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>
20 #include <rte_vhost_async.h>
25 #define MAX_BATCH_LEN 256
27 #define VHOST_ASYNC_BATCH_THRESHOLD 32
29 static __rte_always_inline bool
30 rxvq_is_mergeable(struct virtio_net *dev)
32 return dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF);
35 static __rte_always_inline bool
36 virtio_net_is_inorder(struct virtio_net *dev)
38 return dev->features & (1ULL << VIRTIO_F_IN_ORDER);
42 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
44 return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
48 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
50 struct batch_copy_elem *elem = vq->batch_copy_elems;
51 uint16_t count = vq->batch_copy_nb_elems;
54 for (i = 0; i < count; i++) {
55 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
56 vhost_log_cache_write_iova(dev, vq, elem[i].log_addr,
58 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
61 vq->batch_copy_nb_elems = 0;
65 do_data_copy_dequeue(struct vhost_virtqueue *vq)
67 struct batch_copy_elem *elem = vq->batch_copy_elems;
68 uint16_t count = vq->batch_copy_nb_elems;
71 for (i = 0; i < count; i++)
72 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
74 vq->batch_copy_nb_elems = 0;
77 static __rte_always_inline void
78 do_flush_shadow_used_ring_split(struct virtio_net *dev,
79 struct vhost_virtqueue *vq,
80 uint16_t to, uint16_t from, uint16_t size)
82 rte_memcpy(&vq->used->ring[to],
83 &vq->shadow_used_split[from],
84 size * sizeof(struct vring_used_elem));
85 vhost_log_cache_used_vring(dev, vq,
86 offsetof(struct vring_used, ring[to]),
87 size * sizeof(struct vring_used_elem));
90 static __rte_always_inline void
91 flush_shadow_used_ring_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
93 uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
95 if (used_idx + vq->shadow_used_idx <= vq->size) {
96 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0,
101 /* update used ring interval [used_idx, vq->size] */
102 size = vq->size - used_idx;
103 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0, size);
105 /* update the left half used ring interval [0, left_size] */
106 do_flush_shadow_used_ring_split(dev, vq, 0, size,
107 vq->shadow_used_idx - size);
109 vq->last_used_idx += vq->shadow_used_idx;
111 vhost_log_cache_sync(dev, vq);
113 __atomic_add_fetch(&vq->used->idx, vq->shadow_used_idx,
115 vq->shadow_used_idx = 0;
116 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
117 sizeof(vq->used->idx));
120 static __rte_always_inline void
121 async_flush_shadow_used_ring_split(struct virtio_net *dev,
122 struct vhost_virtqueue *vq)
124 uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
126 if (used_idx + vq->shadow_used_idx <= vq->size) {
127 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0,
128 vq->shadow_used_idx);
132 /* update used ring interval [used_idx, vq->size] */
133 size = vq->size - used_idx;
134 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0, size);
136 /* update the left half used ring interval [0, left_size] */
137 do_flush_shadow_used_ring_split(dev, vq, 0, size,
138 vq->shadow_used_idx - size);
141 vq->last_used_idx += vq->shadow_used_idx;
142 vq->shadow_used_idx = 0;
145 static __rte_always_inline void
146 update_shadow_used_ring_split(struct vhost_virtqueue *vq,
147 uint16_t desc_idx, uint32_t len)
149 uint16_t i = vq->shadow_used_idx++;
151 vq->shadow_used_split[i].id = desc_idx;
152 vq->shadow_used_split[i].len = len;
155 static __rte_always_inline void
156 vhost_flush_enqueue_shadow_packed(struct virtio_net *dev,
157 struct vhost_virtqueue *vq)
160 uint16_t used_idx = vq->last_used_idx;
161 uint16_t head_idx = vq->last_used_idx;
162 uint16_t head_flags = 0;
164 /* Split loop in two to save memory barriers */
165 for (i = 0; i < vq->shadow_used_idx; i++) {
166 vq->desc_packed[used_idx].id = vq->shadow_used_packed[i].id;
167 vq->desc_packed[used_idx].len = vq->shadow_used_packed[i].len;
169 used_idx += vq->shadow_used_packed[i].count;
170 if (used_idx >= vq->size)
171 used_idx -= vq->size;
176 for (i = 0; i < vq->shadow_used_idx; i++) {
179 if (vq->shadow_used_packed[i].len)
180 flags = VRING_DESC_F_WRITE;
184 if (vq->used_wrap_counter) {
185 flags |= VRING_DESC_F_USED;
186 flags |= VRING_DESC_F_AVAIL;
188 flags &= ~VRING_DESC_F_USED;
189 flags &= ~VRING_DESC_F_AVAIL;
193 vq->desc_packed[vq->last_used_idx].flags = flags;
195 vhost_log_cache_used_vring(dev, vq,
197 sizeof(struct vring_packed_desc),
198 sizeof(struct vring_packed_desc));
200 head_idx = vq->last_used_idx;
204 vq_inc_last_used_packed(vq, vq->shadow_used_packed[i].count);
207 vq->desc_packed[head_idx].flags = head_flags;
209 vhost_log_cache_used_vring(dev, vq,
211 sizeof(struct vring_packed_desc),
212 sizeof(struct vring_packed_desc));
214 vq->shadow_used_idx = 0;
215 vhost_log_cache_sync(dev, vq);
218 static __rte_always_inline void
219 vhost_flush_dequeue_shadow_packed(struct virtio_net *dev,
220 struct vhost_virtqueue *vq)
222 struct vring_used_elem_packed *used_elem = &vq->shadow_used_packed[0];
224 vq->desc_packed[vq->shadow_last_used_idx].id = used_elem->id;
226 vq->desc_packed[vq->shadow_last_used_idx].flags = used_elem->flags;
228 vhost_log_cache_used_vring(dev, vq, vq->shadow_last_used_idx *
229 sizeof(struct vring_packed_desc),
230 sizeof(struct vring_packed_desc));
231 vq->shadow_used_idx = 0;
232 vhost_log_cache_sync(dev, vq);
235 static __rte_always_inline void
236 vhost_flush_enqueue_batch_packed(struct virtio_net *dev,
237 struct vhost_virtqueue *vq,
244 if (vq->shadow_used_idx) {
245 do_data_copy_enqueue(dev, vq);
246 vhost_flush_enqueue_shadow_packed(dev, vq);
249 flags = PACKED_DESC_ENQUEUE_USED_FLAG(vq->used_wrap_counter);
251 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
252 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
253 vq->desc_packed[vq->last_used_idx + i].len = lens[i];
258 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
259 vq->desc_packed[vq->last_used_idx + i].flags = flags;
261 vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
262 sizeof(struct vring_packed_desc),
263 sizeof(struct vring_packed_desc) *
265 vhost_log_cache_sync(dev, vq);
267 vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
270 static __rte_always_inline void
271 vhost_shadow_dequeue_batch_packed_inorder(struct vhost_virtqueue *vq,
274 vq->shadow_used_packed[0].id = id;
276 if (!vq->shadow_used_idx) {
277 vq->shadow_last_used_idx = vq->last_used_idx;
278 vq->shadow_used_packed[0].flags =
279 PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
280 vq->shadow_used_packed[0].len = 0;
281 vq->shadow_used_packed[0].count = 1;
282 vq->shadow_used_idx++;
285 vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
288 static __rte_always_inline void
289 vhost_shadow_dequeue_batch_packed(struct virtio_net *dev,
290 struct vhost_virtqueue *vq,
297 flags = PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
299 if (!vq->shadow_used_idx) {
300 vq->shadow_last_used_idx = vq->last_used_idx;
301 vq->shadow_used_packed[0].id = ids[0];
302 vq->shadow_used_packed[0].len = 0;
303 vq->shadow_used_packed[0].count = 1;
304 vq->shadow_used_packed[0].flags = flags;
305 vq->shadow_used_idx++;
310 vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE) {
311 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
312 vq->desc_packed[vq->last_used_idx + i].len = 0;
316 vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE)
317 vq->desc_packed[vq->last_used_idx + i].flags = flags;
319 vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
320 sizeof(struct vring_packed_desc),
321 sizeof(struct vring_packed_desc) *
323 vhost_log_cache_sync(dev, vq);
325 vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
328 static __rte_always_inline void
329 vhost_shadow_dequeue_single_packed(struct vhost_virtqueue *vq,
335 flags = vq->desc_packed[vq->last_used_idx].flags;
336 if (vq->used_wrap_counter) {
337 flags |= VRING_DESC_F_USED;
338 flags |= VRING_DESC_F_AVAIL;
340 flags &= ~VRING_DESC_F_USED;
341 flags &= ~VRING_DESC_F_AVAIL;
344 if (!vq->shadow_used_idx) {
345 vq->shadow_last_used_idx = vq->last_used_idx;
347 vq->shadow_used_packed[0].id = buf_id;
348 vq->shadow_used_packed[0].len = 0;
349 vq->shadow_used_packed[0].flags = flags;
350 vq->shadow_used_idx++;
352 vq->desc_packed[vq->last_used_idx].id = buf_id;
353 vq->desc_packed[vq->last_used_idx].len = 0;
354 vq->desc_packed[vq->last_used_idx].flags = flags;
357 vq_inc_last_used_packed(vq, count);
360 static __rte_always_inline void
361 vhost_shadow_dequeue_single_packed_inorder(struct vhost_virtqueue *vq,
367 vq->shadow_used_packed[0].id = buf_id;
369 flags = vq->desc_packed[vq->last_used_idx].flags;
370 if (vq->used_wrap_counter) {
371 flags |= VRING_DESC_F_USED;
372 flags |= VRING_DESC_F_AVAIL;
374 flags &= ~VRING_DESC_F_USED;
375 flags &= ~VRING_DESC_F_AVAIL;
378 if (!vq->shadow_used_idx) {
379 vq->shadow_last_used_idx = vq->last_used_idx;
380 vq->shadow_used_packed[0].len = 0;
381 vq->shadow_used_packed[0].flags = flags;
382 vq->shadow_used_idx++;
385 vq_inc_last_used_packed(vq, count);
388 static __rte_always_inline void
389 vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
390 struct vhost_virtqueue *vq,
394 uint16_t num_buffers)
397 for (i = 0; i < num_buffers; i++) {
398 /* enqueue shadow flush action aligned with batch num */
399 if (!vq->shadow_used_idx)
400 vq->shadow_aligned_idx = vq->last_used_idx &
402 vq->shadow_used_packed[vq->shadow_used_idx].id = id[i];
403 vq->shadow_used_packed[vq->shadow_used_idx].len = len[i];
404 vq->shadow_used_packed[vq->shadow_used_idx].count = count[i];
405 vq->shadow_aligned_idx += count[i];
406 vq->shadow_used_idx++;
409 if (vq->shadow_aligned_idx >= PACKED_BATCH_SIZE) {
410 do_data_copy_enqueue(dev, vq);
411 vhost_flush_enqueue_shadow_packed(dev, vq);
415 /* avoid write operation when necessary, to lessen cache issues */
416 #define ASSIGN_UNLESS_EQUAL(var, val) do { \
417 if ((var) != (val)) \
421 static __rte_always_inline void
422 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
424 uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
426 if (m_buf->ol_flags & PKT_TX_TCP_SEG)
427 csum_l4 |= PKT_TX_TCP_CKSUM;
430 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
431 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
434 case PKT_TX_TCP_CKSUM:
435 net_hdr->csum_offset = (offsetof(struct rte_tcp_hdr,
438 case PKT_TX_UDP_CKSUM:
439 net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
442 case PKT_TX_SCTP_CKSUM:
443 net_hdr->csum_offset = (offsetof(struct rte_sctp_hdr,
448 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
449 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
450 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
453 /* IP cksum verification cannot be bypassed, then calculate here */
454 if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
455 struct rte_ipv4_hdr *ipv4_hdr;
457 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_ipv4_hdr *,
459 ipv4_hdr->hdr_checksum = 0;
460 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
463 if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
464 if (m_buf->ol_flags & PKT_TX_IPV4)
465 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
467 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
468 net_hdr->gso_size = m_buf->tso_segsz;
469 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
471 } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
472 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
473 net_hdr->gso_size = m_buf->tso_segsz;
474 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
477 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
478 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
479 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
483 static __rte_always_inline int
484 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
485 struct buf_vector *buf_vec, uint16_t *vec_idx,
486 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
488 uint16_t vec_id = *vec_idx;
492 uint64_t desc_chunck_len = desc_len;
494 if (unlikely(vec_id >= BUF_VECTOR_MAX))
497 desc_addr = vhost_iova_to_vva(dev, vq,
501 if (unlikely(!desc_addr))
504 rte_prefetch0((void *)(uintptr_t)desc_addr);
506 buf_vec[vec_id].buf_iova = desc_iova;
507 buf_vec[vec_id].buf_addr = desc_addr;
508 buf_vec[vec_id].buf_len = desc_chunck_len;
510 desc_len -= desc_chunck_len;
511 desc_iova += desc_chunck_len;
519 static __rte_always_inline int
520 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
521 uint32_t avail_idx, uint16_t *vec_idx,
522 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
523 uint32_t *desc_chain_len, uint8_t perm)
525 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
526 uint16_t vec_id = *vec_idx;
529 uint32_t nr_descs = vq->size;
531 struct vring_desc *descs = vq->desc;
532 struct vring_desc *idesc = NULL;
534 if (unlikely(idx >= vq->size))
537 *desc_chain_head = idx;
539 if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
540 dlen = vq->desc[idx].len;
541 nr_descs = dlen / sizeof(struct vring_desc);
542 if (unlikely(nr_descs > vq->size))
545 descs = (struct vring_desc *)(uintptr_t)
546 vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
549 if (unlikely(!descs))
552 if (unlikely(dlen < vq->desc[idx].len)) {
554 * The indirect desc table is not contiguous
555 * in process VA space, we have to copy it.
557 idesc = vhost_alloc_copy_ind_table(dev, vq,
558 vq->desc[idx].addr, vq->desc[idx].len);
559 if (unlikely(!idesc))
569 if (unlikely(idx >= nr_descs || cnt++ >= nr_descs)) {
570 free_ind_table(idesc);
574 len += descs[idx].len;
576 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
577 descs[idx].addr, descs[idx].len,
579 free_ind_table(idesc);
583 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
586 idx = descs[idx].next;
589 *desc_chain_len = len;
592 if (unlikely(!!idesc))
593 free_ind_table(idesc);
599 * Returns -1 on fail, 0 on success
602 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
603 uint32_t size, struct buf_vector *buf_vec,
604 uint16_t *num_buffers, uint16_t avail_head,
608 uint16_t vec_idx = 0;
609 uint16_t max_tries, tries = 0;
611 uint16_t head_idx = 0;
615 cur_idx = vq->last_avail_idx;
617 if (rxvq_is_mergeable(dev))
618 max_tries = vq->size - 1;
623 if (unlikely(cur_idx == avail_head))
626 * if we tried all available ring items, and still
627 * can't get enough buf, it means something abnormal
630 if (unlikely(++tries > max_tries))
633 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
636 VHOST_ACCESS_RW) < 0))
638 len = RTE_MIN(len, size);
639 update_shadow_used_ring_split(vq, head_idx, len);
651 static __rte_always_inline int
652 fill_vec_buf_packed_indirect(struct virtio_net *dev,
653 struct vhost_virtqueue *vq,
654 struct vring_packed_desc *desc, uint16_t *vec_idx,
655 struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
659 uint16_t vec_id = *vec_idx;
661 struct vring_packed_desc *descs, *idescs = NULL;
664 descs = (struct vring_packed_desc *)(uintptr_t)
665 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
666 if (unlikely(!descs))
669 if (unlikely(dlen < desc->len)) {
671 * The indirect desc table is not contiguous
672 * in process VA space, we have to copy it.
674 idescs = vhost_alloc_copy_ind_table(dev,
675 vq, desc->addr, desc->len);
676 if (unlikely(!idescs))
682 nr_descs = desc->len / sizeof(struct vring_packed_desc);
683 if (unlikely(nr_descs >= vq->size)) {
684 free_ind_table(idescs);
688 for (i = 0; i < nr_descs; i++) {
689 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
690 free_ind_table(idescs);
694 *len += descs[i].len;
695 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
696 descs[i].addr, descs[i].len,
702 if (unlikely(!!idescs))
703 free_ind_table(idescs);
708 static __rte_always_inline int
709 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
710 uint16_t avail_idx, uint16_t *desc_count,
711 struct buf_vector *buf_vec, uint16_t *vec_idx,
712 uint16_t *buf_id, uint32_t *len, uint8_t perm)
714 bool wrap_counter = vq->avail_wrap_counter;
715 struct vring_packed_desc *descs = vq->desc_packed;
716 uint16_t vec_id = *vec_idx;
718 if (avail_idx < vq->last_avail_idx)
722 * Perform a load-acquire barrier in desc_is_avail to
723 * enforce the ordering between desc flags and desc
726 if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
733 if (unlikely(vec_id >= BUF_VECTOR_MAX))
736 if (unlikely(*desc_count >= vq->size))
740 *buf_id = descs[avail_idx].id;
742 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
743 if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
749 *len += descs[avail_idx].len;
751 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
752 descs[avail_idx].addr,
753 descs[avail_idx].len,
758 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
761 if (++avail_idx >= vq->size) {
762 avail_idx -= vq->size;
772 static __rte_noinline void
773 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
774 struct buf_vector *buf_vec,
775 struct virtio_net_hdr_mrg_rxbuf *hdr)
778 uint64_t remain = dev->vhost_hlen;
779 uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
780 uint64_t iova = buf_vec->buf_iova;
783 len = RTE_MIN(remain,
785 dst = buf_vec->buf_addr;
786 rte_memcpy((void *)(uintptr_t)dst,
787 (void *)(uintptr_t)src,
790 PRINT_PACKET(dev, (uintptr_t)dst,
792 vhost_log_cache_write_iova(dev, vq,
802 static __rte_always_inline int
803 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
804 struct rte_mbuf *m, struct buf_vector *buf_vec,
805 uint16_t nr_vec, uint16_t num_buffers)
807 uint32_t vec_idx = 0;
808 uint32_t mbuf_offset, mbuf_avail;
809 uint32_t buf_offset, buf_avail;
810 uint64_t buf_addr, buf_iova, buf_len;
813 struct rte_mbuf *hdr_mbuf;
814 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
815 struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
818 if (unlikely(m == NULL)) {
823 buf_addr = buf_vec[vec_idx].buf_addr;
824 buf_iova = buf_vec[vec_idx].buf_iova;
825 buf_len = buf_vec[vec_idx].buf_len;
827 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
834 if (unlikely(buf_len < dev->vhost_hlen))
837 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
839 VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
840 dev->vid, num_buffers);
842 if (unlikely(buf_len < dev->vhost_hlen)) {
843 buf_offset = dev->vhost_hlen - buf_len;
845 buf_addr = buf_vec[vec_idx].buf_addr;
846 buf_iova = buf_vec[vec_idx].buf_iova;
847 buf_len = buf_vec[vec_idx].buf_len;
848 buf_avail = buf_len - buf_offset;
850 buf_offset = dev->vhost_hlen;
851 buf_avail = buf_len - dev->vhost_hlen;
854 mbuf_avail = rte_pktmbuf_data_len(m);
856 while (mbuf_avail != 0 || m->next != NULL) {
857 /* done with current buf, get the next one */
858 if (buf_avail == 0) {
860 if (unlikely(vec_idx >= nr_vec)) {
865 buf_addr = buf_vec[vec_idx].buf_addr;
866 buf_iova = buf_vec[vec_idx].buf_iova;
867 buf_len = buf_vec[vec_idx].buf_len;
873 /* done with current mbuf, get the next one */
874 if (mbuf_avail == 0) {
878 mbuf_avail = rte_pktmbuf_data_len(m);
882 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
883 if (rxvq_is_mergeable(dev))
884 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
887 if (unlikely(hdr == &tmp_hdr)) {
888 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
890 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
892 vhost_log_cache_write_iova(dev, vq,
900 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
902 if (likely(cpy_len > MAX_BATCH_LEN ||
903 vq->batch_copy_nb_elems >= vq->size)) {
904 rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
905 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
907 vhost_log_cache_write_iova(dev, vq,
908 buf_iova + buf_offset,
910 PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
913 batch_copy[vq->batch_copy_nb_elems].dst =
914 (void *)((uintptr_t)(buf_addr + buf_offset));
915 batch_copy[vq->batch_copy_nb_elems].src =
916 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
917 batch_copy[vq->batch_copy_nb_elems].log_addr =
918 buf_iova + buf_offset;
919 batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
920 vq->batch_copy_nb_elems++;
923 mbuf_avail -= cpy_len;
924 mbuf_offset += cpy_len;
925 buf_avail -= cpy_len;
926 buf_offset += cpy_len;
934 static __rte_always_inline void
935 async_fill_vec(struct iovec *v, void *base, size_t len)
941 static __rte_always_inline void
942 async_fill_iter(struct rte_vhost_iov_iter *it, size_t count,
943 struct iovec *vec, unsigned long nr_seg)
950 it->nr_segs = nr_seg;
957 static __rte_always_inline void
958 async_fill_desc(struct rte_vhost_async_desc *desc,
959 struct rte_vhost_iov_iter *src, struct rte_vhost_iov_iter *dst)
965 static __rte_always_inline int
966 async_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
967 struct rte_mbuf *m, struct buf_vector *buf_vec,
968 uint16_t nr_vec, uint16_t num_buffers,
969 struct iovec *src_iovec, struct iovec *dst_iovec,
970 struct rte_vhost_iov_iter *src_it,
971 struct rte_vhost_iov_iter *dst_it)
973 uint32_t vec_idx = 0;
974 uint32_t mbuf_offset, mbuf_avail;
975 uint32_t buf_offset, buf_avail;
976 uint64_t buf_addr, buf_iova, buf_len;
977 uint32_t cpy_len, cpy_threshold;
979 struct rte_mbuf *hdr_mbuf;
980 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
981 struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
989 if (unlikely(m == NULL)) {
994 cpy_threshold = vq->async_threshold;
996 buf_addr = buf_vec[vec_idx].buf_addr;
997 buf_iova = buf_vec[vec_idx].buf_iova;
998 buf_len = buf_vec[vec_idx].buf_len;
1000 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
1006 hdr_addr = buf_addr;
1007 if (unlikely(buf_len < dev->vhost_hlen))
1010 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
1012 VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
1013 dev->vid, num_buffers);
1015 if (unlikely(buf_len < dev->vhost_hlen)) {
1016 buf_offset = dev->vhost_hlen - buf_len;
1018 buf_addr = buf_vec[vec_idx].buf_addr;
1019 buf_iova = buf_vec[vec_idx].buf_iova;
1020 buf_len = buf_vec[vec_idx].buf_len;
1021 buf_avail = buf_len - buf_offset;
1023 buf_offset = dev->vhost_hlen;
1024 buf_avail = buf_len - dev->vhost_hlen;
1027 mbuf_avail = rte_pktmbuf_data_len(m);
1030 while (mbuf_avail != 0 || m->next != NULL) {
1031 /* done with current buf, get the next one */
1032 if (buf_avail == 0) {
1034 if (unlikely(vec_idx >= nr_vec)) {
1039 buf_addr = buf_vec[vec_idx].buf_addr;
1040 buf_iova = buf_vec[vec_idx].buf_iova;
1041 buf_len = buf_vec[vec_idx].buf_len;
1044 buf_avail = buf_len;
1047 /* done with current mbuf, get the next one */
1048 if (mbuf_avail == 0) {
1052 mbuf_avail = rte_pktmbuf_data_len(m);
1056 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
1057 if (rxvq_is_mergeable(dev))
1058 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
1061 if (unlikely(hdr == &tmp_hdr)) {
1062 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
1064 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
1065 dev->vhost_hlen, 0);
1066 vhost_log_cache_write_iova(dev, vq,
1067 buf_vec[0].buf_iova,
1074 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1076 while (unlikely(cpy_len && cpy_len >= cpy_threshold)) {
1077 hpa = (void *)(uintptr_t)gpa_to_first_hpa(dev,
1078 buf_iova + buf_offset,
1079 cpy_len, &mapped_len);
1081 if (unlikely(!hpa || mapped_len < cpy_threshold))
1084 async_fill_vec(src_iovec + tvec_idx,
1085 (void *)(uintptr_t)rte_pktmbuf_iova_offset(m,
1086 mbuf_offset), (size_t)mapped_len);
1088 async_fill_vec(dst_iovec + tvec_idx,
1089 hpa, (size_t)mapped_len);
1091 tlen += (uint32_t)mapped_len;
1092 cpy_len -= (uint32_t)mapped_len;
1093 mbuf_avail -= (uint32_t)mapped_len;
1094 mbuf_offset += (uint32_t)mapped_len;
1095 buf_avail -= (uint32_t)mapped_len;
1096 buf_offset += (uint32_t)mapped_len;
1100 if (likely(cpy_len)) {
1101 if (unlikely(vq->batch_copy_nb_elems >= vq->size)) {
1103 (void *)((uintptr_t)(buf_addr + buf_offset)),
1104 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
1108 (uintptr_t)(buf_addr + buf_offset),
1111 batch_copy[vq->batch_copy_nb_elems].dst =
1112 (void *)((uintptr_t)(buf_addr + buf_offset));
1113 batch_copy[vq->batch_copy_nb_elems].src =
1114 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
1115 batch_copy[vq->batch_copy_nb_elems].log_addr =
1116 buf_iova + buf_offset;
1117 batch_copy[vq->batch_copy_nb_elems].len =
1119 vq->batch_copy_nb_elems++;
1122 mbuf_avail -= cpy_len;
1123 mbuf_offset += cpy_len;
1124 buf_avail -= cpy_len;
1125 buf_offset += cpy_len;
1131 async_fill_iter(src_it, tlen, src_iovec, tvec_idx);
1132 async_fill_iter(dst_it, tlen, dst_iovec, tvec_idx);
1137 static __rte_always_inline int
1138 vhost_enqueue_single_packed(struct virtio_net *dev,
1139 struct vhost_virtqueue *vq,
1140 struct rte_mbuf *pkt,
1141 struct buf_vector *buf_vec,
1144 uint16_t nr_vec = 0;
1145 uint16_t avail_idx = vq->last_avail_idx;
1146 uint16_t max_tries, tries = 0;
1147 uint16_t buf_id = 0;
1149 uint16_t desc_count;
1150 uint32_t size = pkt->pkt_len + sizeof(struct virtio_net_hdr_mrg_rxbuf);
1151 uint16_t num_buffers = 0;
1152 uint32_t buffer_len[vq->size];
1153 uint16_t buffer_buf_id[vq->size];
1154 uint16_t buffer_desc_count[vq->size];
1156 if (rxvq_is_mergeable(dev))
1157 max_tries = vq->size - 1;
1163 * if we tried all available ring items, and still
1164 * can't get enough buf, it means something abnormal
1167 if (unlikely(++tries > max_tries))
1170 if (unlikely(fill_vec_buf_packed(dev, vq,
1171 avail_idx, &desc_count,
1174 VHOST_ACCESS_RW) < 0))
1177 len = RTE_MIN(len, size);
1180 buffer_len[num_buffers] = len;
1181 buffer_buf_id[num_buffers] = buf_id;
1182 buffer_desc_count[num_buffers] = desc_count;
1185 *nr_descs += desc_count;
1186 avail_idx += desc_count;
1187 if (avail_idx >= vq->size)
1188 avail_idx -= vq->size;
1191 if (copy_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, num_buffers) < 0)
1194 vhost_shadow_enqueue_single_packed(dev, vq, buffer_len, buffer_buf_id,
1195 buffer_desc_count, num_buffers);
1200 static __rte_noinline uint32_t
1201 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1202 struct rte_mbuf **pkts, uint32_t count)
1204 uint32_t pkt_idx = 0;
1205 uint16_t num_buffers;
1206 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1207 uint16_t avail_head;
1210 * The ordering between avail index and
1211 * desc reads needs to be enforced.
1213 avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1215 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1217 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1218 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1219 uint16_t nr_vec = 0;
1221 if (unlikely(reserve_avail_buf_split(dev, vq,
1222 pkt_len, buf_vec, &num_buffers,
1223 avail_head, &nr_vec) < 0)) {
1224 VHOST_LOG_DATA(DEBUG,
1225 "(%d) failed to get enough desc from vring\n",
1227 vq->shadow_used_idx -= num_buffers;
1231 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1232 dev->vid, vq->last_avail_idx,
1233 vq->last_avail_idx + num_buffers);
1235 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
1238 vq->shadow_used_idx -= num_buffers;
1242 vq->last_avail_idx += num_buffers;
1245 do_data_copy_enqueue(dev, vq);
1247 if (likely(vq->shadow_used_idx)) {
1248 flush_shadow_used_ring_split(dev, vq);
1249 vhost_vring_call_split(dev, vq);
1255 static __rte_always_inline int
1256 virtio_dev_rx_batch_packed(struct virtio_net *dev,
1257 struct vhost_virtqueue *vq,
1258 struct rte_mbuf **pkts)
1260 bool wrap_counter = vq->avail_wrap_counter;
1261 struct vring_packed_desc *descs = vq->desc_packed;
1262 uint16_t avail_idx = vq->last_avail_idx;
1263 uint64_t desc_addrs[PACKED_BATCH_SIZE];
1264 struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
1265 uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1266 uint64_t lens[PACKED_BATCH_SIZE];
1267 uint16_t ids[PACKED_BATCH_SIZE];
1270 if (unlikely(avail_idx & PACKED_BATCH_MASK))
1273 if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1276 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1277 if (unlikely(pkts[i]->next != NULL))
1279 if (unlikely(!desc_is_avail(&descs[avail_idx + i],
1286 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1287 lens[i] = descs[avail_idx + i].len;
1289 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1290 if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
1294 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1295 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1296 descs[avail_idx + i].addr,
1300 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1301 if (unlikely(!desc_addrs[i]))
1303 if (unlikely(lens[i] != descs[avail_idx + i].len))
1307 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1308 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1309 hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)
1310 (uintptr_t)desc_addrs[i];
1311 lens[i] = pkts[i]->pkt_len +
1312 sizeof(struct virtio_net_hdr_mrg_rxbuf);
1315 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1316 virtio_enqueue_offload(pkts[i], &hdrs[i]->hdr);
1318 vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1320 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1321 rte_memcpy((void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1322 rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1326 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1327 vhost_log_cache_write_iova(dev, vq, descs[avail_idx + i].addr,
1330 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1331 ids[i] = descs[avail_idx + i].id;
1333 vhost_flush_enqueue_batch_packed(dev, vq, lens, ids);
1338 static __rte_always_inline int16_t
1339 virtio_dev_rx_single_packed(struct virtio_net *dev,
1340 struct vhost_virtqueue *vq,
1341 struct rte_mbuf *pkt)
1343 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1344 uint16_t nr_descs = 0;
1347 if (unlikely(vhost_enqueue_single_packed(dev, vq, pkt, buf_vec,
1349 VHOST_LOG_DATA(DEBUG,
1350 "(%d) failed to get enough desc from vring\n",
1355 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1356 dev->vid, vq->last_avail_idx,
1357 vq->last_avail_idx + nr_descs);
1359 vq_inc_last_avail_packed(vq, nr_descs);
1364 static __rte_noinline uint32_t
1365 virtio_dev_rx_packed(struct virtio_net *dev,
1366 struct vhost_virtqueue *__rte_restrict vq,
1367 struct rte_mbuf **__rte_restrict pkts,
1370 uint32_t pkt_idx = 0;
1371 uint32_t remained = count;
1374 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1376 if (remained >= PACKED_BATCH_SIZE) {
1377 if (!virtio_dev_rx_batch_packed(dev, vq,
1379 pkt_idx += PACKED_BATCH_SIZE;
1380 remained -= PACKED_BATCH_SIZE;
1385 if (virtio_dev_rx_single_packed(dev, vq, pkts[pkt_idx]))
1390 } while (pkt_idx < count);
1392 if (vq->shadow_used_idx) {
1393 do_data_copy_enqueue(dev, vq);
1394 vhost_flush_enqueue_shadow_packed(dev, vq);
1398 vhost_vring_call_packed(dev, vq);
1403 static __rte_always_inline uint32_t
1404 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
1405 struct rte_mbuf **pkts, uint32_t count)
1407 struct vhost_virtqueue *vq;
1410 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1411 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1412 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1413 dev->vid, __func__, queue_id);
1417 vq = dev->virtqueue[queue_id];
1419 rte_spinlock_lock(&vq->access_lock);
1421 if (unlikely(vq->enabled == 0))
1422 goto out_access_unlock;
1424 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1425 vhost_user_iotlb_rd_lock(vq);
1427 if (unlikely(vq->access_ok == 0))
1428 if (unlikely(vring_translate(dev, vq) < 0))
1431 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1435 if (vq_is_packed(dev))
1436 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
1438 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
1441 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1442 vhost_user_iotlb_rd_unlock(vq);
1445 rte_spinlock_unlock(&vq->access_lock);
1451 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
1452 struct rte_mbuf **__rte_restrict pkts, uint16_t count)
1454 struct virtio_net *dev = get_device(vid);
1459 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1461 "(%d) %s: built-in vhost net backend is disabled.\n",
1462 dev->vid, __func__);
1466 return virtio_dev_rx(dev, queue_id, pkts, count);
1469 static __rte_always_inline uint16_t
1470 virtio_dev_rx_async_get_info_idx(uint16_t pkts_idx,
1471 uint16_t vq_size, uint16_t n_inflight)
1473 return pkts_idx > n_inflight ? (pkts_idx - n_inflight) :
1474 (vq_size - n_inflight + pkts_idx) & (vq_size - 1);
1477 static __rte_always_inline void
1478 virtio_dev_rx_async_submit_split_err(struct virtio_net *dev,
1479 struct vhost_virtqueue *vq, uint16_t queue_id,
1480 uint16_t last_idx, uint16_t shadow_idx)
1482 uint16_t start_idx, pkts_idx, vq_size;
1483 uint64_t *async_pending_info;
1485 pkts_idx = vq->async_pkts_idx;
1486 async_pending_info = vq->async_pending_info;
1488 start_idx = virtio_dev_rx_async_get_info_idx(pkts_idx,
1489 vq_size, vq->async_pkts_inflight_n);
1491 while (likely((start_idx & (vq_size - 1)) != pkts_idx)) {
1493 async_pending_info[(start_idx) & (vq_size - 1)] >>
1494 ASYNC_PENDING_INFO_N_SFT;
1497 n_seg -= vq->async_ops.check_completed_copies(dev->vid,
1501 vq->async_pkts_inflight_n = 0;
1502 vq->batch_copy_nb_elems = 0;
1504 vq->shadow_used_idx = shadow_idx;
1505 vq->last_avail_idx = last_idx;
1508 static __rte_noinline uint32_t
1509 virtio_dev_rx_async_submit_split(struct virtio_net *dev,
1510 struct vhost_virtqueue *vq, uint16_t queue_id,
1511 struct rte_mbuf **pkts, uint32_t count)
1513 uint32_t pkt_idx = 0, pkt_burst_idx = 0;
1514 uint16_t num_buffers;
1515 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1516 uint16_t avail_head, last_idx, shadow_idx;
1518 struct rte_vhost_iov_iter *it_pool = vq->it_pool;
1519 struct iovec *vec_pool = vq->vec_pool;
1520 struct rte_vhost_async_desc tdes[MAX_PKT_BURST];
1521 struct iovec *src_iovec = vec_pool;
1522 struct iovec *dst_iovec = vec_pool + (VHOST_MAX_ASYNC_VEC >> 1);
1523 struct rte_vhost_iov_iter *src_it = it_pool;
1524 struct rte_vhost_iov_iter *dst_it = it_pool + 1;
1525 uint16_t n_free_slot, slot_idx;
1528 avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1529 last_idx = vq->last_avail_idx;
1530 shadow_idx = vq->shadow_used_idx;
1533 * The ordering between avail index and
1534 * desc reads needs to be enforced.
1538 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1540 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1541 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1542 uint16_t nr_vec = 0;
1544 if (unlikely(reserve_avail_buf_split(dev, vq,
1545 pkt_len, buf_vec, &num_buffers,
1546 avail_head, &nr_vec) < 0)) {
1547 VHOST_LOG_DATA(DEBUG,
1548 "(%d) failed to get enough desc from vring\n",
1550 vq->shadow_used_idx -= num_buffers;
1554 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1555 dev->vid, vq->last_avail_idx,
1556 vq->last_avail_idx + num_buffers);
1558 if (async_mbuf_to_desc(dev, vq, pkts[pkt_idx],
1559 buf_vec, nr_vec, num_buffers,
1560 src_iovec, dst_iovec, src_it, dst_it) < 0) {
1561 vq->shadow_used_idx -= num_buffers;
1565 slot_idx = (vq->async_pkts_idx + pkt_idx) & (vq->size - 1);
1566 if (src_it->count) {
1567 async_fill_desc(&tdes[pkt_burst_idx], src_it, dst_it);
1569 vq->async_pending_info[slot_idx] =
1570 num_buffers | (src_it->nr_segs << 16);
1571 src_iovec += src_it->nr_segs;
1572 dst_iovec += dst_it->nr_segs;
1576 vq->async_pending_info[slot_idx] = num_buffers;
1577 vq->async_pkts_inflight_n++;
1580 vq->last_avail_idx += num_buffers;
1582 if (pkt_burst_idx >= VHOST_ASYNC_BATCH_THRESHOLD ||
1583 (pkt_idx == count - 1 && pkt_burst_idx)) {
1584 n_pkts = vq->async_ops.transfer_data(dev->vid,
1585 queue_id, tdes, 0, pkt_burst_idx);
1586 src_iovec = vec_pool;
1587 dst_iovec = vec_pool + (VHOST_MAX_ASYNC_VEC >> 1);
1589 dst_it = it_pool + 1;
1591 if (unlikely(n_pkts < (int)pkt_burst_idx)) {
1592 vq->async_pkts_inflight_n +=
1593 n_pkts > 0 ? n_pkts : 0;
1594 virtio_dev_rx_async_submit_split_err(dev,
1595 vq, queue_id, last_idx, shadow_idx);
1600 vq->async_pkts_inflight_n += n_pkts;
1604 if (pkt_burst_idx) {
1605 n_pkts = vq->async_ops.transfer_data(dev->vid,
1606 queue_id, tdes, 0, pkt_burst_idx);
1607 if (unlikely(n_pkts < (int)pkt_burst_idx)) {
1608 vq->async_pkts_inflight_n += n_pkts > 0 ? n_pkts : 0;
1609 virtio_dev_rx_async_submit_split_err(dev, vq, queue_id,
1610 last_idx, shadow_idx);
1614 vq->async_pkts_inflight_n += n_pkts;
1617 do_data_copy_enqueue(dev, vq);
1619 n_free_slot = vq->size - vq->async_pkts_idx;
1620 if (n_free_slot > pkt_idx) {
1621 rte_memcpy(&vq->async_pkts_pending[vq->async_pkts_idx],
1622 pkts, pkt_idx * sizeof(uintptr_t));
1623 vq->async_pkts_idx += pkt_idx;
1625 rte_memcpy(&vq->async_pkts_pending[vq->async_pkts_idx],
1626 pkts, n_free_slot * sizeof(uintptr_t));
1627 rte_memcpy(&vq->async_pkts_pending[0],
1629 (pkt_idx - n_free_slot) * sizeof(uintptr_t));
1630 vq->async_pkts_idx = pkt_idx - n_free_slot;
1633 if (likely(vq->shadow_used_idx))
1634 async_flush_shadow_used_ring_split(dev, vq);
1639 uint16_t rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
1640 struct rte_mbuf **pkts, uint16_t count)
1642 struct virtio_net *dev = get_device(vid);
1643 struct vhost_virtqueue *vq;
1644 uint16_t n_segs_cpl, n_pkts_put = 0, n_descs = 0;
1645 uint16_t start_idx, pkts_idx, vq_size;
1646 uint16_t n_inflight;
1647 uint64_t *async_pending_info;
1652 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1653 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1654 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1655 dev->vid, __func__, queue_id);
1659 vq = dev->virtqueue[queue_id];
1661 rte_spinlock_lock(&vq->access_lock);
1663 n_inflight = vq->async_pkts_inflight_n;
1664 pkts_idx = vq->async_pkts_idx;
1665 async_pending_info = vq->async_pending_info;
1667 start_idx = virtio_dev_rx_async_get_info_idx(pkts_idx,
1668 vq_size, vq->async_pkts_inflight_n);
1670 n_segs_cpl = vq->async_ops.check_completed_copies(vid, queue_id,
1671 0, ASYNC_MAX_POLL_SEG - vq->async_last_seg_n) +
1672 vq->async_last_seg_n;
1676 while (likely((n_pkts_put < count) && n_inflight)) {
1677 uint64_t info = async_pending_info[
1678 (start_idx + n_pkts_put) & (vq_size - 1)];
1682 n_descs += info & ASYNC_PENDING_INFO_N_MSK;
1683 n_segs = info >> ASYNC_PENDING_INFO_N_SFT;
1686 if (unlikely(n_segs_cpl < n_segs)) {
1689 n_descs -= info & ASYNC_PENDING_INFO_N_MSK;
1692 (start_idx + n_pkts_put) &
1694 ((n_segs - n_segs_cpl) <<
1695 ASYNC_PENDING_INFO_N_SFT) |
1696 (info & ASYNC_PENDING_INFO_N_MSK);
1701 n_segs_cpl -= n_segs;
1705 vq->async_last_seg_n = n_segs_cpl;
1708 vq->async_pkts_inflight_n = n_inflight;
1709 if (likely(vq->enabled && vq->access_ok)) {
1710 __atomic_add_fetch(&vq->used->idx,
1711 n_descs, __ATOMIC_RELEASE);
1712 vhost_vring_call_split(dev, vq);
1716 if (start_idx + n_pkts_put <= vq_size) {
1717 rte_memcpy(pkts, &vq->async_pkts_pending[start_idx],
1718 n_pkts_put * sizeof(uintptr_t));
1720 rte_memcpy(pkts, &vq->async_pkts_pending[start_idx],
1721 (vq_size - start_idx) * sizeof(uintptr_t));
1722 rte_memcpy(&pkts[vq_size - start_idx], vq->async_pkts_pending,
1723 (n_pkts_put - vq_size + start_idx) * sizeof(uintptr_t));
1726 rte_spinlock_unlock(&vq->access_lock);
1731 static __rte_always_inline uint32_t
1732 virtio_dev_rx_async_submit(struct virtio_net *dev, uint16_t queue_id,
1733 struct rte_mbuf **pkts, uint32_t count)
1735 struct vhost_virtqueue *vq;
1737 bool drawback = false;
1739 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1740 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1741 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1742 dev->vid, __func__, queue_id);
1746 vq = dev->virtqueue[queue_id];
1748 rte_spinlock_lock(&vq->access_lock);
1750 if (unlikely(vq->enabled == 0))
1751 goto out_access_unlock;
1753 if (unlikely(!vq->async_registered)) {
1755 goto out_access_unlock;
1758 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1759 vhost_user_iotlb_rd_lock(vq);
1761 if (unlikely(vq->access_ok == 0))
1762 if (unlikely(vring_translate(dev, vq) < 0))
1765 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1769 /* TODO: packed queue not implemented */
1770 if (vq_is_packed(dev))
1773 nb_tx = virtio_dev_rx_async_submit_split(dev,
1774 vq, queue_id, pkts, count);
1777 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1778 vhost_user_iotlb_rd_unlock(vq);
1781 rte_spinlock_unlock(&vq->access_lock);
1784 return rte_vhost_enqueue_burst(dev->vid, queue_id, pkts, count);
1790 rte_vhost_submit_enqueue_burst(int vid, uint16_t queue_id,
1791 struct rte_mbuf **pkts, uint16_t count)
1793 struct virtio_net *dev = get_device(vid);
1798 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1800 "(%d) %s: built-in vhost net backend is disabled.\n",
1801 dev->vid, __func__);
1805 return virtio_dev_rx_async_submit(dev, queue_id, pkts, count);
1809 virtio_net_with_host_offload(struct virtio_net *dev)
1812 ((1ULL << VIRTIO_NET_F_CSUM) |
1813 (1ULL << VIRTIO_NET_F_HOST_ECN) |
1814 (1ULL << VIRTIO_NET_F_HOST_TSO4) |
1815 (1ULL << VIRTIO_NET_F_HOST_TSO6) |
1816 (1ULL << VIRTIO_NET_F_HOST_UFO)))
1823 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
1825 struct rte_ipv4_hdr *ipv4_hdr;
1826 struct rte_ipv6_hdr *ipv6_hdr;
1827 void *l3_hdr = NULL;
1828 struct rte_ether_hdr *eth_hdr;
1831 eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
1833 m->l2_len = sizeof(struct rte_ether_hdr);
1834 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
1836 if (ethertype == RTE_ETHER_TYPE_VLAN) {
1837 struct rte_vlan_hdr *vlan_hdr =
1838 (struct rte_vlan_hdr *)(eth_hdr + 1);
1840 m->l2_len += sizeof(struct rte_vlan_hdr);
1841 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
1844 l3_hdr = (char *)eth_hdr + m->l2_len;
1846 switch (ethertype) {
1847 case RTE_ETHER_TYPE_IPV4:
1849 *l4_proto = ipv4_hdr->next_proto_id;
1850 m->l3_len = rte_ipv4_hdr_len(ipv4_hdr);
1851 *l4_hdr = (char *)l3_hdr + m->l3_len;
1852 m->ol_flags |= PKT_TX_IPV4;
1854 case RTE_ETHER_TYPE_IPV6:
1856 *l4_proto = ipv6_hdr->proto;
1857 m->l3_len = sizeof(struct rte_ipv6_hdr);
1858 *l4_hdr = (char *)l3_hdr + m->l3_len;
1859 m->ol_flags |= PKT_TX_IPV6;
1869 static __rte_always_inline void
1870 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
1872 uint16_t l4_proto = 0;
1873 void *l4_hdr = NULL;
1874 struct rte_tcp_hdr *tcp_hdr = NULL;
1876 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1879 parse_ethernet(m, &l4_proto, &l4_hdr);
1880 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1881 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
1882 switch (hdr->csum_offset) {
1883 case (offsetof(struct rte_tcp_hdr, cksum)):
1884 if (l4_proto == IPPROTO_TCP)
1885 m->ol_flags |= PKT_TX_TCP_CKSUM;
1887 case (offsetof(struct rte_udp_hdr, dgram_cksum)):
1888 if (l4_proto == IPPROTO_UDP)
1889 m->ol_flags |= PKT_TX_UDP_CKSUM;
1891 case (offsetof(struct rte_sctp_hdr, cksum)):
1892 if (l4_proto == IPPROTO_SCTP)
1893 m->ol_flags |= PKT_TX_SCTP_CKSUM;
1901 if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1902 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1903 case VIRTIO_NET_HDR_GSO_TCPV4:
1904 case VIRTIO_NET_HDR_GSO_TCPV6:
1906 m->ol_flags |= PKT_TX_TCP_SEG;
1907 m->tso_segsz = hdr->gso_size;
1908 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
1910 case VIRTIO_NET_HDR_GSO_UDP:
1911 m->ol_flags |= PKT_TX_UDP_SEG;
1912 m->tso_segsz = hdr->gso_size;
1913 m->l4_len = sizeof(struct rte_udp_hdr);
1916 VHOST_LOG_DATA(WARNING,
1917 "unsupported gso type %u.\n", hdr->gso_type);
1923 static __rte_noinline void
1924 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
1925 struct buf_vector *buf_vec)
1928 uint64_t remain = sizeof(struct virtio_net_hdr);
1930 uint64_t dst = (uint64_t)(uintptr_t)hdr;
1933 len = RTE_MIN(remain, buf_vec->buf_len);
1934 src = buf_vec->buf_addr;
1935 rte_memcpy((void *)(uintptr_t)dst,
1936 (void *)(uintptr_t)src, len);
1944 static __rte_always_inline int
1945 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
1946 struct buf_vector *buf_vec, uint16_t nr_vec,
1947 struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
1949 uint32_t buf_avail, buf_offset;
1950 uint64_t buf_addr, buf_len;
1951 uint32_t mbuf_avail, mbuf_offset;
1953 struct rte_mbuf *cur = m, *prev = m;
1954 struct virtio_net_hdr tmp_hdr;
1955 struct virtio_net_hdr *hdr = NULL;
1956 /* A counter to avoid desc dead loop chain */
1957 uint16_t vec_idx = 0;
1958 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
1961 buf_addr = buf_vec[vec_idx].buf_addr;
1962 buf_len = buf_vec[vec_idx].buf_len;
1964 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
1969 if (virtio_net_with_host_offload(dev)) {
1970 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
1972 * No luck, the virtio-net header doesn't fit
1973 * in a contiguous virtual area.
1975 copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
1978 hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
1983 * A virtio driver normally uses at least 2 desc buffers
1984 * for Tx: the first for storing the header, and others
1985 * for storing the data.
1987 if (unlikely(buf_len < dev->vhost_hlen)) {
1988 buf_offset = dev->vhost_hlen - buf_len;
1990 buf_addr = buf_vec[vec_idx].buf_addr;
1991 buf_len = buf_vec[vec_idx].buf_len;
1992 buf_avail = buf_len - buf_offset;
1993 } else if (buf_len == dev->vhost_hlen) {
1994 if (unlikely(++vec_idx >= nr_vec))
1996 buf_addr = buf_vec[vec_idx].buf_addr;
1997 buf_len = buf_vec[vec_idx].buf_len;
2000 buf_avail = buf_len;
2002 buf_offset = dev->vhost_hlen;
2003 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
2007 (uintptr_t)(buf_addr + buf_offset),
2008 (uint32_t)buf_avail, 0);
2011 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
2013 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
2015 if (likely(cpy_len > MAX_BATCH_LEN ||
2016 vq->batch_copy_nb_elems >= vq->size ||
2017 (hdr && cur == m))) {
2018 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
2020 (void *)((uintptr_t)(buf_addr +
2021 buf_offset)), cpy_len);
2023 batch_copy[vq->batch_copy_nb_elems].dst =
2024 rte_pktmbuf_mtod_offset(cur, void *,
2026 batch_copy[vq->batch_copy_nb_elems].src =
2027 (void *)((uintptr_t)(buf_addr + buf_offset));
2028 batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
2029 vq->batch_copy_nb_elems++;
2032 mbuf_avail -= cpy_len;
2033 mbuf_offset += cpy_len;
2034 buf_avail -= cpy_len;
2035 buf_offset += cpy_len;
2037 /* This buf reaches to its end, get the next one */
2038 if (buf_avail == 0) {
2039 if (++vec_idx >= nr_vec)
2042 buf_addr = buf_vec[vec_idx].buf_addr;
2043 buf_len = buf_vec[vec_idx].buf_len;
2046 buf_avail = buf_len;
2048 PRINT_PACKET(dev, (uintptr_t)buf_addr,
2049 (uint32_t)buf_avail, 0);
2053 * This mbuf reaches to its end, get a new one
2054 * to hold more data.
2056 if (mbuf_avail == 0) {
2057 cur = rte_pktmbuf_alloc(mbuf_pool);
2058 if (unlikely(cur == NULL)) {
2059 VHOST_LOG_DATA(ERR, "Failed to "
2060 "allocate memory for mbuf.\n");
2066 prev->data_len = mbuf_offset;
2068 m->pkt_len += mbuf_offset;
2072 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
2076 prev->data_len = mbuf_offset;
2077 m->pkt_len += mbuf_offset;
2080 vhost_dequeue_offload(hdr, m);
2088 virtio_dev_extbuf_free(void *addr __rte_unused, void *opaque)
2094 virtio_dev_extbuf_alloc(struct rte_mbuf *pkt, uint32_t size)
2096 struct rte_mbuf_ext_shared_info *shinfo = NULL;
2097 uint32_t total_len = RTE_PKTMBUF_HEADROOM + size;
2102 total_len += sizeof(*shinfo) + sizeof(uintptr_t);
2103 total_len = RTE_ALIGN_CEIL(total_len, sizeof(uintptr_t));
2105 if (unlikely(total_len > UINT16_MAX))
2108 buf_len = total_len;
2109 buf = rte_malloc(NULL, buf_len, RTE_CACHE_LINE_SIZE);
2110 if (unlikely(buf == NULL))
2113 /* Initialize shinfo */
2114 shinfo = rte_pktmbuf_ext_shinfo_init_helper(buf, &buf_len,
2115 virtio_dev_extbuf_free, buf);
2116 if (unlikely(shinfo == NULL)) {
2118 VHOST_LOG_DATA(ERR, "Failed to init shinfo\n");
2122 iova = rte_malloc_virt2iova(buf);
2123 rte_pktmbuf_attach_extbuf(pkt, buf, iova, buf_len, shinfo);
2124 rte_pktmbuf_reset_headroom(pkt);
2130 * Allocate a host supported pktmbuf.
2132 static __rte_always_inline struct rte_mbuf *
2133 virtio_dev_pktmbuf_alloc(struct virtio_net *dev, struct rte_mempool *mp,
2136 struct rte_mbuf *pkt = rte_pktmbuf_alloc(mp);
2138 if (unlikely(pkt == NULL)) {
2140 "Failed to allocate memory for mbuf.\n");
2144 if (rte_pktmbuf_tailroom(pkt) >= data_len)
2147 /* attach an external buffer if supported */
2148 if (dev->extbuf && !virtio_dev_extbuf_alloc(pkt, data_len))
2151 /* check if chained buffers are allowed */
2152 if (!dev->linearbuf)
2155 /* Data doesn't fit into the buffer and the host supports
2156 * only linear buffers
2158 rte_pktmbuf_free(pkt);
2163 static __rte_noinline uint16_t
2164 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
2165 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
2168 uint16_t free_entries;
2169 uint16_t dropped = 0;
2170 static bool allocerr_warned;
2173 * The ordering between avail index and
2174 * desc reads needs to be enforced.
2176 free_entries = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE) -
2178 if (free_entries == 0)
2181 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
2183 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2185 count = RTE_MIN(count, MAX_PKT_BURST);
2186 count = RTE_MIN(count, free_entries);
2187 VHOST_LOG_DATA(DEBUG, "(%d) about to dequeue %u buffers\n",
2190 for (i = 0; i < count; i++) {
2191 struct buf_vector buf_vec[BUF_VECTOR_MAX];
2194 uint16_t nr_vec = 0;
2197 if (unlikely(fill_vec_buf_split(dev, vq,
2198 vq->last_avail_idx + i,
2200 &head_idx, &buf_len,
2201 VHOST_ACCESS_RO) < 0))
2204 update_shadow_used_ring_split(vq, head_idx, 0);
2206 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
2207 if (unlikely(pkts[i] == NULL)) {
2209 * mbuf allocation fails for jumbo packets when external
2210 * buffer allocation is not allowed and linear buffer
2211 * is required. Drop this packet.
2213 if (!allocerr_warned) {
2215 "Failed mbuf alloc of size %d from %s on %s.\n",
2216 buf_len, mbuf_pool->name, dev->ifname);
2217 allocerr_warned = true;
2224 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
2226 if (unlikely(err)) {
2227 rte_pktmbuf_free(pkts[i]);
2228 if (!allocerr_warned) {
2230 "Failed to copy desc to mbuf on %s.\n",
2232 allocerr_warned = true;
2240 vq->last_avail_idx += i;
2242 do_data_copy_dequeue(vq);
2243 if (unlikely(i < count))
2244 vq->shadow_used_idx = i;
2245 if (likely(vq->shadow_used_idx)) {
2246 flush_shadow_used_ring_split(dev, vq);
2247 vhost_vring_call_split(dev, vq);
2250 return (i - dropped);
2253 static __rte_always_inline int
2254 vhost_reserve_avail_batch_packed(struct virtio_net *dev,
2255 struct vhost_virtqueue *vq,
2256 struct rte_mempool *mbuf_pool,
2257 struct rte_mbuf **pkts,
2259 uintptr_t *desc_addrs,
2262 bool wrap = vq->avail_wrap_counter;
2263 struct vring_packed_desc *descs = vq->desc_packed;
2264 struct virtio_net_hdr *hdr;
2265 uint64_t lens[PACKED_BATCH_SIZE];
2266 uint64_t buf_lens[PACKED_BATCH_SIZE];
2267 uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2270 if (unlikely(avail_idx & PACKED_BATCH_MASK))
2272 if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
2275 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2276 flags = descs[avail_idx + i].flags;
2277 if (unlikely((wrap != !!(flags & VRING_DESC_F_AVAIL)) ||
2278 (wrap == !!(flags & VRING_DESC_F_USED)) ||
2279 (flags & PACKED_DESC_SINGLE_DEQUEUE_FLAG)))
2285 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2286 lens[i] = descs[avail_idx + i].len;
2288 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2289 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
2290 descs[avail_idx + i].addr,
2291 &lens[i], VHOST_ACCESS_RW);
2294 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2295 if (unlikely(!desc_addrs[i]))
2297 if (unlikely((lens[i] != descs[avail_idx + i].len)))
2301 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2302 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, lens[i]);
2307 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2308 buf_lens[i] = pkts[i]->buf_len - pkts[i]->data_off;
2310 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2311 if (unlikely(buf_lens[i] < (lens[i] - buf_offset)))
2315 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2316 pkts[i]->pkt_len = descs[avail_idx + i].len - buf_offset;
2317 pkts[i]->data_len = pkts[i]->pkt_len;
2318 ids[i] = descs[avail_idx + i].id;
2321 if (virtio_net_with_host_offload(dev)) {
2322 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2323 hdr = (struct virtio_net_hdr *)(desc_addrs[i]);
2324 vhost_dequeue_offload(hdr, pkts[i]);
2331 for (i = 0; i < PACKED_BATCH_SIZE; i++)
2332 rte_pktmbuf_free(pkts[i]);
2337 static __rte_always_inline int
2338 virtio_dev_tx_batch_packed(struct virtio_net *dev,
2339 struct vhost_virtqueue *vq,
2340 struct rte_mempool *mbuf_pool,
2341 struct rte_mbuf **pkts)
2343 uint16_t avail_idx = vq->last_avail_idx;
2344 uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2345 uintptr_t desc_addrs[PACKED_BATCH_SIZE];
2346 uint16_t ids[PACKED_BATCH_SIZE];
2349 if (vhost_reserve_avail_batch_packed(dev, vq, mbuf_pool, pkts,
2350 avail_idx, desc_addrs, ids))
2353 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2354 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
2356 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2357 rte_memcpy(rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
2358 (void *)(uintptr_t)(desc_addrs[i] + buf_offset),
2361 if (virtio_net_is_inorder(dev))
2362 vhost_shadow_dequeue_batch_packed_inorder(vq,
2363 ids[PACKED_BATCH_SIZE - 1]);
2365 vhost_shadow_dequeue_batch_packed(dev, vq, ids);
2367 vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
2372 static __rte_always_inline int
2373 vhost_dequeue_single_packed(struct virtio_net *dev,
2374 struct vhost_virtqueue *vq,
2375 struct rte_mempool *mbuf_pool,
2376 struct rte_mbuf **pkts,
2378 uint16_t *desc_count)
2380 struct buf_vector buf_vec[BUF_VECTOR_MAX];
2382 uint16_t nr_vec = 0;
2384 static bool allocerr_warned;
2386 if (unlikely(fill_vec_buf_packed(dev, vq,
2387 vq->last_avail_idx, desc_count,
2390 VHOST_ACCESS_RO) < 0))
2393 *pkts = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
2394 if (unlikely(*pkts == NULL)) {
2395 if (!allocerr_warned) {
2397 "Failed mbuf alloc of size %d from %s on %s.\n",
2398 buf_len, mbuf_pool->name, dev->ifname);
2399 allocerr_warned = true;
2404 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, *pkts,
2406 if (unlikely(err)) {
2407 if (!allocerr_warned) {
2409 "Failed to copy desc to mbuf on %s.\n",
2411 allocerr_warned = true;
2413 rte_pktmbuf_free(*pkts);
2420 static __rte_always_inline int
2421 virtio_dev_tx_single_packed(struct virtio_net *dev,
2422 struct vhost_virtqueue *vq,
2423 struct rte_mempool *mbuf_pool,
2424 struct rte_mbuf **pkts)
2427 uint16_t buf_id, desc_count = 0;
2430 ret = vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
2433 if (likely(desc_count > 0)) {
2434 if (virtio_net_is_inorder(dev))
2435 vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
2438 vhost_shadow_dequeue_single_packed(vq, buf_id,
2441 vq_inc_last_avail_packed(vq, desc_count);
2447 static __rte_noinline uint16_t
2448 virtio_dev_tx_packed(struct virtio_net *dev,
2449 struct vhost_virtqueue *__rte_restrict vq,
2450 struct rte_mempool *mbuf_pool,
2451 struct rte_mbuf **__rte_restrict pkts,
2454 uint32_t pkt_idx = 0;
2455 uint32_t remained = count;
2458 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
2460 if (remained >= PACKED_BATCH_SIZE) {
2461 if (!virtio_dev_tx_batch_packed(dev, vq, mbuf_pool,
2463 pkt_idx += PACKED_BATCH_SIZE;
2464 remained -= PACKED_BATCH_SIZE;
2469 if (virtio_dev_tx_single_packed(dev, vq, mbuf_pool,
2477 if (vq->shadow_used_idx) {
2478 do_data_copy_dequeue(vq);
2480 vhost_flush_dequeue_shadow_packed(dev, vq);
2481 vhost_vring_call_packed(dev, vq);
2488 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
2489 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
2491 struct virtio_net *dev;
2492 struct rte_mbuf *rarp_mbuf = NULL;
2493 struct vhost_virtqueue *vq;
2494 int16_t success = 1;
2496 dev = get_device(vid);
2500 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
2502 "(%d) %s: built-in vhost net backend is disabled.\n",
2503 dev->vid, __func__);
2507 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
2509 "(%d) %s: invalid virtqueue idx %d.\n",
2510 dev->vid, __func__, queue_id);
2514 vq = dev->virtqueue[queue_id];
2516 if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
2519 if (unlikely(vq->enabled == 0)) {
2521 goto out_access_unlock;
2524 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2525 vhost_user_iotlb_rd_lock(vq);
2527 if (unlikely(vq->access_ok == 0))
2528 if (unlikely(vring_translate(dev, vq) < 0)) {
2534 * Construct a RARP broadcast packet, and inject it to the "pkts"
2535 * array, to looks like that guest actually send such packet.
2537 * Check user_send_rarp() for more information.
2539 * broadcast_rarp shares a cacheline in the virtio_net structure
2540 * with some fields that are accessed during enqueue and
2541 * __atomic_compare_exchange_n causes a write if performed compare
2542 * and exchange. This could result in false sharing between enqueue
2545 * Prevent unnecessary false sharing by reading broadcast_rarp first
2546 * and only performing compare and exchange if the read indicates it
2547 * is likely to be set.
2549 if (unlikely(__atomic_load_n(&dev->broadcast_rarp, __ATOMIC_ACQUIRE) &&
2550 __atomic_compare_exchange_n(&dev->broadcast_rarp,
2551 &success, 0, 0, __ATOMIC_RELEASE, __ATOMIC_RELAXED))) {
2553 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
2554 if (rarp_mbuf == NULL) {
2555 VHOST_LOG_DATA(ERR, "Failed to make RARP packet.\n");
2562 if (vq_is_packed(dev))
2563 count = virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count);
2565 count = virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count);
2568 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2569 vhost_user_iotlb_rd_unlock(vq);
2572 rte_spinlock_unlock(&vq->access_lock);
2574 if (unlikely(rarp_mbuf != NULL)) {
2576 * Inject it to the head of "pkts" array, so that switch's mac
2577 * learning table will get updated first.
2579 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
2580 pkts[0] = rarp_mbuf;