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 inline void __attribute__((always_inline))
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 inline void __attribute__((always_inline))
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 inline void __attribute__((always_inline))
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;
109 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
111 if (m_buf->ol_flags & PKT_TX_L4_MASK) {
112 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
113 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
115 switch (m_buf->ol_flags & PKT_TX_L4_MASK) {
116 case PKT_TX_TCP_CKSUM:
117 net_hdr->csum_offset = (offsetof(struct tcp_hdr,
120 case PKT_TX_UDP_CKSUM:
121 net_hdr->csum_offset = (offsetof(struct udp_hdr,
124 case PKT_TX_SCTP_CKSUM:
125 net_hdr->csum_offset = (offsetof(struct sctp_hdr,
131 if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
132 if (m_buf->ol_flags & PKT_TX_IPV4)
133 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
135 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
136 net_hdr->gso_size = m_buf->tso_segsz;
137 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
143 copy_virtio_net_hdr(struct virtio_net *dev, uint64_t desc_addr,
144 struct virtio_net_hdr_mrg_rxbuf hdr)
146 if (dev->vhost_hlen == sizeof(struct virtio_net_hdr_mrg_rxbuf))
147 *(struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr = hdr;
149 *(struct virtio_net_hdr *)(uintptr_t)desc_addr = hdr.hdr;
152 static inline int __attribute__((always_inline))
153 copy_mbuf_to_desc(struct virtio_net *dev, struct vring_desc *descs,
154 struct rte_mbuf *m, uint16_t desc_idx, uint32_t size)
156 uint32_t desc_avail, desc_offset;
157 uint32_t mbuf_avail, mbuf_offset;
159 struct vring_desc *desc;
161 struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
162 /* A counter to avoid desc dead loop chain */
163 uint16_t nr_desc = 1;
165 desc = &descs[desc_idx];
166 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
168 * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
169 * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
170 * otherwise stores offset on the stack instead of in a register.
172 if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr)
175 rte_prefetch0((void *)(uintptr_t)desc_addr);
177 virtio_enqueue_offload(m, &virtio_hdr.hdr);
178 copy_virtio_net_hdr(dev, desc_addr, virtio_hdr);
179 vhost_log_write(dev, desc->addr, dev->vhost_hlen);
180 PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
182 desc_offset = dev->vhost_hlen;
183 desc_avail = desc->len - dev->vhost_hlen;
185 mbuf_avail = rte_pktmbuf_data_len(m);
187 while (mbuf_avail != 0 || m->next != NULL) {
188 /* done with current mbuf, fetch next */
189 if (mbuf_avail == 0) {
193 mbuf_avail = rte_pktmbuf_data_len(m);
196 /* done with current desc buf, fetch next */
197 if (desc_avail == 0) {
198 if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
199 /* Room in vring buffer is not enough */
202 if (unlikely(desc->next >= size || ++nr_desc > size))
205 desc = &descs[desc->next];
206 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
207 if (unlikely(!desc_addr))
211 desc_avail = desc->len;
214 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
215 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
216 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
218 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
219 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
222 mbuf_avail -= cpy_len;
223 mbuf_offset += cpy_len;
224 desc_avail -= cpy_len;
225 desc_offset += cpy_len;
232 * This function adds buffers to the virtio devices RX virtqueue. Buffers can
233 * be received from the physical port or from another virtio device. A packet
234 * count is returned to indicate the number of packets that are succesfully
235 * added to the RX queue. This function works when the mbuf is scattered, but
236 * it doesn't support the mergeable feature.
238 static inline uint32_t __attribute__((always_inline))
239 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
240 struct rte_mbuf **pkts, uint32_t count)
242 struct vhost_virtqueue *vq;
243 uint16_t avail_idx, free_entries, start_idx;
244 uint16_t desc_indexes[MAX_PKT_BURST];
245 struct vring_desc *descs;
249 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
250 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
251 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
252 dev->vid, __func__, queue_id);
256 vq = dev->virtqueue[queue_id];
257 if (unlikely(vq->enabled == 0))
260 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
261 start_idx = vq->last_used_idx;
262 free_entries = avail_idx - start_idx;
263 count = RTE_MIN(count, free_entries);
264 count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
268 LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
269 dev->vid, start_idx, start_idx + count);
271 /* Retrieve all of the desc indexes first to avoid caching issues. */
272 rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
273 for (i = 0; i < count; i++) {
274 used_idx = (start_idx + i) & (vq->size - 1);
275 desc_indexes[i] = vq->avail->ring[used_idx];
276 vq->used->ring[used_idx].id = desc_indexes[i];
277 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
279 vhost_log_used_vring(dev, vq,
280 offsetof(struct vring_used, ring[used_idx]),
281 sizeof(vq->used->ring[used_idx]));
284 rte_prefetch0(&vq->desc[desc_indexes[0]]);
285 for (i = 0; i < count; i++) {
286 uint16_t desc_idx = desc_indexes[i];
289 if (vq->desc[desc_idx].flags & VRING_DESC_F_INDIRECT) {
290 descs = (struct vring_desc *)(uintptr_t)
291 rte_vhost_gpa_to_vva(dev->mem,
292 vq->desc[desc_idx].addr);
293 if (unlikely(!descs)) {
299 sz = vq->desc[desc_idx].len / sizeof(*descs);
305 err = copy_mbuf_to_desc(dev, descs, pkts[i], desc_idx, sz);
307 used_idx = (start_idx + i) & (vq->size - 1);
308 vq->used->ring[used_idx].len = dev->vhost_hlen;
309 vhost_log_used_vring(dev, vq,
310 offsetof(struct vring_used, ring[used_idx]),
311 sizeof(vq->used->ring[used_idx]));
315 rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
320 *(volatile uint16_t *)&vq->used->idx += count;
321 vq->last_used_idx += count;
322 vhost_log_used_vring(dev, vq,
323 offsetof(struct vring_used, idx),
324 sizeof(vq->used->idx));
326 /* flush used->idx update before we read avail->flags. */
329 /* Kick the guest if necessary. */
330 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
331 && (vq->callfd >= 0))
332 eventfd_write(vq->callfd, (eventfd_t)1);
336 static inline int __attribute__((always_inline))
337 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
338 uint32_t avail_idx, uint32_t *vec_idx,
339 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
340 uint16_t *desc_chain_len)
342 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
343 uint32_t vec_id = *vec_idx;
345 struct vring_desc *descs = vq->desc;
347 *desc_chain_head = idx;
349 if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
350 descs = (struct vring_desc *)(uintptr_t)
351 rte_vhost_gpa_to_vva(dev->mem, vq->desc[idx].addr);
352 if (unlikely(!descs))
359 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
362 len += descs[idx].len;
363 buf_vec[vec_id].buf_addr = descs[idx].addr;
364 buf_vec[vec_id].buf_len = descs[idx].len;
365 buf_vec[vec_id].desc_idx = idx;
368 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
371 idx = descs[idx].next;
374 *desc_chain_len = len;
381 * Returns -1 on fail, 0 on success
384 reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
385 uint32_t size, struct buf_vector *buf_vec,
386 uint16_t *num_buffers, uint16_t avail_head)
389 uint32_t vec_idx = 0;
392 uint16_t head_idx = 0;
396 cur_idx = vq->last_avail_idx;
399 if (unlikely(cur_idx == avail_head))
402 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
403 &head_idx, &len) < 0))
405 len = RTE_MIN(len, size);
406 update_shadow_used_ring(vq, head_idx, len);
414 * if we tried all available ring items, and still
415 * can't get enough buf, it means something abnormal
418 if (unlikely(tries >= vq->size))
425 static inline int __attribute__((always_inline))
426 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct rte_mbuf *m,
427 struct buf_vector *buf_vec, uint16_t num_buffers)
429 struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
430 uint32_t vec_idx = 0;
432 uint32_t mbuf_offset, mbuf_avail;
433 uint32_t desc_offset, desc_avail;
435 uint64_t hdr_addr, hdr_phys_addr;
436 struct rte_mbuf *hdr_mbuf;
438 if (unlikely(m == NULL))
441 desc_addr = rte_vhost_gpa_to_vva(dev->mem, buf_vec[vec_idx].buf_addr);
442 if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr)
446 hdr_addr = desc_addr;
447 hdr_phys_addr = buf_vec[vec_idx].buf_addr;
448 rte_prefetch0((void *)(uintptr_t)hdr_addr);
450 virtio_hdr.num_buffers = num_buffers;
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 virtio_enqueue_offload(hdr_mbuf, &virtio_hdr.hdr);
484 copy_virtio_net_hdr(dev, hdr_addr, virtio_hdr);
485 vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
486 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
492 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
493 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
494 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
496 vhost_log_write(dev, buf_vec[vec_idx].buf_addr + desc_offset,
498 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
501 mbuf_avail -= cpy_len;
502 mbuf_offset += cpy_len;
503 desc_avail -= cpy_len;
504 desc_offset += cpy_len;
510 static inline uint32_t __attribute__((always_inline))
511 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
512 struct rte_mbuf **pkts, uint32_t count)
514 struct vhost_virtqueue *vq;
515 uint32_t pkt_idx = 0;
516 uint16_t num_buffers;
517 struct buf_vector buf_vec[BUF_VECTOR_MAX];
520 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
521 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
522 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
523 dev->vid, __func__, queue_id);
527 vq = dev->virtqueue[queue_id];
528 if (unlikely(vq->enabled == 0))
531 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
535 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
537 vq->shadow_used_idx = 0;
538 avail_head = *((volatile uint16_t *)&vq->avail->idx);
539 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
540 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
542 if (unlikely(reserve_avail_buf_mergeable(dev, vq,
543 pkt_len, buf_vec, &num_buffers,
545 LOG_DEBUG(VHOST_DATA,
546 "(%d) failed to get enough desc from vring\n",
548 vq->shadow_used_idx -= num_buffers;
552 LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
553 dev->vid, vq->last_avail_idx,
554 vq->last_avail_idx + num_buffers);
556 if (copy_mbuf_to_desc_mergeable(dev, pkts[pkt_idx],
557 buf_vec, num_buffers) < 0) {
558 vq->shadow_used_idx -= num_buffers;
562 vq->last_avail_idx += num_buffers;
565 if (likely(vq->shadow_used_idx)) {
566 flush_shadow_used_ring(dev, vq);
568 /* flush used->idx update before we read avail->flags. */
571 /* Kick the guest if necessary. */
572 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
573 && (vq->callfd >= 0))
574 eventfd_write(vq->callfd, (eventfd_t)1);
581 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
582 struct rte_mbuf **pkts, uint16_t count)
584 struct virtio_net *dev = get_device(vid);
589 if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
590 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
592 return virtio_dev_rx(dev, queue_id, pkts, count);
596 virtio_net_with_host_offload(struct virtio_net *dev)
599 (VIRTIO_NET_F_CSUM | VIRTIO_NET_F_HOST_ECN |
600 VIRTIO_NET_F_HOST_TSO4 | VIRTIO_NET_F_HOST_TSO6 |
601 VIRTIO_NET_F_HOST_UFO))
608 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
610 struct ipv4_hdr *ipv4_hdr;
611 struct ipv6_hdr *ipv6_hdr;
613 struct ether_hdr *eth_hdr;
616 eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
618 m->l2_len = sizeof(struct ether_hdr);
619 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
621 if (ethertype == ETHER_TYPE_VLAN) {
622 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
624 m->l2_len += sizeof(struct vlan_hdr);
625 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
628 l3_hdr = (char *)eth_hdr + m->l2_len;
631 case ETHER_TYPE_IPv4:
632 ipv4_hdr = (struct ipv4_hdr *)l3_hdr;
633 *l4_proto = ipv4_hdr->next_proto_id;
634 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
635 *l4_hdr = (char *)l3_hdr + m->l3_len;
636 m->ol_flags |= PKT_TX_IPV4;
638 case ETHER_TYPE_IPv6:
639 ipv6_hdr = (struct ipv6_hdr *)l3_hdr;
640 *l4_proto = ipv6_hdr->proto;
641 m->l3_len = sizeof(struct ipv6_hdr);
642 *l4_hdr = (char *)l3_hdr + m->l3_len;
643 m->ol_flags |= PKT_TX_IPV6;
653 static inline void __attribute__((always_inline))
654 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
656 uint16_t l4_proto = 0;
658 struct tcp_hdr *tcp_hdr = NULL;
660 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
663 parse_ethernet(m, &l4_proto, &l4_hdr);
664 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
665 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
666 switch (hdr->csum_offset) {
667 case (offsetof(struct tcp_hdr, cksum)):
668 if (l4_proto == IPPROTO_TCP)
669 m->ol_flags |= PKT_TX_TCP_CKSUM;
671 case (offsetof(struct udp_hdr, dgram_cksum)):
672 if (l4_proto == IPPROTO_UDP)
673 m->ol_flags |= PKT_TX_UDP_CKSUM;
675 case (offsetof(struct sctp_hdr, cksum)):
676 if (l4_proto == IPPROTO_SCTP)
677 m->ol_flags |= PKT_TX_SCTP_CKSUM;
685 if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
686 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
687 case VIRTIO_NET_HDR_GSO_TCPV4:
688 case VIRTIO_NET_HDR_GSO_TCPV6:
689 tcp_hdr = (struct tcp_hdr *)l4_hdr;
690 m->ol_flags |= PKT_TX_TCP_SEG;
691 m->tso_segsz = hdr->gso_size;
692 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
695 RTE_LOG(WARNING, VHOST_DATA,
696 "unsupported gso type %u.\n", hdr->gso_type);
702 #define RARP_PKT_SIZE 64
705 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
707 struct ether_hdr *eth_hdr;
708 struct arp_hdr *rarp;
710 if (rarp_mbuf->buf_len < 64) {
711 RTE_LOG(WARNING, VHOST_DATA,
712 "failed to make RARP; mbuf size too small %u (< %d)\n",
713 rarp_mbuf->buf_len, RARP_PKT_SIZE);
717 /* Ethernet header. */
718 eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
719 memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
720 ether_addr_copy(mac, ð_hdr->s_addr);
721 eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
724 rarp = (struct arp_hdr *)(eth_hdr + 1);
725 rarp->arp_hrd = htons(ARP_HRD_ETHER);
726 rarp->arp_pro = htons(ETHER_TYPE_IPv4);
727 rarp->arp_hln = ETHER_ADDR_LEN;
729 rarp->arp_op = htons(ARP_OP_REVREQUEST);
731 ether_addr_copy(mac, &rarp->arp_data.arp_sha);
732 ether_addr_copy(mac, &rarp->arp_data.arp_tha);
733 memset(&rarp->arp_data.arp_sip, 0x00, 4);
734 memset(&rarp->arp_data.arp_tip, 0x00, 4);
736 rarp_mbuf->pkt_len = rarp_mbuf->data_len = RARP_PKT_SIZE;
741 static inline void __attribute__((always_inline))
742 put_zmbuf(struct zcopy_mbuf *zmbuf)
747 static inline int __attribute__((always_inline))
748 copy_desc_to_mbuf(struct virtio_net *dev, struct vring_desc *descs,
749 uint16_t max_desc, struct rte_mbuf *m, uint16_t desc_idx,
750 struct rte_mempool *mbuf_pool)
752 struct vring_desc *desc;
754 uint32_t desc_avail, desc_offset;
755 uint32_t mbuf_avail, mbuf_offset;
757 struct rte_mbuf *cur = m, *prev = m;
758 struct virtio_net_hdr *hdr = NULL;
759 /* A counter to avoid desc dead loop chain */
760 uint32_t nr_desc = 1;
762 desc = &descs[desc_idx];
763 if (unlikely((desc->len < dev->vhost_hlen)) ||
764 (desc->flags & VRING_DESC_F_INDIRECT))
767 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
768 if (unlikely(!desc_addr))
771 if (virtio_net_with_host_offload(dev)) {
772 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
777 * A virtio driver normally uses at least 2 desc buffers
778 * for Tx: the first for storing the header, and others
779 * for storing the data.
781 if (likely((desc->len == dev->vhost_hlen) &&
782 (desc->flags & VRING_DESC_F_NEXT) != 0)) {
783 desc = &descs[desc->next];
784 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
787 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
788 if (unlikely(!desc_addr))
792 desc_avail = desc->len;
795 desc_avail = desc->len - dev->vhost_hlen;
796 desc_offset = dev->vhost_hlen;
799 rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
801 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
804 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
808 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
811 * A desc buf might across two host physical pages that are
812 * not continuous. In such case (gpa_to_hpa returns 0), data
813 * will be copied even though zero copy is enabled.
815 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
816 desc->addr + desc_offset, cpy_len)))) {
817 cur->data_len = cpy_len;
819 cur->buf_addr = (void *)(uintptr_t)desc_addr;
820 cur->buf_physaddr = hpa;
823 * In zero copy mode, one mbuf can only reference data
824 * for one or partial of one desc buff.
826 mbuf_avail = cpy_len;
828 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
830 (void *)((uintptr_t)(desc_addr + desc_offset)),
834 mbuf_avail -= cpy_len;
835 mbuf_offset += cpy_len;
836 desc_avail -= cpy_len;
837 desc_offset += cpy_len;
839 /* This desc reaches to its end, get the next one */
840 if (desc_avail == 0) {
841 if ((desc->flags & VRING_DESC_F_NEXT) == 0)
844 if (unlikely(desc->next >= max_desc ||
845 ++nr_desc > max_desc))
847 desc = &descs[desc->next];
848 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
851 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
852 if (unlikely(!desc_addr))
855 rte_prefetch0((void *)(uintptr_t)desc_addr);
858 desc_avail = desc->len;
860 PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
864 * This mbuf reaches to its end, get a new one
867 if (mbuf_avail == 0) {
868 cur = rte_pktmbuf_alloc(mbuf_pool);
869 if (unlikely(cur == NULL)) {
870 RTE_LOG(ERR, VHOST_DATA, "Failed to "
871 "allocate memory for mbuf.\n");
876 prev->data_len = mbuf_offset;
878 m->pkt_len += mbuf_offset;
882 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
886 prev->data_len = mbuf_offset;
887 m->pkt_len += mbuf_offset;
890 vhost_dequeue_offload(hdr, m);
895 static inline void __attribute__((always_inline))
896 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
897 uint32_t used_idx, uint32_t desc_idx)
899 vq->used->ring[used_idx].id = desc_idx;
900 vq->used->ring[used_idx].len = 0;
901 vhost_log_used_vring(dev, vq,
902 offsetof(struct vring_used, ring[used_idx]),
903 sizeof(vq->used->ring[used_idx]));
906 static inline void __attribute__((always_inline))
907 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
910 if (unlikely(count == 0))
916 vq->used->idx += count;
917 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
918 sizeof(vq->used->idx));
920 /* Kick guest if required. */
921 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
922 && (vq->callfd >= 0))
923 eventfd_write(vq->callfd, (eventfd_t)1);
926 static inline struct zcopy_mbuf *__attribute__((always_inline))
927 get_zmbuf(struct vhost_virtqueue *vq)
933 /* search [last_zmbuf_idx, zmbuf_size) */
934 i = vq->last_zmbuf_idx;
935 last = vq->zmbuf_size;
938 for (; i < last; i++) {
939 if (vq->zmbufs[i].in_use == 0) {
940 vq->last_zmbuf_idx = i + 1;
941 vq->zmbufs[i].in_use = 1;
942 return &vq->zmbufs[i];
948 /* search [0, last_zmbuf_idx) */
950 last = vq->last_zmbuf_idx;
957 static inline bool __attribute__((always_inline))
958 mbuf_is_consumed(struct rte_mbuf *m)
961 if (rte_mbuf_refcnt_read(m) > 1)
970 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
971 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
973 struct virtio_net *dev;
974 struct rte_mbuf *rarp_mbuf = NULL;
975 struct vhost_virtqueue *vq;
976 uint32_t desc_indexes[MAX_PKT_BURST];
979 uint16_t free_entries;
982 dev = get_device(vid);
986 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
987 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
988 dev->vid, __func__, queue_id);
992 vq = dev->virtqueue[queue_id];
993 if (unlikely(vq->enabled == 0))
996 if (unlikely(dev->dequeue_zero_copy)) {
997 struct zcopy_mbuf *zmbuf, *next;
1000 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1001 zmbuf != NULL; zmbuf = next) {
1002 next = TAILQ_NEXT(zmbuf, next);
1004 if (mbuf_is_consumed(zmbuf->mbuf)) {
1005 used_idx = vq->last_used_idx++ & (vq->size - 1);
1006 update_used_ring(dev, vq, used_idx,
1010 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1011 rte_pktmbuf_free(zmbuf->mbuf);
1017 update_used_idx(dev, vq, nr_updated);
1021 * Construct a RARP broadcast packet, and inject it to the "pkts"
1022 * array, to looks like that guest actually send such packet.
1024 * Check user_send_rarp() for more information.
1026 * broadcast_rarp shares a cacheline in the virtio_net structure
1027 * with some fields that are accessed during enqueue and
1028 * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1029 * result in false sharing between enqueue and dequeue.
1031 * Prevent unnecessary false sharing by reading broadcast_rarp first
1032 * and only performing cmpset if the read indicates it is likely to
1036 if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1037 rte_atomic16_cmpset((volatile uint16_t *)
1038 &dev->broadcast_rarp.cnt, 1, 0))) {
1040 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
1041 if (rarp_mbuf == NULL) {
1042 RTE_LOG(ERR, VHOST_DATA,
1043 "Failed to allocate memory for mbuf.\n");
1047 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
1048 rte_pktmbuf_free(rarp_mbuf);
1055 free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1057 if (free_entries == 0)
1060 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1062 /* Prefetch available and used ring */
1063 avail_idx = vq->last_avail_idx & (vq->size - 1);
1064 used_idx = vq->last_used_idx & (vq->size - 1);
1065 rte_prefetch0(&vq->avail->ring[avail_idx]);
1066 rte_prefetch0(&vq->used->ring[used_idx]);
1068 count = RTE_MIN(count, MAX_PKT_BURST);
1069 count = RTE_MIN(count, free_entries);
1070 LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1073 /* Retrieve all of the head indexes first to avoid caching issues. */
1074 for (i = 0; i < count; i++) {
1075 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1076 used_idx = (vq->last_used_idx + i) & (vq->size - 1);
1077 desc_indexes[i] = vq->avail->ring[avail_idx];
1079 if (likely(dev->dequeue_zero_copy == 0))
1080 update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1083 /* Prefetch descriptor index. */
1084 rte_prefetch0(&vq->desc[desc_indexes[0]]);
1085 for (i = 0; i < count; i++) {
1086 struct vring_desc *desc;
1090 if (likely(i + 1 < count))
1091 rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1093 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1094 desc = (struct vring_desc *)(uintptr_t)
1095 rte_vhost_gpa_to_vva(dev->mem,
1096 vq->desc[desc_indexes[i]].addr);
1097 if (unlikely(!desc))
1100 rte_prefetch0(desc);
1101 sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1106 idx = desc_indexes[i];
1109 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1110 if (unlikely(pkts[i] == NULL)) {
1111 RTE_LOG(ERR, VHOST_DATA,
1112 "Failed to allocate memory for mbuf.\n");
1116 err = copy_desc_to_mbuf(dev, desc, sz, pkts[i], idx, mbuf_pool);
1117 if (unlikely(err)) {
1118 rte_pktmbuf_free(pkts[i]);
1122 if (unlikely(dev->dequeue_zero_copy)) {
1123 struct zcopy_mbuf *zmbuf;
1125 zmbuf = get_zmbuf(vq);
1127 rte_pktmbuf_free(pkts[i]);
1130 zmbuf->mbuf = pkts[i];
1131 zmbuf->desc_idx = desc_indexes[i];
1134 * Pin lock the mbuf; we will check later to see
1135 * whether the mbuf is freed (when we are the last
1136 * user) or not. If that's the case, we then could
1137 * update the used ring safely.
1139 rte_mbuf_refcnt_update(pkts[i], 1);
1142 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1145 vq->last_avail_idx += i;
1147 if (likely(dev->dequeue_zero_copy == 0)) {
1148 vq->last_used_idx += i;
1149 update_used_idx(dev, vq, i);
1153 if (unlikely(rarp_mbuf != NULL)) {
1155 * Inject it to the head of "pkts" array, so that switch's mac
1156 * learning table will get updated first.
1158 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1159 pkts[0] = rarp_mbuf;