xref: /dpdk/drivers/net/virtio/virtio_rxtx.c (revision 200bc52e5aa0d72e70464c9cd22b55cf536ed13c)
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 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->vq_packed.ring.desc;
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->vq_packed.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_split.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_split.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->vq_packed.ring.desc;
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->vq_packed.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->vq_packed.ring.desc;
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->vq_packed.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_split.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_split.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_split.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->vq_packed.ring.desc;
434 	uint16_t flags = vq->vq_packed.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->vq_packed.cached_flags ^=
464 				VRING_PACKED_DESC_F_AVAIL_USED;
465 			flags = vq->vq_packed.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 rte_ipv4_hdr *iph;
483 		struct rte_ipv6_hdr *ip6h;
484 		struct rte_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,
489 					struct rte_ipv4_hdr *, m->l2_len);
490 		th = RTE_PTR_ADD(iph, m->l3_len);
491 		if ((iph->version_ihl >> 4) == 4) {
492 			iph->hdr_checksum = 0;
493 			iph->hdr_checksum = rte_ipv4_cksum(iph);
494 			ip_len = iph->total_length;
495 			ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
496 				m->l3_len);
497 		} else {
498 			ip6h = (struct rte_ipv6_hdr *)iph;
499 			ip_paylen = ip6h->payload_len;
500 		}
501 
502 		/* calculate the new phdr checksum not including ip_paylen */
503 		prev_cksum = th->cksum;
504 		tmp = prev_cksum;
505 		tmp += ip_paylen;
506 		tmp = (tmp & 0xffff) + (tmp >> 16);
507 		new_cksum = tmp;
508 
509 		/* replace it in the packet */
510 		th->cksum = new_cksum;
511 	}
512 }
513 
514 
515 /* avoid write operation when necessary, to lessen cache issues */
516 #define ASSIGN_UNLESS_EQUAL(var, val) do {	\
517 	if ((var) != (val))			\
518 		(var) = (val);			\
519 } while (0)
520 
521 #define virtqueue_clear_net_hdr(_hdr) do {		\
522 	ASSIGN_UNLESS_EQUAL((_hdr)->csum_start, 0);	\
523 	ASSIGN_UNLESS_EQUAL((_hdr)->csum_offset, 0);	\
524 	ASSIGN_UNLESS_EQUAL((_hdr)->flags, 0);		\
525 	ASSIGN_UNLESS_EQUAL((_hdr)->gso_type, 0);	\
526 	ASSIGN_UNLESS_EQUAL((_hdr)->gso_size, 0);	\
527 	ASSIGN_UNLESS_EQUAL((_hdr)->hdr_len, 0);	\
528 } while (0)
529 
530 static inline void
531 virtqueue_xmit_offload(struct virtio_net_hdr *hdr,
532 			struct rte_mbuf *cookie,
533 			bool offload)
534 {
535 	if (offload) {
536 		if (cookie->ol_flags & PKT_TX_TCP_SEG)
537 			cookie->ol_flags |= PKT_TX_TCP_CKSUM;
538 
539 		switch (cookie->ol_flags & PKT_TX_L4_MASK) {
540 		case PKT_TX_UDP_CKSUM:
541 			hdr->csum_start = cookie->l2_len + cookie->l3_len;
542 			hdr->csum_offset = offsetof(struct rte_udp_hdr,
543 				dgram_cksum);
544 			hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
545 			break;
546 
547 		case PKT_TX_TCP_CKSUM:
548 			hdr->csum_start = cookie->l2_len + cookie->l3_len;
549 			hdr->csum_offset = offsetof(struct rte_tcp_hdr, cksum);
550 			hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
551 			break;
552 
553 		default:
554 			ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
555 			ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
556 			ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
557 			break;
558 		}
559 
560 		/* TCP Segmentation Offload */
561 		if (cookie->ol_flags & PKT_TX_TCP_SEG) {
562 			virtio_tso_fix_cksum(cookie);
563 			hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
564 				VIRTIO_NET_HDR_GSO_TCPV6 :
565 				VIRTIO_NET_HDR_GSO_TCPV4;
566 			hdr->gso_size = cookie->tso_segsz;
567 			hdr->hdr_len =
568 				cookie->l2_len +
569 				cookie->l3_len +
570 				cookie->l4_len;
571 		} else {
572 			ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
573 			ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
574 			ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
575 		}
576 	}
577 }
578 
579 static inline void
580 virtqueue_enqueue_xmit_inorder(struct virtnet_tx *txvq,
581 			struct rte_mbuf **cookies,
582 			uint16_t num)
583 {
584 	struct vq_desc_extra *dxp;
585 	struct virtqueue *vq = txvq->vq;
586 	struct vring_desc *start_dp;
587 	struct virtio_net_hdr *hdr;
588 	uint16_t idx;
589 	uint16_t head_size = vq->hw->vtnet_hdr_size;
590 	uint16_t i = 0;
591 
592 	idx = vq->vq_desc_head_idx;
593 	start_dp = vq->vq_split.ring.desc;
594 
595 	while (i < num) {
596 		idx = idx & (vq->vq_nentries - 1);
597 		dxp = &vq->vq_descx[vq->vq_avail_idx & (vq->vq_nentries - 1)];
598 		dxp->cookie = (void *)cookies[i];
599 		dxp->ndescs = 1;
600 
601 		hdr = (struct virtio_net_hdr *)
602 			rte_pktmbuf_prepend(cookies[i], head_size);
603 		cookies[i]->pkt_len -= head_size;
604 
605 		/* if offload disabled, hdr is not zeroed yet, do it now */
606 		if (!vq->hw->has_tx_offload)
607 			virtqueue_clear_net_hdr(hdr);
608 		else
609 			virtqueue_xmit_offload(hdr, cookies[i], true);
610 
611 		start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookies[i], vq);
612 		start_dp[idx].len   = cookies[i]->data_len;
613 		start_dp[idx].flags = 0;
614 
615 		vq_update_avail_ring(vq, idx);
616 
617 		idx++;
618 		i++;
619 	};
620 
621 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
622 	vq->vq_desc_head_idx = idx & (vq->vq_nentries - 1);
623 }
624 
625 static inline void
626 virtqueue_enqueue_xmit_packed_fast(struct virtnet_tx *txvq,
627 				   struct rte_mbuf *cookie,
628 				   int in_order)
629 {
630 	struct virtqueue *vq = txvq->vq;
631 	struct vring_packed_desc *dp;
632 	struct vq_desc_extra *dxp;
633 	uint16_t idx, id, flags;
634 	uint16_t head_size = vq->hw->vtnet_hdr_size;
635 	struct virtio_net_hdr *hdr;
636 
637 	id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx;
638 	idx = vq->vq_avail_idx;
639 	dp = &vq->vq_packed.ring.desc[idx];
640 
641 	dxp = &vq->vq_descx[id];
642 	dxp->ndescs = 1;
643 	dxp->cookie = cookie;
644 
645 	flags = vq->vq_packed.cached_flags;
646 
647 	/* prepend cannot fail, checked by caller */
648 	hdr = (struct virtio_net_hdr *)
649 		rte_pktmbuf_prepend(cookie, head_size);
650 	cookie->pkt_len -= head_size;
651 
652 	/* if offload disabled, hdr is not zeroed yet, do it now */
653 	if (!vq->hw->has_tx_offload)
654 		virtqueue_clear_net_hdr(hdr);
655 	else
656 		virtqueue_xmit_offload(hdr, cookie, true);
657 
658 	dp->addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
659 	dp->len  = cookie->data_len;
660 	dp->id   = id;
661 
662 	if (++vq->vq_avail_idx >= vq->vq_nentries) {
663 		vq->vq_avail_idx -= vq->vq_nentries;
664 		vq->vq_packed.cached_flags ^= VRING_PACKED_DESC_F_AVAIL_USED;
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->vq_packed.ring.desc;
702 
703 	head_dp = &vq->vq_packed.ring.desc[idx];
704 	head_flags = cookie->next ? VRING_DESC_F_NEXT : 0;
705 	head_flags |= vq->vq_packed.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->vq_packed.cached_flags ^=
731 				VRING_PACKED_DESC_F_AVAIL_USED;
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->vq_packed.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->vq_packed.cached_flags ^=
752 				VRING_PACKED_DESC_F_AVAIL_USED;
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_split.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_split.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_split.ring.avail->ring[desc_idx] = desc_idx;
923 			vq->vq_split.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_split.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 rte_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
1114 			stats->size_bins[7]++;
1115 	}
1116 
1117 	ea = rte_pktmbuf_mtod(mbuf, struct rte_ether_addr *);
1118 	if (rte_is_multicast_ether_addr(ea)) {
1119 		if (rte_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;
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 + RTE_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 	if (likely(!virtqueue_full(vq))) {
1285 		uint16_t free_cnt = vq->vq_free_cnt;
1286 		struct rte_mbuf *new_pkts[free_cnt];
1287 
1288 		if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1289 						free_cnt) == 0)) {
1290 			error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1291 					free_cnt);
1292 			if (unlikely(error)) {
1293 				for (i = 0; i < free_cnt; i++)
1294 					rte_pktmbuf_free(new_pkts[i]);
1295 			}
1296 			nb_enqueued += free_cnt;
1297 		} else {
1298 			struct rte_eth_dev *dev =
1299 				&rte_eth_devices[rxvq->port_id];
1300 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1301 		}
1302 	}
1303 
1304 	if (likely(nb_enqueued)) {
1305 		vq_update_avail_idx(vq);
1306 
1307 		if (unlikely(virtqueue_kick_prepare(vq))) {
1308 			virtqueue_notify(vq);
1309 			PMD_RX_LOG(DEBUG, "Notified");
1310 		}
1311 	}
1312 
1313 	return nb_rx;
1314 }
1315 
1316 uint16_t
1317 virtio_recv_pkts_packed(void *rx_queue, struct rte_mbuf **rx_pkts,
1318 			uint16_t nb_pkts)
1319 {
1320 	struct virtnet_rx *rxvq = rx_queue;
1321 	struct virtqueue *vq = rxvq->vq;
1322 	struct virtio_hw *hw = vq->hw;
1323 	struct rte_mbuf *rxm;
1324 	uint16_t num, nb_rx;
1325 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1326 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1327 	int error;
1328 	uint32_t i, nb_enqueued;
1329 	uint32_t hdr_size;
1330 	struct virtio_net_hdr *hdr;
1331 
1332 	nb_rx = 0;
1333 	if (unlikely(hw->started == 0))
1334 		return nb_rx;
1335 
1336 	num = RTE_MIN(VIRTIO_MBUF_BURST_SZ, nb_pkts);
1337 	if (likely(num > DESC_PER_CACHELINE))
1338 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1339 
1340 	num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1341 	PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1342 
1343 	nb_enqueued = 0;
1344 	hdr_size = hw->vtnet_hdr_size;
1345 
1346 	for (i = 0; i < num; i++) {
1347 		rxm = rcv_pkts[i];
1348 
1349 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1350 
1351 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1352 			PMD_RX_LOG(ERR, "Packet drop");
1353 			nb_enqueued++;
1354 			virtio_discard_rxbuf(vq, rxm);
1355 			rxvq->stats.errors++;
1356 			continue;
1357 		}
1358 
1359 		rxm->port = rxvq->port_id;
1360 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1361 		rxm->ol_flags = 0;
1362 		rxm->vlan_tci = 0;
1363 
1364 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1365 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1366 
1367 		hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
1368 			RTE_PKTMBUF_HEADROOM - hdr_size);
1369 
1370 		if (hw->vlan_strip)
1371 			rte_vlan_strip(rxm);
1372 
1373 		if (hw->has_rx_offload && virtio_rx_offload(rxm, hdr) < 0) {
1374 			virtio_discard_rxbuf(vq, rxm);
1375 			rxvq->stats.errors++;
1376 			continue;
1377 		}
1378 
1379 		virtio_rx_stats_updated(rxvq, rxm);
1380 
1381 		rx_pkts[nb_rx++] = rxm;
1382 	}
1383 
1384 	rxvq->stats.packets += nb_rx;
1385 
1386 	/* Allocate new mbuf for the used descriptor */
1387 	if (likely(!virtqueue_full(vq))) {
1388 		uint16_t free_cnt = vq->vq_free_cnt;
1389 		struct rte_mbuf *new_pkts[free_cnt];
1390 
1391 		if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
1392 						free_cnt) == 0)) {
1393 			error = virtqueue_enqueue_recv_refill_packed(vq,
1394 					new_pkts, free_cnt);
1395 			if (unlikely(error)) {
1396 				for (i = 0; i < free_cnt; i++)
1397 					rte_pktmbuf_free(new_pkts[i]);
1398 			}
1399 			nb_enqueued += free_cnt;
1400 		} else {
1401 			struct rte_eth_dev *dev =
1402 				&rte_eth_devices[rxvq->port_id];
1403 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1404 		}
1405 	}
1406 
1407 	if (likely(nb_enqueued)) {
1408 		if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1409 			virtqueue_notify(vq);
1410 			PMD_RX_LOG(DEBUG, "Notified");
1411 		}
1412 	}
1413 
1414 	return nb_rx;
1415 }
1416 
1417 
1418 uint16_t
1419 virtio_recv_pkts_inorder(void *rx_queue,
1420 			struct rte_mbuf **rx_pkts,
1421 			uint16_t nb_pkts)
1422 {
1423 	struct virtnet_rx *rxvq = rx_queue;
1424 	struct virtqueue *vq = rxvq->vq;
1425 	struct virtio_hw *hw = vq->hw;
1426 	struct rte_mbuf *rxm;
1427 	struct rte_mbuf *prev;
1428 	uint16_t nb_used, num, nb_rx;
1429 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1430 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1431 	int error;
1432 	uint32_t nb_enqueued;
1433 	uint32_t seg_num;
1434 	uint32_t seg_res;
1435 	uint32_t hdr_size;
1436 	int32_t i;
1437 
1438 	nb_rx = 0;
1439 	if (unlikely(hw->started == 0))
1440 		return nb_rx;
1441 
1442 	nb_used = VIRTQUEUE_NUSED(vq);
1443 	nb_used = RTE_MIN(nb_used, nb_pkts);
1444 	nb_used = RTE_MIN(nb_used, VIRTIO_MBUF_BURST_SZ);
1445 
1446 	virtio_rmb(hw->weak_barriers);
1447 
1448 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1449 
1450 	nb_enqueued = 0;
1451 	seg_num = 1;
1452 	seg_res = 0;
1453 	hdr_size = hw->vtnet_hdr_size;
1454 
1455 	num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len, nb_used);
1456 
1457 	for (i = 0; i < num; i++) {
1458 		struct virtio_net_hdr_mrg_rxbuf *header;
1459 
1460 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1461 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1462 
1463 		rxm = rcv_pkts[i];
1464 
1465 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1466 			PMD_RX_LOG(ERR, "Packet drop");
1467 			nb_enqueued++;
1468 			virtio_discard_rxbuf_inorder(vq, rxm);
1469 			rxvq->stats.errors++;
1470 			continue;
1471 		}
1472 
1473 		header = (struct virtio_net_hdr_mrg_rxbuf *)
1474 			 ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1475 			 - hdr_size);
1476 
1477 		if (vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
1478 			seg_num = header->num_buffers;
1479 			if (seg_num == 0)
1480 				seg_num = 1;
1481 		} else {
1482 			seg_num = 1;
1483 		}
1484 
1485 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1486 		rxm->nb_segs = seg_num;
1487 		rxm->ol_flags = 0;
1488 		rxm->vlan_tci = 0;
1489 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1490 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1491 
1492 		rxm->port = rxvq->port_id;
1493 
1494 		rx_pkts[nb_rx] = rxm;
1495 		prev = rxm;
1496 
1497 		if (vq->hw->has_rx_offload &&
1498 				virtio_rx_offload(rxm, &header->hdr) < 0) {
1499 			virtio_discard_rxbuf_inorder(vq, rxm);
1500 			rxvq->stats.errors++;
1501 			continue;
1502 		}
1503 
1504 		if (hw->vlan_strip)
1505 			rte_vlan_strip(rx_pkts[nb_rx]);
1506 
1507 		seg_res = seg_num - 1;
1508 
1509 		/* Merge remaining segments */
1510 		while (seg_res != 0 && i < (num - 1)) {
1511 			i++;
1512 
1513 			rxm = rcv_pkts[i];
1514 			rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1515 			rxm->pkt_len = (uint32_t)(len[i]);
1516 			rxm->data_len = (uint16_t)(len[i]);
1517 
1518 			rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1519 			rx_pkts[nb_rx]->data_len += (uint16_t)(len[i]);
1520 
1521 			if (prev)
1522 				prev->next = rxm;
1523 
1524 			prev = rxm;
1525 			seg_res -= 1;
1526 		}
1527 
1528 		if (!seg_res) {
1529 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1530 			nb_rx++;
1531 		}
1532 	}
1533 
1534 	/* Last packet still need merge segments */
1535 	while (seg_res != 0) {
1536 		uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1537 					VIRTIO_MBUF_BURST_SZ);
1538 
1539 		prev = rcv_pkts[nb_rx];
1540 		if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
1541 			virtio_rmb(hw->weak_barriers);
1542 			num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len,
1543 							   rcv_cnt);
1544 			uint16_t extra_idx = 0;
1545 
1546 			rcv_cnt = num;
1547 			while (extra_idx < rcv_cnt) {
1548 				rxm = rcv_pkts[extra_idx];
1549 				rxm->data_off =
1550 					RTE_PKTMBUF_HEADROOM - hdr_size;
1551 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
1552 				rxm->data_len = (uint16_t)(len[extra_idx]);
1553 				prev->next = rxm;
1554 				prev = rxm;
1555 				rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1556 				rx_pkts[nb_rx]->data_len += len[extra_idx];
1557 				extra_idx += 1;
1558 			};
1559 			seg_res -= rcv_cnt;
1560 
1561 			if (!seg_res) {
1562 				virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1563 				nb_rx++;
1564 			}
1565 		} else {
1566 			PMD_RX_LOG(ERR,
1567 					"No enough segments for packet.");
1568 			virtio_discard_rxbuf_inorder(vq, prev);
1569 			rxvq->stats.errors++;
1570 			break;
1571 		}
1572 	}
1573 
1574 	rxvq->stats.packets += nb_rx;
1575 
1576 	/* Allocate new mbuf for the used descriptor */
1577 
1578 	if (likely(!virtqueue_full(vq))) {
1579 		/* free_cnt may include mrg descs */
1580 		uint16_t free_cnt = vq->vq_free_cnt;
1581 		struct rte_mbuf *new_pkts[free_cnt];
1582 
1583 		if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1584 			error = virtqueue_enqueue_refill_inorder(vq, new_pkts,
1585 					free_cnt);
1586 			if (unlikely(error)) {
1587 				for (i = 0; i < free_cnt; i++)
1588 					rte_pktmbuf_free(new_pkts[i]);
1589 			}
1590 			nb_enqueued += free_cnt;
1591 		} else {
1592 			struct rte_eth_dev *dev =
1593 				&rte_eth_devices[rxvq->port_id];
1594 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1595 		}
1596 	}
1597 
1598 	if (likely(nb_enqueued)) {
1599 		vq_update_avail_idx(vq);
1600 
1601 		if (unlikely(virtqueue_kick_prepare(vq))) {
1602 			virtqueue_notify(vq);
1603 			PMD_RX_LOG(DEBUG, "Notified");
1604 		}
1605 	}
1606 
1607 	return nb_rx;
1608 }
1609 
1610 uint16_t
1611 virtio_recv_mergeable_pkts(void *rx_queue,
1612 			struct rte_mbuf **rx_pkts,
1613 			uint16_t nb_pkts)
1614 {
1615 	struct virtnet_rx *rxvq = rx_queue;
1616 	struct virtqueue *vq = rxvq->vq;
1617 	struct virtio_hw *hw = vq->hw;
1618 	struct rte_mbuf *rxm;
1619 	struct rte_mbuf *prev;
1620 	uint16_t nb_used, num, nb_rx = 0;
1621 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1622 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1623 	int error;
1624 	uint32_t nb_enqueued = 0;
1625 	uint32_t seg_num = 0;
1626 	uint32_t seg_res = 0;
1627 	uint32_t hdr_size = hw->vtnet_hdr_size;
1628 	int32_t i;
1629 
1630 	if (unlikely(hw->started == 0))
1631 		return nb_rx;
1632 
1633 	nb_used = VIRTQUEUE_NUSED(vq);
1634 
1635 	virtio_rmb(hw->weak_barriers);
1636 
1637 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1638 
1639 	num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
1640 	if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1641 		num = VIRTIO_MBUF_BURST_SZ;
1642 	if (likely(num > DESC_PER_CACHELINE))
1643 		num = num - ((vq->vq_used_cons_idx + num) %
1644 				DESC_PER_CACHELINE);
1645 
1646 
1647 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
1648 
1649 	for (i = 0; i < num; i++) {
1650 		struct virtio_net_hdr_mrg_rxbuf *header;
1651 
1652 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1653 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1654 
1655 		rxm = rcv_pkts[i];
1656 
1657 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1658 			PMD_RX_LOG(ERR, "Packet drop");
1659 			nb_enqueued++;
1660 			virtio_discard_rxbuf(vq, rxm);
1661 			rxvq->stats.errors++;
1662 			continue;
1663 		}
1664 
1665 		header = (struct virtio_net_hdr_mrg_rxbuf *)
1666 			 ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1667 			 - hdr_size);
1668 		seg_num = header->num_buffers;
1669 		if (seg_num == 0)
1670 			seg_num = 1;
1671 
1672 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1673 		rxm->nb_segs = seg_num;
1674 		rxm->ol_flags = 0;
1675 		rxm->vlan_tci = 0;
1676 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1677 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1678 
1679 		rxm->port = rxvq->port_id;
1680 
1681 		rx_pkts[nb_rx] = rxm;
1682 		prev = rxm;
1683 
1684 		if (hw->has_rx_offload &&
1685 				virtio_rx_offload(rxm, &header->hdr) < 0) {
1686 			virtio_discard_rxbuf(vq, rxm);
1687 			rxvq->stats.errors++;
1688 			continue;
1689 		}
1690 
1691 		if (hw->vlan_strip)
1692 			rte_vlan_strip(rx_pkts[nb_rx]);
1693 
1694 		seg_res = seg_num - 1;
1695 
1696 		/* Merge remaining segments */
1697 		while (seg_res != 0 && i < (num - 1)) {
1698 			i++;
1699 
1700 			rxm = rcv_pkts[i];
1701 			rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1702 			rxm->pkt_len = (uint32_t)(len[i]);
1703 			rxm->data_len = (uint16_t)(len[i]);
1704 
1705 			rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1706 			rx_pkts[nb_rx]->data_len += (uint16_t)(len[i]);
1707 
1708 			if (prev)
1709 				prev->next = rxm;
1710 
1711 			prev = rxm;
1712 			seg_res -= 1;
1713 		}
1714 
1715 		if (!seg_res) {
1716 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1717 			nb_rx++;
1718 		}
1719 	}
1720 
1721 	/* Last packet still need merge segments */
1722 	while (seg_res != 0) {
1723 		uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1724 					VIRTIO_MBUF_BURST_SZ);
1725 
1726 		prev = rcv_pkts[nb_rx];
1727 		if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
1728 			virtio_rmb(hw->weak_barriers);
1729 			num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len,
1730 							   rcv_cnt);
1731 			uint16_t extra_idx = 0;
1732 
1733 			rcv_cnt = num;
1734 			while (extra_idx < rcv_cnt) {
1735 				rxm = rcv_pkts[extra_idx];
1736 				rxm->data_off =
1737 					RTE_PKTMBUF_HEADROOM - hdr_size;
1738 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
1739 				rxm->data_len = (uint16_t)(len[extra_idx]);
1740 				prev->next = rxm;
1741 				prev = rxm;
1742 				rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1743 				rx_pkts[nb_rx]->data_len += len[extra_idx];
1744 				extra_idx += 1;
1745 			};
1746 			seg_res -= rcv_cnt;
1747 
1748 			if (!seg_res) {
1749 				virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1750 				nb_rx++;
1751 			}
1752 		} else {
1753 			PMD_RX_LOG(ERR,
1754 					"No enough segments for packet.");
1755 			virtio_discard_rxbuf(vq, prev);
1756 			rxvq->stats.errors++;
1757 			break;
1758 		}
1759 	}
1760 
1761 	rxvq->stats.packets += nb_rx;
1762 
1763 	/* Allocate new mbuf for the used descriptor */
1764 	if (likely(!virtqueue_full(vq))) {
1765 		/* free_cnt may include mrg descs */
1766 		uint16_t free_cnt = vq->vq_free_cnt;
1767 		struct rte_mbuf *new_pkts[free_cnt];
1768 
1769 		if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1770 			error = virtqueue_enqueue_recv_refill(vq, new_pkts,
1771 					free_cnt);
1772 			if (unlikely(error)) {
1773 				for (i = 0; i < free_cnt; i++)
1774 					rte_pktmbuf_free(new_pkts[i]);
1775 			}
1776 			nb_enqueued += free_cnt;
1777 		} else {
1778 			struct rte_eth_dev *dev =
1779 				&rte_eth_devices[rxvq->port_id];
1780 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1781 		}
1782 	}
1783 
1784 	if (likely(nb_enqueued)) {
1785 		vq_update_avail_idx(vq);
1786 
1787 		if (unlikely(virtqueue_kick_prepare(vq))) {
1788 			virtqueue_notify(vq);
1789 			PMD_RX_LOG(DEBUG, "Notified");
1790 		}
1791 	}
1792 
1793 	return nb_rx;
1794 }
1795 
1796 uint16_t
1797 virtio_recv_mergeable_pkts_packed(void *rx_queue,
1798 			struct rte_mbuf **rx_pkts,
1799 			uint16_t nb_pkts)
1800 {
1801 	struct virtnet_rx *rxvq = rx_queue;
1802 	struct virtqueue *vq = rxvq->vq;
1803 	struct virtio_hw *hw = vq->hw;
1804 	struct rte_mbuf *rxm;
1805 	struct rte_mbuf *prev = NULL;
1806 	uint16_t num, nb_rx = 0;
1807 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
1808 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1809 	uint32_t nb_enqueued = 0;
1810 	uint32_t seg_num = 0;
1811 	uint32_t seg_res = 0;
1812 	uint32_t hdr_size = hw->vtnet_hdr_size;
1813 	int32_t i;
1814 	int error;
1815 
1816 	if (unlikely(hw->started == 0))
1817 		return nb_rx;
1818 
1819 
1820 	num = nb_pkts;
1821 	if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
1822 		num = VIRTIO_MBUF_BURST_SZ;
1823 	if (likely(num > DESC_PER_CACHELINE))
1824 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
1825 
1826 	num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts, len, num);
1827 
1828 	for (i = 0; i < num; i++) {
1829 		struct virtio_net_hdr_mrg_rxbuf *header;
1830 
1831 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1832 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1833 
1834 		rxm = rcv_pkts[i];
1835 
1836 		if (unlikely(len[i] < hdr_size + RTE_ETHER_HDR_LEN)) {
1837 			PMD_RX_LOG(ERR, "Packet drop");
1838 			nb_enqueued++;
1839 			virtio_discard_rxbuf(vq, rxm);
1840 			rxvq->stats.errors++;
1841 			continue;
1842 		}
1843 
1844 		header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)
1845 			  rxm->buf_addr + RTE_PKTMBUF_HEADROOM - hdr_size);
1846 		seg_num = header->num_buffers;
1847 
1848 		if (seg_num == 0)
1849 			seg_num = 1;
1850 
1851 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1852 		rxm->nb_segs = seg_num;
1853 		rxm->ol_flags = 0;
1854 		rxm->vlan_tci = 0;
1855 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1856 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
1857 
1858 		rxm->port = rxvq->port_id;
1859 		rx_pkts[nb_rx] = rxm;
1860 		prev = rxm;
1861 
1862 		if (hw->has_rx_offload &&
1863 				virtio_rx_offload(rxm, &header->hdr) < 0) {
1864 			virtio_discard_rxbuf(vq, rxm);
1865 			rxvq->stats.errors++;
1866 			continue;
1867 		}
1868 
1869 		if (hw->vlan_strip)
1870 			rte_vlan_strip(rx_pkts[nb_rx]);
1871 
1872 		seg_res = seg_num - 1;
1873 
1874 		/* Merge remaining segments */
1875 		while (seg_res != 0 && i < (num - 1)) {
1876 			i++;
1877 
1878 			rxm = rcv_pkts[i];
1879 			rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1880 			rxm->pkt_len = (uint32_t)(len[i]);
1881 			rxm->data_len = (uint16_t)(len[i]);
1882 
1883 			rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1884 			rx_pkts[nb_rx]->data_len += (uint16_t)(len[i]);
1885 
1886 			if (prev)
1887 				prev->next = rxm;
1888 
1889 			prev = rxm;
1890 			seg_res -= 1;
1891 		}
1892 
1893 		if (!seg_res) {
1894 			virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1895 			nb_rx++;
1896 		}
1897 	}
1898 
1899 	/* Last packet still need merge segments */
1900 	while (seg_res != 0) {
1901 		uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1902 					VIRTIO_MBUF_BURST_SZ);
1903 		if (likely(vq->vq_free_cnt >= rcv_cnt)) {
1904 			num = virtqueue_dequeue_burst_rx_packed(vq, rcv_pkts,
1905 					len, rcv_cnt);
1906 			uint16_t extra_idx = 0;
1907 
1908 			rcv_cnt = num;
1909 
1910 			while (extra_idx < rcv_cnt) {
1911 				rxm = rcv_pkts[extra_idx];
1912 
1913 				rxm->data_off =
1914 					RTE_PKTMBUF_HEADROOM - hdr_size;
1915 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
1916 				rxm->data_len = (uint16_t)(len[extra_idx]);
1917 
1918 				prev->next = rxm;
1919 				prev = rxm;
1920 				rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1921 				rx_pkts[nb_rx]->data_len += len[extra_idx];
1922 				extra_idx += 1;
1923 			}
1924 			seg_res -= rcv_cnt;
1925 			if (!seg_res) {
1926 				virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1927 				nb_rx++;
1928 			}
1929 		} else {
1930 			PMD_RX_LOG(ERR,
1931 					"No enough segments for packet.");
1932 			if (prev)
1933 				virtio_discard_rxbuf(vq, prev);
1934 			rxvq->stats.errors++;
1935 			break;
1936 		}
1937 	}
1938 
1939 	rxvq->stats.packets += nb_rx;
1940 
1941 	/* Allocate new mbuf for the used descriptor */
1942 	if (likely(!virtqueue_full(vq))) {
1943 		/* free_cnt may include mrg descs */
1944 		uint16_t free_cnt = vq->vq_free_cnt;
1945 		struct rte_mbuf *new_pkts[free_cnt];
1946 
1947 		if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1948 			error = virtqueue_enqueue_recv_refill_packed(vq,
1949 					new_pkts, free_cnt);
1950 			if (unlikely(error)) {
1951 				for (i = 0; i < free_cnt; i++)
1952 					rte_pktmbuf_free(new_pkts[i]);
1953 			}
1954 			nb_enqueued += free_cnt;
1955 		} else {
1956 			struct rte_eth_dev *dev =
1957 				&rte_eth_devices[rxvq->port_id];
1958 			dev->data->rx_mbuf_alloc_failed += free_cnt;
1959 		}
1960 	}
1961 
1962 	if (likely(nb_enqueued)) {
1963 		if (unlikely(virtqueue_kick_prepare_packed(vq))) {
1964 			virtqueue_notify(vq);
1965 			PMD_RX_LOG(DEBUG, "Notified");
1966 		}
1967 	}
1968 
1969 	return nb_rx;
1970 }
1971 
1972 uint16_t
1973 virtio_xmit_pkts_packed(void *tx_queue, struct rte_mbuf **tx_pkts,
1974 			uint16_t nb_pkts)
1975 {
1976 	struct virtnet_tx *txvq = tx_queue;
1977 	struct virtqueue *vq = txvq->vq;
1978 	struct virtio_hw *hw = vq->hw;
1979 	uint16_t hdr_size = hw->vtnet_hdr_size;
1980 	uint16_t nb_tx = 0;
1981 	bool in_order = hw->use_inorder_tx;
1982 	int error;
1983 
1984 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1985 		return nb_tx;
1986 
1987 	if (unlikely(nb_pkts < 1))
1988 		return nb_pkts;
1989 
1990 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1991 
1992 	if (nb_pkts > vq->vq_free_cnt)
1993 		virtio_xmit_cleanup_packed(vq, nb_pkts - vq->vq_free_cnt,
1994 					   in_order);
1995 
1996 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1997 		struct rte_mbuf *txm = tx_pkts[nb_tx];
1998 		int can_push = 0, slots, need;
1999 
2000 		/* Do VLAN tag insertion */
2001 		if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
2002 			error = rte_vlan_insert(&txm);
2003 			if (unlikely(error)) {
2004 				rte_pktmbuf_free(txm);
2005 				continue;
2006 			}
2007 			/* vlan_insert may add a header mbuf */
2008 			tx_pkts[nb_tx] = txm;
2009 		}
2010 
2011 		/* optimize ring usage */
2012 		if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
2013 		      vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
2014 		    rte_mbuf_refcnt_read(txm) == 1 &&
2015 		    RTE_MBUF_DIRECT(txm) &&
2016 		    txm->nb_segs == 1 &&
2017 		    rte_pktmbuf_headroom(txm) >= hdr_size &&
2018 		    rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
2019 			   __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
2020 			can_push = 1;
2021 
2022 		/* How many main ring entries are needed to this Tx?
2023 		 * any_layout => number of segments
2024 		 * default    => number of segments + 1
2025 		 */
2026 		slots = txm->nb_segs + !can_push;
2027 		need = slots - vq->vq_free_cnt;
2028 
2029 		/* Positive value indicates it need free vring descriptors */
2030 		if (unlikely(need > 0)) {
2031 			virtio_xmit_cleanup_packed(vq, need, in_order);
2032 			need = slots - vq->vq_free_cnt;
2033 			if (unlikely(need > 0)) {
2034 				PMD_TX_LOG(ERR,
2035 					   "No free tx descriptors to transmit");
2036 				break;
2037 			}
2038 		}
2039 
2040 		/* Enqueue Packet buffers */
2041 		if (can_push)
2042 			virtqueue_enqueue_xmit_packed_fast(txvq, txm, in_order);
2043 		else
2044 			virtqueue_enqueue_xmit_packed(txvq, txm, slots, 0,
2045 						      in_order);
2046 
2047 		virtio_update_packet_stats(&txvq->stats, txm);
2048 	}
2049 
2050 	txvq->stats.packets += nb_tx;
2051 
2052 	if (likely(nb_tx)) {
2053 		if (unlikely(virtqueue_kick_prepare_packed(vq))) {
2054 			virtqueue_notify(vq);
2055 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2056 		}
2057 	}
2058 
2059 	return nb_tx;
2060 }
2061 
2062 uint16_t
2063 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
2064 {
2065 	struct virtnet_tx *txvq = tx_queue;
2066 	struct virtqueue *vq = txvq->vq;
2067 	struct virtio_hw *hw = vq->hw;
2068 	uint16_t hdr_size = hw->vtnet_hdr_size;
2069 	uint16_t nb_used, nb_tx = 0;
2070 	int error;
2071 
2072 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
2073 		return nb_tx;
2074 
2075 	if (unlikely(nb_pkts < 1))
2076 		return nb_pkts;
2077 
2078 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
2079 	nb_used = VIRTQUEUE_NUSED(vq);
2080 
2081 	virtio_rmb(hw->weak_barriers);
2082 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
2083 		virtio_xmit_cleanup(vq, nb_used);
2084 
2085 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
2086 		struct rte_mbuf *txm = tx_pkts[nb_tx];
2087 		int can_push = 0, use_indirect = 0, slots, need;
2088 
2089 		/* Do VLAN tag insertion */
2090 		if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
2091 			error = rte_vlan_insert(&txm);
2092 			if (unlikely(error)) {
2093 				rte_pktmbuf_free(txm);
2094 				continue;
2095 			}
2096 			/* vlan_insert may add a header mbuf */
2097 			tx_pkts[nb_tx] = txm;
2098 		}
2099 
2100 		/* optimize ring usage */
2101 		if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
2102 		      vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
2103 		    rte_mbuf_refcnt_read(txm) == 1 &&
2104 		    RTE_MBUF_DIRECT(txm) &&
2105 		    txm->nb_segs == 1 &&
2106 		    rte_pktmbuf_headroom(txm) >= hdr_size &&
2107 		    rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
2108 				   __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
2109 			can_push = 1;
2110 		else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
2111 			 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
2112 			use_indirect = 1;
2113 
2114 		/* How many main ring entries are needed to this Tx?
2115 		 * any_layout => number of segments
2116 		 * indirect   => 1
2117 		 * default    => number of segments + 1
2118 		 */
2119 		slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
2120 		need = slots - vq->vq_free_cnt;
2121 
2122 		/* Positive value indicates it need free vring descriptors */
2123 		if (unlikely(need > 0)) {
2124 			nb_used = VIRTQUEUE_NUSED(vq);
2125 			virtio_rmb(hw->weak_barriers);
2126 			need = RTE_MIN(need, (int)nb_used);
2127 
2128 			virtio_xmit_cleanup(vq, need);
2129 			need = slots - vq->vq_free_cnt;
2130 			if (unlikely(need > 0)) {
2131 				PMD_TX_LOG(ERR,
2132 					   "No free tx descriptors to transmit");
2133 				break;
2134 			}
2135 		}
2136 
2137 		/* Enqueue Packet buffers */
2138 		virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
2139 			can_push, 0);
2140 
2141 		virtio_update_packet_stats(&txvq->stats, txm);
2142 	}
2143 
2144 	txvq->stats.packets += nb_tx;
2145 
2146 	if (likely(nb_tx)) {
2147 		vq_update_avail_idx(vq);
2148 
2149 		if (unlikely(virtqueue_kick_prepare(vq))) {
2150 			virtqueue_notify(vq);
2151 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2152 		}
2153 	}
2154 
2155 	return nb_tx;
2156 }
2157 
2158 uint16_t
2159 virtio_xmit_pkts_inorder(void *tx_queue,
2160 			struct rte_mbuf **tx_pkts,
2161 			uint16_t nb_pkts)
2162 {
2163 	struct virtnet_tx *txvq = tx_queue;
2164 	struct virtqueue *vq = txvq->vq;
2165 	struct virtio_hw *hw = vq->hw;
2166 	uint16_t hdr_size = hw->vtnet_hdr_size;
2167 	uint16_t nb_used, nb_avail, nb_tx = 0, nb_inorder_pkts = 0;
2168 	struct rte_mbuf *inorder_pkts[nb_pkts];
2169 	int error;
2170 
2171 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
2172 		return nb_tx;
2173 
2174 	if (unlikely(nb_pkts < 1))
2175 		return nb_pkts;
2176 
2177 	VIRTQUEUE_DUMP(vq);
2178 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
2179 	nb_used = VIRTQUEUE_NUSED(vq);
2180 
2181 	virtio_rmb(hw->weak_barriers);
2182 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
2183 		virtio_xmit_cleanup_inorder(vq, nb_used);
2184 
2185 	if (unlikely(!vq->vq_free_cnt))
2186 		virtio_xmit_cleanup_inorder(vq, nb_used);
2187 
2188 	nb_avail = RTE_MIN(vq->vq_free_cnt, nb_pkts);
2189 
2190 	for (nb_tx = 0; nb_tx < nb_avail; nb_tx++) {
2191 		struct rte_mbuf *txm = tx_pkts[nb_tx];
2192 		int slots, need;
2193 
2194 		/* Do VLAN tag insertion */
2195 		if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
2196 			error = rte_vlan_insert(&txm);
2197 			if (unlikely(error)) {
2198 				rte_pktmbuf_free(txm);
2199 				continue;
2200 			}
2201 			/* vlan_insert may add a header mbuf */
2202 			tx_pkts[nb_tx] = txm;
2203 		}
2204 
2205 		/* optimize ring usage */
2206 		if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
2207 		     vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
2208 		     rte_mbuf_refcnt_read(txm) == 1 &&
2209 		     RTE_MBUF_DIRECT(txm) &&
2210 		     txm->nb_segs == 1 &&
2211 		     rte_pktmbuf_headroom(txm) >= hdr_size &&
2212 		     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
2213 				__alignof__(struct virtio_net_hdr_mrg_rxbuf))) {
2214 			inorder_pkts[nb_inorder_pkts] = txm;
2215 			nb_inorder_pkts++;
2216 
2217 			virtio_update_packet_stats(&txvq->stats, txm);
2218 			continue;
2219 		}
2220 
2221 		if (nb_inorder_pkts) {
2222 			virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2223 							nb_inorder_pkts);
2224 			nb_inorder_pkts = 0;
2225 		}
2226 
2227 		slots = txm->nb_segs + 1;
2228 		need = slots - vq->vq_free_cnt;
2229 		if (unlikely(need > 0)) {
2230 			nb_used = VIRTQUEUE_NUSED(vq);
2231 			virtio_rmb(hw->weak_barriers);
2232 			need = RTE_MIN(need, (int)nb_used);
2233 
2234 			virtio_xmit_cleanup_inorder(vq, need);
2235 
2236 			need = slots - vq->vq_free_cnt;
2237 
2238 			if (unlikely(need > 0)) {
2239 				PMD_TX_LOG(ERR,
2240 					"No free tx descriptors to transmit");
2241 				break;
2242 			}
2243 		}
2244 		/* Enqueue Packet buffers */
2245 		virtqueue_enqueue_xmit(txvq, txm, slots, 0, 0, 1);
2246 
2247 		virtio_update_packet_stats(&txvq->stats, txm);
2248 	}
2249 
2250 	/* Transmit all inorder packets */
2251 	if (nb_inorder_pkts)
2252 		virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
2253 						nb_inorder_pkts);
2254 
2255 	txvq->stats.packets += nb_tx;
2256 
2257 	if (likely(nb_tx)) {
2258 		vq_update_avail_idx(vq);
2259 
2260 		if (unlikely(virtqueue_kick_prepare(vq))) {
2261 			virtqueue_notify(vq);
2262 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
2263 		}
2264 	}
2265 
2266 	VIRTQUEUE_DUMP(vq);
2267 
2268 	return nb_tx;
2269 }
2270