xref: /dpdk/drivers/net/virtio/virtio_rxtx.c (revision 9897896ec78b4568ff60ed1dab27fd01e15c96a4)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2010-2014 Intel Corporation
3  */
4 
5 #include <stdint.h>
6 #include <stdio.h>
7 #include <stdlib.h>
8 #include <string.h>
9 #include <errno.h>
10 
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>
16 #include <rte_mbuf.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>
23 #include <rte_net.h>
24 #include <rte_ip.h>
25 #include <rte_udp.h>
26 #include <rte_tcp.h>
27 
28 #include "virtio_logs.h"
29 #include "virtio_ethdev.h"
30 #include "virtio.h"
31 #include "virtqueue.h"
32 #include "virtio_rxtx.h"
33 #include "virtio_rxtx_simple.h"
34 #include "virtio_ring.h"
35 
36 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
37 #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
38 #else
39 #define  VIRTIO_DUMP_PACKET(m, len) do { } while (0)
40 #endif
41 
42 int
43 virtio_dev_rx_queue_done(void *rxq, uint16_t offset)
44 {
45 	struct virtnet_rx *rxvq = rxq;
46 	struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
47 
48 	return virtqueue_nused(vq) >= offset;
49 }
50 
51 void
52 vq_ring_free_inorder(struct virtqueue *vq, uint16_t desc_idx, uint16_t num)
53 {
54 	vq->vq_free_cnt += num;
55 	vq->vq_desc_tail_idx = desc_idx & (vq->vq_nentries - 1);
56 }
57 
58 void
59 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
60 {
61 	struct vring_desc *dp, *dp_tail;
62 	struct vq_desc_extra *dxp;
63 	uint16_t desc_idx_last = desc_idx;
64 
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];
72 		}
73 	}
74 	dxp->ndescs = 0;
75 
76 	/*
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).
80 	 */
81 	if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
82 		vq->vq_desc_head_idx = desc_idx;
83 	} else {
84 		dp_tail = &vq->vq_split.ring.desc[vq->vq_desc_tail_idx];
85 		dp_tail->next = desc_idx;
86 	}
87 
88 	vq->vq_desc_tail_idx = desc_idx_last;
89 	dp->next = VQ_RING_DESC_CHAIN_END;
90 }
91 
92 void
93 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
94 {
95 	uint32_t s = mbuf->pkt_len;
96 	struct rte_ether_addr *ea;
97 
98 	stats->bytes += s;
99 
100 	if (s == 64) {
101 		stats->size_bins[1]++;
102 	} else if (s > 64 && s < 1024) {
103 		uint32_t bin;
104 
105 		/* count zeros, and offset into correct bin */
106 		bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
107 		stats->size_bins[bin]++;
108 	} else {
109 		if (s < 64)
110 			stats->size_bins[0]++;
111 		else if (s < 1519)
112 			stats->size_bins[6]++;
113 		else
114 			stats->size_bins[7]++;
115 	}
116 
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))
120 			stats->broadcast++;
121 		else
122 			stats->multicast++;
123 	}
124 }
125 
126 static inline void
127 virtio_rx_stats_updated(struct virtnet_rx *rxvq, struct rte_mbuf *m)
128 {
129 	VIRTIO_DUMP_PACKET(m, m->data_len);
130 
131 	virtio_update_packet_stats(&rxvq->stats, m);
132 }
133 
134 static uint16_t
135 virtqueue_dequeue_burst_rx_packed(struct virtqueue *vq,
136 				  struct rte_mbuf **rx_pkts,
137 				  uint32_t *len,
138 				  uint16_t num)
139 {
140 	struct rte_mbuf *cookie;
141 	uint16_t used_idx;
142 	uint16_t id;
143 	struct vring_packed_desc *desc;
144 	uint16_t i;
145 
146 	desc = vq->vq_packed.ring.desc;
147 
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.
152 		 */
153 		if (!desc_is_used(&desc[used_idx], vq))
154 			return i;
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);
161 			break;
162 		}
163 		rte_prefetch0(cookie);
164 		rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
165 		rx_pkts[i] = cookie;
166 
167 		vq->vq_free_cnt++;
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;
172 		}
173 	}
174 
175 	return i;
176 }
177 
178 static uint16_t
179 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
180 			   uint32_t *len, uint16_t num)
181 {
182 	struct vring_used_elem *uep;
183 	struct rte_mbuf *cookie;
184 	uint16_t used_idx, desc_idx;
185 	uint16_t i;
186 
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;
192 		len[i] = uep->len;
193 		cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
194 
195 		if (unlikely(cookie == NULL)) {
196 			PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
197 				vq->vq_used_cons_idx);
198 			break;
199 		}
200 
201 		rte_prefetch0(cookie);
202 		rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
203 		rx_pkts[i]  = cookie;
204 		vq->vq_used_cons_idx++;
205 		vq_ring_free_chain(vq, desc_idx);
206 		vq->vq_descx[desc_idx].cookie = NULL;
207 	}
208 
209 	return i;
210 }
211 
212 static uint16_t
213 virtqueue_dequeue_rx_inorder(struct virtqueue *vq,
214 			struct rte_mbuf **rx_pkts,
215 			uint32_t *len,
216 			uint16_t num)
217 {
218 	struct vring_used_elem *uep;
219 	struct rte_mbuf *cookie;
220 	uint16_t used_idx = 0;
221 	uint16_t i;
222 
223 	if (unlikely(num == 0))
224 		return 0;
225 
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];
230 		len[i] = uep->len;
231 		cookie = (struct rte_mbuf *)vq->vq_descx[used_idx].cookie;
232 
233 		if (unlikely(cookie == NULL)) {
234 			PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
235 				vq->vq_used_cons_idx);
236 			break;
237 		}
238 
239 		rte_prefetch0(cookie);
240 		rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
241 		rx_pkts[i]  = cookie;
242 		vq->vq_used_cons_idx++;
243 		vq->vq_descx[used_idx].cookie = NULL;
244 	}
245 
246 	vq_ring_free_inorder(vq, used_idx, i);
247 	return i;
248 }
249 
250 static inline int
251 virtqueue_enqueue_refill_inorder(struct virtqueue *vq,
252 			struct rte_mbuf **cookies,
253 			uint16_t num)
254 {
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;
259 
260 	if (unlikely(vq->vq_free_cnt == 0))
261 		return -ENOSPC;
262 	if (unlikely(vq->vq_free_cnt < num))
263 		return -EMSGSIZE;
264 
265 	head_idx = vq->vq_desc_head_idx & (vq->vq_nentries - 1);
266 	start_dp = vq->vq_split.ring.desc;
267 
268 	while (i < num) {
269 		idx = head_idx & (vq->vq_nentries - 1);
270 		dxp = &vq->vq_descx[idx];
271 		dxp->cookie = (void *)cookies[i];
272 		dxp->ndescs = 1;
273 
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;
279 
280 		vq_update_avail_ring(vq, idx);
281 		head_idx++;
282 		i++;
283 	}
284 
285 	vq->vq_desc_head_idx += num;
286 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
287 	return 0;
288 }
289 
290 static inline int
291 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf **cookie,
292 				uint16_t num)
293 {
294 	struct vq_desc_extra *dxp;
295 	struct virtio_hw *hw = vq->hw;
296 	struct vring_desc *start_dp = vq->vq_split.ring.desc;
297 	uint16_t idx, i;
298 
299 	if (unlikely(vq->vq_free_cnt == 0))
300 		return -ENOSPC;
301 	if (unlikely(vq->vq_free_cnt < num))
302 		return -EMSGSIZE;
303 
304 	if (unlikely(vq->vq_desc_head_idx >= vq->vq_nentries))
305 		return -EFAULT;
306 
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];
311 		dxp->ndescs = 1;
312 
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;
322 			break;
323 		}
324 	}
325 
326 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
327 
328 	return 0;
329 }
330 
331 static inline void
332 virtqueue_refill_single_packed(struct virtqueue *vq,
333 			       struct vring_packed_desc *dp,
334 			       struct rte_mbuf *cookie)
335 {
336 	uint16_t flags = vq->vq_packed.cached_flags;
337 	struct virtio_hw *hw = vq->hw;
338 
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;
343 
344 	virtqueue_store_flags_packed(dp, flags,
345 				     hw->weak_barriers);
346 
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;
352 	}
353 }
354 
355 static inline int
356 virtqueue_enqueue_recv_refill_packed_init(struct virtqueue *vq,
357 				     struct rte_mbuf **cookie, uint16_t num)
358 {
359 	struct vring_packed_desc *start_dp = vq->vq_packed.ring.desc;
360 	struct vq_desc_extra *dxp;
361 	uint16_t idx;
362 	int i;
363 
364 	if (unlikely(vq->vq_free_cnt == 0))
365 		return -ENOSPC;
366 	if (unlikely(vq->vq_free_cnt < num))
367 		return -EMSGSIZE;
368 
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];
373 		dxp->ndescs = 1;
374 
375 		virtqueue_refill_single_packed(vq, &start_dp[idx], cookie[i]);
376 	}
377 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
378 	return 0;
379 }
380 
381 static inline int
382 virtqueue_enqueue_recv_refill_packed(struct virtqueue *vq,
383 				     struct rte_mbuf **cookie, uint16_t num)
384 {
385 	struct vring_packed_desc *start_dp = vq->vq_packed.ring.desc;
386 	struct vq_desc_extra *dxp;
387 	uint16_t idx, did;
388 	int i;
389 
390 	if (unlikely(vq->vq_free_cnt == 0))
391 		return -ENOSPC;
392 	if (unlikely(vq->vq_free_cnt < num))
393 		return -EMSGSIZE;
394 
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];
400 		dxp->ndescs = 1;
401 
402 		virtqueue_refill_single_packed(vq, &start_dp[idx], cookie[i]);
403 	}
404 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
405 	return 0;
406 }
407 
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
410  * expected.
411  */
412 static void
413 virtio_tso_fix_cksum(struct rte_mbuf *m)
414 {
415 	/* common case: header is not fragmented */
416 	if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
417 			m->l4_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;
422 		uint32_t tmp;
423 
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) -
432 				m->l3_len);
433 		} else {
434 			ip6h = (struct rte_ipv6_hdr *)iph;
435 			ip_paylen = ip6h->payload_len;
436 		}
437 
438 		/* calculate the new phdr checksum not including ip_paylen */
439 		prev_cksum = th->cksum;
440 		tmp = prev_cksum;
441 		tmp += ip_paylen;
442 		tmp = (tmp & 0xffff) + (tmp >> 16);
443 		new_cksum = tmp;
444 
445 		/* replace it in the packet */
446 		th->cksum = new_cksum;
447 	}
448 }
449 
450 
451 
452 
453 static inline void
454 virtqueue_enqueue_xmit_inorder(struct virtnet_tx *txvq,
455 			struct rte_mbuf **cookies,
456 			uint16_t num)
457 {
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;
462 	uint16_t idx;
463 	int16_t head_size = vq->hw->vtnet_hdr_size;
464 	uint16_t i = 0;
465 
466 	idx = vq->vq_desc_head_idx;
467 	start_dp = vq->vq_split.ring.desc;
468 
469 	while (i < num) {
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];
473 		dxp->ndescs = 1;
474 		virtio_update_packet_stats(&txvq->stats, cookies[i]);
475 
476 		hdr = rte_pktmbuf_mtod_offset(cookies[i],
477 				struct virtio_net_hdr *, -head_size);
478 
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);
482 		else
483 			virtqueue_xmit_offload(hdr, cookies[i]);
484 
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;
488 
489 
490 		vq_update_avail_ring(vq, idx);
491 
492 		idx++;
493 		i++;
494 	};
495 
496 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
497 	vq->vq_desc_head_idx = idx & (vq->vq_nentries - 1);
498 }
499 
500 static inline void
501 virtqueue_enqueue_xmit_packed_fast(struct virtnet_tx *txvq,
502 				   struct rte_mbuf *cookie,
503 				   int in_order)
504 {
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;
511 
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];
515 
516 	dxp = &vq->vq_descx[id];
517 	dxp->ndescs = 1;
518 	dxp->cookie = cookie;
519 
520 	flags = vq->vq_packed.cached_flags;
521 
522 	/* prepend cannot fail, checked by caller */
523 	hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
524 				      -head_size);
525 
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);
529 	else
530 		virtqueue_xmit_offload(hdr, cookie);
531 
532 	dp->addr = rte_mbuf_data_iova(cookie) - head_size;
533 	dp->len  = cookie->data_len + head_size;
534 	dp->id   = id;
535 
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;
539 	}
540 
541 	vq->vq_free_cnt--;
542 
543 	if (!in_order) {
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;
547 	}
548 
549 	virtqueue_store_flags_packed(dp, flags, vq->hw->weak_barriers);
550 }
551 
552 static inline void
553 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
554 			uint16_t needed, int use_indirect, int can_push,
555 			int in_order)
556 {
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;
566 
567 	head_idx = vq->vq_desc_head_idx;
568 	idx = head_idx;
569 	if (in_order)
570 		dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
571 	else
572 		dxp = &vq->vq_descx[idx];
573 	dxp->cookie = (void *)cookie;
574 	dxp->ndescs = needed;
575 
576 	start_dp = vq->vq_split.ring.desc;
577 
578 	if (can_push) {
579 		/* prepend cannot fail, checked by caller */
580 		hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
581 					      -head_size);
582 		prepend_header = true;
583 
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.
590 		 *
591 		 * the first slot in indirect ring is already preset
592 		 * to point to the header in reserved region
593 		 */
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;
599 
600 		/* loop below will fill in rest of the indirect elements */
601 		start_dp = txr[idx].tx_indir;
602 		idx = 1;
603 	} else {
604 		/* setup first tx ring slot to point to header
605 		 * stored in reserved region.
606 		 */
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;
612 
613 		idx = start_dp[idx].next;
614 	}
615 
616 	if (vq->hw->has_tx_offload)
617 		virtqueue_xmit_offload(hdr, cookie);
618 
619 	do {
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;
626 		}
627 		start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
628 		idx = start_dp[idx].next;
629 	} while ((cookie = cookie->next) != NULL);
630 
631 	if (use_indirect)
632 		idx = vq->vq_split.ring.desc[head_idx].next;
633 
634 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
635 
636 	vq->vq_desc_head_idx = idx;
637 	vq_update_avail_ring(vq, head_idx);
638 
639 	if (!in_order) {
640 		if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
641 			vq->vq_desc_tail_idx = idx;
642 	}
643 }
644 
645 void
646 virtio_dev_cq_start(struct rte_eth_dev *dev)
647 {
648 	struct virtio_hw *hw = dev->data->dev_private;
649 
650 	if (hw->cvq) {
651 		rte_spinlock_init(&hw->cvq->lock);
652 		VIRTQUEUE_DUMP(virtnet_cq_to_vq(hw->cvq));
653 	}
654 }
655 
656 int
657 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
658 			uint16_t queue_idx,
659 			uint16_t nb_desc,
660 			unsigned int socket_id __rte_unused,
661 			const struct rte_eth_rxconf *rx_conf,
662 			struct rte_mempool *mp)
663 {
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;
669 	uint16_t buf_size;
670 	const char *error;
671 
672 	PMD_INIT_FUNC_TRACE();
673 
674 	if (rx_conf->rx_deferred_start) {
675 		PMD_INIT_LOG(ERR, "Rx deferred start is not supported");
676 		return -EINVAL;
677 	}
678 
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",
683 			     queue_idx, error);
684 		return -EINVAL;
685 	}
686 
687 	rx_free_thresh = rx_conf->rx_free_thresh;
688 	if (rx_free_thresh == 0)
689 		rx_free_thresh =
690 			RTE_MIN(vq->vq_nentries / 4, DEFAULT_RX_FREE_THRESH);
691 
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);
696 		return -EINVAL;
697 	}
698 
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",
703 			vq->vq_nentries,
704 			rx_free_thresh, dev->data->port_id, queue_idx);
705 		return -EINVAL;
706 	}
707 	vq->vq_free_thresh = rx_free_thresh;
708 
709 	if (nb_desc > vq->vq_nentries)
710 		nb_desc = vq->vq_nentries;
711 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
712 
713 	rxvq = &vq->rxq;
714 	rxvq->queue_id = queue_idx;
715 	rxvq->mpool = mp;
716 	dev->data->rx_queues[queue_idx] = rxvq;
717 
718 	return 0;
719 }
720 
721 int
722 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
723 {
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;
728 	struct rte_mbuf *m;
729 	uint16_t desc_idx;
730 	int error, nbufs, i;
731 	bool in_order = virtio_with_feature(hw, VIRTIO_F_IN_ORDER);
732 
733 	PMD_INIT_FUNC_TRACE();
734 
735 	/* Allocate blank mbufs for the each rx descriptor */
736 	nbufs = 0;
737 
738 	if (hw->use_vec_rx && !virtio_with_packed_queue(hw)) {
739 		for (desc_idx = 0; desc_idx < vq->vq_nentries;
740 		     desc_idx++) {
741 			vq->vq_split.ring.avail->ring[desc_idx] = desc_idx;
742 			vq->vq_split.ring.desc[desc_idx].flags =
743 				VRING_DESC_F_WRITE;
744 		}
745 
746 		virtio_rxq_vec_setup(rxvq);
747 	}
748 
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;
752 
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;
757 		}
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];
762 
763 			if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, pkts,
764 				free_cnt)) {
765 				error = virtqueue_enqueue_refill_inorder(vq,
766 						pkts,
767 						free_cnt);
768 				if (unlikely(error)) {
769 					for (i = 0; i < free_cnt; i++)
770 						rte_pktmbuf_free(pkts[i]);
771 				} else {
772 					nbufs += free_cnt;
773 				}
774 			}
775 
776 			vq_update_avail_idx(vq);
777 		}
778 	} else {
779 		while (!virtqueue_full(vq)) {
780 			m = rte_mbuf_raw_alloc(rxvq->mpool);
781 			if (m == NULL)
782 				break;
783 
784 			/* Enqueue allocated buffers */
785 			if (virtio_with_packed_queue(vq->hw))
786 				error = virtqueue_enqueue_recv_refill_packed_init(vq,
787 						&m, 1);
788 			else
789 				error = virtqueue_enqueue_recv_refill(vq,
790 						&m, 1);
791 			if (error) {
792 				rte_pktmbuf_free(m);
793 				break;
794 			}
795 			nbufs++;
796 		}
797 
798 		if (!virtio_with_packed_queue(vq->hw))
799 			vq_update_avail_idx(vq);
800 	}
801 
802 	PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
803 
804 	VIRTQUEUE_DUMP(vq);
805 
806 	return 0;
807 }
808 
809 /*
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
815  */
816 int
817 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
818 			uint16_t queue_idx,
819 			uint16_t nb_desc,
820 			unsigned int socket_id __rte_unused,
821 			const struct rte_eth_txconf *tx_conf)
822 {
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;
828 
829 	PMD_INIT_FUNC_TRACE();
830 
831 	if (tx_conf->tx_deferred_start) {
832 		PMD_INIT_LOG(ERR, "Tx deferred start is not supported");
833 		return -EINVAL;
834 	}
835 
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);
839 
840 	txvq = &vq->txq;
841 	txvq->queue_id = queue_idx;
842 
843 	tx_free_thresh = tx_conf->tx_free_thresh;
844 	if (tx_free_thresh == 0)
845 		tx_free_thresh =
846 			RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
847 
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",
852 			vq->vq_nentries - 3,
853 			tx_free_thresh, dev->data->port_id, queue_idx);
854 		return -EINVAL;
855 	}
856 
857 	vq->vq_free_thresh = tx_free_thresh;
858 
859 	dev->data->tx_queues[queue_idx] = txvq;
860 	return 0;
861 }
862 
863 int
864 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
865 				uint16_t queue_idx)
866 {
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];
870 
871 	PMD_INIT_FUNC_TRACE();
872 
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;
876 	}
877 
878 	VIRTQUEUE_DUMP(vq);
879 
880 	return 0;
881 }
882 
883 static inline void
884 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
885 {
886 	int error;
887 	/*
888 	 * Requeue the discarded mbuf. This should always be
889 	 * successful since it was just dequeued.
890 	 */
891 	if (virtio_with_packed_queue(vq->hw))
892 		error = virtqueue_enqueue_recv_refill_packed(vq, &m, 1);
893 	else
894 		error = virtqueue_enqueue_recv_refill(vq, &m, 1);
895 
896 	if (unlikely(error)) {
897 		PMD_DRV_LOG(ERR, "cannot requeue discarded mbuf");
898 		rte_pktmbuf_free(m);
899 	}
900 }
901 
902 static inline void
903 virtio_discard_rxbuf_inorder(struct virtqueue *vq, struct rte_mbuf *m)
904 {
905 	int error;
906 
907 	error = virtqueue_enqueue_refill_inorder(vq, &m, 1);
908 	if (unlikely(error)) {
909 		PMD_DRV_LOG(ERR, "cannot requeue discarded mbuf");
910 		rte_pktmbuf_free(m);
911 	}
912 }
913 
914 /* Optionally fill offload information in structure */
915 static inline int
916 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
917 {
918 	struct rte_net_hdr_lens hdr_lens;
919 	uint32_t hdrlen, ptype;
920 	int l4_supported = 0;
921 
922 	/* nothing to do */
923 	if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
924 		return 0;
925 
926 	m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
927 
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)
933 		l4_supported = 1;
934 
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;
939 		} else {
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.
945 			 */
946 			uint16_t csum = 0, off;
947 
948 			if (rte_raw_cksum_mbuf(m, hdr->csum_start,
949 				rte_pktmbuf_pkt_len(m) - hdr->csum_start,
950 				&csum) < 0)
951 				return -EINVAL;
952 			if (likely(csum != 0xffff))
953 				csum = ~csum;
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 *,
957 					off) = csum;
958 		}
959 	} else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
960 		m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
961 	}
962 
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)) {
968 			return -EINVAL;
969 		}
970 
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;
978 				break;
979 			default:
980 				return -EINVAL;
981 		}
982 	}
983 
984 	return 0;
985 }
986 
987 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
988 uint16_t
989 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
990 {
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];
998 	int error;
999 	uint32_t i, nb_enqueued;
1000 	uint32_t hdr_size;
1001 	struct virtio_net_hdr *hdr;
1002 
1003 	nb_rx = 0;
1004 	if (unlikely(hw->started == 0))
1005 		return nb_rx;
1006 
1007 	nb_used = virtqueue_nused(vq);
1008 
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);
1014 
1015 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1016 	PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
1017 
1018 	nb_enqueued = 0;
1019 	hdr_size = hw->vtnet_hdr_size;
1020 
1021 	for (i = 0; i < num ; i++) {
1022 		rxm = rcv_pkts[i];
1023 
1024 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1025 
1026 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1027 			PMD_RX_LOG(ERR, "Packet drop");
1028 			nb_enqueued++;
1029 			virtio_discard_rxbuf(vq, rxm);
1030 			rxvq->stats.errors++;
1031 			continue;
1032 		}
1033 
1034 		rxm->port = rxvq->port_id;
1035 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1036 		rxm->ol_flags = 0;
1037 		rxm->vlan_tci = 0;
1038 
1039 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1040 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1041 
1042 		hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1043 			RTE_PKTMBUF_HEADROOM - hdr_size);
1044 
1045 		if (hw->vlan_strip)
1046 			rte_vlan_strip(rxm);
1047 
1048 		if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1049 			virtio_discard_rxbuf(vq, rxm);
1050 			rxvq->stats.errors++;
1051 			continue;
1052 		}
1053 
1054 		virtio_rx_stats_updated(rxvq, rxm);
1055 
1056 		rx_pkts[nb_rx++] = rxm;
1057 	}
1058 
1059 	rxvq->stats.packets += nb_rx;
1060 
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];
1065 
1066 		if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1067 						free_cnt) == 0)) {
1068 			error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1069 					free_cnt);
1070 			if (unlikely(error)) {
1071 				for (i = 0; i < free_cnt; i++)
1072 					rte_pktmbuf_free(new_pkts[i]);
1073 			}
1074 			nb_enqueued += free_cnt;
1075 		} else {
1076 			struct rte_eth_dev *dev =
1077 				&rte_eth_devices[rxvq->port_id];
1078 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1079 		}
1080 	}
1081 
1082 	if (likely(nb_enqueued)) {
1083 		vq_update_avail_idx(vq);
1084 
1085 		if (unlikely(virtqueue_kick_prepare(vq))) {
1086 			virtqueue_notify(vq);
1087 			PMD_RX_LOG(DEBUG, "Notified");
1088 		}
1089 	}
1090 
1091 	return nb_rx;
1092 }
1093 
1094 uint16_t
1095 virtio_recv_pkts_packed(void *rx_queue, struct rte_mbuf **rx_pkts,
1096 			uint16_t nb_pkts)
1097 {
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];
1105 	int error;
1106 	uint32_t i, nb_enqueued;
1107 	uint32_t hdr_size;
1108 	struct virtio_net_hdr *hdr;
1109 
1110 	nb_rx = 0;
1111 	if (unlikely(hw->started == 0))
1112 		return nb_rx;
1113 
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);
1117 
1118 	num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1119 	PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1120 
1121 	nb_enqueued = 0;
1122 	hdr_size = hw->vtnet_hdr_size;
1123 
1124 	for (i = 0; i < num; i++) {
1125 		rxm = rcv_pkts[i];
1126 
1127 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1128 
1129 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1130 			PMD_RX_LOG(ERR, "Packet drop");
1131 			nb_enqueued++;
1132 			virtio_discard_rxbuf(vq, rxm);
1133 			rxvq->stats.errors++;
1134 			continue;
1135 		}
1136 
1137 		rxm->port = rxvq->port_id;
1138 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1139 		rxm->ol_flags = 0;
1140 		rxm->vlan_tci = 0;
1141 
1142 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1143 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1144 
1145 		hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1146 			RTE_PKTMBUF_HEADROOM - hdr_size);
1147 
1148 		if (hw->vlan_strip)
1149 			rte_vlan_strip(rxm);
1150 
1151 		if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1152 			virtio_discard_rxbuf(vq, rxm);
1153 			rxvq->stats.errors++;
1154 			continue;
1155 		}
1156 
1157 		virtio_rx_stats_updated(rxvq, rxm);
1158 
1159 		rx_pkts[nb_rx++] = rxm;
1160 	}
1161 
1162 	rxvq->stats.packets += nb_rx;
1163 
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];
1168 
1169 		if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1170 						free_cnt) == 0)) {
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]);
1176 			}
1177 			nb_enqueued += free_cnt;
1178 		} else {
1179 			struct rte_eth_dev *dev =
1180 				&rte_eth_devices[rxvq->port_id];
1181 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1182 		}
1183 	}
1184 
1185 	if (likely(nb_enqueued)) {
1186 		if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1187 			virtqueue_notify(vq);
1188 			PMD_RX_LOG(DEBUG, "Notified");
1189 		}
1190 	}
1191 
1192 	return nb_rx;
1193 }
1194 
1195 
1196 uint16_t
1197 virtio_recv_pkts_inorder(void *rx_queue,
1198 			struct rte_mbuf **rx_pkts,
1199 			uint16_t nb_pkts)
1200 {
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];
1209 	int error;
1210 	uint32_t nb_enqueued;
1211 	uint32_t seg_num;
1212 	uint32_t seg_res;
1213 	uint32_t hdr_size;
1214 	int32_t i;
1215 
1216 	nb_rx = 0;
1217 	if (unlikely(hw->started == 0))
1218 		return nb_rx;
1219 
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);
1223 
1224 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1225 
1226 	nb_enqueued = 0;
1227 	seg_num = 1;
1228 	seg_res = 0;
1229 	hdr_size = hw->vtnet_hdr_size;
1230 
1231 	num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len, nb_used);
1232 
1233 	for (i = 0; i < num; i++) {
1234 		struct virtio_net_hdr_mrg_rxbuf *header;
1235 
1236 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1237 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1238 
1239 		rxm = rcv_pkts[i];
1240 
1241 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1242 			PMD_RX_LOG(ERR, "Packet drop");
1243 			nb_enqueued++;
1244 			virtio_discard_rxbuf_inorder(vq, rxm);
1245 			rxvq->stats.errors++;
1246 			continue;
1247 		}
1248 
1249 		header = (struct virtio_net_hdr_mrg_rxbuf *)
1250 			 ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1251 			 - hdr_size);
1252 
1253 		if (virtio_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
1254 			seg_num = header->num_buffers;
1255 			if (seg_num == 0)
1256 				seg_num = 1;
1257 		} else {
1258 			seg_num = 1;
1259 		}
1260 
1261 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1262 		rxm->nb_segs = seg_num;
1263 		rxm->ol_flags = 0;
1264 		rxm->vlan_tci = 0;
1265 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1266 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1267 
1268 		rxm->port = rxvq->port_id;
1269 
1270 		rx_pkts[nb_rx] = rxm;
1271 		prev = rxm;
1272 
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++;
1277 			continue;
1278 		}
1279 
1280 		if (hw->vlan_strip)
1281 			rte_vlan_strip(rx_pkts[nb_rx]);
1282 
1283 		seg_res = seg_num - 1;
1284 
1285 		/* Merge remaining segments */
1286 		while (seg_res != 0 && i < (num - 1)) {
1287 			i++;
1288 
1289 			rxm = rcv_pkts[i];
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]);
1293 
1294 			rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1295 
1296 			prev->next = rxm;
1297 			prev = rxm;
1298 			seg_res -= 1;
1299 		}
1300 
1301 		if (!seg_res) {
1302 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1303 			nb_rx++;
1304 		}
1305 	}
1306 
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);
1311 
1312 		if (likely(virtqueue_nused(vq) >= rcv_cnt)) {
1313 			num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len,
1314 							   rcv_cnt);
1315 			uint16_t extra_idx = 0;
1316 
1317 			rcv_cnt = num;
1318 			while (extra_idx < rcv_cnt) {
1319 				rxm = rcv_pkts[extra_idx];
1320 				rxm->data_off =
1321 					RTE_PKTMBUF_HEADROOM - hdr_size;
1322 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
1323 				rxm->data_len = (uint16_t)(len[extra_idx]);
1324 				prev->next = rxm;
1325 				prev = rxm;
1326 				rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1327 				extra_idx += 1;
1328 			};
1329 			seg_res -= rcv_cnt;
1330 
1331 			if (!seg_res) {
1332 				virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1333 				nb_rx++;
1334 			}
1335 		} else {
1336 			PMD_RX_LOG(ERR,
1337 					"No enough segments for packet.");
1338 			rte_pktmbuf_free(rx_pkts[nb_rx]);
1339 			rxvq->stats.errors++;
1340 			break;
1341 		}
1342 	}
1343 
1344 	rxvq->stats.packets += nb_rx;
1345 
1346 	/* Allocate new mbuf for the used descriptor */
1347 
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];
1352 
1353 		if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1354 			error = virtqueue_enqueue_refill_inorder(vq, new_pkts,
1355 					free_cnt);
1356 			if (unlikely(error)) {
1357 				for (i = 0; i < free_cnt; i++)
1358 					rte_pktmbuf_free(new_pkts[i]);
1359 			}
1360 			nb_enqueued += free_cnt;
1361 		} else {
1362 			struct rte_eth_dev *dev =
1363 				&rte_eth_devices[rxvq->port_id];
1364 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1365 		}
1366 	}
1367 
1368 	if (likely(nb_enqueued)) {
1369 		vq_update_avail_idx(vq);
1370 
1371 		if (unlikely(virtqueue_kick_prepare(vq))) {
1372 			virtqueue_notify(vq);
1373 			PMD_RX_LOG(DEBUG, "Notified");
1374 		}
1375 	}
1376 
1377 	return nb_rx;
1378 }
1379 
1380 uint16_t
1381 virtio_recv_mergeable_pkts(void *rx_queue,
1382 			struct rte_mbuf **rx_pkts,
1383 			uint16_t nb_pkts)
1384 {
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];
1393 	int error;
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;
1398 	int32_t i;
1399 
1400 	if (unlikely(hw->started == 0))
1401 		return nb_rx;
1402 
1403 	nb_used = virtqueue_nused(vq);
1404 
1405 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1406 
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);
1413 
1414 
1415 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1416 
1417 	for (i = 0; i < num; i++) {
1418 		struct virtio_net_hdr_mrg_rxbuf *header;
1419 
1420 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1421 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1422 
1423 		rxm = rcv_pkts[i];
1424 
1425 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1426 			PMD_RX_LOG(ERR, "Packet drop");
1427 			nb_enqueued++;
1428 			virtio_discard_rxbuf(vq, rxm);
1429 			rxvq->stats.errors++;
1430 			continue;
1431 		}
1432 
1433 		header = (struct virtio_net_hdr_mrg_rxbuf *)
1434 			 ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1435 			 - hdr_size);
1436 		seg_num = header->num_buffers;
1437 		if (seg_num == 0)
1438 			seg_num = 1;
1439 
1440 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1441 		rxm->nb_segs = seg_num;
1442 		rxm->ol_flags = 0;
1443 		rxm->vlan_tci = 0;
1444 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1445 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1446 
1447 		rxm->port = rxvq->port_id;
1448 
1449 		rx_pkts[nb_rx] = rxm;
1450 		prev = rxm;
1451 
1452 		if (hw->has_rx_offload &&
1453 				virtio_rx_offload(rxm, &header->hdr) < 0) {
1454 			virtio_discard_rxbuf(vq, rxm);
1455 			rxvq->stats.errors++;
1456 			continue;
1457 		}
1458 
1459 		if (hw->vlan_strip)
1460 			rte_vlan_strip(rx_pkts[nb_rx]);
1461 
1462 		seg_res = seg_num - 1;
1463 
1464 		/* Merge remaining segments */
1465 		while (seg_res != 0 && i < (num - 1)) {
1466 			i++;
1467 
1468 			rxm = rcv_pkts[i];
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]);
1472 
1473 			rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1474 
1475 			prev->next = rxm;
1476 			prev = rxm;
1477 			seg_res -= 1;
1478 		}
1479 
1480 		if (!seg_res) {
1481 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1482 			nb_rx++;
1483 		}
1484 	}
1485 
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);
1490 
1491 		if (likely(virtqueue_nused(vq) >= rcv_cnt)) {
1492 			num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len,
1493 							   rcv_cnt);
1494 			uint16_t extra_idx = 0;
1495 
1496 			rcv_cnt = num;
1497 			while (extra_idx < rcv_cnt) {
1498 				rxm = rcv_pkts[extra_idx];
1499 				rxm->data_off =
1500 					RTE_PKTMBUF_HEADROOM - hdr_size;
1501 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
1502 				rxm->data_len = (uint16_t)(len[extra_idx]);
1503 				prev->next = rxm;
1504 				prev = rxm;
1505 				rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1506 				extra_idx += 1;
1507 			};
1508 			seg_res -= rcv_cnt;
1509 
1510 			if (!seg_res) {
1511 				virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1512 				nb_rx++;
1513 			}
1514 		} else {
1515 			PMD_RX_LOG(ERR,
1516 					"No enough segments for packet.");
1517 			rte_pktmbuf_free(rx_pkts[nb_rx]);
1518 			rxvq->stats.errors++;
1519 			break;
1520 		}
1521 	}
1522 
1523 	rxvq->stats.packets += nb_rx;
1524 
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];
1530 
1531 		if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1532 			error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1533 					free_cnt);
1534 			if (unlikely(error)) {
1535 				for (i = 0; i < free_cnt; i++)
1536 					rte_pktmbuf_free(new_pkts[i]);
1537 			}
1538 			nb_enqueued += free_cnt;
1539 		} else {
1540 			struct rte_eth_dev *dev =
1541 				&rte_eth_devices[rxvq->port_id];
1542 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1543 		}
1544 	}
1545 
1546 	if (likely(nb_enqueued)) {
1547 		vq_update_avail_idx(vq);
1548 
1549 		if (unlikely(virtqueue_kick_prepare(vq))) {
1550 			virtqueue_notify(vq);
1551 			PMD_RX_LOG(DEBUG, "Notified");
1552 		}
1553 	}
1554 
1555 	return nb_rx;
1556 }
1557 
1558 uint16_t
1559 virtio_recv_mergeable_pkts_packed(void *rx_queue,
1560 			struct rte_mbuf **rx_pkts,
1561 			uint16_t nb_pkts)
1562 {
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;
1575 	int32_t i;
1576 	int error;
1577 
1578 	if (unlikely(hw->started == 0))
1579 		return nb_rx;
1580 
1581 
1582 	num = nb_pkts;
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);
1587 
1588 	num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1589 
1590 	for (i = 0; i < num; i++) {
1591 		struct virtio_net_hdr_mrg_rxbuf *header;
1592 
1593 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1594 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1595 
1596 		rxm = rcv_pkts[i];
1597 
1598 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1599 			PMD_RX_LOG(ERR, "Packet drop");
1600 			nb_enqueued++;
1601 			virtio_discard_rxbuf(vq, rxm);
1602 			rxvq->stats.errors++;
1603 			continue;
1604 		}
1605 
1606 		header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)
1607 			  rxm->buf_addr + RTE_PKTMBUF_HEADROOM - hdr_size);
1608 		seg_num = header->num_buffers;
1609 
1610 		if (seg_num == 0)
1611 			seg_num = 1;
1612 
1613 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1614 		rxm->nb_segs = seg_num;
1615 		rxm->ol_flags = 0;
1616 		rxm->vlan_tci = 0;
1617 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1618 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1619 
1620 		rxm->port = rxvq->port_id;
1621 		rx_pkts[nb_rx] = rxm;
1622 		prev = rxm;
1623 
1624 		if (hw->has_rx_offload &&
1625 				virtio_rx_offload(rxm, &header->hdr) < 0) {
1626 			virtio_discard_rxbuf(vq, rxm);
1627 			rxvq->stats.errors++;
1628 			continue;
1629 		}
1630 
1631 		if (hw->vlan_strip)
1632 			rte_vlan_strip(rx_pkts[nb_rx]);
1633 
1634 		seg_res = seg_num - 1;
1635 
1636 		/* Merge remaining segments */
1637 		while (seg_res != 0 && i < (num - 1)) {
1638 			i++;
1639 
1640 			rxm = rcv_pkts[i];
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]);
1644 
1645 			rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1646 
1647 			prev->next = rxm;
1648 			prev = rxm;
1649 			seg_res -= 1;
1650 		}
1651 
1652 		if (!seg_res) {
1653 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1654 			nb_rx++;
1655 		}
1656 	}
1657 
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;
1663 
1664 		rcv_cnt = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts,
1665 				len, rcv_cnt);
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++;
1670 			break;
1671 		}
1672 
1673 		while (extra_idx < rcv_cnt) {
1674 			rxm = rcv_pkts[extra_idx];
1675 
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]);
1679 
1680 			prev->next = rxm;
1681 			prev = rxm;
1682 			rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1683 			extra_idx += 1;
1684 		}
1685 		seg_res -= rcv_cnt;
1686 		if (!seg_res) {
1687 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1688 			nb_rx++;
1689 		}
1690 	}
1691 
1692 	rxvq->stats.packets += nb_rx;
1693 
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];
1699 
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]);
1706 			}
1707 			nb_enqueued += free_cnt;
1708 		} else {
1709 			struct rte_eth_dev *dev =
1710 				&rte_eth_devices[rxvq->port_id];
1711 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1712 		}
1713 	}
1714 
1715 	if (likely(nb_enqueued)) {
1716 		if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1717 			virtqueue_notify(vq);
1718 			PMD_RX_LOG(DEBUG, "Notified");
1719 		}
1720 	}
1721 
1722 	return nb_rx;
1723 }
1724 
1725 uint16_t
1726 virtio_xmit_pkts_prepare(void *tx_queue __rte_unused, struct rte_mbuf **tx_pkts,
1727 			uint16_t nb_pkts)
1728 {
1729 	uint16_t nb_tx;
1730 	int error;
1731 
1732 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1733 		struct rte_mbuf *m = tx_pkts[nb_tx];
1734 
1735 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
1736 		error = rte_validate_tx_offload(m);
1737 		if (unlikely(error)) {
1738 			rte_errno = -error;
1739 			break;
1740 		}
1741 #endif
1742 
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
1748 			 */
1749 			tx_pkts[nb_tx] = m;
1750 
1751 			if (unlikely(error)) {
1752 				rte_errno = -error;
1753 				break;
1754 			}
1755 		}
1756 
1757 		error = rte_net_intel_cksum_prepare(m);
1758 		if (unlikely(error)) {
1759 			rte_errno = -error;
1760 			break;
1761 		}
1762 
1763 		if (m->ol_flags & PKT_TX_TCP_SEG)
1764 			virtio_tso_fix_cksum(m);
1765 	}
1766 
1767 	return nb_tx;
1768 }
1769 
1770 uint16_t
1771 virtio_xmit_pkts_packed(void *tx_queue, struct rte_mbuf **tx_pkts,
1772 			uint16_t nb_pkts)
1773 {
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;
1778 	uint16_t nb_tx = 0;
1779 	bool in_order = virtio_with_feature(hw, VIRTIO_F_IN_ORDER);
1780 
1781 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1782 		return nb_tx;
1783 
1784 	if (unlikely(nb_pkts < 1))
1785 		return nb_pkts;
1786 
1787 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1788 
1789 	if (nb_pkts > vq->vq_free_cnt)
1790 		virtio_xmit_cleanup_packed(vq, nb_pkts - vq->vq_free_cnt,
1791 					   in_order);
1792 
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;
1796 
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)))
1806 			can_push = 1;
1807 		else if (virtio_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1808 			 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1809 			use_indirect = 1;
1810 		/* How many main ring entries are needed to this Tx?
1811 		 * indirect   => 1
1812 		 * any_layout => number of segments
1813 		 * default    => number of segments + 1
1814 		 */
1815 		slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1816 		need = slots - vq->vq_free_cnt;
1817 
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)) {
1823 				PMD_TX_LOG(ERR,
1824 					   "No free tx descriptors to transmit");
1825 				break;
1826 			}
1827 		}
1828 
1829 		/* Enqueue Packet buffers */
1830 		if (can_push)
1831 			virtqueue_enqueue_xmit_packed_fast(txvq, txm, in_order);
1832 		else
1833 			virtqueue_enqueue_xmit_packed(txvq, txm, slots,
1834 						      use_indirect, 0,
1835 						      in_order);
1836 
1837 		virtio_update_packet_stats(&txvq->stats, txm);
1838 	}
1839 
1840 	txvq->stats.packets += nb_tx;
1841 
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");
1846 		}
1847 	}
1848 
1849 	return nb_tx;
1850 }
1851 
1852 uint16_t
1853 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1854 {
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;
1860 
1861 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1862 		return nb_tx;
1863 
1864 	if (unlikely(nb_pkts < 1))
1865 		return nb_pkts;
1866 
1867 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1868 
1869 	nb_used = virtqueue_nused(vq);
1870 
1871 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1872 		virtio_xmit_cleanup(vq, nb_used);
1873 
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;
1877 
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)))
1887 			can_push = 1;
1888 		else if (virtio_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1889 			 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1890 			use_indirect = 1;
1891 
1892 		/* How many main ring entries are needed to this Tx?
1893 		 * any_layout => number of segments
1894 		 * indirect   => 1
1895 		 * default    => number of segments + 1
1896 		 */
1897 		slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1898 		need = slots - vq->vq_free_cnt;
1899 
1900 		/* Positive value indicates it need free vring descriptors */
1901 		if (unlikely(need > 0)) {
1902 			nb_used = virtqueue_nused(vq);
1903 
1904 			need = RTE_MIN(need, (int)nb_used);
1905 
1906 			virtio_xmit_cleanup(vq, need);
1907 			need = slots - vq->vq_free_cnt;
1908 			if (unlikely(need > 0)) {
1909 				PMD_TX_LOG(ERR,
1910 					   "No free tx descriptors to transmit");
1911 				break;
1912 			}
1913 		}
1914 
1915 		/* Enqueue Packet buffers */
1916 		virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
1917 			can_push, 0);
1918 
1919 		virtio_update_packet_stats(&txvq->stats, txm);
1920 	}
1921 
1922 	txvq->stats.packets += nb_tx;
1923 
1924 	if (likely(nb_tx)) {
1925 		vq_update_avail_idx(vq);
1926 
1927 		if (unlikely(virtqueue_kick_prepare(vq))) {
1928 			virtqueue_notify(vq);
1929 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1930 		}
1931 	}
1932 
1933 	return nb_tx;
1934 }
1935 
1936 static __rte_always_inline int
1937 virtio_xmit_try_cleanup_inorder(struct virtqueue *vq, uint16_t need)
1938 {
1939 	uint16_t nb_used, nb_clean, nb_descs;
1940 
1941 	nb_descs = vq->vq_free_cnt + need;
1942 	nb_used = virtqueue_nused(vq);
1943 	nb_clean = RTE_MIN(need, (int)nb_used);
1944 
1945 	virtio_xmit_cleanup_inorder(vq, nb_clean);
1946 
1947 	return nb_descs - vq->vq_free_cnt;
1948 }
1949 
1950 uint16_t
1951 virtio_xmit_pkts_inorder(void *tx_queue,
1952 			struct rte_mbuf **tx_pkts,
1953 			uint16_t nb_pkts)
1954 {
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];
1961 	int need;
1962 
1963 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1964 		return nb_tx;
1965 
1966 	if (unlikely(nb_pkts < 1))
1967 		return nb_pkts;
1968 
1969 	VIRTQUEUE_DUMP(vq);
1970 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1971 	nb_used = virtqueue_nused(vq);
1972 
1973 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1974 		virtio_xmit_cleanup_inorder(vq, nb_used);
1975 
1976 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1977 		struct rte_mbuf *txm = tx_pkts[nb_tx];
1978 		int slots;
1979 
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;
1990 			nb_inorder_pkts++;
1991 
1992 			continue;
1993 		}
1994 
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,
1999 								       need);
2000 				if (unlikely(need > 0)) {
2001 					PMD_TX_LOG(ERR,
2002 						"No free tx descriptors to "
2003 						"transmit");
2004 					break;
2005 				}
2006 			}
2007 			virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2008 							nb_inorder_pkts);
2009 			nb_inorder_pkts = 0;
2010 		}
2011 
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);
2016 
2017 			if (unlikely(need > 0)) {
2018 				PMD_TX_LOG(ERR,
2019 					"No free tx descriptors to transmit");
2020 				break;
2021 			}
2022 		}
2023 		/* Enqueue Packet buffers */
2024 		virtqueue_enqueue_xmit(txvq, txm, slots, 0, 0, 1);
2025 
2026 		virtio_update_packet_stats(&txvq->stats, txm);
2027 	}
2028 
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,
2034 								  need);
2035 			if (unlikely(need > 0)) {
2036 				PMD_TX_LOG(ERR,
2037 					"No free tx descriptors to transmit");
2038 				nb_inorder_pkts = vq->vq_free_cnt;
2039 				nb_tx -= need;
2040 			}
2041 		}
2042 
2043 		virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2044 						nb_inorder_pkts);
2045 	}
2046 
2047 	txvq->stats.packets += nb_tx;
2048 
2049 	if (likely(nb_tx)) {
2050 		vq_update_avail_idx(vq);
2051 
2052 		if (unlikely(virtqueue_kick_prepare(vq))) {
2053 			virtqueue_notify(vq);
2054 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2055 		}
2056 	}
2057 
2058 	VIRTQUEUE_DUMP(vq);
2059 
2060 	return nb_tx;
2061 }
2062 
2063 __rte_weak uint16_t
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)
2067 {
2068 	return 0;
2069 }
2070 
2071 __rte_weak uint16_t
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)
2075 {
2076 	return 0;
2077 }
2078