_inline int
344 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
345 uint32_t avail_idx, uint32_t *vec_idx,
346 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
347 uint16_t *desc_chain_len)
349 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
350 uint32_t vec_id = *vec_idx;
352 struct vring_desc *descs = vq->desc;
354 *desc_chain_head = idx;
356 if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
357 descs = (struct vring_desc *)(uintptr_t)
358 rte_vhost_gpa_to_vva(dev->mem, vq->desc[idx].addr);
359 if (unlikely(!descs))
366 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
369 len += descs[idx].len;
370 buf_vec[vec_id].buf_addr = descs[idx].addr;
371 buf_vec[vec_id].buf_len = descs[idx].len;
372 buf_vec[vec_id].desc_idx = idx;
375 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
378 idx = descs[idx].next;
381 *desc_chain_len = len;
388 * Returns -1 on fail, 0 on success
391 reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
392 uint32_t size, struct buf_vector *buf_vec,
393 uint16_t *num_buffers, uint16_t avail_head)
396 uint32_t vec_idx = 0;
399 uint16_t head_idx = 0;
403 cur_idx = vq->last_avail_idx;
406 if (unlikely(cur_idx == avail_head))
409 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
410 &head_idx, &len) < 0))
412 len = RTE_MIN(len, size);
413 update_shadow_used_ring(vq, head_idx, len);
421 * if we tried all available ring items, and still
422 * can't get enough buf, it means something abnormal
425 if (unlikely(tries >= vq->size))
432 static __rte_always_inline int
433 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct rte_mbuf *m,
434 struct buf_vector *buf_vec, uint16_t num_buffers)
436 uint32_t vec_idx = 0;
438 uint32_t mbuf_offset, mbuf_avail;
439 uint32_t desc_offset, desc_avail;
441 uint64_t hdr_addr, hdr_phys_addr;
442 struct rte_mbuf *hdr_mbuf;
444 if (unlikely(m == NULL))
447 desc_addr = rte_vhost_gpa_to_vva(dev->mem, buf_vec[vec_idx].buf_addr);
448 if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr)
452 hdr_addr = desc_addr;
453 hdr_phys_addr = buf_vec[vec_idx].buf_addr;
454 rte_prefetch0((void *)(uintptr_t)hdr_addr);
456 LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
457 dev->vid, num_buffers);
459 desc_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
460 desc_offset = dev->vhost_hlen;
462 mbuf_avail = rte_pktmbuf_data_len(m);
464 while (mbuf_avail != 0 || m->next != NULL) {
465 /* done with current desc buf, get the next one */
466 if (desc_avail == 0) {
468 desc_addr = rte_vhost_gpa_to_vva(dev->mem,
469 buf_vec[vec_idx].buf_addr);
470 if (unlikely(!desc_addr))
473 /* Prefetch buffer address. */
474 rte_prefetch0((void *)(uintptr_t)desc_addr);
476 desc_avail = buf_vec[vec_idx].buf_len;
479 /* done with current mbuf, get the next one */
480 if (mbuf_avail == 0) {
484 mbuf_avail = rte_pktmbuf_data_len(m);
488 struct virtio_net_hdr_mrg_rxbuf *hdr;
490 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)
492 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
493 ASSIGN_UNLESS_EQUAL(hdr->num_buffers, num_buffers);
495 vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
496 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
502 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
503 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
504 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
506 vhost_log_write(dev, buf_vec[vec_idx].buf_addr + desc_offset,
508 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
511 mbuf_avail -= cpy_len;
512 mbuf_offset += cpy_len;
513 desc_avail -= cpy_len;
514 desc_offset += cpy_len;
520 static __rte_always_inline uint32_t
521 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
522 struct rte_mbuf **pkts, uint32_t count)
524 struct vhost_virtqueue *vq;
525 uint32_t pkt_idx = 0;
526 uint16_t num_buffers;
527 struct buf_vector buf_vec[BUF_VECTOR_MAX];
530 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
531 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
532 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
533 dev->vid, __func__, queue_id);
537 vq = dev->virtqueue[queue_id];
538 if (unlikely(vq->enabled == 0))
541 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
545 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
547 vq->shadow_used_idx = 0;
548 avail_head = *((volatile uint16_t *)&vq->avail->idx);
549 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
550 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
552 if (unlikely(reserve_avail_buf_mergeable(dev, vq,
553 pkt_len, buf_vec, &num_buffers,
555 LOG_DEBUG(VHOST_DATA,
556 "(%d) failed to get enough desc from vring\n",
558 vq->shadow_used_idx -= num_buffers;
562 LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
563 dev->vid, vq->last_avail_idx,
564 vq->last_avail_idx + num_buffers);
566 if (copy_mbuf_to_desc_mergeable(dev, pkts[pkt_idx],
567 buf_vec, num_buffers) < 0) {
568 vq->shadow_used_idx -= num_buffers;
572 vq->last_avail_idx += num_buffers;
575 if (likely(vq->shadow_used_idx)) {
576 flush_shadow_used_ring(dev, vq);
578 /* flush used->idx update before we read avail->flags. */
581 /* Kick the guest if necessary. */
582 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
583 && (vq->callfd >= 0))
584 eventfd_write(vq->callfd, (eventfd_t)1);
591 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
592 struct rte_mbuf **pkts, uint16_t count)
594 struct virtio_net *dev = get_device(vid);
599 if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
600 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
602 return virtio_dev_rx(dev, queue_id, pkts, count);
606 virtio_net_with_host_offload(struct virtio_net *dev)
609 (VIRTIO_NET_F_CSUM | VIRTIO_NET_F_HOST_ECN |
610 VIRTIO_NET_F_HOST_TSO4 | VIRTIO_NET_F_HOST_TSO6 |
611 VIRTIO_NET_F_HOST_UFO))
618 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
620 struct ipv4_hdr *ipv4_hdr;
621 struct ipv6_hdr *ipv6_hdr;
623 struct ether_hdr *eth_hdr;
626 eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
628 m->l2_len = sizeof(struct ether_hdr);
629 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
631 if (ethertype == ETHER_TYPE_VLAN) {
632 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
634 m->l2_len += sizeof(struct vlan_hdr);
635 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
638 l3_hdr = (char *)eth_hdr + m->l2_len;
641 case ETHER_TYPE_IPv4:
643 *l4_proto = ipv4_hdr->next_proto_id;
644 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
645 *l4_hdr = (char *)l3_hdr + m->l3_len;
646 m->ol_flags |= PKT_TX_IPV4;
648 case ETHER_TYPE_IPv6:
650 *l4_proto = ipv6_hdr->proto;
651 m->l3_len = sizeof(struct ipv6_hdr);
652 *l4_hdr = (char *)l3_hdr + m->l3_len;
653 m->ol_flags |= PKT_TX_IPV6;
663 static __rte_always_inline void
664 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
666 uint16_t l4_proto = 0;
668 struct tcp_hdr *tcp_hdr = NULL;
670 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
673 parse_ethernet(m, &l4_proto, &l4_hdr);
674 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
675 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
676 switch (hdr->csum_offset) {
677 case (offsetof(struct tcp_hdr, cksum)):
678 if (l4_proto == IPPROTO_TCP)
679 m->ol_flags |= PKT_TX_TCP_CKSUM;
681 case (offsetof(struct udp_hdr, dgram_cksum)):
682 if (l4_proto == IPPROTO_UDP)
683 m->ol_flags |= PKT_TX_UDP_CKSUM;
685 case (offsetof(struct sctp_hdr, cksum)):
686 if (l4_proto == IPPROTO_SCTP)
687 m->ol_flags |= PKT_TX_SCTP_CKSUM;
695 if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
696 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
697 case VIRTIO_NET_HDR_GSO_TCPV4:
698 case VIRTIO_NET_HDR_GSO_TCPV6:
700 m->ol_flags |= PKT_TX_TCP_SEG;
701 m->tso_segsz = hdr->gso_size;
702 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
705 RTE_LOG(WARNING, VHOST_DATA,
706 "unsupported gso type %u.\n", hdr->gso_type);
712 #define RARP_PKT_SIZE 64
715 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
717 struct ether_hdr *eth_hdr;
718 struct arp_hdr *rarp;
720 if (rarp_mbuf->buf_len < 64) {
721 RTE_LOG(WARNING, VHOST_DATA,
722 "failed to make RARP; mbuf size too small %u (< %d)\n",
723 rarp_mbuf->buf_len, RARP_PKT_SIZE);
727 /* Ethernet header. */
728 eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
729 memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
730 ether_addr_copy(mac, ð_hdr->s_addr);
731 eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
734 rarp = (struct arp_hdr *)(eth_hdr + 1);
735 rarp->arp_hrd = htons(ARP_HRD_ETHER);
736 rarp->arp_pro = htons(ETHER_TYPE_IPv4);
737 rarp->arp_hln = ETHER_ADDR_LEN;
739 rarp->arp_op = htons(ARP_OP_REVREQUEST);
741 ether_addr_copy(mac, &rarp->arp_data.arp_sha);
742 ether_addr_copy(mac, &rarp->arp_data.arp_tha);
743 memset(&rarp->arp_data.arp_sip, 0x00, 4);
744 memset(&rarp->arp_data.arp_tip, 0x00, 4);
746 rarp_mbuf->pkt_len = rarp_mbuf->data_len = RARP_PKT_SIZE;
751 static __rte_always_inline void
752 put_zmbuf(struct zcopy_mbuf *zmbuf)
757 static __rte_always_inline int
758 copy_desc_to_mbuf(struct virtio_net *dev, struct vring_desc *descs,
759 uint16_t max_desc, struct rte_mbuf *m, uint16_t desc_idx,
760 struct rte_mempool *mbuf_pool)
762 struct vring_desc *desc;
764 uint32_t desc_avail, desc_offset;
765 uint32_t mbuf_avail, mbuf_offset;
767 struct rte_mbuf *cur = m, *prev = m;
768 struct virtio_net_hdr *hdr = NULL;
769 /* A counter to avoid desc dead loop chain */
770 uint32_t nr_desc = 1;
772 desc = &descs[desc_idx];
773 if (unlikely((desc->len < dev->vhost_hlen)) ||
774 (desc->flags & VRING_DESC_F_INDIRECT))
777 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
778 if (unlikely(!desc_addr))
781 if (virtio_net_with_host_offload(dev)) {
782 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
787 * A virtio driver normally uses at least 2 desc buffers
788 * for Tx: the first for storing the header, and others
789 * for storing the data.
791 if (likely((desc->len == dev->vhost_hlen) &&
792 (desc->flags & VRING_DESC_F_NEXT) != 0)) {
793 desc = &descs[desc->next];
794 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
797 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
798 if (unlikely(!desc_addr))
802 desc_avail = desc->len;
805 desc_avail = desc->len - dev->vhost_hlen;
806 desc_offset = dev->vhost_hlen;
809 rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
811 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
814 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
818 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
821 * A desc buf might across two host physical pages that are
822 * not continuous. In such case (gpa_to_hpa returns 0), data
823 * will be copied even though zero copy is enabled.
825 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
826 desc->addr + desc_offset, cpy_len)))) {
827 cur->data_len = cpy_len;
829 cur->buf_addr = (void *)(uintptr_t)desc_addr;
830 cur->buf_physaddr = hpa;
833 * In zero copy mode, one mbuf can only reference data
834 * for one or partial of one desc buff.
836 mbuf_avail = cpy_len;
838 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
840 (void *)((uintptr_t)(desc_addr + desc_offset)),
844 mbuf_avail -= cpy_len;
845 mbuf_offset += cpy_len;
846 desc_avail -= cpy_len;
847 desc_offset += cpy_len;
849 /* This desc reaches to its end, get the next one */
850 if (desc_avail == 0) {
851 if ((desc->flags & VRING_DESC_F_NEXT) == 0)
854 if (unlikely(desc->next >= max_desc ||
855 ++nr_desc > max_desc))
857 desc = &descs[desc->next];
858 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
861 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
862 if (unlikely(!desc_addr))
865 rte_prefetch0((void *)(uintptr_t)desc_addr);
868 desc_avail = desc->len;
870 PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
874 * This mbuf reaches to its end, get a new one
877 if (mbuf_avail == 0) {
878 cur = rte_pktmbuf_alloc(mbuf_pool);
879 if (unlikely(cur == NULL)) {
880 RTE_LOG(ERR, VHOST_DATA, "Failed to "
881 "allocate memory for mbuf.\n");
884 if (unlikely(dev->dequeue_zero_copy))
885 rte_mbuf_refcnt_update(cur, 1);
888 prev->data_len = mbuf_offset;
890 m->pkt_len += mbuf_offset;
894 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
898 prev->data_len = mbuf_offset;
899 m->pkt_len += mbuf_offset;
902 vhost_dequeue_offload(hdr, m);
907 static __rte_always_inline void
908 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
909 uint32_t used_idx, uint32_t desc_idx)
911 vq->used->ring[used_idx].id = desc_idx;
912 vq->used->ring[used_idx].len = 0;
913 vhost_log_used_vring(dev, vq,
914 offsetof(struct vring_used, ring[used_idx]),
915 sizeof(vq->used->ring[used_idx]));
918 static __rte_always_inline void
919 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
922 if (unlikely(count == 0))
928 vq->used->idx += count;
929 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
930 sizeof(vq->used->idx));
932 /* Kick guest if required. */
933 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
934 && (vq->callfd >= 0))
935 eventfd_write(vq->callfd, (eventfd_t)1);
938 static __rte_always_inline struct zcopy_mbuf *
939 get_zmbuf(struct vhost_virtqueue *vq)
945 /* search [last_zmbuf_idx, zmbuf_size) */
946 i = vq->last_zmbuf_idx;
947 last = vq->zmbuf_size;
950 for (; i < last; i++) {
951 if (vq->zmbufs[i].in_use == 0) {
952 vq->last_zmbuf_idx = i + 1;
953 vq->zmbufs[i].in_use = 1;
954 return &vq->zmbufs[i];
960 /* search [0, last_zmbuf_idx) */
962 last = vq->last_zmbuf_idx;
969 static __rte_always_inline bool
970 mbuf_is_consumed(struct rte_mbuf *m)
973 if (rte_mbuf_refcnt_read(m) > 1)
982 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
983 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
985 struct virtio_net *dev;
986 struct rte_mbuf *rarp_mbuf = NULL;
987 struct vhost_virtqueue *vq;
988 uint32_t desc_indexes[MAX_PKT_BURST];
991 uint16_t free_entries;
994 dev = get_device(vid);
998 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
999 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1000 dev->vid, __func__, queue_id);
1004 vq = dev->virtqueue[queue_id];
1005 if (unlikely(vq->enabled == 0))
1008 if (unlikely(dev->dequeue_zero_copy)) {
1009 struct zcopy_mbuf *zmbuf, *next;
1012 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1013 zmbuf != NULL; zmbuf = next) {
1014 next = TAILQ_NEXT(zmbuf, next);
1016 if (mbuf_is_consumed(zmbuf->mbuf)) {
1017 used_idx = vq->last_used_idx++ & (vq->size - 1);
1018 update_used_ring(dev, vq, used_idx,
1022 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1023 rte_pktmbuf_free(zmbuf->mbuf);
1029 update_used_idx(dev, vq, nr_updated);
1033 * Construct a RARP broadcast packet, and inject it to the "pkts"
1034 * array, to looks like that guest actually send such packet.
1036 * Check user_send_rarp() for more information.
1038 * broadcast_rarp shares a cacheline in the virtio_net structure
1039 * with some fields that are accessed during enqueue and
1040 * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1041 * result in false sharing between enqueue and dequeue.
1043 * Prevent unnecessary false sharing by reading broadcast_rarp first
1044 * and only performing cmpset if the read indicates it is likely to
1048 if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1049 rte_atomic16_cmpset((volatile uint16_t *)
1050 &dev->broadcast_rarp.cnt, 1, 0))) {
1052 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
1053 if (rarp_mbuf == NULL) {
1054 RTE_LOG(ERR, VHOST_DATA,
1055 "Failed to allocate memory for mbuf.\n");
1059 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
1060 rte_pktmbuf_free(rarp_mbuf);
1067 free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1069 if (free_entries == 0)
1072 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1074 /* Prefetch available and used ring */
1075 avail_idx = vq->last_avail_idx & (vq->size - 1);
1076 used_idx = vq->last_used_idx & (vq->size - 1);
1077 rte_prefetch0(&vq->avail->ring[avail_idx]);
1078 rte_prefetch0(&vq->used->ring[used_idx]);
1080 count = RTE_MIN(count, MAX_PKT_BURST);
1081 count = RTE_MIN(count, free_entries);
1082 LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1085 /* Retrieve all of the head indexes first to avoid caching issues. */
1086 for (i = 0; i < count; i++) {
1087 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1088 used_idx = (vq->last_used_idx + i) & (vq->size - 1);
1089 desc_indexes[i] = vq->avail->ring[avail_idx];
1091 if (likely(dev->dequeue_zero_copy == 0))
1092 update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1095 /* Prefetch descriptor index. */
1096 rte_prefetch0(&vq->desc[desc_indexes[0]]);
1097 for (i = 0; i < count; i++) {
1098 struct vring_desc *desc;
1102 if (likely(i + 1 < count))
1103 rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1105 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1106 desc = (struct vring_desc *)(uintptr_t)
1107 rte_vhost_gpa_to_vva(dev->mem,
1108 vq->desc[desc_indexes[i]].addr);
1109 if (unlikely(!desc))
1112 rte_prefetch0(desc);
1113 sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1118 idx = desc_indexes[i];
1121 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1122 if (unlikely(pkts[i] == NULL)) {
1123 RTE_LOG(ERR, VHOST_DATA,
1124 "Failed to allocate memory for mbuf.\n");
1128 err = copy_desc_to_mbuf(dev, desc, sz, pkts[i], idx, mbuf_pool);
1129 if (unlikely(err)) {
1130 rte_pktmbuf_free(pkts[i]);
1134 if (unlikely(dev->dequeue_zero_copy)) {
1135 struct zcopy_mbuf *zmbuf;
1137 zmbuf = get_zmbuf(vq);
1139 rte_pktmbuf_free(pkts[i]);
1142 zmbuf->mbuf = pkts[i];
1143 zmbuf->desc_idx = desc_indexes[i];
1146 * Pin lock the mbuf; we will check later to see
1147 * whether the mbuf is freed (when we are the last
1148 * user) or not. If that's the case, we then could
1149 * update the used ring safely.
1151 rte_mbuf_refcnt_update(pkts[i], 1);
1154 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1157 vq->last_avail_idx += i;
1159 if (likely(dev->dequeue_zero_copy == 0)) {
1160 vq->last_used_idx += i;
1161 update_used_idx(dev, vq, i);
1165 if (unlikely(rarp_mbuf != NULL)) {
1167 * Inject it to the head of "pkts" array, so that switch's mac
1168 * learning table will get updated first.
1170 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1171 pkts[0] = rarp_mbuf;