xref: /dpdk/drivers/net/virtio/virtio_rxtx.c (revision 03ab51eafda992874a48c392ca66ffb577fe2b71)
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 = VIRTIO_MBUF_ADDR(cookies[i], vq) +
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 = VIRTIO_MBUF_ADDR(cookie[i], vq) +
314 			RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
315 		start_dp[idx].len = cookie[i]->buf_len - RTE_PKTMBUF_HEADROOM +
316 			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 = VIRTIO_MBUF_ADDR(cookie, vq) + RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
340 	dp->len = cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
341 
342 	virtqueue_store_flags_packed(dp, flags, hw->weak_barriers);
343 
344 	if (++vq->vq_avail_idx >= vq->vq_nentries) {
345 		vq->vq_avail_idx -= vq->vq_nentries;
346 		vq->vq_packed.cached_flags ^=
347 			VRING_PACKED_DESC_F_AVAIL_USED;
348 		flags = vq->vq_packed.cached_flags;
349 	}
350 }
351 
352 static inline int
353 virtqueue_enqueue_recv_refill_packed_init(struct virtqueue *vq,
354 				     struct rte_mbuf **cookie, uint16_t num)
355 {
356 	struct vring_packed_desc *start_dp = vq->vq_packed.ring.desc;
357 	struct vq_desc_extra *dxp;
358 	uint16_t idx;
359 	int i;
360 
361 	if (unlikely(vq->vq_free_cnt == 0))
362 		return -ENOSPC;
363 	if (unlikely(vq->vq_free_cnt < num))
364 		return -EMSGSIZE;
365 
366 	for (i = 0; i < num; i++) {
367 		idx = vq->vq_avail_idx;
368 		dxp = &vq->vq_descx[idx];
369 		dxp->cookie = (void *)cookie[i];
370 		dxp->ndescs = 1;
371 
372 		virtqueue_refill_single_packed(vq, &start_dp[idx], cookie[i]);
373 	}
374 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
375 	return 0;
376 }
377 
378 static inline int
379 virtqueue_enqueue_recv_refill_packed(struct virtqueue *vq,
380 				     struct rte_mbuf **cookie, uint16_t num)
381 {
382 	struct vring_packed_desc *start_dp = vq->vq_packed.ring.desc;
383 	struct vq_desc_extra *dxp;
384 	uint16_t idx, did;
385 	int i;
386 
387 	if (unlikely(vq->vq_free_cnt == 0))
388 		return -ENOSPC;
389 	if (unlikely(vq->vq_free_cnt < num))
390 		return -EMSGSIZE;
391 
392 	for (i = 0; i < num; i++) {
393 		idx = vq->vq_avail_idx;
394 		did = start_dp[idx].id;
395 		dxp = &vq->vq_descx[did];
396 		dxp->cookie = (void *)cookie[i];
397 		dxp->ndescs = 1;
398 
399 		virtqueue_refill_single_packed(vq, &start_dp[idx], cookie[i]);
400 	}
401 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
402 	return 0;
403 }
404 
405 /* When doing TSO, the IP length is not included in the pseudo header
406  * checksum of the packet given to the PMD, but for virtio it is
407  * expected.
408  */
409 static void
410 virtio_tso_fix_cksum(struct rte_mbuf *m)
411 {
412 	/* common case: header is not fragmented */
413 	if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
414 			m->l4_len)) {
415 		struct rte_ipv4_hdr *iph;
416 		struct rte_ipv6_hdr *ip6h;
417 		struct rte_tcp_hdr *th;
418 		uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
419 		uint32_t tmp;
420 
421 		iph = rte_pktmbuf_mtod_offset(m,
422 					struct rte_ipv4_hdr *, m->l2_len);
423 		th = RTE_PTR_ADD(iph, m->l3_len);
424 		if ((iph->version_ihl >> 4) == 4) {
425 			iph->hdr_checksum = 0;
426 			iph->hdr_checksum = rte_ipv4_cksum(iph);
427 			ip_len = iph->total_length;
428 			ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
429 				m->l3_len);
430 		} else {
431 			ip6h = (struct rte_ipv6_hdr *)iph;
432 			ip_paylen = ip6h->payload_len;
433 		}
434 
435 		/* calculate the new phdr checksum not including ip_paylen */
436 		prev_cksum = th->cksum;
437 		tmp = prev_cksum;
438 		tmp += ip_paylen;
439 		tmp = (tmp & 0xffff) + (tmp >> 16);
440 		new_cksum = tmp;
441 
442 		/* replace it in the packet */
443 		th->cksum = new_cksum;
444 	}
445 }
446 
447 
448 
449 
450 static inline void
451 virtqueue_enqueue_xmit_inorder(struct virtnet_tx *txvq,
452 			struct rte_mbuf **cookies,
453 			uint16_t num)
454 {
455 	struct vq_desc_extra *dxp;
456 	struct virtqueue *vq = virtnet_txq_to_vq(txvq);
457 	struct vring_desc *start_dp;
458 	struct virtio_net_hdr *hdr;
459 	uint16_t idx;
460 	int16_t head_size = vq->hw->vtnet_hdr_size;
461 	uint16_t i = 0;
462 
463 	idx = vq->vq_desc_head_idx;
464 	start_dp = vq->vq_split.ring.desc;
465 
466 	while (i < num) {
467 		idx = idx & (vq->vq_nentries - 1);
468 		dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
469 		dxp->cookie = (void *)cookies[i];
470 		dxp->ndescs = 1;
471 		virtio_update_packet_stats(&txvq->stats, cookies[i]);
472 
473 		hdr = rte_pktmbuf_mtod_offset(cookies[i],
474 				struct virtio_net_hdr *, -head_size);
475 
476 		/* if offload disabled, hdr is not zeroed yet, do it now */
477 		if (!vq->hw->has_tx_offload)
478 			virtqueue_clear_net_hdr(hdr);
479 		else
480 			virtqueue_xmit_offload(hdr, cookies[i]);
481 
482 		start_dp[idx].addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookies[i], vq) - head_size;
483 		start_dp[idx].len = cookies[i]->data_len + head_size;
484 		start_dp[idx].flags = 0;
485 
486 
487 		vq_update_avail_ring(vq, idx);
488 
489 		idx++;
490 		i++;
491 	};
492 
493 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
494 	vq->vq_desc_head_idx = idx & (vq->vq_nentries - 1);
495 }
496 
497 static inline void
498 virtqueue_enqueue_xmit_packed_fast(struct virtnet_tx *txvq,
499 				   struct rte_mbuf *cookie,
500 				   int in_order)
501 {
502 	struct virtqueue *vq = virtnet_txq_to_vq(txvq);
503 	struct vring_packed_desc *dp;
504 	struct vq_desc_extra *dxp;
505 	uint16_t idx, id, flags;
506 	int16_t head_size = vq->hw->vtnet_hdr_size;
507 	struct virtio_net_hdr *hdr;
508 
509 	id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx;
510 	idx = vq->vq_avail_idx;
511 	dp = &vq->vq_packed.ring.desc[idx];
512 
513 	dxp = &vq->vq_descx[id];
514 	dxp->ndescs = 1;
515 	dxp->cookie = cookie;
516 
517 	flags = vq->vq_packed.cached_flags;
518 
519 	/* prepend cannot fail, checked by caller */
520 	hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
521 				      -head_size);
522 
523 	/* if offload disabled, hdr is not zeroed yet, do it now */
524 	if (!vq->hw->has_tx_offload)
525 		virtqueue_clear_net_hdr(hdr);
526 	else
527 		virtqueue_xmit_offload(hdr, cookie);
528 
529 	dp->addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq) - head_size;
530 	dp->len = cookie->data_len + head_size;
531 	dp->id = id;
532 
533 	if (++vq->vq_avail_idx >= vq->vq_nentries) {
534 		vq->vq_avail_idx -= vq->vq_nentries;
535 		vq->vq_packed.cached_flags ^= VRING_PACKED_DESC_F_AVAIL_USED;
536 	}
537 
538 	vq->vq_free_cnt--;
539 
540 	if (!in_order) {
541 		vq->vq_desc_head_idx = dxp->next;
542 		if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
543 			vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
544 	}
545 
546 	virtqueue_store_flags_packed(dp, flags, vq->hw->weak_barriers);
547 }
548 
549 static inline void
550 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
551 			uint16_t needed, int use_indirect, int can_push,
552 			int in_order)
553 {
554 	struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
555 	struct vq_desc_extra *dxp;
556 	struct virtqueue *vq = virtnet_txq_to_vq(txvq);
557 	struct vring_desc *start_dp;
558 	uint16_t seg_num = cookie->nb_segs;
559 	uint16_t head_idx, idx;
560 	int16_t head_size = vq->hw->vtnet_hdr_size;
561 	bool prepend_header = false;
562 	struct virtio_net_hdr *hdr;
563 
564 	head_idx = vq->vq_desc_head_idx;
565 	idx = head_idx;
566 	if (in_order)
567 		dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
568 	else
569 		dxp = &vq->vq_descx[idx];
570 	dxp->cookie = (void *)cookie;
571 	dxp->ndescs = needed;
572 
573 	start_dp = vq->vq_split.ring.desc;
574 
575 	if (can_push) {
576 		/* prepend cannot fail, checked by caller */
577 		hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
578 					      -head_size);
579 		prepend_header = true;
580 
581 		/* if offload disabled, it is not zeroed below, do it now */
582 		if (!vq->hw->has_tx_offload)
583 			virtqueue_clear_net_hdr(hdr);
584 	} else if (use_indirect) {
585 		/* setup tx ring slot to point to indirect
586 		 * descriptor list stored in reserved region.
587 		 *
588 		 * the first slot in indirect ring is already preset
589 		 * to point to the header in reserved region
590 		 */
591 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
592 			RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
593 		start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
594 		start_dp[idx].flags = VRING_DESC_F_INDIRECT;
595 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
596 
597 		/* loop below will fill in rest of the indirect elements */
598 		start_dp = txr[idx].tx_indir;
599 		idx = 1;
600 	} else {
601 		/* setup first tx ring slot to point to header
602 		 * stored in reserved region.
603 		 */
604 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
605 			RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
606 		start_dp[idx].len   = vq->hw->vtnet_hdr_size;
607 		start_dp[idx].flags = VRING_DESC_F_NEXT;
608 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
609 
610 		idx = start_dp[idx].next;
611 	}
612 
613 	if (vq->hw->has_tx_offload)
614 		virtqueue_xmit_offload(hdr, cookie);
615 
616 	do {
617 		start_dp[idx].addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
618 		start_dp[idx].len = cookie->data_len;
619 		if (prepend_header) {
620 			start_dp[idx].addr -= head_size;
621 			start_dp[idx].len += head_size;
622 			prepend_header = false;
623 		}
624 		start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
625 		idx = start_dp[idx].next;
626 	} while ((cookie = cookie->next) != NULL);
627 
628 	if (use_indirect)
629 		idx = vq->vq_split.ring.desc[head_idx].next;
630 
631 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
632 
633 	vq->vq_desc_head_idx = idx;
634 	vq_update_avail_ring(vq, head_idx);
635 
636 	if (!in_order) {
637 		if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
638 			vq->vq_desc_tail_idx = idx;
639 	}
640 }
641 
642 void
643 virtio_dev_cq_start(struct rte_eth_dev *dev)
644 {
645 	struct virtio_hw *hw = dev->data->dev_private;
646 
647 	if (hw->cvq) {
648 		rte_spinlock_init(&hw->cvq->lock);
649 		VIRTQUEUE_DUMP(virtnet_cq_to_vq(hw->cvq));
650 	}
651 }
652 
653 int
654 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
655 			uint16_t queue_idx,
656 			uint16_t nb_desc,
657 			unsigned int socket_id __rte_unused,
658 			const struct rte_eth_rxconf *rx_conf,
659 			struct rte_mempool *mp)
660 {
661 	uint16_t vq_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
662 	struct virtio_hw *hw = dev->data->dev_private;
663 	struct virtqueue *vq = hw->vqs[vq_idx];
664 	struct virtnet_rx *rxvq;
665 	uint16_t rx_free_thresh;
666 	uint16_t buf_size;
667 	const char *error;
668 
669 	PMD_INIT_FUNC_TRACE();
670 
671 	if (rx_conf->rx_deferred_start) {
672 		PMD_INIT_LOG(ERR, "Rx deferred start is not supported");
673 		return -EINVAL;
674 	}
675 
676 	buf_size = virtio_rx_mem_pool_buf_size(mp);
677 	if (!virtio_rx_check_scatter(hw->max_rx_pkt_len, buf_size,
678 				     hw->rx_ol_scatter, &error)) {
679 		PMD_INIT_LOG(ERR, "RxQ %u Rx scatter check failed: %s",
680 			     queue_idx, error);
681 		return -EINVAL;
682 	}
683 
684 	rx_free_thresh = rx_conf->rx_free_thresh;
685 	if (rx_free_thresh == 0)
686 		rx_free_thresh =
687 			RTE_MIN(vq->vq_nentries / 4, DEFAULT_RX_FREE_THRESH);
688 
689 	if (rx_free_thresh & 0x3) {
690 		PMD_INIT_LOG(ERR, "rx_free_thresh must be multiples of four."
691 			" (rx_free_thresh=%u port=%u queue=%u)",
692 			rx_free_thresh, dev->data->port_id, queue_idx);
693 		return -EINVAL;
694 	}
695 
696 	if (rx_free_thresh >= vq->vq_nentries) {
697 		PMD_INIT_LOG(ERR, "rx_free_thresh must be less than the "
698 			"number of RX entries (%u)."
699 			" (rx_free_thresh=%u port=%u queue=%u)",
700 			vq->vq_nentries,
701 			rx_free_thresh, dev->data->port_id, queue_idx);
702 		return -EINVAL;
703 	}
704 	vq->vq_free_thresh = rx_free_thresh;
705 
706 	/*
707 	 * For split ring vectorized path descriptors number must be
708 	 * equal to the ring size.
709 	 */
710 	if (nb_desc > vq->vq_nentries ||
711 	    (!virtio_with_packed_queue(hw) && hw->use_vec_rx)) {
712 		nb_desc = vq->vq_nentries;
713 	}
714 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
715 
716 	rxvq = &vq->rxq;
717 	rxvq->queue_id = queue_idx;
718 	rxvq->mpool = mp;
719 	dev->data->rx_queues[queue_idx] = rxvq;
720 
721 	return 0;
722 }
723 
724 int
725 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
726 {
727 	uint16_t vq_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
728 	struct virtio_hw *hw = dev->data->dev_private;
729 	struct virtqueue *vq = hw->vqs[vq_idx];
730 	struct virtnet_rx *rxvq = &vq->rxq;
731 	struct rte_mbuf *m;
732 	uint16_t desc_idx;
733 	int error, nbufs, i;
734 	bool in_order = virtio_with_feature(hw, VIRTIO_F_IN_ORDER);
735 
736 	PMD_INIT_FUNC_TRACE();
737 
738 	/* Allocate blank mbufs for the each rx descriptor */
739 	nbufs = 0;
740 
741 	if (hw->use_vec_rx && !virtio_with_packed_queue(hw)) {
742 		for (desc_idx = 0; desc_idx < vq->vq_nentries;
743 		     desc_idx++) {
744 			vq->vq_split.ring.avail->ring[desc_idx] = desc_idx;
745 			vq->vq_split.ring.desc[desc_idx].flags =
746 				VRING_DESC_F_WRITE;
747 		}
748 
749 		virtio_rxq_vec_setup(rxvq);
750 	}
751 
752 	memset(rxvq->fake_mbuf, 0, sizeof(*rxvq->fake_mbuf));
753 	for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST; desc_idx++)
754 		vq->sw_ring[vq->vq_nentries + desc_idx] = rxvq->fake_mbuf;
755 
756 	if (hw->use_vec_rx && !virtio_with_packed_queue(hw)) {
757 		while (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
758 			virtio_rxq_rearm_vec(rxvq);
759 			nbufs += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
760 		}
761 	} else if (!virtio_with_packed_queue(vq->hw) && in_order) {
762 		if ((!virtqueue_full(vq))) {
763 			uint16_t free_cnt = vq->vq_free_cnt;
764 			struct rte_mbuf *pkts[free_cnt];
765 
766 			if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, pkts,
767 				free_cnt)) {
768 				error = virtqueue_enqueue_refill_inorder(vq,
769 						pkts,
770 						free_cnt);
771 				if (unlikely(error)) {
772 					for (i = 0; i < free_cnt; i++)
773 						rte_pktmbuf_free(pkts[i]);
774 				} else {
775 					nbufs += free_cnt;
776 				}
777 			}
778 
779 			vq_update_avail_idx(vq);
780 		}
781 	} else {
782 		while (!virtqueue_full(vq)) {
783 			m = rte_mbuf_raw_alloc(rxvq->mpool);
784 			if (m == NULL)
785 				break;
786 
787 			/* Enqueue allocated buffers */
788 			if (virtio_with_packed_queue(vq->hw))
789 				error = virtqueue_enqueue_recv_refill_packed_init(vq,
790 						&m, 1);
791 			else
792 				error = virtqueue_enqueue_recv_refill(vq,
793 						&m, 1);
794 			if (error) {
795 				rte_pktmbuf_free(m);
796 				break;
797 			}
798 			nbufs++;
799 		}
800 
801 		if (!virtio_with_packed_queue(vq->hw))
802 			vq_update_avail_idx(vq);
803 	}
804 
805 	PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
806 
807 	VIRTQUEUE_DUMP(vq);
808 
809 	return 0;
810 }
811 
812 /*
813  * struct rte_eth_dev *dev: Used to update dev
814  * uint16_t nb_desc: Defaults to values read from config space
815  * unsigned int socket_id: Used to allocate memzone
816  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
817  * uint16_t queue_idx: Just used as an index in dev txq list
818  */
819 int
820 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
821 			uint16_t queue_idx,
822 			uint16_t nb_desc,
823 			unsigned int socket_id __rte_unused,
824 			const struct rte_eth_txconf *tx_conf)
825 {
826 	uint8_t vq_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
827 	struct virtio_hw *hw = dev->data->dev_private;
828 	struct virtqueue *vq = hw->vqs[vq_idx];
829 	struct virtnet_tx *txvq;
830 	uint16_t tx_free_thresh;
831 
832 	PMD_INIT_FUNC_TRACE();
833 
834 	if (tx_conf->tx_deferred_start) {
835 		PMD_INIT_LOG(ERR, "Tx deferred start is not supported");
836 		return -EINVAL;
837 	}
838 
839 	if (nb_desc == 0 || nb_desc > vq->vq_nentries)
840 		nb_desc = vq->vq_nentries;
841 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
842 
843 	txvq = &vq->txq;
844 	txvq->queue_id = queue_idx;
845 
846 	tx_free_thresh = tx_conf->tx_free_thresh;
847 	if (tx_free_thresh == 0)
848 		tx_free_thresh =
849 			RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
850 
851 	if (tx_free_thresh >= (vq->vq_nentries - 3)) {
852 		PMD_DRV_LOG(ERR, "tx_free_thresh must be less than the "
853 			"number of TX entries minus 3 (%u)."
854 			" (tx_free_thresh=%u port=%u queue=%u)",
855 			vq->vq_nentries - 3,
856 			tx_free_thresh, dev->data->port_id, queue_idx);
857 		return -EINVAL;
858 	}
859 
860 	vq->vq_free_thresh = tx_free_thresh;
861 
862 	dev->data->tx_queues[queue_idx] = txvq;
863 	return 0;
864 }
865 
866 int
867 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
868 				uint16_t queue_idx)
869 {
870 	uint8_t vq_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
871 	struct virtio_hw *hw = dev->data->dev_private;
872 	struct virtqueue *vq = hw->vqs[vq_idx];
873 
874 	PMD_INIT_FUNC_TRACE();
875 
876 	if (!virtio_with_packed_queue(hw)) {
877 		if (virtio_with_feature(hw, VIRTIO_F_IN_ORDER))
878 			vq->vq_split.ring.desc[vq->vq_nentries - 1].next = 0;
879 	}
880 
881 	VIRTQUEUE_DUMP(vq);
882 
883 	return 0;
884 }
885 
886 static inline void
887 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
888 {
889 	int error;
890 	/*
891 	 * Requeue the discarded mbuf. This should always be
892 	 * successful since it was just dequeued.
893 	 */
894 	if (virtio_with_packed_queue(vq->hw))
895 		error = virtqueue_enqueue_recv_refill_packed(vq, &m, 1);
896 	else
897 		error = virtqueue_enqueue_recv_refill(vq, &m, 1);
898 
899 	if (unlikely(error)) {
900 		PMD_DRV_LOG(ERR, "cannot requeue discarded mbuf");
901 		rte_pktmbuf_free(m);
902 	}
903 }
904 
905 static inline void
906 virtio_discard_rxbuf_inorder(struct virtqueue *vq, struct rte_mbuf *m)
907 {
908 	int error;
909 
910 	error = virtqueue_enqueue_refill_inorder(vq, &m, 1);
911 	if (unlikely(error)) {
912 		PMD_DRV_LOG(ERR, "cannot requeue discarded mbuf");
913 		rte_pktmbuf_free(m);
914 	}
915 }
916 
917 /* Optionally fill offload information in structure */
918 static inline int
919 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
920 {
921 	struct rte_net_hdr_lens hdr_lens;
922 	uint32_t hdrlen, ptype;
923 	int l4_supported = 0;
924 
925 	/* nothing to do */
926 	if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
927 		return 0;
928 
929 	m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
930 
931 	ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
932 	m->packet_type = ptype;
933 	if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
934 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
935 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
936 		l4_supported = 1;
937 
938 	if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
939 		hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
940 		if (hdr->csum_start <= hdrlen && l4_supported) {
941 			m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
942 		} else {
943 			/* Unknown proto or tunnel, do sw cksum. We can assume
944 			 * the cksum field is in the first segment since the
945 			 * buffers we provided to the host are large enough.
946 			 * In case of SCTP, this will be wrong since it's a CRC
947 			 * but there's nothing we can do.
948 			 */
949 			uint16_t csum = 0, off;
950 
951 			if (rte_raw_cksum_mbuf(m, hdr->csum_start,
952 				rte_pktmbuf_pkt_len(m) - hdr->csum_start,
953 				&csum) < 0)
954 				return -EINVAL;
955 			if (likely(csum != 0xffff))
956 				csum = ~csum;
957 			off = hdr->csum_offset + hdr->csum_start;
958 			if (rte_pktmbuf_data_len(m) >= off + 1)
959 				*rte_pktmbuf_mtod_offset(m, uint16_t *,
960 					off) = csum;
961 		}
962 	} else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
963 		m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
964 	}
965 
966 	/* GSO request, save required information in mbuf */
967 	if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
968 		/* Check unsupported modes */
969 		if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
970 		    (hdr->gso_size == 0)) {
971 			return -EINVAL;
972 		}
973 
974 		/* Update mss lengthes in mbuf */
975 		m->tso_segsz = hdr->gso_size;
976 		switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
977 			case VIRTIO_NET_HDR_GSO_TCPV4:
978 			case VIRTIO_NET_HDR_GSO_TCPV6:
979 				m->ol_flags |= PKT_RX_LRO | \
980 					PKT_RX_L4_CKSUM_NONE;
981 				break;
982 			default:
983 				return -EINVAL;
984 		}
985 	}
986 
987 	return 0;
988 }
989 
990 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
991 uint16_t
992 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
993 {
994 	struct virtnet_rx *rxvq = rx_queue;
995 	struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
996 	struct virtio_hw *hw = vq->hw;
997 	struct rte_mbuf *rxm;
998 	uint16_t nb_used, num, nb_rx;
999 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1000 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1001 	int error;
1002 	uint32_t i, nb_enqueued;
1003 	uint32_t hdr_size;
1004 	struct virtio_net_hdr *hdr;
1005 
1006 	nb_rx = 0;
1007 	if (unlikely(hw->started == 0))
1008 		return nb_rx;
1009 
1010 	nb_used = virtqueue_nused(vq);
1011 
1012 	num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1013 	if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1014 		num = VIRTIO_MBUF_BURST_SZ;
1015 	if (likely(num > DESC_PER_CACHELINE))
1016 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1017 
1018 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1019 	PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
1020 
1021 	nb_enqueued = 0;
1022 	hdr_size = hw->vtnet_hdr_size;
1023 
1024 	for (i = 0; i < num ; i++) {
1025 		rxm = rcv_pkts[i];
1026 
1027 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1028 
1029 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1030 			PMD_RX_LOG(ERR, "Packet drop");
1031 			nb_enqueued++;
1032 			virtio_discard_rxbuf(vq, rxm);
1033 			rxvq->stats.errors++;
1034 			continue;
1035 		}
1036 
1037 		rxm->port = rxvq->port_id;
1038 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1039 		rxm->ol_flags = 0;
1040 		rxm->vlan_tci = 0;
1041 
1042 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1043 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1044 
1045 		hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1046 			RTE_PKTMBUF_HEADROOM - hdr_size);
1047 
1048 		if (hw->vlan_strip)
1049 			rte_vlan_strip(rxm);
1050 
1051 		if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1052 			virtio_discard_rxbuf(vq, rxm);
1053 			rxvq->stats.errors++;
1054 			continue;
1055 		}
1056 
1057 		virtio_rx_stats_updated(rxvq, rxm);
1058 
1059 		rx_pkts[nb_rx++] = rxm;
1060 	}
1061 
1062 	rxvq->stats.packets += nb_rx;
1063 
1064 	/* Allocate new mbuf for the used descriptor */
1065 	if (likely(!virtqueue_full(vq))) {
1066 		uint16_t free_cnt = vq->vq_free_cnt;
1067 		struct rte_mbuf *new_pkts[free_cnt];
1068 
1069 		if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1070 						free_cnt) == 0)) {
1071 			error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1072 					free_cnt);
1073 			if (unlikely(error)) {
1074 				for (i = 0; i < free_cnt; i++)
1075 					rte_pktmbuf_free(new_pkts[i]);
1076 			}
1077 			nb_enqueued += free_cnt;
1078 		} else {
1079 			struct rte_eth_dev *dev =
1080 				&rte_eth_devices[rxvq->port_id];
1081 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1082 		}
1083 	}
1084 
1085 	if (likely(nb_enqueued)) {
1086 		vq_update_avail_idx(vq);
1087 
1088 		if (unlikely(virtqueue_kick_prepare(vq))) {
1089 			virtqueue_notify(vq);
1090 			PMD_RX_LOG(DEBUG, "Notified");
1091 		}
1092 	}
1093 
1094 	return nb_rx;
1095 }
1096 
1097 uint16_t
1098 virtio_recv_pkts_packed(void *rx_queue, struct rte_mbuf **rx_pkts,
1099 			uint16_t nb_pkts)
1100 {
1101 	struct virtnet_rx *rxvq = rx_queue;
1102 	struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1103 	struct virtio_hw *hw = vq->hw;
1104 	struct rte_mbuf *rxm;
1105 	uint16_t num, nb_rx;
1106 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1107 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1108 	int error;
1109 	uint32_t i, nb_enqueued;
1110 	uint32_t hdr_size;
1111 	struct virtio_net_hdr *hdr;
1112 
1113 	nb_rx = 0;
1114 	if (unlikely(hw->started == 0))
1115 		return nb_rx;
1116 
1117 	num = RTE_MIN(VIRTIO_MBUF_BURST_SZ, nb_pkts);
1118 	if (likely(num > DESC_PER_CACHELINE))
1119 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1120 
1121 	num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1122 	PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1123 
1124 	nb_enqueued = 0;
1125 	hdr_size = hw->vtnet_hdr_size;
1126 
1127 	for (i = 0; i < num; i++) {
1128 		rxm = rcv_pkts[i];
1129 
1130 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1131 
1132 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1133 			PMD_RX_LOG(ERR, "Packet drop");
1134 			nb_enqueued++;
1135 			virtio_discard_rxbuf(vq, rxm);
1136 			rxvq->stats.errors++;
1137 			continue;
1138 		}
1139 
1140 		rxm->port = rxvq->port_id;
1141 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1142 		rxm->ol_flags = 0;
1143 		rxm->vlan_tci = 0;
1144 
1145 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1146 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1147 
1148 		hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1149 			RTE_PKTMBUF_HEADROOM - hdr_size);
1150 
1151 		if (hw->vlan_strip)
1152 			rte_vlan_strip(rxm);
1153 
1154 		if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1155 			virtio_discard_rxbuf(vq, rxm);
1156 			rxvq->stats.errors++;
1157 			continue;
1158 		}
1159 
1160 		virtio_rx_stats_updated(rxvq, rxm);
1161 
1162 		rx_pkts[nb_rx++] = rxm;
1163 	}
1164 
1165 	rxvq->stats.packets += nb_rx;
1166 
1167 	/* Allocate new mbuf for the used descriptor */
1168 	if (likely(!virtqueue_full(vq))) {
1169 		uint16_t free_cnt = vq->vq_free_cnt;
1170 		struct rte_mbuf *new_pkts[free_cnt];
1171 
1172 		if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1173 						free_cnt) == 0)) {
1174 			error = virtqueue_enqueue_recv_refill_packed(vq,
1175 					new_pkts, free_cnt);
1176 			if (unlikely(error)) {
1177 				for (i = 0; i < free_cnt; i++)
1178 					rte_pktmbuf_free(new_pkts[i]);
1179 			}
1180 			nb_enqueued += free_cnt;
1181 		} else {
1182 			struct rte_eth_dev *dev =
1183 				&rte_eth_devices[rxvq->port_id];
1184 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1185 		}
1186 	}
1187 
1188 	if (likely(nb_enqueued)) {
1189 		if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1190 			virtqueue_notify(vq);
1191 			PMD_RX_LOG(DEBUG, "Notified");
1192 		}
1193 	}
1194 
1195 	return nb_rx;
1196 }
1197 
1198 
1199 uint16_t
1200 virtio_recv_pkts_inorder(void *rx_queue,
1201 			struct rte_mbuf **rx_pkts,
1202 			uint16_t nb_pkts)
1203 {
1204 	struct virtnet_rx *rxvq = rx_queue;
1205 	struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1206 	struct virtio_hw *hw = vq->hw;
1207 	struct rte_mbuf *rxm;
1208 	struct rte_mbuf *prev = NULL;
1209 	uint16_t nb_used, num, nb_rx;
1210 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1211 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1212 	int error;
1213 	uint32_t nb_enqueued;
1214 	uint32_t seg_num;
1215 	uint32_t seg_res;
1216 	uint32_t hdr_size;
1217 	int32_t i;
1218 
1219 	nb_rx = 0;
1220 	if (unlikely(hw->started == 0))
1221 		return nb_rx;
1222 
1223 	nb_used = virtqueue_nused(vq);
1224 	nb_used = RTE_MIN(nb_used, nb_pkts);
1225 	nb_used = RTE_MIN(nb_used, VIRTIO_MBUF_BURST_SZ);
1226 
1227 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1228 
1229 	nb_enqueued = 0;
1230 	seg_num = 1;
1231 	seg_res = 0;
1232 	hdr_size = hw->vtnet_hdr_size;
1233 
1234 	num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len, nb_used);
1235 
1236 	for (i = 0; i < num; i++) {
1237 		struct virtio_net_hdr_mrg_rxbuf *header;
1238 
1239 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1240 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1241 
1242 		rxm = rcv_pkts[i];
1243 
1244 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1245 			PMD_RX_LOG(ERR, "Packet drop");
1246 			nb_enqueued++;
1247 			virtio_discard_rxbuf_inorder(vq, rxm);
1248 			rxvq->stats.errors++;
1249 			continue;
1250 		}
1251 
1252 		header = (struct virtio_net_hdr_mrg_rxbuf *)
1253 			 ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1254 			 - hdr_size);
1255 
1256 		if (virtio_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
1257 			seg_num = header->num_buffers;
1258 			if (seg_num == 0)
1259 				seg_num = 1;
1260 		} else {
1261 			seg_num = 1;
1262 		}
1263 
1264 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1265 		rxm->nb_segs = seg_num;
1266 		rxm->ol_flags = 0;
1267 		rxm->vlan_tci = 0;
1268 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1269 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1270 
1271 		rxm->port = rxvq->port_id;
1272 
1273 		rx_pkts[nb_rx] = rxm;
1274 		prev = rxm;
1275 
1276 		if (vq->hw->has_rx_offload &&
1277 				virtio_rx_offload(rxm, &header->hdr) < 0) {
1278 			virtio_discard_rxbuf_inorder(vq, rxm);
1279 			rxvq->stats.errors++;
1280 			continue;
1281 		}
1282 
1283 		if (hw->vlan_strip)
1284 			rte_vlan_strip(rx_pkts[nb_rx]);
1285 
1286 		seg_res = seg_num - 1;
1287 
1288 		/* Merge remaining segments */
1289 		while (seg_res != 0 && i < (num - 1)) {
1290 			i++;
1291 
1292 			rxm = rcv_pkts[i];
1293 			rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1294 			rxm->pkt_len = (uint32_t)(len[i]);
1295 			rxm->data_len = (uint16_t)(len[i]);
1296 
1297 			rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1298 
1299 			prev->next = rxm;
1300 			prev = rxm;
1301 			seg_res -= 1;
1302 		}
1303 
1304 		if (!seg_res) {
1305 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1306 			nb_rx++;
1307 		}
1308 	}
1309 
1310 	/* Last packet still need merge segments */
1311 	while (seg_res != 0) {
1312 		uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1313 					VIRTIO_MBUF_BURST_SZ);
1314 
1315 		if (likely(virtqueue_nused(vq) >= rcv_cnt)) {
1316 			num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len,
1317 							   rcv_cnt);
1318 			uint16_t extra_idx = 0;
1319 
1320 			rcv_cnt = num;
1321 			while (extra_idx < rcv_cnt) {
1322 				rxm = rcv_pkts[extra_idx];
1323 				rxm->data_off =
1324 					RTE_PKTMBUF_HEADROOM - hdr_size;
1325 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
1326 				rxm->data_len = (uint16_t)(len[extra_idx]);
1327 				prev->next = rxm;
1328 				prev = rxm;
1329 				rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1330 				extra_idx += 1;
1331 			};
1332 			seg_res -= rcv_cnt;
1333 
1334 			if (!seg_res) {
1335 				virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1336 				nb_rx++;
1337 			}
1338 		} else {
1339 			PMD_RX_LOG(ERR,
1340 					"No enough segments for packet.");
1341 			rte_pktmbuf_free(rx_pkts[nb_rx]);
1342 			rxvq->stats.errors++;
1343 			break;
1344 		}
1345 	}
1346 
1347 	rxvq->stats.packets += nb_rx;
1348 
1349 	/* Allocate new mbuf for the used descriptor */
1350 
1351 	if (likely(!virtqueue_full(vq))) {
1352 		/* free_cnt may include mrg descs */
1353 		uint16_t free_cnt = vq->vq_free_cnt;
1354 		struct rte_mbuf *new_pkts[free_cnt];
1355 
1356 		if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1357 			error = virtqueue_enqueue_refill_inorder(vq, new_pkts,
1358 					free_cnt);
1359 			if (unlikely(error)) {
1360 				for (i = 0; i < free_cnt; i++)
1361 					rte_pktmbuf_free(new_pkts[i]);
1362 			}
1363 			nb_enqueued += free_cnt;
1364 		} else {
1365 			struct rte_eth_dev *dev =
1366 				&rte_eth_devices[rxvq->port_id];
1367 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1368 		}
1369 	}
1370 
1371 	if (likely(nb_enqueued)) {
1372 		vq_update_avail_idx(vq);
1373 
1374 		if (unlikely(virtqueue_kick_prepare(vq))) {
1375 			virtqueue_notify(vq);
1376 			PMD_RX_LOG(DEBUG, "Notified");
1377 		}
1378 	}
1379 
1380 	return nb_rx;
1381 }
1382 
1383 uint16_t
1384 virtio_recv_mergeable_pkts(void *rx_queue,
1385 			struct rte_mbuf **rx_pkts,
1386 			uint16_t nb_pkts)
1387 {
1388 	struct virtnet_rx *rxvq = rx_queue;
1389 	struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1390 	struct virtio_hw *hw = vq->hw;
1391 	struct rte_mbuf *rxm;
1392 	struct rte_mbuf *prev = NULL;
1393 	uint16_t nb_used, num, nb_rx = 0;
1394 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1395 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1396 	int error;
1397 	uint32_t nb_enqueued = 0;
1398 	uint32_t seg_num = 0;
1399 	uint32_t seg_res = 0;
1400 	uint32_t hdr_size = hw->vtnet_hdr_size;
1401 	int32_t i;
1402 
1403 	if (unlikely(hw->started == 0))
1404 		return nb_rx;
1405 
1406 	nb_used = virtqueue_nused(vq);
1407 
1408 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1409 
1410 	num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1411 	if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1412 		num = VIRTIO_MBUF_BURST_SZ;
1413 	if (likely(num > DESC_PER_CACHELINE))
1414 		num = num - ((vq->vq_used_cons_idx + num) %
1415 				DESC_PER_CACHELINE);
1416 
1417 
1418 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1419 
1420 	for (i = 0; i < num; i++) {
1421 		struct virtio_net_hdr_mrg_rxbuf *header;
1422 
1423 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1424 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1425 
1426 		rxm = rcv_pkts[i];
1427 
1428 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1429 			PMD_RX_LOG(ERR, "Packet drop");
1430 			nb_enqueued++;
1431 			virtio_discard_rxbuf(vq, rxm);
1432 			rxvq->stats.errors++;
1433 			continue;
1434 		}
1435 
1436 		header = (struct virtio_net_hdr_mrg_rxbuf *)
1437 			 ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1438 			 - hdr_size);
1439 		seg_num = header->num_buffers;
1440 		if (seg_num == 0)
1441 			seg_num = 1;
1442 
1443 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1444 		rxm->nb_segs = seg_num;
1445 		rxm->ol_flags = 0;
1446 		rxm->vlan_tci = 0;
1447 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1448 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1449 
1450 		rxm->port = rxvq->port_id;
1451 
1452 		rx_pkts[nb_rx] = rxm;
1453 		prev = rxm;
1454 
1455 		if (hw->has_rx_offload &&
1456 				virtio_rx_offload(rxm, &header->hdr) < 0) {
1457 			virtio_discard_rxbuf(vq, rxm);
1458 			rxvq->stats.errors++;
1459 			continue;
1460 		}
1461 
1462 		if (hw->vlan_strip)
1463 			rte_vlan_strip(rx_pkts[nb_rx]);
1464 
1465 		seg_res = seg_num - 1;
1466 
1467 		/* Merge remaining segments */
1468 		while (seg_res != 0 && i < (num - 1)) {
1469 			i++;
1470 
1471 			rxm = rcv_pkts[i];
1472 			rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1473 			rxm->pkt_len = (uint32_t)(len[i]);
1474 			rxm->data_len = (uint16_t)(len[i]);
1475 
1476 			rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1477 
1478 			prev->next = rxm;
1479 			prev = rxm;
1480 			seg_res -= 1;
1481 		}
1482 
1483 		if (!seg_res) {
1484 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1485 			nb_rx++;
1486 		}
1487 	}
1488 
1489 	/* Last packet still need merge segments */
1490 	while (seg_res != 0) {
1491 		uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1492 					VIRTIO_MBUF_BURST_SZ);
1493 
1494 		if (likely(virtqueue_nused(vq) >= rcv_cnt)) {
1495 			num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len,
1496 							   rcv_cnt);
1497 			uint16_t extra_idx = 0;
1498 
1499 			rcv_cnt = num;
1500 			while (extra_idx < rcv_cnt) {
1501 				rxm = rcv_pkts[extra_idx];
1502 				rxm->data_off =
1503 					RTE_PKTMBUF_HEADROOM - hdr_size;
1504 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
1505 				rxm->data_len = (uint16_t)(len[extra_idx]);
1506 				prev->next = rxm;
1507 				prev = rxm;
1508 				rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1509 				extra_idx += 1;
1510 			};
1511 			seg_res -= rcv_cnt;
1512 
1513 			if (!seg_res) {
1514 				virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1515 				nb_rx++;
1516 			}
1517 		} else {
1518 			PMD_RX_LOG(ERR,
1519 					"No enough segments for packet.");
1520 			rte_pktmbuf_free(rx_pkts[nb_rx]);
1521 			rxvq->stats.errors++;
1522 			break;
1523 		}
1524 	}
1525 
1526 	rxvq->stats.packets += nb_rx;
1527 
1528 	/* Allocate new mbuf for the used descriptor */
1529 	if (likely(!virtqueue_full(vq))) {
1530 		/* free_cnt may include mrg descs */
1531 		uint16_t free_cnt = vq->vq_free_cnt;
1532 		struct rte_mbuf *new_pkts[free_cnt];
1533 
1534 		if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1535 			error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1536 					free_cnt);
1537 			if (unlikely(error)) {
1538 				for (i = 0; i < free_cnt; i++)
1539 					rte_pktmbuf_free(new_pkts[i]);
1540 			}
1541 			nb_enqueued += free_cnt;
1542 		} else {
1543 			struct rte_eth_dev *dev =
1544 				&rte_eth_devices[rxvq->port_id];
1545 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1546 		}
1547 	}
1548 
1549 	if (likely(nb_enqueued)) {
1550 		vq_update_avail_idx(vq);
1551 
1552 		if (unlikely(virtqueue_kick_prepare(vq))) {
1553 			virtqueue_notify(vq);
1554 			PMD_RX_LOG(DEBUG, "Notified");
1555 		}
1556 	}
1557 
1558 	return nb_rx;
1559 }
1560 
1561 uint16_t
1562 virtio_recv_mergeable_pkts_packed(void *rx_queue,
1563 			struct rte_mbuf **rx_pkts,
1564 			uint16_t nb_pkts)
1565 {
1566 	struct virtnet_rx *rxvq = rx_queue;
1567 	struct virtqueue *vq = virtnet_rxq_to_vq(rxvq);
1568 	struct virtio_hw *hw = vq->hw;
1569 	struct rte_mbuf *rxm;
1570 	struct rte_mbuf *prev = NULL;
1571 	uint16_t num, nb_rx = 0;
1572 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1573 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1574 	uint32_t nb_enqueued = 0;
1575 	uint32_t seg_num = 0;
1576 	uint32_t seg_res = 0;
1577 	uint32_t hdr_size = hw->vtnet_hdr_size;
1578 	int32_t i;
1579 	int error;
1580 
1581 	if (unlikely(hw->started == 0))
1582 		return nb_rx;
1583 
1584 
1585 	num = nb_pkts;
1586 	if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1587 		num = VIRTIO_MBUF_BURST_SZ;
1588 	if (likely(num > DESC_PER_CACHELINE))
1589 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1590 
1591 	num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1592 
1593 	for (i = 0; i < num; i++) {
1594 		struct virtio_net_hdr_mrg_rxbuf *header;
1595 
1596 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1597 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1598 
1599 		rxm = rcv_pkts[i];
1600 
1601 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1602 			PMD_RX_LOG(ERR, "Packet drop");
1603 			nb_enqueued++;
1604 			virtio_discard_rxbuf(vq, rxm);
1605 			rxvq->stats.errors++;
1606 			continue;
1607 		}
1608 
1609 		header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)
1610 			  rxm->buf_addr + RTE_PKTMBUF_HEADROOM - hdr_size);
1611 		seg_num = header->num_buffers;
1612 
1613 		if (seg_num == 0)
1614 			seg_num = 1;
1615 
1616 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1617 		rxm->nb_segs = seg_num;
1618 		rxm->ol_flags = 0;
1619 		rxm->vlan_tci = 0;
1620 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1621 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1622 
1623 		rxm->port = rxvq->port_id;
1624 		rx_pkts[nb_rx] = rxm;
1625 		prev = rxm;
1626 
1627 		if (hw->has_rx_offload &&
1628 				virtio_rx_offload(rxm, &header->hdr) < 0) {
1629 			virtio_discard_rxbuf(vq, rxm);
1630 			rxvq->stats.errors++;
1631 			continue;
1632 		}
1633 
1634 		if (hw->vlan_strip)
1635 			rte_vlan_strip(rx_pkts[nb_rx]);
1636 
1637 		seg_res = seg_num - 1;
1638 
1639 		/* Merge remaining segments */
1640 		while (seg_res != 0 && i < (num - 1)) {
1641 			i++;
1642 
1643 			rxm = rcv_pkts[i];
1644 			rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1645 			rxm->pkt_len = (uint32_t)(len[i]);
1646 			rxm->data_len = (uint16_t)(len[i]);
1647 
1648 			rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1649 
1650 			prev->next = rxm;
1651 			prev = rxm;
1652 			seg_res -= 1;
1653 		}
1654 
1655 		if (!seg_res) {
1656 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1657 			nb_rx++;
1658 		}
1659 	}
1660 
1661 	/* Last packet still need merge segments */
1662 	while (seg_res != 0) {
1663 		uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1664 					VIRTIO_MBUF_BURST_SZ);
1665 		uint16_t extra_idx = 0;
1666 
1667 		rcv_cnt = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts,
1668 				len, rcv_cnt);
1669 		if (unlikely(rcv_cnt == 0)) {
1670 			PMD_RX_LOG(ERR, "No enough segments for packet.");
1671 			rte_pktmbuf_free(rx_pkts[nb_rx]);
1672 			rxvq->stats.errors++;
1673 			break;
1674 		}
1675 
1676 		while (extra_idx < rcv_cnt) {
1677 			rxm = rcv_pkts[extra_idx];
1678 
1679 			rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1680 			rxm->pkt_len = (uint32_t)(len[extra_idx]);
1681 			rxm->data_len = (uint16_t)(len[extra_idx]);
1682 
1683 			prev->next = rxm;
1684 			prev = rxm;
1685 			rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1686 			extra_idx += 1;
1687 		}
1688 		seg_res -= rcv_cnt;
1689 		if (!seg_res) {
1690 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1691 			nb_rx++;
1692 		}
1693 	}
1694 
1695 	rxvq->stats.packets += nb_rx;
1696 
1697 	/* Allocate new mbuf for the used descriptor */
1698 	if (likely(!virtqueue_full(vq))) {
1699 		/* free_cnt may include mrg descs */
1700 		uint16_t free_cnt = vq->vq_free_cnt;
1701 		struct rte_mbuf *new_pkts[free_cnt];
1702 
1703 		if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1704 			error = virtqueue_enqueue_recv_refill_packed(vq,
1705 					new_pkts, free_cnt);
1706 			if (unlikely(error)) {
1707 				for (i = 0; i < free_cnt; i++)
1708 					rte_pktmbuf_free(new_pkts[i]);
1709 			}
1710 			nb_enqueued += free_cnt;
1711 		} else {
1712 			struct rte_eth_dev *dev =
1713 				&rte_eth_devices[rxvq->port_id];
1714 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1715 		}
1716 	}
1717 
1718 	if (likely(nb_enqueued)) {
1719 		if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1720 			virtqueue_notify(vq);
1721 			PMD_RX_LOG(DEBUG, "Notified");
1722 		}
1723 	}
1724 
1725 	return nb_rx;
1726 }
1727 
1728 uint16_t
1729 virtio_xmit_pkts_prepare(void *tx_queue __rte_unused, struct rte_mbuf **tx_pkts,
1730 			uint16_t nb_pkts)
1731 {
1732 	uint16_t nb_tx;
1733 	int error;
1734 
1735 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1736 		struct rte_mbuf *m = tx_pkts[nb_tx];
1737 
1738 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
1739 		error = rte_validate_tx_offload(m);
1740 		if (unlikely(error)) {
1741 			rte_errno = -error;
1742 			break;
1743 		}
1744 #endif
1745 
1746 		/* Do VLAN tag insertion */
1747 		if (unlikely(m->ol_flags & PKT_TX_VLAN_PKT)) {
1748 			error = rte_vlan_insert(&m);
1749 			/* rte_vlan_insert() may change pointer
1750 			 * even in the case of failure
1751 			 */
1752 			tx_pkts[nb_tx] = m;
1753 
1754 			if (unlikely(error)) {
1755 				rte_errno = -error;
1756 				break;
1757 			}
1758 		}
1759 
1760 		error = rte_net_intel_cksum_prepare(m);
1761 		if (unlikely(error)) {
1762 			rte_errno = -error;
1763 			break;
1764 		}
1765 
1766 		if (m->ol_flags & PKT_TX_TCP_SEG)
1767 			virtio_tso_fix_cksum(m);
1768 	}
1769 
1770 	return nb_tx;
1771 }
1772 
1773 uint16_t
1774 virtio_xmit_pkts_packed(void *tx_queue, struct rte_mbuf **tx_pkts,
1775 			uint16_t nb_pkts)
1776 {
1777 	struct virtnet_tx *txvq = tx_queue;
1778 	struct virtqueue *vq = virtnet_txq_to_vq(txvq);
1779 	struct virtio_hw *hw = vq->hw;
1780 	uint16_t hdr_size = hw->vtnet_hdr_size;
1781 	uint16_t nb_tx = 0;
1782 	bool in_order = virtio_with_feature(hw, VIRTIO_F_IN_ORDER);
1783 
1784 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1785 		return nb_tx;
1786 
1787 	if (unlikely(nb_pkts < 1))
1788 		return nb_pkts;
1789 
1790 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1791 
1792 	if (nb_pkts > vq->vq_free_cnt)
1793 		virtio_xmit_cleanup_packed(vq, nb_pkts - vq->vq_free_cnt,
1794 					   in_order);
1795 
1796 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1797 		struct rte_mbuf *txm = tx_pkts[nb_tx];
1798 		int can_push = 0, use_indirect = 0, slots, need;
1799 
1800 		/* optimize ring usage */
1801 		if ((virtio_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1802 		      virtio_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1803 		    rte_mbuf_refcnt_read(txm) == 1 &&
1804 		    RTE_MBUF_DIRECT(txm) &&
1805 		    txm->nb_segs == 1 &&
1806 		    rte_pktmbuf_headroom(txm) >= hdr_size &&
1807 		    rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1808 			   __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1809 			can_push = 1;
1810 		else if (virtio_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1811 			 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1812 			use_indirect = 1;
1813 		/* How many main ring entries are needed to this Tx?
1814 		 * indirect   => 1
1815 		 * any_layout => number of segments
1816 		 * default    => number of segments + 1
1817 		 */
1818 		slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1819 		need = slots - vq->vq_free_cnt;
1820 
1821 		/* Positive value indicates it need free vring descriptors */
1822 		if (unlikely(need > 0)) {
1823 			virtio_xmit_cleanup_packed(vq, need, in_order);
1824 			need = slots - vq->vq_free_cnt;
1825 			if (unlikely(need > 0)) {
1826 				PMD_TX_LOG(ERR,
1827 					   "No free tx descriptors to transmit");
1828 				break;
1829 			}
1830 		}
1831 
1832 		/* Enqueue Packet buffers */
1833 		if (can_push)
1834 			virtqueue_enqueue_xmit_packed_fast(txvq, txm, in_order);
1835 		else
1836 			virtqueue_enqueue_xmit_packed(txvq, txm, slots,
1837 						      use_indirect, 0,
1838 						      in_order);
1839 
1840 		virtio_update_packet_stats(&txvq->stats, txm);
1841 	}
1842 
1843 	txvq->stats.packets += nb_tx;
1844 
1845 	if (likely(nb_tx)) {
1846 		if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1847 			virtqueue_notify(vq);
1848 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1849 		}
1850 	}
1851 
1852 	return nb_tx;
1853 }
1854 
1855 uint16_t
1856 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1857 {
1858 	struct virtnet_tx *txvq = tx_queue;
1859 	struct virtqueue *vq = virtnet_txq_to_vq(txvq);
1860 	struct virtio_hw *hw = vq->hw;
1861 	uint16_t hdr_size = hw->vtnet_hdr_size;
1862 	uint16_t nb_used, nb_tx = 0;
1863 
1864 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1865 		return nb_tx;
1866 
1867 	if (unlikely(nb_pkts < 1))
1868 		return nb_pkts;
1869 
1870 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1871 
1872 	nb_used = virtqueue_nused(vq);
1873 
1874 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1875 		virtio_xmit_cleanup(vq, nb_used);
1876 
1877 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1878 		struct rte_mbuf *txm = tx_pkts[nb_tx];
1879 		int can_push = 0, use_indirect = 0, slots, need;
1880 
1881 		/* optimize ring usage */
1882 		if ((virtio_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1883 		      virtio_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1884 		    rte_mbuf_refcnt_read(txm) == 1 &&
1885 		    RTE_MBUF_DIRECT(txm) &&
1886 		    txm->nb_segs == 1 &&
1887 		    rte_pktmbuf_headroom(txm) >= hdr_size &&
1888 		    rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1889 				   __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1890 			can_push = 1;
1891 		else if (virtio_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1892 			 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1893 			use_indirect = 1;
1894 
1895 		/* How many main ring entries are needed to this Tx?
1896 		 * any_layout => number of segments
1897 		 * indirect   => 1
1898 		 * default    => number of segments + 1
1899 		 */
1900 		slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1901 		need = slots - vq->vq_free_cnt;
1902 
1903 		/* Positive value indicates it need free vring descriptors */
1904 		if (unlikely(need > 0)) {
1905 			nb_used = virtqueue_nused(vq);
1906 
1907 			need = RTE_MIN(need, (int)nb_used);
1908 
1909 			virtio_xmit_cleanup(vq, need);
1910 			need = slots - vq->vq_free_cnt;
1911 			if (unlikely(need > 0)) {
1912 				PMD_TX_LOG(ERR,
1913 					   "No free tx descriptors to transmit");
1914 				break;
1915 			}
1916 		}
1917 
1918 		/* Enqueue Packet buffers */
1919 		virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
1920 			can_push, 0);
1921 
1922 		virtio_update_packet_stats(&txvq->stats, txm);
1923 	}
1924 
1925 	txvq->stats.packets += nb_tx;
1926 
1927 	if (likely(nb_tx)) {
1928 		vq_update_avail_idx(vq);
1929 
1930 		if (unlikely(virtqueue_kick_prepare(vq))) {
1931 			virtqueue_notify(vq);
1932 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1933 		}
1934 	}
1935 
1936 	return nb_tx;
1937 }
1938 
1939 static __rte_always_inline int
1940 virtio_xmit_try_cleanup_inorder(struct virtqueue *vq, uint16_t need)
1941 {
1942 	uint16_t nb_used, nb_clean, nb_descs;
1943 
1944 	nb_descs = vq->vq_free_cnt + need;
1945 	nb_used = virtqueue_nused(vq);
1946 	nb_clean = RTE_MIN(need, (int)nb_used);
1947 
1948 	virtio_xmit_cleanup_inorder(vq, nb_clean);
1949 
1950 	return nb_descs - vq->vq_free_cnt;
1951 }
1952 
1953 uint16_t
1954 virtio_xmit_pkts_inorder(void *tx_queue,
1955 			struct rte_mbuf **tx_pkts,
1956 			uint16_t nb_pkts)
1957 {
1958 	struct virtnet_tx *txvq = tx_queue;
1959 	struct virtqueue *vq = virtnet_txq_to_vq(txvq);
1960 	struct virtio_hw *hw = vq->hw;
1961 	uint16_t hdr_size = hw->vtnet_hdr_size;
1962 	uint16_t nb_used, nb_tx = 0, nb_inorder_pkts = 0;
1963 	struct rte_mbuf *inorder_pkts[nb_pkts];
1964 	int need;
1965 
1966 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1967 		return nb_tx;
1968 
1969 	if (unlikely(nb_pkts < 1))
1970 		return nb_pkts;
1971 
1972 	VIRTQUEUE_DUMP(vq);
1973 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1974 	nb_used = virtqueue_nused(vq);
1975 
1976 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1977 		virtio_xmit_cleanup_inorder(vq, nb_used);
1978 
1979 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1980 		struct rte_mbuf *txm = tx_pkts[nb_tx];
1981 		int slots;
1982 
1983 		/* optimize ring usage */
1984 		if ((virtio_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1985 		     virtio_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1986 		     rte_mbuf_refcnt_read(txm) == 1 &&
1987 		     RTE_MBUF_DIRECT(txm) &&
1988 		     txm->nb_segs == 1 &&
1989 		     rte_pktmbuf_headroom(txm) >= hdr_size &&
1990 		     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1991 				__alignof__(struct virtio_net_hdr_mrg_rxbuf))) {
1992 			inorder_pkts[nb_inorder_pkts] = txm;
1993 			nb_inorder_pkts++;
1994 
1995 			continue;
1996 		}
1997 
1998 		if (nb_inorder_pkts) {
1999 			need = nb_inorder_pkts - vq->vq_free_cnt;
2000 			if (unlikely(need > 0)) {
2001 				need = virtio_xmit_try_cleanup_inorder(vq,
2002 								       need);
2003 				if (unlikely(need > 0)) {
2004 					PMD_TX_LOG(ERR,
2005 						"No free tx descriptors to "
2006 						"transmit");
2007 					break;
2008 				}
2009 			}
2010 			virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2011 							nb_inorder_pkts);
2012 			nb_inorder_pkts = 0;
2013 		}
2014 
2015 		slots = txm->nb_segs + 1;
2016 		need = slots - vq->vq_free_cnt;
2017 		if (unlikely(need > 0)) {
2018 			need = virtio_xmit_try_cleanup_inorder(vq, slots);
2019 
2020 			if (unlikely(need > 0)) {
2021 				PMD_TX_LOG(ERR,
2022 					"No free tx descriptors to transmit");
2023 				break;
2024 			}
2025 		}
2026 		/* Enqueue Packet buffers */
2027 		virtqueue_enqueue_xmit(txvq, txm, slots, 0, 0, 1);
2028 
2029 		virtio_update_packet_stats(&txvq->stats, txm);
2030 	}
2031 
2032 	/* Transmit all inorder packets */
2033 	if (nb_inorder_pkts) {
2034 		need = nb_inorder_pkts - vq->vq_free_cnt;
2035 		if (unlikely(need > 0)) {
2036 			need = virtio_xmit_try_cleanup_inorder(vq,
2037 								  need);
2038 			if (unlikely(need > 0)) {
2039 				PMD_TX_LOG(ERR,
2040 					"No free tx descriptors to transmit");
2041 				nb_inorder_pkts = vq->vq_free_cnt;
2042 				nb_tx -= need;
2043 			}
2044 		}
2045 
2046 		virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2047 						nb_inorder_pkts);
2048 	}
2049 
2050 	txvq->stats.packets += nb_tx;
2051 
2052 	if (likely(nb_tx)) {
2053 		vq_update_avail_idx(vq);
2054 
2055 		if (unlikely(virtqueue_kick_prepare(vq))) {
2056 			virtqueue_notify(vq);
2057 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2058 		}
2059 	}
2060 
2061 	VIRTQUEUE_DUMP(vq);
2062 
2063 	return nb_tx;
2064 }
2065 
2066 __rte_weak uint16_t
2067 virtio_recv_pkts_packed_vec(void *rx_queue __rte_unused,
2068 			    struct rte_mbuf **rx_pkts __rte_unused,
2069 			    uint16_t nb_pkts __rte_unused)
2070 {
2071 	return 0;
2072 }
2073 
2074 __rte_weak uint16_t
2075 virtio_xmit_pkts_packed_vec(void *tx_queue __rte_unused,
2076 			    struct rte_mbuf **tx_pkts __rte_unused,
2077 			    uint16_t nb_pkts __rte_unused)
2078 {
2079 	return 0;
2080 }
2081