1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation
11 #include <rte_cycles.h>
12 #include <rte_memory.h>
13 #include <rte_branch_prediction.h>
14 #include <rte_mempool.h>
15 #include <rte_malloc.h>
17 #include <rte_ether.h>
18 #include <rte_ethdev_driver.h>
19 #include <rte_prefetch.h>
20 #include <rte_string_fns.h>
21 #include <rte_errno.h>
22 #include <rte_byteorder.h>
28 #include "virtio_logs.h"
29 #include "virtio_ethdev.h"
30 #include "virtio_pci.h"
31 #include "virtqueue.h"
32 #include "virtio_rxtx.h"
33 #include "virtio_rxtx_simple.h"
34 #include "virtio_ring.h"
36 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
37 #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
39 #define VIRTIO_DUMP_PACKET(m, len) do { } while (0)
43 virtio_dev_rx_queue_done(void *rxq, uint16_t offset)
45 struct virtnet_rx *rxvq = rxq;
46 struct virtqueue *vq = rxvq->vq;
48 return VIRTQUEUE_NUSED(vq) >= offset;
52 vq_ring_free_inorder(struct virtqueue *vq, uint16_t desc_idx, uint16_t num)
54 vq->vq_free_cnt += num;
55 vq->vq_desc_tail_idx = desc_idx & (vq->vq_nentries - 1);
59 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
61 struct vring_desc *dp, *dp_tail;
62 struct vq_desc_extra *dxp;
63 uint16_t desc_idx_last = desc_idx;
65 dp = &vq->vq_split.ring.desc[desc_idx];
66 dxp = &vq->vq_descx[desc_idx];
67 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
68 if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
69 while (dp->flags & VRING_DESC_F_NEXT) {
70 desc_idx_last = dp->next;
71 dp = &vq->vq_split.ring.desc[dp->next];
77 * We must append the existing free chain, if any, to the end of
78 * newly freed chain. If the virtqueue was completely used, then
79 * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
81 if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
82 vq->vq_desc_head_idx = desc_idx;
84 dp_tail = &vq->vq_split.ring.desc[vq->vq_desc_tail_idx];
85 dp_tail->next = desc_idx;
88 vq->vq_desc_tail_idx = desc_idx_last;
89 dp->next = VQ_RING_DESC_CHAIN_END;
93 vq_ring_free_id_packed(struct virtqueue *vq, uint16_t id)
95 struct vq_desc_extra *dxp;
97 dxp = &vq->vq_descx[id];
98 vq->vq_free_cnt += dxp->ndescs;
100 if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END)
101 vq->vq_desc_head_idx = id;
103 vq->vq_descx[vq->vq_desc_tail_idx].next = id;
105 vq->vq_desc_tail_idx = id;
106 dxp->next = VQ_RING_DESC_CHAIN_END;
110 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
112 uint32_t s = mbuf->pkt_len;
113 struct rte_ether_addr *ea;
118 stats->size_bins[1]++;
119 } else if (s > 64 && s < 1024) {
122 /* count zeros, and offset into correct bin */
123 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
124 stats->size_bins[bin]++;
127 stats->size_bins[0]++;
129 stats->size_bins[6]++;
131 stats->size_bins[7]++;
134 ea = rte_pktmbuf_mtod(mbuf, struct rte_ether_addr *);
135 if (rte_is_multicast_ether_addr(ea)) {
136 if (rte_is_broadcast_ether_addr(ea))
144 virtio_rx_stats_updated(struct virtnet_rx *rxvq, struct rte_mbuf *m)
146 VIRTIO_DUMP_PACKET(m, m->data_len);
148 virtio_update_packet_stats(&rxvq->stats, m);
152 virtqueue_dequeue_burst_rx_packed(struct virtqueue *vq,
153 struct rte_mbuf **rx_pkts,
157 struct rte_mbuf *cookie;
160 struct vring_packed_desc *desc;
163 desc = vq->vq_packed.ring.desc;
165 for (i = 0; i < num; i++) {
166 used_idx = vq->vq_used_cons_idx;
167 /* desc_is_used has a load-acquire or rte_cio_rmb inside
168 * and wait for used desc in virtqueue.
170 if (!desc_is_used(&desc[used_idx], vq))
172 len[i] = desc[used_idx].len;
173 id = desc[used_idx].id;
174 cookie = (struct rte_mbuf *)vq->vq_descx[id].cookie;
175 if (unlikely(cookie == NULL)) {
176 PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
177 vq->vq_used_cons_idx);
180 rte_prefetch0(cookie);
181 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
185 vq->vq_used_cons_idx++;
186 if (vq->vq_used_cons_idx >= vq->vq_nentries) {
187 vq->vq_used_cons_idx -= vq->vq_nentries;
188 vq->vq_packed.used_wrap_counter ^= 1;
196 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
197 uint32_t *len, uint16_t num)
199 struct vring_used_elem *uep;
200 struct rte_mbuf *cookie;
201 uint16_t used_idx, desc_idx;
204 /* Caller does the check */
205 for (i = 0; i < num ; i++) {
206 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
207 uep = &vq->vq_split.ring.used->ring[used_idx];
208 desc_idx = (uint16_t) uep->id;
210 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
212 if (unlikely(cookie == NULL)) {
213 PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
214 vq->vq_used_cons_idx);
218 rte_prefetch0(cookie);
219 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
221 vq->vq_used_cons_idx++;
222 vq_ring_free_chain(vq, desc_idx);
223 vq->vq_descx[desc_idx].cookie = NULL;
230 virtqueue_dequeue_rx_inorder(struct virtqueue *vq,
231 struct rte_mbuf **rx_pkts,
235 struct vring_used_elem *uep;
236 struct rte_mbuf *cookie;
237 uint16_t used_idx = 0;
240 if (unlikely(num == 0))
243 for (i = 0; i < num; i++) {
244 used_idx = vq->vq_used_cons_idx & (vq->vq_nentries - 1);
245 /* Desc idx same as used idx */
246 uep = &vq->vq_split.ring.used->ring[used_idx];
248 cookie = (struct rte_mbuf *)vq->vq_descx[used_idx].cookie;
250 if (unlikely(cookie == NULL)) {
251 PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
252 vq->vq_used_cons_idx);
256 rte_prefetch0(cookie);
257 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
259 vq->vq_used_cons_idx++;
260 vq->vq_descx[used_idx].cookie = NULL;
263 vq_ring_free_inorder(vq, used_idx, i);
267 #ifndef DEFAULT_TX_FREE_THRESH
268 #define DEFAULT_TX_FREE_THRESH 32
272 virtio_xmit_cleanup_inorder_packed(struct virtqueue *vq, int num)
274 uint16_t used_idx, id, curr_id, free_cnt = 0;
275 uint16_t size = vq->vq_nentries;
276 struct vring_packed_desc *desc = vq->vq_packed.ring.desc;
277 struct vq_desc_extra *dxp;
279 used_idx = vq->vq_used_cons_idx;
280 /* desc_is_used has a load-acquire or rte_cio_rmb inside
281 * and wait for used desc in virtqueue.
283 while (num > 0 && desc_is_used(&desc[used_idx], vq)) {
284 id = desc[used_idx].id;
287 dxp = &vq->vq_descx[used_idx];
288 used_idx += dxp->ndescs;
289 free_cnt += dxp->ndescs;
291 if (used_idx >= size) {
293 vq->vq_packed.used_wrap_counter ^= 1;
295 if (dxp->cookie != NULL) {
296 rte_pktmbuf_free(dxp->cookie);
299 } while (curr_id != id);
301 vq->vq_used_cons_idx = used_idx;
302 vq->vq_free_cnt += free_cnt;
306 virtio_xmit_cleanup_normal_packed(struct virtqueue *vq, int num)
308 uint16_t used_idx, id;
309 uint16_t size = vq->vq_nentries;
310 struct vring_packed_desc *desc = vq->vq_packed.ring.desc;
311 struct vq_desc_extra *dxp;
313 used_idx = vq->vq_used_cons_idx;
314 /* desc_is_used has a load-acquire or rte_cio_rmb inside
315 * and wait for used desc in virtqueue.
317 while (num-- && desc_is_used(&desc[used_idx], vq)) {
318 id = desc[used_idx].id;
319 dxp = &vq->vq_descx[id];
320 vq->vq_used_cons_idx += dxp->ndescs;
321 if (vq->vq_used_cons_idx >= size) {
322 vq->vq_used_cons_idx -= size;
323 vq->vq_packed.used_wrap_counter ^= 1;
325 vq_ring_free_id_packed(vq, id);
326 if (dxp->cookie != NULL) {
327 rte_pktmbuf_free(dxp->cookie);
330 used_idx = vq->vq_used_cons_idx;
334 /* Cleanup from completed transmits. */
336 virtio_xmit_cleanup_packed(struct virtqueue *vq, int num, int in_order)
339 virtio_xmit_cleanup_inorder_packed(vq, num);
341 virtio_xmit_cleanup_normal_packed(vq, num);
345 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
347 uint16_t i, used_idx, desc_idx;
348 for (i = 0; i < num; i++) {
349 struct vring_used_elem *uep;
350 struct vq_desc_extra *dxp;
352 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
353 uep = &vq->vq_split.ring.used->ring[used_idx];
355 desc_idx = (uint16_t) uep->id;
356 dxp = &vq->vq_descx[desc_idx];
357 vq->vq_used_cons_idx++;
358 vq_ring_free_chain(vq, desc_idx);
360 if (dxp->cookie != NULL) {
361 rte_pktmbuf_free(dxp->cookie);
367 /* Cleanup from completed inorder transmits. */
368 static __rte_always_inline void
369 virtio_xmit_cleanup_inorder(struct virtqueue *vq, uint16_t num)
371 uint16_t i, idx = vq->vq_used_cons_idx;
372 int16_t free_cnt = 0;
373 struct vq_desc_extra *dxp = NULL;
375 if (unlikely(num == 0))
378 for (i = 0; i < num; i++) {
379 dxp = &vq->vq_descx[idx++ & (vq->vq_nentries - 1)];
380 free_cnt += dxp->ndescs;
381 if (dxp->cookie != NULL) {
382 rte_pktmbuf_free(dxp->cookie);
387 vq->vq_free_cnt += free_cnt;
388 vq->vq_used_cons_idx = idx;
392 virtqueue_enqueue_refill_inorder(struct virtqueue *vq,
393 struct rte_mbuf **cookies,
396 struct vq_desc_extra *dxp;
397 struct virtio_hw *hw = vq->hw;
398 struct vring_desc *start_dp;
399 uint16_t head_idx, idx, i = 0;
401 if (unlikely(vq->vq_free_cnt == 0))
403 if (unlikely(vq->vq_free_cnt < num))
406 head_idx = vq->vq_desc_head_idx & (vq->vq_nentries - 1);
407 start_dp = vq->vq_split.ring.desc;
410 idx = head_idx & (vq->vq_nentries - 1);
411 dxp = &vq->vq_descx[idx];
412 dxp->cookie = (void *)cookies[i];
416 VIRTIO_MBUF_ADDR(cookies[i], vq) +
417 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
419 cookies[i]->buf_len -
420 RTE_PKTMBUF_HEADROOM +
422 start_dp[idx].flags = VRING_DESC_F_WRITE;
424 vq_update_avail_ring(vq, idx);
429 vq->vq_desc_head_idx += num;
430 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
435 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf **cookie,
438 struct vq_desc_extra *dxp;
439 struct virtio_hw *hw = vq->hw;
440 struct vring_desc *start_dp = vq->vq_split.ring.desc;
443 if (unlikely(vq->vq_free_cnt == 0))
445 if (unlikely(vq->vq_free_cnt < num))
448 if (unlikely(vq->vq_desc_head_idx >= vq->vq_nentries))
451 for (i = 0; i < num; i++) {
452 idx = vq->vq_desc_head_idx;
453 dxp = &vq->vq_descx[idx];
454 dxp->cookie = (void *)cookie[i];
458 VIRTIO_MBUF_ADDR(cookie[i], vq) +
459 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
461 cookie[i]->buf_len - RTE_PKTMBUF_HEADROOM +
463 start_dp[idx].flags = VRING_DESC_F_WRITE;
464 vq->vq_desc_head_idx = start_dp[idx].next;
465 vq_update_avail_ring(vq, idx);
466 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END) {
467 vq->vq_desc_tail_idx = vq->vq_desc_head_idx;
472 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
478 virtqueue_enqueue_recv_refill_packed(struct virtqueue *vq,
479 struct rte_mbuf **cookie, uint16_t num)
481 struct vring_packed_desc *start_dp = vq->vq_packed.ring.desc;
482 uint16_t flags = vq->vq_packed.cached_flags;
483 struct virtio_hw *hw = vq->hw;
484 struct vq_desc_extra *dxp;
488 if (unlikely(vq->vq_free_cnt == 0))
490 if (unlikely(vq->vq_free_cnt < num))
493 for (i = 0; i < num; i++) {
494 idx = vq->vq_avail_idx;
495 dxp = &vq->vq_descx[idx];
496 dxp->cookie = (void *)cookie[i];
499 start_dp[idx].addr = VIRTIO_MBUF_ADDR(cookie[i], vq) +
500 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
501 start_dp[idx].len = cookie[i]->buf_len - RTE_PKTMBUF_HEADROOM
502 + hw->vtnet_hdr_size;
504 vq->vq_desc_head_idx = dxp->next;
505 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
506 vq->vq_desc_tail_idx = vq->vq_desc_head_idx;
508 virtqueue_store_flags_packed(&start_dp[idx], flags,
511 if (++vq->vq_avail_idx >= vq->vq_nentries) {
512 vq->vq_avail_idx -= vq->vq_nentries;
513 vq->vq_packed.cached_flags ^=
514 VRING_PACKED_DESC_F_AVAIL_USED;
515 flags = vq->vq_packed.cached_flags;
518 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
522 /* When doing TSO, the IP length is not included in the pseudo header
523 * checksum of the packet given to the PMD, but for virtio it is
527 virtio_tso_fix_cksum(struct rte_mbuf *m)
529 /* common case: header is not fragmented */
530 if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
532 struct rte_ipv4_hdr *iph;
533 struct rte_ipv6_hdr *ip6h;
534 struct rte_tcp_hdr *th;
535 uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
538 iph = rte_pktmbuf_mtod_offset(m,
539 struct rte_ipv4_hdr *, m->l2_len);
540 th = RTE_PTR_ADD(iph, m->l3_len);
541 if ((iph->version_ihl >> 4) == 4) {
542 iph->hdr_checksum = 0;
543 iph->hdr_checksum = rte_ipv4_cksum(iph);
544 ip_len = iph->total_length;
545 ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
548 ip6h = (struct rte_ipv6_hdr *)iph;
549 ip_paylen = ip6h->payload_len;
552 /* calculate the new phdr checksum not including ip_paylen */
553 prev_cksum = th->cksum;
556 tmp = (tmp & 0xffff) + (tmp >> 16);
559 /* replace it in the packet */
560 th->cksum = new_cksum;
565 /* avoid write operation when necessary, to lessen cache issues */
566 #define ASSIGN_UNLESS_EQUAL(var, val) do { \
567 if ((var) != (val)) \
571 #define virtqueue_clear_net_hdr(_hdr) do { \
572 ASSIGN_UNLESS_EQUAL((_hdr)->csum_start, 0); \
573 ASSIGN_UNLESS_EQUAL((_hdr)->csum_offset, 0); \
574 ASSIGN_UNLESS_EQUAL((_hdr)->flags, 0); \
575 ASSIGN_UNLESS_EQUAL((_hdr)->gso_type, 0); \
576 ASSIGN_UNLESS_EQUAL((_hdr)->gso_size, 0); \
577 ASSIGN_UNLESS_EQUAL((_hdr)->hdr_len, 0); \
581 virtqueue_xmit_offload(struct virtio_net_hdr *hdr,
582 struct rte_mbuf *cookie,
586 if (cookie->ol_flags & PKT_TX_TCP_SEG)
587 cookie->ol_flags |= PKT_TX_TCP_CKSUM;
589 switch (cookie->ol_flags & PKT_TX_L4_MASK) {
590 case PKT_TX_UDP_CKSUM:
591 hdr->csum_start = cookie->l2_len + cookie->l3_len;
592 hdr->csum_offset = offsetof(struct rte_udp_hdr,
594 hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
597 case PKT_TX_TCP_CKSUM:
598 hdr->csum_start = cookie->l2_len + cookie->l3_len;
599 hdr->csum_offset = offsetof(struct rte_tcp_hdr, cksum);
600 hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
604 ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
605 ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
606 ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
610 /* TCP Segmentation Offload */
611 if (cookie->ol_flags & PKT_TX_TCP_SEG) {
612 hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
613 VIRTIO_NET_HDR_GSO_TCPV6 :
614 VIRTIO_NET_HDR_GSO_TCPV4;
615 hdr->gso_size = cookie->tso_segsz;
621 ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
622 ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
623 ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
629 virtqueue_enqueue_xmit_inorder(struct virtnet_tx *txvq,
630 struct rte_mbuf **cookies,
633 struct vq_desc_extra *dxp;
634 struct virtqueue *vq = txvq->vq;
635 struct vring_desc *start_dp;
636 struct virtio_net_hdr *hdr;
638 int16_t head_size = vq->hw->vtnet_hdr_size;
641 idx = vq->vq_desc_head_idx;
642 start_dp = vq->vq_split.ring.desc;
645 idx = idx & (vq->vq_nentries - 1);
646 dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
647 dxp->cookie = (void *)cookies[i];
649 virtio_update_packet_stats(&txvq->stats, cookies[i]);
651 hdr = rte_pktmbuf_mtod_offset(cookies[i],
652 struct virtio_net_hdr *, -head_size);
654 /* if offload disabled, hdr is not zeroed yet, do it now */
655 if (!vq->hw->has_tx_offload)
656 virtqueue_clear_net_hdr(hdr);
658 virtqueue_xmit_offload(hdr, cookies[i], true);
661 VIRTIO_MBUF_DATA_DMA_ADDR(cookies[i], vq) - head_size;
662 start_dp[idx].len = cookies[i]->data_len + head_size;
663 start_dp[idx].flags = 0;
666 vq_update_avail_ring(vq, idx);
672 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
673 vq->vq_desc_head_idx = idx & (vq->vq_nentries - 1);
677 virtqueue_enqueue_xmit_packed_fast(struct virtnet_tx *txvq,
678 struct rte_mbuf *cookie,
681 struct virtqueue *vq = txvq->vq;
682 struct vring_packed_desc *dp;
683 struct vq_desc_extra *dxp;
684 uint16_t idx, id, flags;
685 int16_t head_size = vq->hw->vtnet_hdr_size;
686 struct virtio_net_hdr *hdr;
688 id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx;
689 idx = vq->vq_avail_idx;
690 dp = &vq->vq_packed.ring.desc[idx];
692 dxp = &vq->vq_descx[id];
694 dxp->cookie = cookie;
696 flags = vq->vq_packed.cached_flags;
698 /* prepend cannot fail, checked by caller */
699 hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
702 /* if offload disabled, hdr is not zeroed yet, do it now */
703 if (!vq->hw->has_tx_offload)
704 virtqueue_clear_net_hdr(hdr);
706 virtqueue_xmit_offload(hdr, cookie, true);
708 dp->addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq) - head_size;
709 dp->len = cookie->data_len + head_size;
712 if (++vq->vq_avail_idx >= vq->vq_nentries) {
713 vq->vq_avail_idx -= vq->vq_nentries;
714 vq->vq_packed.cached_flags ^= VRING_PACKED_DESC_F_AVAIL_USED;
720 vq->vq_desc_head_idx = dxp->next;
721 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
722 vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
725 virtqueue_store_flags_packed(dp, flags, vq->hw->weak_barriers);
729 virtqueue_enqueue_xmit_packed(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
730 uint16_t needed, int can_push, int in_order)
732 struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
733 struct vq_desc_extra *dxp;
734 struct virtqueue *vq = txvq->vq;
735 struct vring_packed_desc *start_dp, *head_dp;
736 uint16_t idx, id, head_idx, head_flags;
737 int16_t head_size = vq->hw->vtnet_hdr_size;
738 struct virtio_net_hdr *hdr;
740 bool prepend_header = false;
742 id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx;
744 dxp = &vq->vq_descx[id];
745 dxp->ndescs = needed;
746 dxp->cookie = cookie;
748 head_idx = vq->vq_avail_idx;
751 start_dp = vq->vq_packed.ring.desc;
753 head_dp = &vq->vq_packed.ring.desc[idx];
754 head_flags = cookie->next ? VRING_DESC_F_NEXT : 0;
755 head_flags |= vq->vq_packed.cached_flags;
758 /* prepend cannot fail, checked by caller */
759 hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
761 prepend_header = true;
763 /* if offload disabled, it is not zeroed below, do it now */
764 if (!vq->hw->has_tx_offload)
765 virtqueue_clear_net_hdr(hdr);
767 /* setup first tx ring slot to point to header
768 * stored in reserved region.
770 start_dp[idx].addr = txvq->virtio_net_hdr_mem +
771 RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
772 start_dp[idx].len = vq->hw->vtnet_hdr_size;
773 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
775 if (idx >= vq->vq_nentries) {
776 idx -= vq->vq_nentries;
777 vq->vq_packed.cached_flags ^=
778 VRING_PACKED_DESC_F_AVAIL_USED;
782 virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload);
787 start_dp[idx].addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
788 start_dp[idx].len = cookie->data_len;
789 if (prepend_header) {
790 start_dp[idx].addr -= head_size;
791 start_dp[idx].len += head_size;
792 prepend_header = false;
795 if (likely(idx != head_idx)) {
796 flags = cookie->next ? VRING_DESC_F_NEXT : 0;
797 flags |= vq->vq_packed.cached_flags;
798 start_dp[idx].flags = flags;
802 if (idx >= vq->vq_nentries) {
803 idx -= vq->vq_nentries;
804 vq->vq_packed.cached_flags ^=
805 VRING_PACKED_DESC_F_AVAIL_USED;
807 } while ((cookie = cookie->next) != NULL);
809 start_dp[prev].id = id;
811 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
812 vq->vq_avail_idx = idx;
815 vq->vq_desc_head_idx = dxp->next;
816 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
817 vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
820 virtqueue_store_flags_packed(head_dp, head_flags,
821 vq->hw->weak_barriers);
825 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
826 uint16_t needed, int use_indirect, int can_push,
829 struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
830 struct vq_desc_extra *dxp;
831 struct virtqueue *vq = txvq->vq;
832 struct vring_desc *start_dp;
833 uint16_t seg_num = cookie->nb_segs;
834 uint16_t head_idx, idx;
835 int16_t head_size = vq->hw->vtnet_hdr_size;
836 bool prepend_header = false;
837 struct virtio_net_hdr *hdr;
839 head_idx = vq->vq_desc_head_idx;
842 dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
844 dxp = &vq->vq_descx[idx];
845 dxp->cookie = (void *)cookie;
846 dxp->ndescs = needed;
848 start_dp = vq->vq_split.ring.desc;
851 /* prepend cannot fail, checked by caller */
852 hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
854 prepend_header = true;
856 /* if offload disabled, it is not zeroed below, do it now */
857 if (!vq->hw->has_tx_offload)
858 virtqueue_clear_net_hdr(hdr);
859 } else if (use_indirect) {
860 /* setup tx ring slot to point to indirect
861 * descriptor list stored in reserved region.
863 * the first slot in indirect ring is already preset
864 * to point to the header in reserved region
866 start_dp[idx].addr = txvq->virtio_net_hdr_mem +
867 RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
868 start_dp[idx].len = (seg_num + 1) * sizeof(struct vring_desc);
869 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
870 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
872 /* loop below will fill in rest of the indirect elements */
873 start_dp = txr[idx].tx_indir;
876 /* setup first tx ring slot to point to header
877 * stored in reserved region.
879 start_dp[idx].addr = txvq->virtio_net_hdr_mem +
880 RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
881 start_dp[idx].len = vq->hw->vtnet_hdr_size;
882 start_dp[idx].flags = VRING_DESC_F_NEXT;
883 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
885 idx = start_dp[idx].next;
888 virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload);
891 start_dp[idx].addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
892 start_dp[idx].len = cookie->data_len;
893 if (prepend_header) {
894 start_dp[idx].addr -= head_size;
895 start_dp[idx].len += head_size;
896 prepend_header = false;
898 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
899 idx = start_dp[idx].next;
900 } while ((cookie = cookie->next) != NULL);
903 idx = vq->vq_split.ring.desc[head_idx].next;
905 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
907 vq->vq_desc_head_idx = idx;
908 vq_update_avail_ring(vq, head_idx);
911 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
912 vq->vq_desc_tail_idx = idx;
917 virtio_dev_cq_start(struct rte_eth_dev *dev)
919 struct virtio_hw *hw = dev->data->dev_private;
921 if (hw->cvq && hw->cvq->vq) {
922 rte_spinlock_init(&hw->cvq->lock);
923 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
928 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
931 unsigned int socket_id __rte_unused,
932 const struct rte_eth_rxconf *rx_conf,
933 struct rte_mempool *mp)
935 uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
936 struct virtio_hw *hw = dev->data->dev_private;
937 struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
938 struct virtnet_rx *rxvq;
939 uint16_t rx_free_thresh;
941 PMD_INIT_FUNC_TRACE();
943 if (rx_conf->rx_deferred_start) {
944 PMD_INIT_LOG(ERR, "Rx deferred start is not supported");
948 rx_free_thresh = rx_conf->rx_free_thresh;
949 if (rx_free_thresh == 0)
951 RTE_MIN(vq->vq_nentries / 4, DEFAULT_RX_FREE_THRESH);
953 if (rx_free_thresh & 0x3) {
954 RTE_LOG(ERR, PMD, "rx_free_thresh must be multiples of four."
955 " (rx_free_thresh=%u port=%u queue=%u)\n",
956 rx_free_thresh, dev->data->port_id, queue_idx);
960 if (rx_free_thresh >= vq->vq_nentries) {
961 RTE_LOG(ERR, PMD, "rx_free_thresh must be less than the "
962 "number of RX entries (%u)."
963 " (rx_free_thresh=%u port=%u queue=%u)\n",
965 rx_free_thresh, dev->data->port_id, queue_idx);
968 vq->vq_free_thresh = rx_free_thresh;
970 if (nb_desc == 0 || nb_desc > vq->vq_nentries)
971 nb_desc = vq->vq_nentries;
972 vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
975 rxvq->queue_id = queue_idx;
977 dev->data->rx_queues[queue_idx] = rxvq;
983 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
985 uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
986 struct virtio_hw *hw = dev->data->dev_private;
987 struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
988 struct virtnet_rx *rxvq = &vq->rxq;
993 PMD_INIT_FUNC_TRACE();
995 /* Allocate blank mbufs for the each rx descriptor */
998 if (hw->use_simple_rx) {
999 for (desc_idx = 0; desc_idx < vq->vq_nentries;
1001 vq->vq_split.ring.avail->ring[desc_idx] = desc_idx;
1002 vq->vq_split.ring.desc[desc_idx].flags =
1006 virtio_rxq_vec_setup(rxvq);
1009 memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
1010 for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
1012 vq->sw_ring[vq->vq_nentries + desc_idx] =
1016 if (hw->use_simple_rx) {
1017 while (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
1018 virtio_rxq_rearm_vec(rxvq);
1019 nbufs += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
1021 } else if (hw->use_inorder_rx) {
1022 if ((!virtqueue_full(vq))) {
1023 uint16_t free_cnt = vq->vq_free_cnt;
1024 struct rte_mbuf *pkts[free_cnt];
1026 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, pkts,
1028 error = virtqueue_enqueue_refill_inorder(vq,
1031 if (unlikely(error)) {
1032 for (i = 0; i < free_cnt; i++)
1033 rte_pktmbuf_free(pkts[i]);
1038 vq_update_avail_idx(vq);
1041 while (!virtqueue_full(vq)) {
1042 m = rte_mbuf_raw_alloc(rxvq->mpool);
1046 /* Enqueue allocated buffers */
1047 if (vtpci_packed_queue(vq->hw))
1048 error = virtqueue_enqueue_recv_refill_packed(vq,
1051 error = virtqueue_enqueue_recv_refill(vq,
1054 rte_pktmbuf_free(m);
1060 if (!vtpci_packed_queue(vq->hw))
1061 vq_update_avail_idx(vq);
1064 PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
1072 * struct rte_eth_dev *dev: Used to update dev
1073 * uint16_t nb_desc: Defaults to values read from config space
1074 * unsigned int socket_id: Used to allocate memzone
1075 * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
1076 * uint16_t queue_idx: Just used as an index in dev txq list
1079 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
1082 unsigned int socket_id __rte_unused,
1083 const struct rte_eth_txconf *tx_conf)
1085 uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
1086 struct virtio_hw *hw = dev->data->dev_private;
1087 struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
1088 struct virtnet_tx *txvq;
1089 uint16_t tx_free_thresh;
1091 PMD_INIT_FUNC_TRACE();
1093 if (tx_conf->tx_deferred_start) {
1094 PMD_INIT_LOG(ERR, "Tx deferred start is not supported");
1098 if (nb_desc == 0 || nb_desc > vq->vq_nentries)
1099 nb_desc = vq->vq_nentries;
1100 vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
1103 txvq->queue_id = queue_idx;
1105 tx_free_thresh = tx_conf->tx_free_thresh;
1106 if (tx_free_thresh == 0)
1108 RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
1110 if (tx_free_thresh >= (vq->vq_nentries - 3)) {
1111 PMD_DRV_LOG(ERR, "tx_free_thresh must be less than the "
1112 "number of TX entries minus 3 (%u)."
1113 " (tx_free_thresh=%u port=%u queue=%u)\n",
1114 vq->vq_nentries - 3,
1115 tx_free_thresh, dev->data->port_id, queue_idx);
1119 vq->vq_free_thresh = tx_free_thresh;
1121 dev->data->tx_queues[queue_idx] = txvq;
1126 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
1129 uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
1130 struct virtio_hw *hw = dev->data->dev_private;
1131 struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
1133 PMD_INIT_FUNC_TRACE();
1135 if (!vtpci_packed_queue(hw)) {
1136 if (hw->use_inorder_tx)
1137 vq->vq_split.ring.desc[vq->vq_nentries - 1].next = 0;
1146 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
1150 * Requeue the discarded mbuf. This should always be
1151 * successful since it was just dequeued.
1153 if (vtpci_packed_queue(vq->hw))
1154 error = virtqueue_enqueue_recv_refill_packed(vq, &m, 1);
1156 error = virtqueue_enqueue_recv_refill(vq, &m, 1);
1158 if (unlikely(error)) {
1159 PMD_DRV_LOG(ERR, "cannot requeue discarded mbuf");
1160 rte_pktmbuf_free(m);
1165 virtio_discard_rxbuf_inorder(struct virtqueue *vq, struct rte_mbuf *m)
1169 error = virtqueue_enqueue_refill_inorder(vq, &m, 1);
1170 if (unlikely(error)) {
1171 PMD_DRV_LOG(ERR, "cannot requeue discarded mbuf");
1172 rte_pktmbuf_free(m);
1176 /* Optionally fill offload information in structure */
1178 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
1180 struct rte_net_hdr_lens hdr_lens;
1181 uint32_t hdrlen, ptype;
1182 int l4_supported = 0;
1185 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1188 m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
1190 ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
1191 m->packet_type = ptype;
1192 if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
1193 (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
1194 (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
1197 if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1198 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
1199 if (hdr->csum_start <= hdrlen && l4_supported) {
1200 m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
1202 /* Unknown proto or tunnel, do sw cksum. We can assume
1203 * the cksum field is in the first segment since the
1204 * buffers we provided to the host are large enough.
1205 * In case of SCTP, this will be wrong since it's a CRC
1206 * but there's nothing we can do.
1208 uint16_t csum = 0, off;
1210 rte_raw_cksum_mbuf(m, hdr->csum_start,
1211 rte_pktmbuf_pkt_len(m) - hdr->csum_start,
1213 if (likely(csum != 0xffff))
1215 off = hdr->csum_offset + hdr->csum_start;
1216 if (rte_pktmbuf_data_len(m) >= off + 1)
1217 *rte_pktmbuf_mtod_offset(m, uint16_t *,
1220 } else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
1221 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
1224 /* GSO request, save required information in mbuf */
1225 if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1226 /* Check unsupported modes */
1227 if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
1228 (hdr->gso_size == 0)) {
1232 /* Update mss lengthes in mbuf */
1233 m->tso_segsz = hdr->gso_size;
1234 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1235 case VIRTIO_NET_HDR_GSO_TCPV4:
1236 case VIRTIO_NET_HDR_GSO_TCPV6:
1237 m->ol_flags |= PKT_RX_LRO | \
1238 PKT_RX_L4_CKSUM_NONE;
1248 #define VIRTIO_MBUF_BURST_SZ 64
1249 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
1251 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1253 struct virtnet_rx *rxvq = rx_queue;
1254 struct virtqueue *vq = rxvq->vq;
1255 struct virtio_hw *hw = vq->hw;
1256 struct rte_mbuf *rxm;
1257 uint16_t nb_used, num, nb_rx;
1258 uint32_t len[VIRTIO_MBUF_BURST_SZ];
1259 struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1261 uint32_t i, nb_enqueued;
1263 struct virtio_net_hdr *hdr;
1266 if (unlikely(hw->started == 0))
1269 nb_used = VIRTQUEUE_NUSED(vq);
1271 virtio_rmb(hw->weak_barriers);
1273 num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1274 if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1275 num = VIRTIO_MBUF_BURST_SZ;
1276 if (likely(num > DESC_PER_CACHELINE))
1277 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1279 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1280 PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
1283 hdr_size = hw->vtnet_hdr_size;
1285 for (i = 0; i < num ; i++) {
1288 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1290 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1291 PMD_RX_LOG(ERR, "Packet drop");
1293 virtio_discard_rxbuf(vq, rxm);
1294 rxvq->stats.errors++;
1298 rxm->port = rxvq->port_id;
1299 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1303 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1304 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1306 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1307 RTE_PKTMBUF_HEADROOM - hdr_size);
1310 rte_vlan_strip(rxm);
1312 if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1313 virtio_discard_rxbuf(vq, rxm);
1314 rxvq->stats.errors++;
1318 virtio_rx_stats_updated(rxvq, rxm);
1320 rx_pkts[nb_rx++] = rxm;
1323 rxvq->stats.packets += nb_rx;
1325 /* Allocate new mbuf for the used descriptor */
1326 if (likely(!virtqueue_full(vq))) {
1327 uint16_t free_cnt = vq->vq_free_cnt;
1328 struct rte_mbuf *new_pkts[free_cnt];
1330 if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1332 error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1334 if (unlikely(error)) {
1335 for (i = 0; i < free_cnt; i++)
1336 rte_pktmbuf_free(new_pkts[i]);
1338 nb_enqueued += free_cnt;
1340 struct rte_eth_dev *dev =
1341 &rte_eth_devices[rxvq->port_id];
1342 dev->data->rx_mbuf_alloc_failed += free_cnt;
1346 if (likely(nb_enqueued)) {
1347 vq_update_avail_idx(vq);
1349 if (unlikely(virtqueue_kick_prepare(vq))) {
1350 virtqueue_notify(vq);
1351 PMD_RX_LOG(DEBUG, "Notified");
1359 virtio_recv_pkts_packed(void *rx_queue, struct rte_mbuf **rx_pkts,
1362 struct virtnet_rx *rxvq = rx_queue;
1363 struct virtqueue *vq = rxvq->vq;
1364 struct virtio_hw *hw = vq->hw;
1365 struct rte_mbuf *rxm;
1366 uint16_t num, nb_rx;
1367 uint32_t len[VIRTIO_MBUF_BURST_SZ];
1368 struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1370 uint32_t i, nb_enqueued;
1372 struct virtio_net_hdr *hdr;
1375 if (unlikely(hw->started == 0))
1378 num = RTE_MIN(VIRTIO_MBUF_BURST_SZ, nb_pkts);
1379 if (likely(num > DESC_PER_CACHELINE))
1380 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1382 num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1383 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1386 hdr_size = hw->vtnet_hdr_size;
1388 for (i = 0; i < num; i++) {
1391 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1393 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1394 PMD_RX_LOG(ERR, "Packet drop");
1396 virtio_discard_rxbuf(vq, rxm);
1397 rxvq->stats.errors++;
1401 rxm->port = rxvq->port_id;
1402 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1406 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1407 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1409 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1410 RTE_PKTMBUF_HEADROOM - hdr_size);
1413 rte_vlan_strip(rxm);
1415 if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1416 virtio_discard_rxbuf(vq, rxm);
1417 rxvq->stats.errors++;
1421 virtio_rx_stats_updated(rxvq, rxm);
1423 rx_pkts[nb_rx++] = rxm;
1426 rxvq->stats.packets += nb_rx;
1428 /* Allocate new mbuf for the used descriptor */
1429 if (likely(!virtqueue_full(vq))) {
1430 uint16_t free_cnt = vq->vq_free_cnt;
1431 struct rte_mbuf *new_pkts[free_cnt];
1433 if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1435 error = virtqueue_enqueue_recv_refill_packed(vq,
1436 new_pkts, free_cnt);
1437 if (unlikely(error)) {
1438 for (i = 0; i < free_cnt; i++)
1439 rte_pktmbuf_free(new_pkts[i]);
1441 nb_enqueued += free_cnt;
1443 struct rte_eth_dev *dev =
1444 &rte_eth_devices[rxvq->port_id];
1445 dev->data->rx_mbuf_alloc_failed += free_cnt;
1449 if (likely(nb_enqueued)) {
1450 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1451 virtqueue_notify(vq);
1452 PMD_RX_LOG(DEBUG, "Notified");
1461 virtio_recv_pkts_inorder(void *rx_queue,
1462 struct rte_mbuf **rx_pkts,
1465 struct virtnet_rx *rxvq = rx_queue;
1466 struct virtqueue *vq = rxvq->vq;
1467 struct virtio_hw *hw = vq->hw;
1468 struct rte_mbuf *rxm;
1469 struct rte_mbuf *prev = NULL;
1470 uint16_t nb_used, num, nb_rx;
1471 uint32_t len[VIRTIO_MBUF_BURST_SZ];
1472 struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1474 uint32_t nb_enqueued;
1481 if (unlikely(hw->started == 0))
1484 nb_used = VIRTQUEUE_NUSED(vq);
1485 nb_used = RTE_MIN(nb_used, nb_pkts);
1486 nb_used = RTE_MIN(nb_used, VIRTIO_MBUF_BURST_SZ);
1488 virtio_rmb(hw->weak_barriers);
1490 PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1495 hdr_size = hw->vtnet_hdr_size;
1497 num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len, nb_used);
1499 for (i = 0; i < num; i++) {
1500 struct virtio_net_hdr_mrg_rxbuf *header;
1502 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1503 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1507 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1508 PMD_RX_LOG(ERR, "Packet drop");
1510 virtio_discard_rxbuf_inorder(vq, rxm);
1511 rxvq->stats.errors++;
1515 header = (struct virtio_net_hdr_mrg_rxbuf *)
1516 ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1519 if (vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
1520 seg_num = header->num_buffers;
1527 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1528 rxm->nb_segs = seg_num;
1531 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1532 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1534 rxm->port = rxvq->port_id;
1536 rx_pkts[nb_rx] = rxm;
1539 if (vq->hw->has_rx_offload &&
1540 virtio_rx_offload(rxm, &header->hdr) < 0) {
1541 virtio_discard_rxbuf_inorder(vq, rxm);
1542 rxvq->stats.errors++;
1547 rte_vlan_strip(rx_pkts[nb_rx]);
1549 seg_res = seg_num - 1;
1551 /* Merge remaining segments */
1552 while (seg_res != 0 && i < (num - 1)) {
1556 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1557 rxm->pkt_len = (uint32_t)(len[i]);
1558 rxm->data_len = (uint16_t)(len[i]);
1560 rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1568 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1573 /* Last packet still need merge segments */
1574 while (seg_res != 0) {
1575 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1576 VIRTIO_MBUF_BURST_SZ);
1578 if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
1579 virtio_rmb(hw->weak_barriers);
1580 num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len,
1582 uint16_t extra_idx = 0;
1585 while (extra_idx < rcv_cnt) {
1586 rxm = rcv_pkts[extra_idx];
1588 RTE_PKTMBUF_HEADROOM - hdr_size;
1589 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1590 rxm->data_len = (uint16_t)(len[extra_idx]);
1593 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1599 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1604 "No enough segments for packet.");
1605 rte_pktmbuf_free(rx_pkts[nb_rx]);
1606 rxvq->stats.errors++;
1611 rxvq->stats.packets += nb_rx;
1613 /* Allocate new mbuf for the used descriptor */
1615 if (likely(!virtqueue_full(vq))) {
1616 /* free_cnt may include mrg descs */
1617 uint16_t free_cnt = vq->vq_free_cnt;
1618 struct rte_mbuf *new_pkts[free_cnt];
1620 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1621 error = virtqueue_enqueue_refill_inorder(vq, new_pkts,
1623 if (unlikely(error)) {
1624 for (i = 0; i < free_cnt; i++)
1625 rte_pktmbuf_free(new_pkts[i]);
1627 nb_enqueued += free_cnt;
1629 struct rte_eth_dev *dev =
1630 &rte_eth_devices[rxvq->port_id];
1631 dev->data->rx_mbuf_alloc_failed += free_cnt;
1635 if (likely(nb_enqueued)) {
1636 vq_update_avail_idx(vq);
1638 if (unlikely(virtqueue_kick_prepare(vq))) {
1639 virtqueue_notify(vq);
1640 PMD_RX_LOG(DEBUG, "Notified");
1648 virtio_recv_mergeable_pkts(void *rx_queue,
1649 struct rte_mbuf **rx_pkts,
1652 struct virtnet_rx *rxvq = rx_queue;
1653 struct virtqueue *vq = rxvq->vq;
1654 struct virtio_hw *hw = vq->hw;
1655 struct rte_mbuf *rxm;
1656 struct rte_mbuf *prev = NULL;
1657 uint16_t nb_used, num, nb_rx = 0;
1658 uint32_t len[VIRTIO_MBUF_BURST_SZ];
1659 struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1661 uint32_t nb_enqueued = 0;
1662 uint32_t seg_num = 0;
1663 uint32_t seg_res = 0;
1664 uint32_t hdr_size = hw->vtnet_hdr_size;
1667 if (unlikely(hw->started == 0))
1670 nb_used = VIRTQUEUE_NUSED(vq);
1672 virtio_rmb(hw->weak_barriers);
1674 PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1676 num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1677 if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1678 num = VIRTIO_MBUF_BURST_SZ;
1679 if (likely(num > DESC_PER_CACHELINE))
1680 num = num - ((vq->vq_used_cons_idx + num) %
1681 DESC_PER_CACHELINE);
1684 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1686 for (i = 0; i < num; i++) {
1687 struct virtio_net_hdr_mrg_rxbuf *header;
1689 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1690 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1694 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1695 PMD_RX_LOG(ERR, "Packet drop");
1697 virtio_discard_rxbuf(vq, rxm);
1698 rxvq->stats.errors++;
1702 header = (struct virtio_net_hdr_mrg_rxbuf *)
1703 ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1705 seg_num = header->num_buffers;
1709 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1710 rxm->nb_segs = seg_num;
1713 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1714 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1716 rxm->port = rxvq->port_id;
1718 rx_pkts[nb_rx] = rxm;
1721 if (hw->has_rx_offload &&
1722 virtio_rx_offload(rxm, &header->hdr) < 0) {
1723 virtio_discard_rxbuf(vq, rxm);
1724 rxvq->stats.errors++;
1729 rte_vlan_strip(rx_pkts[nb_rx]);
1731 seg_res = seg_num - 1;
1733 /* Merge remaining segments */
1734 while (seg_res != 0 && i < (num - 1)) {
1738 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1739 rxm->pkt_len = (uint32_t)(len[i]);
1740 rxm->data_len = (uint16_t)(len[i]);
1742 rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1750 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1755 /* Last packet still need merge segments */
1756 while (seg_res != 0) {
1757 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1758 VIRTIO_MBUF_BURST_SZ);
1760 if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
1761 virtio_rmb(hw->weak_barriers);
1762 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len,
1764 uint16_t extra_idx = 0;
1767 while (extra_idx < rcv_cnt) {
1768 rxm = rcv_pkts[extra_idx];
1770 RTE_PKTMBUF_HEADROOM - hdr_size;
1771 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1772 rxm->data_len = (uint16_t)(len[extra_idx]);
1775 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1781 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1786 "No enough segments for packet.");
1787 rte_pktmbuf_free(rx_pkts[nb_rx]);
1788 rxvq->stats.errors++;
1793 rxvq->stats.packets += nb_rx;
1795 /* Allocate new mbuf for the used descriptor */
1796 if (likely(!virtqueue_full(vq))) {
1797 /* free_cnt may include mrg descs */
1798 uint16_t free_cnt = vq->vq_free_cnt;
1799 struct rte_mbuf *new_pkts[free_cnt];
1801 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1802 error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1804 if (unlikely(error)) {
1805 for (i = 0; i < free_cnt; i++)
1806 rte_pktmbuf_free(new_pkts[i]);
1808 nb_enqueued += free_cnt;
1810 struct rte_eth_dev *dev =
1811 &rte_eth_devices[rxvq->port_id];
1812 dev->data->rx_mbuf_alloc_failed += free_cnt;
1816 if (likely(nb_enqueued)) {
1817 vq_update_avail_idx(vq);
1819 if (unlikely(virtqueue_kick_prepare(vq))) {
1820 virtqueue_notify(vq);
1821 PMD_RX_LOG(DEBUG, "Notified");
1829 virtio_recv_mergeable_pkts_packed(void *rx_queue,
1830 struct rte_mbuf **rx_pkts,
1833 struct virtnet_rx *rxvq = rx_queue;
1834 struct virtqueue *vq = rxvq->vq;
1835 struct virtio_hw *hw = vq->hw;
1836 struct rte_mbuf *rxm;
1837 struct rte_mbuf *prev = NULL;
1838 uint16_t num, nb_rx = 0;
1839 uint32_t len[VIRTIO_MBUF_BURST_SZ];
1840 struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1841 uint32_t nb_enqueued = 0;
1842 uint32_t seg_num = 0;
1843 uint32_t seg_res = 0;
1844 uint32_t hdr_size = hw->vtnet_hdr_size;
1848 if (unlikely(hw->started == 0))
1853 if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1854 num = VIRTIO_MBUF_BURST_SZ;
1855 if (likely(num > DESC_PER_CACHELINE))
1856 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1858 num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1860 for (i = 0; i < num; i++) {
1861 struct virtio_net_hdr_mrg_rxbuf *header;
1863 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1864 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1868 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1869 PMD_RX_LOG(ERR, "Packet drop");
1871 virtio_discard_rxbuf(vq, rxm);
1872 rxvq->stats.errors++;
1876 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)
1877 rxm->buf_addr + RTE_PKTMBUF_HEADROOM - hdr_size);
1878 seg_num = header->num_buffers;
1883 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1884 rxm->nb_segs = seg_num;
1887 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1888 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1890 rxm->port = rxvq->port_id;
1891 rx_pkts[nb_rx] = rxm;
1894 if (hw->has_rx_offload &&
1895 virtio_rx_offload(rxm, &header->hdr) < 0) {
1896 virtio_discard_rxbuf(vq, rxm);
1897 rxvq->stats.errors++;
1902 rte_vlan_strip(rx_pkts[nb_rx]);
1904 seg_res = seg_num - 1;
1906 /* Merge remaining segments */
1907 while (seg_res != 0 && i < (num - 1)) {
1911 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1912 rxm->pkt_len = (uint32_t)(len[i]);
1913 rxm->data_len = (uint16_t)(len[i]);
1915 rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1923 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1928 /* Last packet still need merge segments */
1929 while (seg_res != 0) {
1930 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1931 VIRTIO_MBUF_BURST_SZ);
1932 uint16_t extra_idx = 0;
1934 rcv_cnt = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts,
1936 if (unlikely(rcv_cnt == 0)) {
1937 PMD_RX_LOG(ERR, "No enough segments for packet.");
1938 rte_pktmbuf_free(rx_pkts[nb_rx]);
1939 rxvq->stats.errors++;
1943 while (extra_idx < rcv_cnt) {
1944 rxm = rcv_pkts[extra_idx];
1946 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1947 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1948 rxm->data_len = (uint16_t)(len[extra_idx]);
1952 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1957 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1962 rxvq->stats.packets += nb_rx;
1964 /* Allocate new mbuf for the used descriptor */
1965 if (likely(!virtqueue_full(vq))) {
1966 /* free_cnt may include mrg descs */
1967 uint16_t free_cnt = vq->vq_free_cnt;
1968 struct rte_mbuf *new_pkts[free_cnt];
1970 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1971 error = virtqueue_enqueue_recv_refill_packed(vq,
1972 new_pkts, free_cnt);
1973 if (unlikely(error)) {
1974 for (i = 0; i < free_cnt; i++)
1975 rte_pktmbuf_free(new_pkts[i]);
1977 nb_enqueued += free_cnt;
1979 struct rte_eth_dev *dev =
1980 &rte_eth_devices[rxvq->port_id];
1981 dev->data->rx_mbuf_alloc_failed += free_cnt;
1985 if (likely(nb_enqueued)) {
1986 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1987 virtqueue_notify(vq);
1988 PMD_RX_LOG(DEBUG, "Notified");
1996 virtio_xmit_pkts_prepare(void *tx_queue __rte_unused, struct rte_mbuf **tx_pkts,
2002 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
2003 struct rte_mbuf *m = tx_pkts[nb_tx];
2005 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
2006 error = rte_validate_tx_offload(m);
2007 if (unlikely(error)) {
2013 /* Do VLAN tag insertion */
2014 if (unlikely(m->ol_flags & PKT_TX_VLAN_PKT)) {
2015 error = rte_vlan_insert(&m);
2016 /* rte_vlan_insert() may change pointer
2017 * even in the case of failure
2021 if (unlikely(error)) {
2027 error = rte_net_intel_cksum_prepare(m);
2028 if (unlikely(error)) {
2033 if (m->ol_flags & PKT_TX_TCP_SEG)
2034 virtio_tso_fix_cksum(m);
2041 virtio_xmit_pkts_packed(void *tx_queue, struct rte_mbuf **tx_pkts,
2044 struct virtnet_tx *txvq = tx_queue;
2045 struct virtqueue *vq = txvq->vq;
2046 struct virtio_hw *hw = vq->hw;
2047 uint16_t hdr_size = hw->vtnet_hdr_size;
2049 bool in_order = hw->use_inorder_tx;
2051 if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
2054 if (unlikely(nb_pkts < 1))
2057 PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
2059 if (nb_pkts > vq->vq_free_cnt)
2060 virtio_xmit_cleanup_packed(vq, nb_pkts - vq->vq_free_cnt,
2063 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
2064 struct rte_mbuf *txm = tx_pkts[nb_tx];
2065 int can_push = 0, slots, need;
2067 /* optimize ring usage */
2068 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
2069 vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
2070 rte_mbuf_refcnt_read(txm) == 1 &&
2071 RTE_MBUF_DIRECT(txm) &&
2072 txm->nb_segs == 1 &&
2073 rte_pktmbuf_headroom(txm) >= hdr_size &&
2074 rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
2075 __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
2078 /* How many main ring entries are needed to this Tx?
2079 * any_layout => number of segments
2080 * default => number of segments + 1
2082 slots = txm->nb_segs + !can_push;
2083 need = slots - vq->vq_free_cnt;
2085 /* Positive value indicates it need free vring descriptors */
2086 if (unlikely(need > 0)) {
2087 virtio_xmit_cleanup_packed(vq, need, in_order);
2088 need = slots - vq->vq_free_cnt;
2089 if (unlikely(need > 0)) {
2091 "No free tx descriptors to transmit");
2096 /* Enqueue Packet buffers */
2098 virtqueue_enqueue_xmit_packed_fast(txvq, txm, in_order);
2100 virtqueue_enqueue_xmit_packed(txvq, txm, slots, 0,
2103 virtio_update_packet_stats(&txvq->stats, txm);
2106 txvq->stats.packets += nb_tx;
2108 if (likely(nb_tx)) {
2109 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
2110 virtqueue_notify(vq);
2111 PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2119 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
2121 struct virtnet_tx *txvq = tx_queue;
2122 struct virtqueue *vq = txvq->vq;
2123 struct virtio_hw *hw = vq->hw;
2124 uint16_t hdr_size = hw->vtnet_hdr_size;
2125 uint16_t nb_used, nb_tx = 0;
2127 if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
2130 if (unlikely(nb_pkts < 1))
2133 PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
2134 nb_used = VIRTQUEUE_NUSED(vq);
2136 virtio_rmb(hw->weak_barriers);
2137 if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
2138 virtio_xmit_cleanup(vq, nb_used);
2140 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
2141 struct rte_mbuf *txm = tx_pkts[nb_tx];
2142 int can_push = 0, use_indirect = 0, slots, need;
2144 /* optimize ring usage */
2145 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
2146 vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
2147 rte_mbuf_refcnt_read(txm) == 1 &&
2148 RTE_MBUF_DIRECT(txm) &&
2149 txm->nb_segs == 1 &&
2150 rte_pktmbuf_headroom(txm) >= hdr_size &&
2151 rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
2152 __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
2154 else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
2155 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
2158 /* How many main ring entries are needed to this Tx?
2159 * any_layout => number of segments
2161 * default => number of segments + 1
2163 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
2164 need = slots - vq->vq_free_cnt;
2166 /* Positive value indicates it need free vring descriptors */
2167 if (unlikely(need > 0)) {
2168 nb_used = VIRTQUEUE_NUSED(vq);
2169 virtio_rmb(hw->weak_barriers);
2170 need = RTE_MIN(need, (int)nb_used);
2172 virtio_xmit_cleanup(vq, need);
2173 need = slots - vq->vq_free_cnt;
2174 if (unlikely(need > 0)) {
2176 "No free tx descriptors to transmit");
2181 /* Enqueue Packet buffers */
2182 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
2185 virtio_update_packet_stats(&txvq->stats, txm);
2188 txvq->stats.packets += nb_tx;
2190 if (likely(nb_tx)) {
2191 vq_update_avail_idx(vq);
2193 if (unlikely(virtqueue_kick_prepare(vq))) {
2194 virtqueue_notify(vq);
2195 PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2202 static __rte_always_inline int
2203 virtio_xmit_try_cleanup_inorder(struct virtqueue *vq, uint16_t need)
2205 uint16_t nb_used, nb_clean, nb_descs;
2206 struct virtio_hw *hw = vq->hw;
2208 nb_descs = vq->vq_free_cnt + need;
2209 nb_used = VIRTQUEUE_NUSED(vq);
2210 virtio_rmb(hw->weak_barriers);
2211 nb_clean = RTE_MIN(need, (int)nb_used);
2213 virtio_xmit_cleanup_inorder(vq, nb_clean);
2215 return nb_descs - vq->vq_free_cnt;
2219 virtio_xmit_pkts_inorder(void *tx_queue,
2220 struct rte_mbuf **tx_pkts,
2223 struct virtnet_tx *txvq = tx_queue;
2224 struct virtqueue *vq = txvq->vq;
2225 struct virtio_hw *hw = vq->hw;
2226 uint16_t hdr_size = hw->vtnet_hdr_size;
2227 uint16_t nb_used, nb_tx = 0, nb_inorder_pkts = 0;
2228 struct rte_mbuf *inorder_pkts[nb_pkts];
2231 if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
2234 if (unlikely(nb_pkts < 1))
2238 PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
2239 nb_used = VIRTQUEUE_NUSED(vq);
2241 virtio_rmb(hw->weak_barriers);
2242 if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
2243 virtio_xmit_cleanup_inorder(vq, nb_used);
2245 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
2246 struct rte_mbuf *txm = tx_pkts[nb_tx];
2249 /* optimize ring usage */
2250 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
2251 vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
2252 rte_mbuf_refcnt_read(txm) == 1 &&
2253 RTE_MBUF_DIRECT(txm) &&
2254 txm->nb_segs == 1 &&
2255 rte_pktmbuf_headroom(txm) >= hdr_size &&
2256 rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
2257 __alignof__(struct virtio_net_hdr_mrg_rxbuf))) {
2258 inorder_pkts[nb_inorder_pkts] = txm;
2264 if (nb_inorder_pkts) {
2265 need = nb_inorder_pkts - vq->vq_free_cnt;
2266 if (unlikely(need > 0)) {
2267 need = virtio_xmit_try_cleanup_inorder(vq,
2269 if (unlikely(need > 0)) {
2271 "No free tx descriptors to "
2276 virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2278 nb_inorder_pkts = 0;
2281 slots = txm->nb_segs + 1;
2282 need = slots - vq->vq_free_cnt;
2283 if (unlikely(need > 0)) {
2284 need = virtio_xmit_try_cleanup_inorder(vq, slots);
2286 if (unlikely(need > 0)) {
2288 "No free tx descriptors to transmit");
2292 /* Enqueue Packet buffers */
2293 virtqueue_enqueue_xmit(txvq, txm, slots, 0, 0, 1);
2295 virtio_update_packet_stats(&txvq->stats, txm);
2298 /* Transmit all inorder packets */
2299 if (nb_inorder_pkts) {
2300 need = nb_inorder_pkts - vq->vq_free_cnt;
2301 if (unlikely(need > 0)) {
2302 need = virtio_xmit_try_cleanup_inorder(vq,
2304 if (unlikely(need > 0)) {
2306 "No free tx descriptors to transmit");
2307 nb_inorder_pkts = vq->vq_free_cnt;
2312 virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2316 txvq->stats.packets += nb_tx;
2318 if (likely(nb_tx)) {
2319 vq_update_avail_idx(vq);
2321 if (unlikely(virtqueue_kick_prepare(vq))) {
2322 virtqueue_notify(vq);
2323 PMD_TX_LOG(DEBUG, "Notified backend after xmit");