1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2016 Intel Corporation
7 #include <linux/virtio_net.h>
10 #include <rte_memcpy.h>
11 #include <rte_ether.h>
13 #include <rte_vhost.h>
18 #include <rte_spinlock.h>
19 #include <rte_malloc.h>
24 #define MAX_PKT_BURST 32
26 #define MAX_BATCH_LEN 256
28 static __rte_always_inline bool
29 rxvq_is_mergeable(struct virtio_net *dev)
31 return dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF);
34 static __rte_always_inline bool
35 virtio_net_is_inorder(struct virtio_net *dev)
37 return dev->features & (1ULL << VIRTIO_F_IN_ORDER);
41 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
43 return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
46 static __rte_always_inline void
47 do_flush_shadow_used_ring_split(struct virtio_net *dev,
48 struct vhost_virtqueue *vq,
49 uint16_t to, uint16_t from, uint16_t size)
51 rte_memcpy(&vq->used->ring[to],
52 &vq->shadow_used_split[from],
53 size * sizeof(struct vring_used_elem));
54 vhost_log_cache_used_vring(dev, vq,
55 offsetof(struct vring_used, ring[to]),
56 size * sizeof(struct vring_used_elem));
59 static __rte_always_inline void
60 flush_shadow_used_ring_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
62 uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
64 if (used_idx + vq->shadow_used_idx <= vq->size) {
65 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0,
70 /* update used ring interval [used_idx, vq->size] */
71 size = vq->size - used_idx;
72 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0, size);
74 /* update the left half used ring interval [0, left_size] */
75 do_flush_shadow_used_ring_split(dev, vq, 0, size,
76 vq->shadow_used_idx - size);
78 vq->last_used_idx += vq->shadow_used_idx;
82 vhost_log_cache_sync(dev, vq);
84 *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
85 vq->shadow_used_idx = 0;
86 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
87 sizeof(vq->used->idx));
90 static __rte_always_inline void
91 update_shadow_used_ring_split(struct vhost_virtqueue *vq,
92 uint16_t desc_idx, uint32_t len)
94 uint16_t i = vq->shadow_used_idx++;
96 vq->shadow_used_split[i].id = desc_idx;
97 vq->shadow_used_split[i].len = len;
100 static __rte_always_inline void
101 vhost_flush_enqueue_shadow_packed(struct virtio_net *dev,
102 struct vhost_virtqueue *vq)
105 uint16_t used_idx = vq->last_used_idx;
106 uint16_t head_idx = vq->last_used_idx;
107 uint16_t head_flags = 0;
109 /* Split loop in two to save memory barriers */
110 for (i = 0; i < vq->shadow_used_idx; i++) {
111 vq->desc_packed[used_idx].id = vq->shadow_used_packed[i].id;
112 vq->desc_packed[used_idx].len = vq->shadow_used_packed[i].len;
114 used_idx += vq->shadow_used_packed[i].count;
115 if (used_idx >= vq->size)
116 used_idx -= vq->size;
121 for (i = 0; i < vq->shadow_used_idx; i++) {
124 if (vq->shadow_used_packed[i].len)
125 flags = VRING_DESC_F_WRITE;
129 if (vq->used_wrap_counter) {
130 flags |= VRING_DESC_F_USED;
131 flags |= VRING_DESC_F_AVAIL;
133 flags &= ~VRING_DESC_F_USED;
134 flags &= ~VRING_DESC_F_AVAIL;
138 vq->desc_packed[vq->last_used_idx].flags = flags;
140 vhost_log_cache_used_vring(dev, vq,
142 sizeof(struct vring_packed_desc),
143 sizeof(struct vring_packed_desc));
145 head_idx = vq->last_used_idx;
149 vq_inc_last_used_packed(vq, vq->shadow_used_packed[i].count);
152 vq->desc_packed[head_idx].flags = head_flags;
154 vhost_log_cache_used_vring(dev, vq,
156 sizeof(struct vring_packed_desc),
157 sizeof(struct vring_packed_desc));
159 vq->shadow_used_idx = 0;
160 vhost_log_cache_sync(dev, vq);
163 static __rte_always_inline void
164 vhost_flush_dequeue_shadow_packed(struct virtio_net *dev,
165 struct vhost_virtqueue *vq)
167 struct vring_used_elem_packed *used_elem = &vq->shadow_used_packed[0];
169 vq->desc_packed[vq->shadow_last_used_idx].id = used_elem->id;
171 vq->desc_packed[vq->shadow_last_used_idx].flags = used_elem->flags;
173 vhost_log_cache_used_vring(dev, vq, vq->shadow_last_used_idx *
174 sizeof(struct vring_packed_desc),
175 sizeof(struct vring_packed_desc));
176 vq->shadow_used_idx = 0;
177 vhost_log_cache_sync(dev, vq);
180 static __rte_always_inline void
181 vhost_flush_enqueue_batch_packed(struct virtio_net *dev,
182 struct vhost_virtqueue *vq,
189 flags = PACKED_DESC_ENQUEUE_USED_FLAG(vq->used_wrap_counter);
191 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
192 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
193 vq->desc_packed[vq->last_used_idx + i].len = lens[i];
198 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
199 vq->desc_packed[vq->last_used_idx + i].flags = flags;
201 vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
202 sizeof(struct vring_packed_desc),
203 sizeof(struct vring_packed_desc) *
205 vhost_log_cache_sync(dev, vq);
207 vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
210 static __rte_always_inline void
211 vhost_shadow_dequeue_batch_packed_inorder(struct vhost_virtqueue *vq,
214 vq->shadow_used_packed[0].id = id;
216 if (!vq->shadow_used_idx) {
217 vq->shadow_last_used_idx = vq->last_used_idx;
218 vq->shadow_used_packed[0].flags =
219 PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
220 vq->shadow_used_packed[0].len = 0;
221 vq->shadow_used_packed[0].count = 1;
222 vq->shadow_used_idx++;
225 vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
228 static __rte_always_inline void
229 vhost_shadow_dequeue_batch_packed(struct virtio_net *dev,
230 struct vhost_virtqueue *vq,
237 flags = PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
239 if (!vq->shadow_used_idx) {
240 vq->shadow_last_used_idx = vq->last_used_idx;
241 vq->shadow_used_packed[0].id = ids[0];
242 vq->shadow_used_packed[0].len = 0;
243 vq->shadow_used_packed[0].count = 1;
244 vq->shadow_used_packed[0].flags = flags;
245 vq->shadow_used_idx++;
250 vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE) {
251 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
252 vq->desc_packed[vq->last_used_idx + i].len = 0;
256 vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE)
257 vq->desc_packed[vq->last_used_idx + i].flags = flags;
259 vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
260 sizeof(struct vring_packed_desc),
261 sizeof(struct vring_packed_desc) *
263 vhost_log_cache_sync(dev, vq);
265 vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
268 static __rte_always_inline void
269 vhost_shadow_dequeue_single_packed(struct vhost_virtqueue *vq,
275 flags = vq->desc_packed[vq->last_used_idx].flags;
276 if (vq->used_wrap_counter) {
277 flags |= VRING_DESC_F_USED;
278 flags |= VRING_DESC_F_AVAIL;
280 flags &= ~VRING_DESC_F_USED;
281 flags &= ~VRING_DESC_F_AVAIL;
284 if (!vq->shadow_used_idx) {
285 vq->shadow_last_used_idx = vq->last_used_idx;
287 vq->shadow_used_packed[0].id = buf_id;
288 vq->shadow_used_packed[0].len = 0;
289 vq->shadow_used_packed[0].flags = flags;
290 vq->shadow_used_idx++;
292 vq->desc_packed[vq->last_used_idx].id = buf_id;
293 vq->desc_packed[vq->last_used_idx].len = 0;
294 vq->desc_packed[vq->last_used_idx].flags = flags;
297 vq_inc_last_used_packed(vq, count);
300 static __rte_always_inline void
301 vhost_shadow_dequeue_single_packed_inorder(struct vhost_virtqueue *vq,
307 vq->shadow_used_packed[0].id = buf_id;
309 flags = vq->desc_packed[vq->last_used_idx].flags;
310 if (vq->used_wrap_counter) {
311 flags |= VRING_DESC_F_USED;
312 flags |= VRING_DESC_F_AVAIL;
314 flags &= ~VRING_DESC_F_USED;
315 flags &= ~VRING_DESC_F_AVAIL;
318 if (!vq->shadow_used_idx) {
319 vq->shadow_last_used_idx = vq->last_used_idx;
320 vq->shadow_used_packed[0].len = 0;
321 vq->shadow_used_packed[0].flags = flags;
322 vq->shadow_used_idx++;
325 vq_inc_last_used_packed(vq, count);
329 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
331 struct batch_copy_elem *elem = vq->batch_copy_elems;
332 uint16_t count = vq->batch_copy_nb_elems;
335 for (i = 0; i < count; i++) {
336 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
337 vhost_log_cache_write_iova(dev, vq, elem[i].log_addr,
339 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
342 vq->batch_copy_nb_elems = 0;
346 do_data_copy_dequeue(struct vhost_virtqueue *vq)
348 struct batch_copy_elem *elem = vq->batch_copy_elems;
349 uint16_t count = vq->batch_copy_nb_elems;
352 for (i = 0; i < count; i++)
353 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
355 vq->batch_copy_nb_elems = 0;
358 static __rte_always_inline void
359 vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
360 struct vhost_virtqueue *vq,
364 uint16_t num_buffers)
367 for (i = 0; i < num_buffers; i++) {
368 /* enqueue shadow flush action aligned with batch num */
369 if (!vq->shadow_used_idx)
370 vq->shadow_aligned_idx = vq->last_used_idx &
372 vq->shadow_used_packed[vq->shadow_used_idx].id = id[i];
373 vq->shadow_used_packed[vq->shadow_used_idx].len = len[i];
374 vq->shadow_used_packed[vq->shadow_used_idx].count = count[i];
375 vq->shadow_aligned_idx += count[i];
376 vq->shadow_used_idx++;
379 if (vq->shadow_aligned_idx >= PACKED_BATCH_SIZE) {
380 do_data_copy_enqueue(dev, vq);
381 vhost_flush_enqueue_shadow_packed(dev, vq);
385 /* avoid write operation when necessary, to lessen cache issues */
386 #define ASSIGN_UNLESS_EQUAL(var, val) do { \
387 if ((var) != (val)) \
391 static __rte_always_inline void
392 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
394 uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
396 if (m_buf->ol_flags & PKT_TX_TCP_SEG)
397 csum_l4 |= PKT_TX_TCP_CKSUM;
400 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
401 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
404 case PKT_TX_TCP_CKSUM:
405 net_hdr->csum_offset = (offsetof(struct rte_tcp_hdr,
408 case PKT_TX_UDP_CKSUM:
409 net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
412 case PKT_TX_SCTP_CKSUM:
413 net_hdr->csum_offset = (offsetof(struct rte_sctp_hdr,
418 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
419 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
420 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
423 /* IP cksum verification cannot be bypassed, then calculate here */
424 if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
425 struct rte_ipv4_hdr *ipv4_hdr;
427 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_ipv4_hdr *,
429 ipv4_hdr->hdr_checksum = 0;
430 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
433 if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
434 if (m_buf->ol_flags & PKT_TX_IPV4)
435 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
437 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
438 net_hdr->gso_size = m_buf->tso_segsz;
439 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
441 } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
442 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
443 net_hdr->gso_size = m_buf->tso_segsz;
444 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
447 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
448 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
449 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
453 static __rte_always_inline int
454 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
455 struct buf_vector *buf_vec, uint16_t *vec_idx,
456 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
458 uint16_t vec_id = *vec_idx;
462 uint64_t desc_chunck_len = desc_len;
464 if (unlikely(vec_id >= BUF_VECTOR_MAX))
467 desc_addr = vhost_iova_to_vva(dev, vq,
471 if (unlikely(!desc_addr))
474 rte_prefetch0((void *)(uintptr_t)desc_addr);
476 buf_vec[vec_id].buf_iova = desc_iova;
477 buf_vec[vec_id].buf_addr = desc_addr;
478 buf_vec[vec_id].buf_len = desc_chunck_len;
480 desc_len -= desc_chunck_len;
481 desc_iova += desc_chunck_len;
489 static __rte_always_inline int
490 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
491 uint32_t avail_idx, uint16_t *vec_idx,
492 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
493 uint32_t *desc_chain_len, uint8_t perm)
495 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
496 uint16_t vec_id = *vec_idx;
499 uint32_t nr_descs = vq->size;
501 struct vring_desc *descs = vq->desc;
502 struct vring_desc *idesc = NULL;
504 if (unlikely(idx >= vq->size))
507 *desc_chain_head = idx;
509 if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
510 dlen = vq->desc[idx].len;
511 nr_descs = dlen / sizeof(struct vring_desc);
512 if (unlikely(nr_descs > vq->size))
515 descs = (struct vring_desc *)(uintptr_t)
516 vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
519 if (unlikely(!descs))
522 if (unlikely(dlen < vq->desc[idx].len)) {
524 * The indirect desc table is not contiguous
525 * in process VA space, we have to copy it.
527 idesc = vhost_alloc_copy_ind_table(dev, vq,
528 vq->desc[idx].addr, vq->desc[idx].len);
529 if (unlikely(!idesc))
539 if (unlikely(idx >= nr_descs || cnt++ >= nr_descs)) {
540 free_ind_table(idesc);
544 len += descs[idx].len;
546 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
547 descs[idx].addr, descs[idx].len,
549 free_ind_table(idesc);
553 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
556 idx = descs[idx].next;
559 *desc_chain_len = len;
562 if (unlikely(!!idesc))
563 free_ind_table(idesc);
569 * Returns -1 on fail, 0 on success
572 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
573 uint32_t size, struct buf_vector *buf_vec,
574 uint16_t *num_buffers, uint16_t avail_head,
578 uint16_t vec_idx = 0;
579 uint16_t max_tries, tries = 0;
581 uint16_t head_idx = 0;
585 cur_idx = vq->last_avail_idx;
587 if (rxvq_is_mergeable(dev))
588 max_tries = vq->size - 1;
593 if (unlikely(cur_idx == avail_head))
596 * if we tried all available ring items, and still
597 * can't get enough buf, it means something abnormal
600 if (unlikely(++tries > max_tries))
603 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
606 VHOST_ACCESS_RW) < 0))
608 len = RTE_MIN(len, size);
609 update_shadow_used_ring_split(vq, head_idx, len);
621 static __rte_always_inline int
622 fill_vec_buf_packed_indirect(struct virtio_net *dev,
623 struct vhost_virtqueue *vq,
624 struct vring_packed_desc *desc, uint16_t *vec_idx,
625 struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
629 uint16_t vec_id = *vec_idx;
631 struct vring_packed_desc *descs, *idescs = NULL;
634 descs = (struct vring_packed_desc *)(uintptr_t)
635 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
636 if (unlikely(!descs))
639 if (unlikely(dlen < desc->len)) {
641 * The indirect desc table is not contiguous
642 * in process VA space, we have to copy it.
644 idescs = vhost_alloc_copy_ind_table(dev,
645 vq, desc->addr, desc->len);
646 if (unlikely(!idescs))
652 nr_descs = desc->len / sizeof(struct vring_packed_desc);
653 if (unlikely(nr_descs >= vq->size)) {
654 free_ind_table(idescs);
658 for (i = 0; i < nr_descs; i++) {
659 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
660 free_ind_table(idescs);
664 *len += descs[i].len;
665 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
666 descs[i].addr, descs[i].len,
672 if (unlikely(!!idescs))
673 free_ind_table(idescs);
678 static __rte_always_inline int
679 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
680 uint16_t avail_idx, uint16_t *desc_count,
681 struct buf_vector *buf_vec, uint16_t *vec_idx,
682 uint16_t *buf_id, uint32_t *len, uint8_t perm)
684 bool wrap_counter = vq->avail_wrap_counter;
685 struct vring_packed_desc *descs = vq->desc_packed;
686 uint16_t vec_id = *vec_idx;
688 if (avail_idx < vq->last_avail_idx)
692 * Perform a load-acquire barrier in desc_is_avail to
693 * enforce the ordering between desc flags and desc
696 if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
703 if (unlikely(vec_id >= BUF_VECTOR_MAX))
706 if (unlikely(*desc_count >= vq->size))
710 *buf_id = descs[avail_idx].id;
712 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
713 if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
719 *len += descs[avail_idx].len;
721 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
722 descs[avail_idx].addr,
723 descs[avail_idx].len,
728 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
731 if (++avail_idx >= vq->size) {
732 avail_idx -= vq->size;
742 static __rte_noinline void
743 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
744 struct buf_vector *buf_vec,
745 struct virtio_net_hdr_mrg_rxbuf *hdr)
748 uint64_t remain = dev->vhost_hlen;
749 uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
750 uint64_t iova = buf_vec->buf_iova;
753 len = RTE_MIN(remain,
755 dst = buf_vec->buf_addr;
756 rte_memcpy((void *)(uintptr_t)dst,
757 (void *)(uintptr_t)src,
760 PRINT_PACKET(dev, (uintptr_t)dst,
762 vhost_log_cache_write_iova(dev, vq,
772 static __rte_always_inline int
773 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
774 struct rte_mbuf *m, struct buf_vector *buf_vec,
775 uint16_t nr_vec, uint16_t num_buffers)
777 uint32_t vec_idx = 0;
778 uint32_t mbuf_offset, mbuf_avail;
779 uint32_t buf_offset, buf_avail;
780 uint64_t buf_addr, buf_iova, buf_len;
783 struct rte_mbuf *hdr_mbuf;
784 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
785 struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
788 if (unlikely(m == NULL)) {
793 buf_addr = buf_vec[vec_idx].buf_addr;
794 buf_iova = buf_vec[vec_idx].buf_iova;
795 buf_len = buf_vec[vec_idx].buf_len;
797 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
804 if (unlikely(buf_len < dev->vhost_hlen))
807 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
809 VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
810 dev->vid, num_buffers);
812 if (unlikely(buf_len < dev->vhost_hlen)) {
813 buf_offset = dev->vhost_hlen - buf_len;
815 buf_addr = buf_vec[vec_idx].buf_addr;
816 buf_iova = buf_vec[vec_idx].buf_iova;
817 buf_len = buf_vec[vec_idx].buf_len;
818 buf_avail = buf_len - buf_offset;
820 buf_offset = dev->vhost_hlen;
821 buf_avail = buf_len - dev->vhost_hlen;
824 mbuf_avail = rte_pktmbuf_data_len(m);
826 while (mbuf_avail != 0 || m->next != NULL) {
827 /* done with current buf, get the next one */
828 if (buf_avail == 0) {
830 if (unlikely(vec_idx >= nr_vec)) {
835 buf_addr = buf_vec[vec_idx].buf_addr;
836 buf_iova = buf_vec[vec_idx].buf_iova;
837 buf_len = buf_vec[vec_idx].buf_len;
843 /* done with current mbuf, get the next one */
844 if (mbuf_avail == 0) {
848 mbuf_avail = rte_pktmbuf_data_len(m);
852 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
853 if (rxvq_is_mergeable(dev))
854 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
857 if (unlikely(hdr == &tmp_hdr)) {
858 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
860 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
862 vhost_log_cache_write_iova(dev, vq,
870 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
872 if (likely(cpy_len > MAX_BATCH_LEN ||
873 vq->batch_copy_nb_elems >= vq->size)) {
874 rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
875 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
877 vhost_log_cache_write_iova(dev, vq,
878 buf_iova + buf_offset,
880 PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
883 batch_copy[vq->batch_copy_nb_elems].dst =
884 (void *)((uintptr_t)(buf_addr + buf_offset));
885 batch_copy[vq->batch_copy_nb_elems].src =
886 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
887 batch_copy[vq->batch_copy_nb_elems].log_addr =
888 buf_iova + buf_offset;
889 batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
890 vq->batch_copy_nb_elems++;
893 mbuf_avail -= cpy_len;
894 mbuf_offset += cpy_len;
895 buf_avail -= cpy_len;
896 buf_offset += cpy_len;
904 static __rte_always_inline int
905 vhost_enqueue_single_packed(struct virtio_net *dev,
906 struct vhost_virtqueue *vq,
907 struct rte_mbuf *pkt,
908 struct buf_vector *buf_vec,
912 uint16_t avail_idx = vq->last_avail_idx;
913 uint16_t max_tries, tries = 0;
917 uint32_t size = pkt->pkt_len + dev->vhost_hlen;
918 uint16_t num_buffers = 0;
919 uint32_t buffer_len[vq->size];
920 uint16_t buffer_buf_id[vq->size];
921 uint16_t buffer_desc_count[vq->size];
923 if (rxvq_is_mergeable(dev))
924 max_tries = vq->size - 1;
930 * if we tried all available ring items, and still
931 * can't get enough buf, it means something abnormal
934 if (unlikely(++tries > max_tries))
937 if (unlikely(fill_vec_buf_packed(dev, vq,
938 avail_idx, &desc_count,
941 VHOST_ACCESS_RW) < 0))
944 len = RTE_MIN(len, size);
947 buffer_len[num_buffers] = len;
948 buffer_buf_id[num_buffers] = buf_id;
949 buffer_desc_count[num_buffers] = desc_count;
952 *nr_descs += desc_count;
953 avail_idx += desc_count;
954 if (avail_idx >= vq->size)
955 avail_idx -= vq->size;
958 if (copy_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, num_buffers) < 0)
961 vhost_shadow_enqueue_single_packed(dev, vq, buffer_len, buffer_buf_id,
962 buffer_desc_count, num_buffers);
967 static __rte_noinline uint32_t
968 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
969 struct rte_mbuf **pkts, uint32_t count)
971 uint32_t pkt_idx = 0;
972 uint16_t num_buffers;
973 struct buf_vector buf_vec[BUF_VECTOR_MAX];
976 avail_head = *((volatile uint16_t *)&vq->avail->idx);
979 * The ordering between avail index and
980 * desc reads needs to be enforced.
984 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
986 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
987 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
990 if (unlikely(reserve_avail_buf_split(dev, vq,
991 pkt_len, buf_vec, &num_buffers,
992 avail_head, &nr_vec) < 0)) {
993 VHOST_LOG_DATA(DEBUG,
994 "(%d) failed to get enough desc from vring\n",
996 vq->shadow_used_idx -= num_buffers;
1000 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1001 dev->vid, vq->last_avail_idx,
1002 vq->last_avail_idx + num_buffers);
1004 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
1007 vq->shadow_used_idx -= num_buffers;
1011 vq->last_avail_idx += num_buffers;
1014 do_data_copy_enqueue(dev, vq);
1016 if (likely(vq->shadow_used_idx)) {
1017 flush_shadow_used_ring_split(dev, vq);
1018 vhost_vring_call_split(dev, vq);
1024 static __rte_always_inline int
1025 virtio_dev_rx_batch_packed(struct virtio_net *dev,
1026 struct vhost_virtqueue *vq,
1027 struct rte_mbuf **pkts)
1029 bool wrap_counter = vq->avail_wrap_counter;
1030 struct vring_packed_desc *descs = vq->desc_packed;
1031 uint16_t avail_idx = vq->last_avail_idx;
1032 uint64_t desc_addrs[PACKED_BATCH_SIZE];
1033 struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
1034 uint32_t buf_offset = dev->vhost_hlen;
1035 uint64_t lens[PACKED_BATCH_SIZE];
1036 uint16_t ids[PACKED_BATCH_SIZE];
1039 if (unlikely(avail_idx & PACKED_BATCH_MASK))
1042 if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1045 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1046 if (unlikely(pkts[i]->next != NULL))
1048 if (unlikely(!desc_is_avail(&descs[avail_idx + i],
1055 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1056 lens[i] = descs[avail_idx + i].len;
1058 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1059 if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
1063 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1064 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1065 descs[avail_idx + i].addr,
1069 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1070 if (unlikely(lens[i] != descs[avail_idx + i].len))
1074 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1075 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1076 hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)
1077 (uintptr_t)desc_addrs[i];
1078 lens[i] = pkts[i]->pkt_len + dev->vhost_hlen;
1081 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1082 virtio_enqueue_offload(pkts[i], &hdrs[i]->hdr);
1084 vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1086 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1087 rte_memcpy((void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1088 rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1092 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1093 vhost_log_cache_write_iova(dev, vq, descs[avail_idx + i].addr,
1096 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1097 ids[i] = descs[avail_idx + i].id;
1099 vhost_flush_enqueue_batch_packed(dev, vq, lens, ids);
1104 static __rte_always_inline int16_t
1105 virtio_dev_rx_single_packed(struct virtio_net *dev,
1106 struct vhost_virtqueue *vq,
1107 struct rte_mbuf *pkt)
1109 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1110 uint16_t nr_descs = 0;
1113 if (unlikely(vhost_enqueue_single_packed(dev, vq, pkt, buf_vec,
1115 VHOST_LOG_DATA(DEBUG,
1116 "(%d) failed to get enough desc from vring\n",
1121 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1122 dev->vid, vq->last_avail_idx,
1123 vq->last_avail_idx + nr_descs);
1125 vq_inc_last_avail_packed(vq, nr_descs);
1130 static __rte_noinline uint32_t
1131 virtio_dev_rx_packed(struct virtio_net *dev,
1132 struct vhost_virtqueue *vq,
1133 struct rte_mbuf **pkts,
1136 uint32_t pkt_idx = 0;
1137 uint32_t remained = count;
1140 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1142 if (remained >= PACKED_BATCH_SIZE) {
1143 if (!virtio_dev_rx_batch_packed(dev, vq,
1145 pkt_idx += PACKED_BATCH_SIZE;
1146 remained -= PACKED_BATCH_SIZE;
1151 if (virtio_dev_rx_single_packed(dev, vq, pkts[pkt_idx]))
1156 } while (pkt_idx < count);
1158 if (vq->shadow_used_idx) {
1159 do_data_copy_enqueue(dev, vq);
1160 vhost_flush_enqueue_shadow_packed(dev, vq);
1164 vhost_vring_call_packed(dev, vq);
1169 static __rte_always_inline uint32_t
1170 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
1171 struct rte_mbuf **pkts, uint32_t count)
1173 struct vhost_virtqueue *vq;
1176 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1177 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1178 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1179 dev->vid, __func__, queue_id);
1183 vq = dev->virtqueue[queue_id];
1185 rte_spinlock_lock(&vq->access_lock);
1187 if (unlikely(vq->enabled == 0))
1188 goto out_access_unlock;
1190 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1191 vhost_user_iotlb_rd_lock(vq);
1193 if (unlikely(vq->access_ok == 0))
1194 if (unlikely(vring_translate(dev, vq) < 0))
1197 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1201 if (vq_is_packed(dev))
1202 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
1204 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
1207 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1208 vhost_user_iotlb_rd_unlock(vq);
1211 rte_spinlock_unlock(&vq->access_lock);
1217 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
1218 struct rte_mbuf **pkts, uint16_t count)
1220 struct virtio_net *dev = get_device(vid);
1225 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1227 "(%d) %s: built-in vhost net backend is disabled.\n",
1228 dev->vid, __func__);
1232 return virtio_dev_rx(dev, queue_id, pkts, count);
1236 virtio_net_with_host_offload(struct virtio_net *dev)
1239 ((1ULL << VIRTIO_NET_F_CSUM) |
1240 (1ULL << VIRTIO_NET_F_HOST_ECN) |
1241 (1ULL << VIRTIO_NET_F_HOST_TSO4) |
1242 (1ULL << VIRTIO_NET_F_HOST_TSO6) |
1243 (1ULL << VIRTIO_NET_F_HOST_UFO)))
1250 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
1252 struct rte_ipv4_hdr *ipv4_hdr;
1253 struct rte_ipv6_hdr *ipv6_hdr;
1254 void *l3_hdr = NULL;
1255 struct rte_ether_hdr *eth_hdr;
1258 eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
1260 m->l2_len = sizeof(struct rte_ether_hdr);
1261 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
1263 if (ethertype == RTE_ETHER_TYPE_VLAN) {
1264 struct rte_vlan_hdr *vlan_hdr =
1265 (struct rte_vlan_hdr *)(eth_hdr + 1);
1267 m->l2_len += sizeof(struct rte_vlan_hdr);
1268 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
1271 l3_hdr = (char *)eth_hdr + m->l2_len;
1273 switch (ethertype) {
1274 case RTE_ETHER_TYPE_IPV4:
1276 *l4_proto = ipv4_hdr->next_proto_id;
1277 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
1278 *l4_hdr = (char *)l3_hdr + m->l3_len;
1279 m->ol_flags |= PKT_TX_IPV4;
1281 case RTE_ETHER_TYPE_IPV6:
1283 *l4_proto = ipv6_hdr->proto;
1284 m->l3_len = sizeof(struct rte_ipv6_hdr);
1285 *l4_hdr = (char *)l3_hdr + m->l3_len;
1286 m->ol_flags |= PKT_TX_IPV6;
1296 static __rte_always_inline void
1297 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
1299 uint16_t l4_proto = 0;
1300 void *l4_hdr = NULL;
1301 struct rte_tcp_hdr *tcp_hdr = NULL;
1303 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1306 parse_ethernet(m, &l4_proto, &l4_hdr);
1307 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1308 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
1309 switch (hdr->csum_offset) {
1310 case (offsetof(struct rte_tcp_hdr, cksum)):
1311 if (l4_proto == IPPROTO_TCP)
1312 m->ol_flags |= PKT_TX_TCP_CKSUM;
1314 case (offsetof(struct rte_udp_hdr, dgram_cksum)):
1315 if (l4_proto == IPPROTO_UDP)
1316 m->ol_flags |= PKT_TX_UDP_CKSUM;
1318 case (offsetof(struct rte_sctp_hdr, cksum)):
1319 if (l4_proto == IPPROTO_SCTP)
1320 m->ol_flags |= PKT_TX_SCTP_CKSUM;
1328 if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1329 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1330 case VIRTIO_NET_HDR_GSO_TCPV4:
1331 case VIRTIO_NET_HDR_GSO_TCPV6:
1333 m->ol_flags |= PKT_TX_TCP_SEG;
1334 m->tso_segsz = hdr->gso_size;
1335 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
1337 case VIRTIO_NET_HDR_GSO_UDP:
1338 m->ol_flags |= PKT_TX_UDP_SEG;
1339 m->tso_segsz = hdr->gso_size;
1340 m->l4_len = sizeof(struct rte_udp_hdr);
1343 VHOST_LOG_DATA(WARNING,
1344 "unsupported gso type %u.\n", hdr->gso_type);
1350 static __rte_noinline void
1351 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
1352 struct buf_vector *buf_vec)
1355 uint64_t remain = sizeof(struct virtio_net_hdr);
1357 uint64_t dst = (uint64_t)(uintptr_t)hdr;
1360 len = RTE_MIN(remain, buf_vec->buf_len);
1361 src = buf_vec->buf_addr;
1362 rte_memcpy((void *)(uintptr_t)dst,
1363 (void *)(uintptr_t)src, len);
1371 static __rte_always_inline int
1372 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
1373 struct buf_vector *buf_vec, uint16_t nr_vec,
1374 struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
1376 uint32_t buf_avail, buf_offset;
1377 uint64_t buf_addr, buf_iova, buf_len;
1378 uint32_t mbuf_avail, mbuf_offset;
1380 struct rte_mbuf *cur = m, *prev = m;
1381 struct virtio_net_hdr tmp_hdr;
1382 struct virtio_net_hdr *hdr = NULL;
1383 /* A counter to avoid desc dead loop chain */
1384 uint16_t vec_idx = 0;
1385 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
1388 buf_addr = buf_vec[vec_idx].buf_addr;
1389 buf_iova = buf_vec[vec_idx].buf_iova;
1390 buf_len = buf_vec[vec_idx].buf_len;
1392 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
1397 if (virtio_net_with_host_offload(dev)) {
1398 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
1400 * No luck, the virtio-net header doesn't fit
1401 * in a contiguous virtual area.
1403 copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
1406 hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
1411 * A virtio driver normally uses at least 2 desc buffers
1412 * for Tx: the first for storing the header, and others
1413 * for storing the data.
1415 if (unlikely(buf_len < dev->vhost_hlen)) {
1416 buf_offset = dev->vhost_hlen - buf_len;
1418 buf_addr = buf_vec[vec_idx].buf_addr;
1419 buf_iova = buf_vec[vec_idx].buf_iova;
1420 buf_len = buf_vec[vec_idx].buf_len;
1421 buf_avail = buf_len - buf_offset;
1422 } else if (buf_len == dev->vhost_hlen) {
1423 if (unlikely(++vec_idx >= nr_vec))
1425 buf_addr = buf_vec[vec_idx].buf_addr;
1426 buf_iova = buf_vec[vec_idx].buf_iova;
1427 buf_len = buf_vec[vec_idx].buf_len;
1430 buf_avail = buf_len;
1432 buf_offset = dev->vhost_hlen;
1433 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
1437 (uintptr_t)(buf_addr + buf_offset),
1438 (uint32_t)buf_avail, 0);
1441 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
1445 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1448 * A desc buf might across two host physical pages that are
1449 * not continuous. In such case (gpa_to_hpa returns 0), data
1450 * will be copied even though zero copy is enabled.
1452 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
1453 buf_iova + buf_offset, cpy_len)))) {
1454 cur->data_len = cpy_len;
1457 (void *)(uintptr_t)(buf_addr + buf_offset);
1458 cur->buf_iova = hpa;
1461 * In zero copy mode, one mbuf can only reference data
1462 * for one or partial of one desc buff.
1464 mbuf_avail = cpy_len;
1466 if (likely(cpy_len > MAX_BATCH_LEN ||
1467 vq->batch_copy_nb_elems >= vq->size ||
1468 (hdr && cur == m))) {
1469 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
1471 (void *)((uintptr_t)(buf_addr +
1475 batch_copy[vq->batch_copy_nb_elems].dst =
1476 rte_pktmbuf_mtod_offset(cur, void *,
1478 batch_copy[vq->batch_copy_nb_elems].src =
1479 (void *)((uintptr_t)(buf_addr +
1481 batch_copy[vq->batch_copy_nb_elems].len =
1483 vq->batch_copy_nb_elems++;
1487 mbuf_avail -= cpy_len;
1488 mbuf_offset += cpy_len;
1489 buf_avail -= cpy_len;
1490 buf_offset += cpy_len;
1492 /* This buf reaches to its end, get the next one */
1493 if (buf_avail == 0) {
1494 if (++vec_idx >= nr_vec)
1497 buf_addr = buf_vec[vec_idx].buf_addr;
1498 buf_iova = buf_vec[vec_idx].buf_iova;
1499 buf_len = buf_vec[vec_idx].buf_len;
1502 buf_avail = buf_len;
1504 PRINT_PACKET(dev, (uintptr_t)buf_addr,
1505 (uint32_t)buf_avail, 0);
1509 * This mbuf reaches to its end, get a new one
1510 * to hold more data.
1512 if (mbuf_avail == 0) {
1513 cur = rte_pktmbuf_alloc(mbuf_pool);
1514 if (unlikely(cur == NULL)) {
1515 VHOST_LOG_DATA(ERR, "Failed to "
1516 "allocate memory for mbuf.\n");
1520 if (unlikely(dev->dequeue_zero_copy))
1521 rte_mbuf_refcnt_update(cur, 1);
1524 prev->data_len = mbuf_offset;
1526 m->pkt_len += mbuf_offset;
1530 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1534 prev->data_len = mbuf_offset;
1535 m->pkt_len += mbuf_offset;
1538 vhost_dequeue_offload(hdr, m);
1545 static __rte_always_inline struct zcopy_mbuf *
1546 get_zmbuf(struct vhost_virtqueue *vq)
1552 /* search [last_zmbuf_idx, zmbuf_size) */
1553 i = vq->last_zmbuf_idx;
1554 last = vq->zmbuf_size;
1557 for (; i < last; i++) {
1558 if (vq->zmbufs[i].in_use == 0) {
1559 vq->last_zmbuf_idx = i + 1;
1560 vq->zmbufs[i].in_use = 1;
1561 return &vq->zmbufs[i];
1567 /* search [0, last_zmbuf_idx) */
1569 last = vq->last_zmbuf_idx;
1577 virtio_dev_extbuf_free(void *addr __rte_unused, void *opaque)
1583 virtio_dev_extbuf_alloc(struct rte_mbuf *pkt, uint32_t size)
1585 struct rte_mbuf_ext_shared_info *shinfo = NULL;
1586 uint32_t total_len = RTE_PKTMBUF_HEADROOM + size;
1591 /* Try to use pkt buffer to store shinfo to reduce the amount of memory
1592 * required, otherwise store shinfo in the new buffer.
1594 if (rte_pktmbuf_tailroom(pkt) >= sizeof(*shinfo))
1595 shinfo = rte_pktmbuf_mtod(pkt,
1596 struct rte_mbuf_ext_shared_info *);
1598 total_len += sizeof(*shinfo) + sizeof(uintptr_t);
1599 total_len = RTE_ALIGN_CEIL(total_len, sizeof(uintptr_t));
1602 if (unlikely(total_len > UINT16_MAX))
1605 buf_len = total_len;
1606 buf = rte_malloc(NULL, buf_len, RTE_CACHE_LINE_SIZE);
1607 if (unlikely(buf == NULL))
1610 /* Initialize shinfo */
1612 shinfo->free_cb = virtio_dev_extbuf_free;
1613 shinfo->fcb_opaque = buf;
1614 rte_mbuf_ext_refcnt_set(shinfo, 1);
1616 shinfo = rte_pktmbuf_ext_shinfo_init_helper(buf, &buf_len,
1617 virtio_dev_extbuf_free, buf);
1618 if (unlikely(shinfo == NULL)) {
1620 VHOST_LOG_DATA(ERR, "Failed to init shinfo\n");
1625 iova = rte_malloc_virt2iova(buf);
1626 rte_pktmbuf_attach_extbuf(pkt, buf, iova, buf_len, shinfo);
1627 rte_pktmbuf_reset_headroom(pkt);
1633 * Allocate a host supported pktmbuf.
1635 static __rte_always_inline struct rte_mbuf *
1636 virtio_dev_pktmbuf_alloc(struct virtio_net *dev, struct rte_mempool *mp,
1639 struct rte_mbuf *pkt = rte_pktmbuf_alloc(mp);
1641 if (unlikely(pkt == NULL)) {
1643 "Failed to allocate memory for mbuf.\n");
1647 if (rte_pktmbuf_tailroom(pkt) >= data_len)
1650 /* attach an external buffer if supported */
1651 if (dev->extbuf && !virtio_dev_extbuf_alloc(pkt, data_len))
1654 /* check if chained buffers are allowed */
1655 if (!dev->linearbuf)
1658 /* Data doesn't fit into the buffer and the host supports
1659 * only linear buffers
1661 rte_pktmbuf_free(pkt);
1666 static __rte_noinline uint16_t
1667 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1668 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1671 uint16_t free_entries;
1673 if (unlikely(dev->dequeue_zero_copy)) {
1674 struct zcopy_mbuf *zmbuf, *next;
1676 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1677 zmbuf != NULL; zmbuf = next) {
1678 next = TAILQ_NEXT(zmbuf, next);
1680 if (mbuf_is_consumed(zmbuf->mbuf)) {
1681 update_shadow_used_ring_split(vq,
1682 zmbuf->desc_idx, 0);
1683 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1684 restore_mbuf(zmbuf->mbuf);
1685 rte_pktmbuf_free(zmbuf->mbuf);
1691 if (likely(vq->shadow_used_idx)) {
1692 flush_shadow_used_ring_split(dev, vq);
1693 vhost_vring_call_split(dev, vq);
1697 free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1699 if (free_entries == 0)
1703 * The ordering between avail index and
1704 * desc reads needs to be enforced.
1708 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1710 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1712 count = RTE_MIN(count, MAX_PKT_BURST);
1713 count = RTE_MIN(count, free_entries);
1714 VHOST_LOG_DATA(DEBUG, "(%d) about to dequeue %u buffers\n",
1717 for (i = 0; i < count; i++) {
1718 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1721 uint16_t nr_vec = 0;
1724 if (unlikely(fill_vec_buf_split(dev, vq,
1725 vq->last_avail_idx + i,
1727 &head_idx, &buf_len,
1728 VHOST_ACCESS_RO) < 0))
1731 if (likely(dev->dequeue_zero_copy == 0))
1732 update_shadow_used_ring_split(vq, head_idx, 0);
1734 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
1735 if (unlikely(pkts[i] == NULL))
1738 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1740 if (unlikely(err)) {
1741 rte_pktmbuf_free(pkts[i]);
1745 if (unlikely(dev->dequeue_zero_copy)) {
1746 struct zcopy_mbuf *zmbuf;
1748 zmbuf = get_zmbuf(vq);
1750 rte_pktmbuf_free(pkts[i]);
1753 zmbuf->mbuf = pkts[i];
1754 zmbuf->desc_idx = head_idx;
1757 * Pin lock the mbuf; we will check later to see
1758 * whether the mbuf is freed (when we are the last
1759 * user) or not. If that's the case, we then could
1760 * update the used ring safely.
1762 rte_mbuf_refcnt_update(pkts[i], 1);
1765 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1768 vq->last_avail_idx += i;
1770 if (likely(dev->dequeue_zero_copy == 0)) {
1771 do_data_copy_dequeue(vq);
1772 if (unlikely(i < count))
1773 vq->shadow_used_idx = i;
1774 if (likely(vq->shadow_used_idx)) {
1775 flush_shadow_used_ring_split(dev, vq);
1776 vhost_vring_call_split(dev, vq);
1783 static __rte_always_inline int
1784 vhost_reserve_avail_batch_packed(struct virtio_net *dev,
1785 struct vhost_virtqueue *vq,
1786 struct rte_mempool *mbuf_pool,
1787 struct rte_mbuf **pkts,
1789 uintptr_t *desc_addrs,
1792 bool wrap = vq->avail_wrap_counter;
1793 struct vring_packed_desc *descs = vq->desc_packed;
1794 struct virtio_net_hdr *hdr;
1795 uint64_t lens[PACKED_BATCH_SIZE];
1796 uint64_t buf_lens[PACKED_BATCH_SIZE];
1797 uint32_t buf_offset = dev->vhost_hlen;
1800 if (unlikely(avail_idx & PACKED_BATCH_MASK))
1802 if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1805 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1806 flags = descs[avail_idx + i].flags;
1807 if (unlikely((wrap != !!(flags & VRING_DESC_F_AVAIL)) ||
1808 (wrap == !!(flags & VRING_DESC_F_USED)) ||
1809 (flags & PACKED_DESC_SINGLE_DEQUEUE_FLAG)))
1815 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1816 lens[i] = descs[avail_idx + i].len;
1818 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1819 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1820 descs[avail_idx + i].addr,
1821 &lens[i], VHOST_ACCESS_RW);
1824 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1825 if (unlikely((lens[i] != descs[avail_idx + i].len)))
1829 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1830 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, lens[i]);
1835 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1836 buf_lens[i] = pkts[i]->buf_len - pkts[i]->data_off;
1838 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1839 if (unlikely(buf_lens[i] < (lens[i] - buf_offset)))
1843 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1844 pkts[i]->pkt_len = descs[avail_idx + i].len - buf_offset;
1845 pkts[i]->data_len = pkts[i]->pkt_len;
1846 ids[i] = descs[avail_idx + i].id;
1849 if (virtio_net_with_host_offload(dev)) {
1850 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1851 hdr = (struct virtio_net_hdr *)(desc_addrs[i]);
1852 vhost_dequeue_offload(hdr, pkts[i]);
1859 for (i = 0; i < PACKED_BATCH_SIZE; i++)
1860 rte_pktmbuf_free(pkts[i]);
1865 static __rte_always_inline int
1866 virtio_dev_tx_batch_packed(struct virtio_net *dev,
1867 struct vhost_virtqueue *vq,
1868 struct rte_mempool *mbuf_pool,
1869 struct rte_mbuf **pkts)
1871 uint16_t avail_idx = vq->last_avail_idx;
1872 uint32_t buf_offset = dev->vhost_hlen;
1873 uintptr_t desc_addrs[PACKED_BATCH_SIZE];
1874 uint16_t ids[PACKED_BATCH_SIZE];
1877 if (vhost_reserve_avail_batch_packed(dev, vq, mbuf_pool, pkts,
1878 avail_idx, desc_addrs, ids))
1881 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1882 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1884 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1885 rte_memcpy(rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1886 (void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1889 if (virtio_net_is_inorder(dev))
1890 vhost_shadow_dequeue_batch_packed_inorder(vq,
1891 ids[PACKED_BATCH_SIZE - 1]);
1893 vhost_shadow_dequeue_batch_packed(dev, vq, ids);
1895 vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1900 static __rte_always_inline int
1901 vhost_dequeue_single_packed(struct virtio_net *dev,
1902 struct vhost_virtqueue *vq,
1903 struct rte_mempool *mbuf_pool,
1904 struct rte_mbuf **pkts,
1906 uint16_t *desc_count)
1908 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1910 uint16_t nr_vec = 0;
1913 if (unlikely(fill_vec_buf_packed(dev, vq,
1914 vq->last_avail_idx, desc_count,
1917 VHOST_ACCESS_RO) < 0))
1920 *pkts = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
1921 if (unlikely(*pkts == NULL)) {
1923 "Failed to allocate memory for mbuf.\n");
1927 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, *pkts,
1929 if (unlikely(err)) {
1930 rte_pktmbuf_free(*pkts);
1937 static __rte_always_inline int
1938 virtio_dev_tx_single_packed(struct virtio_net *dev,
1939 struct vhost_virtqueue *vq,
1940 struct rte_mempool *mbuf_pool,
1941 struct rte_mbuf **pkts)
1944 uint16_t buf_id, desc_count;
1946 if (vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
1950 if (virtio_net_is_inorder(dev))
1951 vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
1954 vhost_shadow_dequeue_single_packed(vq, buf_id, desc_count);
1956 vq_inc_last_avail_packed(vq, desc_count);
1961 static __rte_always_inline int
1962 virtio_dev_tx_batch_packed_zmbuf(struct virtio_net *dev,
1963 struct vhost_virtqueue *vq,
1964 struct rte_mempool *mbuf_pool,
1965 struct rte_mbuf **pkts)
1967 struct zcopy_mbuf *zmbufs[PACKED_BATCH_SIZE];
1968 uintptr_t desc_addrs[PACKED_BATCH_SIZE];
1969 uint16_t ids[PACKED_BATCH_SIZE];
1972 uint16_t avail_idx = vq->last_avail_idx;
1974 if (vhost_reserve_avail_batch_packed(dev, vq, mbuf_pool, pkts,
1975 avail_idx, desc_addrs, ids))
1978 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1979 zmbufs[i] = get_zmbuf(vq);
1981 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1986 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1987 zmbufs[i]->mbuf = pkts[i];
1988 zmbufs[i]->desc_idx = ids[i];
1989 zmbufs[i]->desc_count = 1;
1992 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1993 rte_mbuf_refcnt_update(pkts[i], 1);
1995 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1996 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbufs[i], next);
1998 vq->nr_zmbuf += PACKED_BATCH_SIZE;
1999 vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
2004 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2005 rte_pktmbuf_free(pkts[i]);
2010 static __rte_always_inline int
2011 virtio_dev_tx_single_packed_zmbuf(struct virtio_net *dev,
2012 struct vhost_virtqueue *vq,
2013 struct rte_mempool *mbuf_pool,
2014 struct rte_mbuf **pkts)
2016 uint16_t buf_id, desc_count;
2017 struct zcopy_mbuf *zmbuf;
2019 if (vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
2023 zmbuf = get_zmbuf(vq);
2025 rte_pktmbuf_free(*pkts);
2028 zmbuf->mbuf = *pkts;
2029 zmbuf->desc_idx = buf_id;
2030 zmbuf->desc_count = desc_count;
2032 rte_mbuf_refcnt_update(*pkts, 1);
2035 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
2037 vq_inc_last_avail_packed(vq, desc_count);
2041 static __rte_always_inline void
2042 free_zmbuf(struct vhost_virtqueue *vq)
2044 struct zcopy_mbuf *next = NULL;
2045 struct zcopy_mbuf *zmbuf;
2047 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
2048 zmbuf != NULL; zmbuf = next) {
2049 next = TAILQ_NEXT(zmbuf, next);
2051 uint16_t last_used_idx = vq->last_used_idx;
2053 if (mbuf_is_consumed(zmbuf->mbuf)) {
2055 flags = vq->desc_packed[last_used_idx].flags;
2056 if (vq->used_wrap_counter) {
2057 flags |= VRING_DESC_F_USED;
2058 flags |= VRING_DESC_F_AVAIL;
2060 flags &= ~VRING_DESC_F_USED;
2061 flags &= ~VRING_DESC_F_AVAIL;
2064 vq->desc_packed[last_used_idx].id = zmbuf->desc_idx;
2065 vq->desc_packed[last_used_idx].len = 0;
2068 vq->desc_packed[last_used_idx].flags = flags;
2070 vq_inc_last_used_packed(vq, zmbuf->desc_count);
2072 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
2073 restore_mbuf(zmbuf->mbuf);
2074 rte_pktmbuf_free(zmbuf->mbuf);
2081 static __rte_noinline uint16_t
2082 virtio_dev_tx_packed_zmbuf(struct virtio_net *dev,
2083 struct vhost_virtqueue *vq,
2084 struct rte_mempool *mbuf_pool,
2085 struct rte_mbuf **pkts,
2088 uint32_t pkt_idx = 0;
2089 uint32_t remained = count;
2094 if (remained >= PACKED_BATCH_SIZE) {
2095 if (!virtio_dev_tx_batch_packed_zmbuf(dev, vq,
2096 mbuf_pool, &pkts[pkt_idx])) {
2097 pkt_idx += PACKED_BATCH_SIZE;
2098 remained -= PACKED_BATCH_SIZE;
2103 if (virtio_dev_tx_single_packed_zmbuf(dev, vq, mbuf_pool,
2112 vhost_vring_call_packed(dev, vq);
2117 static __rte_noinline uint16_t
2118 virtio_dev_tx_packed(struct virtio_net *dev,
2119 struct vhost_virtqueue *vq,
2120 struct rte_mempool *mbuf_pool,
2121 struct rte_mbuf **pkts,
2124 uint32_t pkt_idx = 0;
2125 uint32_t remained = count;
2128 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
2130 if (remained >= PACKED_BATCH_SIZE) {
2131 if (!virtio_dev_tx_batch_packed(dev, vq, mbuf_pool,
2133 pkt_idx += PACKED_BATCH_SIZE;
2134 remained -= PACKED_BATCH_SIZE;
2139 if (virtio_dev_tx_single_packed(dev, vq, mbuf_pool,
2147 if (vq->shadow_used_idx) {
2148 do_data_copy_dequeue(vq);
2150 vhost_flush_dequeue_shadow_packed(dev, vq);
2151 vhost_vring_call_packed(dev, vq);
2158 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
2159 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
2161 struct virtio_net *dev;
2162 struct rte_mbuf *rarp_mbuf = NULL;
2163 struct vhost_virtqueue *vq;
2165 dev = get_device(vid);
2169 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
2171 "(%d) %s: built-in vhost net backend is disabled.\n",
2172 dev->vid, __func__);
2176 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
2178 "(%d) %s: invalid virtqueue idx %d.\n",
2179 dev->vid, __func__, queue_id);
2183 vq = dev->virtqueue[queue_id];
2185 if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
2188 if (unlikely(vq->enabled == 0)) {
2190 goto out_access_unlock;
2193 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2194 vhost_user_iotlb_rd_lock(vq);
2196 if (unlikely(vq->access_ok == 0))
2197 if (unlikely(vring_translate(dev, vq) < 0)) {
2203 * Construct a RARP broadcast packet, and inject it to the "pkts"
2204 * array, to looks like that guest actually send such packet.
2206 * Check user_send_rarp() for more information.
2208 * broadcast_rarp shares a cacheline in the virtio_net structure
2209 * with some fields that are accessed during enqueue and
2210 * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
2211 * result in false sharing between enqueue and dequeue.
2213 * Prevent unnecessary false sharing by reading broadcast_rarp first
2214 * and only performing cmpset if the read indicates it is likely to
2217 if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
2218 rte_atomic16_cmpset((volatile uint16_t *)
2219 &dev->broadcast_rarp.cnt, 1, 0))) {
2221 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
2222 if (rarp_mbuf == NULL) {
2223 VHOST_LOG_DATA(ERR, "Failed to make RARP packet.\n");
2230 if (vq_is_packed(dev)) {
2231 if (unlikely(dev->dequeue_zero_copy))
2232 count = virtio_dev_tx_packed_zmbuf(dev, vq, mbuf_pool,
2235 count = virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts,
2238 count = virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count);
2241 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2242 vhost_user_iotlb_rd_unlock(vq);
2245 rte_spinlock_unlock(&vq->access_lock);
2247 if (unlikely(rarp_mbuf != NULL)) {
2249 * Inject it to the head of "pkts" array, so that switch's mac
2250 * learning table will get updated first.
2252 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
2253 pkts[0] = rarp_mbuf;