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 update_shadow_used_ring_split(struct vhost_virtqueue *vq,
122 uint16_t desc_idx, uint32_t len)
124 uint16_t i = vq->shadow_used_idx++;
126 vq->shadow_used_split[i].id = desc_idx;
127 vq->shadow_used_split[i].len = len;
130 static __rte_always_inline void
131 vhost_flush_enqueue_shadow_packed(struct virtio_net *dev,
132 struct vhost_virtqueue *vq)
135 uint16_t used_idx = vq->last_used_idx;
136 uint16_t head_idx = vq->last_used_idx;
137 uint16_t head_flags = 0;
139 /* Split loop in two to save memory barriers */
140 for (i = 0; i < vq->shadow_used_idx; i++) {
141 vq->desc_packed[used_idx].id = vq->shadow_used_packed[i].id;
142 vq->desc_packed[used_idx].len = vq->shadow_used_packed[i].len;
144 used_idx += vq->shadow_used_packed[i].count;
145 if (used_idx >= vq->size)
146 used_idx -= vq->size;
149 /* The ordering for storing desc flags needs to be enforced. */
150 rte_atomic_thread_fence(__ATOMIC_RELEASE);
152 for (i = 0; i < vq->shadow_used_idx; i++) {
155 if (vq->shadow_used_packed[i].len)
156 flags = VRING_DESC_F_WRITE;
160 if (vq->used_wrap_counter) {
161 flags |= VRING_DESC_F_USED;
162 flags |= VRING_DESC_F_AVAIL;
164 flags &= ~VRING_DESC_F_USED;
165 flags &= ~VRING_DESC_F_AVAIL;
169 vq->desc_packed[vq->last_used_idx].flags = flags;
171 vhost_log_cache_used_vring(dev, vq,
173 sizeof(struct vring_packed_desc),
174 sizeof(struct vring_packed_desc));
176 head_idx = vq->last_used_idx;
180 vq_inc_last_used_packed(vq, vq->shadow_used_packed[i].count);
183 vq->desc_packed[head_idx].flags = head_flags;
185 vhost_log_cache_used_vring(dev, vq,
187 sizeof(struct vring_packed_desc),
188 sizeof(struct vring_packed_desc));
190 vq->shadow_used_idx = 0;
191 vhost_log_cache_sync(dev, vq);
194 static __rte_always_inline void
195 vhost_flush_dequeue_shadow_packed(struct virtio_net *dev,
196 struct vhost_virtqueue *vq)
198 struct vring_used_elem_packed *used_elem = &vq->shadow_used_packed[0];
200 vq->desc_packed[vq->shadow_last_used_idx].id = used_elem->id;
201 /* desc flags is the synchronization point for virtio packed vring */
202 __atomic_store_n(&vq->desc_packed[vq->shadow_last_used_idx].flags,
203 used_elem->flags, __ATOMIC_RELEASE);
205 vhost_log_cache_used_vring(dev, vq, vq->shadow_last_used_idx *
206 sizeof(struct vring_packed_desc),
207 sizeof(struct vring_packed_desc));
208 vq->shadow_used_idx = 0;
209 vhost_log_cache_sync(dev, vq);
212 static __rte_always_inline void
213 vhost_flush_enqueue_batch_packed(struct virtio_net *dev,
214 struct vhost_virtqueue *vq,
221 if (vq->shadow_used_idx) {
222 do_data_copy_enqueue(dev, vq);
223 vhost_flush_enqueue_shadow_packed(dev, vq);
226 flags = PACKED_DESC_ENQUEUE_USED_FLAG(vq->used_wrap_counter);
228 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
229 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
230 vq->desc_packed[vq->last_used_idx + i].len = lens[i];
233 rte_atomic_thread_fence(__ATOMIC_RELEASE);
235 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
236 vq->desc_packed[vq->last_used_idx + i].flags = flags;
238 vhost_log_cache_used_vring(dev, vq, 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_single_packed(struct virtio_net *dev,
367 struct vhost_virtqueue *vq,
371 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++;
386 if (vq->shadow_aligned_idx >= PACKED_BATCH_SIZE) {
387 do_data_copy_enqueue(dev, vq);
388 vhost_flush_enqueue_shadow_packed(dev, vq);
392 /* avoid write operation when necessary, to lessen cache issues */
393 #define ASSIGN_UNLESS_EQUAL(var, val) do { \
394 if ((var) != (val)) \
398 static __rte_always_inline void
399 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
401 uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
403 if (m_buf->ol_flags & PKT_TX_TCP_SEG)
404 csum_l4 |= PKT_TX_TCP_CKSUM;
407 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
408 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
411 case PKT_TX_TCP_CKSUM:
412 net_hdr->csum_offset = (offsetof(struct rte_tcp_hdr,
415 case PKT_TX_UDP_CKSUM:
416 net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
419 case PKT_TX_SCTP_CKSUM:
420 net_hdr->csum_offset = (offsetof(struct rte_sctp_hdr,
425 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
426 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
427 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
430 /* IP cksum verification cannot be bypassed, then calculate here */
431 if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
432 struct rte_ipv4_hdr *ipv4_hdr;
434 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_ipv4_hdr *,
436 ipv4_hdr->hdr_checksum = 0;
437 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
440 if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
441 if (m_buf->ol_flags & PKT_TX_IPV4)
442 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
444 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
445 net_hdr->gso_size = m_buf->tso_segsz;
446 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
448 } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
449 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
450 net_hdr->gso_size = m_buf->tso_segsz;
451 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
454 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
455 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
456 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
460 static __rte_always_inline int
461 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
462 struct buf_vector *buf_vec, uint16_t *vec_idx,
463 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
465 uint16_t vec_id = *vec_idx;
469 uint64_t desc_chunck_len = desc_len;
471 if (unlikely(vec_id >= BUF_VECTOR_MAX))
474 desc_addr = vhost_iova_to_vva(dev, vq,
478 if (unlikely(!desc_addr))
481 rte_prefetch0((void *)(uintptr_t)desc_addr);
483 buf_vec[vec_id].buf_iova = desc_iova;
484 buf_vec[vec_id].buf_addr = desc_addr;
485 buf_vec[vec_id].buf_len = desc_chunck_len;
487 desc_len -= desc_chunck_len;
488 desc_iova += desc_chunck_len;
496 static __rte_always_inline int
497 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
498 uint32_t avail_idx, uint16_t *vec_idx,
499 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
500 uint32_t *desc_chain_len, uint8_t perm)
502 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
503 uint16_t vec_id = *vec_idx;
506 uint32_t nr_descs = vq->size;
508 struct vring_desc *descs = vq->desc;
509 struct vring_desc *idesc = NULL;
511 if (unlikely(idx >= vq->size))
514 *desc_chain_head = idx;
516 if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
517 dlen = vq->desc[idx].len;
518 nr_descs = dlen / sizeof(struct vring_desc);
519 if (unlikely(nr_descs > vq->size))
522 descs = (struct vring_desc *)(uintptr_t)
523 vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
526 if (unlikely(!descs))
529 if (unlikely(dlen < vq->desc[idx].len)) {
531 * The indirect desc table is not contiguous
532 * in process VA space, we have to copy it.
534 idesc = vhost_alloc_copy_ind_table(dev, vq,
535 vq->desc[idx].addr, vq->desc[idx].len);
536 if (unlikely(!idesc))
546 if (unlikely(idx >= nr_descs || cnt++ >= nr_descs)) {
547 free_ind_table(idesc);
551 len += descs[idx].len;
553 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
554 descs[idx].addr, descs[idx].len,
556 free_ind_table(idesc);
560 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
563 idx = descs[idx].next;
566 *desc_chain_len = len;
569 if (unlikely(!!idesc))
570 free_ind_table(idesc);
576 * Returns -1 on fail, 0 on success
579 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
580 uint32_t size, struct buf_vector *buf_vec,
581 uint16_t *num_buffers, uint16_t avail_head,
585 uint16_t vec_idx = 0;
586 uint16_t max_tries, tries = 0;
588 uint16_t head_idx = 0;
592 cur_idx = vq->last_avail_idx;
594 if (rxvq_is_mergeable(dev))
595 max_tries = vq->size - 1;
600 if (unlikely(cur_idx == avail_head))
603 * if we tried all available ring items, and still
604 * can't get enough buf, it means something abnormal
607 if (unlikely(++tries > max_tries))
610 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
613 VHOST_ACCESS_RW) < 0))
615 len = RTE_MIN(len, size);
616 update_shadow_used_ring_split(vq, head_idx, len);
628 static __rte_always_inline int
629 fill_vec_buf_packed_indirect(struct virtio_net *dev,
630 struct vhost_virtqueue *vq,
631 struct vring_packed_desc *desc, uint16_t *vec_idx,
632 struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
636 uint16_t vec_id = *vec_idx;
638 struct vring_packed_desc *descs, *idescs = NULL;
641 descs = (struct vring_packed_desc *)(uintptr_t)
642 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
643 if (unlikely(!descs))
646 if (unlikely(dlen < desc->len)) {
648 * The indirect desc table is not contiguous
649 * in process VA space, we have to copy it.
651 idescs = vhost_alloc_copy_ind_table(dev,
652 vq, desc->addr, desc->len);
653 if (unlikely(!idescs))
659 nr_descs = desc->len / sizeof(struct vring_packed_desc);
660 if (unlikely(nr_descs >= vq->size)) {
661 free_ind_table(idescs);
665 for (i = 0; i < nr_descs; i++) {
666 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
667 free_ind_table(idescs);
671 *len += descs[i].len;
672 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
673 descs[i].addr, descs[i].len,
679 if (unlikely(!!idescs))
680 free_ind_table(idescs);
685 static __rte_always_inline int
686 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
687 uint16_t avail_idx, uint16_t *desc_count,
688 struct buf_vector *buf_vec, uint16_t *vec_idx,
689 uint16_t *buf_id, uint32_t *len, uint8_t perm)
691 bool wrap_counter = vq->avail_wrap_counter;
692 struct vring_packed_desc *descs = vq->desc_packed;
693 uint16_t vec_id = *vec_idx;
695 if (avail_idx < vq->last_avail_idx)
699 * Perform a load-acquire barrier in desc_is_avail to
700 * enforce the ordering between desc flags and desc
703 if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
710 if (unlikely(vec_id >= BUF_VECTOR_MAX))
713 if (unlikely(*desc_count >= vq->size))
717 *buf_id = descs[avail_idx].id;
719 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
720 if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
726 *len += descs[avail_idx].len;
728 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
729 descs[avail_idx].addr,
730 descs[avail_idx].len,
735 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
738 if (++avail_idx >= vq->size) {
739 avail_idx -= vq->size;
749 static __rte_noinline void
750 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
751 struct buf_vector *buf_vec,
752 struct virtio_net_hdr_mrg_rxbuf *hdr)
755 uint64_t remain = dev->vhost_hlen;
756 uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
757 uint64_t iova = buf_vec->buf_iova;
760 len = RTE_MIN(remain,
762 dst = buf_vec->buf_addr;
763 rte_memcpy((void *)(uintptr_t)dst,
764 (void *)(uintptr_t)src,
767 PRINT_PACKET(dev, (uintptr_t)dst,
769 vhost_log_cache_write_iova(dev, vq,
779 static __rte_always_inline int
780 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
781 struct rte_mbuf *m, struct buf_vector *buf_vec,
782 uint16_t nr_vec, uint16_t num_buffers)
784 uint32_t vec_idx = 0;
785 uint32_t mbuf_offset, mbuf_avail;
786 uint32_t buf_offset, buf_avail;
787 uint64_t buf_addr, buf_iova, buf_len;
790 struct rte_mbuf *hdr_mbuf;
791 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
792 struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
795 if (unlikely(m == NULL)) {
800 buf_addr = buf_vec[vec_idx].buf_addr;
801 buf_iova = buf_vec[vec_idx].buf_iova;
802 buf_len = buf_vec[vec_idx].buf_len;
804 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
811 if (unlikely(buf_len < dev->vhost_hlen))
814 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
816 VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
817 dev->vid, num_buffers);
819 if (unlikely(buf_len < dev->vhost_hlen)) {
820 buf_offset = dev->vhost_hlen - buf_len;
822 buf_addr = buf_vec[vec_idx].buf_addr;
823 buf_iova = buf_vec[vec_idx].buf_iova;
824 buf_len = buf_vec[vec_idx].buf_len;
825 buf_avail = buf_len - buf_offset;
827 buf_offset = dev->vhost_hlen;
828 buf_avail = buf_len - dev->vhost_hlen;
831 mbuf_avail = rte_pktmbuf_data_len(m);
833 while (mbuf_avail != 0 || m->next != NULL) {
834 /* done with current buf, get the next one */
835 if (buf_avail == 0) {
837 if (unlikely(vec_idx >= nr_vec)) {
842 buf_addr = buf_vec[vec_idx].buf_addr;
843 buf_iova = buf_vec[vec_idx].buf_iova;
844 buf_len = buf_vec[vec_idx].buf_len;
850 /* done with current mbuf, get the next one */
851 if (mbuf_avail == 0) {
855 mbuf_avail = rte_pktmbuf_data_len(m);
859 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
860 if (rxvq_is_mergeable(dev))
861 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
864 if (unlikely(hdr == &tmp_hdr)) {
865 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
867 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
869 vhost_log_cache_write_iova(dev, vq,
877 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
879 if (likely(cpy_len > MAX_BATCH_LEN ||
880 vq->batch_copy_nb_elems >= vq->size)) {
881 rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
882 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
884 vhost_log_cache_write_iova(dev, vq,
885 buf_iova + buf_offset,
887 PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
890 batch_copy[vq->batch_copy_nb_elems].dst =
891 (void *)((uintptr_t)(buf_addr + buf_offset));
892 batch_copy[vq->batch_copy_nb_elems].src =
893 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
894 batch_copy[vq->batch_copy_nb_elems].log_addr =
895 buf_iova + buf_offset;
896 batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
897 vq->batch_copy_nb_elems++;
900 mbuf_avail -= cpy_len;
901 mbuf_offset += cpy_len;
902 buf_avail -= cpy_len;
903 buf_offset += cpy_len;
911 static __rte_always_inline void
912 async_fill_vec(struct iovec *v, void *base, size_t len)
918 static __rte_always_inline void
919 async_fill_iter(struct rte_vhost_iov_iter *it, size_t count,
920 struct iovec *vec, unsigned long nr_seg)
927 it->nr_segs = nr_seg;
934 static __rte_always_inline void
935 async_fill_desc(struct rte_vhost_async_desc *desc,
936 struct rte_vhost_iov_iter *src, struct rte_vhost_iov_iter *dst)
942 static __rte_always_inline int
943 async_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
944 struct rte_mbuf *m, struct buf_vector *buf_vec,
945 uint16_t nr_vec, uint16_t num_buffers,
946 struct iovec *src_iovec, struct iovec *dst_iovec,
947 struct rte_vhost_iov_iter *src_it,
948 struct rte_vhost_iov_iter *dst_it)
950 uint32_t vec_idx = 0;
951 uint32_t mbuf_offset, mbuf_avail;
952 uint32_t buf_offset, buf_avail;
953 uint64_t buf_addr, buf_iova, buf_len;
954 uint32_t cpy_len, cpy_threshold;
956 struct rte_mbuf *hdr_mbuf;
957 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
958 struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
966 if (unlikely(m == NULL)) {
971 cpy_threshold = vq->async_threshold;
973 buf_addr = buf_vec[vec_idx].buf_addr;
974 buf_iova = buf_vec[vec_idx].buf_iova;
975 buf_len = buf_vec[vec_idx].buf_len;
977 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
984 if (unlikely(buf_len < dev->vhost_hlen))
987 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
989 VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
990 dev->vid, num_buffers);
992 if (unlikely(buf_len < dev->vhost_hlen)) {
993 buf_offset = dev->vhost_hlen - buf_len;
995 buf_addr = buf_vec[vec_idx].buf_addr;
996 buf_iova = buf_vec[vec_idx].buf_iova;
997 buf_len = buf_vec[vec_idx].buf_len;
998 buf_avail = buf_len - buf_offset;
1000 buf_offset = dev->vhost_hlen;
1001 buf_avail = buf_len - dev->vhost_hlen;
1004 mbuf_avail = rte_pktmbuf_data_len(m);
1007 while (mbuf_avail != 0 || m->next != NULL) {
1008 /* done with current buf, get the next one */
1009 if (buf_avail == 0) {
1011 if (unlikely(vec_idx >= nr_vec)) {
1016 buf_addr = buf_vec[vec_idx].buf_addr;
1017 buf_iova = buf_vec[vec_idx].buf_iova;
1018 buf_len = buf_vec[vec_idx].buf_len;
1021 buf_avail = buf_len;
1024 /* done with current mbuf, get the next one */
1025 if (mbuf_avail == 0) {
1029 mbuf_avail = rte_pktmbuf_data_len(m);
1033 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
1034 if (rxvq_is_mergeable(dev))
1035 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
1038 if (unlikely(hdr == &tmp_hdr)) {
1039 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
1041 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
1042 dev->vhost_hlen, 0);
1043 vhost_log_cache_write_iova(dev, vq,
1044 buf_vec[0].buf_iova,
1051 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1053 while (unlikely(cpy_len && cpy_len >= cpy_threshold)) {
1054 hpa = (void *)(uintptr_t)gpa_to_first_hpa(dev,
1055 buf_iova + buf_offset,
1056 cpy_len, &mapped_len);
1058 if (unlikely(!hpa || mapped_len < cpy_threshold))
1061 async_fill_vec(src_iovec + tvec_idx,
1062 (void *)(uintptr_t)rte_pktmbuf_iova_offset(m,
1063 mbuf_offset), (size_t)mapped_len);
1065 async_fill_vec(dst_iovec + tvec_idx,
1066 hpa, (size_t)mapped_len);
1068 tlen += (uint32_t)mapped_len;
1069 cpy_len -= (uint32_t)mapped_len;
1070 mbuf_avail -= (uint32_t)mapped_len;
1071 mbuf_offset += (uint32_t)mapped_len;
1072 buf_avail -= (uint32_t)mapped_len;
1073 buf_offset += (uint32_t)mapped_len;
1077 if (likely(cpy_len)) {
1078 if (unlikely(vq->batch_copy_nb_elems >= vq->size)) {
1080 (void *)((uintptr_t)(buf_addr + buf_offset)),
1081 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
1085 (uintptr_t)(buf_addr + buf_offset),
1088 batch_copy[vq->batch_copy_nb_elems].dst =
1089 (void *)((uintptr_t)(buf_addr + buf_offset));
1090 batch_copy[vq->batch_copy_nb_elems].src =
1091 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
1092 batch_copy[vq->batch_copy_nb_elems].log_addr =
1093 buf_iova + buf_offset;
1094 batch_copy[vq->batch_copy_nb_elems].len =
1096 vq->batch_copy_nb_elems++;
1099 mbuf_avail -= cpy_len;
1100 mbuf_offset += cpy_len;
1101 buf_avail -= cpy_len;
1102 buf_offset += cpy_len;
1109 async_fill_iter(src_it, tlen, src_iovec, tvec_idx);
1110 async_fill_iter(dst_it, tlen, dst_iovec, tvec_idx);
1118 static __rte_always_inline int
1119 vhost_enqueue_single_packed(struct virtio_net *dev,
1120 struct vhost_virtqueue *vq,
1121 struct rte_mbuf *pkt,
1122 struct buf_vector *buf_vec,
1125 uint16_t nr_vec = 0;
1126 uint16_t avail_idx = vq->last_avail_idx;
1127 uint16_t max_tries, tries = 0;
1128 uint16_t buf_id = 0;
1130 uint16_t desc_count;
1131 uint32_t size = pkt->pkt_len + sizeof(struct virtio_net_hdr_mrg_rxbuf);
1132 uint16_t num_buffers = 0;
1133 uint32_t buffer_len[vq->size];
1134 uint16_t buffer_buf_id[vq->size];
1135 uint16_t buffer_desc_count[vq->size];
1137 if (rxvq_is_mergeable(dev))
1138 max_tries = vq->size - 1;
1144 * if we tried all available ring items, and still
1145 * can't get enough buf, it means something abnormal
1148 if (unlikely(++tries > max_tries))
1151 if (unlikely(fill_vec_buf_packed(dev, vq,
1152 avail_idx, &desc_count,
1155 VHOST_ACCESS_RW) < 0))
1158 len = RTE_MIN(len, size);
1161 buffer_len[num_buffers] = len;
1162 buffer_buf_id[num_buffers] = buf_id;
1163 buffer_desc_count[num_buffers] = desc_count;
1166 *nr_descs += desc_count;
1167 avail_idx += desc_count;
1168 if (avail_idx >= vq->size)
1169 avail_idx -= vq->size;
1172 if (copy_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, num_buffers) < 0)
1175 vhost_shadow_enqueue_single_packed(dev, vq, buffer_len, buffer_buf_id,
1176 buffer_desc_count, num_buffers);
1181 static __rte_noinline uint32_t
1182 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1183 struct rte_mbuf **pkts, uint32_t count)
1185 uint32_t pkt_idx = 0;
1186 uint16_t num_buffers;
1187 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1188 uint16_t avail_head;
1191 * The ordering between avail index and
1192 * desc reads needs to be enforced.
1194 avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1196 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1198 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1199 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1200 uint16_t nr_vec = 0;
1202 if (unlikely(reserve_avail_buf_split(dev, vq,
1203 pkt_len, buf_vec, &num_buffers,
1204 avail_head, &nr_vec) < 0)) {
1205 VHOST_LOG_DATA(DEBUG,
1206 "(%d) failed to get enough desc from vring\n",
1208 vq->shadow_used_idx -= num_buffers;
1212 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1213 dev->vid, vq->last_avail_idx,
1214 vq->last_avail_idx + num_buffers);
1216 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
1219 vq->shadow_used_idx -= num_buffers;
1223 vq->last_avail_idx += num_buffers;
1226 do_data_copy_enqueue(dev, vq);
1228 if (likely(vq->shadow_used_idx)) {
1229 flush_shadow_used_ring_split(dev, vq);
1230 vhost_vring_call_split(dev, vq);
1236 static __rte_always_inline int
1237 virtio_dev_rx_batch_packed(struct virtio_net *dev,
1238 struct vhost_virtqueue *vq,
1239 struct rte_mbuf **pkts)
1241 bool wrap_counter = vq->avail_wrap_counter;
1242 struct vring_packed_desc *descs = vq->desc_packed;
1243 uint16_t avail_idx = vq->last_avail_idx;
1244 uint64_t desc_addrs[PACKED_BATCH_SIZE];
1245 struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
1246 uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1247 uint64_t lens[PACKED_BATCH_SIZE];
1248 uint16_t ids[PACKED_BATCH_SIZE];
1251 if (unlikely(avail_idx & PACKED_BATCH_MASK))
1254 if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1257 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1258 if (unlikely(pkts[i]->next != NULL))
1260 if (unlikely(!desc_is_avail(&descs[avail_idx + i],
1265 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1266 lens[i] = descs[avail_idx + i].len;
1268 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1269 if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
1273 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1274 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1275 descs[avail_idx + i].addr,
1279 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1280 if (unlikely(!desc_addrs[i]))
1282 if (unlikely(lens[i] != descs[avail_idx + i].len))
1286 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1287 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1288 hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)
1289 (uintptr_t)desc_addrs[i];
1290 lens[i] = pkts[i]->pkt_len +
1291 sizeof(struct virtio_net_hdr_mrg_rxbuf);
1294 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1295 virtio_enqueue_offload(pkts[i], &hdrs[i]->hdr);
1297 vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1299 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1300 rte_memcpy((void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1301 rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1305 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1306 vhost_log_cache_write_iova(dev, vq, descs[avail_idx + i].addr,
1309 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1310 ids[i] = descs[avail_idx + i].id;
1312 vhost_flush_enqueue_batch_packed(dev, vq, lens, ids);
1317 static __rte_always_inline int16_t
1318 virtio_dev_rx_single_packed(struct virtio_net *dev,
1319 struct vhost_virtqueue *vq,
1320 struct rte_mbuf *pkt)
1322 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1323 uint16_t nr_descs = 0;
1325 if (unlikely(vhost_enqueue_single_packed(dev, vq, pkt, buf_vec,
1327 VHOST_LOG_DATA(DEBUG,
1328 "(%d) failed to get enough desc from vring\n",
1333 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1334 dev->vid, vq->last_avail_idx,
1335 vq->last_avail_idx + nr_descs);
1337 vq_inc_last_avail_packed(vq, nr_descs);
1342 static __rte_noinline uint32_t
1343 virtio_dev_rx_packed(struct virtio_net *dev,
1344 struct vhost_virtqueue *__rte_restrict vq,
1345 struct rte_mbuf **__rte_restrict pkts,
1348 uint32_t pkt_idx = 0;
1349 uint32_t remained = count;
1352 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1354 if (remained >= PACKED_BATCH_SIZE) {
1355 if (!virtio_dev_rx_batch_packed(dev, vq,
1357 pkt_idx += PACKED_BATCH_SIZE;
1358 remained -= PACKED_BATCH_SIZE;
1363 if (virtio_dev_rx_single_packed(dev, vq, pkts[pkt_idx]))
1368 } while (pkt_idx < count);
1370 if (vq->shadow_used_idx) {
1371 do_data_copy_enqueue(dev, vq);
1372 vhost_flush_enqueue_shadow_packed(dev, vq);
1376 vhost_vring_call_packed(dev, vq);
1381 static __rte_always_inline uint32_t
1382 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
1383 struct rte_mbuf **pkts, uint32_t count)
1385 struct vhost_virtqueue *vq;
1388 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1389 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1390 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1391 dev->vid, __func__, queue_id);
1395 vq = dev->virtqueue[queue_id];
1397 rte_spinlock_lock(&vq->access_lock);
1399 if (unlikely(vq->enabled == 0))
1400 goto out_access_unlock;
1402 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1403 vhost_user_iotlb_rd_lock(vq);
1405 if (unlikely(vq->access_ok == 0))
1406 if (unlikely(vring_translate(dev, vq) < 0))
1409 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1413 if (vq_is_packed(dev))
1414 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
1416 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
1419 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1420 vhost_user_iotlb_rd_unlock(vq);
1423 rte_spinlock_unlock(&vq->access_lock);
1429 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
1430 struct rte_mbuf **__rte_restrict pkts, uint16_t count)
1432 struct virtio_net *dev = get_device(vid);
1437 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1439 "(%d) %s: built-in vhost net backend is disabled.\n",
1440 dev->vid, __func__);
1444 return virtio_dev_rx(dev, queue_id, pkts, count);
1447 static __rte_always_inline uint16_t
1448 virtio_dev_rx_async_get_info_idx(uint16_t pkts_idx,
1449 uint16_t vq_size, uint16_t n_inflight)
1451 return pkts_idx > n_inflight ? (pkts_idx - n_inflight) :
1452 (vq_size - n_inflight + pkts_idx) & (vq_size - 1);
1455 static __rte_noinline uint32_t
1456 virtio_dev_rx_async_submit_split(struct virtio_net *dev,
1457 struct vhost_virtqueue *vq, uint16_t queue_id,
1458 struct rte_mbuf **pkts, uint32_t count,
1459 struct rte_mbuf **comp_pkts, uint32_t *comp_count)
1461 uint32_t pkt_idx = 0, pkt_burst_idx = 0;
1462 uint16_t num_buffers;
1463 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1464 uint16_t avail_head;
1466 struct rte_vhost_iov_iter *it_pool = vq->it_pool;
1467 struct iovec *vec_pool = vq->vec_pool;
1468 struct rte_vhost_async_desc tdes[MAX_PKT_BURST];
1469 struct iovec *src_iovec = vec_pool;
1470 struct iovec *dst_iovec = vec_pool + (VHOST_MAX_ASYNC_VEC >> 1);
1471 struct rte_vhost_iov_iter *src_it = it_pool;
1472 struct rte_vhost_iov_iter *dst_it = it_pool + 1;
1473 uint16_t slot_idx = 0;
1474 uint16_t segs_await = 0;
1475 struct async_inflight_info *pkts_info = vq->async_pkts_info;
1476 uint32_t n_pkts = 0, pkt_err = 0;
1477 uint32_t num_async_pkts = 0, num_done_pkts = 0;
1480 uint16_t last_avail_idx;
1481 } async_pkts_log[MAX_PKT_BURST];
1484 * The ordering between avail index and desc reads need to be enforced.
1486 avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1488 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1490 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1491 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1492 uint16_t nr_vec = 0;
1494 if (unlikely(reserve_avail_buf_split(dev, vq,
1495 pkt_len, buf_vec, &num_buffers,
1496 avail_head, &nr_vec) < 0)) {
1497 VHOST_LOG_DATA(DEBUG,
1498 "(%d) failed to get enough desc from vring\n",
1500 vq->shadow_used_idx -= num_buffers;
1504 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1505 dev->vid, vq->last_avail_idx,
1506 vq->last_avail_idx + num_buffers);
1508 if (async_mbuf_to_desc(dev, vq, pkts[pkt_idx],
1509 buf_vec, nr_vec, num_buffers,
1510 src_iovec, dst_iovec, src_it, dst_it) < 0) {
1511 vq->shadow_used_idx -= num_buffers;
1515 slot_idx = (vq->async_pkts_idx + num_async_pkts) &
1517 if (src_it->count) {
1520 async_fill_desc(&tdes[pkt_burst_idx++], src_it, dst_it);
1521 pkts_info[slot_idx].descs = num_buffers;
1522 pkts_info[slot_idx].mbuf = pkts[pkt_idx];
1523 async_pkts_log[num_async_pkts].pkt_idx = pkt_idx;
1524 async_pkts_log[num_async_pkts++].last_avail_idx =
1526 src_iovec += src_it->nr_segs;
1527 dst_iovec += dst_it->nr_segs;
1530 segs_await += src_it->nr_segs;
1533 * recover shadow used ring and keep DMA-occupied
1536 from = vq->shadow_used_idx - num_buffers;
1537 to = vq->async_desc_idx & (vq->size - 1);
1538 if (num_buffers + to <= vq->size) {
1539 rte_memcpy(&vq->async_descs_split[to],
1540 &vq->shadow_used_split[from],
1542 sizeof(struct vring_used_elem));
1544 int size = vq->size - to;
1546 rte_memcpy(&vq->async_descs_split[to],
1547 &vq->shadow_used_split[from],
1549 sizeof(struct vring_used_elem));
1550 rte_memcpy(vq->async_descs_split,
1551 &vq->shadow_used_split[from +
1552 size], (num_buffers - size) *
1553 sizeof(struct vring_used_elem));
1555 vq->async_desc_idx += num_buffers;
1556 vq->shadow_used_idx -= num_buffers;
1558 comp_pkts[num_done_pkts++] = pkts[pkt_idx];
1560 vq->last_avail_idx += num_buffers;
1563 * conditions to trigger async device transfer:
1564 * - buffered packet number reaches transfer threshold
1565 * - unused async iov number is less than max vhost vector
1567 if (unlikely(pkt_burst_idx >= VHOST_ASYNC_BATCH_THRESHOLD ||
1568 ((VHOST_MAX_ASYNC_VEC >> 1) - segs_await <
1570 n_pkts = vq->async_ops.transfer_data(dev->vid,
1571 queue_id, tdes, 0, pkt_burst_idx);
1572 src_iovec = vec_pool;
1573 dst_iovec = vec_pool + (VHOST_MAX_ASYNC_VEC >> 1);
1575 dst_it = it_pool + 1;
1577 vq->async_pkts_inflight_n += n_pkts;
1579 if (unlikely(n_pkts < pkt_burst_idx)) {
1581 * log error packets number here and do actual
1582 * error processing when applications poll
1585 pkt_err = pkt_burst_idx - n_pkts;
1594 if (pkt_burst_idx) {
1595 n_pkts = vq->async_ops.transfer_data(dev->vid,
1596 queue_id, tdes, 0, pkt_burst_idx);
1597 vq->async_pkts_inflight_n += n_pkts;
1599 if (unlikely(n_pkts < pkt_burst_idx))
1600 pkt_err = pkt_burst_idx - n_pkts;
1603 do_data_copy_enqueue(dev, vq);
1605 if (unlikely(pkt_err)) {
1606 uint16_t num_descs = 0;
1608 num_async_pkts -= pkt_err;
1609 /* calculate the sum of descriptors of DMA-error packets. */
1610 while (pkt_err-- > 0) {
1611 num_descs += pkts_info[slot_idx & (vq->size - 1)].descs;
1614 vq->async_desc_idx -= num_descs;
1615 /* recover shadow used ring and available ring */
1616 vq->shadow_used_idx -= (vq->last_avail_idx -
1617 async_pkts_log[num_async_pkts].last_avail_idx -
1619 vq->last_avail_idx =
1620 async_pkts_log[num_async_pkts].last_avail_idx;
1621 pkt_idx = async_pkts_log[num_async_pkts].pkt_idx;
1622 num_done_pkts = pkt_idx - num_async_pkts;
1625 vq->async_pkts_idx += num_async_pkts;
1626 *comp_count = num_done_pkts;
1628 if (likely(vq->shadow_used_idx)) {
1629 flush_shadow_used_ring_split(dev, vq);
1630 vhost_vring_call_split(dev, vq);
1636 uint16_t rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
1637 struct rte_mbuf **pkts, uint16_t count)
1639 struct virtio_net *dev = get_device(vid);
1640 struct vhost_virtqueue *vq;
1641 uint16_t n_pkts_cpl = 0, n_pkts_put = 0, n_descs = 0;
1642 uint16_t start_idx, pkts_idx, vq_size;
1643 struct async_inflight_info *pkts_info;
1649 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1650 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1651 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1652 dev->vid, __func__, queue_id);
1656 vq = dev->virtqueue[queue_id];
1658 if (unlikely(!vq->async_registered)) {
1659 VHOST_LOG_DATA(ERR, "(%d) %s: async not registered for queue id %d.\n",
1660 dev->vid, __func__, queue_id);
1664 rte_spinlock_lock(&vq->access_lock);
1666 pkts_idx = vq->async_pkts_idx & (vq->size - 1);
1667 pkts_info = vq->async_pkts_info;
1669 start_idx = virtio_dev_rx_async_get_info_idx(pkts_idx,
1670 vq_size, vq->async_pkts_inflight_n);
1672 if (count > vq->async_last_pkts_n)
1673 n_pkts_cpl = vq->async_ops.check_completed_copies(vid,
1674 queue_id, 0, count - vq->async_last_pkts_n);
1675 n_pkts_cpl += vq->async_last_pkts_n;
1677 n_pkts_put = RTE_MIN(count, n_pkts_cpl);
1678 if (unlikely(n_pkts_put == 0)) {
1679 vq->async_last_pkts_n = n_pkts_cpl;
1683 for (i = 0; i < n_pkts_put; i++) {
1684 from = (start_idx + i) & (vq_size - 1);
1685 n_descs += pkts_info[from].descs;
1686 pkts[i] = pkts_info[from].mbuf;
1688 vq->async_last_pkts_n = n_pkts_cpl - n_pkts_put;
1689 vq->async_pkts_inflight_n -= n_pkts_put;
1691 if (likely(vq->enabled && vq->access_ok)) {
1692 uint16_t nr_left = n_descs;
1696 /* write back completed descriptors to used ring */
1698 from = vq->last_async_desc_idx & (vq->size - 1);
1699 nr_copy = nr_left + from <= vq->size ? nr_left :
1701 to = vq->last_used_idx & (vq->size - 1);
1703 if (to + nr_copy <= vq->size) {
1704 rte_memcpy(&vq->used->ring[to],
1705 &vq->async_descs_split[from],
1707 sizeof(struct vring_used_elem));
1709 uint16_t size = vq->size - to;
1711 rte_memcpy(&vq->used->ring[to],
1712 &vq->async_descs_split[from],
1714 sizeof(struct vring_used_elem));
1715 rte_memcpy(vq->used->ring,
1716 &vq->async_descs_split[from +
1717 size], (nr_copy - size) *
1718 sizeof(struct vring_used_elem));
1721 vq->last_async_desc_idx += nr_copy;
1722 vq->last_used_idx += nr_copy;
1724 } while (nr_left > 0);
1726 __atomic_add_fetch(&vq->used->idx, n_descs, __ATOMIC_RELEASE);
1727 vhost_vring_call_split(dev, vq);
1729 vq->last_async_desc_idx += n_descs;
1732 rte_spinlock_unlock(&vq->access_lock);
1737 static __rte_always_inline uint32_t
1738 virtio_dev_rx_async_submit(struct virtio_net *dev, uint16_t queue_id,
1739 struct rte_mbuf **pkts, uint32_t count,
1740 struct rte_mbuf **comp_pkts, uint32_t *comp_count)
1742 struct vhost_virtqueue *vq;
1745 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1746 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1747 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1748 dev->vid, __func__, queue_id);
1752 vq = dev->virtqueue[queue_id];
1754 rte_spinlock_lock(&vq->access_lock);
1756 if (unlikely(vq->enabled == 0 || !vq->async_registered))
1757 goto out_access_unlock;
1759 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1760 vhost_user_iotlb_rd_lock(vq);
1762 if (unlikely(vq->access_ok == 0))
1763 if (unlikely(vring_translate(dev, vq) < 0))
1766 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1770 /* TODO: packed queue not implemented */
1771 if (vq_is_packed(dev))
1774 nb_tx = virtio_dev_rx_async_submit_split(dev,
1775 vq, queue_id, pkts, count, comp_pkts,
1779 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1780 vhost_user_iotlb_rd_unlock(vq);
1783 rte_spinlock_unlock(&vq->access_lock);
1789 rte_vhost_submit_enqueue_burst(int vid, uint16_t queue_id,
1790 struct rte_mbuf **pkts, uint16_t count,
1791 struct rte_mbuf **comp_pkts, uint32_t *comp_count)
1793 struct virtio_net *dev = get_device(vid);
1799 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1801 "(%d) %s: built-in vhost net backend is disabled.\n",
1802 dev->vid, __func__);
1806 return virtio_dev_rx_async_submit(dev, queue_id, pkts, count, comp_pkts,
1811 virtio_net_with_host_offload(struct virtio_net *dev)
1814 ((1ULL << VIRTIO_NET_F_CSUM) |
1815 (1ULL << VIRTIO_NET_F_HOST_ECN) |
1816 (1ULL << VIRTIO_NET_F_HOST_TSO4) |
1817 (1ULL << VIRTIO_NET_F_HOST_TSO6) |
1818 (1ULL << VIRTIO_NET_F_HOST_UFO)))
1825 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
1827 struct rte_ipv4_hdr *ipv4_hdr;
1828 struct rte_ipv6_hdr *ipv6_hdr;
1829 void *l3_hdr = NULL;
1830 struct rte_ether_hdr *eth_hdr;
1833 eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
1835 m->l2_len = sizeof(struct rte_ether_hdr);
1836 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
1838 if (ethertype == RTE_ETHER_TYPE_VLAN) {
1839 struct rte_vlan_hdr *vlan_hdr =
1840 (struct rte_vlan_hdr *)(eth_hdr + 1);
1842 m->l2_len += sizeof(struct rte_vlan_hdr);
1843 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
1846 l3_hdr = (char *)eth_hdr + m->l2_len;
1848 switch (ethertype) {
1849 case RTE_ETHER_TYPE_IPV4:
1851 *l4_proto = ipv4_hdr->next_proto_id;
1852 m->l3_len = rte_ipv4_hdr_len(ipv4_hdr);
1853 *l4_hdr = (char *)l3_hdr + m->l3_len;
1854 m->ol_flags |= PKT_TX_IPV4;
1856 case RTE_ETHER_TYPE_IPV6:
1858 *l4_proto = ipv6_hdr->proto;
1859 m->l3_len = sizeof(struct rte_ipv6_hdr);
1860 *l4_hdr = (char *)l3_hdr + m->l3_len;
1861 m->ol_flags |= PKT_TX_IPV6;
1871 static __rte_always_inline void
1872 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
1874 uint16_t l4_proto = 0;
1875 void *l4_hdr = NULL;
1876 struct rte_tcp_hdr *tcp_hdr = NULL;
1878 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1881 parse_ethernet(m, &l4_proto, &l4_hdr);
1882 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1883 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
1884 switch (hdr->csum_offset) {
1885 case (offsetof(struct rte_tcp_hdr, cksum)):
1886 if (l4_proto == IPPROTO_TCP)
1887 m->ol_flags |= PKT_TX_TCP_CKSUM;
1889 case (offsetof(struct rte_udp_hdr, dgram_cksum)):
1890 if (l4_proto == IPPROTO_UDP)
1891 m->ol_flags |= PKT_TX_UDP_CKSUM;
1893 case (offsetof(struct rte_sctp_hdr, cksum)):
1894 if (l4_proto == IPPROTO_SCTP)
1895 m->ol_flags |= PKT_TX_SCTP_CKSUM;
1903 if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1904 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1905 case VIRTIO_NET_HDR_GSO_TCPV4:
1906 case VIRTIO_NET_HDR_GSO_TCPV6:
1908 m->ol_flags |= PKT_TX_TCP_SEG;
1909 m->tso_segsz = hdr->gso_size;
1910 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
1912 case VIRTIO_NET_HDR_GSO_UDP:
1913 m->ol_flags |= PKT_TX_UDP_SEG;
1914 m->tso_segsz = hdr->gso_size;
1915 m->l4_len = sizeof(struct rte_udp_hdr);
1918 VHOST_LOG_DATA(WARNING,
1919 "unsupported gso type %u.\n", hdr->gso_type);
1925 static __rte_noinline void
1926 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
1927 struct buf_vector *buf_vec)
1930 uint64_t remain = sizeof(struct virtio_net_hdr);
1932 uint64_t dst = (uint64_t)(uintptr_t)hdr;
1935 len = RTE_MIN(remain, buf_vec->buf_len);
1936 src = buf_vec->buf_addr;
1937 rte_memcpy((void *)(uintptr_t)dst,
1938 (void *)(uintptr_t)src, len);
1946 static __rte_always_inline int
1947 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
1948 struct buf_vector *buf_vec, uint16_t nr_vec,
1949 struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
1951 uint32_t buf_avail, buf_offset;
1952 uint64_t buf_addr, buf_len;
1953 uint32_t mbuf_avail, mbuf_offset;
1955 struct rte_mbuf *cur = m, *prev = m;
1956 struct virtio_net_hdr tmp_hdr;
1957 struct virtio_net_hdr *hdr = NULL;
1958 /* A counter to avoid desc dead loop chain */
1959 uint16_t vec_idx = 0;
1960 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
1963 buf_addr = buf_vec[vec_idx].buf_addr;
1964 buf_len = buf_vec[vec_idx].buf_len;
1966 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
1971 if (virtio_net_with_host_offload(dev)) {
1972 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
1974 * No luck, the virtio-net header doesn't fit
1975 * in a contiguous virtual area.
1977 copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
1980 hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
1985 * A virtio driver normally uses at least 2 desc buffers
1986 * for Tx: the first for storing the header, and others
1987 * for storing the data.
1989 if (unlikely(buf_len < dev->vhost_hlen)) {
1990 buf_offset = dev->vhost_hlen - buf_len;
1992 buf_addr = buf_vec[vec_idx].buf_addr;
1993 buf_len = buf_vec[vec_idx].buf_len;
1994 buf_avail = buf_len - buf_offset;
1995 } else if (buf_len == dev->vhost_hlen) {
1996 if (unlikely(++vec_idx >= nr_vec))
1998 buf_addr = buf_vec[vec_idx].buf_addr;
1999 buf_len = buf_vec[vec_idx].buf_len;
2002 buf_avail = buf_len;
2004 buf_offset = dev->vhost_hlen;
2005 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
2009 (uintptr_t)(buf_addr + buf_offset),
2010 (uint32_t)buf_avail, 0);
2013 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
2015 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
2017 if (likely(cpy_len > MAX_BATCH_LEN ||
2018 vq->batch_copy_nb_elems >= vq->size ||
2019 (hdr && cur == m))) {
2020 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
2022 (void *)((uintptr_t)(buf_addr +
2023 buf_offset)), cpy_len);
2025 batch_copy[vq->batch_copy_nb_elems].dst =
2026 rte_pktmbuf_mtod_offset(cur, void *,
2028 batch_copy[vq->batch_copy_nb_elems].src =
2029 (void *)((uintptr_t)(buf_addr + buf_offset));
2030 batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
2031 vq->batch_copy_nb_elems++;
2034 mbuf_avail -= cpy_len;
2035 mbuf_offset += cpy_len;
2036 buf_avail -= cpy_len;
2037 buf_offset += cpy_len;
2039 /* This buf reaches to its end, get the next one */
2040 if (buf_avail == 0) {
2041 if (++vec_idx >= nr_vec)
2044 buf_addr = buf_vec[vec_idx].buf_addr;
2045 buf_len = buf_vec[vec_idx].buf_len;
2048 buf_avail = buf_len;
2050 PRINT_PACKET(dev, (uintptr_t)buf_addr,
2051 (uint32_t)buf_avail, 0);
2055 * This mbuf reaches to its end, get a new one
2056 * to hold more data.
2058 if (mbuf_avail == 0) {
2059 cur = rte_pktmbuf_alloc(mbuf_pool);
2060 if (unlikely(cur == NULL)) {
2061 VHOST_LOG_DATA(ERR, "Failed to "
2062 "allocate memory for mbuf.\n");
2068 prev->data_len = mbuf_offset;
2070 m->pkt_len += mbuf_offset;
2074 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
2078 prev->data_len = mbuf_offset;
2079 m->pkt_len += mbuf_offset;
2082 vhost_dequeue_offload(hdr, m);
2090 virtio_dev_extbuf_free(void *addr __rte_unused, void *opaque)
2096 virtio_dev_extbuf_alloc(struct rte_mbuf *pkt, uint32_t size)
2098 struct rte_mbuf_ext_shared_info *shinfo = NULL;
2099 uint32_t total_len = RTE_PKTMBUF_HEADROOM + size;
2104 total_len += sizeof(*shinfo) + sizeof(uintptr_t);
2105 total_len = RTE_ALIGN_CEIL(total_len, sizeof(uintptr_t));
2107 if (unlikely(total_len > UINT16_MAX))
2110 buf_len = total_len;
2111 buf = rte_malloc(NULL, buf_len, RTE_CACHE_LINE_SIZE);
2112 if (unlikely(buf == NULL))
2115 /* Initialize shinfo */
2116 shinfo = rte_pktmbuf_ext_shinfo_init_helper(buf, &buf_len,
2117 virtio_dev_extbuf_free, buf);
2118 if (unlikely(shinfo == NULL)) {
2120 VHOST_LOG_DATA(ERR, "Failed to init shinfo\n");
2124 iova = rte_malloc_virt2iova(buf);
2125 rte_pktmbuf_attach_extbuf(pkt, buf, iova, buf_len, shinfo);
2126 rte_pktmbuf_reset_headroom(pkt);
2132 * Allocate a host supported pktmbuf.
2134 static __rte_always_inline struct rte_mbuf *
2135 virtio_dev_pktmbuf_alloc(struct virtio_net *dev, struct rte_mempool *mp,
2138 struct rte_mbuf *pkt = rte_pktmbuf_alloc(mp);
2140 if (unlikely(pkt == NULL)) {
2142 "Failed to allocate memory for mbuf.\n");
2146 if (rte_pktmbuf_tailroom(pkt) >= data_len)
2149 /* attach an external buffer if supported */
2150 if (dev->extbuf && !virtio_dev_extbuf_alloc(pkt, data_len))
2153 /* check if chained buffers are allowed */
2154 if (!dev->linearbuf)
2157 /* Data doesn't fit into the buffer and the host supports
2158 * only linear buffers
2160 rte_pktmbuf_free(pkt);
2165 static __rte_noinline uint16_t
2166 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
2167 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
2170 uint16_t free_entries;
2171 uint16_t dropped = 0;
2172 static bool allocerr_warned;
2175 * The ordering between avail index and
2176 * desc reads needs to be enforced.
2178 free_entries = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE) -
2180 if (free_entries == 0)
2183 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
2185 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2187 count = RTE_MIN(count, MAX_PKT_BURST);
2188 count = RTE_MIN(count, free_entries);
2189 VHOST_LOG_DATA(DEBUG, "(%d) about to dequeue %u buffers\n",
2192 for (i = 0; i < count; i++) {
2193 struct buf_vector buf_vec[BUF_VECTOR_MAX];
2196 uint16_t nr_vec = 0;
2199 if (unlikely(fill_vec_buf_split(dev, vq,
2200 vq->last_avail_idx + i,
2202 &head_idx, &buf_len,
2203 VHOST_ACCESS_RO) < 0))
2206 update_shadow_used_ring_split(vq, head_idx, 0);
2208 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
2209 if (unlikely(pkts[i] == NULL)) {
2211 * mbuf allocation fails for jumbo packets when external
2212 * buffer allocation is not allowed and linear buffer
2213 * is required. Drop this packet.
2215 if (!allocerr_warned) {
2217 "Failed mbuf alloc of size %d from %s on %s.\n",
2218 buf_len, mbuf_pool->name, dev->ifname);
2219 allocerr_warned = true;
2226 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
2228 if (unlikely(err)) {
2229 rte_pktmbuf_free(pkts[i]);
2230 if (!allocerr_warned) {
2232 "Failed to copy desc to mbuf on %s.\n",
2234 allocerr_warned = true;
2242 vq->last_avail_idx += i;
2244 do_data_copy_dequeue(vq);
2245 if (unlikely(i < count))
2246 vq->shadow_used_idx = i;
2247 if (likely(vq->shadow_used_idx)) {
2248 flush_shadow_used_ring_split(dev, vq);
2249 vhost_vring_call_split(dev, vq);
2252 return (i - dropped);
2255 static __rte_always_inline int
2256 vhost_reserve_avail_batch_packed(struct virtio_net *dev,
2257 struct vhost_virtqueue *vq,
2258 struct rte_mempool *mbuf_pool,
2259 struct rte_mbuf **pkts,
2261 uintptr_t *desc_addrs,
2264 bool wrap = vq->avail_wrap_counter;
2265 struct vring_packed_desc *descs = vq->desc_packed;
2266 uint64_t lens[PACKED_BATCH_SIZE];
2267 uint64_t buf_lens[PACKED_BATCH_SIZE];
2268 uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2271 if (unlikely(avail_idx & PACKED_BATCH_MASK))
2273 if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
2276 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2277 flags = descs[avail_idx + i].flags;
2278 if (unlikely((wrap != !!(flags & VRING_DESC_F_AVAIL)) ||
2279 (wrap == !!(flags & VRING_DESC_F_USED)) ||
2280 (flags & PACKED_DESC_SINGLE_DEQUEUE_FLAG)))
2284 rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
2286 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2287 lens[i] = descs[avail_idx + i].len;
2289 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2290 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
2291 descs[avail_idx + i].addr,
2292 &lens[i], VHOST_ACCESS_RW);
2295 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2296 if (unlikely(!desc_addrs[i]))
2298 if (unlikely((lens[i] != descs[avail_idx + i].len)))
2302 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2303 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, lens[i]);
2308 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2309 buf_lens[i] = pkts[i]->buf_len - pkts[i]->data_off;
2311 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2312 if (unlikely(buf_lens[i] < (lens[i] - buf_offset)))
2316 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2317 pkts[i]->pkt_len = descs[avail_idx + i].len - buf_offset;
2318 pkts[i]->data_len = pkts[i]->pkt_len;
2319 ids[i] = descs[avail_idx + i].id;
2325 for (i = 0; i < PACKED_BATCH_SIZE; i++)
2326 rte_pktmbuf_free(pkts[i]);
2331 static __rte_always_inline int
2332 virtio_dev_tx_batch_packed(struct virtio_net *dev,
2333 struct vhost_virtqueue *vq,
2334 struct rte_mempool *mbuf_pool,
2335 struct rte_mbuf **pkts)
2337 uint16_t avail_idx = vq->last_avail_idx;
2338 uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2339 struct virtio_net_hdr *hdr;
2340 uintptr_t desc_addrs[PACKED_BATCH_SIZE];
2341 uint16_t ids[PACKED_BATCH_SIZE];
2344 if (vhost_reserve_avail_batch_packed(dev, vq, mbuf_pool, pkts,
2345 avail_idx, desc_addrs, ids))
2348 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2349 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
2351 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2352 rte_memcpy(rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
2353 (void *)(uintptr_t)(desc_addrs[i] + buf_offset),
2356 if (virtio_net_with_host_offload(dev)) {
2357 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2358 hdr = (struct virtio_net_hdr *)(desc_addrs[i]);
2359 vhost_dequeue_offload(hdr, pkts[i]);
2363 if (virtio_net_is_inorder(dev))
2364 vhost_shadow_dequeue_batch_packed_inorder(vq,
2365 ids[PACKED_BATCH_SIZE - 1]);
2367 vhost_shadow_dequeue_batch_packed(dev, vq, ids);
2369 vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
2374 static __rte_always_inline int
2375 vhost_dequeue_single_packed(struct virtio_net *dev,
2376 struct vhost_virtqueue *vq,
2377 struct rte_mempool *mbuf_pool,
2378 struct rte_mbuf **pkts,
2380 uint16_t *desc_count)
2382 struct buf_vector buf_vec[BUF_VECTOR_MAX];
2384 uint16_t nr_vec = 0;
2386 static bool allocerr_warned;
2388 if (unlikely(fill_vec_buf_packed(dev, vq,
2389 vq->last_avail_idx, desc_count,
2392 VHOST_ACCESS_RO) < 0))
2395 *pkts = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
2396 if (unlikely(*pkts == NULL)) {
2397 if (!allocerr_warned) {
2399 "Failed mbuf alloc of size %d from %s on %s.\n",
2400 buf_len, mbuf_pool->name, dev->ifname);
2401 allocerr_warned = true;
2406 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, *pkts,
2408 if (unlikely(err)) {
2409 if (!allocerr_warned) {
2411 "Failed to copy desc to mbuf on %s.\n",
2413 allocerr_warned = true;
2415 rte_pktmbuf_free(*pkts);
2422 static __rte_always_inline int
2423 virtio_dev_tx_single_packed(struct virtio_net *dev,
2424 struct vhost_virtqueue *vq,
2425 struct rte_mempool *mbuf_pool,
2426 struct rte_mbuf **pkts)
2429 uint16_t buf_id, desc_count = 0;
2432 ret = vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
2435 if (likely(desc_count > 0)) {
2436 if (virtio_net_is_inorder(dev))
2437 vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
2440 vhost_shadow_dequeue_single_packed(vq, buf_id,
2443 vq_inc_last_avail_packed(vq, desc_count);
2449 static __rte_noinline uint16_t
2450 virtio_dev_tx_packed(struct virtio_net *dev,
2451 struct vhost_virtqueue *__rte_restrict vq,
2452 struct rte_mempool *mbuf_pool,
2453 struct rte_mbuf **__rte_restrict pkts,
2456 uint32_t pkt_idx = 0;
2457 uint32_t remained = count;
2460 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
2462 if (remained >= PACKED_BATCH_SIZE) {
2463 if (!virtio_dev_tx_batch_packed(dev, vq, mbuf_pool,
2465 pkt_idx += PACKED_BATCH_SIZE;
2466 remained -= PACKED_BATCH_SIZE;
2471 if (virtio_dev_tx_single_packed(dev, vq, mbuf_pool,
2479 if (vq->shadow_used_idx) {
2480 do_data_copy_dequeue(vq);
2482 vhost_flush_dequeue_shadow_packed(dev, vq);
2483 vhost_vring_call_packed(dev, vq);
2490 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
2491 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
2493 struct virtio_net *dev;
2494 struct rte_mbuf *rarp_mbuf = NULL;
2495 struct vhost_virtqueue *vq;
2496 int16_t success = 1;
2498 dev = get_device(vid);
2502 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
2504 "(%d) %s: built-in vhost net backend is disabled.\n",
2505 dev->vid, __func__);
2509 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
2511 "(%d) %s: invalid virtqueue idx %d.\n",
2512 dev->vid, __func__, queue_id);
2516 vq = dev->virtqueue[queue_id];
2518 if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
2521 if (unlikely(vq->enabled == 0)) {
2523 goto out_access_unlock;
2526 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2527 vhost_user_iotlb_rd_lock(vq);
2529 if (unlikely(vq->access_ok == 0))
2530 if (unlikely(vring_translate(dev, vq) < 0)) {
2536 * Construct a RARP broadcast packet, and inject it to the "pkts"
2537 * array, to looks like that guest actually send such packet.
2539 * Check user_send_rarp() for more information.
2541 * broadcast_rarp shares a cacheline in the virtio_net structure
2542 * with some fields that are accessed during enqueue and
2543 * __atomic_compare_exchange_n causes a write if performed compare
2544 * and exchange. This could result in false sharing between enqueue
2547 * Prevent unnecessary false sharing by reading broadcast_rarp first
2548 * and only performing compare and exchange if the read indicates it
2549 * is likely to be set.
2551 if (unlikely(__atomic_load_n(&dev->broadcast_rarp, __ATOMIC_ACQUIRE) &&
2552 __atomic_compare_exchange_n(&dev->broadcast_rarp,
2553 &success, 0, 0, __ATOMIC_RELEASE, __ATOMIC_RELAXED))) {
2555 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
2556 if (rarp_mbuf == NULL) {
2557 VHOST_LOG_DATA(ERR, "Failed to make RARP packet.\n");
2564 if (vq_is_packed(dev))
2565 count = virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count);
2567 count = virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count);
2570 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2571 vhost_user_iotlb_rd_unlock(vq);
2574 rte_spinlock_unlock(&vq->access_lock);
2576 if (unlikely(rarp_mbuf != NULL)) {
2578 * Inject it to the head of "pkts" array, so that switch's mac
2579 * learning table will get updated first.
2581 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
2582 pkts[0] = rarp_mbuf;