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 <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"
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 = virtnet_rxq_to_vq(rxvq);
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 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
95 uint32_t s = mbuf->pkt_len;
96 struct rte_ether_addr *ea;
101 stats->size_bins[1]++;
102 } else if (s > 64 && s < 1024) {
105 /* count zeros, and offset into correct bin */
106 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
107 stats->size_bins[bin]++;
110 stats->size_bins[0]++;
112 stats->size_bins[6]++;
114 stats->size_bins[7]++;
117 ea = rte_pktmbuf_mtod(mbuf, struct rte_ether_addr *);
118 if (rte_is_multicast_ether_addr(ea)) {
119 if (rte_is_broadcast_ether_addr(ea))
127 virtio_rx_stats_updated(struct virtnet_rx *rxvq, struct rte_mbuf *m)
129 VIRTIO_DUMP_PACKET(m, m->data_len);
131 virtio_update_packet_stats(&rxvq->stats, m);
135 virtqueue_dequeue_burst_rx_packed(struct virtqueue *vq,
136 struct rte_mbuf **rx_pkts,
140 struct rte_mbuf *cookie;
143 struct vring_packed_desc *desc;
146 desc = vq->vq_packed.ring.desc;
148 for (i = 0; i < num; i++) {
149 used_idx = vq->vq_used_cons_idx;
150 /* desc_is_used has a load-acquire or rte_io_rmb inside
151 * and wait for used desc in virtqueue.
153 if (!desc_is_used(&desc[used_idx], vq))
155 len[i] = desc[used_idx].len;
156 id = desc[used_idx].id;
157 cookie = (struct rte_mbuf *)vq->vq_descx[id].cookie;
158 if (unlikely(cookie == NULL)) {
159 PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
160 vq->vq_used_cons_idx);
163 rte_prefetch0(cookie);
164 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
168 vq->vq_used_cons_idx++;
169 if (vq->vq_used_cons_idx >= vq->vq_nentries) {
170 vq->vq_used_cons_idx -= vq->vq_nentries;
171 vq->vq_packed.used_wrap_counter ^= 1;
179 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
180 uint32_t *len, uint16_t num)
182 struct vring_used_elem *uep;
183 struct rte_mbuf *cookie;
184 uint16_t used_idx, desc_idx;
187 /* Caller does the check */
188 for (i = 0; i < num ; i++) {
189 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
190 uep = &vq->vq_split.ring.used->ring[used_idx];
191 desc_idx = (uint16_t) uep->id;
193 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
195 if (unlikely(cookie == NULL)) {
196 PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
197 vq->vq_used_cons_idx);
201 rte_prefetch0(cookie);
202 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
204 vq->vq_used_cons_idx++;
205 vq_ring_free_chain(vq, desc_idx);
206 vq->vq_descx[desc_idx].cookie = NULL;
213 virtqueue_dequeue_rx_inorder(struct virtqueue *vq,
214 struct rte_mbuf **rx_pkts,
218 struct vring_used_elem *uep;
219 struct rte_mbuf *cookie;
220 uint16_t used_idx = 0;
223 if (unlikely(num == 0))
226 for (i = 0; i < num; i++) {
227 used_idx = vq->vq_used_cons_idx & (vq->vq_nentries - 1);
228 /* Desc idx same as used idx */
229 uep = &vq->vq_split.ring.used->ring[used_idx];
231 cookie = (struct rte_mbuf *)vq->vq_descx[used_idx].cookie;
233 if (unlikely(cookie == NULL)) {
234 PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
235 vq->vq_used_cons_idx);
239 rte_prefetch0(cookie);
240 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
242 vq->vq_used_cons_idx++;
243 vq->vq_descx[used_idx].cookie = NULL;
246 vq_ring_free_inorder(vq, used_idx, i);
251 virtqueue_enqueue_refill_inorder(struct virtqueue *vq,
252 struct rte_mbuf **cookies,
255 struct vq_desc_extra *dxp;
256 struct virtio_hw *hw = vq->hw;
257 struct vring_desc *start_dp;
258 uint16_t head_idx, idx, i = 0;
260 if (unlikely(vq->vq_free_cnt == 0))
262 if (unlikely(vq->vq_free_cnt < num))
265 head_idx = vq->vq_desc_head_idx & (vq->vq_nentries - 1);
266 start_dp = vq->vq_split.ring.desc;
269 idx = head_idx & (vq->vq_nentries - 1);
270 dxp = &vq->vq_descx[idx];
271 dxp->cookie = (void *)cookies[i];
274 start_dp[idx].addr = cookies[i]->buf_iova +
275 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
276 start_dp[idx].len = cookies[i]->buf_len -
277 RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
278 start_dp[idx].flags = VRING_DESC_F_WRITE;
280 vq_update_avail_ring(vq, idx);
285 vq->vq_desc_head_idx += num;
286 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
291 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf **cookie,
294 struct vq_desc_extra *dxp;
295 struct virtio_hw *hw = vq->hw;
296 struct vring_desc *start_dp = vq->vq_split.ring.desc;
299 if (unlikely(vq->vq_free_cnt == 0))
301 if (unlikely(vq->vq_free_cnt < num))
304 if (unlikely(vq->vq_desc_head_idx >= vq->vq_nentries))
307 for (i = 0; i < num; i++) {
308 idx = vq->vq_desc_head_idx;
309 dxp = &vq->vq_descx[idx];
310 dxp->cookie = (void *)cookie[i];
313 start_dp[idx].addr = cookie[i]->buf_iova +
314 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
315 start_dp[idx].len = cookie[i]->buf_len -
316 RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
317 start_dp[idx].flags = VRING_DESC_F_WRITE;
318 vq->vq_desc_head_idx = start_dp[idx].next;
319 vq_update_avail_ring(vq, idx);
320 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END) {
321 vq->vq_desc_tail_idx = vq->vq_desc_head_idx;
326 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
332 virtqueue_refill_single_packed(struct virtqueue *vq,
333 struct vring_packed_desc *dp,
334 struct rte_mbuf *cookie)
336 uint16_t flags = vq->vq_packed.cached_flags;
337 struct virtio_hw *hw = vq->hw;
339 dp->addr = cookie->buf_iova +
340 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
341 dp->len = cookie->buf_len -
342 RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
344 virtqueue_store_flags_packed(dp, flags,
347 if (++vq->vq_avail_idx >= vq->vq_nentries) {
348 vq->vq_avail_idx -= vq->vq_nentries;
349 vq->vq_packed.cached_flags ^=
350 VRING_PACKED_DESC_F_AVAIL_USED;
351 flags = vq->vq_packed.cached_flags;
356 virtqueue_enqueue_recv_refill_packed_init(struct virtqueue *vq,
357 struct rte_mbuf **cookie, uint16_t num)
359 struct vring_packed_desc *start_dp = vq->vq_packed.ring.desc;
360 struct vq_desc_extra *dxp;
364 if (unlikely(vq->vq_free_cnt == 0))
366 if (unlikely(vq->vq_free_cnt < num))
369 for (i = 0; i < num; i++) {
370 idx = vq->vq_avail_idx;
371 dxp = &vq->vq_descx[idx];
372 dxp->cookie = (void *)cookie[i];
375 virtqueue_refill_single_packed(vq, &start_dp[idx], cookie[i]);
377 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
382 virtqueue_enqueue_recv_refill_packed(struct virtqueue *vq,
383 struct rte_mbuf **cookie, uint16_t num)
385 struct vring_packed_desc *start_dp = vq->vq_packed.ring.desc;
386 struct vq_desc_extra *dxp;
390 if (unlikely(vq->vq_free_cnt == 0))
392 if (unlikely(vq->vq_free_cnt < num))
395 for (i = 0; i < num; i++) {
396 idx = vq->vq_avail_idx;
397 did = start_dp[idx].id;
398 dxp = &vq->vq_descx[did];
399 dxp->cookie = (void *)cookie[i];
402 virtqueue_refill_single_packed(vq, &start_dp[idx], cookie[i]);
404 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
408 /* When doing TSO, the IP length is not included in the pseudo header
409 * checksum of the packet given to the PMD, but for virtio it is
413 virtio_tso_fix_cksum(struct rte_mbuf *m)
415 /* common case: header is not fragmented */
416 if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
418 struct rte_ipv4_hdr *iph;
419 struct rte_ipv6_hdr *ip6h;
420 struct rte_tcp_hdr *th;
421 uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
424 iph = rte_pktmbuf_mtod_offset(m,
425 struct rte_ipv4_hdr *, m->l2_len);
426 th = RTE_PTR_ADD(iph, m->l3_len);
427 if ((iph->version_ihl >> 4) == 4) {
428 iph->hdr_checksum = 0;
429 iph->hdr_checksum = rte_ipv4_cksum(iph);
430 ip_len = iph->total_length;
431 ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
434 ip6h = (struct rte_ipv6_hdr *)iph;
435 ip_paylen = ip6h->payload_len;
438 /* calculate the new phdr checksum not including ip_paylen */
439 prev_cksum = th->cksum;
442 tmp = (tmp & 0xffff) + (tmp >> 16);
445 /* replace it in the packet */
446 th->cksum = new_cksum;
454 virtqueue_enqueue_xmit_inorder(struct virtnet_tx *txvq,
455 struct rte_mbuf **cookies,
458 struct vq_desc_extra *dxp;
459 struct virtqueue *vq = virtnet_txq_to_vq(txvq);
460 struct vring_desc *start_dp;
461 struct virtio_net_hdr *hdr;
463 int16_t head_size = vq->hw->vtnet_hdr_size;
466 idx = vq->vq_desc_head_idx;
467 start_dp = vq->vq_split.ring.desc;
470 idx = idx & (vq->vq_nentries - 1);
471 dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
472 dxp->cookie = (void *)cookies[i];
474 virtio_update_packet_stats(&txvq->stats, cookies[i]);
476 hdr = rte_pktmbuf_mtod_offset(cookies[i],
477 struct virtio_net_hdr *, -head_size);
479 /* if offload disabled, hdr is not zeroed yet, do it now */
480 if (!vq->hw->has_tx_offload)
481 virtqueue_clear_net_hdr(hdr);
483 virtqueue_xmit_offload(hdr, cookies[i]);
485 start_dp[idx].addr = rte_mbuf_data_iova(cookies[i]) - head_size;
486 start_dp[idx].len = cookies[i]->data_len + head_size;
487 start_dp[idx].flags = 0;
490 vq_update_avail_ring(vq, idx);
496 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
497 vq->vq_desc_head_idx = idx & (vq->vq_nentries - 1);
501 virtqueue_enqueue_xmit_packed_fast(struct virtnet_tx *txvq,
502 struct rte_mbuf *cookie,
505 struct virtqueue *vq = virtnet_txq_to_vq(txvq);
506 struct vring_packed_desc *dp;
507 struct vq_desc_extra *dxp;
508 uint16_t idx, id, flags;
509 int16_t head_size = vq->hw->vtnet_hdr_size;
510 struct virtio_net_hdr *hdr;
512 id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx;
513 idx = vq->vq_avail_idx;
514 dp = &vq->vq_packed.ring.desc[idx];
516 dxp = &vq->vq_descx[id];
518 dxp->cookie = cookie;
520 flags = vq->vq_packed.cached_flags;
522 /* prepend cannot fail, checked by caller */
523 hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
526 /* if offload disabled, hdr is not zeroed yet, do it now */
527 if (!vq->hw->has_tx_offload)
528 virtqueue_clear_net_hdr(hdr);
530 virtqueue_xmit_offload(hdr, cookie);
532 dp->addr = rte_mbuf_data_iova(cookie) - head_size;
533 dp->len = cookie->data_len + head_size;
536 if (++vq->vq_avail_idx >= vq->vq_nentries) {
537 vq->vq_avail_idx -= vq->vq_nentries;
538 vq->vq_packed.cached_flags ^= VRING_PACKED_DESC_F_AVAIL_USED;
544 vq->vq_desc_head_idx = dxp->next;
545 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
546 vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
549 virtqueue_store_flags_packed(dp, flags, vq->hw->weak_barriers);
553 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
554 uint16_t needed, int use_indirect, int can_push,
557 struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
558 struct vq_desc_extra *dxp;
559 struct virtqueue *vq = virtnet_txq_to_vq(txvq);
560 struct vring_desc *start_dp;
561 uint16_t seg_num = cookie->nb_segs;
562 uint16_t head_idx, idx;
563 int16_t head_size = vq->hw->vtnet_hdr_size;
564 bool prepend_header = false;
565 struct virtio_net_hdr *hdr;
567 head_idx = vq->vq_desc_head_idx;
570 dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
572 dxp = &vq->vq_descx[idx];
573 dxp->cookie = (void *)cookie;
574 dxp->ndescs = needed;
576 start_dp = vq->vq_split.ring.desc;
579 /* prepend cannot fail, checked by caller */
580 hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
582 prepend_header = true;
584 /* if offload disabled, it is not zeroed below, do it now */
585 if (!vq->hw->has_tx_offload)
586 virtqueue_clear_net_hdr(hdr);
587 } else if (use_indirect) {
588 /* setup tx ring slot to point to indirect
589 * descriptor list stored in reserved region.
591 * the first slot in indirect ring is already preset
592 * to point to the header in reserved region
594 start_dp[idx].addr = txvq->virtio_net_hdr_mem +
595 RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
596 start_dp[idx].len = (seg_num + 1) * sizeof(struct vring_desc);
597 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
598 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
600 /* loop below will fill in rest of the indirect elements */
601 start_dp = txr[idx].tx_indir;
604 /* setup first tx ring slot to point to header
605 * stored in reserved region.
607 start_dp[idx].addr = txvq->virtio_net_hdr_mem +
608 RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
609 start_dp[idx].len = vq->hw->vtnet_hdr_size;
610 start_dp[idx].flags = VRING_DESC_F_NEXT;
611 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
613 idx = start_dp[idx].next;
616 if (vq->hw->has_tx_offload)
617 virtqueue_xmit_offload(hdr, cookie);
620 start_dp[idx].addr = rte_mbuf_data_iova(cookie);
621 start_dp[idx].len = cookie->data_len;
622 if (prepend_header) {
623 start_dp[idx].addr -= head_size;
624 start_dp[idx].len += head_size;
625 prepend_header = false;
627 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
628 idx = start_dp[idx].next;
629 } while ((cookie = cookie->next) != NULL);
632 idx = vq->vq_split.ring.desc[head_idx].next;
634 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
636 vq->vq_desc_head_idx = idx;
637 vq_update_avail_ring(vq, head_idx);
640 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
641 vq->vq_desc_tail_idx = idx;
646 virtio_dev_cq_start(struct rte_eth_dev *dev)
648 struct virtio_hw *hw = dev->data->dev_private;
651 rte_spinlock_init(&hw->cvq->lock);
652 VIRTQUEUE_DUMP(virtnet_cq_to_vq(hw->cvq));
657 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
660 unsigned int socket_id __rte_unused,
661 const struct rte_eth_rxconf *rx_conf,
662 struct rte_mempool *mp)
664 uint16_t vq_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
665 struct virtio_hw *hw = dev->data->dev_private;
666 struct virtqueue *vq = hw->vqs[vq_idx];
667 struct virtnet_rx *rxvq;
668 uint16_t rx_free_thresh;
672 PMD_INIT_FUNC_TRACE();
674 if (rx_conf->rx_deferred_start) {
675 PMD_INIT_LOG(ERR, "Rx deferred start is not supported");
679 buf_size = virtio_rx_mem_pool_buf_size(mp);
680 if (!virtio_rx_check_scatter(hw->max_rx_pkt_len, buf_size,
681 hw->rx_ol_scatter, &error)) {
682 PMD_INIT_LOG(ERR, "RxQ %u Rx scatter check failed: %s",
687 rx_free_thresh = rx_conf->rx_free_thresh;
688 if (rx_free_thresh == 0)
690 RTE_MIN(vq->vq_nentries / 4, DEFAULT_RX_FREE_THRESH);
692 if (rx_free_thresh & 0x3) {
693 RTE_LOG(ERR, PMD, "rx_free_thresh must be multiples of four."
694 " (rx_free_thresh=%u port=%u queue=%u)\n",
695 rx_free_thresh, dev->data->port_id, queue_idx);
699 if (rx_free_thresh >= vq->vq_nentries) {
700 RTE_LOG(ERR, PMD, "rx_free_thresh must be less than the "
701 "number of RX entries (%u)."
702 " (rx_free_thresh=%u port=%u queue=%u)\n",
704 rx_free_thresh, dev->data->port_id, queue_idx);
707 vq->vq_free_thresh = rx_free_thresh;
709 if (nb_desc == 0 || nb_desc > vq->vq_nentries)
710 nb_desc = vq->vq_nentries;
711 vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
714 rxvq->queue_id = queue_idx;
716 dev->data->rx_queues[queue_idx] = rxvq;
722 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
724 uint16_t vq_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
725 struct virtio_hw *hw = dev->data->dev_private;
726 struct virtqueue *vq = hw->vqs[vq_idx];
727 struct virtnet_rx *rxvq = &vq->rxq;
731 bool in_order = virtio_with_feature(hw, VIRTIO_F_IN_ORDER);
733 PMD_INIT_FUNC_TRACE();
735 /* Allocate blank mbufs for the each rx descriptor */
738 if (hw->use_vec_rx && !virtio_with_packed_queue(hw)) {
739 for (desc_idx = 0; desc_idx < vq->vq_nentries;
741 vq->vq_split.ring.avail->ring[desc_idx] = desc_idx;
742 vq->vq_split.ring.desc[desc_idx].flags =
746 virtio_rxq_vec_setup(rxvq);
749 memset(rxvq->fake_mbuf, 0, sizeof(*rxvq->fake_mbuf));
750 for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST; desc_idx++)
751 vq->sw_ring[vq->vq_nentries + desc_idx] = rxvq->fake_mbuf;
753 if (hw->use_vec_rx && !virtio_with_packed_queue(hw)) {
754 while (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
755 virtio_rxq_rearm_vec(rxvq);
756 nbufs += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
758 } else if (!virtio_with_packed_queue(vq->hw) && in_order) {
759 if ((!virtqueue_full(vq))) {
760 uint16_t free_cnt = vq->vq_free_cnt;
761 struct rte_mbuf *pkts[free_cnt];
763 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, pkts,
765 error = virtqueue_enqueue_refill_inorder(vq,
768 if (unlikely(error)) {
769 for (i = 0; i < free_cnt; i++)
770 rte_pktmbuf_free(pkts[i]);
776 vq_update_avail_idx(vq);
779 while (!virtqueue_full(vq)) {
780 m = rte_mbuf_raw_alloc(rxvq->mpool);
784 /* Enqueue allocated buffers */
785 if (virtio_with_packed_queue(vq->hw))
786 error = virtqueue_enqueue_recv_refill_packed_init(vq,
789 error = virtqueue_enqueue_recv_refill(vq,
798 if (!virtio_with_packed_queue(vq->hw))
799 vq_update_avail_idx(vq);
802 PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
810 * struct rte_eth_dev *dev: Used to update dev
811 * uint16_t nb_desc: Defaults to values read from config space
812 * unsigned int socket_id: Used to allocate memzone
813 * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
814 * uint16_t queue_idx: Just used as an index in dev txq list
817 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
820 unsigned int socket_id __rte_unused,
821 const struct rte_eth_txconf *tx_conf)
823 uint8_t vq_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
824 struct virtio_hw *hw = dev->data->dev_private;
825 struct virtqueue *vq = hw->vqs[vq_idx];
826 struct virtnet_tx *txvq;
827 uint16_t tx_free_thresh;
829 PMD_INIT_FUNC_TRACE();
831 if (tx_conf->tx_deferred_start) {
832 PMD_INIT_LOG(ERR, "Tx deferred start is not supported");
836 if (nb_desc == 0 || nb_desc > vq->vq_nentries)
837 nb_desc = vq->vq_nentries;
838 vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
841 txvq->queue_id = queue_idx;
843 tx_free_thresh = tx_conf->tx_free_thresh;
844 if (tx_free_thresh == 0)
846 RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
848 if (tx_free_thresh >= (vq->vq_nentries - 3)) {
849 PMD_DRV_LOG(ERR, "tx_free_thresh must be less than the "
850 "number of TX entries minus 3 (%u)."
851 " (tx_free_thresh=%u port=%u queue=%u)\n",
853 tx_free_thresh, dev->data->port_id, queue_idx);
857 vq->vq_free_thresh = tx_free_thresh;
859 dev->data->tx_queues[queue_idx] = txvq;
864 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
867 uint8_t vq_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
868 struct virtio_hw *hw = dev->data->dev_private;
869 struct virtqueue *vq = hw->vqs[vq_idx];
871 PMD_INIT_FUNC_TRACE();
873 if (!virtio_with_packed_queue(hw)) {
874 if (virtio_with_feature(hw, VIRTIO_F_IN_ORDER))
875 vq->vq_split.ring.desc[vq->vq_nentries - 1].next = 0;
884 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
888 * Requeue the discarded mbuf. This should always be
889 * successful since it was just dequeued.
891 if (virtio_with_packed_queue(vq->hw))
892 error = virtqueue_enqueue_recv_refill_packed(vq, &m, 1);
894 error = virtqueue_enqueue_recv_refill(vq, &m, 1);
896 if (unlikely(error)) {
897 PMD_DRV_LOG(ERR, "cannot requeue discarded mbuf");
903 virtio_discard_rxbuf_inorder(struct virtqueue *vq, struct rte_mbuf *m)
907 error = virtqueue_enqueue_refill_inorder(vq, &m, 1);
908 if (unlikely(error)) {
909 PMD_DRV_LOG(ERR, "cannot requeue discarded mbuf");
914 /* Optionally fill offload information in structure */
916 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
918 struct rte_net_hdr_lens hdr_lens;
919 uint32_t hdrlen, ptype;
920 int l4_supported = 0;
923 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
926 m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
928 ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
929 m->packet_type = ptype;
930 if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
931 (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
932 (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
935 if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
936 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
937 if (hdr->csum_start <= hdrlen && l4_supported) {
938 m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
940 /* Unknown proto or tunnel, do sw cksum. We can assume
941 * the cksum field is in the first segment since the
942 * buffers we provided to the host are large enough.
943 * In case of SCTP, this will be wrong since it's a CRC
944 * but there's nothing we can do.
946 uint16_t csum = 0, off;
948 if (rte_raw_cksum_mbuf(m, hdr->csum_start,
949 rte_pktmbuf_pkt_len(m) - hdr->csum_start,
952 if (likely(csum != 0xffff))
954 off = hdr->csum_offset + hdr->csum_start;
955 if (rte_pktmbuf_data_len(m) >= off + 1)
956 *rte_pktmbuf_mtod_offset(m, uint16_t *,
959 } else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
960 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
963 /* GSO request, save required information in mbuf */
964 if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
965 /* Check unsupported modes */
966 if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
967 (hdr->gso_size == 0)) {
971 /* Update mss lengthes in mbuf */
972 m->tso_segsz = hdr->gso_size;
973 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
974 case VIRTIO_NET_HDR_GSO_TCPV4:
975 case VIRTIO_NET_HDR_GSO_TCPV6:
976 m->ol_flags |= PKT_RX_LRO | \
977 PKT_RX_L4_CKSUM_NONE;
987 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
989 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
991 struct virtnet_rx *rxvq = rx_queue;
992 struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
993 struct virtio_hw *hw = vq->hw;
994 struct rte_mbuf *rxm;
995 uint16_t nb_used, num, nb_rx;
996 uint32_t len[VIRTIO_MBUF_BURST_SZ];
997 struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
999 uint32_t i, nb_enqueued;
1001 struct virtio_net_hdr *hdr;
1004 if (unlikely(hw->started == 0))
1007 nb_used = virtqueue_nused(vq);
1009 num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1010 if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1011 num = VIRTIO_MBUF_BURST_SZ;
1012 if (likely(num > DESC_PER_CACHELINE))
1013 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1015 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1016 PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
1019 hdr_size = hw->vtnet_hdr_size;
1021 for (i = 0; i < num ; i++) {
1024 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1026 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1027 PMD_RX_LOG(ERR, "Packet drop");
1029 virtio_discard_rxbuf(vq, rxm);
1030 rxvq->stats.errors++;
1034 rxm->port = rxvq->port_id;
1035 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1039 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1040 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1042 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1043 RTE_PKTMBUF_HEADROOM - hdr_size);
1046 rte_vlan_strip(rxm);
1048 if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1049 virtio_discard_rxbuf(vq, rxm);
1050 rxvq->stats.errors++;
1054 virtio_rx_stats_updated(rxvq, rxm);
1056 rx_pkts[nb_rx++] = rxm;
1059 rxvq->stats.packets += nb_rx;
1061 /* Allocate new mbuf for the used descriptor */
1062 if (likely(!virtqueue_full(vq))) {
1063 uint16_t free_cnt = vq->vq_free_cnt;
1064 struct rte_mbuf *new_pkts[free_cnt];
1066 if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1068 error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1070 if (unlikely(error)) {
1071 for (i = 0; i < free_cnt; i++)
1072 rte_pktmbuf_free(new_pkts[i]);
1074 nb_enqueued += free_cnt;
1076 struct rte_eth_dev *dev =
1077 &rte_eth_devices[rxvq->port_id];
1078 dev->data->rx_mbuf_alloc_failed += free_cnt;
1082 if (likely(nb_enqueued)) {
1083 vq_update_avail_idx(vq);
1085 if (unlikely(virtqueue_kick_prepare(vq))) {
1086 virtqueue_notify(vq);
1087 PMD_RX_LOG(DEBUG, "Notified");
1095 virtio_recv_pkts_packed(void *rx_queue, struct rte_mbuf **rx_pkts,
1098 struct virtnet_rx *rxvq = rx_queue;
1099 struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1100 struct virtio_hw *hw = vq->hw;
1101 struct rte_mbuf *rxm;
1102 uint16_t num, nb_rx;
1103 uint32_t len[VIRTIO_MBUF_BURST_SZ];
1104 struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1106 uint32_t i, nb_enqueued;
1108 struct virtio_net_hdr *hdr;
1111 if (unlikely(hw->started == 0))
1114 num = RTE_MIN(VIRTIO_MBUF_BURST_SZ, nb_pkts);
1115 if (likely(num > DESC_PER_CACHELINE))
1116 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1118 num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1119 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1122 hdr_size = hw->vtnet_hdr_size;
1124 for (i = 0; i < num; i++) {
1127 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1129 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1130 PMD_RX_LOG(ERR, "Packet drop");
1132 virtio_discard_rxbuf(vq, rxm);
1133 rxvq->stats.errors++;
1137 rxm->port = rxvq->port_id;
1138 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1142 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1143 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1145 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1146 RTE_PKTMBUF_HEADROOM - hdr_size);
1149 rte_vlan_strip(rxm);
1151 if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1152 virtio_discard_rxbuf(vq, rxm);
1153 rxvq->stats.errors++;
1157 virtio_rx_stats_updated(rxvq, rxm);
1159 rx_pkts[nb_rx++] = rxm;
1162 rxvq->stats.packets += nb_rx;
1164 /* Allocate new mbuf for the used descriptor */
1165 if (likely(!virtqueue_full(vq))) {
1166 uint16_t free_cnt = vq->vq_free_cnt;
1167 struct rte_mbuf *new_pkts[free_cnt];
1169 if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1171 error = virtqueue_enqueue_recv_refill_packed(vq,
1172 new_pkts, free_cnt);
1173 if (unlikely(error)) {
1174 for (i = 0; i < free_cnt; i++)
1175 rte_pktmbuf_free(new_pkts[i]);
1177 nb_enqueued += free_cnt;
1179 struct rte_eth_dev *dev =
1180 &rte_eth_devices[rxvq->port_id];
1181 dev->data->rx_mbuf_alloc_failed += free_cnt;
1185 if (likely(nb_enqueued)) {
1186 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1187 virtqueue_notify(vq);
1188 PMD_RX_LOG(DEBUG, "Notified");
1197 virtio_recv_pkts_inorder(void *rx_queue,
1198 struct rte_mbuf **rx_pkts,
1201 struct virtnet_rx *rxvq = rx_queue;
1202 struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1203 struct virtio_hw *hw = vq->hw;
1204 struct rte_mbuf *rxm;
1205 struct rte_mbuf *prev = NULL;
1206 uint16_t nb_used, num, nb_rx;
1207 uint32_t len[VIRTIO_MBUF_BURST_SZ];
1208 struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1210 uint32_t nb_enqueued;
1217 if (unlikely(hw->started == 0))
1220 nb_used = virtqueue_nused(vq);
1221 nb_used = RTE_MIN(nb_used, nb_pkts);
1222 nb_used = RTE_MIN(nb_used, VIRTIO_MBUF_BURST_SZ);
1224 PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1229 hdr_size = hw->vtnet_hdr_size;
1231 num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len, nb_used);
1233 for (i = 0; i < num; i++) {
1234 struct virtio_net_hdr_mrg_rxbuf *header;
1236 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1237 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1241 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1242 PMD_RX_LOG(ERR, "Packet drop");
1244 virtio_discard_rxbuf_inorder(vq, rxm);
1245 rxvq->stats.errors++;
1249 header = (struct virtio_net_hdr_mrg_rxbuf *)
1250 ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1253 if (virtio_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
1254 seg_num = header->num_buffers;
1261 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1262 rxm->nb_segs = seg_num;
1265 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1266 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1268 rxm->port = rxvq->port_id;
1270 rx_pkts[nb_rx] = rxm;
1273 if (vq->hw->has_rx_offload &&
1274 virtio_rx_offload(rxm, &header->hdr) < 0) {
1275 virtio_discard_rxbuf_inorder(vq, rxm);
1276 rxvq->stats.errors++;
1281 rte_vlan_strip(rx_pkts[nb_rx]);
1283 seg_res = seg_num - 1;
1285 /* Merge remaining segments */
1286 while (seg_res != 0 && i < (num - 1)) {
1290 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1291 rxm->pkt_len = (uint32_t)(len[i]);
1292 rxm->data_len = (uint16_t)(len[i]);
1294 rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1302 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1307 /* Last packet still need merge segments */
1308 while (seg_res != 0) {
1309 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1310 VIRTIO_MBUF_BURST_SZ);
1312 if (likely(virtqueue_nused(vq) >= rcv_cnt)) {
1313 num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len,
1315 uint16_t extra_idx = 0;
1318 while (extra_idx < rcv_cnt) {
1319 rxm = rcv_pkts[extra_idx];
1321 RTE_PKTMBUF_HEADROOM - hdr_size;
1322 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1323 rxm->data_len = (uint16_t)(len[extra_idx]);
1326 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1332 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1337 "No enough segments for packet.");
1338 rte_pktmbuf_free(rx_pkts[nb_rx]);
1339 rxvq->stats.errors++;
1344 rxvq->stats.packets += nb_rx;
1346 /* Allocate new mbuf for the used descriptor */
1348 if (likely(!virtqueue_full(vq))) {
1349 /* free_cnt may include mrg descs */
1350 uint16_t free_cnt = vq->vq_free_cnt;
1351 struct rte_mbuf *new_pkts[free_cnt];
1353 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1354 error = virtqueue_enqueue_refill_inorder(vq, new_pkts,
1356 if (unlikely(error)) {
1357 for (i = 0; i < free_cnt; i++)
1358 rte_pktmbuf_free(new_pkts[i]);
1360 nb_enqueued += free_cnt;
1362 struct rte_eth_dev *dev =
1363 &rte_eth_devices[rxvq->port_id];
1364 dev->data->rx_mbuf_alloc_failed += free_cnt;
1368 if (likely(nb_enqueued)) {
1369 vq_update_avail_idx(vq);
1371 if (unlikely(virtqueue_kick_prepare(vq))) {
1372 virtqueue_notify(vq);
1373 PMD_RX_LOG(DEBUG, "Notified");
1381 virtio_recv_mergeable_pkts(void *rx_queue,
1382 struct rte_mbuf **rx_pkts,
1385 struct virtnet_rx *rxvq = rx_queue;
1386 struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1387 struct virtio_hw *hw = vq->hw;
1388 struct rte_mbuf *rxm;
1389 struct rte_mbuf *prev = NULL;
1390 uint16_t nb_used, num, nb_rx = 0;
1391 uint32_t len[VIRTIO_MBUF_BURST_SZ];
1392 struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1394 uint32_t nb_enqueued = 0;
1395 uint32_t seg_num = 0;
1396 uint32_t seg_res = 0;
1397 uint32_t hdr_size = hw->vtnet_hdr_size;
1400 if (unlikely(hw->started == 0))
1403 nb_used = virtqueue_nused(vq);
1405 PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1407 num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1408 if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1409 num = VIRTIO_MBUF_BURST_SZ;
1410 if (likely(num > DESC_PER_CACHELINE))
1411 num = num - ((vq->vq_used_cons_idx + num) %
1412 DESC_PER_CACHELINE);
1415 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1417 for (i = 0; i < num; i++) {
1418 struct virtio_net_hdr_mrg_rxbuf *header;
1420 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1421 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1425 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1426 PMD_RX_LOG(ERR, "Packet drop");
1428 virtio_discard_rxbuf(vq, rxm);
1429 rxvq->stats.errors++;
1433 header = (struct virtio_net_hdr_mrg_rxbuf *)
1434 ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1436 seg_num = header->num_buffers;
1440 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1441 rxm->nb_segs = seg_num;
1444 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1445 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1447 rxm->port = rxvq->port_id;
1449 rx_pkts[nb_rx] = rxm;
1452 if (hw->has_rx_offload &&
1453 virtio_rx_offload(rxm, &header->hdr) < 0) {
1454 virtio_discard_rxbuf(vq, rxm);
1455 rxvq->stats.errors++;
1460 rte_vlan_strip(rx_pkts[nb_rx]);
1462 seg_res = seg_num - 1;
1464 /* Merge remaining segments */
1465 while (seg_res != 0 && i < (num - 1)) {
1469 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1470 rxm->pkt_len = (uint32_t)(len[i]);
1471 rxm->data_len = (uint16_t)(len[i]);
1473 rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1481 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1486 /* Last packet still need merge segments */
1487 while (seg_res != 0) {
1488 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1489 VIRTIO_MBUF_BURST_SZ);
1491 if (likely(virtqueue_nused(vq) >= rcv_cnt)) {
1492 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len,
1494 uint16_t extra_idx = 0;
1497 while (extra_idx < rcv_cnt) {
1498 rxm = rcv_pkts[extra_idx];
1500 RTE_PKTMBUF_HEADROOM - hdr_size;
1501 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1502 rxm->data_len = (uint16_t)(len[extra_idx]);
1505 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1511 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1516 "No enough segments for packet.");
1517 rte_pktmbuf_free(rx_pkts[nb_rx]);
1518 rxvq->stats.errors++;
1523 rxvq->stats.packets += nb_rx;
1525 /* Allocate new mbuf for the used descriptor */
1526 if (likely(!virtqueue_full(vq))) {
1527 /* free_cnt may include mrg descs */
1528 uint16_t free_cnt = vq->vq_free_cnt;
1529 struct rte_mbuf *new_pkts[free_cnt];
1531 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1532 error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1534 if (unlikely(error)) {
1535 for (i = 0; i < free_cnt; i++)
1536 rte_pktmbuf_free(new_pkts[i]);
1538 nb_enqueued += free_cnt;
1540 struct rte_eth_dev *dev =
1541 &rte_eth_devices[rxvq->port_id];
1542 dev->data->rx_mbuf_alloc_failed += free_cnt;
1546 if (likely(nb_enqueued)) {
1547 vq_update_avail_idx(vq);
1549 if (unlikely(virtqueue_kick_prepare(vq))) {
1550 virtqueue_notify(vq);
1551 PMD_RX_LOG(DEBUG, "Notified");
1559 virtio_recv_mergeable_pkts_packed(void *rx_queue,
1560 struct rte_mbuf **rx_pkts,
1563 struct virtnet_rx *rxvq = rx_queue;
1564 struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1565 struct virtio_hw *hw = vq->hw;
1566 struct rte_mbuf *rxm;
1567 struct rte_mbuf *prev = NULL;
1568 uint16_t num, nb_rx = 0;
1569 uint32_t len[VIRTIO_MBUF_BURST_SZ];
1570 struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1571 uint32_t nb_enqueued = 0;
1572 uint32_t seg_num = 0;
1573 uint32_t seg_res = 0;
1574 uint32_t hdr_size = hw->vtnet_hdr_size;
1578 if (unlikely(hw->started == 0))
1583 if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1584 num = VIRTIO_MBUF_BURST_SZ;
1585 if (likely(num > DESC_PER_CACHELINE))
1586 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1588 num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1590 for (i = 0; i < num; i++) {
1591 struct virtio_net_hdr_mrg_rxbuf *header;
1593 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1594 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1598 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1599 PMD_RX_LOG(ERR, "Packet drop");
1601 virtio_discard_rxbuf(vq, rxm);
1602 rxvq->stats.errors++;
1606 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)
1607 rxm->buf_addr + RTE_PKTMBUF_HEADROOM - hdr_size);
1608 seg_num = header->num_buffers;
1613 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1614 rxm->nb_segs = seg_num;
1617 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1618 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1620 rxm->port = rxvq->port_id;
1621 rx_pkts[nb_rx] = rxm;
1624 if (hw->has_rx_offload &&
1625 virtio_rx_offload(rxm, &header->hdr) < 0) {
1626 virtio_discard_rxbuf(vq, rxm);
1627 rxvq->stats.errors++;
1632 rte_vlan_strip(rx_pkts[nb_rx]);
1634 seg_res = seg_num - 1;
1636 /* Merge remaining segments */
1637 while (seg_res != 0 && i < (num - 1)) {
1641 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1642 rxm->pkt_len = (uint32_t)(len[i]);
1643 rxm->data_len = (uint16_t)(len[i]);
1645 rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1653 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1658 /* Last packet still need merge segments */
1659 while (seg_res != 0) {
1660 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1661 VIRTIO_MBUF_BURST_SZ);
1662 uint16_t extra_idx = 0;
1664 rcv_cnt = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts,
1666 if (unlikely(rcv_cnt == 0)) {
1667 PMD_RX_LOG(ERR, "No enough segments for packet.");
1668 rte_pktmbuf_free(rx_pkts[nb_rx]);
1669 rxvq->stats.errors++;
1673 while (extra_idx < rcv_cnt) {
1674 rxm = rcv_pkts[extra_idx];
1676 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1677 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1678 rxm->data_len = (uint16_t)(len[extra_idx]);
1682 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1687 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1692 rxvq->stats.packets += nb_rx;
1694 /* Allocate new mbuf for the used descriptor */
1695 if (likely(!virtqueue_full(vq))) {
1696 /* free_cnt may include mrg descs */
1697 uint16_t free_cnt = vq->vq_free_cnt;
1698 struct rte_mbuf *new_pkts[free_cnt];
1700 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1701 error = virtqueue_enqueue_recv_refill_packed(vq,
1702 new_pkts, free_cnt);
1703 if (unlikely(error)) {
1704 for (i = 0; i < free_cnt; i++)
1705 rte_pktmbuf_free(new_pkts[i]);
1707 nb_enqueued += free_cnt;
1709 struct rte_eth_dev *dev =
1710 &rte_eth_devices[rxvq->port_id];
1711 dev->data->rx_mbuf_alloc_failed += free_cnt;
1715 if (likely(nb_enqueued)) {
1716 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1717 virtqueue_notify(vq);
1718 PMD_RX_LOG(DEBUG, "Notified");
1726 virtio_xmit_pkts_prepare(void *tx_queue __rte_unused, struct rte_mbuf **tx_pkts,
1732 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1733 struct rte_mbuf *m = tx_pkts[nb_tx];
1735 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
1736 error = rte_validate_tx_offload(m);
1737 if (unlikely(error)) {
1743 /* Do VLAN tag insertion */
1744 if (unlikely(m->ol_flags & PKT_TX_VLAN_PKT)) {
1745 error = rte_vlan_insert(&m);
1746 /* rte_vlan_insert() may change pointer
1747 * even in the case of failure
1751 if (unlikely(error)) {
1757 error = rte_net_intel_cksum_prepare(m);
1758 if (unlikely(error)) {
1763 if (m->ol_flags & PKT_TX_TCP_SEG)
1764 virtio_tso_fix_cksum(m);
1771 virtio_xmit_pkts_packed(void *tx_queue, struct rte_mbuf **tx_pkts,
1774 struct virtnet_tx *txvq = tx_queue;
1775 struct virtqueue *vq = virtnet_txq_to_vq(txvq);
1776 struct virtio_hw *hw = vq->hw;
1777 uint16_t hdr_size = hw->vtnet_hdr_size;
1779 bool in_order = virtio_with_feature(hw, VIRTIO_F_IN_ORDER);
1781 if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1784 if (unlikely(nb_pkts < 1))
1787 PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1789 if (nb_pkts > vq->vq_free_cnt)
1790 virtio_xmit_cleanup_packed(vq, nb_pkts - vq->vq_free_cnt,
1793 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1794 struct rte_mbuf *txm = tx_pkts[nb_tx];
1795 int can_push = 0, use_indirect = 0, slots, need;
1797 /* optimize ring usage */
1798 if ((virtio_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1799 virtio_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1800 rte_mbuf_refcnt_read(txm) == 1 &&
1801 RTE_MBUF_DIRECT(txm) &&
1802 txm->nb_segs == 1 &&
1803 rte_pktmbuf_headroom(txm) >= hdr_size &&
1804 rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1805 __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1807 else if (virtio_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1808 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1810 /* How many main ring entries are needed to this Tx?
1812 * any_layout => number of segments
1813 * default => number of segments + 1
1815 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1816 need = slots - vq->vq_free_cnt;
1818 /* Positive value indicates it need free vring descriptors */
1819 if (unlikely(need > 0)) {
1820 virtio_xmit_cleanup_packed(vq, need, in_order);
1821 need = slots - vq->vq_free_cnt;
1822 if (unlikely(need > 0)) {
1824 "No free tx descriptors to transmit");
1829 /* Enqueue Packet buffers */
1831 virtqueue_enqueue_xmit_packed_fast(txvq, txm, in_order);
1833 virtqueue_enqueue_xmit_packed(txvq, txm, slots,
1837 virtio_update_packet_stats(&txvq->stats, txm);
1840 txvq->stats.packets += nb_tx;
1842 if (likely(nb_tx)) {
1843 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1844 virtqueue_notify(vq);
1845 PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1853 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1855 struct virtnet_tx *txvq = tx_queue;
1856 struct virtqueue *vq = virtnet_txq_to_vq(txvq);
1857 struct virtio_hw *hw = vq->hw;
1858 uint16_t hdr_size = hw->vtnet_hdr_size;
1859 uint16_t nb_used, nb_tx = 0;
1861 if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1864 if (unlikely(nb_pkts < 1))
1867 PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1869 nb_used = virtqueue_nused(vq);
1871 if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1872 virtio_xmit_cleanup(vq, nb_used);
1874 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1875 struct rte_mbuf *txm = tx_pkts[nb_tx];
1876 int can_push = 0, use_indirect = 0, slots, need;
1878 /* optimize ring usage */
1879 if ((virtio_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1880 virtio_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1881 rte_mbuf_refcnt_read(txm) == 1 &&
1882 RTE_MBUF_DIRECT(txm) &&
1883 txm->nb_segs == 1 &&
1884 rte_pktmbuf_headroom(txm) >= hdr_size &&
1885 rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1886 __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1888 else if (virtio_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1889 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1892 /* How many main ring entries are needed to this Tx?
1893 * any_layout => number of segments
1895 * default => number of segments + 1
1897 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1898 need = slots - vq->vq_free_cnt;
1900 /* Positive value indicates it need free vring descriptors */
1901 if (unlikely(need > 0)) {
1902 nb_used = virtqueue_nused(vq);
1904 need = RTE_MIN(need, (int)nb_used);
1906 virtio_xmit_cleanup(vq, need);
1907 need = slots - vq->vq_free_cnt;
1908 if (unlikely(need > 0)) {
1910 "No free tx descriptors to transmit");
1915 /* Enqueue Packet buffers */
1916 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
1919 virtio_update_packet_stats(&txvq->stats, txm);
1922 txvq->stats.packets += nb_tx;
1924 if (likely(nb_tx)) {
1925 vq_update_avail_idx(vq);
1927 if (unlikely(virtqueue_kick_prepare(vq))) {
1928 virtqueue_notify(vq);
1929 PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1936 static __rte_always_inline int
1937 virtio_xmit_try_cleanup_inorder(struct virtqueue *vq, uint16_t need)
1939 uint16_t nb_used, nb_clean, nb_descs;
1941 nb_descs = vq->vq_free_cnt + need;
1942 nb_used = virtqueue_nused(vq);
1943 nb_clean = RTE_MIN(need, (int)nb_used);
1945 virtio_xmit_cleanup_inorder(vq, nb_clean);
1947 return nb_descs - vq->vq_free_cnt;
1951 virtio_xmit_pkts_inorder(void *tx_queue,
1952 struct rte_mbuf **tx_pkts,
1955 struct virtnet_tx *txvq = tx_queue;
1956 struct virtqueue *vq = virtnet_txq_to_vq(txvq);
1957 struct virtio_hw *hw = vq->hw;
1958 uint16_t hdr_size = hw->vtnet_hdr_size;
1959 uint16_t nb_used, nb_tx = 0, nb_inorder_pkts = 0;
1960 struct rte_mbuf *inorder_pkts[nb_pkts];
1963 if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1966 if (unlikely(nb_pkts < 1))
1970 PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1971 nb_used = virtqueue_nused(vq);
1973 if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1974 virtio_xmit_cleanup_inorder(vq, nb_used);
1976 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1977 struct rte_mbuf *txm = tx_pkts[nb_tx];
1980 /* optimize ring usage */
1981 if ((virtio_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1982 virtio_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1983 rte_mbuf_refcnt_read(txm) == 1 &&
1984 RTE_MBUF_DIRECT(txm) &&
1985 txm->nb_segs == 1 &&
1986 rte_pktmbuf_headroom(txm) >= hdr_size &&
1987 rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1988 __alignof__(struct virtio_net_hdr_mrg_rxbuf))) {
1989 inorder_pkts[nb_inorder_pkts] = txm;
1995 if (nb_inorder_pkts) {
1996 need = nb_inorder_pkts - vq->vq_free_cnt;
1997 if (unlikely(need > 0)) {
1998 need = virtio_xmit_try_cleanup_inorder(vq,
2000 if (unlikely(need > 0)) {
2002 "No free tx descriptors to "
2007 virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2009 nb_inorder_pkts = 0;
2012 slots = txm->nb_segs + 1;
2013 need = slots - vq->vq_free_cnt;
2014 if (unlikely(need > 0)) {
2015 need = virtio_xmit_try_cleanup_inorder(vq, slots);
2017 if (unlikely(need > 0)) {
2019 "No free tx descriptors to transmit");
2023 /* Enqueue Packet buffers */
2024 virtqueue_enqueue_xmit(txvq, txm, slots, 0, 0, 1);
2026 virtio_update_packet_stats(&txvq->stats, txm);
2029 /* Transmit all inorder packets */
2030 if (nb_inorder_pkts) {
2031 need = nb_inorder_pkts - vq->vq_free_cnt;
2032 if (unlikely(need > 0)) {
2033 need = virtio_xmit_try_cleanup_inorder(vq,
2035 if (unlikely(need > 0)) {
2037 "No free tx descriptors to transmit");
2038 nb_inorder_pkts = vq->vq_free_cnt;
2043 virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2047 txvq->stats.packets += nb_tx;
2049 if (likely(nb_tx)) {
2050 vq_update_avail_idx(vq);
2052 if (unlikely(virtqueue_kick_prepare(vq))) {
2053 virtqueue_notify(vq);
2054 PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2064 virtio_recv_pkts_packed_vec(void *rx_queue __rte_unused,
2065 struct rte_mbuf **rx_pkts __rte_unused,
2066 uint16_t nb_pkts __rte_unused)
2072 virtio_xmit_pkts_packed_vec(void *tx_queue __rte_unused,
2073 struct rte_mbuf **tx_pkts __rte_unused,
2074 uint16_t nb_pkts __rte_unused)