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 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];
275 VIRTIO_MBUF_ADDR(cookies[i], vq) +
276 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
278 cookies[i]->buf_len -
279 RTE_PKTMBUF_HEADROOM +
281 start_dp[idx].flags = VRING_DESC_F_WRITE;
283 vq_update_avail_ring(vq, idx);
288 vq->vq_desc_head_idx += num;
289 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
294 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf **cookie,
297 struct vq_desc_extra *dxp;
298 struct virtio_hw *hw = vq->hw;
299 struct vring_desc *start_dp = vq->vq_split.ring.desc;
302 if (unlikely(vq->vq_free_cnt == 0))
304 if (unlikely(vq->vq_free_cnt < num))
307 if (unlikely(vq->vq_desc_head_idx >= vq->vq_nentries))
310 for (i = 0; i < num; i++) {
311 idx = vq->vq_desc_head_idx;
312 dxp = &vq->vq_descx[idx];
313 dxp->cookie = (void *)cookie[i];
317 VIRTIO_MBUF_ADDR(cookie[i], vq) +
318 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
320 cookie[i]->buf_len - RTE_PKTMBUF_HEADROOM +
322 start_dp[idx].flags = VRING_DESC_F_WRITE;
323 vq->vq_desc_head_idx = start_dp[idx].next;
324 vq_update_avail_ring(vq, idx);
325 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END) {
326 vq->vq_desc_tail_idx = vq->vq_desc_head_idx;
331 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
337 virtqueue_enqueue_recv_refill_packed(struct virtqueue *vq,
338 struct rte_mbuf **cookie, uint16_t num)
340 struct vring_packed_desc *start_dp = vq->vq_packed.ring.desc;
341 uint16_t flags = vq->vq_packed.cached_flags;
342 struct virtio_hw *hw = vq->hw;
343 struct vq_desc_extra *dxp;
347 if (unlikely(vq->vq_free_cnt == 0))
349 if (unlikely(vq->vq_free_cnt < num))
352 for (i = 0; i < num; i++) {
353 idx = vq->vq_avail_idx;
354 dxp = &vq->vq_descx[idx];
355 dxp->cookie = (void *)cookie[i];
358 start_dp[idx].addr = VIRTIO_MBUF_ADDR(cookie[i], vq) +
359 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
360 start_dp[idx].len = cookie[i]->buf_len - RTE_PKTMBUF_HEADROOM
361 + hw->vtnet_hdr_size;
363 vq->vq_desc_head_idx = dxp->next;
364 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
365 vq->vq_desc_tail_idx = vq->vq_desc_head_idx;
367 virtqueue_store_flags_packed(&start_dp[idx], flags,
370 if (++vq->vq_avail_idx >= vq->vq_nentries) {
371 vq->vq_avail_idx -= vq->vq_nentries;
372 vq->vq_packed.cached_flags ^=
373 VRING_PACKED_DESC_F_AVAIL_USED;
374 flags = vq->vq_packed.cached_flags;
377 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
381 /* When doing TSO, the IP length is not included in the pseudo header
382 * checksum of the packet given to the PMD, but for virtio it is
386 virtio_tso_fix_cksum(struct rte_mbuf *m)
388 /* common case: header is not fragmented */
389 if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
391 struct rte_ipv4_hdr *iph;
392 struct rte_ipv6_hdr *ip6h;
393 struct rte_tcp_hdr *th;
394 uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
397 iph = rte_pktmbuf_mtod_offset(m,
398 struct rte_ipv4_hdr *, m->l2_len);
399 th = RTE_PTR_ADD(iph, m->l3_len);
400 if ((iph->version_ihl >> 4) == 4) {
401 iph->hdr_checksum = 0;
402 iph->hdr_checksum = rte_ipv4_cksum(iph);
403 ip_len = iph->total_length;
404 ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
407 ip6h = (struct rte_ipv6_hdr *)iph;
408 ip_paylen = ip6h->payload_len;
411 /* calculate the new phdr checksum not including ip_paylen */
412 prev_cksum = th->cksum;
415 tmp = (tmp & 0xffff) + (tmp >> 16);
418 /* replace it in the packet */
419 th->cksum = new_cksum;
427 virtqueue_enqueue_xmit_inorder(struct virtnet_tx *txvq,
428 struct rte_mbuf **cookies,
431 struct vq_desc_extra *dxp;
432 struct virtqueue *vq = txvq->vq;
433 struct vring_desc *start_dp;
434 struct virtio_net_hdr *hdr;
436 int16_t head_size = vq->hw->vtnet_hdr_size;
439 idx = vq->vq_desc_head_idx;
440 start_dp = vq->vq_split.ring.desc;
443 idx = idx & (vq->vq_nentries - 1);
444 dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
445 dxp->cookie = (void *)cookies[i];
447 virtio_update_packet_stats(&txvq->stats, cookies[i]);
449 hdr = rte_pktmbuf_mtod_offset(cookies[i],
450 struct virtio_net_hdr *, -head_size);
452 /* if offload disabled, hdr is not zeroed yet, do it now */
453 if (!vq->hw->has_tx_offload)
454 virtqueue_clear_net_hdr(hdr);
456 virtqueue_xmit_offload(hdr, cookies[i], true);
459 VIRTIO_MBUF_DATA_DMA_ADDR(cookies[i], vq) - head_size;
460 start_dp[idx].len = cookies[i]->data_len + head_size;
461 start_dp[idx].flags = 0;
464 vq_update_avail_ring(vq, idx);
470 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
471 vq->vq_desc_head_idx = idx & (vq->vq_nentries - 1);
475 virtqueue_enqueue_xmit_packed_fast(struct virtnet_tx *txvq,
476 struct rte_mbuf *cookie,
479 struct virtqueue *vq = txvq->vq;
480 struct vring_packed_desc *dp;
481 struct vq_desc_extra *dxp;
482 uint16_t idx, id, flags;
483 int16_t head_size = vq->hw->vtnet_hdr_size;
484 struct virtio_net_hdr *hdr;
486 id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx;
487 idx = vq->vq_avail_idx;
488 dp = &vq->vq_packed.ring.desc[idx];
490 dxp = &vq->vq_descx[id];
492 dxp->cookie = cookie;
494 flags = vq->vq_packed.cached_flags;
496 /* prepend cannot fail, checked by caller */
497 hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
500 /* if offload disabled, hdr is not zeroed yet, do it now */
501 if (!vq->hw->has_tx_offload)
502 virtqueue_clear_net_hdr(hdr);
504 virtqueue_xmit_offload(hdr, cookie, true);
506 dp->addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq) - head_size;
507 dp->len = cookie->data_len + head_size;
510 if (++vq->vq_avail_idx >= vq->vq_nentries) {
511 vq->vq_avail_idx -= vq->vq_nentries;
512 vq->vq_packed.cached_flags ^= VRING_PACKED_DESC_F_AVAIL_USED;
518 vq->vq_desc_head_idx = dxp->next;
519 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
520 vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
523 virtqueue_store_flags_packed(dp, flags, vq->hw->weak_barriers);
527 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
528 uint16_t needed, int use_indirect, int can_push,
531 struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
532 struct vq_desc_extra *dxp;
533 struct virtqueue *vq = txvq->vq;
534 struct vring_desc *start_dp;
535 uint16_t seg_num = cookie->nb_segs;
536 uint16_t head_idx, idx;
537 int16_t head_size = vq->hw->vtnet_hdr_size;
538 bool prepend_header = false;
539 struct virtio_net_hdr *hdr;
541 head_idx = vq->vq_desc_head_idx;
544 dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
546 dxp = &vq->vq_descx[idx];
547 dxp->cookie = (void *)cookie;
548 dxp->ndescs = needed;
550 start_dp = vq->vq_split.ring.desc;
553 /* prepend cannot fail, checked by caller */
554 hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
556 prepend_header = true;
558 /* if offload disabled, it is not zeroed below, do it now */
559 if (!vq->hw->has_tx_offload)
560 virtqueue_clear_net_hdr(hdr);
561 } else if (use_indirect) {
562 /* setup tx ring slot to point to indirect
563 * descriptor list stored in reserved region.
565 * the first slot in indirect ring is already preset
566 * to point to the header in reserved region
568 start_dp[idx].addr = txvq->virtio_net_hdr_mem +
569 RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
570 start_dp[idx].len = (seg_num + 1) * sizeof(struct vring_desc);
571 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
572 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
574 /* loop below will fill in rest of the indirect elements */
575 start_dp = txr[idx].tx_indir;
578 /* setup first tx ring slot to point to header
579 * stored in reserved region.
581 start_dp[idx].addr = txvq->virtio_net_hdr_mem +
582 RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
583 start_dp[idx].len = vq->hw->vtnet_hdr_size;
584 start_dp[idx].flags = VRING_DESC_F_NEXT;
585 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
587 idx = start_dp[idx].next;
590 virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload);
593 start_dp[idx].addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
594 start_dp[idx].len = cookie->data_len;
595 if (prepend_header) {
596 start_dp[idx].addr -= head_size;
597 start_dp[idx].len += head_size;
598 prepend_header = false;
600 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
601 idx = start_dp[idx].next;
602 } while ((cookie = cookie->next) != NULL);
605 idx = vq->vq_split.ring.desc[head_idx].next;
607 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
609 vq->vq_desc_head_idx = idx;
610 vq_update_avail_ring(vq, head_idx);
613 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
614 vq->vq_desc_tail_idx = idx;
619 virtio_dev_cq_start(struct rte_eth_dev *dev)
621 struct virtio_hw *hw = dev->data->dev_private;
623 if (hw->cvq && hw->cvq->vq) {
624 rte_spinlock_init(&hw->cvq->lock);
625 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
630 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
633 unsigned int socket_id __rte_unused,
634 const struct rte_eth_rxconf *rx_conf,
635 struct rte_mempool *mp)
637 uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
638 struct virtio_hw *hw = dev->data->dev_private;
639 struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
640 struct virtnet_rx *rxvq;
641 uint16_t rx_free_thresh;
643 PMD_INIT_FUNC_TRACE();
645 if (rx_conf->rx_deferred_start) {
646 PMD_INIT_LOG(ERR, "Rx deferred start is not supported");
650 rx_free_thresh = rx_conf->rx_free_thresh;
651 if (rx_free_thresh == 0)
653 RTE_MIN(vq->vq_nentries / 4, DEFAULT_RX_FREE_THRESH);
655 if (rx_free_thresh & 0x3) {
656 RTE_LOG(ERR, PMD, "rx_free_thresh must be multiples of four."
657 " (rx_free_thresh=%u port=%u queue=%u)\n",
658 rx_free_thresh, dev->data->port_id, queue_idx);
662 if (rx_free_thresh >= vq->vq_nentries) {
663 RTE_LOG(ERR, PMD, "rx_free_thresh must be less than the "
664 "number of RX entries (%u)."
665 " (rx_free_thresh=%u port=%u queue=%u)\n",
667 rx_free_thresh, dev->data->port_id, queue_idx);
670 vq->vq_free_thresh = rx_free_thresh;
672 if (nb_desc == 0 || nb_desc > vq->vq_nentries)
673 nb_desc = vq->vq_nentries;
674 vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
677 rxvq->queue_id = queue_idx;
679 dev->data->rx_queues[queue_idx] = rxvq;
685 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
687 uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
688 struct virtio_hw *hw = dev->data->dev_private;
689 struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
690 struct virtnet_rx *rxvq = &vq->rxq;
694 bool in_order = vtpci_with_feature(hw, VIRTIO_F_IN_ORDER);
696 PMD_INIT_FUNC_TRACE();
698 /* Allocate blank mbufs for the each rx descriptor */
701 if (hw->use_vec_rx && !vtpci_packed_queue(hw)) {
702 for (desc_idx = 0; desc_idx < vq->vq_nentries;
704 vq->vq_split.ring.avail->ring[desc_idx] = desc_idx;
705 vq->vq_split.ring.desc[desc_idx].flags =
709 virtio_rxq_vec_setup(rxvq);
712 memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
713 for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
715 vq->sw_ring[vq->vq_nentries + desc_idx] =
719 if (hw->use_vec_rx && !vtpci_packed_queue(hw)) {
720 while (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
721 virtio_rxq_rearm_vec(rxvq);
722 nbufs += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
724 } else if (!vtpci_packed_queue(vq->hw) && in_order) {
725 if ((!virtqueue_full(vq))) {
726 uint16_t free_cnt = vq->vq_free_cnt;
727 struct rte_mbuf *pkts[free_cnt];
729 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, pkts,
731 error = virtqueue_enqueue_refill_inorder(vq,
734 if (unlikely(error)) {
735 for (i = 0; i < free_cnt; i++)
736 rte_pktmbuf_free(pkts[i]);
741 vq_update_avail_idx(vq);
744 while (!virtqueue_full(vq)) {
745 m = rte_mbuf_raw_alloc(rxvq->mpool);
749 /* Enqueue allocated buffers */
750 if (vtpci_packed_queue(vq->hw))
751 error = virtqueue_enqueue_recv_refill_packed(vq,
754 error = virtqueue_enqueue_recv_refill(vq,
763 if (!vtpci_packed_queue(vq->hw))
764 vq_update_avail_idx(vq);
767 PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
775 * struct rte_eth_dev *dev: Used to update dev
776 * uint16_t nb_desc: Defaults to values read from config space
777 * unsigned int socket_id: Used to allocate memzone
778 * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
779 * uint16_t queue_idx: Just used as an index in dev txq list
782 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
785 unsigned int socket_id __rte_unused,
786 const struct rte_eth_txconf *tx_conf)
788 uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
789 struct virtio_hw *hw = dev->data->dev_private;
790 struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
791 struct virtnet_tx *txvq;
792 uint16_t tx_free_thresh;
794 PMD_INIT_FUNC_TRACE();
796 if (tx_conf->tx_deferred_start) {
797 PMD_INIT_LOG(ERR, "Tx deferred start is not supported");
801 if (nb_desc == 0 || nb_desc > vq->vq_nentries)
802 nb_desc = vq->vq_nentries;
803 vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
806 txvq->queue_id = queue_idx;
808 tx_free_thresh = tx_conf->tx_free_thresh;
809 if (tx_free_thresh == 0)
811 RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
813 if (tx_free_thresh >= (vq->vq_nentries - 3)) {
814 PMD_DRV_LOG(ERR, "tx_free_thresh must be less than the "
815 "number of TX entries minus 3 (%u)."
816 " (tx_free_thresh=%u port=%u queue=%u)\n",
818 tx_free_thresh, dev->data->port_id, queue_idx);
822 vq->vq_free_thresh = tx_free_thresh;
824 dev->data->tx_queues[queue_idx] = txvq;
829 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
832 uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
833 struct virtio_hw *hw = dev->data->dev_private;
834 struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
836 PMD_INIT_FUNC_TRACE();
838 if (!vtpci_packed_queue(hw)) {
839 if (vtpci_with_feature(hw, VIRTIO_F_IN_ORDER))
840 vq->vq_split.ring.desc[vq->vq_nentries - 1].next = 0;
849 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
853 * Requeue the discarded mbuf. This should always be
854 * successful since it was just dequeued.
856 if (vtpci_packed_queue(vq->hw))
857 error = virtqueue_enqueue_recv_refill_packed(vq, &m, 1);
859 error = virtqueue_enqueue_recv_refill(vq, &m, 1);
861 if (unlikely(error)) {
862 PMD_DRV_LOG(ERR, "cannot requeue discarded mbuf");
868 virtio_discard_rxbuf_inorder(struct virtqueue *vq, struct rte_mbuf *m)
872 error = virtqueue_enqueue_refill_inorder(vq, &m, 1);
873 if (unlikely(error)) {
874 PMD_DRV_LOG(ERR, "cannot requeue discarded mbuf");
879 /* Optionally fill offload information in structure */
881 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
883 struct rte_net_hdr_lens hdr_lens;
884 uint32_t hdrlen, ptype;
885 int l4_supported = 0;
888 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
891 m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
893 ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
894 m->packet_type = ptype;
895 if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
896 (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
897 (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
900 if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
901 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
902 if (hdr->csum_start <= hdrlen && l4_supported) {
903 m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
905 /* Unknown proto or tunnel, do sw cksum. We can assume
906 * the cksum field is in the first segment since the
907 * buffers we provided to the host are large enough.
908 * In case of SCTP, this will be wrong since it's a CRC
909 * but there's nothing we can do.
911 uint16_t csum = 0, off;
913 if (rte_raw_cksum_mbuf(m, hdr->csum_start,
914 rte_pktmbuf_pkt_len(m) - hdr->csum_start,
917 if (likely(csum != 0xffff))
919 off = hdr->csum_offset + hdr->csum_start;
920 if (rte_pktmbuf_data_len(m) >= off + 1)
921 *rte_pktmbuf_mtod_offset(m, uint16_t *,
924 } else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
925 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
928 /* GSO request, save required information in mbuf */
929 if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
930 /* Check unsupported modes */
931 if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
932 (hdr->gso_size == 0)) {
936 /* Update mss lengthes in mbuf */
937 m->tso_segsz = hdr->gso_size;
938 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
939 case VIRTIO_NET_HDR_GSO_TCPV4:
940 case VIRTIO_NET_HDR_GSO_TCPV6:
941 m->ol_flags |= PKT_RX_LRO | \
942 PKT_RX_L4_CKSUM_NONE;
952 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
954 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
956 struct virtnet_rx *rxvq = rx_queue;
957 struct virtqueue *vq = rxvq->vq;
958 struct virtio_hw *hw = vq->hw;
959 struct rte_mbuf *rxm;
960 uint16_t nb_used, num, nb_rx;
961 uint32_t len[VIRTIO_MBUF_BURST_SZ];
962 struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
964 uint32_t i, nb_enqueued;
966 struct virtio_net_hdr *hdr;
969 if (unlikely(hw->started == 0))
972 nb_used = virtqueue_nused(vq);
974 num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
975 if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
976 num = VIRTIO_MBUF_BURST_SZ;
977 if (likely(num > DESC_PER_CACHELINE))
978 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
980 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
981 PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
984 hdr_size = hw->vtnet_hdr_size;
986 for (i = 0; i < num ; i++) {
989 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
991 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
992 PMD_RX_LOG(ERR, "Packet drop");
994 virtio_discard_rxbuf(vq, rxm);
995 rxvq->stats.errors++;
999 rxm->port = rxvq->port_id;
1000 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1004 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1005 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1007 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1008 RTE_PKTMBUF_HEADROOM - hdr_size);
1011 rte_vlan_strip(rxm);
1013 if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1014 virtio_discard_rxbuf(vq, rxm);
1015 rxvq->stats.errors++;
1019 virtio_rx_stats_updated(rxvq, rxm);
1021 rx_pkts[nb_rx++] = rxm;
1024 rxvq->stats.packets += nb_rx;
1026 /* Allocate new mbuf for the used descriptor */
1027 if (likely(!virtqueue_full(vq))) {
1028 uint16_t free_cnt = vq->vq_free_cnt;
1029 struct rte_mbuf *new_pkts[free_cnt];
1031 if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1033 error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1035 if (unlikely(error)) {
1036 for (i = 0; i < free_cnt; i++)
1037 rte_pktmbuf_free(new_pkts[i]);
1039 nb_enqueued += free_cnt;
1041 struct rte_eth_dev *dev =
1042 &rte_eth_devices[rxvq->port_id];
1043 dev->data->rx_mbuf_alloc_failed += free_cnt;
1047 if (likely(nb_enqueued)) {
1048 vq_update_avail_idx(vq);
1050 if (unlikely(virtqueue_kick_prepare(vq))) {
1051 virtqueue_notify(vq);
1052 PMD_RX_LOG(DEBUG, "Notified");
1060 virtio_recv_pkts_packed(void *rx_queue, struct rte_mbuf **rx_pkts,
1063 struct virtnet_rx *rxvq = rx_queue;
1064 struct virtqueue *vq = rxvq->vq;
1065 struct virtio_hw *hw = vq->hw;
1066 struct rte_mbuf *rxm;
1067 uint16_t num, nb_rx;
1068 uint32_t len[VIRTIO_MBUF_BURST_SZ];
1069 struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1071 uint32_t i, nb_enqueued;
1073 struct virtio_net_hdr *hdr;
1076 if (unlikely(hw->started == 0))
1079 num = RTE_MIN(VIRTIO_MBUF_BURST_SZ, nb_pkts);
1080 if (likely(num > DESC_PER_CACHELINE))
1081 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1083 num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1084 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1087 hdr_size = hw->vtnet_hdr_size;
1089 for (i = 0; i < num; i++) {
1092 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1094 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1095 PMD_RX_LOG(ERR, "Packet drop");
1097 virtio_discard_rxbuf(vq, rxm);
1098 rxvq->stats.errors++;
1102 rxm->port = rxvq->port_id;
1103 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1107 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1108 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1110 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1111 RTE_PKTMBUF_HEADROOM - hdr_size);
1114 rte_vlan_strip(rxm);
1116 if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1117 virtio_discard_rxbuf(vq, rxm);
1118 rxvq->stats.errors++;
1122 virtio_rx_stats_updated(rxvq, rxm);
1124 rx_pkts[nb_rx++] = rxm;
1127 rxvq->stats.packets += nb_rx;
1129 /* Allocate new mbuf for the used descriptor */
1130 if (likely(!virtqueue_full(vq))) {
1131 uint16_t free_cnt = vq->vq_free_cnt;
1132 struct rte_mbuf *new_pkts[free_cnt];
1134 if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1136 error = virtqueue_enqueue_recv_refill_packed(vq,
1137 new_pkts, free_cnt);
1138 if (unlikely(error)) {
1139 for (i = 0; i < free_cnt; i++)
1140 rte_pktmbuf_free(new_pkts[i]);
1142 nb_enqueued += free_cnt;
1144 struct rte_eth_dev *dev =
1145 &rte_eth_devices[rxvq->port_id];
1146 dev->data->rx_mbuf_alloc_failed += free_cnt;
1150 if (likely(nb_enqueued)) {
1151 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1152 virtqueue_notify(vq);
1153 PMD_RX_LOG(DEBUG, "Notified");
1162 virtio_recv_pkts_inorder(void *rx_queue,
1163 struct rte_mbuf **rx_pkts,
1166 struct virtnet_rx *rxvq = rx_queue;
1167 struct virtqueue *vq = rxvq->vq;
1168 struct virtio_hw *hw = vq->hw;
1169 struct rte_mbuf *rxm;
1170 struct rte_mbuf *prev = NULL;
1171 uint16_t nb_used, num, nb_rx;
1172 uint32_t len[VIRTIO_MBUF_BURST_SZ];
1173 struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1175 uint32_t nb_enqueued;
1182 if (unlikely(hw->started == 0))
1185 nb_used = virtqueue_nused(vq);
1186 nb_used = RTE_MIN(nb_used, nb_pkts);
1187 nb_used = RTE_MIN(nb_used, VIRTIO_MBUF_BURST_SZ);
1189 PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1194 hdr_size = hw->vtnet_hdr_size;
1196 num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len, nb_used);
1198 for (i = 0; i < num; i++) {
1199 struct virtio_net_hdr_mrg_rxbuf *header;
1201 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1202 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1206 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1207 PMD_RX_LOG(ERR, "Packet drop");
1209 virtio_discard_rxbuf_inorder(vq, rxm);
1210 rxvq->stats.errors++;
1214 header = (struct virtio_net_hdr_mrg_rxbuf *)
1215 ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1218 if (vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
1219 seg_num = header->num_buffers;
1226 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1227 rxm->nb_segs = seg_num;
1230 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1231 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1233 rxm->port = rxvq->port_id;
1235 rx_pkts[nb_rx] = rxm;
1238 if (vq->hw->has_rx_offload &&
1239 virtio_rx_offload(rxm, &header->hdr) < 0) {
1240 virtio_discard_rxbuf_inorder(vq, rxm);
1241 rxvq->stats.errors++;
1246 rte_vlan_strip(rx_pkts[nb_rx]);
1248 seg_res = seg_num - 1;
1250 /* Merge remaining segments */
1251 while (seg_res != 0 && i < (num - 1)) {
1255 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1256 rxm->pkt_len = (uint32_t)(len[i]);
1257 rxm->data_len = (uint16_t)(len[i]);
1259 rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1267 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1272 /* Last packet still need merge segments */
1273 while (seg_res != 0) {
1274 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1275 VIRTIO_MBUF_BURST_SZ);
1277 if (likely(virtqueue_nused(vq) >= rcv_cnt)) {
1278 num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len,
1280 uint16_t extra_idx = 0;
1283 while (extra_idx < rcv_cnt) {
1284 rxm = rcv_pkts[extra_idx];
1286 RTE_PKTMBUF_HEADROOM - hdr_size;
1287 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1288 rxm->data_len = (uint16_t)(len[extra_idx]);
1291 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1297 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1302 "No enough segments for packet.");
1303 rte_pktmbuf_free(rx_pkts[nb_rx]);
1304 rxvq->stats.errors++;
1309 rxvq->stats.packets += nb_rx;
1311 /* Allocate new mbuf for the used descriptor */
1313 if (likely(!virtqueue_full(vq))) {
1314 /* free_cnt may include mrg descs */
1315 uint16_t free_cnt = vq->vq_free_cnt;
1316 struct rte_mbuf *new_pkts[free_cnt];
1318 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1319 error = virtqueue_enqueue_refill_inorder(vq, new_pkts,
1321 if (unlikely(error)) {
1322 for (i = 0; i < free_cnt; i++)
1323 rte_pktmbuf_free(new_pkts[i]);
1325 nb_enqueued += free_cnt;
1327 struct rte_eth_dev *dev =
1328 &rte_eth_devices[rxvq->port_id];
1329 dev->data->rx_mbuf_alloc_failed += free_cnt;
1333 if (likely(nb_enqueued)) {
1334 vq_update_avail_idx(vq);
1336 if (unlikely(virtqueue_kick_prepare(vq))) {
1337 virtqueue_notify(vq);
1338 PMD_RX_LOG(DEBUG, "Notified");
1346 virtio_recv_mergeable_pkts(void *rx_queue,
1347 struct rte_mbuf **rx_pkts,
1350 struct virtnet_rx *rxvq = rx_queue;
1351 struct virtqueue *vq = rxvq->vq;
1352 struct virtio_hw *hw = vq->hw;
1353 struct rte_mbuf *rxm;
1354 struct rte_mbuf *prev = NULL;
1355 uint16_t nb_used, num, nb_rx = 0;
1356 uint32_t len[VIRTIO_MBUF_BURST_SZ];
1357 struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1359 uint32_t nb_enqueued = 0;
1360 uint32_t seg_num = 0;
1361 uint32_t seg_res = 0;
1362 uint32_t hdr_size = hw->vtnet_hdr_size;
1365 if (unlikely(hw->started == 0))
1368 nb_used = virtqueue_nused(vq);
1370 PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1372 num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1373 if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1374 num = VIRTIO_MBUF_BURST_SZ;
1375 if (likely(num > DESC_PER_CACHELINE))
1376 num = num - ((vq->vq_used_cons_idx + num) %
1377 DESC_PER_CACHELINE);
1380 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1382 for (i = 0; i < num; i++) {
1383 struct virtio_net_hdr_mrg_rxbuf *header;
1385 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1386 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1390 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1391 PMD_RX_LOG(ERR, "Packet drop");
1393 virtio_discard_rxbuf(vq, rxm);
1394 rxvq->stats.errors++;
1398 header = (struct virtio_net_hdr_mrg_rxbuf *)
1399 ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1401 seg_num = header->num_buffers;
1405 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1406 rxm->nb_segs = seg_num;
1409 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1410 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1412 rxm->port = rxvq->port_id;
1414 rx_pkts[nb_rx] = rxm;
1417 if (hw->has_rx_offload &&
1418 virtio_rx_offload(rxm, &header->hdr) < 0) {
1419 virtio_discard_rxbuf(vq, rxm);
1420 rxvq->stats.errors++;
1425 rte_vlan_strip(rx_pkts[nb_rx]);
1427 seg_res = seg_num - 1;
1429 /* Merge remaining segments */
1430 while (seg_res != 0 && i < (num - 1)) {
1434 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1435 rxm->pkt_len = (uint32_t)(len[i]);
1436 rxm->data_len = (uint16_t)(len[i]);
1438 rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1446 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1451 /* Last packet still need merge segments */
1452 while (seg_res != 0) {
1453 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1454 VIRTIO_MBUF_BURST_SZ);
1456 if (likely(virtqueue_nused(vq) >= rcv_cnt)) {
1457 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len,
1459 uint16_t extra_idx = 0;
1462 while (extra_idx < rcv_cnt) {
1463 rxm = rcv_pkts[extra_idx];
1465 RTE_PKTMBUF_HEADROOM - hdr_size;
1466 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1467 rxm->data_len = (uint16_t)(len[extra_idx]);
1470 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1476 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1481 "No enough segments for packet.");
1482 rte_pktmbuf_free(rx_pkts[nb_rx]);
1483 rxvq->stats.errors++;
1488 rxvq->stats.packets += nb_rx;
1490 /* Allocate new mbuf for the used descriptor */
1491 if (likely(!virtqueue_full(vq))) {
1492 /* free_cnt may include mrg descs */
1493 uint16_t free_cnt = vq->vq_free_cnt;
1494 struct rte_mbuf *new_pkts[free_cnt];
1496 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1497 error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1499 if (unlikely(error)) {
1500 for (i = 0; i < free_cnt; i++)
1501 rte_pktmbuf_free(new_pkts[i]);
1503 nb_enqueued += free_cnt;
1505 struct rte_eth_dev *dev =
1506 &rte_eth_devices[rxvq->port_id];
1507 dev->data->rx_mbuf_alloc_failed += free_cnt;
1511 if (likely(nb_enqueued)) {
1512 vq_update_avail_idx(vq);
1514 if (unlikely(virtqueue_kick_prepare(vq))) {
1515 virtqueue_notify(vq);
1516 PMD_RX_LOG(DEBUG, "Notified");
1524 virtio_recv_mergeable_pkts_packed(void *rx_queue,
1525 struct rte_mbuf **rx_pkts,
1528 struct virtnet_rx *rxvq = rx_queue;
1529 struct virtqueue *vq = rxvq->vq;
1530 struct virtio_hw *hw = vq->hw;
1531 struct rte_mbuf *rxm;
1532 struct rte_mbuf *prev = NULL;
1533 uint16_t num, nb_rx = 0;
1534 uint32_t len[VIRTIO_MBUF_BURST_SZ];
1535 struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1536 uint32_t nb_enqueued = 0;
1537 uint32_t seg_num = 0;
1538 uint32_t seg_res = 0;
1539 uint32_t hdr_size = hw->vtnet_hdr_size;
1543 if (unlikely(hw->started == 0))
1548 if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1549 num = VIRTIO_MBUF_BURST_SZ;
1550 if (likely(num > DESC_PER_CACHELINE))
1551 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1553 num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1555 for (i = 0; i < num; i++) {
1556 struct virtio_net_hdr_mrg_rxbuf *header;
1558 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1559 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1563 if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1564 PMD_RX_LOG(ERR, "Packet drop");
1566 virtio_discard_rxbuf(vq, rxm);
1567 rxvq->stats.errors++;
1571 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)
1572 rxm->buf_addr + RTE_PKTMBUF_HEADROOM - hdr_size);
1573 seg_num = header->num_buffers;
1578 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1579 rxm->nb_segs = seg_num;
1582 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1583 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1585 rxm->port = rxvq->port_id;
1586 rx_pkts[nb_rx] = rxm;
1589 if (hw->has_rx_offload &&
1590 virtio_rx_offload(rxm, &header->hdr) < 0) {
1591 virtio_discard_rxbuf(vq, rxm);
1592 rxvq->stats.errors++;
1597 rte_vlan_strip(rx_pkts[nb_rx]);
1599 seg_res = seg_num - 1;
1601 /* Merge remaining segments */
1602 while (seg_res != 0 && i < (num - 1)) {
1606 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1607 rxm->pkt_len = (uint32_t)(len[i]);
1608 rxm->data_len = (uint16_t)(len[i]);
1610 rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1618 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1623 /* Last packet still need merge segments */
1624 while (seg_res != 0) {
1625 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1626 VIRTIO_MBUF_BURST_SZ);
1627 uint16_t extra_idx = 0;
1629 rcv_cnt = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts,
1631 if (unlikely(rcv_cnt == 0)) {
1632 PMD_RX_LOG(ERR, "No enough segments for packet.");
1633 rte_pktmbuf_free(rx_pkts[nb_rx]);
1634 rxvq->stats.errors++;
1638 while (extra_idx < rcv_cnt) {
1639 rxm = rcv_pkts[extra_idx];
1641 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1642 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1643 rxm->data_len = (uint16_t)(len[extra_idx]);
1647 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1652 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1657 rxvq->stats.packets += nb_rx;
1659 /* Allocate new mbuf for the used descriptor */
1660 if (likely(!virtqueue_full(vq))) {
1661 /* free_cnt may include mrg descs */
1662 uint16_t free_cnt = vq->vq_free_cnt;
1663 struct rte_mbuf *new_pkts[free_cnt];
1665 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1666 error = virtqueue_enqueue_recv_refill_packed(vq,
1667 new_pkts, free_cnt);
1668 if (unlikely(error)) {
1669 for (i = 0; i < free_cnt; i++)
1670 rte_pktmbuf_free(new_pkts[i]);
1672 nb_enqueued += free_cnt;
1674 struct rte_eth_dev *dev =
1675 &rte_eth_devices[rxvq->port_id];
1676 dev->data->rx_mbuf_alloc_failed += free_cnt;
1680 if (likely(nb_enqueued)) {
1681 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1682 virtqueue_notify(vq);
1683 PMD_RX_LOG(DEBUG, "Notified");
1691 virtio_xmit_pkts_prepare(void *tx_queue __rte_unused, struct rte_mbuf **tx_pkts,
1697 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1698 struct rte_mbuf *m = tx_pkts[nb_tx];
1700 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
1701 error = rte_validate_tx_offload(m);
1702 if (unlikely(error)) {
1708 /* Do VLAN tag insertion */
1709 if (unlikely(m->ol_flags & PKT_TX_VLAN_PKT)) {
1710 error = rte_vlan_insert(&m);
1711 /* rte_vlan_insert() may change pointer
1712 * even in the case of failure
1716 if (unlikely(error)) {
1722 error = rte_net_intel_cksum_prepare(m);
1723 if (unlikely(error)) {
1728 if (m->ol_flags & PKT_TX_TCP_SEG)
1729 virtio_tso_fix_cksum(m);
1736 virtio_xmit_pkts_packed(void *tx_queue, struct rte_mbuf **tx_pkts,
1739 struct virtnet_tx *txvq = tx_queue;
1740 struct virtqueue *vq = txvq->vq;
1741 struct virtio_hw *hw = vq->hw;
1742 uint16_t hdr_size = hw->vtnet_hdr_size;
1744 bool in_order = vtpci_with_feature(hw, VIRTIO_F_IN_ORDER);
1746 if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1749 if (unlikely(nb_pkts < 1))
1752 PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1754 if (nb_pkts > vq->vq_free_cnt)
1755 virtio_xmit_cleanup_packed(vq, nb_pkts - vq->vq_free_cnt,
1758 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1759 struct rte_mbuf *txm = tx_pkts[nb_tx];
1760 int can_push = 0, use_indirect = 0, slots, need;
1762 /* optimize ring usage */
1763 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1764 vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1765 rte_mbuf_refcnt_read(txm) == 1 &&
1766 RTE_MBUF_DIRECT(txm) &&
1767 txm->nb_segs == 1 &&
1768 rte_pktmbuf_headroom(txm) >= hdr_size &&
1769 rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1770 __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1772 else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1773 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1775 /* How many main ring entries are needed to this Tx?
1777 * any_layout => number of segments
1778 * default => number of segments + 1
1780 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1781 need = slots - vq->vq_free_cnt;
1783 /* Positive value indicates it need free vring descriptors */
1784 if (unlikely(need > 0)) {
1785 virtio_xmit_cleanup_packed(vq, need, in_order);
1786 need = slots - vq->vq_free_cnt;
1787 if (unlikely(need > 0)) {
1789 "No free tx descriptors to transmit");
1794 /* Enqueue Packet buffers */
1796 virtqueue_enqueue_xmit_packed_fast(txvq, txm, in_order);
1798 virtqueue_enqueue_xmit_packed(txvq, txm, slots,
1802 virtio_update_packet_stats(&txvq->stats, txm);
1805 txvq->stats.packets += nb_tx;
1807 if (likely(nb_tx)) {
1808 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1809 virtqueue_notify(vq);
1810 PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1818 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1820 struct virtnet_tx *txvq = tx_queue;
1821 struct virtqueue *vq = txvq->vq;
1822 struct virtio_hw *hw = vq->hw;
1823 uint16_t hdr_size = hw->vtnet_hdr_size;
1824 uint16_t nb_used, nb_tx = 0;
1826 if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1829 if (unlikely(nb_pkts < 1))
1832 PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1834 nb_used = virtqueue_nused(vq);
1836 if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1837 virtio_xmit_cleanup(vq, nb_used);
1839 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1840 struct rte_mbuf *txm = tx_pkts[nb_tx];
1841 int can_push = 0, use_indirect = 0, slots, need;
1843 /* optimize ring usage */
1844 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1845 vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1846 rte_mbuf_refcnt_read(txm) == 1 &&
1847 RTE_MBUF_DIRECT(txm) &&
1848 txm->nb_segs == 1 &&
1849 rte_pktmbuf_headroom(txm) >= hdr_size &&
1850 rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1851 __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1853 else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1854 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1857 /* How many main ring entries are needed to this Tx?
1858 * any_layout => number of segments
1860 * default => number of segments + 1
1862 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1863 need = slots - vq->vq_free_cnt;
1865 /* Positive value indicates it need free vring descriptors */
1866 if (unlikely(need > 0)) {
1867 nb_used = virtqueue_nused(vq);
1869 need = RTE_MIN(need, (int)nb_used);
1871 virtio_xmit_cleanup(vq, need);
1872 need = slots - vq->vq_free_cnt;
1873 if (unlikely(need > 0)) {
1875 "No free tx descriptors to transmit");
1880 /* Enqueue Packet buffers */
1881 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
1884 virtio_update_packet_stats(&txvq->stats, txm);
1887 txvq->stats.packets += nb_tx;
1889 if (likely(nb_tx)) {
1890 vq_update_avail_idx(vq);
1892 if (unlikely(virtqueue_kick_prepare(vq))) {
1893 virtqueue_notify(vq);
1894 PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1901 static __rte_always_inline int
1902 virtio_xmit_try_cleanup_inorder(struct virtqueue *vq, uint16_t need)
1904 uint16_t nb_used, nb_clean, nb_descs;
1906 nb_descs = vq->vq_free_cnt + need;
1907 nb_used = virtqueue_nused(vq);
1908 nb_clean = RTE_MIN(need, (int)nb_used);
1910 virtio_xmit_cleanup_inorder(vq, nb_clean);
1912 return nb_descs - vq->vq_free_cnt;
1916 virtio_xmit_pkts_inorder(void *tx_queue,
1917 struct rte_mbuf **tx_pkts,
1920 struct virtnet_tx *txvq = tx_queue;
1921 struct virtqueue *vq = txvq->vq;
1922 struct virtio_hw *hw = vq->hw;
1923 uint16_t hdr_size = hw->vtnet_hdr_size;
1924 uint16_t nb_used, nb_tx = 0, nb_inorder_pkts = 0;
1925 struct rte_mbuf *inorder_pkts[nb_pkts];
1928 if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1931 if (unlikely(nb_pkts < 1))
1935 PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1936 nb_used = virtqueue_nused(vq);
1938 if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1939 virtio_xmit_cleanup_inorder(vq, nb_used);
1941 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1942 struct rte_mbuf *txm = tx_pkts[nb_tx];
1945 /* optimize ring usage */
1946 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1947 vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1948 rte_mbuf_refcnt_read(txm) == 1 &&
1949 RTE_MBUF_DIRECT(txm) &&
1950 txm->nb_segs == 1 &&
1951 rte_pktmbuf_headroom(txm) >= hdr_size &&
1952 rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1953 __alignof__(struct virtio_net_hdr_mrg_rxbuf))) {
1954 inorder_pkts[nb_inorder_pkts] = txm;
1960 if (nb_inorder_pkts) {
1961 need = nb_inorder_pkts - vq->vq_free_cnt;
1962 if (unlikely(need > 0)) {
1963 need = virtio_xmit_try_cleanup_inorder(vq,
1965 if (unlikely(need > 0)) {
1967 "No free tx descriptors to "
1972 virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
1974 nb_inorder_pkts = 0;
1977 slots = txm->nb_segs + 1;
1978 need = slots - vq->vq_free_cnt;
1979 if (unlikely(need > 0)) {
1980 need = virtio_xmit_try_cleanup_inorder(vq, slots);
1982 if (unlikely(need > 0)) {
1984 "No free tx descriptors to transmit");
1988 /* Enqueue Packet buffers */
1989 virtqueue_enqueue_xmit(txvq, txm, slots, 0, 0, 1);
1991 virtio_update_packet_stats(&txvq->stats, txm);
1994 /* Transmit all inorder packets */
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 transmit");
2003 nb_inorder_pkts = vq->vq_free_cnt;
2008 virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2012 txvq->stats.packets += nb_tx;
2014 if (likely(nb_tx)) {
2015 vq_update_avail_idx(vq);
2017 if (unlikely(virtqueue_kick_prepare(vq))) {
2018 virtqueue_notify(vq);
2019 PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2028 #ifndef CC_AVX512_SUPPORT
2030 virtio_recv_pkts_packed_vec(void *rx_queue __rte_unused,
2031 struct rte_mbuf **rx_pkts __rte_unused,
2032 uint16_t nb_pkts __rte_unused)
2038 virtio_xmit_pkts_packed_vec(void *tx_queue __rte_unused,
2039 struct rte_mbuf **tx_pkts __rte_unused,
2040 uint16_t nb_pkts __rte_unused)
2044 #endif /* ifndef CC_AVX512_SUPPORT */