xref: /dpdk/drivers/net/virtio/virtio_rxtx.c (revision 77d66da8383448dc7d99eb19f4b1e33aa562743e)
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 <rte_ethdev_driver.h>
19 #include <rte_prefetch.h>
20 #include <rte_string_fns.h>
21 #include <rte_errno.h>
22 #include <rte_byteorder.h>
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_pci.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 = rxvq->vq;
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_cio_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 =
275 				VIRTIO_MBUF_ADDR(cookies[i], vq) +
276 				RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
277 		start_dp[idx].len =
278 				cookies[i]->buf_len -
279 				RTE_PKTMBUF_HEADROOM +
280 				hw->vtnet_hdr_size;
281 		start_dp[idx].flags =  VRING_DESC_F_WRITE;
282 
283 		vq_update_avail_ring(vq, idx);
284 		head_idx++;
285 		i++;
286 	}
287 
288 	vq->vq_desc_head_idx += num;
289 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
290 	return 0;
291 }
292 
293 static inline int
294 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf **cookie,
295 				uint16_t num)
296 {
297 	struct vq_desc_extra *dxp;
298 	struct virtio_hw *hw = vq->hw;
299 	struct vring_desc *start_dp = vq->vq_split.ring.desc;
300 	uint16_t idx, i;
301 
302 	if (unlikely(vq->vq_free_cnt == 0))
303 		return -ENOSPC;
304 	if (unlikely(vq->vq_free_cnt < num))
305 		return -EMSGSIZE;
306 
307 	if (unlikely(vq->vq_desc_head_idx >= vq->vq_nentries))
308 		return -EFAULT;
309 
310 	for (i = 0; i < num; i++) {
311 		idx = vq->vq_desc_head_idx;
312 		dxp = &vq->vq_descx[idx];
313 		dxp->cookie = (void *)cookie[i];
314 		dxp->ndescs = 1;
315 
316 		start_dp[idx].addr =
317 			VIRTIO_MBUF_ADDR(cookie[i], vq) +
318 			RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
319 		start_dp[idx].len =
320 			cookie[i]->buf_len - RTE_PKTMBUF_HEADROOM +
321 			hw->vtnet_hdr_size;
322 		start_dp[idx].flags = VRING_DESC_F_WRITE;
323 		vq->vq_desc_head_idx = start_dp[idx].next;
324 		vq_update_avail_ring(vq, idx);
325 		if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END) {
326 			vq->vq_desc_tail_idx = vq->vq_desc_head_idx;
327 			break;
328 		}
329 	}
330 
331 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
332 
333 	return 0;
334 }
335 
336 static inline int
337 virtqueue_enqueue_recv_refill_packed(struct virtqueue *vq,
338 				     struct rte_mbuf **cookie, uint16_t num)
339 {
340 	struct vring_packed_desc *start_dp = vq->vq_packed.ring.desc;
341 	uint16_t flags = vq->vq_packed.cached_flags;
342 	struct virtio_hw *hw = vq->hw;
343 	struct vq_desc_extra *dxp;
344 	uint16_t idx;
345 	int i;
346 
347 	if (unlikely(vq->vq_free_cnt == 0))
348 		return -ENOSPC;
349 	if (unlikely(vq->vq_free_cnt < num))
350 		return -EMSGSIZE;
351 
352 	for (i = 0; i < num; i++) {
353 		idx = vq->vq_avail_idx;
354 		dxp = &vq->vq_descx[idx];
355 		dxp->cookie = (void *)cookie[i];
356 		dxp->ndescs = 1;
357 
358 		start_dp[idx].addr = VIRTIO_MBUF_ADDR(cookie[i], vq) +
359 				RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
360 		start_dp[idx].len = cookie[i]->buf_len - RTE_PKTMBUF_HEADROOM
361 					+ hw->vtnet_hdr_size;
362 
363 		vq->vq_desc_head_idx = dxp->next;
364 		if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
365 			vq->vq_desc_tail_idx = vq->vq_desc_head_idx;
366 
367 		virtqueue_store_flags_packed(&start_dp[idx], flags,
368 					     hw->weak_barriers);
369 
370 		if (++vq->vq_avail_idx >= vq->vq_nentries) {
371 			vq->vq_avail_idx -= vq->vq_nentries;
372 			vq->vq_packed.cached_flags ^=
373 				VRING_PACKED_DESC_F_AVAIL_USED;
374 			flags = vq->vq_packed.cached_flags;
375 		}
376 	}
377 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
378 	return 0;
379 }
380 
381 /* When doing TSO, the IP length is not included in the pseudo header
382  * checksum of the packet given to the PMD, but for virtio it is
383  * expected.
384  */
385 static void
386 virtio_tso_fix_cksum(struct rte_mbuf *m)
387 {
388 	/* common case: header is not fragmented */
389 	if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
390 			m->l4_len)) {
391 		struct rte_ipv4_hdr *iph;
392 		struct rte_ipv6_hdr *ip6h;
393 		struct rte_tcp_hdr *th;
394 		uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
395 		uint32_t tmp;
396 
397 		iph = rte_pktmbuf_mtod_offset(m,
398 					struct rte_ipv4_hdr *, m->l2_len);
399 		th = RTE_PTR_ADD(iph, m->l3_len);
400 		if ((iph->version_ihl >> 4) == 4) {
401 			iph->hdr_checksum = 0;
402 			iph->hdr_checksum = rte_ipv4_cksum(iph);
403 			ip_len = iph->total_length;
404 			ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
405 				m->l3_len);
406 		} else {
407 			ip6h = (struct rte_ipv6_hdr *)iph;
408 			ip_paylen = ip6h->payload_len;
409 		}
410 
411 		/* calculate the new phdr checksum not including ip_paylen */
412 		prev_cksum = th->cksum;
413 		tmp = prev_cksum;
414 		tmp += ip_paylen;
415 		tmp = (tmp & 0xffff) + (tmp >> 16);
416 		new_cksum = tmp;
417 
418 		/* replace it in the packet */
419 		th->cksum = new_cksum;
420 	}
421 }
422 
423 
424 
425 
426 static inline void
427 virtqueue_enqueue_xmit_inorder(struct virtnet_tx *txvq,
428 			struct rte_mbuf **cookies,
429 			uint16_t num)
430 {
431 	struct vq_desc_extra *dxp;
432 	struct virtqueue *vq = txvq->vq;
433 	struct vring_desc *start_dp;
434 	struct virtio_net_hdr *hdr;
435 	uint16_t idx;
436 	int16_t head_size = vq->hw->vtnet_hdr_size;
437 	uint16_t i = 0;
438 
439 	idx = vq->vq_desc_head_idx;
440 	start_dp = vq->vq_split.ring.desc;
441 
442 	while (i < num) {
443 		idx = idx & (vq->vq_nentries - 1);
444 		dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
445 		dxp->cookie = (void *)cookies[i];
446 		dxp->ndescs = 1;
447 		virtio_update_packet_stats(&txvq->stats, cookies[i]);
448 
449 		hdr = rte_pktmbuf_mtod_offset(cookies[i],
450 				struct virtio_net_hdr *, -head_size);
451 
452 		/* if offload disabled, hdr is not zeroed yet, do it now */
453 		if (!vq->hw->has_tx_offload)
454 			virtqueue_clear_net_hdr(hdr);
455 		else
456 			virtqueue_xmit_offload(hdr, cookies[i], true);
457 
458 		start_dp[idx].addr  =
459 			VIRTIO_MBUF_DATA_DMA_ADDR(cookies[i], vq) - head_size;
460 		start_dp[idx].len   = cookies[i]->data_len + head_size;
461 		start_dp[idx].flags = 0;
462 
463 
464 		vq_update_avail_ring(vq, idx);
465 
466 		idx++;
467 		i++;
468 	};
469 
470 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
471 	vq->vq_desc_head_idx = idx & (vq->vq_nentries - 1);
472 }
473 
474 static inline void
475 virtqueue_enqueue_xmit_packed_fast(struct virtnet_tx *txvq,
476 				   struct rte_mbuf *cookie,
477 				   int in_order)
478 {
479 	struct virtqueue *vq = txvq->vq;
480 	struct vring_packed_desc *dp;
481 	struct vq_desc_extra *dxp;
482 	uint16_t idx, id, flags;
483 	int16_t head_size = vq->hw->vtnet_hdr_size;
484 	struct virtio_net_hdr *hdr;
485 
486 	id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx;
487 	idx = vq->vq_avail_idx;
488 	dp = &vq->vq_packed.ring.desc[idx];
489 
490 	dxp = &vq->vq_descx[id];
491 	dxp->ndescs = 1;
492 	dxp->cookie = cookie;
493 
494 	flags = vq->vq_packed.cached_flags;
495 
496 	/* prepend cannot fail, checked by caller */
497 	hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
498 				      -head_size);
499 
500 	/* if offload disabled, hdr is not zeroed yet, do it now */
501 	if (!vq->hw->has_tx_offload)
502 		virtqueue_clear_net_hdr(hdr);
503 	else
504 		virtqueue_xmit_offload(hdr, cookie, true);
505 
506 	dp->addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq) - head_size;
507 	dp->len  = cookie->data_len + head_size;
508 	dp->id   = id;
509 
510 	if (++vq->vq_avail_idx >= vq->vq_nentries) {
511 		vq->vq_avail_idx -= vq->vq_nentries;
512 		vq->vq_packed.cached_flags ^= VRING_PACKED_DESC_F_AVAIL_USED;
513 	}
514 
515 	vq->vq_free_cnt--;
516 
517 	if (!in_order) {
518 		vq->vq_desc_head_idx = dxp->next;
519 		if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
520 			vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
521 	}
522 
523 	virtqueue_store_flags_packed(dp, flags, vq->hw->weak_barriers);
524 }
525 
526 static inline void
527 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
528 			uint16_t needed, int use_indirect, int can_push,
529 			int in_order)
530 {
531 	struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
532 	struct vq_desc_extra *dxp;
533 	struct virtqueue *vq = txvq->vq;
534 	struct vring_desc *start_dp;
535 	uint16_t seg_num = cookie->nb_segs;
536 	uint16_t head_idx, idx;
537 	int16_t head_size = vq->hw->vtnet_hdr_size;
538 	bool prepend_header = false;
539 	struct virtio_net_hdr *hdr;
540 
541 	head_idx = vq->vq_desc_head_idx;
542 	idx = head_idx;
543 	if (in_order)
544 		dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
545 	else
546 		dxp = &vq->vq_descx[idx];
547 	dxp->cookie = (void *)cookie;
548 	dxp->ndescs = needed;
549 
550 	start_dp = vq->vq_split.ring.desc;
551 
552 	if (can_push) {
553 		/* prepend cannot fail, checked by caller */
554 		hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
555 					      -head_size);
556 		prepend_header = true;
557 
558 		/* if offload disabled, it is not zeroed below, do it now */
559 		if (!vq->hw->has_tx_offload)
560 			virtqueue_clear_net_hdr(hdr);
561 	} else if (use_indirect) {
562 		/* setup tx ring slot to point to indirect
563 		 * descriptor list stored in reserved region.
564 		 *
565 		 * the first slot in indirect ring is already preset
566 		 * to point to the header in reserved region
567 		 */
568 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
569 			RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
570 		start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
571 		start_dp[idx].flags = VRING_DESC_F_INDIRECT;
572 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
573 
574 		/* loop below will fill in rest of the indirect elements */
575 		start_dp = txr[idx].tx_indir;
576 		idx = 1;
577 	} else {
578 		/* setup first tx ring slot to point to header
579 		 * stored in reserved region.
580 		 */
581 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
582 			RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
583 		start_dp[idx].len   = vq->hw->vtnet_hdr_size;
584 		start_dp[idx].flags = VRING_DESC_F_NEXT;
585 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
586 
587 		idx = start_dp[idx].next;
588 	}
589 
590 	virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload);
591 
592 	do {
593 		start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
594 		start_dp[idx].len   = cookie->data_len;
595 		if (prepend_header) {
596 			start_dp[idx].addr -= head_size;
597 			start_dp[idx].len += head_size;
598 			prepend_header = false;
599 		}
600 		start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
601 		idx = start_dp[idx].next;
602 	} while ((cookie = cookie->next) != NULL);
603 
604 	if (use_indirect)
605 		idx = vq->vq_split.ring.desc[head_idx].next;
606 
607 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
608 
609 	vq->vq_desc_head_idx = idx;
610 	vq_update_avail_ring(vq, head_idx);
611 
612 	if (!in_order) {
613 		if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
614 			vq->vq_desc_tail_idx = idx;
615 	}
616 }
617 
618 void
619 virtio_dev_cq_start(struct rte_eth_dev *dev)
620 {
621 	struct virtio_hw *hw = dev->data->dev_private;
622 
623 	if (hw->cvq && hw->cvq->vq) {
624 		rte_spinlock_init(&hw->cvq->lock);
625 		VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
626 	}
627 }
628 
629 int
630 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
631 			uint16_t queue_idx,
632 			uint16_t nb_desc,
633 			unsigned int socket_id __rte_unused,
634 			const struct rte_eth_rxconf *rx_conf,
635 			struct rte_mempool *mp)
636 {
637 	uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
638 	struct virtio_hw *hw = dev->data->dev_private;
639 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
640 	struct virtnet_rx *rxvq;
641 	uint16_t rx_free_thresh;
642 
643 	PMD_INIT_FUNC_TRACE();
644 
645 	if (rx_conf->rx_deferred_start) {
646 		PMD_INIT_LOG(ERR, "Rx deferred start is not supported");
647 		return -EINVAL;
648 	}
649 
650 	rx_free_thresh = rx_conf->rx_free_thresh;
651 	if (rx_free_thresh == 0)
652 		rx_free_thresh =
653 			RTE_MIN(vq->vq_nentries / 4, DEFAULT_RX_FREE_THRESH);
654 
655 	if (rx_free_thresh & 0x3) {
656 		RTE_LOG(ERR, PMD, "rx_free_thresh must be multiples of four."
657 			" (rx_free_thresh=%u port=%u queue=%u)\n",
658 			rx_free_thresh, dev->data->port_id, queue_idx);
659 		return -EINVAL;
660 	}
661 
662 	if (rx_free_thresh >= vq->vq_nentries) {
663 		RTE_LOG(ERR, PMD, "rx_free_thresh must be less than the "
664 			"number of RX entries (%u)."
665 			" (rx_free_thresh=%u port=%u queue=%u)\n",
666 			vq->vq_nentries,
667 			rx_free_thresh, dev->data->port_id, queue_idx);
668 		return -EINVAL;
669 	}
670 	vq->vq_free_thresh = rx_free_thresh;
671 
672 	if (nb_desc == 0 || nb_desc > vq->vq_nentries)
673 		nb_desc = vq->vq_nentries;
674 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
675 
676 	rxvq = &vq->rxq;
677 	rxvq->queue_id = queue_idx;
678 	rxvq->mpool = mp;
679 	dev->data->rx_queues[queue_idx] = rxvq;
680 
681 	return 0;
682 }
683 
684 int
685 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
686 {
687 	uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
688 	struct virtio_hw *hw = dev->data->dev_private;
689 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
690 	struct virtnet_rx *rxvq = &vq->rxq;
691 	struct rte_mbuf *m;
692 	uint16_t desc_idx;
693 	int error, nbufs, i;
694 	bool in_order = vtpci_with_feature(hw, VIRTIO_F_IN_ORDER);
695 
696 	PMD_INIT_FUNC_TRACE();
697 
698 	/* Allocate blank mbufs for the each rx descriptor */
699 	nbufs = 0;
700 
701 	if (hw->use_vec_rx && !vtpci_packed_queue(hw)) {
702 		for (desc_idx = 0; desc_idx < vq->vq_nentries;
703 		     desc_idx++) {
704 			vq->vq_split.ring.avail->ring[desc_idx] = desc_idx;
705 			vq->vq_split.ring.desc[desc_idx].flags =
706 				VRING_DESC_F_WRITE;
707 		}
708 
709 		virtio_rxq_vec_setup(rxvq);
710 	}
711 
712 	memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
713 	for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
714 	     desc_idx++) {
715 		vq->sw_ring[vq->vq_nentries + desc_idx] =
716 			&rxvq->fake_mbuf;
717 	}
718 
719 	if (hw->use_vec_rx && !vtpci_packed_queue(hw)) {
720 		while (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
721 			virtio_rxq_rearm_vec(rxvq);
722 			nbufs += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
723 		}
724 	} else if (!vtpci_packed_queue(vq->hw) && in_order) {
725 		if ((!virtqueue_full(vq))) {
726 			uint16_t free_cnt = vq->vq_free_cnt;
727 			struct rte_mbuf *pkts[free_cnt];
728 
729 			if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, pkts,
730 				free_cnt)) {
731 				error = virtqueue_enqueue_refill_inorder(vq,
732 						pkts,
733 						free_cnt);
734 				if (unlikely(error)) {
735 					for (i = 0; i < free_cnt; i++)
736 						rte_pktmbuf_free(pkts[i]);
737 				}
738 			}
739 
740 			nbufs += free_cnt;
741 			vq_update_avail_idx(vq);
742 		}
743 	} else {
744 		while (!virtqueue_full(vq)) {
745 			m = rte_mbuf_raw_alloc(rxvq->mpool);
746 			if (m == NULL)
747 				break;
748 
749 			/* Enqueue allocated buffers */
750 			if (vtpci_packed_queue(vq->hw))
751 				error = virtqueue_enqueue_recv_refill_packed(vq,
752 						&m, 1);
753 			else
754 				error = virtqueue_enqueue_recv_refill(vq,
755 						&m, 1);
756 			if (error) {
757 				rte_pktmbuf_free(m);
758 				break;
759 			}
760 			nbufs++;
761 		}
762 
763 		if (!vtpci_packed_queue(vq->hw))
764 			vq_update_avail_idx(vq);
765 	}
766 
767 	PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
768 
769 	VIRTQUEUE_DUMP(vq);
770 
771 	return 0;
772 }
773 
774 /*
775  * struct rte_eth_dev *dev: Used to update dev
776  * uint16_t nb_desc: Defaults to values read from config space
777  * unsigned int socket_id: Used to allocate memzone
778  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
779  * uint16_t queue_idx: Just used as an index in dev txq list
780  */
781 int
782 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
783 			uint16_t queue_idx,
784 			uint16_t nb_desc,
785 			unsigned int socket_id __rte_unused,
786 			const struct rte_eth_txconf *tx_conf)
787 {
788 	uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
789 	struct virtio_hw *hw = dev->data->dev_private;
790 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
791 	struct virtnet_tx *txvq;
792 	uint16_t tx_free_thresh;
793 
794 	PMD_INIT_FUNC_TRACE();
795 
796 	if (tx_conf->tx_deferred_start) {
797 		PMD_INIT_LOG(ERR, "Tx deferred start is not supported");
798 		return -EINVAL;
799 	}
800 
801 	if (nb_desc == 0 || nb_desc > vq->vq_nentries)
802 		nb_desc = vq->vq_nentries;
803 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
804 
805 	txvq = &vq->txq;
806 	txvq->queue_id = queue_idx;
807 
808 	tx_free_thresh = tx_conf->tx_free_thresh;
809 	if (tx_free_thresh == 0)
810 		tx_free_thresh =
811 			RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
812 
813 	if (tx_free_thresh >= (vq->vq_nentries - 3)) {
814 		PMD_DRV_LOG(ERR, "tx_free_thresh must be less than the "
815 			"number of TX entries minus 3 (%u)."
816 			" (tx_free_thresh=%u port=%u queue=%u)\n",
817 			vq->vq_nentries - 3,
818 			tx_free_thresh, dev->data->port_id, queue_idx);
819 		return -EINVAL;
820 	}
821 
822 	vq->vq_free_thresh = tx_free_thresh;
823 
824 	dev->data->tx_queues[queue_idx] = txvq;
825 	return 0;
826 }
827 
828 int
829 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
830 				uint16_t queue_idx)
831 {
832 	uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
833 	struct virtio_hw *hw = dev->data->dev_private;
834 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
835 
836 	PMD_INIT_FUNC_TRACE();
837 
838 	if (!vtpci_packed_queue(hw)) {
839 		if (vtpci_with_feature(hw, VIRTIO_F_IN_ORDER))
840 			vq->vq_split.ring.desc[vq->vq_nentries - 1].next = 0;
841 	}
842 
843 	VIRTQUEUE_DUMP(vq);
844 
845 	return 0;
846 }
847 
848 static inline void
849 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
850 {
851 	int error;
852 	/*
853 	 * Requeue the discarded mbuf. This should always be
854 	 * successful since it was just dequeued.
855 	 */
856 	if (vtpci_packed_queue(vq->hw))
857 		error = virtqueue_enqueue_recv_refill_packed(vq, &m, 1);
858 	else
859 		error = virtqueue_enqueue_recv_refill(vq, &m, 1);
860 
861 	if (unlikely(error)) {
862 		PMD_DRV_LOG(ERR, "cannot requeue discarded mbuf");
863 		rte_pktmbuf_free(m);
864 	}
865 }
866 
867 static inline void
868 virtio_discard_rxbuf_inorder(struct virtqueue *vq, struct rte_mbuf *m)
869 {
870 	int error;
871 
872 	error = virtqueue_enqueue_refill_inorder(vq, &m, 1);
873 	if (unlikely(error)) {
874 		PMD_DRV_LOG(ERR, "cannot requeue discarded mbuf");
875 		rte_pktmbuf_free(m);
876 	}
877 }
878 
879 /* Optionally fill offload information in structure */
880 static inline int
881 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
882 {
883 	struct rte_net_hdr_lens hdr_lens;
884 	uint32_t hdrlen, ptype;
885 	int l4_supported = 0;
886 
887 	/* nothing to do */
888 	if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
889 		return 0;
890 
891 	m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
892 
893 	ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
894 	m->packet_type = ptype;
895 	if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
896 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
897 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
898 		l4_supported = 1;
899 
900 	if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
901 		hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
902 		if (hdr->csum_start <= hdrlen && l4_supported) {
903 			m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
904 		} else {
905 			/* Unknown proto or tunnel, do sw cksum. We can assume
906 			 * the cksum field is in the first segment since the
907 			 * buffers we provided to the host are large enough.
908 			 * In case of SCTP, this will be wrong since it's a CRC
909 			 * but there's nothing we can do.
910 			 */
911 			uint16_t csum = 0, off;
912 
913 			rte_raw_cksum_mbuf(m, hdr->csum_start,
914 				rte_pktmbuf_pkt_len(m) - hdr->csum_start,
915 				&csum);
916 			if (likely(csum != 0xffff))
917 				csum = ~csum;
918 			off = hdr->csum_offset + hdr->csum_start;
919 			if (rte_pktmbuf_data_len(m) >= off + 1)
920 				*rte_pktmbuf_mtod_offset(m, uint16_t *,
921 					off) = csum;
922 		}
923 	} else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
924 		m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
925 	}
926 
927 	/* GSO request, save required information in mbuf */
928 	if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
929 		/* Check unsupported modes */
930 		if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
931 		    (hdr->gso_size == 0)) {
932 			return -EINVAL;
933 		}
934 
935 		/* Update mss lengthes in mbuf */
936 		m->tso_segsz = hdr->gso_size;
937 		switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
938 			case VIRTIO_NET_HDR_GSO_TCPV4:
939 			case VIRTIO_NET_HDR_GSO_TCPV6:
940 				m->ol_flags |= PKT_RX_LRO | \
941 					PKT_RX_L4_CKSUM_NONE;
942 				break;
943 			default:
944 				return -EINVAL;
945 		}
946 	}
947 
948 	return 0;
949 }
950 
951 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
952 uint16_t
953 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
954 {
955 	struct virtnet_rx *rxvq = rx_queue;
956 	struct virtqueue *vq = rxvq->vq;
957 	struct virtio_hw *hw = vq->hw;
958 	struct rte_mbuf *rxm;
959 	uint16_t nb_used, num, nb_rx;
960 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
961 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
962 	int error;
963 	uint32_t i, nb_enqueued;
964 	uint32_t hdr_size;
965 	struct virtio_net_hdr *hdr;
966 
967 	nb_rx = 0;
968 	if (unlikely(hw->started == 0))
969 		return nb_rx;
970 
971 	nb_used = VIRTQUEUE_NUSED(vq);
972 
973 	virtio_rmb(hw->weak_barriers);
974 
975 	num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
976 	if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
977 		num = VIRTIO_MBUF_BURST_SZ;
978 	if (likely(num > DESC_PER_CACHELINE))
979 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
980 
981 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
982 	PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
983 
984 	nb_enqueued = 0;
985 	hdr_size = hw->vtnet_hdr_size;
986 
987 	for (i = 0; i < num ; i++) {
988 		rxm = rcv_pkts[i];
989 
990 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
991 
992 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
993 			PMD_RX_LOG(ERR, "Packet drop");
994 			nb_enqueued++;
995 			virtio_discard_rxbuf(vq, rxm);
996 			rxvq->stats.errors++;
997 			continue;
998 		}
999 
1000 		rxm->port = rxvq->port_id;
1001 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1002 		rxm->ol_flags = 0;
1003 		rxm->vlan_tci = 0;
1004 
1005 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1006 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1007 
1008 		hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1009 			RTE_PKTMBUF_HEADROOM - hdr_size);
1010 
1011 		if (hw->vlan_strip)
1012 			rte_vlan_strip(rxm);
1013 
1014 		if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1015 			virtio_discard_rxbuf(vq, rxm);
1016 			rxvq->stats.errors++;
1017 			continue;
1018 		}
1019 
1020 		virtio_rx_stats_updated(rxvq, rxm);
1021 
1022 		rx_pkts[nb_rx++] = rxm;
1023 	}
1024 
1025 	rxvq->stats.packets += nb_rx;
1026 
1027 	/* Allocate new mbuf for the used descriptor */
1028 	if (likely(!virtqueue_full(vq))) {
1029 		uint16_t free_cnt = vq->vq_free_cnt;
1030 		struct rte_mbuf *new_pkts[free_cnt];
1031 
1032 		if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1033 						free_cnt) == 0)) {
1034 			error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1035 					free_cnt);
1036 			if (unlikely(error)) {
1037 				for (i = 0; i < free_cnt; i++)
1038 					rte_pktmbuf_free(new_pkts[i]);
1039 			}
1040 			nb_enqueued += free_cnt;
1041 		} else {
1042 			struct rte_eth_dev *dev =
1043 				&rte_eth_devices[rxvq->port_id];
1044 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1045 		}
1046 	}
1047 
1048 	if (likely(nb_enqueued)) {
1049 		vq_update_avail_idx(vq);
1050 
1051 		if (unlikely(virtqueue_kick_prepare(vq))) {
1052 			virtqueue_notify(vq);
1053 			PMD_RX_LOG(DEBUG, "Notified");
1054 		}
1055 	}
1056 
1057 	return nb_rx;
1058 }
1059 
1060 uint16_t
1061 virtio_recv_pkts_packed(void *rx_queue, struct rte_mbuf **rx_pkts,
1062 			uint16_t nb_pkts)
1063 {
1064 	struct virtnet_rx *rxvq = rx_queue;
1065 	struct virtqueue *vq = rxvq->vq;
1066 	struct virtio_hw *hw = vq->hw;
1067 	struct rte_mbuf *rxm;
1068 	uint16_t num, nb_rx;
1069 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1070 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1071 	int error;
1072 	uint32_t i, nb_enqueued;
1073 	uint32_t hdr_size;
1074 	struct virtio_net_hdr *hdr;
1075 
1076 	nb_rx = 0;
1077 	if (unlikely(hw->started == 0))
1078 		return nb_rx;
1079 
1080 	num = RTE_MIN(VIRTIO_MBUF_BURST_SZ, nb_pkts);
1081 	if (likely(num > DESC_PER_CACHELINE))
1082 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1083 
1084 	num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1085 	PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1086 
1087 	nb_enqueued = 0;
1088 	hdr_size = hw->vtnet_hdr_size;
1089 
1090 	for (i = 0; i < num; i++) {
1091 		rxm = rcv_pkts[i];
1092 
1093 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1094 
1095 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1096 			PMD_RX_LOG(ERR, "Packet drop");
1097 			nb_enqueued++;
1098 			virtio_discard_rxbuf(vq, rxm);
1099 			rxvq->stats.errors++;
1100 			continue;
1101 		}
1102 
1103 		rxm->port = rxvq->port_id;
1104 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1105 		rxm->ol_flags = 0;
1106 		rxm->vlan_tci = 0;
1107 
1108 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1109 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1110 
1111 		hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1112 			RTE_PKTMBUF_HEADROOM - hdr_size);
1113 
1114 		if (hw->vlan_strip)
1115 			rte_vlan_strip(rxm);
1116 
1117 		if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1118 			virtio_discard_rxbuf(vq, rxm);
1119 			rxvq->stats.errors++;
1120 			continue;
1121 		}
1122 
1123 		virtio_rx_stats_updated(rxvq, rxm);
1124 
1125 		rx_pkts[nb_rx++] = rxm;
1126 	}
1127 
1128 	rxvq->stats.packets += nb_rx;
1129 
1130 	/* Allocate new mbuf for the used descriptor */
1131 	if (likely(!virtqueue_full(vq))) {
1132 		uint16_t free_cnt = vq->vq_free_cnt;
1133 		struct rte_mbuf *new_pkts[free_cnt];
1134 
1135 		if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1136 						free_cnt) == 0)) {
1137 			error = virtqueue_enqueue_recv_refill_packed(vq,
1138 					new_pkts, free_cnt);
1139 			if (unlikely(error)) {
1140 				for (i = 0; i < free_cnt; i++)
1141 					rte_pktmbuf_free(new_pkts[i]);
1142 			}
1143 			nb_enqueued += free_cnt;
1144 		} else {
1145 			struct rte_eth_dev *dev =
1146 				&rte_eth_devices[rxvq->port_id];
1147 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1148 		}
1149 	}
1150 
1151 	if (likely(nb_enqueued)) {
1152 		if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1153 			virtqueue_notify(vq);
1154 			PMD_RX_LOG(DEBUG, "Notified");
1155 		}
1156 	}
1157 
1158 	return nb_rx;
1159 }
1160 
1161 
1162 uint16_t
1163 virtio_recv_pkts_inorder(void *rx_queue,
1164 			struct rte_mbuf **rx_pkts,
1165 			uint16_t nb_pkts)
1166 {
1167 	struct virtnet_rx *rxvq = rx_queue;
1168 	struct virtqueue *vq = rxvq->vq;
1169 	struct virtio_hw *hw = vq->hw;
1170 	struct rte_mbuf *rxm;
1171 	struct rte_mbuf *prev = NULL;
1172 	uint16_t nb_used, num, nb_rx;
1173 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1174 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1175 	int error;
1176 	uint32_t nb_enqueued;
1177 	uint32_t seg_num;
1178 	uint32_t seg_res;
1179 	uint32_t hdr_size;
1180 	int32_t i;
1181 
1182 	nb_rx = 0;
1183 	if (unlikely(hw->started == 0))
1184 		return nb_rx;
1185 
1186 	nb_used = VIRTQUEUE_NUSED(vq);
1187 	nb_used = RTE_MIN(nb_used, nb_pkts);
1188 	nb_used = RTE_MIN(nb_used, VIRTIO_MBUF_BURST_SZ);
1189 
1190 	virtio_rmb(hw->weak_barriers);
1191 
1192 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1193 
1194 	nb_enqueued = 0;
1195 	seg_num = 1;
1196 	seg_res = 0;
1197 	hdr_size = hw->vtnet_hdr_size;
1198 
1199 	num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len, nb_used);
1200 
1201 	for (i = 0; i < num; i++) {
1202 		struct virtio_net_hdr_mrg_rxbuf *header;
1203 
1204 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1205 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1206 
1207 		rxm = rcv_pkts[i];
1208 
1209 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1210 			PMD_RX_LOG(ERR, "Packet drop");
1211 			nb_enqueued++;
1212 			virtio_discard_rxbuf_inorder(vq, rxm);
1213 			rxvq->stats.errors++;
1214 			continue;
1215 		}
1216 
1217 		header = (struct virtio_net_hdr_mrg_rxbuf *)
1218 			 ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1219 			 - hdr_size);
1220 
1221 		if (vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
1222 			seg_num = header->num_buffers;
1223 			if (seg_num == 0)
1224 				seg_num = 1;
1225 		} else {
1226 			seg_num = 1;
1227 		}
1228 
1229 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1230 		rxm->nb_segs = seg_num;
1231 		rxm->ol_flags = 0;
1232 		rxm->vlan_tci = 0;
1233 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1234 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1235 
1236 		rxm->port = rxvq->port_id;
1237 
1238 		rx_pkts[nb_rx] = rxm;
1239 		prev = rxm;
1240 
1241 		if (vq->hw->has_rx_offload &&
1242 				virtio_rx_offload(rxm, &header->hdr) < 0) {
1243 			virtio_discard_rxbuf_inorder(vq, rxm);
1244 			rxvq->stats.errors++;
1245 			continue;
1246 		}
1247 
1248 		if (hw->vlan_strip)
1249 			rte_vlan_strip(rx_pkts[nb_rx]);
1250 
1251 		seg_res = seg_num - 1;
1252 
1253 		/* Merge remaining segments */
1254 		while (seg_res != 0 && i < (num - 1)) {
1255 			i++;
1256 
1257 			rxm = rcv_pkts[i];
1258 			rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1259 			rxm->pkt_len = (uint32_t)(len[i]);
1260 			rxm->data_len = (uint16_t)(len[i]);
1261 
1262 			rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1263 
1264 			prev->next = rxm;
1265 			prev = rxm;
1266 			seg_res -= 1;
1267 		}
1268 
1269 		if (!seg_res) {
1270 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1271 			nb_rx++;
1272 		}
1273 	}
1274 
1275 	/* Last packet still need merge segments */
1276 	while (seg_res != 0) {
1277 		uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1278 					VIRTIO_MBUF_BURST_SZ);
1279 
1280 		if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
1281 			virtio_rmb(hw->weak_barriers);
1282 			num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len,
1283 							   rcv_cnt);
1284 			uint16_t extra_idx = 0;
1285 
1286 			rcv_cnt = num;
1287 			while (extra_idx < rcv_cnt) {
1288 				rxm = rcv_pkts[extra_idx];
1289 				rxm->data_off =
1290 					RTE_PKTMBUF_HEADROOM - hdr_size;
1291 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
1292 				rxm->data_len = (uint16_t)(len[extra_idx]);
1293 				prev->next = rxm;
1294 				prev = rxm;
1295 				rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1296 				extra_idx += 1;
1297 			};
1298 			seg_res -= rcv_cnt;
1299 
1300 			if (!seg_res) {
1301 				virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1302 				nb_rx++;
1303 			}
1304 		} else {
1305 			PMD_RX_LOG(ERR,
1306 					"No enough segments for packet.");
1307 			rte_pktmbuf_free(rx_pkts[nb_rx]);
1308 			rxvq->stats.errors++;
1309 			break;
1310 		}
1311 	}
1312 
1313 	rxvq->stats.packets += nb_rx;
1314 
1315 	/* Allocate new mbuf for the used descriptor */
1316 
1317 	if (likely(!virtqueue_full(vq))) {
1318 		/* free_cnt may include mrg descs */
1319 		uint16_t free_cnt = vq->vq_free_cnt;
1320 		struct rte_mbuf *new_pkts[free_cnt];
1321 
1322 		if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1323 			error = virtqueue_enqueue_refill_inorder(vq, new_pkts,
1324 					free_cnt);
1325 			if (unlikely(error)) {
1326 				for (i = 0; i < free_cnt; i++)
1327 					rte_pktmbuf_free(new_pkts[i]);
1328 			}
1329 			nb_enqueued += free_cnt;
1330 		} else {
1331 			struct rte_eth_dev *dev =
1332 				&rte_eth_devices[rxvq->port_id];
1333 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1334 		}
1335 	}
1336 
1337 	if (likely(nb_enqueued)) {
1338 		vq_update_avail_idx(vq);
1339 
1340 		if (unlikely(virtqueue_kick_prepare(vq))) {
1341 			virtqueue_notify(vq);
1342 			PMD_RX_LOG(DEBUG, "Notified");
1343 		}
1344 	}
1345 
1346 	return nb_rx;
1347 }
1348 
1349 uint16_t
1350 virtio_recv_mergeable_pkts(void *rx_queue,
1351 			struct rte_mbuf **rx_pkts,
1352 			uint16_t nb_pkts)
1353 {
1354 	struct virtnet_rx *rxvq = rx_queue;
1355 	struct virtqueue *vq = rxvq->vq;
1356 	struct virtio_hw *hw = vq->hw;
1357 	struct rte_mbuf *rxm;
1358 	struct rte_mbuf *prev = NULL;
1359 	uint16_t nb_used, num, nb_rx = 0;
1360 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1361 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1362 	int error;
1363 	uint32_t nb_enqueued = 0;
1364 	uint32_t seg_num = 0;
1365 	uint32_t seg_res = 0;
1366 	uint32_t hdr_size = hw->vtnet_hdr_size;
1367 	int32_t i;
1368 
1369 	if (unlikely(hw->started == 0))
1370 		return nb_rx;
1371 
1372 	nb_used = VIRTQUEUE_NUSED(vq);
1373 
1374 	virtio_rmb(hw->weak_barriers);
1375 
1376 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1377 
1378 	num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1379 	if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1380 		num = VIRTIO_MBUF_BURST_SZ;
1381 	if (likely(num > DESC_PER_CACHELINE))
1382 		num = num - ((vq->vq_used_cons_idx + num) %
1383 				DESC_PER_CACHELINE);
1384 
1385 
1386 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1387 
1388 	for (i = 0; i < num; i++) {
1389 		struct virtio_net_hdr_mrg_rxbuf *header;
1390 
1391 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1392 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1393 
1394 		rxm = rcv_pkts[i];
1395 
1396 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1397 			PMD_RX_LOG(ERR, "Packet drop");
1398 			nb_enqueued++;
1399 			virtio_discard_rxbuf(vq, rxm);
1400 			rxvq->stats.errors++;
1401 			continue;
1402 		}
1403 
1404 		header = (struct virtio_net_hdr_mrg_rxbuf *)
1405 			 ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1406 			 - hdr_size);
1407 		seg_num = header->num_buffers;
1408 		if (seg_num == 0)
1409 			seg_num = 1;
1410 
1411 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1412 		rxm->nb_segs = seg_num;
1413 		rxm->ol_flags = 0;
1414 		rxm->vlan_tci = 0;
1415 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1416 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1417 
1418 		rxm->port = rxvq->port_id;
1419 
1420 		rx_pkts[nb_rx] = rxm;
1421 		prev = rxm;
1422 
1423 		if (hw->has_rx_offload &&
1424 				virtio_rx_offload(rxm, &header->hdr) < 0) {
1425 			virtio_discard_rxbuf(vq, rxm);
1426 			rxvq->stats.errors++;
1427 			continue;
1428 		}
1429 
1430 		if (hw->vlan_strip)
1431 			rte_vlan_strip(rx_pkts[nb_rx]);
1432 
1433 		seg_res = seg_num - 1;
1434 
1435 		/* Merge remaining segments */
1436 		while (seg_res != 0 && i < (num - 1)) {
1437 			i++;
1438 
1439 			rxm = rcv_pkts[i];
1440 			rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1441 			rxm->pkt_len = (uint32_t)(len[i]);
1442 			rxm->data_len = (uint16_t)(len[i]);
1443 
1444 			rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1445 
1446 			prev->next = rxm;
1447 			prev = rxm;
1448 			seg_res -= 1;
1449 		}
1450 
1451 		if (!seg_res) {
1452 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1453 			nb_rx++;
1454 		}
1455 	}
1456 
1457 	/* Last packet still need merge segments */
1458 	while (seg_res != 0) {
1459 		uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1460 					VIRTIO_MBUF_BURST_SZ);
1461 
1462 		if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
1463 			virtio_rmb(hw->weak_barriers);
1464 			num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len,
1465 							   rcv_cnt);
1466 			uint16_t extra_idx = 0;
1467 
1468 			rcv_cnt = num;
1469 			while (extra_idx < rcv_cnt) {
1470 				rxm = rcv_pkts[extra_idx];
1471 				rxm->data_off =
1472 					RTE_PKTMBUF_HEADROOM - hdr_size;
1473 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
1474 				rxm->data_len = (uint16_t)(len[extra_idx]);
1475 				prev->next = rxm;
1476 				prev = rxm;
1477 				rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1478 				extra_idx += 1;
1479 			};
1480 			seg_res -= rcv_cnt;
1481 
1482 			if (!seg_res) {
1483 				virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1484 				nb_rx++;
1485 			}
1486 		} else {
1487 			PMD_RX_LOG(ERR,
1488 					"No enough segments for packet.");
1489 			rte_pktmbuf_free(rx_pkts[nb_rx]);
1490 			rxvq->stats.errors++;
1491 			break;
1492 		}
1493 	}
1494 
1495 	rxvq->stats.packets += nb_rx;
1496 
1497 	/* Allocate new mbuf for the used descriptor */
1498 	if (likely(!virtqueue_full(vq))) {
1499 		/* free_cnt may include mrg descs */
1500 		uint16_t free_cnt = vq->vq_free_cnt;
1501 		struct rte_mbuf *new_pkts[free_cnt];
1502 
1503 		if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1504 			error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1505 					free_cnt);
1506 			if (unlikely(error)) {
1507 				for (i = 0; i < free_cnt; i++)
1508 					rte_pktmbuf_free(new_pkts[i]);
1509 			}
1510 			nb_enqueued += free_cnt;
1511 		} else {
1512 			struct rte_eth_dev *dev =
1513 				&rte_eth_devices[rxvq->port_id];
1514 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1515 		}
1516 	}
1517 
1518 	if (likely(nb_enqueued)) {
1519 		vq_update_avail_idx(vq);
1520 
1521 		if (unlikely(virtqueue_kick_prepare(vq))) {
1522 			virtqueue_notify(vq);
1523 			PMD_RX_LOG(DEBUG, "Notified");
1524 		}
1525 	}
1526 
1527 	return nb_rx;
1528 }
1529 
1530 uint16_t
1531 virtio_recv_mergeable_pkts_packed(void *rx_queue,
1532 			struct rte_mbuf **rx_pkts,
1533 			uint16_t nb_pkts)
1534 {
1535 	struct virtnet_rx *rxvq = rx_queue;
1536 	struct virtqueue *vq = rxvq->vq;
1537 	struct virtio_hw *hw = vq->hw;
1538 	struct rte_mbuf *rxm;
1539 	struct rte_mbuf *prev = NULL;
1540 	uint16_t num, nb_rx = 0;
1541 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1542 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1543 	uint32_t nb_enqueued = 0;
1544 	uint32_t seg_num = 0;
1545 	uint32_t seg_res = 0;
1546 	uint32_t hdr_size = hw->vtnet_hdr_size;
1547 	int32_t i;
1548 	int error;
1549 
1550 	if (unlikely(hw->started == 0))
1551 		return nb_rx;
1552 
1553 
1554 	num = nb_pkts;
1555 	if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1556 		num = VIRTIO_MBUF_BURST_SZ;
1557 	if (likely(num > DESC_PER_CACHELINE))
1558 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1559 
1560 	num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1561 
1562 	for (i = 0; i < num; i++) {
1563 		struct virtio_net_hdr_mrg_rxbuf *header;
1564 
1565 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1566 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1567 
1568 		rxm = rcv_pkts[i];
1569 
1570 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1571 			PMD_RX_LOG(ERR, "Packet drop");
1572 			nb_enqueued++;
1573 			virtio_discard_rxbuf(vq, rxm);
1574 			rxvq->stats.errors++;
1575 			continue;
1576 		}
1577 
1578 		header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)
1579 			  rxm->buf_addr + RTE_PKTMBUF_HEADROOM - hdr_size);
1580 		seg_num = header->num_buffers;
1581 
1582 		if (seg_num == 0)
1583 			seg_num = 1;
1584 
1585 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1586 		rxm->nb_segs = seg_num;
1587 		rxm->ol_flags = 0;
1588 		rxm->vlan_tci = 0;
1589 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1590 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1591 
1592 		rxm->port = rxvq->port_id;
1593 		rx_pkts[nb_rx] = rxm;
1594 		prev = rxm;
1595 
1596 		if (hw->has_rx_offload &&
1597 				virtio_rx_offload(rxm, &header->hdr) < 0) {
1598 			virtio_discard_rxbuf(vq, rxm);
1599 			rxvq->stats.errors++;
1600 			continue;
1601 		}
1602 
1603 		if (hw->vlan_strip)
1604 			rte_vlan_strip(rx_pkts[nb_rx]);
1605 
1606 		seg_res = seg_num - 1;
1607 
1608 		/* Merge remaining segments */
1609 		while (seg_res != 0 && i < (num - 1)) {
1610 			i++;
1611 
1612 			rxm = rcv_pkts[i];
1613 			rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1614 			rxm->pkt_len = (uint32_t)(len[i]);
1615 			rxm->data_len = (uint16_t)(len[i]);
1616 
1617 			rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1618 
1619 			prev->next = rxm;
1620 			prev = rxm;
1621 			seg_res -= 1;
1622 		}
1623 
1624 		if (!seg_res) {
1625 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1626 			nb_rx++;
1627 		}
1628 	}
1629 
1630 	/* Last packet still need merge segments */
1631 	while (seg_res != 0) {
1632 		uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1633 					VIRTIO_MBUF_BURST_SZ);
1634 		uint16_t extra_idx = 0;
1635 
1636 		rcv_cnt = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts,
1637 				len, rcv_cnt);
1638 		if (unlikely(rcv_cnt == 0)) {
1639 			PMD_RX_LOG(ERR, "No enough segments for packet.");
1640 			rte_pktmbuf_free(rx_pkts[nb_rx]);
1641 			rxvq->stats.errors++;
1642 			break;
1643 		}
1644 
1645 		while (extra_idx < rcv_cnt) {
1646 			rxm = rcv_pkts[extra_idx];
1647 
1648 			rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1649 			rxm->pkt_len = (uint32_t)(len[extra_idx]);
1650 			rxm->data_len = (uint16_t)(len[extra_idx]);
1651 
1652 			prev->next = rxm;
1653 			prev = rxm;
1654 			rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1655 			extra_idx += 1;
1656 		}
1657 		seg_res -= rcv_cnt;
1658 		if (!seg_res) {
1659 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1660 			nb_rx++;
1661 		}
1662 	}
1663 
1664 	rxvq->stats.packets += nb_rx;
1665 
1666 	/* Allocate new mbuf for the used descriptor */
1667 	if (likely(!virtqueue_full(vq))) {
1668 		/* free_cnt may include mrg descs */
1669 		uint16_t free_cnt = vq->vq_free_cnt;
1670 		struct rte_mbuf *new_pkts[free_cnt];
1671 
1672 		if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1673 			error = virtqueue_enqueue_recv_refill_packed(vq,
1674 					new_pkts, free_cnt);
1675 			if (unlikely(error)) {
1676 				for (i = 0; i < free_cnt; i++)
1677 					rte_pktmbuf_free(new_pkts[i]);
1678 			}
1679 			nb_enqueued += free_cnt;
1680 		} else {
1681 			struct rte_eth_dev *dev =
1682 				&rte_eth_devices[rxvq->port_id];
1683 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1684 		}
1685 	}
1686 
1687 	if (likely(nb_enqueued)) {
1688 		if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1689 			virtqueue_notify(vq);
1690 			PMD_RX_LOG(DEBUG, "Notified");
1691 		}
1692 	}
1693 
1694 	return nb_rx;
1695 }
1696 
1697 uint16_t
1698 virtio_xmit_pkts_prepare(void *tx_queue __rte_unused, struct rte_mbuf **tx_pkts,
1699 			uint16_t nb_pkts)
1700 {
1701 	uint16_t nb_tx;
1702 	int error;
1703 
1704 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1705 		struct rte_mbuf *m = tx_pkts[nb_tx];
1706 
1707 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
1708 		error = rte_validate_tx_offload(m);
1709 		if (unlikely(error)) {
1710 			rte_errno = -error;
1711 			break;
1712 		}
1713 #endif
1714 
1715 		/* Do VLAN tag insertion */
1716 		if (unlikely(m->ol_flags & PKT_TX_VLAN_PKT)) {
1717 			error = rte_vlan_insert(&m);
1718 			/* rte_vlan_insert() may change pointer
1719 			 * even in the case of failure
1720 			 */
1721 			tx_pkts[nb_tx] = m;
1722 
1723 			if (unlikely(error)) {
1724 				rte_errno = -error;
1725 				break;
1726 			}
1727 		}
1728 
1729 		error = rte_net_intel_cksum_prepare(m);
1730 		if (unlikely(error)) {
1731 			rte_errno = -error;
1732 			break;
1733 		}
1734 
1735 		if (m->ol_flags & PKT_TX_TCP_SEG)
1736 			virtio_tso_fix_cksum(m);
1737 	}
1738 
1739 	return nb_tx;
1740 }
1741 
1742 uint16_t
1743 virtio_xmit_pkts_packed(void *tx_queue, struct rte_mbuf **tx_pkts,
1744 			uint16_t nb_pkts)
1745 {
1746 	struct virtnet_tx *txvq = tx_queue;
1747 	struct virtqueue *vq = txvq->vq;
1748 	struct virtio_hw *hw = vq->hw;
1749 	uint16_t hdr_size = hw->vtnet_hdr_size;
1750 	uint16_t nb_tx = 0;
1751 	bool in_order = vtpci_with_feature(hw, VIRTIO_F_IN_ORDER);
1752 
1753 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1754 		return nb_tx;
1755 
1756 	if (unlikely(nb_pkts < 1))
1757 		return nb_pkts;
1758 
1759 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1760 
1761 	if (nb_pkts > vq->vq_free_cnt)
1762 		virtio_xmit_cleanup_packed(vq, nb_pkts - vq->vq_free_cnt,
1763 					   in_order);
1764 
1765 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1766 		struct rte_mbuf *txm = tx_pkts[nb_tx];
1767 		int can_push = 0, slots, need;
1768 
1769 		/* optimize ring usage */
1770 		if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1771 		      vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1772 		    rte_mbuf_refcnt_read(txm) == 1 &&
1773 		    RTE_MBUF_DIRECT(txm) &&
1774 		    txm->nb_segs == 1 &&
1775 		    rte_pktmbuf_headroom(txm) >= hdr_size &&
1776 		    rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1777 			   __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1778 			can_push = 1;
1779 
1780 		/* How many main ring entries are needed to this Tx?
1781 		 * any_layout => number of segments
1782 		 * default    => number of segments + 1
1783 		 */
1784 		slots = txm->nb_segs + !can_push;
1785 		need = slots - vq->vq_free_cnt;
1786 
1787 		/* Positive value indicates it need free vring descriptors */
1788 		if (unlikely(need > 0)) {
1789 			virtio_xmit_cleanup_packed(vq, need, in_order);
1790 			need = slots - vq->vq_free_cnt;
1791 			if (unlikely(need > 0)) {
1792 				PMD_TX_LOG(ERR,
1793 					   "No free tx descriptors to transmit");
1794 				break;
1795 			}
1796 		}
1797 
1798 		/* Enqueue Packet buffers */
1799 		if (can_push)
1800 			virtqueue_enqueue_xmit_packed_fast(txvq, txm, in_order);
1801 		else
1802 			virtqueue_enqueue_xmit_packed(txvq, txm, slots, 0,
1803 						      in_order);
1804 
1805 		virtio_update_packet_stats(&txvq->stats, txm);
1806 	}
1807 
1808 	txvq->stats.packets += nb_tx;
1809 
1810 	if (likely(nb_tx)) {
1811 		if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1812 			virtqueue_notify(vq);
1813 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1814 		}
1815 	}
1816 
1817 	return nb_tx;
1818 }
1819 
1820 uint16_t
1821 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1822 {
1823 	struct virtnet_tx *txvq = tx_queue;
1824 	struct virtqueue *vq = txvq->vq;
1825 	struct virtio_hw *hw = vq->hw;
1826 	uint16_t hdr_size = hw->vtnet_hdr_size;
1827 	uint16_t nb_used, nb_tx = 0;
1828 
1829 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1830 		return nb_tx;
1831 
1832 	if (unlikely(nb_pkts < 1))
1833 		return nb_pkts;
1834 
1835 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1836 	nb_used = VIRTQUEUE_NUSED(vq);
1837 
1838 	virtio_rmb(hw->weak_barriers);
1839 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1840 		virtio_xmit_cleanup(vq, nb_used);
1841 
1842 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1843 		struct rte_mbuf *txm = tx_pkts[nb_tx];
1844 		int can_push = 0, use_indirect = 0, slots, need;
1845 
1846 		/* optimize ring usage */
1847 		if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1848 		      vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1849 		    rte_mbuf_refcnt_read(txm) == 1 &&
1850 		    RTE_MBUF_DIRECT(txm) &&
1851 		    txm->nb_segs == 1 &&
1852 		    rte_pktmbuf_headroom(txm) >= hdr_size &&
1853 		    rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1854 				   __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1855 			can_push = 1;
1856 		else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1857 			 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1858 			use_indirect = 1;
1859 
1860 		/* How many main ring entries are needed to this Tx?
1861 		 * any_layout => number of segments
1862 		 * indirect   => 1
1863 		 * default    => number of segments + 1
1864 		 */
1865 		slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1866 		need = slots - vq->vq_free_cnt;
1867 
1868 		/* Positive value indicates it need free vring descriptors */
1869 		if (unlikely(need > 0)) {
1870 			nb_used = VIRTQUEUE_NUSED(vq);
1871 			virtio_rmb(hw->weak_barriers);
1872 			need = RTE_MIN(need, (int)nb_used);
1873 
1874 			virtio_xmit_cleanup(vq, need);
1875 			need = slots - vq->vq_free_cnt;
1876 			if (unlikely(need > 0)) {
1877 				PMD_TX_LOG(ERR,
1878 					   "No free tx descriptors to transmit");
1879 				break;
1880 			}
1881 		}
1882 
1883 		/* Enqueue Packet buffers */
1884 		virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
1885 			can_push, 0);
1886 
1887 		virtio_update_packet_stats(&txvq->stats, txm);
1888 	}
1889 
1890 	txvq->stats.packets += nb_tx;
1891 
1892 	if (likely(nb_tx)) {
1893 		vq_update_avail_idx(vq);
1894 
1895 		if (unlikely(virtqueue_kick_prepare(vq))) {
1896 			virtqueue_notify(vq);
1897 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1898 		}
1899 	}
1900 
1901 	return nb_tx;
1902 }
1903 
1904 static __rte_always_inline int
1905 virtio_xmit_try_cleanup_inorder(struct virtqueue *vq, uint16_t need)
1906 {
1907 	uint16_t nb_used, nb_clean, nb_descs;
1908 	struct virtio_hw *hw = vq->hw;
1909 
1910 	nb_descs = vq->vq_free_cnt + need;
1911 	nb_used = VIRTQUEUE_NUSED(vq);
1912 	virtio_rmb(hw->weak_barriers);
1913 	nb_clean = RTE_MIN(need, (int)nb_used);
1914 
1915 	virtio_xmit_cleanup_inorder(vq, nb_clean);
1916 
1917 	return nb_descs - vq->vq_free_cnt;
1918 }
1919 
1920 uint16_t
1921 virtio_xmit_pkts_inorder(void *tx_queue,
1922 			struct rte_mbuf **tx_pkts,
1923 			uint16_t nb_pkts)
1924 {
1925 	struct virtnet_tx *txvq = tx_queue;
1926 	struct virtqueue *vq = txvq->vq;
1927 	struct virtio_hw *hw = vq->hw;
1928 	uint16_t hdr_size = hw->vtnet_hdr_size;
1929 	uint16_t nb_used, nb_tx = 0, nb_inorder_pkts = 0;
1930 	struct rte_mbuf *inorder_pkts[nb_pkts];
1931 	int need;
1932 
1933 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1934 		return nb_tx;
1935 
1936 	if (unlikely(nb_pkts < 1))
1937 		return nb_pkts;
1938 
1939 	VIRTQUEUE_DUMP(vq);
1940 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1941 	nb_used = VIRTQUEUE_NUSED(vq);
1942 
1943 	virtio_rmb(hw->weak_barriers);
1944 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1945 		virtio_xmit_cleanup_inorder(vq, nb_used);
1946 
1947 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1948 		struct rte_mbuf *txm = tx_pkts[nb_tx];
1949 		int slots;
1950 
1951 		/* optimize ring usage */
1952 		if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1953 		     vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1954 		     rte_mbuf_refcnt_read(txm) == 1 &&
1955 		     RTE_MBUF_DIRECT(txm) &&
1956 		     txm->nb_segs == 1 &&
1957 		     rte_pktmbuf_headroom(txm) >= hdr_size &&
1958 		     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1959 				__alignof__(struct virtio_net_hdr_mrg_rxbuf))) {
1960 			inorder_pkts[nb_inorder_pkts] = txm;
1961 			nb_inorder_pkts++;
1962 
1963 			continue;
1964 		}
1965 
1966 		if (nb_inorder_pkts) {
1967 			need = nb_inorder_pkts - vq->vq_free_cnt;
1968 			if (unlikely(need > 0)) {
1969 				need = virtio_xmit_try_cleanup_inorder(vq,
1970 								       need);
1971 				if (unlikely(need > 0)) {
1972 					PMD_TX_LOG(ERR,
1973 						"No free tx descriptors to "
1974 						"transmit");
1975 					break;
1976 				}
1977 			}
1978 			virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
1979 							nb_inorder_pkts);
1980 			nb_inorder_pkts = 0;
1981 		}
1982 
1983 		slots = txm->nb_segs + 1;
1984 		need = slots - vq->vq_free_cnt;
1985 		if (unlikely(need > 0)) {
1986 			need = virtio_xmit_try_cleanup_inorder(vq, slots);
1987 
1988 			if (unlikely(need > 0)) {
1989 				PMD_TX_LOG(ERR,
1990 					"No free tx descriptors to transmit");
1991 				break;
1992 			}
1993 		}
1994 		/* Enqueue Packet buffers */
1995 		virtqueue_enqueue_xmit(txvq, txm, slots, 0, 0, 1);
1996 
1997 		virtio_update_packet_stats(&txvq->stats, txm);
1998 	}
1999 
2000 	/* Transmit all inorder packets */
2001 	if (nb_inorder_pkts) {
2002 		need = nb_inorder_pkts - vq->vq_free_cnt;
2003 		if (unlikely(need > 0)) {
2004 			need = virtio_xmit_try_cleanup_inorder(vq,
2005 								  need);
2006 			if (unlikely(need > 0)) {
2007 				PMD_TX_LOG(ERR,
2008 					"No free tx descriptors to transmit");
2009 				nb_inorder_pkts = vq->vq_free_cnt;
2010 				nb_tx -= need;
2011 			}
2012 		}
2013 
2014 		virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2015 						nb_inorder_pkts);
2016 	}
2017 
2018 	txvq->stats.packets += nb_tx;
2019 
2020 	if (likely(nb_tx)) {
2021 		vq_update_avail_idx(vq);
2022 
2023 		if (unlikely(virtqueue_kick_prepare(vq))) {
2024 			virtqueue_notify(vq);
2025 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2026 		}
2027 	}
2028 
2029 	VIRTQUEUE_DUMP(vq);
2030 
2031 	return nb_tx;
2032 }
2033 
2034 #ifndef CC_AVX512_SUPPORT
2035 uint16_t
2036 virtio_recv_pkts_packed_vec(void *rx_queue __rte_unused,
2037 			    struct rte_mbuf **rx_pkts __rte_unused,
2038 			    uint16_t nb_pkts __rte_unused)
2039 {
2040 	return 0;
2041 }
2042 #endif /* ifndef CC_AVX512_SUPPORT */
2043