xref: /dpdk/drivers/net/virtio/virtio_rxtx.c (revision cf8a8a8f4896c0885d3996716f73513c4317e545)
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 == 0 || nb_desc > vq->vq_nentries)
710 		nb_desc = vq->vq_nentries;
711 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
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 				}
772 			}
773 
774 			nbufs += free_cnt;
775 			vq_update_avail_idx(vq);
776 		}
777 	} else {
778 		while (!virtqueue_full(vq)) {
779 			m = rte_mbuf_raw_alloc(rxvq->mpool);
780 			if (m == NULL)
781 				break;
782 
783 			/* Enqueue allocated buffers */
784 			if (virtio_with_packed_queue(vq->hw))
785 				error = virtqueue_enqueue_recv_refill_packed_init(vq,
786 						&m, 1);
787 			else
788 				error = virtqueue_enqueue_recv_refill(vq,
789 						&m, 1);
790 			if (error) {
791 				rte_pktmbuf_free(m);
792 				break;
793 			}
794 			nbufs++;
795 		}
796 
797 		if (!virtio_with_packed_queue(vq->hw))
798 			vq_update_avail_idx(vq);
799 	}
800 
801 	PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
802 
803 	VIRTQUEUE_DUMP(vq);
804 
805 	return 0;
806 }
807 
808 /*
809  * struct rte_eth_dev *dev: Used to update dev
810  * uint16_t nb_desc: Defaults to values read from config space
811  * unsigned int socket_id: Used to allocate memzone
812  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
813  * uint16_t queue_idx: Just used as an index in dev txq list
814  */
815 int
816 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
817 			uint16_t queue_idx,
818 			uint16_t nb_desc,
819 			unsigned int socket_id __rte_unused,
820 			const struct rte_eth_txconf *tx_conf)
821 {
822 	uint8_t vq_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
823 	struct virtio_hw *hw = dev->data->dev_private;
824 	struct virtqueue *vq = hw->vqs[vq_idx];
825 	struct virtnet_tx *txvq;
826 	uint16_t tx_free_thresh;
827 
828 	PMD_INIT_FUNC_TRACE();
829 
830 	if (tx_conf->tx_deferred_start) {
831 		PMD_INIT_LOG(ERR, "Tx deferred start is not supported");
832 		return -EINVAL;
833 	}
834 
835 	if (nb_desc == 0 || nb_desc > vq->vq_nentries)
836 		nb_desc = vq->vq_nentries;
837 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
838 
839 	txvq = &vq->txq;
840 	txvq->queue_id = queue_idx;
841 
842 	tx_free_thresh = tx_conf->tx_free_thresh;
843 	if (tx_free_thresh == 0)
844 		tx_free_thresh =
845 			RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
846 
847 	if (tx_free_thresh >= (vq->vq_nentries - 3)) {
848 		PMD_DRV_LOG(ERR, "tx_free_thresh must be less than the "
849 			"number of TX entries minus 3 (%u)."
850 			" (tx_free_thresh=%u port=%u queue=%u)\n",
851 			vq->vq_nentries - 3,
852 			tx_free_thresh, dev->data->port_id, queue_idx);
853 		return -EINVAL;
854 	}
855 
856 	vq->vq_free_thresh = tx_free_thresh;
857 
858 	dev->data->tx_queues[queue_idx] = txvq;
859 	return 0;
860 }
861 
862 int
863 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
864 				uint16_t queue_idx)
865 {
866 	uint8_t vq_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
867 	struct virtio_hw *hw = dev->data->dev_private;
868 	struct virtqueue *vq = hw->vqs[vq_idx];
869 
870 	PMD_INIT_FUNC_TRACE();
871 
872 	if (!virtio_with_packed_queue(hw)) {
873 		if (virtio_with_feature(hw, VIRTIO_F_IN_ORDER))
874 			vq->vq_split.ring.desc[vq->vq_nentries - 1].next = 0;
875 	}
876 
877 	VIRTQUEUE_DUMP(vq);
878 
879 	return 0;
880 }
881 
882 static inline void
883 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
884 {
885 	int error;
886 	/*
887 	 * Requeue the discarded mbuf. This should always be
888 	 * successful since it was just dequeued.
889 	 */
890 	if (virtio_with_packed_queue(vq->hw))
891 		error = virtqueue_enqueue_recv_refill_packed(vq, &m, 1);
892 	else
893 		error = virtqueue_enqueue_recv_refill(vq, &m, 1);
894 
895 	if (unlikely(error)) {
896 		PMD_DRV_LOG(ERR, "cannot requeue discarded mbuf");
897 		rte_pktmbuf_free(m);
898 	}
899 }
900 
901 static inline void
902 virtio_discard_rxbuf_inorder(struct virtqueue *vq, struct rte_mbuf *m)
903 {
904 	int error;
905 
906 	error = virtqueue_enqueue_refill_inorder(vq, &m, 1);
907 	if (unlikely(error)) {
908 		PMD_DRV_LOG(ERR, "cannot requeue discarded mbuf");
909 		rte_pktmbuf_free(m);
910 	}
911 }
912 
913 /* Optionally fill offload information in structure */
914 static inline int
915 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
916 {
917 	struct rte_net_hdr_lens hdr_lens;
918 	uint32_t hdrlen, ptype;
919 	int l4_supported = 0;
920 
921 	/* nothing to do */
922 	if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
923 		return 0;
924 
925 	m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
926 
927 	ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
928 	m->packet_type = ptype;
929 	if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
930 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
931 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
932 		l4_supported = 1;
933 
934 	if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
935 		hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
936 		if (hdr->csum_start <= hdrlen && l4_supported) {
937 			m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
938 		} else {
939 			/* Unknown proto or tunnel, do sw cksum. We can assume
940 			 * the cksum field is in the first segment since the
941 			 * buffers we provided to the host are large enough.
942 			 * In case of SCTP, this will be wrong since it's a CRC
943 			 * but there's nothing we can do.
944 			 */
945 			uint16_t csum = 0, off;
946 
947 			if (rte_raw_cksum_mbuf(m, hdr->csum_start,
948 				rte_pktmbuf_pkt_len(m) - hdr->csum_start,
949 				&csum) < 0)
950 				return -EINVAL;
951 			if (likely(csum != 0xffff))
952 				csum = ~csum;
953 			off = hdr->csum_offset + hdr->csum_start;
954 			if (rte_pktmbuf_data_len(m) >= off + 1)
955 				*rte_pktmbuf_mtod_offset(m, uint16_t *,
956 					off) = csum;
957 		}
958 	} else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
959 		m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
960 	}
961 
962 	/* GSO request, save required information in mbuf */
963 	if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
964 		/* Check unsupported modes */
965 		if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
966 		    (hdr->gso_size == 0)) {
967 			return -EINVAL;
968 		}
969 
970 		/* Update mss lengthes in mbuf */
971 		m->tso_segsz = hdr->gso_size;
972 		switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
973 			case VIRTIO_NET_HDR_GSO_TCPV4:
974 			case VIRTIO_NET_HDR_GSO_TCPV6:
975 				m->ol_flags |= PKT_RX_LRO | \
976 					PKT_RX_L4_CKSUM_NONE;
977 				break;
978 			default:
979 				return -EINVAL;
980 		}
981 	}
982 
983 	return 0;
984 }
985 
986 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
987 uint16_t
988 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
989 {
990 	struct virtnet_rx *rxvq = rx_queue;
991 	struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
992 	struct virtio_hw *hw = vq->hw;
993 	struct rte_mbuf *rxm;
994 	uint16_t nb_used, num, nb_rx;
995 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
996 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
997 	int error;
998 	uint32_t i, nb_enqueued;
999 	uint32_t hdr_size;
1000 	struct virtio_net_hdr *hdr;
1001 
1002 	nb_rx = 0;
1003 	if (unlikely(hw->started == 0))
1004 		return nb_rx;
1005 
1006 	nb_used = virtqueue_nused(vq);
1007 
1008 	num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1009 	if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1010 		num = VIRTIO_MBUF_BURST_SZ;
1011 	if (likely(num > DESC_PER_CACHELINE))
1012 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1013 
1014 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1015 	PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
1016 
1017 	nb_enqueued = 0;
1018 	hdr_size = hw->vtnet_hdr_size;
1019 
1020 	for (i = 0; i < num ; i++) {
1021 		rxm = rcv_pkts[i];
1022 
1023 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1024 
1025 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1026 			PMD_RX_LOG(ERR, "Packet drop");
1027 			nb_enqueued++;
1028 			virtio_discard_rxbuf(vq, rxm);
1029 			rxvq->stats.errors++;
1030 			continue;
1031 		}
1032 
1033 		rxm->port = rxvq->port_id;
1034 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1035 		rxm->ol_flags = 0;
1036 		rxm->vlan_tci = 0;
1037 
1038 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1039 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1040 
1041 		hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1042 			RTE_PKTMBUF_HEADROOM - hdr_size);
1043 
1044 		if (hw->vlan_strip)
1045 			rte_vlan_strip(rxm);
1046 
1047 		if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1048 			virtio_discard_rxbuf(vq, rxm);
1049 			rxvq->stats.errors++;
1050 			continue;
1051 		}
1052 
1053 		virtio_rx_stats_updated(rxvq, rxm);
1054 
1055 		rx_pkts[nb_rx++] = rxm;
1056 	}
1057 
1058 	rxvq->stats.packets += nb_rx;
1059 
1060 	/* Allocate new mbuf for the used descriptor */
1061 	if (likely(!virtqueue_full(vq))) {
1062 		uint16_t free_cnt = vq->vq_free_cnt;
1063 		struct rte_mbuf *new_pkts[free_cnt];
1064 
1065 		if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1066 						free_cnt) == 0)) {
1067 			error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1068 					free_cnt);
1069 			if (unlikely(error)) {
1070 				for (i = 0; i < free_cnt; i++)
1071 					rte_pktmbuf_free(new_pkts[i]);
1072 			}
1073 			nb_enqueued += free_cnt;
1074 		} else {
1075 			struct rte_eth_dev *dev =
1076 				&rte_eth_devices[rxvq->port_id];
1077 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1078 		}
1079 	}
1080 
1081 	if (likely(nb_enqueued)) {
1082 		vq_update_avail_idx(vq);
1083 
1084 		if (unlikely(virtqueue_kick_prepare(vq))) {
1085 			virtqueue_notify(vq);
1086 			PMD_RX_LOG(DEBUG, "Notified");
1087 		}
1088 	}
1089 
1090 	return nb_rx;
1091 }
1092 
1093 uint16_t
1094 virtio_recv_pkts_packed(void *rx_queue, struct rte_mbuf **rx_pkts,
1095 			uint16_t nb_pkts)
1096 {
1097 	struct virtnet_rx *rxvq = rx_queue;
1098 	struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1099 	struct virtio_hw *hw = vq->hw;
1100 	struct rte_mbuf *rxm;
1101 	uint16_t num, nb_rx;
1102 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1103 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1104 	int error;
1105 	uint32_t i, nb_enqueued;
1106 	uint32_t hdr_size;
1107 	struct virtio_net_hdr *hdr;
1108 
1109 	nb_rx = 0;
1110 	if (unlikely(hw->started == 0))
1111 		return nb_rx;
1112 
1113 	num = RTE_MIN(VIRTIO_MBUF_BURST_SZ, nb_pkts);
1114 	if (likely(num > DESC_PER_CACHELINE))
1115 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1116 
1117 	num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1118 	PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1119 
1120 	nb_enqueued = 0;
1121 	hdr_size = hw->vtnet_hdr_size;
1122 
1123 	for (i = 0; i < num; i++) {
1124 		rxm = rcv_pkts[i];
1125 
1126 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1127 
1128 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1129 			PMD_RX_LOG(ERR, "Packet drop");
1130 			nb_enqueued++;
1131 			virtio_discard_rxbuf(vq, rxm);
1132 			rxvq->stats.errors++;
1133 			continue;
1134 		}
1135 
1136 		rxm->port = rxvq->port_id;
1137 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1138 		rxm->ol_flags = 0;
1139 		rxm->vlan_tci = 0;
1140 
1141 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1142 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1143 
1144 		hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1145 			RTE_PKTMBUF_HEADROOM - hdr_size);
1146 
1147 		if (hw->vlan_strip)
1148 			rte_vlan_strip(rxm);
1149 
1150 		if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1151 			virtio_discard_rxbuf(vq, rxm);
1152 			rxvq->stats.errors++;
1153 			continue;
1154 		}
1155 
1156 		virtio_rx_stats_updated(rxvq, rxm);
1157 
1158 		rx_pkts[nb_rx++] = rxm;
1159 	}
1160 
1161 	rxvq->stats.packets += nb_rx;
1162 
1163 	/* Allocate new mbuf for the used descriptor */
1164 	if (likely(!virtqueue_full(vq))) {
1165 		uint16_t free_cnt = vq->vq_free_cnt;
1166 		struct rte_mbuf *new_pkts[free_cnt];
1167 
1168 		if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1169 						free_cnt) == 0)) {
1170 			error = virtqueue_enqueue_recv_refill_packed(vq,
1171 					new_pkts, free_cnt);
1172 			if (unlikely(error)) {
1173 				for (i = 0; i < free_cnt; i++)
1174 					rte_pktmbuf_free(new_pkts[i]);
1175 			}
1176 			nb_enqueued += free_cnt;
1177 		} else {
1178 			struct rte_eth_dev *dev =
1179 				&rte_eth_devices[rxvq->port_id];
1180 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1181 		}
1182 	}
1183 
1184 	if (likely(nb_enqueued)) {
1185 		if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1186 			virtqueue_notify(vq);
1187 			PMD_RX_LOG(DEBUG, "Notified");
1188 		}
1189 	}
1190 
1191 	return nb_rx;
1192 }
1193 
1194 
1195 uint16_t
1196 virtio_recv_pkts_inorder(void *rx_queue,
1197 			struct rte_mbuf **rx_pkts,
1198 			uint16_t nb_pkts)
1199 {
1200 	struct virtnet_rx *rxvq = rx_queue;
1201 	struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1202 	struct virtio_hw *hw = vq->hw;
1203 	struct rte_mbuf *rxm;
1204 	struct rte_mbuf *prev = NULL;
1205 	uint16_t nb_used, num, nb_rx;
1206 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1207 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1208 	int error;
1209 	uint32_t nb_enqueued;
1210 	uint32_t seg_num;
1211 	uint32_t seg_res;
1212 	uint32_t hdr_size;
1213 	int32_t i;
1214 
1215 	nb_rx = 0;
1216 	if (unlikely(hw->started == 0))
1217 		return nb_rx;
1218 
1219 	nb_used = virtqueue_nused(vq);
1220 	nb_used = RTE_MIN(nb_used, nb_pkts);
1221 	nb_used = RTE_MIN(nb_used, VIRTIO_MBUF_BURST_SZ);
1222 
1223 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1224 
1225 	nb_enqueued = 0;
1226 	seg_num = 1;
1227 	seg_res = 0;
1228 	hdr_size = hw->vtnet_hdr_size;
1229 
1230 	num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len, nb_used);
1231 
1232 	for (i = 0; i < num; i++) {
1233 		struct virtio_net_hdr_mrg_rxbuf *header;
1234 
1235 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1236 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1237 
1238 		rxm = rcv_pkts[i];
1239 
1240 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1241 			PMD_RX_LOG(ERR, "Packet drop");
1242 			nb_enqueued++;
1243 			virtio_discard_rxbuf_inorder(vq, rxm);
1244 			rxvq->stats.errors++;
1245 			continue;
1246 		}
1247 
1248 		header = (struct virtio_net_hdr_mrg_rxbuf *)
1249 			 ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1250 			 - hdr_size);
1251 
1252 		if (virtio_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
1253 			seg_num = header->num_buffers;
1254 			if (seg_num == 0)
1255 				seg_num = 1;
1256 		} else {
1257 			seg_num = 1;
1258 		}
1259 
1260 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1261 		rxm->nb_segs = seg_num;
1262 		rxm->ol_flags = 0;
1263 		rxm->vlan_tci = 0;
1264 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1265 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1266 
1267 		rxm->port = rxvq->port_id;
1268 
1269 		rx_pkts[nb_rx] = rxm;
1270 		prev = rxm;
1271 
1272 		if (vq->hw->has_rx_offload &&
1273 				virtio_rx_offload(rxm, &header->hdr) < 0) {
1274 			virtio_discard_rxbuf_inorder(vq, rxm);
1275 			rxvq->stats.errors++;
1276 			continue;
1277 		}
1278 
1279 		if (hw->vlan_strip)
1280 			rte_vlan_strip(rx_pkts[nb_rx]);
1281 
1282 		seg_res = seg_num - 1;
1283 
1284 		/* Merge remaining segments */
1285 		while (seg_res != 0 && i < (num - 1)) {
1286 			i++;
1287 
1288 			rxm = rcv_pkts[i];
1289 			rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1290 			rxm->pkt_len = (uint32_t)(len[i]);
1291 			rxm->data_len = (uint16_t)(len[i]);
1292 
1293 			rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1294 
1295 			prev->next = rxm;
1296 			prev = rxm;
1297 			seg_res -= 1;
1298 		}
1299 
1300 		if (!seg_res) {
1301 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1302 			nb_rx++;
1303 		}
1304 	}
1305 
1306 	/* Last packet still need merge segments */
1307 	while (seg_res != 0) {
1308 		uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1309 					VIRTIO_MBUF_BURST_SZ);
1310 
1311 		if (likely(virtqueue_nused(vq) >= rcv_cnt)) {
1312 			num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len,
1313 							   rcv_cnt);
1314 			uint16_t extra_idx = 0;
1315 
1316 			rcv_cnt = num;
1317 			while (extra_idx < rcv_cnt) {
1318 				rxm = rcv_pkts[extra_idx];
1319 				rxm->data_off =
1320 					RTE_PKTMBUF_HEADROOM - hdr_size;
1321 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
1322 				rxm->data_len = (uint16_t)(len[extra_idx]);
1323 				prev->next = rxm;
1324 				prev = rxm;
1325 				rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1326 				extra_idx += 1;
1327 			};
1328 			seg_res -= rcv_cnt;
1329 
1330 			if (!seg_res) {
1331 				virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1332 				nb_rx++;
1333 			}
1334 		} else {
1335 			PMD_RX_LOG(ERR,
1336 					"No enough segments for packet.");
1337 			rte_pktmbuf_free(rx_pkts[nb_rx]);
1338 			rxvq->stats.errors++;
1339 			break;
1340 		}
1341 	}
1342 
1343 	rxvq->stats.packets += nb_rx;
1344 
1345 	/* Allocate new mbuf for the used descriptor */
1346 
1347 	if (likely(!virtqueue_full(vq))) {
1348 		/* free_cnt may include mrg descs */
1349 		uint16_t free_cnt = vq->vq_free_cnt;
1350 		struct rte_mbuf *new_pkts[free_cnt];
1351 
1352 		if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1353 			error = virtqueue_enqueue_refill_inorder(vq, new_pkts,
1354 					free_cnt);
1355 			if (unlikely(error)) {
1356 				for (i = 0; i < free_cnt; i++)
1357 					rte_pktmbuf_free(new_pkts[i]);
1358 			}
1359 			nb_enqueued += free_cnt;
1360 		} else {
1361 			struct rte_eth_dev *dev =
1362 				&rte_eth_devices[rxvq->port_id];
1363 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1364 		}
1365 	}
1366 
1367 	if (likely(nb_enqueued)) {
1368 		vq_update_avail_idx(vq);
1369 
1370 		if (unlikely(virtqueue_kick_prepare(vq))) {
1371 			virtqueue_notify(vq);
1372 			PMD_RX_LOG(DEBUG, "Notified");
1373 		}
1374 	}
1375 
1376 	return nb_rx;
1377 }
1378 
1379 uint16_t
1380 virtio_recv_mergeable_pkts(void *rx_queue,
1381 			struct rte_mbuf **rx_pkts,
1382 			uint16_t nb_pkts)
1383 {
1384 	struct virtnet_rx *rxvq = rx_queue;
1385 	struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1386 	struct virtio_hw *hw = vq->hw;
1387 	struct rte_mbuf *rxm;
1388 	struct rte_mbuf *prev = NULL;
1389 	uint16_t nb_used, num, nb_rx = 0;
1390 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1391 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1392 	int error;
1393 	uint32_t nb_enqueued = 0;
1394 	uint32_t seg_num = 0;
1395 	uint32_t seg_res = 0;
1396 	uint32_t hdr_size = hw->vtnet_hdr_size;
1397 	int32_t i;
1398 
1399 	if (unlikely(hw->started == 0))
1400 		return nb_rx;
1401 
1402 	nb_used = virtqueue_nused(vq);
1403 
1404 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1405 
1406 	num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1407 	if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1408 		num = VIRTIO_MBUF_BURST_SZ;
1409 	if (likely(num > DESC_PER_CACHELINE))
1410 		num = num - ((vq->vq_used_cons_idx + num) %
1411 				DESC_PER_CACHELINE);
1412 
1413 
1414 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1415 
1416 	for (i = 0; i < num; i++) {
1417 		struct virtio_net_hdr_mrg_rxbuf *header;
1418 
1419 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1420 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1421 
1422 		rxm = rcv_pkts[i];
1423 
1424 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1425 			PMD_RX_LOG(ERR, "Packet drop");
1426 			nb_enqueued++;
1427 			virtio_discard_rxbuf(vq, rxm);
1428 			rxvq->stats.errors++;
1429 			continue;
1430 		}
1431 
1432 		header = (struct virtio_net_hdr_mrg_rxbuf *)
1433 			 ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1434 			 - hdr_size);
1435 		seg_num = header->num_buffers;
1436 		if (seg_num == 0)
1437 			seg_num = 1;
1438 
1439 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1440 		rxm->nb_segs = seg_num;
1441 		rxm->ol_flags = 0;
1442 		rxm->vlan_tci = 0;
1443 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1444 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1445 
1446 		rxm->port = rxvq->port_id;
1447 
1448 		rx_pkts[nb_rx] = rxm;
1449 		prev = rxm;
1450 
1451 		if (hw->has_rx_offload &&
1452 				virtio_rx_offload(rxm, &header->hdr) < 0) {
1453 			virtio_discard_rxbuf(vq, rxm);
1454 			rxvq->stats.errors++;
1455 			continue;
1456 		}
1457 
1458 		if (hw->vlan_strip)
1459 			rte_vlan_strip(rx_pkts[nb_rx]);
1460 
1461 		seg_res = seg_num - 1;
1462 
1463 		/* Merge remaining segments */
1464 		while (seg_res != 0 && i < (num - 1)) {
1465 			i++;
1466 
1467 			rxm = rcv_pkts[i];
1468 			rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1469 			rxm->pkt_len = (uint32_t)(len[i]);
1470 			rxm->data_len = (uint16_t)(len[i]);
1471 
1472 			rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1473 
1474 			prev->next = rxm;
1475 			prev = rxm;
1476 			seg_res -= 1;
1477 		}
1478 
1479 		if (!seg_res) {
1480 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1481 			nb_rx++;
1482 		}
1483 	}
1484 
1485 	/* Last packet still need merge segments */
1486 	while (seg_res != 0) {
1487 		uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1488 					VIRTIO_MBUF_BURST_SZ);
1489 
1490 		if (likely(virtqueue_nused(vq) >= rcv_cnt)) {
1491 			num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len,
1492 							   rcv_cnt);
1493 			uint16_t extra_idx = 0;
1494 
1495 			rcv_cnt = num;
1496 			while (extra_idx < rcv_cnt) {
1497 				rxm = rcv_pkts[extra_idx];
1498 				rxm->data_off =
1499 					RTE_PKTMBUF_HEADROOM - hdr_size;
1500 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
1501 				rxm->data_len = (uint16_t)(len[extra_idx]);
1502 				prev->next = rxm;
1503 				prev = rxm;
1504 				rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1505 				extra_idx += 1;
1506 			};
1507 			seg_res -= rcv_cnt;
1508 
1509 			if (!seg_res) {
1510 				virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1511 				nb_rx++;
1512 			}
1513 		} else {
1514 			PMD_RX_LOG(ERR,
1515 					"No enough segments for packet.");
1516 			rte_pktmbuf_free(rx_pkts[nb_rx]);
1517 			rxvq->stats.errors++;
1518 			break;
1519 		}
1520 	}
1521 
1522 	rxvq->stats.packets += nb_rx;
1523 
1524 	/* Allocate new mbuf for the used descriptor */
1525 	if (likely(!virtqueue_full(vq))) {
1526 		/* free_cnt may include mrg descs */
1527 		uint16_t free_cnt = vq->vq_free_cnt;
1528 		struct rte_mbuf *new_pkts[free_cnt];
1529 
1530 		if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1531 			error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1532 					free_cnt);
1533 			if (unlikely(error)) {
1534 				for (i = 0; i < free_cnt; i++)
1535 					rte_pktmbuf_free(new_pkts[i]);
1536 			}
1537 			nb_enqueued += free_cnt;
1538 		} else {
1539 			struct rte_eth_dev *dev =
1540 				&rte_eth_devices[rxvq->port_id];
1541 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1542 		}
1543 	}
1544 
1545 	if (likely(nb_enqueued)) {
1546 		vq_update_avail_idx(vq);
1547 
1548 		if (unlikely(virtqueue_kick_prepare(vq))) {
1549 			virtqueue_notify(vq);
1550 			PMD_RX_LOG(DEBUG, "Notified");
1551 		}
1552 	}
1553 
1554 	return nb_rx;
1555 }
1556 
1557 uint16_t
1558 virtio_recv_mergeable_pkts_packed(void *rx_queue,
1559 			struct rte_mbuf **rx_pkts,
1560 			uint16_t nb_pkts)
1561 {
1562 	struct virtnet_rx *rxvq = rx_queue;
1563 	struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1564 	struct virtio_hw *hw = vq->hw;
1565 	struct rte_mbuf *rxm;
1566 	struct rte_mbuf *prev = NULL;
1567 	uint16_t num, nb_rx = 0;
1568 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1569 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1570 	uint32_t nb_enqueued = 0;
1571 	uint32_t seg_num = 0;
1572 	uint32_t seg_res = 0;
1573 	uint32_t hdr_size = hw->vtnet_hdr_size;
1574 	int32_t i;
1575 	int error;
1576 
1577 	if (unlikely(hw->started == 0))
1578 		return nb_rx;
1579 
1580 
1581 	num = nb_pkts;
1582 	if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1583 		num = VIRTIO_MBUF_BURST_SZ;
1584 	if (likely(num > DESC_PER_CACHELINE))
1585 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1586 
1587 	num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1588 
1589 	for (i = 0; i < num; i++) {
1590 		struct virtio_net_hdr_mrg_rxbuf *header;
1591 
1592 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1593 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1594 
1595 		rxm = rcv_pkts[i];
1596 
1597 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1598 			PMD_RX_LOG(ERR, "Packet drop");
1599 			nb_enqueued++;
1600 			virtio_discard_rxbuf(vq, rxm);
1601 			rxvq->stats.errors++;
1602 			continue;
1603 		}
1604 
1605 		header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)
1606 			  rxm->buf_addr + RTE_PKTMBUF_HEADROOM - hdr_size);
1607 		seg_num = header->num_buffers;
1608 
1609 		if (seg_num == 0)
1610 			seg_num = 1;
1611 
1612 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1613 		rxm->nb_segs = seg_num;
1614 		rxm->ol_flags = 0;
1615 		rxm->vlan_tci = 0;
1616 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1617 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1618 
1619 		rxm->port = rxvq->port_id;
1620 		rx_pkts[nb_rx] = rxm;
1621 		prev = rxm;
1622 
1623 		if (hw->has_rx_offload &&
1624 				virtio_rx_offload(rxm, &header->hdr) < 0) {
1625 			virtio_discard_rxbuf(vq, rxm);
1626 			rxvq->stats.errors++;
1627 			continue;
1628 		}
1629 
1630 		if (hw->vlan_strip)
1631 			rte_vlan_strip(rx_pkts[nb_rx]);
1632 
1633 		seg_res = seg_num - 1;
1634 
1635 		/* Merge remaining segments */
1636 		while (seg_res != 0 && i < (num - 1)) {
1637 			i++;
1638 
1639 			rxm = rcv_pkts[i];
1640 			rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1641 			rxm->pkt_len = (uint32_t)(len[i]);
1642 			rxm->data_len = (uint16_t)(len[i]);
1643 
1644 			rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1645 
1646 			prev->next = rxm;
1647 			prev = rxm;
1648 			seg_res -= 1;
1649 		}
1650 
1651 		if (!seg_res) {
1652 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1653 			nb_rx++;
1654 		}
1655 	}
1656 
1657 	/* Last packet still need merge segments */
1658 	while (seg_res != 0) {
1659 		uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1660 					VIRTIO_MBUF_BURST_SZ);
1661 		uint16_t extra_idx = 0;
1662 
1663 		rcv_cnt = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts,
1664 				len, rcv_cnt);
1665 		if (unlikely(rcv_cnt == 0)) {
1666 			PMD_RX_LOG(ERR, "No enough segments for packet.");
1667 			rte_pktmbuf_free(rx_pkts[nb_rx]);
1668 			rxvq->stats.errors++;
1669 			break;
1670 		}
1671 
1672 		while (extra_idx < rcv_cnt) {
1673 			rxm = rcv_pkts[extra_idx];
1674 
1675 			rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1676 			rxm->pkt_len = (uint32_t)(len[extra_idx]);
1677 			rxm->data_len = (uint16_t)(len[extra_idx]);
1678 
1679 			prev->next = rxm;
1680 			prev = rxm;
1681 			rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1682 			extra_idx += 1;
1683 		}
1684 		seg_res -= rcv_cnt;
1685 		if (!seg_res) {
1686 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1687 			nb_rx++;
1688 		}
1689 	}
1690 
1691 	rxvq->stats.packets += nb_rx;
1692 
1693 	/* Allocate new mbuf for the used descriptor */
1694 	if (likely(!virtqueue_full(vq))) {
1695 		/* free_cnt may include mrg descs */
1696 		uint16_t free_cnt = vq->vq_free_cnt;
1697 		struct rte_mbuf *new_pkts[free_cnt];
1698 
1699 		if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1700 			error = virtqueue_enqueue_recv_refill_packed(vq,
1701 					new_pkts, free_cnt);
1702 			if (unlikely(error)) {
1703 				for (i = 0; i < free_cnt; i++)
1704 					rte_pktmbuf_free(new_pkts[i]);
1705 			}
1706 			nb_enqueued += free_cnt;
1707 		} else {
1708 			struct rte_eth_dev *dev =
1709 				&rte_eth_devices[rxvq->port_id];
1710 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1711 		}
1712 	}
1713 
1714 	if (likely(nb_enqueued)) {
1715 		if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1716 			virtqueue_notify(vq);
1717 			PMD_RX_LOG(DEBUG, "Notified");
1718 		}
1719 	}
1720 
1721 	return nb_rx;
1722 }
1723 
1724 uint16_t
1725 virtio_xmit_pkts_prepare(void *tx_queue __rte_unused, struct rte_mbuf **tx_pkts,
1726 			uint16_t nb_pkts)
1727 {
1728 	uint16_t nb_tx;
1729 	int error;
1730 
1731 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1732 		struct rte_mbuf *m = tx_pkts[nb_tx];
1733 
1734 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
1735 		error = rte_validate_tx_offload(m);
1736 		if (unlikely(error)) {
1737 			rte_errno = -error;
1738 			break;
1739 		}
1740 #endif
1741 
1742 		/* Do VLAN tag insertion */
1743 		if (unlikely(m->ol_flags & PKT_TX_VLAN_PKT)) {
1744 			error = rte_vlan_insert(&m);
1745 			/* rte_vlan_insert() may change pointer
1746 			 * even in the case of failure
1747 			 */
1748 			tx_pkts[nb_tx] = m;
1749 
1750 			if (unlikely(error)) {
1751 				rte_errno = -error;
1752 				break;
1753 			}
1754 		}
1755 
1756 		error = rte_net_intel_cksum_prepare(m);
1757 		if (unlikely(error)) {
1758 			rte_errno = -error;
1759 			break;
1760 		}
1761 
1762 		if (m->ol_flags & PKT_TX_TCP_SEG)
1763 			virtio_tso_fix_cksum(m);
1764 	}
1765 
1766 	return nb_tx;
1767 }
1768 
1769 uint16_t
1770 virtio_xmit_pkts_packed(void *tx_queue, struct rte_mbuf **tx_pkts,
1771 			uint16_t nb_pkts)
1772 {
1773 	struct virtnet_tx *txvq = tx_queue;
1774 	struct virtqueue *vq = virtnet_txq_to_vq(txvq);
1775 	struct virtio_hw *hw = vq->hw;
1776 	uint16_t hdr_size = hw->vtnet_hdr_size;
1777 	uint16_t nb_tx = 0;
1778 	bool in_order = virtio_with_feature(hw, VIRTIO_F_IN_ORDER);
1779 
1780 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1781 		return nb_tx;
1782 
1783 	if (unlikely(nb_pkts < 1))
1784 		return nb_pkts;
1785 
1786 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1787 
1788 	if (nb_pkts > vq->vq_free_cnt)
1789 		virtio_xmit_cleanup_packed(vq, nb_pkts - vq->vq_free_cnt,
1790 					   in_order);
1791 
1792 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1793 		struct rte_mbuf *txm = tx_pkts[nb_tx];
1794 		int can_push = 0, use_indirect = 0, slots, need;
1795 
1796 		/* optimize ring usage */
1797 		if ((virtio_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1798 		      virtio_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1799 		    rte_mbuf_refcnt_read(txm) == 1 &&
1800 		    RTE_MBUF_DIRECT(txm) &&
1801 		    txm->nb_segs == 1 &&
1802 		    rte_pktmbuf_headroom(txm) >= hdr_size &&
1803 		    rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1804 			   __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1805 			can_push = 1;
1806 		else if (virtio_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1807 			 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1808 			use_indirect = 1;
1809 		/* How many main ring entries are needed to this Tx?
1810 		 * indirect   => 1
1811 		 * any_layout => number of segments
1812 		 * default    => number of segments + 1
1813 		 */
1814 		slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1815 		need = slots - vq->vq_free_cnt;
1816 
1817 		/* Positive value indicates it need free vring descriptors */
1818 		if (unlikely(need > 0)) {
1819 			virtio_xmit_cleanup_packed(vq, need, in_order);
1820 			need = slots - vq->vq_free_cnt;
1821 			if (unlikely(need > 0)) {
1822 				PMD_TX_LOG(ERR,
1823 					   "No free tx descriptors to transmit");
1824 				break;
1825 			}
1826 		}
1827 
1828 		/* Enqueue Packet buffers */
1829 		if (can_push)
1830 			virtqueue_enqueue_xmit_packed_fast(txvq, txm, in_order);
1831 		else
1832 			virtqueue_enqueue_xmit_packed(txvq, txm, slots,
1833 						      use_indirect, 0,
1834 						      in_order);
1835 
1836 		virtio_update_packet_stats(&txvq->stats, txm);
1837 	}
1838 
1839 	txvq->stats.packets += nb_tx;
1840 
1841 	if (likely(nb_tx)) {
1842 		if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1843 			virtqueue_notify(vq);
1844 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1845 		}
1846 	}
1847 
1848 	return nb_tx;
1849 }
1850 
1851 uint16_t
1852 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1853 {
1854 	struct virtnet_tx *txvq = tx_queue;
1855 	struct virtqueue *vq = virtnet_txq_to_vq(txvq);
1856 	struct virtio_hw *hw = vq->hw;
1857 	uint16_t hdr_size = hw->vtnet_hdr_size;
1858 	uint16_t nb_used, nb_tx = 0;
1859 
1860 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1861 		return nb_tx;
1862 
1863 	if (unlikely(nb_pkts < 1))
1864 		return nb_pkts;
1865 
1866 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1867 
1868 	nb_used = virtqueue_nused(vq);
1869 
1870 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1871 		virtio_xmit_cleanup(vq, nb_used);
1872 
1873 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1874 		struct rte_mbuf *txm = tx_pkts[nb_tx];
1875 		int can_push = 0, use_indirect = 0, slots, need;
1876 
1877 		/* optimize ring usage */
1878 		if ((virtio_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1879 		      virtio_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1880 		    rte_mbuf_refcnt_read(txm) == 1 &&
1881 		    RTE_MBUF_DIRECT(txm) &&
1882 		    txm->nb_segs == 1 &&
1883 		    rte_pktmbuf_headroom(txm) >= hdr_size &&
1884 		    rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1885 				   __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1886 			can_push = 1;
1887 		else if (virtio_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1888 			 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1889 			use_indirect = 1;
1890 
1891 		/* How many main ring entries are needed to this Tx?
1892 		 * any_layout => number of segments
1893 		 * indirect   => 1
1894 		 * default    => number of segments + 1
1895 		 */
1896 		slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1897 		need = slots - vq->vq_free_cnt;
1898 
1899 		/* Positive value indicates it need free vring descriptors */
1900 		if (unlikely(need > 0)) {
1901 			nb_used = virtqueue_nused(vq);
1902 
1903 			need = RTE_MIN(need, (int)nb_used);
1904 
1905 			virtio_xmit_cleanup(vq, need);
1906 			need = slots - vq->vq_free_cnt;
1907 			if (unlikely(need > 0)) {
1908 				PMD_TX_LOG(ERR,
1909 					   "No free tx descriptors to transmit");
1910 				break;
1911 			}
1912 		}
1913 
1914 		/* Enqueue Packet buffers */
1915 		virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
1916 			can_push, 0);
1917 
1918 		virtio_update_packet_stats(&txvq->stats, txm);
1919 	}
1920 
1921 	txvq->stats.packets += nb_tx;
1922 
1923 	if (likely(nb_tx)) {
1924 		vq_update_avail_idx(vq);
1925 
1926 		if (unlikely(virtqueue_kick_prepare(vq))) {
1927 			virtqueue_notify(vq);
1928 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1929 		}
1930 	}
1931 
1932 	return nb_tx;
1933 }
1934 
1935 static __rte_always_inline int
1936 virtio_xmit_try_cleanup_inorder(struct virtqueue *vq, uint16_t need)
1937 {
1938 	uint16_t nb_used, nb_clean, nb_descs;
1939 
1940 	nb_descs = vq->vq_free_cnt + need;
1941 	nb_used = virtqueue_nused(vq);
1942 	nb_clean = RTE_MIN(need, (int)nb_used);
1943 
1944 	virtio_xmit_cleanup_inorder(vq, nb_clean);
1945 
1946 	return nb_descs - vq->vq_free_cnt;
1947 }
1948 
1949 uint16_t
1950 virtio_xmit_pkts_inorder(void *tx_queue,
1951 			struct rte_mbuf **tx_pkts,
1952 			uint16_t nb_pkts)
1953 {
1954 	struct virtnet_tx *txvq = tx_queue;
1955 	struct virtqueue *vq = virtnet_txq_to_vq(txvq);
1956 	struct virtio_hw *hw = vq->hw;
1957 	uint16_t hdr_size = hw->vtnet_hdr_size;
1958 	uint16_t nb_used, nb_tx = 0, nb_inorder_pkts = 0;
1959 	struct rte_mbuf *inorder_pkts[nb_pkts];
1960 	int need;
1961 
1962 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1963 		return nb_tx;
1964 
1965 	if (unlikely(nb_pkts < 1))
1966 		return nb_pkts;
1967 
1968 	VIRTQUEUE_DUMP(vq);
1969 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1970 	nb_used = virtqueue_nused(vq);
1971 
1972 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1973 		virtio_xmit_cleanup_inorder(vq, nb_used);
1974 
1975 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1976 		struct rte_mbuf *txm = tx_pkts[nb_tx];
1977 		int slots;
1978 
1979 		/* optimize ring usage */
1980 		if ((virtio_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1981 		     virtio_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1982 		     rte_mbuf_refcnt_read(txm) == 1 &&
1983 		     RTE_MBUF_DIRECT(txm) &&
1984 		     txm->nb_segs == 1 &&
1985 		     rte_pktmbuf_headroom(txm) >= hdr_size &&
1986 		     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1987 				__alignof__(struct virtio_net_hdr_mrg_rxbuf))) {
1988 			inorder_pkts[nb_inorder_pkts] = txm;
1989 			nb_inorder_pkts++;
1990 
1991 			continue;
1992 		}
1993 
1994 		if (nb_inorder_pkts) {
1995 			need = nb_inorder_pkts - vq->vq_free_cnt;
1996 			if (unlikely(need > 0)) {
1997 				need = virtio_xmit_try_cleanup_inorder(vq,
1998 								       need);
1999 				if (unlikely(need > 0)) {
2000 					PMD_TX_LOG(ERR,
2001 						"No free tx descriptors to "
2002 						"transmit");
2003 					break;
2004 				}
2005 			}
2006 			virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2007 							nb_inorder_pkts);
2008 			nb_inorder_pkts = 0;
2009 		}
2010 
2011 		slots = txm->nb_segs + 1;
2012 		need = slots - vq->vq_free_cnt;
2013 		if (unlikely(need > 0)) {
2014 			need = virtio_xmit_try_cleanup_inorder(vq, slots);
2015 
2016 			if (unlikely(need > 0)) {
2017 				PMD_TX_LOG(ERR,
2018 					"No free tx descriptors to transmit");
2019 				break;
2020 			}
2021 		}
2022 		/* Enqueue Packet buffers */
2023 		virtqueue_enqueue_xmit(txvq, txm, slots, 0, 0, 1);
2024 
2025 		virtio_update_packet_stats(&txvq->stats, txm);
2026 	}
2027 
2028 	/* Transmit all inorder packets */
2029 	if (nb_inorder_pkts) {
2030 		need = nb_inorder_pkts - vq->vq_free_cnt;
2031 		if (unlikely(need > 0)) {
2032 			need = virtio_xmit_try_cleanup_inorder(vq,
2033 								  need);
2034 			if (unlikely(need > 0)) {
2035 				PMD_TX_LOG(ERR,
2036 					"No free tx descriptors to transmit");
2037 				nb_inorder_pkts = vq->vq_free_cnt;
2038 				nb_tx -= need;
2039 			}
2040 		}
2041 
2042 		virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2043 						nb_inorder_pkts);
2044 	}
2045 
2046 	txvq->stats.packets += nb_tx;
2047 
2048 	if (likely(nb_tx)) {
2049 		vq_update_avail_idx(vq);
2050 
2051 		if (unlikely(virtqueue_kick_prepare(vq))) {
2052 			virtqueue_notify(vq);
2053 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2054 		}
2055 	}
2056 
2057 	VIRTQUEUE_DUMP(vq);
2058 
2059 	return nb_tx;
2060 }
2061 
2062 __rte_weak uint16_t
2063 virtio_recv_pkts_packed_vec(void *rx_queue __rte_unused,
2064 			    struct rte_mbuf **rx_pkts __rte_unused,
2065 			    uint16_t nb_pkts __rte_unused)
2066 {
2067 	return 0;
2068 }
2069 
2070 __rte_weak uint16_t
2071 virtio_xmit_pkts_packed_vec(void *tx_queue __rte_unused,
2072 			    struct rte_mbuf **tx_pkts __rte_unused,
2073 			    uint16_t nb_pkts __rte_unused)
2074 {
2075 	return 0;
2076 }
2077