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>
12 #include <rte_ether.h>
14 #include <rte_vhost.h>
19 #include <rte_spinlock.h>
20 #include <rte_malloc.h>
21 #include <rte_vhost_async.h>
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;
47 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
49 struct batch_copy_elem *elem = vq->batch_copy_elems;
50 uint16_t count = vq->batch_copy_nb_elems;
53 for (i = 0; i < count; i++) {
54 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
55 vhost_log_cache_write_iova(dev, vq, elem[i].log_addr,
57 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
60 vq->batch_copy_nb_elems = 0;
64 do_data_copy_dequeue(struct vhost_virtqueue *vq)
66 struct batch_copy_elem *elem = vq->batch_copy_elems;
67 uint16_t count = vq->batch_copy_nb_elems;
70 for (i = 0; i < count; i++)
71 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
73 vq->batch_copy_nb_elems = 0;
76 static __rte_always_inline void
77 do_flush_shadow_used_ring_split(struct virtio_net *dev,
78 struct vhost_virtqueue *vq,
79 uint16_t to, uint16_t from, uint16_t size)
81 rte_memcpy(&vq->used->ring[to],
82 &vq->shadow_used_split[from],
83 size * sizeof(struct vring_used_elem));
84 vhost_log_cache_used_vring(dev, vq,
85 offsetof(struct vring_used, ring[to]),
86 size * sizeof(struct vring_used_elem));
89 static __rte_always_inline void
90 flush_shadow_used_ring_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
92 uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
94 if (used_idx + vq->shadow_used_idx <= vq->size) {
95 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0,
100 /* update used ring interval [used_idx, vq->size] */
101 size = vq->size - used_idx;
102 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0, size);
104 /* update the left half used ring interval [0, left_size] */
105 do_flush_shadow_used_ring_split(dev, vq, 0, size,
106 vq->shadow_used_idx - size);
108 vq->last_used_idx += vq->shadow_used_idx;
110 vhost_log_cache_sync(dev, vq);
112 __atomic_add_fetch(&vq->used->idx, vq->shadow_used_idx,
114 vq->shadow_used_idx = 0;
115 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
116 sizeof(vq->used->idx));
119 static __rte_always_inline void
120 update_shadow_used_ring_split(struct vhost_virtqueue *vq,
121 uint16_t desc_idx, uint32_t len)
123 uint16_t i = vq->shadow_used_idx++;
125 vq->shadow_used_split[i].id = desc_idx;
126 vq->shadow_used_split[i].len = len;
129 static __rte_always_inline void
130 vhost_flush_enqueue_shadow_packed(struct virtio_net *dev,
131 struct vhost_virtqueue *vq)
134 uint16_t used_idx = vq->last_used_idx;
135 uint16_t head_idx = vq->last_used_idx;
136 uint16_t head_flags = 0;
138 /* Split loop in two to save memory barriers */
139 for (i = 0; i < vq->shadow_used_idx; i++) {
140 vq->desc_packed[used_idx].id = vq->shadow_used_packed[i].id;
141 vq->desc_packed[used_idx].len = vq->shadow_used_packed[i].len;
143 used_idx += vq->shadow_used_packed[i].count;
144 if (used_idx >= vq->size)
145 used_idx -= vq->size;
148 /* The ordering for storing desc flags needs to be enforced. */
149 rte_atomic_thread_fence(__ATOMIC_RELEASE);
151 for (i = 0; i < vq->shadow_used_idx; i++) {
154 if (vq->shadow_used_packed[i].len)
155 flags = VRING_DESC_F_WRITE;
159 if (vq->used_wrap_counter) {
160 flags |= VRING_DESC_F_USED;
161 flags |= VRING_DESC_F_AVAIL;
163 flags &= ~VRING_DESC_F_USED;
164 flags &= ~VRING_DESC_F_AVAIL;
168 vq->desc_packed[vq->last_used_idx].flags = flags;
170 vhost_log_cache_used_vring(dev, vq,
172 sizeof(struct vring_packed_desc),
173 sizeof(struct vring_packed_desc));
175 head_idx = vq->last_used_idx;
179 vq_inc_last_used_packed(vq, vq->shadow_used_packed[i].count);
182 vq->desc_packed[head_idx].flags = head_flags;
184 vhost_log_cache_used_vring(dev, vq,
186 sizeof(struct vring_packed_desc),
187 sizeof(struct vring_packed_desc));
189 vq->shadow_used_idx = 0;
190 vhost_log_cache_sync(dev, vq);
193 static __rte_always_inline void
194 vhost_flush_dequeue_shadow_packed(struct virtio_net *dev,
195 struct vhost_virtqueue *vq)
197 struct vring_used_elem_packed *used_elem = &vq->shadow_used_packed[0];
199 vq->desc_packed[vq->shadow_last_used_idx].id = used_elem->id;
200 /* desc flags is the synchronization point for virtio packed vring */
201 __atomic_store_n(&vq->desc_packed[vq->shadow_last_used_idx].flags,
202 used_elem->flags, __ATOMIC_RELEASE);
204 vhost_log_cache_used_vring(dev, vq, vq->shadow_last_used_idx *
205 sizeof(struct vring_packed_desc),
206 sizeof(struct vring_packed_desc));
207 vq->shadow_used_idx = 0;
208 vhost_log_cache_sync(dev, vq);
211 static __rte_always_inline void
212 vhost_flush_enqueue_batch_packed(struct virtio_net *dev,
213 struct vhost_virtqueue *vq,
219 uint16_t last_used_idx;
220 struct vring_packed_desc *desc_base;
222 last_used_idx = vq->last_used_idx;
223 desc_base = &vq->desc_packed[last_used_idx];
225 flags = PACKED_DESC_ENQUEUE_USED_FLAG(vq->used_wrap_counter);
227 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
228 desc_base[i].id = ids[i];
229 desc_base[i].len = lens[i];
232 rte_atomic_thread_fence(__ATOMIC_RELEASE);
234 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
235 desc_base[i].flags = flags;
238 vhost_log_cache_used_vring(dev, vq, last_used_idx *
239 sizeof(struct vring_packed_desc),
240 sizeof(struct vring_packed_desc) *
242 vhost_log_cache_sync(dev, vq);
244 vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
247 static __rte_always_inline void
248 vhost_shadow_dequeue_batch_packed_inorder(struct vhost_virtqueue *vq,
251 vq->shadow_used_packed[0].id = id;
253 if (!vq->shadow_used_idx) {
254 vq->shadow_last_used_idx = vq->last_used_idx;
255 vq->shadow_used_packed[0].flags =
256 PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
257 vq->shadow_used_packed[0].len = 0;
258 vq->shadow_used_packed[0].count = 1;
259 vq->shadow_used_idx++;
262 vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
265 static __rte_always_inline void
266 vhost_shadow_dequeue_batch_packed(struct virtio_net *dev,
267 struct vhost_virtqueue *vq,
274 flags = PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
276 if (!vq->shadow_used_idx) {
277 vq->shadow_last_used_idx = vq->last_used_idx;
278 vq->shadow_used_packed[0].id = ids[0];
279 vq->shadow_used_packed[0].len = 0;
280 vq->shadow_used_packed[0].count = 1;
281 vq->shadow_used_packed[0].flags = flags;
282 vq->shadow_used_idx++;
287 vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE) {
288 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
289 vq->desc_packed[vq->last_used_idx + i].len = 0;
292 rte_atomic_thread_fence(__ATOMIC_RELEASE);
293 vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE)
294 vq->desc_packed[vq->last_used_idx + i].flags = flags;
296 vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
297 sizeof(struct vring_packed_desc),
298 sizeof(struct vring_packed_desc) *
300 vhost_log_cache_sync(dev, vq);
302 vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
305 static __rte_always_inline void
306 vhost_shadow_dequeue_single_packed(struct vhost_virtqueue *vq,
312 flags = vq->desc_packed[vq->last_used_idx].flags;
313 if (vq->used_wrap_counter) {
314 flags |= VRING_DESC_F_USED;
315 flags |= VRING_DESC_F_AVAIL;
317 flags &= ~VRING_DESC_F_USED;
318 flags &= ~VRING_DESC_F_AVAIL;
321 if (!vq->shadow_used_idx) {
322 vq->shadow_last_used_idx = vq->last_used_idx;
324 vq->shadow_used_packed[0].id = buf_id;
325 vq->shadow_used_packed[0].len = 0;
326 vq->shadow_used_packed[0].flags = flags;
327 vq->shadow_used_idx++;
329 vq->desc_packed[vq->last_used_idx].id = buf_id;
330 vq->desc_packed[vq->last_used_idx].len = 0;
331 vq->desc_packed[vq->last_used_idx].flags = flags;
334 vq_inc_last_used_packed(vq, count);
337 static __rte_always_inline void
338 vhost_shadow_dequeue_single_packed_inorder(struct vhost_virtqueue *vq,
344 vq->shadow_used_packed[0].id = buf_id;
346 flags = vq->desc_packed[vq->last_used_idx].flags;
347 if (vq->used_wrap_counter) {
348 flags |= VRING_DESC_F_USED;
349 flags |= VRING_DESC_F_AVAIL;
351 flags &= ~VRING_DESC_F_USED;
352 flags &= ~VRING_DESC_F_AVAIL;
355 if (!vq->shadow_used_idx) {
356 vq->shadow_last_used_idx = vq->last_used_idx;
357 vq->shadow_used_packed[0].len = 0;
358 vq->shadow_used_packed[0].flags = flags;
359 vq->shadow_used_idx++;
362 vq_inc_last_used_packed(vq, count);
365 static __rte_always_inline void
366 vhost_shadow_enqueue_packed(struct vhost_virtqueue *vq,
370 uint16_t num_buffers)
374 for (i = 0; i < num_buffers; i++) {
375 /* enqueue shadow flush action aligned with batch num */
376 if (!vq->shadow_used_idx)
377 vq->shadow_aligned_idx = vq->last_used_idx &
379 vq->shadow_used_packed[vq->shadow_used_idx].id = id[i];
380 vq->shadow_used_packed[vq->shadow_used_idx].len = len[i];
381 vq->shadow_used_packed[vq->shadow_used_idx].count = count[i];
382 vq->shadow_aligned_idx += count[i];
383 vq->shadow_used_idx++;
387 static __rte_always_inline void
388 vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
389 struct vhost_virtqueue *vq,
393 uint16_t num_buffers)
395 vhost_shadow_enqueue_packed(vq, len, id, count, num_buffers);
397 if (vq->shadow_aligned_idx >= PACKED_BATCH_SIZE) {
398 do_data_copy_enqueue(dev, vq);
399 vhost_flush_enqueue_shadow_packed(dev, vq);
403 /* avoid write operation when necessary, to lessen cache issues */
404 #define ASSIGN_UNLESS_EQUAL(var, val) do { \
405 if ((var) != (val)) \
409 static __rte_always_inline void
410 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
412 uint64_t csum_l4 = m_buf->ol_flags & RTE_MBUF_F_TX_L4_MASK;
414 if (m_buf->ol_flags & RTE_MBUF_F_TX_TCP_SEG)
415 csum_l4 |= RTE_MBUF_F_TX_TCP_CKSUM;
418 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
419 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
422 case RTE_MBUF_F_TX_TCP_CKSUM:
423 net_hdr->csum_offset = (offsetof(struct rte_tcp_hdr,
426 case RTE_MBUF_F_TX_UDP_CKSUM:
427 net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
430 case RTE_MBUF_F_TX_SCTP_CKSUM:
431 net_hdr->csum_offset = (offsetof(struct rte_sctp_hdr,
436 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
437 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
438 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
441 /* IP cksum verification cannot be bypassed, then calculate here */
442 if (m_buf->ol_flags & RTE_MBUF_F_TX_IP_CKSUM) {
443 struct rte_ipv4_hdr *ipv4_hdr;
445 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_ipv4_hdr *,
447 ipv4_hdr->hdr_checksum = 0;
448 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
451 if (m_buf->ol_flags & RTE_MBUF_F_TX_TCP_SEG) {
452 if (m_buf->ol_flags & RTE_MBUF_F_TX_IPV4)
453 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
455 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
456 net_hdr->gso_size = m_buf->tso_segsz;
457 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
459 } else if (m_buf->ol_flags & RTE_MBUF_F_TX_UDP_SEG) {
460 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
461 net_hdr->gso_size = m_buf->tso_segsz;
462 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
465 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
466 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
467 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
471 static __rte_always_inline int
472 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
473 struct buf_vector *buf_vec, uint16_t *vec_idx,
474 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
476 uint16_t vec_id = *vec_idx;
480 uint64_t desc_chunck_len = desc_len;
482 if (unlikely(vec_id >= BUF_VECTOR_MAX))
485 desc_addr = vhost_iova_to_vva(dev, vq,
489 if (unlikely(!desc_addr))
492 rte_prefetch0((void *)(uintptr_t)desc_addr);
494 buf_vec[vec_id].buf_iova = desc_iova;
495 buf_vec[vec_id].buf_addr = desc_addr;
496 buf_vec[vec_id].buf_len = desc_chunck_len;
498 desc_len -= desc_chunck_len;
499 desc_iova += desc_chunck_len;
507 static __rte_always_inline int
508 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
509 uint32_t avail_idx, uint16_t *vec_idx,
510 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
511 uint32_t *desc_chain_len, uint8_t perm)
513 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
514 uint16_t vec_id = *vec_idx;
517 uint32_t nr_descs = vq->size;
519 struct vring_desc *descs = vq->desc;
520 struct vring_desc *idesc = NULL;
522 if (unlikely(idx >= vq->size))
525 *desc_chain_head = idx;
527 if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
528 dlen = vq->desc[idx].len;
529 nr_descs = dlen / sizeof(struct vring_desc);
530 if (unlikely(nr_descs > vq->size))
533 descs = (struct vring_desc *)(uintptr_t)
534 vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
537 if (unlikely(!descs))
540 if (unlikely(dlen < vq->desc[idx].len)) {
542 * The indirect desc table is not contiguous
543 * in process VA space, we have to copy it.
545 idesc = vhost_alloc_copy_ind_table(dev, vq,
546 vq->desc[idx].addr, vq->desc[idx].len);
547 if (unlikely(!idesc))
557 if (unlikely(idx >= nr_descs || cnt++ >= nr_descs)) {
558 free_ind_table(idesc);
562 dlen = descs[idx].len;
565 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
566 descs[idx].addr, dlen,
568 free_ind_table(idesc);
572 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
575 idx = descs[idx].next;
578 *desc_chain_len = len;
581 if (unlikely(!!idesc))
582 free_ind_table(idesc);
588 * Returns -1 on fail, 0 on success
591 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
592 uint32_t size, struct buf_vector *buf_vec,
593 uint16_t *num_buffers, uint16_t avail_head,
597 uint16_t vec_idx = 0;
598 uint16_t max_tries, tries = 0;
600 uint16_t head_idx = 0;
604 cur_idx = vq->last_avail_idx;
606 if (rxvq_is_mergeable(dev))
607 max_tries = vq->size - 1;
612 if (unlikely(cur_idx == avail_head))
615 * if we tried all available ring items, and still
616 * can't get enough buf, it means something abnormal
619 if (unlikely(++tries > max_tries))
622 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
625 VHOST_ACCESS_RW) < 0))
627 len = RTE_MIN(len, size);
628 update_shadow_used_ring_split(vq, head_idx, len);
640 static __rte_always_inline int
641 fill_vec_buf_packed_indirect(struct virtio_net *dev,
642 struct vhost_virtqueue *vq,
643 struct vring_packed_desc *desc, uint16_t *vec_idx,
644 struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
648 uint16_t vec_id = *vec_idx;
650 struct vring_packed_desc *descs, *idescs = NULL;
653 descs = (struct vring_packed_desc *)(uintptr_t)
654 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
655 if (unlikely(!descs))
658 if (unlikely(dlen < desc->len)) {
660 * The indirect desc table is not contiguous
661 * in process VA space, we have to copy it.
663 idescs = vhost_alloc_copy_ind_table(dev,
664 vq, desc->addr, desc->len);
665 if (unlikely(!idescs))
671 nr_descs = desc->len / sizeof(struct vring_packed_desc);
672 if (unlikely(nr_descs >= vq->size)) {
673 free_ind_table(idescs);
677 for (i = 0; i < nr_descs; i++) {
678 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
679 free_ind_table(idescs);
685 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
692 if (unlikely(!!idescs))
693 free_ind_table(idescs);
698 static __rte_always_inline int
699 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
700 uint16_t avail_idx, uint16_t *desc_count,
701 struct buf_vector *buf_vec, uint16_t *vec_idx,
702 uint16_t *buf_id, uint32_t *len, uint8_t perm)
704 bool wrap_counter = vq->avail_wrap_counter;
705 struct vring_packed_desc *descs = vq->desc_packed;
706 uint16_t vec_id = *vec_idx;
709 if (avail_idx < vq->last_avail_idx)
713 * Perform a load-acquire barrier in desc_is_avail to
714 * enforce the ordering between desc flags and desc
717 if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
724 if (unlikely(vec_id >= BUF_VECTOR_MAX))
727 if (unlikely(*desc_count >= vq->size))
731 *buf_id = descs[avail_idx].id;
733 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
734 if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
740 dlen = descs[avail_idx].len;
743 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
744 descs[avail_idx].addr,
750 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
753 if (++avail_idx >= vq->size) {
754 avail_idx -= vq->size;
764 static __rte_noinline void
765 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
766 struct buf_vector *buf_vec,
767 struct virtio_net_hdr_mrg_rxbuf *hdr)
770 uint64_t remain = dev->vhost_hlen;
771 uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
772 uint64_t iova = buf_vec->buf_iova;
775 len = RTE_MIN(remain,
777 dst = buf_vec->buf_addr;
778 rte_memcpy((void *)(uintptr_t)dst,
779 (void *)(uintptr_t)src,
782 PRINT_PACKET(dev, (uintptr_t)dst,
784 vhost_log_cache_write_iova(dev, vq,
794 static __rte_always_inline int
795 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
796 struct rte_mbuf *m, struct buf_vector *buf_vec,
797 uint16_t nr_vec, uint16_t num_buffers)
799 uint32_t vec_idx = 0;
800 uint32_t mbuf_offset, mbuf_avail;
801 uint32_t buf_offset, buf_avail;
802 uint64_t buf_addr, buf_iova, buf_len;
805 struct rte_mbuf *hdr_mbuf;
806 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
807 struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
810 if (unlikely(m == NULL)) {
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;
819 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
826 if (unlikely(buf_len < dev->vhost_hlen)) {
827 memset(&tmp_hdr, 0, sizeof(struct virtio_net_hdr_mrg_rxbuf));
830 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
832 VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
833 dev->vid, num_buffers);
835 if (unlikely(buf_len < dev->vhost_hlen)) {
836 buf_offset = dev->vhost_hlen - buf_len;
838 buf_addr = buf_vec[vec_idx].buf_addr;
839 buf_iova = buf_vec[vec_idx].buf_iova;
840 buf_len = buf_vec[vec_idx].buf_len;
841 buf_avail = buf_len - buf_offset;
843 buf_offset = dev->vhost_hlen;
844 buf_avail = buf_len - dev->vhost_hlen;
847 mbuf_avail = rte_pktmbuf_data_len(m);
849 while (mbuf_avail != 0 || m->next != NULL) {
850 /* done with current buf, get the next one */
851 if (buf_avail == 0) {
853 if (unlikely(vec_idx >= nr_vec)) {
858 buf_addr = buf_vec[vec_idx].buf_addr;
859 buf_iova = buf_vec[vec_idx].buf_iova;
860 buf_len = buf_vec[vec_idx].buf_len;
866 /* done with current mbuf, get the next one */
867 if (mbuf_avail == 0) {
871 mbuf_avail = rte_pktmbuf_data_len(m);
875 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
876 if (rxvq_is_mergeable(dev))
877 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
880 if (unlikely(hdr == &tmp_hdr)) {
881 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
883 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
885 vhost_log_cache_write_iova(dev, vq,
893 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
895 if (likely(cpy_len > MAX_BATCH_LEN ||
896 vq->batch_copy_nb_elems >= vq->size)) {
897 rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
898 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
900 vhost_log_cache_write_iova(dev, vq,
901 buf_iova + buf_offset,
903 PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
906 batch_copy[vq->batch_copy_nb_elems].dst =
907 (void *)((uintptr_t)(buf_addr + buf_offset));
908 batch_copy[vq->batch_copy_nb_elems].src =
909 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
910 batch_copy[vq->batch_copy_nb_elems].log_addr =
911 buf_iova + buf_offset;
912 batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
913 vq->batch_copy_nb_elems++;
916 mbuf_avail -= cpy_len;
917 mbuf_offset += cpy_len;
918 buf_avail -= cpy_len;
919 buf_offset += cpy_len;
927 static __rte_always_inline int
928 async_iter_initialize(struct vhost_async *async)
930 struct rte_vhost_iov_iter *iter;
932 if (unlikely(async->iovec_idx >= VHOST_MAX_ASYNC_VEC)) {
933 VHOST_LOG_DATA(ERR, "no more async iovec available\n");
937 iter = async->iov_iter + async->iter_idx;
938 iter->iov = async->iovec + async->iovec_idx;
944 static __rte_always_inline int
945 async_iter_add_iovec(struct vhost_async *async, void *src, void *dst, size_t len)
947 struct rte_vhost_iov_iter *iter;
948 struct rte_vhost_iovec *iovec;
950 if (unlikely(async->iovec_idx >= VHOST_MAX_ASYNC_VEC)) {
951 static bool vhost_max_async_vec_log;
953 if (!vhost_max_async_vec_log) {
954 VHOST_LOG_DATA(ERR, "no more async iovec available\n");
955 vhost_max_async_vec_log = true;
961 iter = async->iov_iter + async->iter_idx;
962 iovec = async->iovec + async->iovec_idx;
964 iovec->src_addr = src;
965 iovec->dst_addr = dst;
974 static __rte_always_inline void
975 async_iter_finalize(struct vhost_async *async)
980 static __rte_always_inline void
981 async_iter_cancel(struct vhost_async *async)
983 struct rte_vhost_iov_iter *iter;
985 iter = async->iov_iter + async->iter_idx;
986 async->iovec_idx -= iter->nr_segs;
991 static __rte_always_inline void
992 async_iter_reset(struct vhost_async *async)
995 async->iovec_idx = 0;
998 static __rte_always_inline void
999 async_fill_descs(struct vhost_async *async, struct rte_vhost_async_desc *descs)
1003 for (i = 0; i < async->iter_idx; i++)
1004 descs[i].iter = async->iov_iter + i;
1007 static __rte_always_inline int
1008 async_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
1009 struct rte_mbuf *m, struct buf_vector *buf_vec,
1010 uint16_t nr_vec, uint16_t num_buffers)
1012 struct vhost_async *async = vq->async;
1013 struct rte_mbuf *hdr_mbuf;
1014 struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
1015 uint64_t buf_addr, buf_iova;
1017 uint64_t mapped_len;
1018 uint32_t vec_idx = 0;
1019 uint32_t mbuf_offset, mbuf_avail;
1020 uint32_t buf_offset, buf_avail;
1021 uint32_t cpy_len, buf_len;
1025 if (unlikely(m == NULL))
1028 buf_addr = buf_vec[vec_idx].buf_addr;
1029 buf_iova = buf_vec[vec_idx].buf_iova;
1030 buf_len = buf_vec[vec_idx].buf_len;
1032 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1))
1036 hdr_addr = buf_addr;
1037 if (unlikely(buf_len < dev->vhost_hlen)) {
1038 memset(&tmp_hdr, 0, sizeof(struct virtio_net_hdr_mrg_rxbuf));
1041 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
1043 VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
1044 dev->vid, num_buffers);
1046 if (unlikely(buf_len < dev->vhost_hlen)) {
1047 buf_offset = dev->vhost_hlen - buf_len;
1049 buf_addr = buf_vec[vec_idx].buf_addr;
1050 buf_iova = buf_vec[vec_idx].buf_iova;
1051 buf_len = buf_vec[vec_idx].buf_len;
1052 buf_avail = buf_len - buf_offset;
1054 buf_offset = dev->vhost_hlen;
1055 buf_avail = buf_len - dev->vhost_hlen;
1058 mbuf_avail = rte_pktmbuf_data_len(m);
1061 if (async_iter_initialize(async))
1064 while (mbuf_avail != 0 || m->next != NULL) {
1065 /* done with current buf, get the next one */
1066 if (buf_avail == 0) {
1068 if (unlikely(vec_idx >= nr_vec))
1071 buf_addr = buf_vec[vec_idx].buf_addr;
1072 buf_iova = buf_vec[vec_idx].buf_iova;
1073 buf_len = buf_vec[vec_idx].buf_len;
1076 buf_avail = buf_len;
1079 /* done with current mbuf, get the next one */
1080 if (mbuf_avail == 0) {
1084 mbuf_avail = rte_pktmbuf_data_len(m);
1088 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
1089 if (rxvq_is_mergeable(dev))
1090 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
1093 if (unlikely(hdr == &tmp_hdr)) {
1094 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
1096 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
1097 dev->vhost_hlen, 0);
1098 vhost_log_cache_write_iova(dev, vq,
1099 buf_vec[0].buf_iova,
1106 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1108 while (unlikely(cpy_len)) {
1109 hpa = (void *)(uintptr_t)gpa_to_first_hpa(dev,
1110 buf_iova + buf_offset,
1111 cpy_len, &mapped_len);
1112 if (unlikely(!hpa)) {
1113 VHOST_LOG_DATA(ERR, "(%d) %s: failed to get hpa.\n",
1114 dev->vid, __func__);
1118 if (unlikely(async_iter_add_iovec(async,
1119 (void *)(uintptr_t)rte_pktmbuf_iova_offset(m,
1121 hpa, (size_t)mapped_len)))
1124 cpy_len -= (uint32_t)mapped_len;
1125 mbuf_avail -= (uint32_t)mapped_len;
1126 mbuf_offset += (uint32_t)mapped_len;
1127 buf_avail -= (uint32_t)mapped_len;
1128 buf_offset += (uint32_t)mapped_len;
1132 async_iter_finalize(async);
1136 async_iter_cancel(async);
1141 static __rte_always_inline int
1142 vhost_enqueue_single_packed(struct virtio_net *dev,
1143 struct vhost_virtqueue *vq,
1144 struct rte_mbuf *pkt,
1145 struct buf_vector *buf_vec,
1148 uint16_t nr_vec = 0;
1149 uint16_t avail_idx = vq->last_avail_idx;
1150 uint16_t max_tries, tries = 0;
1151 uint16_t buf_id = 0;
1153 uint16_t desc_count;
1154 uint32_t size = pkt->pkt_len + sizeof(struct virtio_net_hdr_mrg_rxbuf);
1155 uint16_t num_buffers = 0;
1156 uint32_t buffer_len[vq->size];
1157 uint16_t buffer_buf_id[vq->size];
1158 uint16_t buffer_desc_count[vq->size];
1160 if (rxvq_is_mergeable(dev))
1161 max_tries = vq->size - 1;
1167 * if we tried all available ring items, and still
1168 * can't get enough buf, it means something abnormal
1171 if (unlikely(++tries > max_tries))
1174 if (unlikely(fill_vec_buf_packed(dev, vq,
1175 avail_idx, &desc_count,
1178 VHOST_ACCESS_RW) < 0))
1181 len = RTE_MIN(len, size);
1184 buffer_len[num_buffers] = len;
1185 buffer_buf_id[num_buffers] = buf_id;
1186 buffer_desc_count[num_buffers] = desc_count;
1189 *nr_descs += desc_count;
1190 avail_idx += desc_count;
1191 if (avail_idx >= vq->size)
1192 avail_idx -= vq->size;
1195 if (copy_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, num_buffers) < 0)
1198 vhost_shadow_enqueue_single_packed(dev, vq, buffer_len, buffer_buf_id,
1199 buffer_desc_count, num_buffers);
1204 static __rte_noinline uint32_t
1205 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1206 struct rte_mbuf **pkts, uint32_t count)
1208 uint32_t pkt_idx = 0;
1209 uint16_t num_buffers;
1210 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1211 uint16_t avail_head;
1214 * The ordering between avail index and
1215 * desc reads needs to be enforced.
1217 avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1219 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1221 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1222 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1223 uint16_t nr_vec = 0;
1225 if (unlikely(reserve_avail_buf_split(dev, vq,
1226 pkt_len, buf_vec, &num_buffers,
1227 avail_head, &nr_vec) < 0)) {
1228 VHOST_LOG_DATA(DEBUG,
1229 "(%d) failed to get enough desc from vring\n",
1231 vq->shadow_used_idx -= num_buffers;
1235 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1236 dev->vid, vq->last_avail_idx,
1237 vq->last_avail_idx + num_buffers);
1239 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
1242 vq->shadow_used_idx -= num_buffers;
1246 vq->last_avail_idx += num_buffers;
1249 do_data_copy_enqueue(dev, vq);
1251 if (likely(vq->shadow_used_idx)) {
1252 flush_shadow_used_ring_split(dev, vq);
1253 vhost_vring_call_split(dev, vq);
1259 static __rte_always_inline int
1260 virtio_dev_rx_sync_batch_check(struct virtio_net *dev,
1261 struct vhost_virtqueue *vq,
1262 struct rte_mbuf **pkts,
1263 uint64_t *desc_addrs,
1266 bool wrap_counter = vq->avail_wrap_counter;
1267 struct vring_packed_desc *descs = vq->desc_packed;
1268 uint16_t avail_idx = vq->last_avail_idx;
1269 uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1272 if (unlikely(avail_idx & PACKED_BATCH_MASK))
1275 if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1278 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1279 if (unlikely(pkts[i]->next != NULL))
1281 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))
1310 static __rte_always_inline void
1311 virtio_dev_rx_batch_packed_copy(struct virtio_net *dev,
1312 struct vhost_virtqueue *vq,
1313 struct rte_mbuf **pkts,
1314 uint64_t *desc_addrs,
1317 uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1318 struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
1319 struct vring_packed_desc *descs = vq->desc_packed;
1320 uint16_t avail_idx = vq->last_avail_idx;
1321 uint16_t ids[PACKED_BATCH_SIZE];
1324 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1325 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1326 hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)
1327 (uintptr_t)desc_addrs[i];
1328 lens[i] = pkts[i]->pkt_len +
1329 sizeof(struct virtio_net_hdr_mrg_rxbuf);
1332 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1333 virtio_enqueue_offload(pkts[i], &hdrs[i]->hdr);
1335 vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1337 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1338 rte_memcpy((void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1339 rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1343 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1344 vhost_log_cache_write_iova(dev, vq, descs[avail_idx + i].addr,
1347 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1348 ids[i] = descs[avail_idx + i].id;
1350 vhost_flush_enqueue_batch_packed(dev, vq, lens, ids);
1353 static __rte_always_inline int
1354 virtio_dev_rx_sync_batch_packed(struct virtio_net *dev,
1355 struct vhost_virtqueue *vq,
1356 struct rte_mbuf **pkts)
1358 uint64_t desc_addrs[PACKED_BATCH_SIZE];
1359 uint64_t lens[PACKED_BATCH_SIZE];
1361 if (virtio_dev_rx_sync_batch_check(dev, vq, pkts, desc_addrs, lens) == -1)
1364 if (vq->shadow_used_idx) {
1365 do_data_copy_enqueue(dev, vq);
1366 vhost_flush_enqueue_shadow_packed(dev, vq);
1369 virtio_dev_rx_batch_packed_copy(dev, vq, pkts, desc_addrs, lens);
1374 static __rte_always_inline int16_t
1375 virtio_dev_rx_single_packed(struct virtio_net *dev,
1376 struct vhost_virtqueue *vq,
1377 struct rte_mbuf *pkt)
1379 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1380 uint16_t nr_descs = 0;
1382 if (unlikely(vhost_enqueue_single_packed(dev, vq, pkt, buf_vec,
1384 VHOST_LOG_DATA(DEBUG,
1385 "(%d) failed to get enough desc from vring\n",
1390 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1391 dev->vid, vq->last_avail_idx,
1392 vq->last_avail_idx + nr_descs);
1394 vq_inc_last_avail_packed(vq, nr_descs);
1399 static __rte_noinline uint32_t
1400 virtio_dev_rx_packed(struct virtio_net *dev,
1401 struct vhost_virtqueue *__rte_restrict vq,
1402 struct rte_mbuf **__rte_restrict pkts,
1405 uint32_t pkt_idx = 0;
1408 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1410 if (count - pkt_idx >= PACKED_BATCH_SIZE) {
1411 if (!virtio_dev_rx_sync_batch_packed(dev, vq,
1413 pkt_idx += PACKED_BATCH_SIZE;
1418 if (virtio_dev_rx_single_packed(dev, vq, pkts[pkt_idx]))
1422 } while (pkt_idx < count);
1424 if (vq->shadow_used_idx) {
1425 do_data_copy_enqueue(dev, vq);
1426 vhost_flush_enqueue_shadow_packed(dev, vq);
1430 vhost_vring_call_packed(dev, vq);
1435 static __rte_always_inline uint32_t
1436 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
1437 struct rte_mbuf **pkts, uint32_t count)
1439 struct vhost_virtqueue *vq;
1442 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1443 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1444 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1445 dev->vid, __func__, queue_id);
1449 vq = dev->virtqueue[queue_id];
1451 rte_spinlock_lock(&vq->access_lock);
1453 if (unlikely(!vq->enabled))
1454 goto out_access_unlock;
1456 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1457 vhost_user_iotlb_rd_lock(vq);
1459 if (unlikely(!vq->access_ok))
1460 if (unlikely(vring_translate(dev, vq) < 0))
1463 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1467 if (vq_is_packed(dev))
1468 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
1470 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
1473 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1474 vhost_user_iotlb_rd_unlock(vq);
1477 rte_spinlock_unlock(&vq->access_lock);
1483 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
1484 struct rte_mbuf **__rte_restrict pkts, uint16_t count)
1486 struct virtio_net *dev = get_device(vid);
1491 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1493 "(%d) %s: built-in vhost net backend is disabled.\n",
1494 dev->vid, __func__);
1498 return virtio_dev_rx(dev, queue_id, pkts, count);
1501 static __rte_always_inline uint16_t
1502 virtio_dev_rx_async_get_info_idx(uint16_t pkts_idx,
1503 uint16_t vq_size, uint16_t n_inflight)
1505 return pkts_idx > n_inflight ? (pkts_idx - n_inflight) :
1506 (vq_size - n_inflight + pkts_idx) % vq_size;
1509 static __rte_always_inline void
1510 store_dma_desc_info_split(struct vring_used_elem *s_ring, struct vring_used_elem *d_ring,
1511 uint16_t ring_size, uint16_t s_idx, uint16_t d_idx, uint16_t count)
1513 size_t elem_size = sizeof(struct vring_used_elem);
1515 if (d_idx + count <= ring_size) {
1516 rte_memcpy(d_ring + d_idx, s_ring + s_idx, count * elem_size);
1518 uint16_t size = ring_size - d_idx;
1520 rte_memcpy(d_ring + d_idx, s_ring + s_idx, size * elem_size);
1521 rte_memcpy(d_ring, s_ring + s_idx + size, (count - size) * elem_size);
1525 static __rte_always_inline void
1526 store_dma_desc_info_packed(struct vring_used_elem_packed *s_ring,
1527 struct vring_used_elem_packed *d_ring,
1528 uint16_t ring_size, uint16_t s_idx, uint16_t d_idx, uint16_t count)
1530 size_t elem_size = sizeof(struct vring_used_elem_packed);
1532 if (d_idx + count <= ring_size) {
1533 rte_memcpy(d_ring + d_idx, s_ring + s_idx, count * elem_size);
1535 uint16_t size = ring_size - d_idx;
1537 rte_memcpy(d_ring + d_idx, s_ring + s_idx, size * elem_size);
1538 rte_memcpy(d_ring, s_ring + s_idx + size, (count - size) * elem_size);
1542 static __rte_noinline uint32_t
1543 virtio_dev_rx_async_submit_split(struct virtio_net *dev,
1544 struct vhost_virtqueue *vq, uint16_t queue_id,
1545 struct rte_mbuf **pkts, uint32_t count)
1547 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1548 uint32_t pkt_idx = 0;
1549 uint16_t num_buffers;
1550 uint16_t avail_head;
1552 struct vhost_async *async = vq->async;
1553 struct rte_vhost_async_desc async_descs[MAX_PKT_BURST];
1554 struct async_inflight_info *pkts_info = async->pkts_info;
1555 uint32_t pkt_err = 0;
1557 uint16_t slot_idx = 0;
1560 * The ordering between avail index and desc reads need to be enforced.
1562 avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1564 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1566 async_iter_reset(async);
1568 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1569 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1570 uint16_t nr_vec = 0;
1572 if (unlikely(reserve_avail_buf_split(dev, vq, pkt_len, buf_vec,
1573 &num_buffers, avail_head, &nr_vec) < 0)) {
1574 VHOST_LOG_DATA(DEBUG, "(%d) failed to get enough desc from vring\n",
1576 vq->shadow_used_idx -= num_buffers;
1580 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1581 dev->vid, vq->last_avail_idx, vq->last_avail_idx + num_buffers);
1583 if (async_mbuf_to_desc(dev, vq, pkts[pkt_idx], buf_vec, nr_vec, num_buffers) < 0) {
1584 vq->shadow_used_idx -= num_buffers;
1588 slot_idx = (async->pkts_idx + pkt_idx) & (vq->size - 1);
1589 pkts_info[slot_idx].descs = num_buffers;
1590 pkts_info[slot_idx].mbuf = pkts[pkt_idx];
1592 vq->last_avail_idx += num_buffers;
1595 if (unlikely(pkt_idx == 0))
1598 async_fill_descs(async, async_descs);
1600 n_xfer = async->ops.transfer_data(dev->vid, queue_id, async_descs, 0, pkt_idx);
1601 if (unlikely(n_xfer < 0)) {
1602 VHOST_LOG_DATA(ERR, "(%d) %s: failed to transfer data for queue id %d.\n",
1603 dev->vid, __func__, queue_id);
1607 pkt_err = pkt_idx - n_xfer;
1608 if (unlikely(pkt_err)) {
1609 uint16_t num_descs = 0;
1611 /* update number of completed packets */
1614 /* calculate the sum of descriptors to revert */
1615 while (pkt_err-- > 0) {
1616 num_descs += pkts_info[slot_idx & (vq->size - 1)].descs;
1620 /* recover shadow used ring and available ring */
1621 vq->shadow_used_idx -= num_descs;
1622 vq->last_avail_idx -= num_descs;
1625 /* keep used descriptors */
1626 if (likely(vq->shadow_used_idx)) {
1627 uint16_t to = async->desc_idx_split & (vq->size - 1);
1629 store_dma_desc_info_split(vq->shadow_used_split,
1630 async->descs_split, vq->size, 0, to,
1631 vq->shadow_used_idx);
1633 async->desc_idx_split += vq->shadow_used_idx;
1634 async->pkts_idx += pkt_idx;
1635 async->pkts_inflight_n += pkt_idx;
1636 vq->shadow_used_idx = 0;
1642 static __rte_always_inline void
1643 vhost_update_used_packed(struct vhost_virtqueue *vq,
1644 struct vring_used_elem_packed *shadow_ring,
1648 uint16_t used_idx = vq->last_used_idx;
1649 uint16_t head_idx = vq->last_used_idx;
1650 uint16_t head_flags = 0;
1655 /* Split loop in two to save memory barriers */
1656 for (i = 0; i < count; i++) {
1657 vq->desc_packed[used_idx].id = shadow_ring[i].id;
1658 vq->desc_packed[used_idx].len = shadow_ring[i].len;
1660 used_idx += shadow_ring[i].count;
1661 if (used_idx >= vq->size)
1662 used_idx -= vq->size;
1665 /* The ordering for storing desc flags needs to be enforced. */
1666 rte_atomic_thread_fence(__ATOMIC_RELEASE);
1668 for (i = 0; i < count; i++) {
1671 if (vq->shadow_used_packed[i].len)
1672 flags = VRING_DESC_F_WRITE;
1676 if (vq->used_wrap_counter) {
1677 flags |= VRING_DESC_F_USED;
1678 flags |= VRING_DESC_F_AVAIL;
1680 flags &= ~VRING_DESC_F_USED;
1681 flags &= ~VRING_DESC_F_AVAIL;
1685 vq->desc_packed[vq->last_used_idx].flags = flags;
1687 head_idx = vq->last_used_idx;
1691 vq_inc_last_used_packed(vq, shadow_ring[i].count);
1694 vq->desc_packed[head_idx].flags = head_flags;
1697 static __rte_always_inline int
1698 vhost_enqueue_async_packed(struct virtio_net *dev,
1699 struct vhost_virtqueue *vq,
1700 struct rte_mbuf *pkt,
1701 struct buf_vector *buf_vec,
1703 uint16_t *nr_buffers)
1705 uint16_t nr_vec = 0;
1706 uint16_t avail_idx = vq->last_avail_idx;
1707 uint16_t max_tries, tries = 0;
1708 uint16_t buf_id = 0;
1710 uint16_t desc_count = 0;
1711 uint32_t size = pkt->pkt_len + sizeof(struct virtio_net_hdr_mrg_rxbuf);
1712 uint32_t buffer_len[vq->size];
1713 uint16_t buffer_buf_id[vq->size];
1714 uint16_t buffer_desc_count[vq->size];
1716 if (rxvq_is_mergeable(dev))
1717 max_tries = vq->size - 1;
1723 * if we tried all available ring items, and still
1724 * can't get enough buf, it means something abnormal
1727 if (unlikely(++tries > max_tries))
1730 if (unlikely(fill_vec_buf_packed(dev, vq,
1731 avail_idx, &desc_count,
1734 VHOST_ACCESS_RW) < 0))
1737 len = RTE_MIN(len, size);
1740 buffer_len[*nr_buffers] = len;
1741 buffer_buf_id[*nr_buffers] = buf_id;
1742 buffer_desc_count[*nr_buffers] = desc_count;
1744 *nr_descs += desc_count;
1745 avail_idx += desc_count;
1746 if (avail_idx >= vq->size)
1747 avail_idx -= vq->size;
1750 if (unlikely(async_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec,
1754 vhost_shadow_enqueue_packed(vq, buffer_len, buffer_buf_id, buffer_desc_count, *nr_buffers);
1759 static __rte_always_inline int16_t
1760 virtio_dev_rx_async_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
1761 struct rte_mbuf *pkt, uint16_t *nr_descs, uint16_t *nr_buffers)
1763 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1765 if (unlikely(vhost_enqueue_async_packed(dev, vq, pkt, buf_vec,
1766 nr_descs, nr_buffers) < 0)) {
1767 VHOST_LOG_DATA(DEBUG, "(%d) failed to get enough desc from vring\n", dev->vid);
1771 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1772 dev->vid, vq->last_avail_idx, vq->last_avail_idx + *nr_descs);
1777 static __rte_always_inline void
1778 dma_error_handler_packed(struct vhost_virtqueue *vq, uint16_t slot_idx,
1779 uint32_t nr_err, uint32_t *pkt_idx)
1781 uint16_t descs_err = 0;
1782 uint16_t buffers_err = 0;
1783 struct async_inflight_info *pkts_info = vq->async->pkts_info;
1786 /* calculate the sum of buffers and descs of DMA-error packets. */
1787 while (nr_err-- > 0) {
1788 descs_err += pkts_info[slot_idx % vq->size].descs;
1789 buffers_err += pkts_info[slot_idx % vq->size].nr_buffers;
1793 if (vq->last_avail_idx >= descs_err) {
1794 vq->last_avail_idx -= descs_err;
1796 vq->last_avail_idx = vq->last_avail_idx + vq->size - descs_err;
1797 vq->avail_wrap_counter ^= 1;
1800 vq->shadow_used_idx -= buffers_err;
1803 static __rte_noinline uint32_t
1804 virtio_dev_rx_async_submit_packed(struct virtio_net *dev,
1805 struct vhost_virtqueue *vq, uint16_t queue_id,
1806 struct rte_mbuf **pkts, uint32_t count)
1808 uint32_t pkt_idx = 0;
1809 uint32_t remained = count;
1811 uint16_t num_buffers;
1814 struct vhost_async *async = vq->async;
1815 struct rte_vhost_async_desc async_descs[MAX_PKT_BURST];
1816 struct async_inflight_info *pkts_info = async->pkts_info;
1817 uint32_t pkt_err = 0;
1818 uint16_t slot_idx = 0;
1821 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1825 if (unlikely(virtio_dev_rx_async_packed(dev, vq, pkts[pkt_idx],
1826 &num_descs, &num_buffers) < 0))
1829 slot_idx = (async->pkts_idx + pkt_idx) % vq->size;
1831 pkts_info[slot_idx].descs = num_descs;
1832 pkts_info[slot_idx].nr_buffers = num_buffers;
1833 pkts_info[slot_idx].mbuf = pkts[pkt_idx];
1837 vq_inc_last_avail_packed(vq, num_descs);
1838 } while (pkt_idx < count);
1840 if (unlikely(pkt_idx == 0))
1843 async_fill_descs(async, async_descs);
1845 n_xfer = async->ops.transfer_data(dev->vid, queue_id, async_descs, 0, pkt_idx);
1846 if (unlikely(n_xfer < 0)) {
1847 VHOST_LOG_DATA(ERR, "(%d) %s: failed to transfer data for queue id %d.\n",
1848 dev->vid, __func__, queue_id);
1852 pkt_err = pkt_idx - n_xfer;
1854 async_iter_reset(async);
1856 if (unlikely(pkt_err))
1857 dma_error_handler_packed(vq, slot_idx, pkt_err, &pkt_idx);
1859 if (likely(vq->shadow_used_idx)) {
1860 /* keep used descriptors. */
1861 store_dma_desc_info_packed(vq->shadow_used_packed, async->buffers_packed,
1862 vq->size, 0, async->buffer_idx_packed,
1863 vq->shadow_used_idx);
1865 async->buffer_idx_packed += vq->shadow_used_idx;
1866 if (async->buffer_idx_packed >= vq->size)
1867 async->buffer_idx_packed -= vq->size;
1869 async->pkts_idx += pkt_idx;
1870 if (async->pkts_idx >= vq->size)
1871 async->pkts_idx -= vq->size;
1873 vq->shadow_used_idx = 0;
1874 async->pkts_inflight_n += pkt_idx;
1880 static __rte_always_inline void
1881 write_back_completed_descs_split(struct vhost_virtqueue *vq, uint16_t n_descs)
1883 struct vhost_async *async = vq->async;
1884 uint16_t nr_left = n_descs;
1889 from = async->last_desc_idx_split & (vq->size - 1);
1890 nr_copy = nr_left + from <= vq->size ? nr_left : vq->size - from;
1891 to = vq->last_used_idx & (vq->size - 1);
1893 if (to + nr_copy <= vq->size) {
1894 rte_memcpy(&vq->used->ring[to], &async->descs_split[from],
1895 nr_copy * sizeof(struct vring_used_elem));
1897 uint16_t size = vq->size - to;
1899 rte_memcpy(&vq->used->ring[to], &async->descs_split[from],
1900 size * sizeof(struct vring_used_elem));
1901 rte_memcpy(&vq->used->ring[0], &async->descs_split[from + size],
1902 (nr_copy - size) * sizeof(struct vring_used_elem));
1905 async->last_desc_idx_split += nr_copy;
1906 vq->last_used_idx += nr_copy;
1908 } while (nr_left > 0);
1911 static __rte_always_inline void
1912 write_back_completed_descs_packed(struct vhost_virtqueue *vq,
1915 struct vhost_async *async = vq->async;
1916 uint16_t nr_left = n_buffers;
1920 from = async->last_buffer_idx_packed;
1921 to = (from + nr_left) % vq->size;
1923 vhost_update_used_packed(vq, async->buffers_packed + from, to - from);
1924 async->last_buffer_idx_packed += nr_left;
1927 vhost_update_used_packed(vq, async->buffers_packed + from,
1929 async->last_buffer_idx_packed = 0;
1930 nr_left -= vq->size - from;
1932 } while (nr_left > 0);
1935 static __rte_always_inline uint16_t
1936 vhost_poll_enqueue_completed(struct virtio_net *dev, uint16_t queue_id,
1937 struct rte_mbuf **pkts, uint16_t count)
1939 struct vhost_virtqueue *vq;
1940 struct vhost_async *async;
1941 struct async_inflight_info *pkts_info;
1943 uint16_t n_pkts_cpl = 0, n_pkts_put = 0, n_descs = 0, n_buffers = 0;
1944 uint16_t start_idx, pkts_idx, vq_size;
1947 vq = dev->virtqueue[queue_id];
1949 pkts_idx = async->pkts_idx % vq->size;
1950 pkts_info = async->pkts_info;
1952 start_idx = virtio_dev_rx_async_get_info_idx(pkts_idx,
1953 vq_size, async->pkts_inflight_n);
1955 if (count > async->last_pkts_n) {
1956 n_cpl = async->ops.check_completed_copies(dev->vid,
1957 queue_id, 0, count - async->last_pkts_n);
1958 if (likely(n_cpl >= 0)) {
1962 "(%d) %s: failed to check completed copies for queue id %d.\n",
1963 dev->vid, __func__, queue_id);
1968 n_pkts_cpl += async->last_pkts_n;
1969 n_pkts_put = RTE_MIN(n_pkts_cpl, count);
1970 if (unlikely(n_pkts_put == 0)) {
1971 async->last_pkts_n = n_pkts_cpl;
1975 if (vq_is_packed(dev)) {
1976 for (i = 0; i < n_pkts_put; i++) {
1977 from = (start_idx + i) % vq_size;
1978 n_buffers += pkts_info[from].nr_buffers;
1979 pkts[i] = pkts_info[from].mbuf;
1982 for (i = 0; i < n_pkts_put; i++) {
1983 from = (start_idx + i) & (vq_size - 1);
1984 n_descs += pkts_info[from].descs;
1985 pkts[i] = pkts_info[from].mbuf;
1988 async->last_pkts_n = n_pkts_cpl - n_pkts_put;
1989 async->pkts_inflight_n -= n_pkts_put;
1991 if (likely(vq->enabled && vq->access_ok)) {
1992 if (vq_is_packed(dev)) {
1993 write_back_completed_descs_packed(vq, n_buffers);
1995 vhost_vring_call_packed(dev, vq);
1997 write_back_completed_descs_split(vq, n_descs);
1999 __atomic_add_fetch(&vq->used->idx, n_descs,
2001 vhost_vring_call_split(dev, vq);
2004 if (vq_is_packed(dev)) {
2005 async->last_buffer_idx_packed += n_buffers;
2006 if (async->last_buffer_idx_packed >= vq->size)
2007 async->last_buffer_idx_packed -= vq->size;
2009 async->last_desc_idx_split += n_descs;
2017 rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
2018 struct rte_mbuf **pkts, uint16_t count)
2020 struct virtio_net *dev = get_device(vid);
2021 struct vhost_virtqueue *vq;
2022 uint16_t n_pkts_cpl = 0;
2027 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2028 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
2029 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
2030 dev->vid, __func__, queue_id);
2034 vq = dev->virtqueue[queue_id];
2036 if (unlikely(!vq->async)) {
2037 VHOST_LOG_DATA(ERR, "(%d) %s: async not registered for queue id %d.\n",
2038 dev->vid, __func__, queue_id);
2042 rte_spinlock_lock(&vq->access_lock);
2044 n_pkts_cpl = vhost_poll_enqueue_completed(dev, queue_id, pkts, count);
2046 rte_spinlock_unlock(&vq->access_lock);
2052 rte_vhost_clear_queue_thread_unsafe(int vid, uint16_t queue_id,
2053 struct rte_mbuf **pkts, uint16_t count)
2055 struct virtio_net *dev = get_device(vid);
2056 struct vhost_virtqueue *vq;
2057 uint16_t n_pkts_cpl = 0;
2062 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2063 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
2064 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
2065 dev->vid, __func__, queue_id);
2069 vq = dev->virtqueue[queue_id];
2071 if (unlikely(!vq->async)) {
2072 VHOST_LOG_DATA(ERR, "(%d) %s: async not registered for queue id %d.\n",
2073 dev->vid, __func__, queue_id);
2077 n_pkts_cpl = vhost_poll_enqueue_completed(dev, queue_id, pkts, count);
2082 static __rte_always_inline uint32_t
2083 virtio_dev_rx_async_submit(struct virtio_net *dev, uint16_t queue_id,
2084 struct rte_mbuf **pkts, uint32_t count)
2086 struct vhost_virtqueue *vq;
2089 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2090 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
2091 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
2092 dev->vid, __func__, queue_id);
2096 vq = dev->virtqueue[queue_id];
2098 rte_spinlock_lock(&vq->access_lock);
2100 if (unlikely(!vq->enabled || !vq->async))
2101 goto out_access_unlock;
2103 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2104 vhost_user_iotlb_rd_lock(vq);
2106 if (unlikely(!vq->access_ok))
2107 if (unlikely(vring_translate(dev, vq) < 0))
2110 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
2114 if (vq_is_packed(dev))
2115 nb_tx = virtio_dev_rx_async_submit_packed(dev, vq, queue_id,
2118 nb_tx = virtio_dev_rx_async_submit_split(dev, vq, queue_id,
2122 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2123 vhost_user_iotlb_rd_unlock(vq);
2126 rte_spinlock_unlock(&vq->access_lock);
2132 rte_vhost_submit_enqueue_burst(int vid, uint16_t queue_id,
2133 struct rte_mbuf **pkts, uint16_t count)
2135 struct virtio_net *dev = get_device(vid);
2140 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
2142 "(%d) %s: built-in vhost net backend is disabled.\n",
2143 dev->vid, __func__);
2147 return virtio_dev_rx_async_submit(dev, queue_id, pkts, count);
2151 virtio_net_with_host_offload(struct virtio_net *dev)
2154 ((1ULL << VIRTIO_NET_F_CSUM) |
2155 (1ULL << VIRTIO_NET_F_HOST_ECN) |
2156 (1ULL << VIRTIO_NET_F_HOST_TSO4) |
2157 (1ULL << VIRTIO_NET_F_HOST_TSO6) |
2158 (1ULL << VIRTIO_NET_F_HOST_UFO)))
2165 parse_headers(struct rte_mbuf *m, uint8_t *l4_proto)
2167 struct rte_ipv4_hdr *ipv4_hdr;
2168 struct rte_ipv6_hdr *ipv6_hdr;
2169 struct rte_ether_hdr *eth_hdr;
2171 uint16_t data_len = rte_pktmbuf_data_len(m);
2173 if (data_len < sizeof(struct rte_ether_hdr))
2176 eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
2178 m->l2_len = sizeof(struct rte_ether_hdr);
2179 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
2181 if (ethertype == RTE_ETHER_TYPE_VLAN) {
2182 if (data_len < sizeof(struct rte_ether_hdr) +
2183 sizeof(struct rte_vlan_hdr))
2186 struct rte_vlan_hdr *vlan_hdr =
2187 (struct rte_vlan_hdr *)(eth_hdr + 1);
2189 m->l2_len += sizeof(struct rte_vlan_hdr);
2190 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
2193 switch (ethertype) {
2194 case RTE_ETHER_TYPE_IPV4:
2195 if (data_len < m->l2_len + sizeof(struct rte_ipv4_hdr))
2197 ipv4_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv4_hdr *,
2199 m->l3_len = rte_ipv4_hdr_len(ipv4_hdr);
2200 if (data_len < m->l2_len + m->l3_len)
2202 m->ol_flags |= RTE_MBUF_F_TX_IPV4;
2203 *l4_proto = ipv4_hdr->next_proto_id;
2205 case RTE_ETHER_TYPE_IPV6:
2206 if (data_len < m->l2_len + sizeof(struct rte_ipv6_hdr))
2208 ipv6_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv6_hdr *,
2210 m->l3_len = sizeof(struct rte_ipv6_hdr);
2211 m->ol_flags |= RTE_MBUF_F_TX_IPV6;
2212 *l4_proto = ipv6_hdr->proto;
2215 /* a valid L3 header is needed for further L4 parsing */
2219 /* both CSUM and GSO need a valid L4 header */
2220 switch (*l4_proto) {
2222 if (data_len < m->l2_len + m->l3_len +
2223 sizeof(struct rte_tcp_hdr))
2227 if (data_len < m->l2_len + m->l3_len +
2228 sizeof(struct rte_udp_hdr))
2232 if (data_len < m->l2_len + m->l3_len +
2233 sizeof(struct rte_sctp_hdr))
2249 static __rte_always_inline void
2250 vhost_dequeue_offload_legacy(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
2252 uint8_t l4_proto = 0;
2253 struct rte_tcp_hdr *tcp_hdr = NULL;
2255 uint16_t data_len = rte_pktmbuf_data_len(m);
2257 if (parse_headers(m, &l4_proto) < 0)
2260 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
2261 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
2262 switch (hdr->csum_offset) {
2263 case (offsetof(struct rte_tcp_hdr, cksum)):
2264 if (l4_proto != IPPROTO_TCP)
2266 m->ol_flags |= RTE_MBUF_F_TX_TCP_CKSUM;
2268 case (offsetof(struct rte_udp_hdr, dgram_cksum)):
2269 if (l4_proto != IPPROTO_UDP)
2271 m->ol_flags |= RTE_MBUF_F_TX_UDP_CKSUM;
2273 case (offsetof(struct rte_sctp_hdr, cksum)):
2274 if (l4_proto != IPPROTO_SCTP)
2276 m->ol_flags |= RTE_MBUF_F_TX_SCTP_CKSUM;
2286 if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
2287 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
2288 case VIRTIO_NET_HDR_GSO_TCPV4:
2289 case VIRTIO_NET_HDR_GSO_TCPV6:
2290 if (l4_proto != IPPROTO_TCP)
2292 tcp_hdr = rte_pktmbuf_mtod_offset(m,
2293 struct rte_tcp_hdr *,
2294 m->l2_len + m->l3_len);
2295 tcp_len = (tcp_hdr->data_off & 0xf0) >> 2;
2296 if (data_len < m->l2_len + m->l3_len + tcp_len)
2298 m->ol_flags |= RTE_MBUF_F_TX_TCP_SEG;
2299 m->tso_segsz = hdr->gso_size;
2300 m->l4_len = tcp_len;
2302 case VIRTIO_NET_HDR_GSO_UDP:
2303 if (l4_proto != IPPROTO_UDP)
2305 m->ol_flags |= RTE_MBUF_F_TX_UDP_SEG;
2306 m->tso_segsz = hdr->gso_size;
2307 m->l4_len = sizeof(struct rte_udp_hdr);
2310 VHOST_LOG_DATA(WARNING,
2311 "unsupported gso type %u.\n", hdr->gso_type);
2323 static __rte_always_inline void
2324 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m,
2325 bool legacy_ol_flags)
2327 struct rte_net_hdr_lens hdr_lens;
2328 int l4_supported = 0;
2331 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
2334 if (legacy_ol_flags) {
2335 vhost_dequeue_offload_legacy(hdr, m);
2339 m->ol_flags |= RTE_MBUF_F_RX_IP_CKSUM_UNKNOWN;
2341 ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
2342 m->packet_type = ptype;
2343 if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
2344 (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
2345 (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
2348 /* According to Virtio 1.1 spec, the device only needs to look at
2349 * VIRTIO_NET_HDR_F_NEEDS_CSUM in the packet transmission path.
2350 * This differs from the processing incoming packets path where the
2351 * driver could rely on VIRTIO_NET_HDR_F_DATA_VALID flag set by the
2354 * 5.1.6.2.1 Driver Requirements: Packet Transmission
2355 * The driver MUST NOT set the VIRTIO_NET_HDR_F_DATA_VALID and
2356 * VIRTIO_NET_HDR_F_RSC_INFO bits in flags.
2358 * 5.1.6.2.2 Device Requirements: Packet Transmission
2359 * The device MUST ignore flag bits that it does not recognize.
2361 if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
2364 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
2365 if (hdr->csum_start <= hdrlen && l4_supported != 0) {
2366 m->ol_flags |= RTE_MBUF_F_RX_L4_CKSUM_NONE;
2368 /* Unknown proto or tunnel, do sw cksum. We can assume
2369 * the cksum field is in the first segment since the
2370 * buffers we provided to the host are large enough.
2371 * In case of SCTP, this will be wrong since it's a CRC
2372 * but there's nothing we can do.
2374 uint16_t csum = 0, off;
2376 if (rte_raw_cksum_mbuf(m, hdr->csum_start,
2377 rte_pktmbuf_pkt_len(m) - hdr->csum_start, &csum) < 0)
2379 if (likely(csum != 0xffff))
2381 off = hdr->csum_offset + hdr->csum_start;
2382 if (rte_pktmbuf_data_len(m) >= off + 1)
2383 *rte_pktmbuf_mtod_offset(m, uint16_t *, off) = csum;
2387 if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
2388 if (hdr->gso_size == 0)
2391 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
2392 case VIRTIO_NET_HDR_GSO_TCPV4:
2393 case VIRTIO_NET_HDR_GSO_TCPV6:
2394 if ((ptype & RTE_PTYPE_L4_MASK) != RTE_PTYPE_L4_TCP)
2396 m->ol_flags |= RTE_MBUF_F_RX_LRO | RTE_MBUF_F_RX_L4_CKSUM_NONE;
2397 m->tso_segsz = hdr->gso_size;
2399 case VIRTIO_NET_HDR_GSO_UDP:
2400 if ((ptype & RTE_PTYPE_L4_MASK) != RTE_PTYPE_L4_UDP)
2402 m->ol_flags |= RTE_MBUF_F_RX_LRO | RTE_MBUF_F_RX_L4_CKSUM_NONE;
2403 m->tso_segsz = hdr->gso_size;
2411 static __rte_noinline void
2412 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
2413 struct buf_vector *buf_vec)
2416 uint64_t remain = sizeof(struct virtio_net_hdr);
2418 uint64_t dst = (uint64_t)(uintptr_t)hdr;
2421 len = RTE_MIN(remain, buf_vec->buf_len);
2422 src = buf_vec->buf_addr;
2423 rte_memcpy((void *)(uintptr_t)dst,
2424 (void *)(uintptr_t)src, len);
2432 static __rte_always_inline int
2433 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
2434 struct buf_vector *buf_vec, uint16_t nr_vec,
2435 struct rte_mbuf *m, struct rte_mempool *mbuf_pool,
2436 bool legacy_ol_flags)
2438 uint32_t buf_avail, buf_offset;
2439 uint64_t buf_addr, buf_len;
2440 uint32_t mbuf_avail, mbuf_offset;
2442 struct rte_mbuf *cur = m, *prev = m;
2443 struct virtio_net_hdr tmp_hdr;
2444 struct virtio_net_hdr *hdr = NULL;
2445 /* A counter to avoid desc dead loop chain */
2446 uint16_t vec_idx = 0;
2447 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
2450 buf_addr = buf_vec[vec_idx].buf_addr;
2451 buf_len = buf_vec[vec_idx].buf_len;
2453 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
2458 if (virtio_net_with_host_offload(dev)) {
2459 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
2461 * No luck, the virtio-net header doesn't fit
2462 * in a contiguous virtual area.
2464 copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
2467 hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
2472 * A virtio driver normally uses at least 2 desc buffers
2473 * for Tx: the first for storing the header, and others
2474 * for storing the data.
2476 if (unlikely(buf_len < dev->vhost_hlen)) {
2477 buf_offset = dev->vhost_hlen - buf_len;
2479 buf_addr = buf_vec[vec_idx].buf_addr;
2480 buf_len = buf_vec[vec_idx].buf_len;
2481 buf_avail = buf_len - buf_offset;
2482 } else if (buf_len == dev->vhost_hlen) {
2483 if (unlikely(++vec_idx >= nr_vec))
2485 buf_addr = buf_vec[vec_idx].buf_addr;
2486 buf_len = buf_vec[vec_idx].buf_len;
2489 buf_avail = buf_len;
2491 buf_offset = dev->vhost_hlen;
2492 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
2496 (uintptr_t)(buf_addr + buf_offset),
2497 (uint32_t)buf_avail, 0);
2500 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
2502 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
2504 if (likely(cpy_len > MAX_BATCH_LEN ||
2505 vq->batch_copy_nb_elems >= vq->size ||
2506 (hdr && cur == m))) {
2507 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
2509 (void *)((uintptr_t)(buf_addr +
2510 buf_offset)), cpy_len);
2512 batch_copy[vq->batch_copy_nb_elems].dst =
2513 rte_pktmbuf_mtod_offset(cur, void *,
2515 batch_copy[vq->batch_copy_nb_elems].src =
2516 (void *)((uintptr_t)(buf_addr + buf_offset));
2517 batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
2518 vq->batch_copy_nb_elems++;
2521 mbuf_avail -= cpy_len;
2522 mbuf_offset += cpy_len;
2523 buf_avail -= cpy_len;
2524 buf_offset += cpy_len;
2526 /* This buf reaches to its end, get the next one */
2527 if (buf_avail == 0) {
2528 if (++vec_idx >= nr_vec)
2531 buf_addr = buf_vec[vec_idx].buf_addr;
2532 buf_len = buf_vec[vec_idx].buf_len;
2535 buf_avail = buf_len;
2537 PRINT_PACKET(dev, (uintptr_t)buf_addr,
2538 (uint32_t)buf_avail, 0);
2542 * This mbuf reaches to its end, get a new one
2543 * to hold more data.
2545 if (mbuf_avail == 0) {
2546 cur = rte_pktmbuf_alloc(mbuf_pool);
2547 if (unlikely(cur == NULL)) {
2548 VHOST_LOG_DATA(ERR, "Failed to "
2549 "allocate memory for mbuf.\n");
2555 prev->data_len = mbuf_offset;
2557 m->pkt_len += mbuf_offset;
2561 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
2565 prev->data_len = mbuf_offset;
2566 m->pkt_len += mbuf_offset;
2569 vhost_dequeue_offload(hdr, m, legacy_ol_flags);
2577 virtio_dev_extbuf_free(void *addr __rte_unused, void *opaque)
2583 virtio_dev_extbuf_alloc(struct rte_mbuf *pkt, uint32_t size)
2585 struct rte_mbuf_ext_shared_info *shinfo = NULL;
2586 uint32_t total_len = RTE_PKTMBUF_HEADROOM + size;
2591 total_len += sizeof(*shinfo) + sizeof(uintptr_t);
2592 total_len = RTE_ALIGN_CEIL(total_len, sizeof(uintptr_t));
2594 if (unlikely(total_len > UINT16_MAX))
2597 buf_len = total_len;
2598 buf = rte_malloc(NULL, buf_len, RTE_CACHE_LINE_SIZE);
2599 if (unlikely(buf == NULL))
2602 /* Initialize shinfo */
2603 shinfo = rte_pktmbuf_ext_shinfo_init_helper(buf, &buf_len,
2604 virtio_dev_extbuf_free, buf);
2605 if (unlikely(shinfo == NULL)) {
2607 VHOST_LOG_DATA(ERR, "Failed to init shinfo\n");
2611 iova = rte_malloc_virt2iova(buf);
2612 rte_pktmbuf_attach_extbuf(pkt, buf, iova, buf_len, shinfo);
2613 rte_pktmbuf_reset_headroom(pkt);
2619 * Prepare a host supported pktmbuf.
2621 static __rte_always_inline int
2622 virtio_dev_pktmbuf_prep(struct virtio_net *dev, struct rte_mbuf *pkt,
2625 if (rte_pktmbuf_tailroom(pkt) >= data_len)
2628 /* attach an external buffer if supported */
2629 if (dev->extbuf && !virtio_dev_extbuf_alloc(pkt, data_len))
2632 /* check if chained buffers are allowed */
2633 if (!dev->linearbuf)
2641 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
2642 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count,
2643 bool legacy_ol_flags)
2646 uint16_t free_entries;
2647 uint16_t dropped = 0;
2648 static bool allocerr_warned;
2651 * The ordering between avail index and
2652 * desc reads needs to be enforced.
2654 free_entries = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE) -
2656 if (free_entries == 0)
2659 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
2661 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2663 count = RTE_MIN(count, MAX_PKT_BURST);
2664 count = RTE_MIN(count, free_entries);
2665 VHOST_LOG_DATA(DEBUG, "(%d) about to dequeue %u buffers\n",
2668 if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts, count))
2671 for (i = 0; i < count; i++) {
2672 struct buf_vector buf_vec[BUF_VECTOR_MAX];
2675 uint16_t nr_vec = 0;
2678 if (unlikely(fill_vec_buf_split(dev, vq,
2679 vq->last_avail_idx + i,
2681 &head_idx, &buf_len,
2682 VHOST_ACCESS_RO) < 0))
2685 update_shadow_used_ring_split(vq, head_idx, 0);
2687 err = virtio_dev_pktmbuf_prep(dev, pkts[i], buf_len);
2688 if (unlikely(err)) {
2690 * mbuf allocation fails for jumbo packets when external
2691 * buffer allocation is not allowed and linear buffer
2692 * is required. Drop this packet.
2694 if (!allocerr_warned) {
2696 "Failed mbuf alloc of size %d from %s on %s.\n",
2697 buf_len, mbuf_pool->name, dev->ifname);
2698 allocerr_warned = true;
2705 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
2706 mbuf_pool, legacy_ol_flags);
2707 if (unlikely(err)) {
2708 if (!allocerr_warned) {
2710 "Failed to copy desc to mbuf on %s.\n",
2712 allocerr_warned = true;
2721 rte_pktmbuf_free_bulk(&pkts[i - 1], count - i + 1);
2723 vq->last_avail_idx += i;
2725 do_data_copy_dequeue(vq);
2726 if (unlikely(i < count))
2727 vq->shadow_used_idx = i;
2728 if (likely(vq->shadow_used_idx)) {
2729 flush_shadow_used_ring_split(dev, vq);
2730 vhost_vring_call_split(dev, vq);
2733 return (i - dropped);
2738 virtio_dev_tx_split_legacy(struct virtio_net *dev,
2739 struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
2740 struct rte_mbuf **pkts, uint16_t count)
2742 return virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count, true);
2747 virtio_dev_tx_split_compliant(struct virtio_net *dev,
2748 struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
2749 struct rte_mbuf **pkts, uint16_t count)
2751 return virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count, false);
2754 static __rte_always_inline int
2755 vhost_reserve_avail_batch_packed(struct virtio_net *dev,
2756 struct vhost_virtqueue *vq,
2757 struct rte_mbuf **pkts,
2759 uintptr_t *desc_addrs,
2762 bool wrap = vq->avail_wrap_counter;
2763 struct vring_packed_desc *descs = vq->desc_packed;
2764 uint64_t lens[PACKED_BATCH_SIZE];
2765 uint64_t buf_lens[PACKED_BATCH_SIZE];
2766 uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2769 if (unlikely(avail_idx & PACKED_BATCH_MASK))
2771 if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
2774 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2775 flags = descs[avail_idx + i].flags;
2776 if (unlikely((wrap != !!(flags & VRING_DESC_F_AVAIL)) ||
2777 (wrap == !!(flags & VRING_DESC_F_USED)) ||
2778 (flags & PACKED_DESC_SINGLE_DEQUEUE_FLAG)))
2782 rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
2784 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2785 lens[i] = descs[avail_idx + i].len;
2787 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2788 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
2789 descs[avail_idx + i].addr,
2790 &lens[i], VHOST_ACCESS_RW);
2793 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2794 if (unlikely(!desc_addrs[i]))
2796 if (unlikely((lens[i] != descs[avail_idx + i].len)))
2800 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2801 if (virtio_dev_pktmbuf_prep(dev, pkts[i], lens[i]))
2805 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2806 buf_lens[i] = pkts[i]->buf_len - pkts[i]->data_off;
2808 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2809 if (unlikely(buf_lens[i] < (lens[i] - buf_offset)))
2813 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2814 pkts[i]->pkt_len = lens[i] - buf_offset;
2815 pkts[i]->data_len = pkts[i]->pkt_len;
2816 ids[i] = descs[avail_idx + i].id;
2825 static __rte_always_inline int
2826 virtio_dev_tx_batch_packed(struct virtio_net *dev,
2827 struct vhost_virtqueue *vq,
2828 struct rte_mbuf **pkts,
2829 bool legacy_ol_flags)
2831 uint16_t avail_idx = vq->last_avail_idx;
2832 uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2833 struct virtio_net_hdr *hdr;
2834 uintptr_t desc_addrs[PACKED_BATCH_SIZE];
2835 uint16_t ids[PACKED_BATCH_SIZE];
2838 if (vhost_reserve_avail_batch_packed(dev, vq, pkts, avail_idx,
2842 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2843 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
2845 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2846 rte_memcpy(rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
2847 (void *)(uintptr_t)(desc_addrs[i] + buf_offset),
2850 if (virtio_net_with_host_offload(dev)) {
2851 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2852 hdr = (struct virtio_net_hdr *)(desc_addrs[i]);
2853 vhost_dequeue_offload(hdr, pkts[i], legacy_ol_flags);
2857 if (virtio_net_is_inorder(dev))
2858 vhost_shadow_dequeue_batch_packed_inorder(vq,
2859 ids[PACKED_BATCH_SIZE - 1]);
2861 vhost_shadow_dequeue_batch_packed(dev, vq, ids);
2863 vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
2868 static __rte_always_inline int
2869 vhost_dequeue_single_packed(struct virtio_net *dev,
2870 struct vhost_virtqueue *vq,
2871 struct rte_mempool *mbuf_pool,
2872 struct rte_mbuf *pkts,
2874 uint16_t *desc_count,
2875 bool legacy_ol_flags)
2877 struct buf_vector buf_vec[BUF_VECTOR_MAX];
2879 uint16_t nr_vec = 0;
2881 static bool allocerr_warned;
2883 if (unlikely(fill_vec_buf_packed(dev, vq,
2884 vq->last_avail_idx, desc_count,
2887 VHOST_ACCESS_RO) < 0))
2890 if (unlikely(virtio_dev_pktmbuf_prep(dev, pkts, buf_len))) {
2891 if (!allocerr_warned) {
2893 "Failed mbuf alloc of size %d from %s on %s.\n",
2894 buf_len, mbuf_pool->name, dev->ifname);
2895 allocerr_warned = true;
2900 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts,
2901 mbuf_pool, legacy_ol_flags);
2902 if (unlikely(err)) {
2903 if (!allocerr_warned) {
2905 "Failed to copy desc to mbuf on %s.\n",
2907 allocerr_warned = true;
2915 static __rte_always_inline int
2916 virtio_dev_tx_single_packed(struct virtio_net *dev,
2917 struct vhost_virtqueue *vq,
2918 struct rte_mempool *mbuf_pool,
2919 struct rte_mbuf *pkts,
2920 bool legacy_ol_flags)
2923 uint16_t buf_id, desc_count = 0;
2926 ret = vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
2927 &desc_count, legacy_ol_flags);
2929 if (likely(desc_count > 0)) {
2930 if (virtio_net_is_inorder(dev))
2931 vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
2934 vhost_shadow_dequeue_single_packed(vq, buf_id,
2937 vq_inc_last_avail_packed(vq, desc_count);
2945 virtio_dev_tx_packed(struct virtio_net *dev,
2946 struct vhost_virtqueue *__rte_restrict vq,
2947 struct rte_mempool *mbuf_pool,
2948 struct rte_mbuf **__rte_restrict pkts,
2950 bool legacy_ol_flags)
2952 uint32_t pkt_idx = 0;
2954 if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts, count))
2958 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
2960 if (count - pkt_idx >= PACKED_BATCH_SIZE) {
2961 if (!virtio_dev_tx_batch_packed(dev, vq,
2964 pkt_idx += PACKED_BATCH_SIZE;
2969 if (virtio_dev_tx_single_packed(dev, vq, mbuf_pool,
2974 } while (pkt_idx < count);
2976 if (pkt_idx != count)
2977 rte_pktmbuf_free_bulk(&pkts[pkt_idx], count - pkt_idx);
2979 if (vq->shadow_used_idx) {
2980 do_data_copy_dequeue(vq);
2982 vhost_flush_dequeue_shadow_packed(dev, vq);
2983 vhost_vring_call_packed(dev, vq);
2991 virtio_dev_tx_packed_legacy(struct virtio_net *dev,
2992 struct vhost_virtqueue *__rte_restrict vq, struct rte_mempool *mbuf_pool,
2993 struct rte_mbuf **__rte_restrict pkts, uint32_t count)
2995 return virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count, true);
3000 virtio_dev_tx_packed_compliant(struct virtio_net *dev,
3001 struct vhost_virtqueue *__rte_restrict vq, struct rte_mempool *mbuf_pool,
3002 struct rte_mbuf **__rte_restrict pkts, uint32_t count)
3004 return virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count, false);
3008 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
3009 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
3011 struct virtio_net *dev;
3012 struct rte_mbuf *rarp_mbuf = NULL;
3013 struct vhost_virtqueue *vq;
3014 int16_t success = 1;
3016 dev = get_device(vid);
3020 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
3022 "(%d) %s: built-in vhost net backend is disabled.\n",
3023 dev->vid, __func__);
3027 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
3029 "(%d) %s: invalid virtqueue idx %d.\n",
3030 dev->vid, __func__, queue_id);
3034 vq = dev->virtqueue[queue_id];
3036 if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
3039 if (unlikely(!vq->enabled)) {
3041 goto out_access_unlock;
3044 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
3045 vhost_user_iotlb_rd_lock(vq);
3047 if (unlikely(!vq->access_ok))
3048 if (unlikely(vring_translate(dev, vq) < 0)) {
3054 * Construct a RARP broadcast packet, and inject it to the "pkts"
3055 * array, to looks like that guest actually send such packet.
3057 * Check user_send_rarp() for more information.
3059 * broadcast_rarp shares a cacheline in the virtio_net structure
3060 * with some fields that are accessed during enqueue and
3061 * __atomic_compare_exchange_n causes a write if performed compare
3062 * and exchange. This could result in false sharing between enqueue
3065 * Prevent unnecessary false sharing by reading broadcast_rarp first
3066 * and only performing compare and exchange if the read indicates it
3067 * is likely to be set.
3069 if (unlikely(__atomic_load_n(&dev->broadcast_rarp, __ATOMIC_ACQUIRE) &&
3070 __atomic_compare_exchange_n(&dev->broadcast_rarp,
3071 &success, 0, 0, __ATOMIC_RELEASE, __ATOMIC_RELAXED))) {
3073 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
3074 if (rarp_mbuf == NULL) {
3075 VHOST_LOG_DATA(ERR, "Failed to make RARP packet.\n");
3080 * Inject it to the head of "pkts" array, so that switch's mac
3081 * learning table will get updated first.
3083 pkts[0] = rarp_mbuf;
3088 if (vq_is_packed(dev)) {
3089 if (dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS)
3090 count = virtio_dev_tx_packed_legacy(dev, vq, mbuf_pool, pkts, count);
3092 count = virtio_dev_tx_packed_compliant(dev, vq, mbuf_pool, pkts, count);
3094 if (dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS)
3095 count = virtio_dev_tx_split_legacy(dev, vq, mbuf_pool, pkts, count);
3097 count = virtio_dev_tx_split_compliant(dev, vq, mbuf_pool, pkts, count);
3101 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
3102 vhost_user_iotlb_rd_unlock(vq);
3105 rte_spinlock_unlock(&vq->access_lock);
3107 if (unlikely(rarp_mbuf != NULL))