xref: /dpdk/drivers/net/virtio/virtio_rxtx.c (revision 8e148e492022f9e0a8bf1c48e59a199c86ccc65b)
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_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_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_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 static void
93 vq_ring_free_id_packed(struct virtqueue *vq, uint16_t id)
94 {
95 	struct vq_desc_extra *dxp;
96 
97 	dxp = &vq->vq_descx[id];
98 	vq->vq_free_cnt += dxp->ndescs;
99 
100 	if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END)
101 		vq->vq_desc_head_idx = id;
102 	else
103 		vq->vq_descx[vq->vq_desc_tail_idx].next = id;
104 
105 	vq->vq_desc_tail_idx = id;
106 	dxp->next = VQ_RING_DESC_CHAIN_END;
107 }
108 
109 static uint16_t
110 virtqueue_dequeue_burst_rx_packed(struct virtqueue *vq,
111 				  struct rte_mbuf **rx_pkts,
112 				  uint32_t *len,
113 				  uint16_t num)
114 {
115 	struct rte_mbuf *cookie;
116 	uint16_t used_idx;
117 	uint16_t id;
118 	struct vring_packed_desc *desc;
119 	uint16_t i;
120 
121 	desc = vq->ring_packed.desc_packed;
122 
123 	for (i = 0; i < num; i++) {
124 		used_idx = vq->vq_used_cons_idx;
125 		if (!desc_is_used(&desc[used_idx], vq))
126 			return i;
127 		virtio_rmb(vq->hw->weak_barriers);
128 		len[i] = desc[used_idx].len;
129 		id = desc[used_idx].id;
130 		cookie = (struct rte_mbuf *)vq->vq_descx[id].cookie;
131 		if (unlikely(cookie == NULL)) {
132 			PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
133 				vq->vq_used_cons_idx);
134 			break;
135 		}
136 		rte_prefetch0(cookie);
137 		rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
138 		rx_pkts[i] = cookie;
139 
140 		vq->vq_free_cnt++;
141 		vq->vq_used_cons_idx++;
142 		if (vq->vq_used_cons_idx >= vq->vq_nentries) {
143 			vq->vq_used_cons_idx -= vq->vq_nentries;
144 			vq->used_wrap_counter ^= 1;
145 		}
146 	}
147 
148 	return i;
149 }
150 
151 static uint16_t
152 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
153 			   uint32_t *len, uint16_t num)
154 {
155 	struct vring_used_elem *uep;
156 	struct rte_mbuf *cookie;
157 	uint16_t used_idx, desc_idx;
158 	uint16_t i;
159 
160 	/*  Caller does the check */
161 	for (i = 0; i < num ; i++) {
162 		used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
163 		uep = &vq->vq_ring.used->ring[used_idx];
164 		desc_idx = (uint16_t) uep->id;
165 		len[i] = uep->len;
166 		cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
167 
168 		if (unlikely(cookie == NULL)) {
169 			PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
170 				vq->vq_used_cons_idx);
171 			break;
172 		}
173 
174 		rte_prefetch0(cookie);
175 		rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
176 		rx_pkts[i]  = cookie;
177 		vq->vq_used_cons_idx++;
178 		vq_ring_free_chain(vq, desc_idx);
179 		vq->vq_descx[desc_idx].cookie = NULL;
180 	}
181 
182 	return i;
183 }
184 
185 static uint16_t
186 virtqueue_dequeue_rx_inorder(struct virtqueue *vq,
187 			struct rte_mbuf **rx_pkts,
188 			uint32_t *len,
189 			uint16_t num)
190 {
191 	struct vring_used_elem *uep;
192 	struct rte_mbuf *cookie;
193 	uint16_t used_idx = 0;
194 	uint16_t i;
195 
196 	if (unlikely(num == 0))
197 		return 0;
198 
199 	for (i = 0; i < num; i++) {
200 		used_idx = vq->vq_used_cons_idx & (vq->vq_nentries - 1);
201 		/* Desc idx same as used idx */
202 		uep = &vq->vq_ring.used->ring[used_idx];
203 		len[i] = uep->len;
204 		cookie = (struct rte_mbuf *)vq->vq_descx[used_idx].cookie;
205 
206 		if (unlikely(cookie == NULL)) {
207 			PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
208 				vq->vq_used_cons_idx);
209 			break;
210 		}
211 
212 		rte_prefetch0(cookie);
213 		rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
214 		rx_pkts[i]  = cookie;
215 		vq->vq_used_cons_idx++;
216 		vq->vq_descx[used_idx].cookie = NULL;
217 	}
218 
219 	vq_ring_free_inorder(vq, used_idx, i);
220 	return i;
221 }
222 
223 #ifndef DEFAULT_TX_FREE_THRESH
224 #define DEFAULT_TX_FREE_THRESH 32
225 #endif
226 
227 static void
228 virtio_xmit_cleanup_inorder_packed(struct virtqueue *vq, int num)
229 {
230 	uint16_t used_idx, id, curr_id, free_cnt = 0;
231 	uint16_t size = vq->vq_nentries;
232 	struct vring_packed_desc *desc = vq->ring_packed.desc_packed;
233 	struct vq_desc_extra *dxp;
234 
235 	used_idx = vq->vq_used_cons_idx;
236 	while (num > 0 && desc_is_used(&desc[used_idx], vq)) {
237 		virtio_rmb(vq->hw->weak_barriers);
238 		id = desc[used_idx].id;
239 		do {
240 			curr_id = used_idx;
241 			dxp = &vq->vq_descx[used_idx];
242 			used_idx += dxp->ndescs;
243 			free_cnt += dxp->ndescs;
244 			num -= dxp->ndescs;
245 			if (used_idx >= size) {
246 				used_idx -= size;
247 				vq->used_wrap_counter ^= 1;
248 			}
249 			if (dxp->cookie != NULL) {
250 				rte_pktmbuf_free(dxp->cookie);
251 				dxp->cookie = NULL;
252 			}
253 		} while (curr_id != id);
254 	}
255 	vq->vq_used_cons_idx = used_idx;
256 	vq->vq_free_cnt += free_cnt;
257 }
258 
259 static void
260 virtio_xmit_cleanup_normal_packed(struct virtqueue *vq, int num)
261 {
262 	uint16_t used_idx, id;
263 	uint16_t size = vq->vq_nentries;
264 	struct vring_packed_desc *desc = vq->ring_packed.desc_packed;
265 	struct vq_desc_extra *dxp;
266 
267 	used_idx = vq->vq_used_cons_idx;
268 	while (num-- && desc_is_used(&desc[used_idx], vq)) {
269 		virtio_rmb(vq->hw->weak_barriers);
270 		id = desc[used_idx].id;
271 		dxp = &vq->vq_descx[id];
272 		vq->vq_used_cons_idx += dxp->ndescs;
273 		if (vq->vq_used_cons_idx >= size) {
274 			vq->vq_used_cons_idx -= size;
275 			vq->used_wrap_counter ^= 1;
276 		}
277 		vq_ring_free_id_packed(vq, id);
278 		if (dxp->cookie != NULL) {
279 			rte_pktmbuf_free(dxp->cookie);
280 			dxp->cookie = NULL;
281 		}
282 		used_idx = vq->vq_used_cons_idx;
283 	}
284 }
285 
286 /* Cleanup from completed transmits. */
287 static inline void
288 virtio_xmit_cleanup_packed(struct virtqueue *vq, int num, int in_order)
289 {
290 	if (in_order)
291 		virtio_xmit_cleanup_inorder_packed(vq, num);
292 	else
293 		virtio_xmit_cleanup_normal_packed(vq, num);
294 }
295 
296 static void
297 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
298 {
299 	uint16_t i, used_idx, desc_idx;
300 	for (i = 0; i < num; i++) {
301 		struct vring_used_elem *uep;
302 		struct vq_desc_extra *dxp;
303 
304 		used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
305 		uep = &vq->vq_ring.used->ring[used_idx];
306 
307 		desc_idx = (uint16_t) uep->id;
308 		dxp = &vq->vq_descx[desc_idx];
309 		vq->vq_used_cons_idx++;
310 		vq_ring_free_chain(vq, desc_idx);
311 
312 		if (dxp->cookie != NULL) {
313 			rte_pktmbuf_free(dxp->cookie);
314 			dxp->cookie = NULL;
315 		}
316 	}
317 }
318 
319 /* Cleanup from completed inorder transmits. */
320 static void
321 virtio_xmit_cleanup_inorder(struct virtqueue *vq, uint16_t num)
322 {
323 	uint16_t i, idx = vq->vq_used_cons_idx;
324 	int16_t free_cnt = 0;
325 	struct vq_desc_extra *dxp = NULL;
326 
327 	if (unlikely(num == 0))
328 		return;
329 
330 	for (i = 0; i < num; i++) {
331 		dxp = &vq->vq_descx[idx++ & (vq->vq_nentries - 1)];
332 		free_cnt += dxp->ndescs;
333 		if (dxp->cookie != NULL) {
334 			rte_pktmbuf_free(dxp->cookie);
335 			dxp->cookie = NULL;
336 		}
337 	}
338 
339 	vq->vq_free_cnt += free_cnt;
340 	vq->vq_used_cons_idx = idx;
341 }
342 
343 static inline int
344 virtqueue_enqueue_refill_inorder(struct virtqueue *vq,
345 			struct rte_mbuf **cookies,
346 			uint16_t num)
347 {
348 	struct vq_desc_extra *dxp;
349 	struct virtio_hw *hw = vq->hw;
350 	struct vring_desc *start_dp;
351 	uint16_t head_idx, idx, i = 0;
352 
353 	if (unlikely(vq->vq_free_cnt == 0))
354 		return -ENOSPC;
355 	if (unlikely(vq->vq_free_cnt < num))
356 		return -EMSGSIZE;
357 
358 	head_idx = vq->vq_desc_head_idx & (vq->vq_nentries - 1);
359 	start_dp = vq->vq_ring.desc;
360 
361 	while (i < num) {
362 		idx = head_idx & (vq->vq_nentries - 1);
363 		dxp = &vq->vq_descx[idx];
364 		dxp->cookie = (void *)cookies[i];
365 		dxp->ndescs = 1;
366 
367 		start_dp[idx].addr =
368 				VIRTIO_MBUF_ADDR(cookies[i], vq) +
369 				RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
370 		start_dp[idx].len =
371 				cookies[i]->buf_len -
372 				RTE_PKTMBUF_HEADROOM +
373 				hw->vtnet_hdr_size;
374 		start_dp[idx].flags =  VRING_DESC_F_WRITE;
375 
376 		vq_update_avail_ring(vq, idx);
377 		head_idx++;
378 		i++;
379 	}
380 
381 	vq->vq_desc_head_idx += num;
382 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
383 	return 0;
384 }
385 
386 static inline int
387 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf **cookie,
388 				uint16_t num)
389 {
390 	struct vq_desc_extra *dxp;
391 	struct virtio_hw *hw = vq->hw;
392 	struct vring_desc *start_dp = vq->vq_ring.desc;
393 	uint16_t idx, i;
394 
395 	if (unlikely(vq->vq_free_cnt == 0))
396 		return -ENOSPC;
397 	if (unlikely(vq->vq_free_cnt < num))
398 		return -EMSGSIZE;
399 
400 	if (unlikely(vq->vq_desc_head_idx >= vq->vq_nentries))
401 		return -EFAULT;
402 
403 	for (i = 0; i < num; i++) {
404 		idx = vq->vq_desc_head_idx;
405 		dxp = &vq->vq_descx[idx];
406 		dxp->cookie = (void *)cookie[i];
407 		dxp->ndescs = 1;
408 
409 		start_dp[idx].addr =
410 			VIRTIO_MBUF_ADDR(cookie[i], vq) +
411 			RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
412 		start_dp[idx].len =
413 			cookie[i]->buf_len - RTE_PKTMBUF_HEADROOM +
414 			hw->vtnet_hdr_size;
415 		start_dp[idx].flags = VRING_DESC_F_WRITE;
416 		vq->vq_desc_head_idx = start_dp[idx].next;
417 		vq_update_avail_ring(vq, idx);
418 		if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END) {
419 			vq->vq_desc_tail_idx = vq->vq_desc_head_idx;
420 			break;
421 		}
422 	}
423 
424 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
425 
426 	return 0;
427 }
428 
429 static inline int
430 virtqueue_enqueue_recv_refill_packed(struct virtqueue *vq,
431 				     struct rte_mbuf **cookie, uint16_t num)
432 {
433 	struct vring_packed_desc *start_dp = vq->ring_packed.desc_packed;
434 	uint16_t flags = vq->cached_flags;
435 	struct virtio_hw *hw = vq->hw;
436 	struct vq_desc_extra *dxp;
437 	uint16_t idx;
438 	int i;
439 
440 	if (unlikely(vq->vq_free_cnt == 0))
441 		return -ENOSPC;
442 	if (unlikely(vq->vq_free_cnt < num))
443 		return -EMSGSIZE;
444 
445 	for (i = 0; i < num; i++) {
446 		idx = vq->vq_avail_idx;
447 		dxp = &vq->vq_descx[idx];
448 		dxp->cookie = (void *)cookie[i];
449 		dxp->ndescs = 1;
450 
451 		start_dp[idx].addr = VIRTIO_MBUF_ADDR(cookie[i], vq) +
452 				RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
453 		start_dp[idx].len = cookie[i]->buf_len - RTE_PKTMBUF_HEADROOM
454 					+ hw->vtnet_hdr_size;
455 
456 		vq->vq_desc_head_idx = dxp->next;
457 		if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
458 			vq->vq_desc_tail_idx = vq->vq_desc_head_idx;
459 		virtio_wmb(hw->weak_barriers);
460 		start_dp[idx].flags = flags;
461 		if (++vq->vq_avail_idx >= vq->vq_nentries) {
462 			vq->vq_avail_idx -= vq->vq_nentries;
463 			vq->cached_flags ^=
464 				VRING_DESC_F_AVAIL(1) | VRING_DESC_F_USED(1);
465 			flags = vq->cached_flags;
466 		}
467 	}
468 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
469 	return 0;
470 }
471 
472 /* When doing TSO, the IP length is not included in the pseudo header
473  * checksum of the packet given to the PMD, but for virtio it is
474  * expected.
475  */
476 static void
477 virtio_tso_fix_cksum(struct rte_mbuf *m)
478 {
479 	/* common case: header is not fragmented */
480 	if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
481 			m->l4_len)) {
482 		struct ipv4_hdr *iph;
483 		struct ipv6_hdr *ip6h;
484 		struct tcp_hdr *th;
485 		uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
486 		uint32_t tmp;
487 
488 		iph = rte_pktmbuf_mtod_offset(m, struct ipv4_hdr *, m->l2_len);
489 		th = RTE_PTR_ADD(iph, m->l3_len);
490 		if ((iph->version_ihl >> 4) == 4) {
491 			iph->hdr_checksum = 0;
492 			iph->hdr_checksum = rte_ipv4_cksum(iph);
493 			ip_len = iph->total_length;
494 			ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
495 				m->l3_len);
496 		} else {
497 			ip6h = (struct ipv6_hdr *)iph;
498 			ip_paylen = ip6h->payload_len;
499 		}
500 
501 		/* calculate the new phdr checksum not including ip_paylen */
502 		prev_cksum = th->cksum;
503 		tmp = prev_cksum;
504 		tmp += ip_paylen;
505 		tmp = (tmp & 0xffff) + (tmp >> 16);
506 		new_cksum = tmp;
507 
508 		/* replace it in the packet */
509 		th->cksum = new_cksum;
510 	}
511 }
512 
513 
514 /* avoid write operation when necessary, to lessen cache issues */
515 #define ASSIGN_UNLESS_EQUAL(var, val) do {	\
516 	if ((var) != (val))			\
517 		(var) = (val);			\
518 } while (0)
519 
520 #define virtqueue_clear_net_hdr(_hdr) do {		\
521 	ASSIGN_UNLESS_EQUAL((_hdr)->csum_start, 0);	\
522 	ASSIGN_UNLESS_EQUAL((_hdr)->csum_offset, 0);	\
523 	ASSIGN_UNLESS_EQUAL((_hdr)->flags, 0);		\
524 	ASSIGN_UNLESS_EQUAL((_hdr)->gso_type, 0);	\
525 	ASSIGN_UNLESS_EQUAL((_hdr)->gso_size, 0);	\
526 	ASSIGN_UNLESS_EQUAL((_hdr)->hdr_len, 0);	\
527 } while (0)
528 
529 static inline void
530 virtqueue_xmit_offload(struct virtio_net_hdr *hdr,
531 			struct rte_mbuf *cookie,
532 			bool offload)
533 {
534 	if (offload) {
535 		if (cookie->ol_flags & PKT_TX_TCP_SEG)
536 			cookie->ol_flags |= PKT_TX_TCP_CKSUM;
537 
538 		switch (cookie->ol_flags & PKT_TX_L4_MASK) {
539 		case PKT_TX_UDP_CKSUM:
540 			hdr->csum_start = cookie->l2_len + cookie->l3_len;
541 			hdr->csum_offset = offsetof(struct udp_hdr,
542 				dgram_cksum);
543 			hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
544 			break;
545 
546 		case PKT_TX_TCP_CKSUM:
547 			hdr->csum_start = cookie->l2_len + cookie->l3_len;
548 			hdr->csum_offset = offsetof(struct tcp_hdr, cksum);
549 			hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
550 			break;
551 
552 		default:
553 			ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
554 			ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
555 			ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
556 			break;
557 		}
558 
559 		/* TCP Segmentation Offload */
560 		if (cookie->ol_flags & PKT_TX_TCP_SEG) {
561 			virtio_tso_fix_cksum(cookie);
562 			hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
563 				VIRTIO_NET_HDR_GSO_TCPV6 :
564 				VIRTIO_NET_HDR_GSO_TCPV4;
565 			hdr->gso_size = cookie->tso_segsz;
566 			hdr->hdr_len =
567 				cookie->l2_len +
568 				cookie->l3_len +
569 				cookie->l4_len;
570 		} else {
571 			ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
572 			ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
573 			ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
574 		}
575 	}
576 }
577 
578 static inline void
579 virtqueue_enqueue_xmit_inorder(struct virtnet_tx *txvq,
580 			struct rte_mbuf **cookies,
581 			uint16_t num)
582 {
583 	struct vq_desc_extra *dxp;
584 	struct virtqueue *vq = txvq->vq;
585 	struct vring_desc *start_dp;
586 	struct virtio_net_hdr *hdr;
587 	uint16_t idx;
588 	uint16_t head_size = vq->hw->vtnet_hdr_size;
589 	uint16_t i = 0;
590 
591 	idx = vq->vq_desc_head_idx;
592 	start_dp = vq->vq_ring.desc;
593 
594 	while (i < num) {
595 		idx = idx & (vq->vq_nentries - 1);
596 		dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
597 		dxp->cookie = (void *)cookies[i];
598 		dxp->ndescs = 1;
599 
600 		hdr = (struct virtio_net_hdr *)
601 			rte_pktmbuf_prepend(cookies[i], head_size);
602 		cookies[i]->pkt_len -= head_size;
603 
604 		/* if offload disabled, hdr is not zeroed yet, do it now */
605 		if (!vq->hw->has_tx_offload)
606 			virtqueue_clear_net_hdr(hdr);
607 		else
608 			virtqueue_xmit_offload(hdr, cookies[i], true);
609 
610 		start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookies[i], vq);
611 		start_dp[idx].len   = cookies[i]->data_len;
612 		start_dp[idx].flags = 0;
613 
614 		vq_update_avail_ring(vq, idx);
615 
616 		idx++;
617 		i++;
618 	};
619 
620 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
621 	vq->vq_desc_head_idx = idx & (vq->vq_nentries - 1);
622 }
623 
624 static inline void
625 virtqueue_enqueue_xmit_packed_fast(struct virtnet_tx *txvq,
626 				   struct rte_mbuf *cookie,
627 				   int in_order)
628 {
629 	struct virtqueue *vq = txvq->vq;
630 	struct vring_packed_desc *dp;
631 	struct vq_desc_extra *dxp;
632 	uint16_t idx, id, flags;
633 	uint16_t head_size = vq->hw->vtnet_hdr_size;
634 	struct virtio_net_hdr *hdr;
635 
636 	id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx;
637 	idx = vq->vq_avail_idx;
638 	dp = &vq->ring_packed.desc_packed[idx];
639 
640 	dxp = &vq->vq_descx[id];
641 	dxp->ndescs = 1;
642 	dxp->cookie = cookie;
643 
644 	flags = vq->cached_flags;
645 
646 	/* prepend cannot fail, checked by caller */
647 	hdr = (struct virtio_net_hdr *)
648 		rte_pktmbuf_prepend(cookie, head_size);
649 	cookie->pkt_len -= head_size;
650 
651 	/* if offload disabled, hdr is not zeroed yet, do it now */
652 	if (!vq->hw->has_tx_offload)
653 		virtqueue_clear_net_hdr(hdr);
654 	else
655 		virtqueue_xmit_offload(hdr, cookie, true);
656 
657 	dp->addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
658 	dp->len  = cookie->data_len;
659 	dp->id   = id;
660 
661 	if (++vq->vq_avail_idx >= vq->vq_nentries) {
662 		vq->vq_avail_idx -= vq->vq_nentries;
663 		vq->cached_flags ^=
664 			VRING_DESC_F_AVAIL(1) | VRING_DESC_F_USED(1);
665 	}
666 
667 	vq->vq_free_cnt--;
668 
669 	if (!in_order) {
670 		vq->vq_desc_head_idx = dxp->next;
671 		if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
672 			vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
673 	}
674 
675 	virtio_wmb(vq->hw->weak_barriers);
676 	dp->flags = flags;
677 }
678 
679 static inline void
680 virtqueue_enqueue_xmit_packed(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
681 			      uint16_t needed, int can_push, int in_order)
682 {
683 	struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
684 	struct vq_desc_extra *dxp;
685 	struct virtqueue *vq = txvq->vq;
686 	struct vring_packed_desc *start_dp, *head_dp;
687 	uint16_t idx, id, head_idx, head_flags;
688 	uint16_t head_size = vq->hw->vtnet_hdr_size;
689 	struct virtio_net_hdr *hdr;
690 	uint16_t prev;
691 
692 	id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx;
693 
694 	dxp = &vq->vq_descx[id];
695 	dxp->ndescs = needed;
696 	dxp->cookie = cookie;
697 
698 	head_idx = vq->vq_avail_idx;
699 	idx = head_idx;
700 	prev = head_idx;
701 	start_dp = vq->ring_packed.desc_packed;
702 
703 	head_dp = &vq->ring_packed.desc_packed[idx];
704 	head_flags = cookie->next ? VRING_DESC_F_NEXT : 0;
705 	head_flags |= vq->cached_flags;
706 
707 	if (can_push) {
708 		/* prepend cannot fail, checked by caller */
709 		hdr = (struct virtio_net_hdr *)
710 			rte_pktmbuf_prepend(cookie, head_size);
711 		/* rte_pktmbuf_prepend() counts the hdr size to the pkt length,
712 		 * which is wrong. Below subtract restores correct pkt size.
713 		 */
714 		cookie->pkt_len -= head_size;
715 
716 		/* if offload disabled, it is not zeroed below, do it now */
717 		if (!vq->hw->has_tx_offload)
718 			virtqueue_clear_net_hdr(hdr);
719 	} else {
720 		/* setup first tx ring slot to point to header
721 		 * stored in reserved region.
722 		 */
723 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
724 			RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
725 		start_dp[idx].len   = vq->hw->vtnet_hdr_size;
726 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
727 		idx++;
728 		if (idx >= vq->vq_nentries) {
729 			idx -= vq->vq_nentries;
730 			vq->cached_flags ^=
731 				VRING_DESC_F_AVAIL(1) | VRING_DESC_F_USED(1);
732 		}
733 	}
734 
735 	virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload);
736 
737 	do {
738 		uint16_t flags;
739 
740 		start_dp[idx].addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
741 		start_dp[idx].len  = cookie->data_len;
742 		if (likely(idx != head_idx)) {
743 			flags = cookie->next ? VRING_DESC_F_NEXT : 0;
744 			flags |= vq->cached_flags;
745 			start_dp[idx].flags = flags;
746 		}
747 		prev = idx;
748 		idx++;
749 		if (idx >= vq->vq_nentries) {
750 			idx -= vq->vq_nentries;
751 			vq->cached_flags ^=
752 				VRING_DESC_F_AVAIL(1) | VRING_DESC_F_USED(1);
753 		}
754 	} while ((cookie = cookie->next) != NULL);
755 
756 	start_dp[prev].id = id;
757 
758 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
759 	vq->vq_avail_idx = idx;
760 
761 	if (!in_order) {
762 		vq->vq_desc_head_idx = dxp->next;
763 		if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
764 			vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
765 	}
766 
767 	virtio_wmb(vq->hw->weak_barriers);
768 	head_dp->flags = head_flags;
769 }
770 
771 static inline void
772 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
773 			uint16_t needed, int use_indirect, int can_push,
774 			int in_order)
775 {
776 	struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
777 	struct vq_desc_extra *dxp;
778 	struct virtqueue *vq = txvq->vq;
779 	struct vring_desc *start_dp;
780 	uint16_t seg_num = cookie->nb_segs;
781 	uint16_t head_idx, idx;
782 	uint16_t head_size = vq->hw->vtnet_hdr_size;
783 	struct virtio_net_hdr *hdr;
784 
785 	head_idx = vq->vq_desc_head_idx;
786 	idx = head_idx;
787 	if (in_order)
788 		dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
789 	else
790 		dxp = &vq->vq_descx[idx];
791 	dxp->cookie = (void *)cookie;
792 	dxp->ndescs = needed;
793 
794 	start_dp = vq->vq_ring.desc;
795 
796 	if (can_push) {
797 		/* prepend cannot fail, checked by caller */
798 		hdr = (struct virtio_net_hdr *)
799 			rte_pktmbuf_prepend(cookie, head_size);
800 		/* rte_pktmbuf_prepend() counts the hdr size to the pkt length,
801 		 * which is wrong. Below subtract restores correct pkt size.
802 		 */
803 		cookie->pkt_len -= head_size;
804 
805 		/* if offload disabled, it is not zeroed below, do it now */
806 		if (!vq->hw->has_tx_offload)
807 			virtqueue_clear_net_hdr(hdr);
808 	} else if (use_indirect) {
809 		/* setup tx ring slot to point to indirect
810 		 * descriptor list stored in reserved region.
811 		 *
812 		 * the first slot in indirect ring is already preset
813 		 * to point to the header in reserved region
814 		 */
815 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
816 			RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
817 		start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
818 		start_dp[idx].flags = VRING_DESC_F_INDIRECT;
819 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
820 
821 		/* loop below will fill in rest of the indirect elements */
822 		start_dp = txr[idx].tx_indir;
823 		idx = 1;
824 	} else {
825 		/* setup first tx ring slot to point to header
826 		 * stored in reserved region.
827 		 */
828 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
829 			RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
830 		start_dp[idx].len   = vq->hw->vtnet_hdr_size;
831 		start_dp[idx].flags = VRING_DESC_F_NEXT;
832 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
833 
834 		idx = start_dp[idx].next;
835 	}
836 
837 	virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload);
838 
839 	do {
840 		start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
841 		start_dp[idx].len   = cookie->data_len;
842 		start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
843 		idx = start_dp[idx].next;
844 	} while ((cookie = cookie->next) != NULL);
845 
846 	if (use_indirect)
847 		idx = vq->vq_ring.desc[head_idx].next;
848 
849 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
850 
851 	vq->vq_desc_head_idx = idx;
852 	vq_update_avail_ring(vq, head_idx);
853 
854 	if (!in_order) {
855 		if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
856 			vq->vq_desc_tail_idx = idx;
857 	}
858 }
859 
860 void
861 virtio_dev_cq_start(struct rte_eth_dev *dev)
862 {
863 	struct virtio_hw *hw = dev->data->dev_private;
864 
865 	if (hw->cvq && hw->cvq->vq) {
866 		rte_spinlock_init(&hw->cvq->lock);
867 		VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
868 	}
869 }
870 
871 int
872 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
873 			uint16_t queue_idx,
874 			uint16_t nb_desc,
875 			unsigned int socket_id __rte_unused,
876 			const struct rte_eth_rxconf *rx_conf __rte_unused,
877 			struct rte_mempool *mp)
878 {
879 	uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
880 	struct virtio_hw *hw = dev->data->dev_private;
881 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
882 	struct virtnet_rx *rxvq;
883 
884 	PMD_INIT_FUNC_TRACE();
885 
886 	if (nb_desc == 0 || nb_desc > vq->vq_nentries)
887 		nb_desc = vq->vq_nentries;
888 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
889 
890 	rxvq = &vq->rxq;
891 	rxvq->queue_id = queue_idx;
892 	rxvq->mpool = mp;
893 	if (rxvq->mpool == NULL) {
894 		rte_exit(EXIT_FAILURE,
895 			"Cannot allocate mbufs for rx virtqueue");
896 	}
897 
898 	dev->data->rx_queues[queue_idx] = rxvq;
899 
900 	return 0;
901 }
902 
903 int
904 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
905 {
906 	uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
907 	struct virtio_hw *hw = dev->data->dev_private;
908 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
909 	struct virtnet_rx *rxvq = &vq->rxq;
910 	struct rte_mbuf *m;
911 	uint16_t desc_idx;
912 	int error, nbufs, i;
913 
914 	PMD_INIT_FUNC_TRACE();
915 
916 	/* Allocate blank mbufs for the each rx descriptor */
917 	nbufs = 0;
918 
919 	if (hw->use_simple_rx) {
920 		for (desc_idx = 0; desc_idx < vq->vq_nentries;
921 		     desc_idx++) {
922 			vq->vq_ring.avail->ring[desc_idx] = desc_idx;
923 			vq->vq_ring.desc[desc_idx].flags =
924 				VRING_DESC_F_WRITE;
925 		}
926 
927 		virtio_rxq_vec_setup(rxvq);
928 	}
929 
930 	memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
931 	for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
932 	     desc_idx++) {
933 		vq->sw_ring[vq->vq_nentries + desc_idx] =
934 			&rxvq->fake_mbuf;
935 	}
936 
937 	if (hw->use_simple_rx) {
938 		while (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
939 			virtio_rxq_rearm_vec(rxvq);
940 			nbufs += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
941 		}
942 	} else if (hw->use_inorder_rx) {
943 		if ((!virtqueue_full(vq))) {
944 			uint16_t free_cnt = vq->vq_free_cnt;
945 			struct rte_mbuf *pkts[free_cnt];
946 
947 			if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, pkts,
948 				free_cnt)) {
949 				error = virtqueue_enqueue_refill_inorder(vq,
950 						pkts,
951 						free_cnt);
952 				if (unlikely(error)) {
953 					for (i = 0; i < free_cnt; i++)
954 						rte_pktmbuf_free(pkts[i]);
955 				}
956 			}
957 
958 			nbufs += free_cnt;
959 			vq_update_avail_idx(vq);
960 		}
961 	} else {
962 		while (!virtqueue_full(vq)) {
963 			m = rte_mbuf_raw_alloc(rxvq->mpool);
964 			if (m == NULL)
965 				break;
966 
967 			/* Enqueue allocated buffers */
968 			if (vtpci_packed_queue(vq->hw))
969 				error = virtqueue_enqueue_recv_refill_packed(vq,
970 						&m, 1);
971 			else
972 				error = virtqueue_enqueue_recv_refill(vq,
973 						&m, 1);
974 			if (error) {
975 				rte_pktmbuf_free(m);
976 				break;
977 			}
978 			nbufs++;
979 		}
980 
981 		if (!vtpci_packed_queue(vq->hw))
982 			vq_update_avail_idx(vq);
983 	}
984 
985 	PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
986 
987 	VIRTQUEUE_DUMP(vq);
988 
989 	return 0;
990 }
991 
992 /*
993  * struct rte_eth_dev *dev: Used to update dev
994  * uint16_t nb_desc: Defaults to values read from config space
995  * unsigned int socket_id: Used to allocate memzone
996  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
997  * uint16_t queue_idx: Just used as an index in dev txq list
998  */
999 int
1000 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
1001 			uint16_t queue_idx,
1002 			uint16_t nb_desc,
1003 			unsigned int socket_id __rte_unused,
1004 			const struct rte_eth_txconf *tx_conf)
1005 {
1006 	uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
1007 	struct virtio_hw *hw = dev->data->dev_private;
1008 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
1009 	struct virtnet_tx *txvq;
1010 	uint16_t tx_free_thresh;
1011 
1012 	PMD_INIT_FUNC_TRACE();
1013 
1014 	if (nb_desc == 0 || nb_desc > vq->vq_nentries)
1015 		nb_desc = vq->vq_nentries;
1016 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
1017 
1018 	txvq = &vq->txq;
1019 	txvq->queue_id = queue_idx;
1020 
1021 	tx_free_thresh = tx_conf->tx_free_thresh;
1022 	if (tx_free_thresh == 0)
1023 		tx_free_thresh =
1024 			RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
1025 
1026 	if (tx_free_thresh >= (vq->vq_nentries - 3)) {
1027 		RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
1028 			"number of TX entries minus 3 (%u)."
1029 			" (tx_free_thresh=%u port=%u queue=%u)\n",
1030 			vq->vq_nentries - 3,
1031 			tx_free_thresh, dev->data->port_id, queue_idx);
1032 		return -EINVAL;
1033 	}
1034 
1035 	vq->vq_free_thresh = tx_free_thresh;
1036 
1037 	dev->data->tx_queues[queue_idx] = txvq;
1038 	return 0;
1039 }
1040 
1041 int
1042 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
1043 				uint16_t queue_idx)
1044 {
1045 	uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
1046 	struct virtio_hw *hw = dev->data->dev_private;
1047 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
1048 
1049 	PMD_INIT_FUNC_TRACE();
1050 
1051 	if (!vtpci_packed_queue(hw)) {
1052 		if (hw->use_inorder_tx)
1053 			vq->vq_ring.desc[vq->vq_nentries - 1].next = 0;
1054 	}
1055 
1056 	VIRTQUEUE_DUMP(vq);
1057 
1058 	return 0;
1059 }
1060 
1061 static inline void
1062 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
1063 {
1064 	int error;
1065 	/*
1066 	 * Requeue the discarded mbuf. This should always be
1067 	 * successful since it was just dequeued.
1068 	 */
1069 	if (vtpci_packed_queue(vq->hw))
1070 		error = virtqueue_enqueue_recv_refill_packed(vq, &m, 1);
1071 	else
1072 		error = virtqueue_enqueue_recv_refill(vq, &m, 1);
1073 
1074 	if (unlikely(error)) {
1075 		RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
1076 		rte_pktmbuf_free(m);
1077 	}
1078 }
1079 
1080 static inline void
1081 virtio_discard_rxbuf_inorder(struct virtqueue *vq, struct rte_mbuf *m)
1082 {
1083 	int error;
1084 
1085 	error = virtqueue_enqueue_refill_inorder(vq, &m, 1);
1086 	if (unlikely(error)) {
1087 		RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
1088 		rte_pktmbuf_free(m);
1089 	}
1090 }
1091 
1092 static inline void
1093 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
1094 {
1095 	uint32_t s = mbuf->pkt_len;
1096 	struct ether_addr *ea;
1097 
1098 	stats->bytes += s;
1099 
1100 	if (s == 64) {
1101 		stats->size_bins[1]++;
1102 	} else if (s > 64 && s < 1024) {
1103 		uint32_t bin;
1104 
1105 		/* count zeros, and offset into correct bin */
1106 		bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
1107 		stats->size_bins[bin]++;
1108 	} else {
1109 		if (s < 64)
1110 			stats->size_bins[0]++;
1111 		else if (s < 1519)
1112 			stats->size_bins[6]++;
1113 		else if (s >= 1519)
1114 			stats->size_bins[7]++;
1115 	}
1116 
1117 	ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
1118 	if (is_multicast_ether_addr(ea)) {
1119 		if (is_broadcast_ether_addr(ea))
1120 			stats->broadcast++;
1121 		else
1122 			stats->multicast++;
1123 	}
1124 }
1125 
1126 static inline void
1127 virtio_rx_stats_updated(struct virtnet_rx *rxvq, struct rte_mbuf *m)
1128 {
1129 	VIRTIO_DUMP_PACKET(m, m->data_len);
1130 
1131 	virtio_update_packet_stats(&rxvq->stats, m);
1132 }
1133 
1134 /* Optionally fill offload information in structure */
1135 static inline int
1136 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
1137 {
1138 	struct rte_net_hdr_lens hdr_lens;
1139 	uint32_t hdrlen, ptype;
1140 	int l4_supported = 0;
1141 
1142 	/* nothing to do */
1143 	if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1144 		return 0;
1145 
1146 	m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
1147 
1148 	ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
1149 	m->packet_type = ptype;
1150 	if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
1151 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
1152 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
1153 		l4_supported = 1;
1154 
1155 	if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1156 		hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
1157 		if (hdr->csum_start <= hdrlen && l4_supported) {
1158 			m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
1159 		} else {
1160 			/* Unknown proto or tunnel, do sw cksum. We can assume
1161 			 * the cksum field is in the first segment since the
1162 			 * buffers we provided to the host are large enough.
1163 			 * In case of SCTP, this will be wrong since it's a CRC
1164 			 * but there's nothing we can do.
1165 			 */
1166 			uint16_t csum = 0, off;
1167 
1168 			rte_raw_cksum_mbuf(m, hdr->csum_start,
1169 				rte_pktmbuf_pkt_len(m) - hdr->csum_start,
1170 				&csum);
1171 			if (likely(csum != 0xffff))
1172 				csum = ~csum;
1173 			off = hdr->csum_offset + hdr->csum_start;
1174 			if (rte_pktmbuf_data_len(m) >= off + 1)
1175 				*rte_pktmbuf_mtod_offset(m, uint16_t *,
1176 					off) = csum;
1177 		}
1178 	} else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
1179 		m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
1180 	}
1181 
1182 	/* GSO request, save required information in mbuf */
1183 	if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1184 		/* Check unsupported modes */
1185 		if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
1186 		    (hdr->gso_size == 0)) {
1187 			return -EINVAL;
1188 		}
1189 
1190 		/* Update mss lengthes in mbuf */
1191 		m->tso_segsz = hdr->gso_size;
1192 		switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1193 			case VIRTIO_NET_HDR_GSO_TCPV4:
1194 			case VIRTIO_NET_HDR_GSO_TCPV6:
1195 				m->ol_flags |= PKT_RX_LRO | \
1196 					PKT_RX_L4_CKSUM_NONE;
1197 				break;
1198 			default:
1199 				return -EINVAL;
1200 		}
1201 	}
1202 
1203 	return 0;
1204 }
1205 
1206 #define VIRTIO_MBUF_BURST_SZ 64
1207 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
1208 uint16_t
1209 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1210 {
1211 	struct virtnet_rx *rxvq = rx_queue;
1212 	struct virtqueue *vq = rxvq->vq;
1213 	struct virtio_hw *hw = vq->hw;
1214 	struct rte_mbuf *rxm, *new_mbuf;
1215 	uint16_t nb_used, num, nb_rx;
1216 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1217 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1218 	int error;
1219 	uint32_t i, nb_enqueued;
1220 	uint32_t hdr_size;
1221 	struct virtio_net_hdr *hdr;
1222 
1223 	nb_rx = 0;
1224 	if (unlikely(hw->started == 0))
1225 		return nb_rx;
1226 
1227 	nb_used = VIRTQUEUE_NUSED(vq);
1228 
1229 	virtio_rmb(hw->weak_barriers);
1230 
1231 	num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1232 	if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1233 		num = VIRTIO_MBUF_BURST_SZ;
1234 	if (likely(num > DESC_PER_CACHELINE))
1235 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1236 
1237 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1238 	PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
1239 
1240 	nb_enqueued = 0;
1241 	hdr_size = hw->vtnet_hdr_size;
1242 
1243 	for (i = 0; i < num ; i++) {
1244 		rxm = rcv_pkts[i];
1245 
1246 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1247 
1248 		if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
1249 			PMD_RX_LOG(ERR, "Packet drop");
1250 			nb_enqueued++;
1251 			virtio_discard_rxbuf(vq, rxm);
1252 			rxvq->stats.errors++;
1253 			continue;
1254 		}
1255 
1256 		rxm->port = rxvq->port_id;
1257 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1258 		rxm->ol_flags = 0;
1259 		rxm->vlan_tci = 0;
1260 
1261 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1262 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1263 
1264 		hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1265 			RTE_PKTMBUF_HEADROOM - hdr_size);
1266 
1267 		if (hw->vlan_strip)
1268 			rte_vlan_strip(rxm);
1269 
1270 		if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1271 			virtio_discard_rxbuf(vq, rxm);
1272 			rxvq->stats.errors++;
1273 			continue;
1274 		}
1275 
1276 		virtio_rx_stats_updated(rxvq, rxm);
1277 
1278 		rx_pkts[nb_rx++] = rxm;
1279 	}
1280 
1281 	rxvq->stats.packets += nb_rx;
1282 
1283 	/* Allocate new mbuf for the used descriptor */
1284 	while (likely(!virtqueue_full(vq))) {
1285 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
1286 		if (unlikely(new_mbuf == NULL)) {
1287 			struct rte_eth_dev *dev
1288 				= &rte_eth_devices[rxvq->port_id];
1289 			dev->data->rx_mbuf_alloc_failed++;
1290 			break;
1291 		}
1292 		error = virtqueue_enqueue_recv_refill(vq, &new_mbuf, 1);
1293 		if (unlikely(error)) {
1294 			rte_pktmbuf_free(new_mbuf);
1295 			break;
1296 		}
1297 		nb_enqueued++;
1298 	}
1299 
1300 	if (likely(nb_enqueued)) {
1301 		vq_update_avail_idx(vq);
1302 
1303 		if (unlikely(virtqueue_kick_prepare(vq))) {
1304 			virtqueue_notify(vq);
1305 			PMD_RX_LOG(DEBUG, "Notified");
1306 		}
1307 	}
1308 
1309 	return nb_rx;
1310 }
1311 
1312 uint16_t
1313 virtio_recv_pkts_packed(void *rx_queue, struct rte_mbuf **rx_pkts,
1314 			uint16_t nb_pkts)
1315 {
1316 	struct virtnet_rx *rxvq = rx_queue;
1317 	struct virtqueue *vq = rxvq->vq;
1318 	struct virtio_hw *hw = vq->hw;
1319 	struct rte_mbuf *rxm, *new_mbuf;
1320 	uint16_t num, nb_rx;
1321 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1322 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1323 	int error;
1324 	uint32_t i, nb_enqueued;
1325 	uint32_t hdr_size;
1326 	struct virtio_net_hdr *hdr;
1327 
1328 	nb_rx = 0;
1329 	if (unlikely(hw->started == 0))
1330 		return nb_rx;
1331 
1332 	num = RTE_MIN(VIRTIO_MBUF_BURST_SZ, nb_pkts);
1333 	if (likely(num > DESC_PER_CACHELINE))
1334 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1335 
1336 	num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1337 	PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1338 
1339 	nb_enqueued = 0;
1340 	hdr_size = hw->vtnet_hdr_size;
1341 
1342 	for (i = 0; i < num; i++) {
1343 		rxm = rcv_pkts[i];
1344 
1345 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1346 
1347 		if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
1348 			PMD_RX_LOG(ERR, "Packet drop");
1349 			nb_enqueued++;
1350 			virtio_discard_rxbuf(vq, rxm);
1351 			rxvq->stats.errors++;
1352 			continue;
1353 		}
1354 
1355 		rxm->port = rxvq->port_id;
1356 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1357 		rxm->ol_flags = 0;
1358 		rxm->vlan_tci = 0;
1359 
1360 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1361 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1362 
1363 		hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1364 			RTE_PKTMBUF_HEADROOM - hdr_size);
1365 
1366 		if (hw->vlan_strip)
1367 			rte_vlan_strip(rxm);
1368 
1369 		if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1370 			virtio_discard_rxbuf(vq, rxm);
1371 			rxvq->stats.errors++;
1372 			continue;
1373 		}
1374 
1375 		virtio_rx_stats_updated(rxvq, rxm);
1376 
1377 		rx_pkts[nb_rx++] = rxm;
1378 	}
1379 
1380 	rxvq->stats.packets += nb_rx;
1381 
1382 	/* Allocate new mbuf for the used descriptor */
1383 	while (likely(!virtqueue_full(vq))) {
1384 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
1385 		if (unlikely(new_mbuf == NULL)) {
1386 			struct rte_eth_dev *dev =
1387 				&rte_eth_devices[rxvq->port_id];
1388 			dev->data->rx_mbuf_alloc_failed++;
1389 			break;
1390 		}
1391 		error = virtqueue_enqueue_recv_refill_packed(vq, &new_mbuf, 1);
1392 		if (unlikely(error)) {
1393 			rte_pktmbuf_free(new_mbuf);
1394 			break;
1395 		}
1396 		nb_enqueued++;
1397 	}
1398 
1399 	if (likely(nb_enqueued)) {
1400 		if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1401 			virtqueue_notify(vq);
1402 			PMD_RX_LOG(DEBUG, "Notified");
1403 		}
1404 	}
1405 
1406 	return nb_rx;
1407 }
1408 
1409 
1410 uint16_t
1411 virtio_recv_pkts_inorder(void *rx_queue,
1412 			struct rte_mbuf **rx_pkts,
1413 			uint16_t nb_pkts)
1414 {
1415 	struct virtnet_rx *rxvq = rx_queue;
1416 	struct virtqueue *vq = rxvq->vq;
1417 	struct virtio_hw *hw = vq->hw;
1418 	struct rte_mbuf *rxm;
1419 	struct rte_mbuf *prev;
1420 	uint16_t nb_used, num, nb_rx;
1421 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1422 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1423 	int error;
1424 	uint32_t nb_enqueued;
1425 	uint32_t seg_num;
1426 	uint32_t seg_res;
1427 	uint32_t hdr_size;
1428 	int32_t i;
1429 
1430 	nb_rx = 0;
1431 	if (unlikely(hw->started == 0))
1432 		return nb_rx;
1433 
1434 	nb_used = VIRTQUEUE_NUSED(vq);
1435 	nb_used = RTE_MIN(nb_used, nb_pkts);
1436 	nb_used = RTE_MIN(nb_used, VIRTIO_MBUF_BURST_SZ);
1437 
1438 	virtio_rmb(hw->weak_barriers);
1439 
1440 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1441 
1442 	nb_enqueued = 0;
1443 	seg_num = 1;
1444 	seg_res = 0;
1445 	hdr_size = hw->vtnet_hdr_size;
1446 
1447 	num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len, nb_used);
1448 
1449 	for (i = 0; i < num; i++) {
1450 		struct virtio_net_hdr_mrg_rxbuf *header;
1451 
1452 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1453 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1454 
1455 		rxm = rcv_pkts[i];
1456 
1457 		if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
1458 			PMD_RX_LOG(ERR, "Packet drop");
1459 			nb_enqueued++;
1460 			virtio_discard_rxbuf_inorder(vq, rxm);
1461 			rxvq->stats.errors++;
1462 			continue;
1463 		}
1464 
1465 		header = (struct virtio_net_hdr_mrg_rxbuf *)
1466 			 ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1467 			 - hdr_size);
1468 
1469 		if (vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
1470 			seg_num = header->num_buffers;
1471 			if (seg_num == 0)
1472 				seg_num = 1;
1473 		} else {
1474 			seg_num = 1;
1475 		}
1476 
1477 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1478 		rxm->nb_segs = seg_num;
1479 		rxm->ol_flags = 0;
1480 		rxm->vlan_tci = 0;
1481 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1482 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1483 
1484 		rxm->port = rxvq->port_id;
1485 
1486 		rx_pkts[nb_rx] = rxm;
1487 		prev = rxm;
1488 
1489 		if (vq->hw->has_rx_offload &&
1490 				virtio_rx_offload(rxm, &header->hdr) < 0) {
1491 			virtio_discard_rxbuf_inorder(vq, rxm);
1492 			rxvq->stats.errors++;
1493 			continue;
1494 		}
1495 
1496 		if (hw->vlan_strip)
1497 			rte_vlan_strip(rx_pkts[nb_rx]);
1498 
1499 		seg_res = seg_num - 1;
1500 
1501 		/* Merge remaining segments */
1502 		while (seg_res != 0 && i < (num - 1)) {
1503 			i++;
1504 
1505 			rxm = rcv_pkts[i];
1506 			rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1507 			rxm->pkt_len = (uint32_t)(len[i]);
1508 			rxm->data_len = (uint16_t)(len[i]);
1509 
1510 			rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1511 			rx_pkts[nb_rx]->data_len += (uint16_t)(len[i]);
1512 
1513 			if (prev)
1514 				prev->next = rxm;
1515 
1516 			prev = rxm;
1517 			seg_res -= 1;
1518 		}
1519 
1520 		if (!seg_res) {
1521 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1522 			nb_rx++;
1523 		}
1524 	}
1525 
1526 	/* Last packet still need merge segments */
1527 	while (seg_res != 0) {
1528 		uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1529 					VIRTIO_MBUF_BURST_SZ);
1530 
1531 		prev = rcv_pkts[nb_rx];
1532 		if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
1533 			virtio_rmb(hw->weak_barriers);
1534 			num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len,
1535 							   rcv_cnt);
1536 			uint16_t extra_idx = 0;
1537 
1538 			rcv_cnt = num;
1539 			while (extra_idx < rcv_cnt) {
1540 				rxm = rcv_pkts[extra_idx];
1541 				rxm->data_off =
1542 					RTE_PKTMBUF_HEADROOM - hdr_size;
1543 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
1544 				rxm->data_len = (uint16_t)(len[extra_idx]);
1545 				prev->next = rxm;
1546 				prev = rxm;
1547 				rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1548 				rx_pkts[nb_rx]->data_len += len[extra_idx];
1549 				extra_idx += 1;
1550 			};
1551 			seg_res -= rcv_cnt;
1552 
1553 			if (!seg_res) {
1554 				virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1555 				nb_rx++;
1556 			}
1557 		} else {
1558 			PMD_RX_LOG(ERR,
1559 					"No enough segments for packet.");
1560 			virtio_discard_rxbuf_inorder(vq, prev);
1561 			rxvq->stats.errors++;
1562 			break;
1563 		}
1564 	}
1565 
1566 	rxvq->stats.packets += nb_rx;
1567 
1568 	/* Allocate new mbuf for the used descriptor */
1569 
1570 	if (likely(!virtqueue_full(vq))) {
1571 		/* free_cnt may include mrg descs */
1572 		uint16_t free_cnt = vq->vq_free_cnt;
1573 		struct rte_mbuf *new_pkts[free_cnt];
1574 
1575 		if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1576 			error = virtqueue_enqueue_refill_inorder(vq, new_pkts,
1577 					free_cnt);
1578 			if (unlikely(error)) {
1579 				for (i = 0; i < free_cnt; i++)
1580 					rte_pktmbuf_free(new_pkts[i]);
1581 			}
1582 			nb_enqueued += free_cnt;
1583 		} else {
1584 			struct rte_eth_dev *dev =
1585 				&rte_eth_devices[rxvq->port_id];
1586 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1587 		}
1588 	}
1589 
1590 	if (likely(nb_enqueued)) {
1591 		vq_update_avail_idx(vq);
1592 
1593 		if (unlikely(virtqueue_kick_prepare(vq))) {
1594 			virtqueue_notify(vq);
1595 			PMD_RX_LOG(DEBUG, "Notified");
1596 		}
1597 	}
1598 
1599 	return nb_rx;
1600 }
1601 
1602 uint16_t
1603 virtio_recv_mergeable_pkts(void *rx_queue,
1604 			struct rte_mbuf **rx_pkts,
1605 			uint16_t nb_pkts)
1606 {
1607 	struct virtnet_rx *rxvq = rx_queue;
1608 	struct virtqueue *vq = rxvq->vq;
1609 	struct virtio_hw *hw = vq->hw;
1610 	struct rte_mbuf *rxm;
1611 	struct rte_mbuf *prev;
1612 	uint16_t nb_used, num, nb_rx = 0;
1613 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1614 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1615 	int error;
1616 	uint32_t nb_enqueued = 0;
1617 	uint32_t seg_num = 0;
1618 	uint32_t seg_res = 0;
1619 	uint32_t hdr_size = hw->vtnet_hdr_size;
1620 	int32_t i;
1621 
1622 	if (unlikely(hw->started == 0))
1623 		return nb_rx;
1624 
1625 	nb_used = VIRTQUEUE_NUSED(vq);
1626 
1627 	virtio_rmb(hw->weak_barriers);
1628 
1629 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1630 
1631 	num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1632 	if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1633 		num = VIRTIO_MBUF_BURST_SZ;
1634 	if (likely(num > DESC_PER_CACHELINE))
1635 		num = num - ((vq->vq_used_cons_idx + num) %
1636 				DESC_PER_CACHELINE);
1637 
1638 
1639 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1640 
1641 	for (i = 0; i < num; i++) {
1642 		struct virtio_net_hdr_mrg_rxbuf *header;
1643 
1644 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1645 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1646 
1647 		rxm = rcv_pkts[i];
1648 
1649 		if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
1650 			PMD_RX_LOG(ERR, "Packet drop");
1651 			nb_enqueued++;
1652 			virtio_discard_rxbuf(vq, rxm);
1653 			rxvq->stats.errors++;
1654 			continue;
1655 		}
1656 
1657 		header = (struct virtio_net_hdr_mrg_rxbuf *)
1658 			 ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1659 			 - hdr_size);
1660 		seg_num = header->num_buffers;
1661 		if (seg_num == 0)
1662 			seg_num = 1;
1663 
1664 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1665 		rxm->nb_segs = seg_num;
1666 		rxm->ol_flags = 0;
1667 		rxm->vlan_tci = 0;
1668 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1669 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1670 
1671 		rxm->port = rxvq->port_id;
1672 
1673 		rx_pkts[nb_rx] = rxm;
1674 		prev = rxm;
1675 
1676 		if (hw->has_rx_offload &&
1677 				virtio_rx_offload(rxm, &header->hdr) < 0) {
1678 			virtio_discard_rxbuf(vq, rxm);
1679 			rxvq->stats.errors++;
1680 			continue;
1681 		}
1682 
1683 		if (hw->vlan_strip)
1684 			rte_vlan_strip(rx_pkts[nb_rx]);
1685 
1686 		seg_res = seg_num - 1;
1687 
1688 		/* Merge remaining segments */
1689 		while (seg_res != 0 && i < (num - 1)) {
1690 			i++;
1691 
1692 			rxm = rcv_pkts[i];
1693 			rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1694 			rxm->pkt_len = (uint32_t)(len[i]);
1695 			rxm->data_len = (uint16_t)(len[i]);
1696 
1697 			rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1698 			rx_pkts[nb_rx]->data_len += (uint16_t)(len[i]);
1699 
1700 			if (prev)
1701 				prev->next = rxm;
1702 
1703 			prev = rxm;
1704 			seg_res -= 1;
1705 		}
1706 
1707 		if (!seg_res) {
1708 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1709 			nb_rx++;
1710 		}
1711 	}
1712 
1713 	/* Last packet still need merge segments */
1714 	while (seg_res != 0) {
1715 		uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1716 					VIRTIO_MBUF_BURST_SZ);
1717 
1718 		prev = rcv_pkts[nb_rx];
1719 		if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
1720 			virtio_rmb(hw->weak_barriers);
1721 			num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len,
1722 							   rcv_cnt);
1723 			uint16_t extra_idx = 0;
1724 
1725 			rcv_cnt = num;
1726 			while (extra_idx < rcv_cnt) {
1727 				rxm = rcv_pkts[extra_idx];
1728 				rxm->data_off =
1729 					RTE_PKTMBUF_HEADROOM - hdr_size;
1730 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
1731 				rxm->data_len = (uint16_t)(len[extra_idx]);
1732 				prev->next = rxm;
1733 				prev = rxm;
1734 				rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1735 				rx_pkts[nb_rx]->data_len += len[extra_idx];
1736 				extra_idx += 1;
1737 			};
1738 			seg_res -= rcv_cnt;
1739 
1740 			if (!seg_res) {
1741 				virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1742 				nb_rx++;
1743 			}
1744 		} else {
1745 			PMD_RX_LOG(ERR,
1746 					"No enough segments for packet.");
1747 			virtio_discard_rxbuf(vq, prev);
1748 			rxvq->stats.errors++;
1749 			break;
1750 		}
1751 	}
1752 
1753 	rxvq->stats.packets += nb_rx;
1754 
1755 	/* Allocate new mbuf for the used descriptor */
1756 	if (likely(!virtqueue_full(vq))) {
1757 		/* free_cnt may include mrg descs */
1758 		uint16_t free_cnt = vq->vq_free_cnt;
1759 		struct rte_mbuf *new_pkts[free_cnt];
1760 
1761 		if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1762 			error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1763 					free_cnt);
1764 			if (unlikely(error)) {
1765 				for (i = 0; i < free_cnt; i++)
1766 					rte_pktmbuf_free(new_pkts[i]);
1767 			}
1768 			nb_enqueued += free_cnt;
1769 		} else {
1770 			struct rte_eth_dev *dev =
1771 				&rte_eth_devices[rxvq->port_id];
1772 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1773 		}
1774 	}
1775 
1776 	if (likely(nb_enqueued)) {
1777 		vq_update_avail_idx(vq);
1778 
1779 		if (unlikely(virtqueue_kick_prepare(vq))) {
1780 			virtqueue_notify(vq);
1781 			PMD_RX_LOG(DEBUG, "Notified");
1782 		}
1783 	}
1784 
1785 	return nb_rx;
1786 }
1787 
1788 uint16_t
1789 virtio_recv_mergeable_pkts_packed(void *rx_queue,
1790 			struct rte_mbuf **rx_pkts,
1791 			uint16_t nb_pkts)
1792 {
1793 	struct virtnet_rx *rxvq = rx_queue;
1794 	struct virtqueue *vq = rxvq->vq;
1795 	struct virtio_hw *hw = vq->hw;
1796 	struct rte_mbuf *rxm;
1797 	struct rte_mbuf *prev = NULL;
1798 	uint16_t num, nb_rx = 0;
1799 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1800 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1801 	uint32_t nb_enqueued = 0;
1802 	uint32_t seg_num = 0;
1803 	uint32_t seg_res = 0;
1804 	uint32_t hdr_size = hw->vtnet_hdr_size;
1805 	int32_t i;
1806 	int error;
1807 
1808 	if (unlikely(hw->started == 0))
1809 		return nb_rx;
1810 
1811 
1812 	num = nb_pkts;
1813 	if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1814 		num = VIRTIO_MBUF_BURST_SZ;
1815 	if (likely(num > DESC_PER_CACHELINE))
1816 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1817 
1818 	num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1819 
1820 	for (i = 0; i < num; i++) {
1821 		struct virtio_net_hdr_mrg_rxbuf *header;
1822 
1823 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1824 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1825 
1826 		rxm = rcv_pkts[i];
1827 
1828 		if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
1829 			PMD_RX_LOG(ERR, "Packet drop");
1830 			nb_enqueued++;
1831 			virtio_discard_rxbuf(vq, rxm);
1832 			rxvq->stats.errors++;
1833 			continue;
1834 		}
1835 
1836 		header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)
1837 			  rxm->buf_addr + RTE_PKTMBUF_HEADROOM - hdr_size);
1838 		seg_num = header->num_buffers;
1839 
1840 		if (seg_num == 0)
1841 			seg_num = 1;
1842 
1843 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1844 		rxm->nb_segs = seg_num;
1845 		rxm->ol_flags = 0;
1846 		rxm->vlan_tci = 0;
1847 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1848 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1849 
1850 		rxm->port = rxvq->port_id;
1851 		rx_pkts[nb_rx] = rxm;
1852 		prev = rxm;
1853 
1854 		if (hw->has_rx_offload &&
1855 				virtio_rx_offload(rxm, &header->hdr) < 0) {
1856 			virtio_discard_rxbuf(vq, rxm);
1857 			rxvq->stats.errors++;
1858 			continue;
1859 		}
1860 
1861 		if (hw->vlan_strip)
1862 			rte_vlan_strip(rx_pkts[nb_rx]);
1863 
1864 		seg_res = seg_num - 1;
1865 
1866 		/* Merge remaining segments */
1867 		while (seg_res != 0 && i < (num - 1)) {
1868 			i++;
1869 
1870 			rxm = rcv_pkts[i];
1871 			rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1872 			rxm->pkt_len = (uint32_t)(len[i]);
1873 			rxm->data_len = (uint16_t)(len[i]);
1874 
1875 			rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1876 			rx_pkts[nb_rx]->data_len += (uint16_t)(len[i]);
1877 
1878 			if (prev)
1879 				prev->next = rxm;
1880 
1881 			prev = rxm;
1882 			seg_res -= 1;
1883 		}
1884 
1885 		if (!seg_res) {
1886 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1887 			nb_rx++;
1888 		}
1889 	}
1890 
1891 	/* Last packet still need merge segments */
1892 	while (seg_res != 0) {
1893 		uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1894 					VIRTIO_MBUF_BURST_SZ);
1895 		if (likely(vq->vq_free_cnt >= rcv_cnt)) {
1896 			num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts,
1897 					len, rcv_cnt);
1898 			uint16_t extra_idx = 0;
1899 
1900 			rcv_cnt = num;
1901 
1902 			while (extra_idx < rcv_cnt) {
1903 				rxm = rcv_pkts[extra_idx];
1904 
1905 				rxm->data_off =
1906 					RTE_PKTMBUF_HEADROOM - hdr_size;
1907 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
1908 				rxm->data_len = (uint16_t)(len[extra_idx]);
1909 
1910 				prev->next = rxm;
1911 				prev = rxm;
1912 				rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1913 				rx_pkts[nb_rx]->data_len += len[extra_idx];
1914 				extra_idx += 1;
1915 			}
1916 			seg_res -= rcv_cnt;
1917 			if (!seg_res) {
1918 				virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1919 				nb_rx++;
1920 			}
1921 		} else {
1922 			PMD_RX_LOG(ERR,
1923 					"No enough segments for packet.");
1924 			if (prev)
1925 				virtio_discard_rxbuf(vq, prev);
1926 			rxvq->stats.errors++;
1927 			break;
1928 		}
1929 	}
1930 
1931 	rxvq->stats.packets += nb_rx;
1932 
1933 	/* Allocate new mbuf for the used descriptor */
1934 	if (likely(!virtqueue_full(vq))) {
1935 		/* free_cnt may include mrg descs */
1936 		uint16_t free_cnt = vq->vq_free_cnt;
1937 		struct rte_mbuf *new_pkts[free_cnt];
1938 
1939 		if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1940 			error = virtqueue_enqueue_recv_refill_packed(vq,
1941 					new_pkts, free_cnt);
1942 			if (unlikely(error)) {
1943 				for (i = 0; i < free_cnt; i++)
1944 					rte_pktmbuf_free(new_pkts[i]);
1945 			}
1946 			nb_enqueued += free_cnt;
1947 		} else {
1948 			struct rte_eth_dev *dev =
1949 				&rte_eth_devices[rxvq->port_id];
1950 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1951 		}
1952 	}
1953 
1954 	if (likely(nb_enqueued)) {
1955 		if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1956 			virtqueue_notify(vq);
1957 			PMD_RX_LOG(DEBUG, "Notified");
1958 		}
1959 	}
1960 
1961 	return nb_rx;
1962 }
1963 
1964 uint16_t
1965 virtio_xmit_pkts_packed(void *tx_queue, struct rte_mbuf **tx_pkts,
1966 			uint16_t nb_pkts)
1967 {
1968 	struct virtnet_tx *txvq = tx_queue;
1969 	struct virtqueue *vq = txvq->vq;
1970 	struct virtio_hw *hw = vq->hw;
1971 	uint16_t hdr_size = hw->vtnet_hdr_size;
1972 	uint16_t nb_tx = 0;
1973 	bool in_order = hw->use_inorder_tx;
1974 	int error;
1975 
1976 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1977 		return nb_tx;
1978 
1979 	if (unlikely(nb_pkts < 1))
1980 		return nb_pkts;
1981 
1982 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1983 
1984 	if (nb_pkts > vq->vq_free_cnt)
1985 		virtio_xmit_cleanup_packed(vq, nb_pkts - vq->vq_free_cnt,
1986 					   in_order);
1987 
1988 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1989 		struct rte_mbuf *txm = tx_pkts[nb_tx];
1990 		int can_push = 0, slots, need;
1991 
1992 		/* Do VLAN tag insertion */
1993 		if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1994 			error = rte_vlan_insert(&txm);
1995 			if (unlikely(error)) {
1996 				rte_pktmbuf_free(txm);
1997 				continue;
1998 			}
1999 		}
2000 
2001 		/* optimize ring usage */
2002 		if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
2003 		      vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
2004 		    rte_mbuf_refcnt_read(txm) == 1 &&
2005 		    RTE_MBUF_DIRECT(txm) &&
2006 		    txm->nb_segs == 1 &&
2007 		    rte_pktmbuf_headroom(txm) >= hdr_size &&
2008 		    rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
2009 			   __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
2010 			can_push = 1;
2011 
2012 		/* How many main ring entries are needed to this Tx?
2013 		 * any_layout => number of segments
2014 		 * default    => number of segments + 1
2015 		 */
2016 		slots = txm->nb_segs + !can_push;
2017 		need = slots - vq->vq_free_cnt;
2018 
2019 		/* Positive value indicates it need free vring descriptors */
2020 		if (unlikely(need > 0)) {
2021 			virtio_xmit_cleanup_packed(vq, need, in_order);
2022 			need = slots - vq->vq_free_cnt;
2023 			if (unlikely(need > 0)) {
2024 				PMD_TX_LOG(ERR,
2025 					   "No free tx descriptors to transmit");
2026 				break;
2027 			}
2028 		}
2029 
2030 		/* Enqueue Packet buffers */
2031 		if (can_push)
2032 			virtqueue_enqueue_xmit_packed_fast(txvq, txm, in_order);
2033 		else
2034 			virtqueue_enqueue_xmit_packed(txvq, txm, slots, 0,
2035 						      in_order);
2036 
2037 		virtio_update_packet_stats(&txvq->stats, txm);
2038 	}
2039 
2040 	txvq->stats.packets += nb_tx;
2041 
2042 	if (likely(nb_tx)) {
2043 		if (unlikely(virtqueue_kick_prepare_packed(vq))) {
2044 			virtqueue_notify(vq);
2045 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2046 		}
2047 	}
2048 
2049 	return nb_tx;
2050 }
2051 
2052 uint16_t
2053 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
2054 {
2055 	struct virtnet_tx *txvq = tx_queue;
2056 	struct virtqueue *vq = txvq->vq;
2057 	struct virtio_hw *hw = vq->hw;
2058 	uint16_t hdr_size = hw->vtnet_hdr_size;
2059 	uint16_t nb_used, nb_tx = 0;
2060 	int error;
2061 
2062 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
2063 		return nb_tx;
2064 
2065 	if (unlikely(nb_pkts < 1))
2066 		return nb_pkts;
2067 
2068 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
2069 	nb_used = VIRTQUEUE_NUSED(vq);
2070 
2071 	virtio_rmb(hw->weak_barriers);
2072 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
2073 		virtio_xmit_cleanup(vq, nb_used);
2074 
2075 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
2076 		struct rte_mbuf *txm = tx_pkts[nb_tx];
2077 		int can_push = 0, use_indirect = 0, slots, need;
2078 
2079 		/* Do VLAN tag insertion */
2080 		if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
2081 			error = rte_vlan_insert(&txm);
2082 			if (unlikely(error)) {
2083 				rte_pktmbuf_free(txm);
2084 				continue;
2085 			}
2086 		}
2087 
2088 		/* optimize ring usage */
2089 		if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
2090 		      vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
2091 		    rte_mbuf_refcnt_read(txm) == 1 &&
2092 		    RTE_MBUF_DIRECT(txm) &&
2093 		    txm->nb_segs == 1 &&
2094 		    rte_pktmbuf_headroom(txm) >= hdr_size &&
2095 		    rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
2096 				   __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
2097 			can_push = 1;
2098 		else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
2099 			 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
2100 			use_indirect = 1;
2101 
2102 		/* How many main ring entries are needed to this Tx?
2103 		 * any_layout => number of segments
2104 		 * indirect   => 1
2105 		 * default    => number of segments + 1
2106 		 */
2107 		slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
2108 		need = slots - vq->vq_free_cnt;
2109 
2110 		/* Positive value indicates it need free vring descriptors */
2111 		if (unlikely(need > 0)) {
2112 			nb_used = VIRTQUEUE_NUSED(vq);
2113 			virtio_rmb(hw->weak_barriers);
2114 			need = RTE_MIN(need, (int)nb_used);
2115 
2116 			virtio_xmit_cleanup(vq, need);
2117 			need = slots - vq->vq_free_cnt;
2118 			if (unlikely(need > 0)) {
2119 				PMD_TX_LOG(ERR,
2120 					   "No free tx descriptors to transmit");
2121 				break;
2122 			}
2123 		}
2124 
2125 		/* Enqueue Packet buffers */
2126 		virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
2127 			can_push, 0);
2128 
2129 		virtio_update_packet_stats(&txvq->stats, txm);
2130 	}
2131 
2132 	txvq->stats.packets += nb_tx;
2133 
2134 	if (likely(nb_tx)) {
2135 		vq_update_avail_idx(vq);
2136 
2137 		if (unlikely(virtqueue_kick_prepare(vq))) {
2138 			virtqueue_notify(vq);
2139 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2140 		}
2141 	}
2142 
2143 	return nb_tx;
2144 }
2145 
2146 uint16_t
2147 virtio_xmit_pkts_inorder(void *tx_queue,
2148 			struct rte_mbuf **tx_pkts,
2149 			uint16_t nb_pkts)
2150 {
2151 	struct virtnet_tx *txvq = tx_queue;
2152 	struct virtqueue *vq = txvq->vq;
2153 	struct virtio_hw *hw = vq->hw;
2154 	uint16_t hdr_size = hw->vtnet_hdr_size;
2155 	uint16_t nb_used, nb_avail, nb_tx = 0, nb_inorder_pkts = 0;
2156 	struct rte_mbuf *inorder_pkts[nb_pkts];
2157 	int error;
2158 
2159 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
2160 		return nb_tx;
2161 
2162 	if (unlikely(nb_pkts < 1))
2163 		return nb_pkts;
2164 
2165 	VIRTQUEUE_DUMP(vq);
2166 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
2167 	nb_used = VIRTQUEUE_NUSED(vq);
2168 
2169 	virtio_rmb(hw->weak_barriers);
2170 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
2171 		virtio_xmit_cleanup_inorder(vq, nb_used);
2172 
2173 	if (unlikely(!vq->vq_free_cnt))
2174 		virtio_xmit_cleanup_inorder(vq, nb_used);
2175 
2176 	nb_avail = RTE_MIN(vq->vq_free_cnt, nb_pkts);
2177 
2178 	for (nb_tx = 0; nb_tx < nb_avail; nb_tx++) {
2179 		struct rte_mbuf *txm = tx_pkts[nb_tx];
2180 		int slots, need;
2181 
2182 		/* Do VLAN tag insertion */
2183 		if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
2184 			error = rte_vlan_insert(&txm);
2185 			if (unlikely(error)) {
2186 				rte_pktmbuf_free(txm);
2187 				continue;
2188 			}
2189 		}
2190 
2191 		/* optimize ring usage */
2192 		if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
2193 		     vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
2194 		     rte_mbuf_refcnt_read(txm) == 1 &&
2195 		     RTE_MBUF_DIRECT(txm) &&
2196 		     txm->nb_segs == 1 &&
2197 		     rte_pktmbuf_headroom(txm) >= hdr_size &&
2198 		     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
2199 				__alignof__(struct virtio_net_hdr_mrg_rxbuf))) {
2200 			inorder_pkts[nb_inorder_pkts] = txm;
2201 			nb_inorder_pkts++;
2202 
2203 			virtio_update_packet_stats(&txvq->stats, txm);
2204 			continue;
2205 		}
2206 
2207 		if (nb_inorder_pkts) {
2208 			virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2209 							nb_inorder_pkts);
2210 			nb_inorder_pkts = 0;
2211 		}
2212 
2213 		slots = txm->nb_segs + 1;
2214 		need = slots - vq->vq_free_cnt;
2215 		if (unlikely(need > 0)) {
2216 			nb_used = VIRTQUEUE_NUSED(vq);
2217 			virtio_rmb(hw->weak_barriers);
2218 			need = RTE_MIN(need, (int)nb_used);
2219 
2220 			virtio_xmit_cleanup_inorder(vq, need);
2221 
2222 			need = slots - vq->vq_free_cnt;
2223 
2224 			if (unlikely(need > 0)) {
2225 				PMD_TX_LOG(ERR,
2226 					"No free tx descriptors to transmit");
2227 				break;
2228 			}
2229 		}
2230 		/* Enqueue Packet buffers */
2231 		virtqueue_enqueue_xmit(txvq, txm, slots, 0, 0, 1);
2232 
2233 		virtio_update_packet_stats(&txvq->stats, txm);
2234 	}
2235 
2236 	/* Transmit all inorder packets */
2237 	if (nb_inorder_pkts)
2238 		virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2239 						nb_inorder_pkts);
2240 
2241 	txvq->stats.packets += nb_tx;
2242 
2243 	if (likely(nb_tx)) {
2244 		vq_update_avail_idx(vq);
2245 
2246 		if (unlikely(virtqueue_kick_prepare(vq))) {
2247 			virtqueue_notify(vq);
2248 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2249 		}
2250 	}
2251 
2252 	VIRTQUEUE_DUMP(vq);
2253 
2254 	return nb_tx;
2255 }
2256