xref: /dpdk/drivers/net/virtio/virtio_rxtx.c (revision 5566a3e35866ce9e5eacf886c27b460ebfcd6ee9)
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.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 
34 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
35 #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
36 #else
37 #define  VIRTIO_DUMP_PACKET(m, len) do { } while (0)
38 #endif
39 
40 
41 #define VIRTIO_SIMPLE_FLAGS ((uint32_t)ETH_TXQ_FLAGS_NOMULTSEGS | \
42 	ETH_TXQ_FLAGS_NOOFFLOADS)
43 
44 int
45 virtio_dev_rx_queue_done(void *rxq, uint16_t offset)
46 {
47 	struct virtnet_rx *rxvq = rxq;
48 	struct virtqueue *vq = rxvq->vq;
49 
50 	return VIRTQUEUE_NUSED(vq) >= offset;
51 }
52 
53 void
54 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
55 {
56 	struct vring_desc *dp, *dp_tail;
57 	struct vq_desc_extra *dxp;
58 	uint16_t desc_idx_last = desc_idx;
59 
60 	dp  = &vq->vq_ring.desc[desc_idx];
61 	dxp = &vq->vq_descx[desc_idx];
62 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
63 	if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
64 		while (dp->flags & VRING_DESC_F_NEXT) {
65 			desc_idx_last = dp->next;
66 			dp = &vq->vq_ring.desc[dp->next];
67 		}
68 	}
69 	dxp->ndescs = 0;
70 
71 	/*
72 	 * We must append the existing free chain, if any, to the end of
73 	 * newly freed chain. If the virtqueue was completely used, then
74 	 * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
75 	 */
76 	if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
77 		vq->vq_desc_head_idx = desc_idx;
78 	} else {
79 		dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
80 		dp_tail->next = desc_idx;
81 	}
82 
83 	vq->vq_desc_tail_idx = desc_idx_last;
84 	dp->next = VQ_RING_DESC_CHAIN_END;
85 }
86 
87 static uint16_t
88 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
89 			   uint32_t *len, uint16_t num)
90 {
91 	struct vring_used_elem *uep;
92 	struct rte_mbuf *cookie;
93 	uint16_t used_idx, desc_idx;
94 	uint16_t i;
95 
96 	/*  Caller does the check */
97 	for (i = 0; i < num ; i++) {
98 		used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
99 		uep = &vq->vq_ring.used->ring[used_idx];
100 		desc_idx = (uint16_t) uep->id;
101 		len[i] = uep->len;
102 		cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
103 
104 		if (unlikely(cookie == NULL)) {
105 			PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
106 				vq->vq_used_cons_idx);
107 			break;
108 		}
109 
110 		rte_prefetch0(cookie);
111 		rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
112 		rx_pkts[i]  = cookie;
113 		vq->vq_used_cons_idx++;
114 		vq_ring_free_chain(vq, desc_idx);
115 		vq->vq_descx[desc_idx].cookie = NULL;
116 	}
117 
118 	return i;
119 }
120 
121 #ifndef DEFAULT_TX_FREE_THRESH
122 #define DEFAULT_TX_FREE_THRESH 32
123 #endif
124 
125 /* Cleanup from completed transmits. */
126 static void
127 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
128 {
129 	uint16_t i, used_idx, desc_idx;
130 	for (i = 0; i < num; i++) {
131 		struct vring_used_elem *uep;
132 		struct vq_desc_extra *dxp;
133 
134 		used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
135 		uep = &vq->vq_ring.used->ring[used_idx];
136 
137 		desc_idx = (uint16_t) uep->id;
138 		dxp = &vq->vq_descx[desc_idx];
139 		vq->vq_used_cons_idx++;
140 		vq_ring_free_chain(vq, desc_idx);
141 
142 		if (dxp->cookie != NULL) {
143 			rte_pktmbuf_free(dxp->cookie);
144 			dxp->cookie = NULL;
145 		}
146 	}
147 }
148 
149 
150 static inline int
151 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
152 {
153 	struct vq_desc_extra *dxp;
154 	struct virtio_hw *hw = vq->hw;
155 	struct vring_desc *start_dp;
156 	uint16_t needed = 1;
157 	uint16_t head_idx, idx;
158 
159 	if (unlikely(vq->vq_free_cnt == 0))
160 		return -ENOSPC;
161 	if (unlikely(vq->vq_free_cnt < needed))
162 		return -EMSGSIZE;
163 
164 	head_idx = vq->vq_desc_head_idx;
165 	if (unlikely(head_idx >= vq->vq_nentries))
166 		return -EFAULT;
167 
168 	idx = head_idx;
169 	dxp = &vq->vq_descx[idx];
170 	dxp->cookie = (void *)cookie;
171 	dxp->ndescs = needed;
172 
173 	start_dp = vq->vq_ring.desc;
174 	start_dp[idx].addr =
175 		VIRTIO_MBUF_ADDR(cookie, vq) +
176 		RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
177 	start_dp[idx].len =
178 		cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
179 	start_dp[idx].flags =  VRING_DESC_F_WRITE;
180 	idx = start_dp[idx].next;
181 	vq->vq_desc_head_idx = idx;
182 	if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
183 		vq->vq_desc_tail_idx = idx;
184 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
185 	vq_update_avail_ring(vq, head_idx);
186 
187 	return 0;
188 }
189 
190 /* When doing TSO, the IP length is not included in the pseudo header
191  * checksum of the packet given to the PMD, but for virtio it is
192  * expected.
193  */
194 static void
195 virtio_tso_fix_cksum(struct rte_mbuf *m)
196 {
197 	/* common case: header is not fragmented */
198 	if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
199 			m->l4_len)) {
200 		struct ipv4_hdr *iph;
201 		struct ipv6_hdr *ip6h;
202 		struct tcp_hdr *th;
203 		uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
204 		uint32_t tmp;
205 
206 		iph = rte_pktmbuf_mtod_offset(m, struct ipv4_hdr *, m->l2_len);
207 		th = RTE_PTR_ADD(iph, m->l3_len);
208 		if ((iph->version_ihl >> 4) == 4) {
209 			iph->hdr_checksum = 0;
210 			iph->hdr_checksum = rte_ipv4_cksum(iph);
211 			ip_len = iph->total_length;
212 			ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
213 				m->l3_len);
214 		} else {
215 			ip6h = (struct ipv6_hdr *)iph;
216 			ip_paylen = ip6h->payload_len;
217 		}
218 
219 		/* calculate the new phdr checksum not including ip_paylen */
220 		prev_cksum = th->cksum;
221 		tmp = prev_cksum;
222 		tmp += ip_paylen;
223 		tmp = (tmp & 0xffff) + (tmp >> 16);
224 		new_cksum = tmp;
225 
226 		/* replace it in the packet */
227 		th->cksum = new_cksum;
228 	}
229 }
230 
231 static inline int
232 tx_offload_enabled(struct virtio_hw *hw)
233 {
234 	return vtpci_with_feature(hw, VIRTIO_NET_F_CSUM) ||
235 		vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO4) ||
236 		vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO6);
237 }
238 
239 /* avoid write operation when necessary, to lessen cache issues */
240 #define ASSIGN_UNLESS_EQUAL(var, val) do {	\
241 	if ((var) != (val))			\
242 		(var) = (val);			\
243 } while (0)
244 
245 static inline void
246 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
247 		       uint16_t needed, int use_indirect, int can_push)
248 {
249 	struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
250 	struct vq_desc_extra *dxp;
251 	struct virtqueue *vq = txvq->vq;
252 	struct vring_desc *start_dp;
253 	uint16_t seg_num = cookie->nb_segs;
254 	uint16_t head_idx, idx;
255 	uint16_t head_size = vq->hw->vtnet_hdr_size;
256 	struct virtio_net_hdr *hdr;
257 	int offload;
258 
259 	offload = tx_offload_enabled(vq->hw);
260 	head_idx = vq->vq_desc_head_idx;
261 	idx = head_idx;
262 	dxp = &vq->vq_descx[idx];
263 	dxp->cookie = (void *)cookie;
264 	dxp->ndescs = needed;
265 
266 	start_dp = vq->vq_ring.desc;
267 
268 	if (can_push) {
269 		/* prepend cannot fail, checked by caller */
270 		hdr = (struct virtio_net_hdr *)
271 			rte_pktmbuf_prepend(cookie, head_size);
272 		/* rte_pktmbuf_prepend() counts the hdr size to the pkt length,
273 		 * which is wrong. Below subtract restores correct pkt size.
274 		 */
275 		cookie->pkt_len -= head_size;
276 		/* if offload disabled, it is not zeroed below, do it now */
277 		if (offload == 0) {
278 			ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
279 			ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
280 			ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
281 			ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
282 			ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
283 			ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
284 		}
285 	} else if (use_indirect) {
286 		/* setup tx ring slot to point to indirect
287 		 * descriptor list stored in reserved region.
288 		 *
289 		 * the first slot in indirect ring is already preset
290 		 * to point to the header in reserved region
291 		 */
292 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
293 			RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
294 		start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
295 		start_dp[idx].flags = VRING_DESC_F_INDIRECT;
296 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
297 
298 		/* loop below will fill in rest of the indirect elements */
299 		start_dp = txr[idx].tx_indir;
300 		idx = 1;
301 	} else {
302 		/* setup first tx ring slot to point to header
303 		 * stored in reserved region.
304 		 */
305 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
306 			RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
307 		start_dp[idx].len   = vq->hw->vtnet_hdr_size;
308 		start_dp[idx].flags = VRING_DESC_F_NEXT;
309 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
310 
311 		idx = start_dp[idx].next;
312 	}
313 
314 	/* Checksum Offload / TSO */
315 	if (offload) {
316 		if (cookie->ol_flags & PKT_TX_TCP_SEG)
317 			cookie->ol_flags |= PKT_TX_TCP_CKSUM;
318 
319 		switch (cookie->ol_flags & PKT_TX_L4_MASK) {
320 		case PKT_TX_UDP_CKSUM:
321 			hdr->csum_start = cookie->l2_len + cookie->l3_len;
322 			hdr->csum_offset = offsetof(struct udp_hdr,
323 				dgram_cksum);
324 			hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
325 			break;
326 
327 		case PKT_TX_TCP_CKSUM:
328 			hdr->csum_start = cookie->l2_len + cookie->l3_len;
329 			hdr->csum_offset = offsetof(struct tcp_hdr, cksum);
330 			hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
331 			break;
332 
333 		default:
334 			ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
335 			ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
336 			ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
337 			break;
338 		}
339 
340 		/* TCP Segmentation Offload */
341 		if (cookie->ol_flags & PKT_TX_TCP_SEG) {
342 			virtio_tso_fix_cksum(cookie);
343 			hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
344 				VIRTIO_NET_HDR_GSO_TCPV6 :
345 				VIRTIO_NET_HDR_GSO_TCPV4;
346 			hdr->gso_size = cookie->tso_segsz;
347 			hdr->hdr_len =
348 				cookie->l2_len +
349 				cookie->l3_len +
350 				cookie->l4_len;
351 		} else {
352 			ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
353 			ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
354 			ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
355 		}
356 	}
357 
358 	do {
359 		start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
360 		start_dp[idx].len   = cookie->data_len;
361 		start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
362 		idx = start_dp[idx].next;
363 	} while ((cookie = cookie->next) != NULL);
364 
365 	if (use_indirect)
366 		idx = vq->vq_ring.desc[head_idx].next;
367 
368 	vq->vq_desc_head_idx = idx;
369 	if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
370 		vq->vq_desc_tail_idx = idx;
371 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
372 	vq_update_avail_ring(vq, head_idx);
373 }
374 
375 void
376 virtio_dev_cq_start(struct rte_eth_dev *dev)
377 {
378 	struct virtio_hw *hw = dev->data->dev_private;
379 
380 	if (hw->cvq && hw->cvq->vq) {
381 		VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
382 	}
383 }
384 
385 int
386 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
387 			uint16_t queue_idx,
388 			uint16_t nb_desc,
389 			unsigned int socket_id __rte_unused,
390 			__rte_unused const struct rte_eth_rxconf *rx_conf,
391 			struct rte_mempool *mp)
392 {
393 	uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
394 	struct virtio_hw *hw = dev->data->dev_private;
395 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
396 	struct virtnet_rx *rxvq;
397 
398 	PMD_INIT_FUNC_TRACE();
399 
400 	if (nb_desc == 0 || nb_desc > vq->vq_nentries)
401 		nb_desc = vq->vq_nentries;
402 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
403 
404 	rxvq = &vq->rxq;
405 	rxvq->queue_id = queue_idx;
406 	rxvq->mpool = mp;
407 	if (rxvq->mpool == NULL) {
408 		rte_exit(EXIT_FAILURE,
409 			"Cannot allocate mbufs for rx virtqueue");
410 	}
411 	dev->data->rx_queues[queue_idx] = rxvq;
412 
413 	return 0;
414 }
415 
416 int
417 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
418 {
419 	uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
420 	struct virtio_hw *hw = dev->data->dev_private;
421 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
422 	struct virtnet_rx *rxvq = &vq->rxq;
423 	struct rte_mbuf *m;
424 	uint16_t desc_idx;
425 	int error, nbufs;
426 
427 	PMD_INIT_FUNC_TRACE();
428 
429 	/* Allocate blank mbufs for the each rx descriptor */
430 	nbufs = 0;
431 
432 	if (hw->use_simple_rx) {
433 		for (desc_idx = 0; desc_idx < vq->vq_nentries;
434 		     desc_idx++) {
435 			vq->vq_ring.avail->ring[desc_idx] = desc_idx;
436 			vq->vq_ring.desc[desc_idx].flags =
437 				VRING_DESC_F_WRITE;
438 		}
439 	}
440 
441 	memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
442 	for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
443 	     desc_idx++) {
444 		vq->sw_ring[vq->vq_nentries + desc_idx] =
445 			&rxvq->fake_mbuf;
446 	}
447 
448 	while (!virtqueue_full(vq)) {
449 		m = rte_mbuf_raw_alloc(rxvq->mpool);
450 		if (m == NULL)
451 			break;
452 
453 		/* Enqueue allocated buffers */
454 		if (hw->use_simple_rx)
455 			error = virtqueue_enqueue_recv_refill_simple(vq, m);
456 		else
457 			error = virtqueue_enqueue_recv_refill(vq, m);
458 
459 		if (error) {
460 			rte_pktmbuf_free(m);
461 			break;
462 		}
463 		nbufs++;
464 	}
465 
466 	vq_update_avail_idx(vq);
467 
468 	PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
469 
470 	virtio_rxq_vec_setup(rxvq);
471 
472 	VIRTQUEUE_DUMP(vq);
473 
474 	return 0;
475 }
476 
477 /*
478  * struct rte_eth_dev *dev: Used to update dev
479  * uint16_t nb_desc: Defaults to values read from config space
480  * unsigned int socket_id: Used to allocate memzone
481  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
482  * uint16_t queue_idx: Just used as an index in dev txq list
483  */
484 int
485 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
486 			uint16_t queue_idx,
487 			uint16_t nb_desc,
488 			unsigned int socket_id __rte_unused,
489 			const struct rte_eth_txconf *tx_conf)
490 {
491 	uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
492 	struct virtio_hw *hw = dev->data->dev_private;
493 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
494 	struct virtnet_tx *txvq;
495 	uint16_t tx_free_thresh;
496 
497 	PMD_INIT_FUNC_TRACE();
498 
499 	/* cannot use simple rxtx funcs with multisegs or offloads */
500 	if ((tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) != VIRTIO_SIMPLE_FLAGS)
501 		hw->use_simple_tx = 0;
502 
503 	if (nb_desc == 0 || nb_desc > vq->vq_nentries)
504 		nb_desc = vq->vq_nentries;
505 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
506 
507 	txvq = &vq->txq;
508 	txvq->queue_id = queue_idx;
509 
510 	tx_free_thresh = tx_conf->tx_free_thresh;
511 	if (tx_free_thresh == 0)
512 		tx_free_thresh =
513 			RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
514 
515 	if (tx_free_thresh >= (vq->vq_nentries - 3)) {
516 		RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
517 			"number of TX entries minus 3 (%u)."
518 			" (tx_free_thresh=%u port=%u queue=%u)\n",
519 			vq->vq_nentries - 3,
520 			tx_free_thresh, dev->data->port_id, queue_idx);
521 		return -EINVAL;
522 	}
523 
524 	vq->vq_free_thresh = tx_free_thresh;
525 
526 	dev->data->tx_queues[queue_idx] = txvq;
527 	return 0;
528 }
529 
530 int
531 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
532 				uint16_t queue_idx)
533 {
534 	uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
535 	struct virtio_hw *hw = dev->data->dev_private;
536 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
537 	uint16_t mid_idx = vq->vq_nentries >> 1;
538 	struct virtnet_tx *txvq = &vq->txq;
539 	uint16_t desc_idx;
540 
541 	PMD_INIT_FUNC_TRACE();
542 
543 	if (hw->use_simple_tx) {
544 		for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
545 			vq->vq_ring.avail->ring[desc_idx] =
546 				desc_idx + mid_idx;
547 			vq->vq_ring.desc[desc_idx + mid_idx].next =
548 				desc_idx;
549 			vq->vq_ring.desc[desc_idx + mid_idx].addr =
550 				txvq->virtio_net_hdr_mem +
551 				offsetof(struct virtio_tx_region, tx_hdr);
552 			vq->vq_ring.desc[desc_idx + mid_idx].len =
553 				vq->hw->vtnet_hdr_size;
554 			vq->vq_ring.desc[desc_idx + mid_idx].flags =
555 				VRING_DESC_F_NEXT;
556 			vq->vq_ring.desc[desc_idx].flags = 0;
557 		}
558 		for (desc_idx = mid_idx; desc_idx < vq->vq_nentries;
559 		     desc_idx++)
560 			vq->vq_ring.avail->ring[desc_idx] = desc_idx;
561 	}
562 
563 	VIRTQUEUE_DUMP(vq);
564 
565 	return 0;
566 }
567 
568 static void
569 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
570 {
571 	int error;
572 	/*
573 	 * Requeue the discarded mbuf. This should always be
574 	 * successful since it was just dequeued.
575 	 */
576 	error = virtqueue_enqueue_recv_refill(vq, m);
577 	if (unlikely(error)) {
578 		RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
579 		rte_pktmbuf_free(m);
580 	}
581 }
582 
583 static void
584 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
585 {
586 	uint32_t s = mbuf->pkt_len;
587 	struct ether_addr *ea;
588 
589 	if (s == 64) {
590 		stats->size_bins[1]++;
591 	} else if (s > 64 && s < 1024) {
592 		uint32_t bin;
593 
594 		/* count zeros, and offset into correct bin */
595 		bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
596 		stats->size_bins[bin]++;
597 	} else {
598 		if (s < 64)
599 			stats->size_bins[0]++;
600 		else if (s < 1519)
601 			stats->size_bins[6]++;
602 		else if (s >= 1519)
603 			stats->size_bins[7]++;
604 	}
605 
606 	ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
607 	if (is_multicast_ether_addr(ea)) {
608 		if (is_broadcast_ether_addr(ea))
609 			stats->broadcast++;
610 		else
611 			stats->multicast++;
612 	}
613 }
614 
615 /* Optionally fill offload information in structure */
616 static int
617 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
618 {
619 	struct rte_net_hdr_lens hdr_lens;
620 	uint32_t hdrlen, ptype;
621 	int l4_supported = 0;
622 
623 	/* nothing to do */
624 	if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
625 		return 0;
626 
627 	m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
628 
629 	ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
630 	m->packet_type = ptype;
631 	if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
632 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
633 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
634 		l4_supported = 1;
635 
636 	if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
637 		hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
638 		if (hdr->csum_start <= hdrlen && l4_supported) {
639 			m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
640 		} else {
641 			/* Unknown proto or tunnel, do sw cksum. We can assume
642 			 * the cksum field is in the first segment since the
643 			 * buffers we provided to the host are large enough.
644 			 * In case of SCTP, this will be wrong since it's a CRC
645 			 * but there's nothing we can do.
646 			 */
647 			uint16_t csum = 0, off;
648 
649 			rte_raw_cksum_mbuf(m, hdr->csum_start,
650 				rte_pktmbuf_pkt_len(m) - hdr->csum_start,
651 				&csum);
652 			if (likely(csum != 0xffff))
653 				csum = ~csum;
654 			off = hdr->csum_offset + hdr->csum_start;
655 			if (rte_pktmbuf_data_len(m) >= off + 1)
656 				*rte_pktmbuf_mtod_offset(m, uint16_t *,
657 					off) = csum;
658 		}
659 	} else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
660 		m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
661 	}
662 
663 	/* GSO request, save required information in mbuf */
664 	if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
665 		/* Check unsupported modes */
666 		if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
667 		    (hdr->gso_size == 0)) {
668 			return -EINVAL;
669 		}
670 
671 		/* Update mss lengthes in mbuf */
672 		m->tso_segsz = hdr->gso_size;
673 		switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
674 			case VIRTIO_NET_HDR_GSO_TCPV4:
675 			case VIRTIO_NET_HDR_GSO_TCPV6:
676 				m->ol_flags |= PKT_RX_LRO | \
677 					PKT_RX_L4_CKSUM_NONE;
678 				break;
679 			default:
680 				return -EINVAL;
681 		}
682 	}
683 
684 	return 0;
685 }
686 
687 static inline int
688 rx_offload_enabled(struct virtio_hw *hw)
689 {
690 	return vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_CSUM) ||
691 		vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO4) ||
692 		vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO6);
693 }
694 
695 #define VIRTIO_MBUF_BURST_SZ 64
696 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
697 uint16_t
698 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
699 {
700 	struct virtnet_rx *rxvq = rx_queue;
701 	struct virtqueue *vq = rxvq->vq;
702 	struct virtio_hw *hw = vq->hw;
703 	struct rte_mbuf *rxm, *new_mbuf;
704 	uint16_t nb_used, num, nb_rx;
705 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
706 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
707 	int error;
708 	uint32_t i, nb_enqueued;
709 	uint32_t hdr_size;
710 	int offload;
711 	struct virtio_net_hdr *hdr;
712 
713 	nb_rx = 0;
714 	if (unlikely(hw->started == 0))
715 		return nb_rx;
716 
717 	nb_used = VIRTQUEUE_NUSED(vq);
718 
719 	virtio_rmb();
720 
721 	num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
722 	if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
723 		num = VIRTIO_MBUF_BURST_SZ;
724 	if (likely(num > DESC_PER_CACHELINE))
725 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
726 
727 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
728 	PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
729 
730 	nb_enqueued = 0;
731 	hdr_size = hw->vtnet_hdr_size;
732 	offload = rx_offload_enabled(hw);
733 
734 	for (i = 0; i < num ; i++) {
735 		rxm = rcv_pkts[i];
736 
737 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
738 
739 		if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
740 			PMD_RX_LOG(ERR, "Packet drop");
741 			nb_enqueued++;
742 			virtio_discard_rxbuf(vq, rxm);
743 			rxvq->stats.errors++;
744 			continue;
745 		}
746 
747 		rxm->port = rxvq->port_id;
748 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
749 		rxm->ol_flags = 0;
750 		rxm->vlan_tci = 0;
751 
752 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
753 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
754 
755 		hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
756 			RTE_PKTMBUF_HEADROOM - hdr_size);
757 
758 		if (hw->vlan_strip)
759 			rte_vlan_strip(rxm);
760 
761 		if (offload && virtio_rx_offload(rxm, hdr) < 0) {
762 			virtio_discard_rxbuf(vq, rxm);
763 			rxvq->stats.errors++;
764 			continue;
765 		}
766 
767 		VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
768 
769 		rx_pkts[nb_rx++] = rxm;
770 
771 		rxvq->stats.bytes += rxm->pkt_len;
772 		virtio_update_packet_stats(&rxvq->stats, rxm);
773 	}
774 
775 	rxvq->stats.packets += nb_rx;
776 
777 	/* Allocate new mbuf for the used descriptor */
778 	error = ENOSPC;
779 	while (likely(!virtqueue_full(vq))) {
780 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
781 		if (unlikely(new_mbuf == NULL)) {
782 			struct rte_eth_dev *dev
783 				= &rte_eth_devices[rxvq->port_id];
784 			dev->data->rx_mbuf_alloc_failed++;
785 			break;
786 		}
787 		error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
788 		if (unlikely(error)) {
789 			rte_pktmbuf_free(new_mbuf);
790 			break;
791 		}
792 		nb_enqueued++;
793 	}
794 
795 	if (likely(nb_enqueued)) {
796 		vq_update_avail_idx(vq);
797 
798 		if (unlikely(virtqueue_kick_prepare(vq))) {
799 			virtqueue_notify(vq);
800 			PMD_RX_LOG(DEBUG, "Notified");
801 		}
802 	}
803 
804 	return nb_rx;
805 }
806 
807 uint16_t
808 virtio_recv_mergeable_pkts(void *rx_queue,
809 			struct rte_mbuf **rx_pkts,
810 			uint16_t nb_pkts)
811 {
812 	struct virtnet_rx *rxvq = rx_queue;
813 	struct virtqueue *vq = rxvq->vq;
814 	struct virtio_hw *hw = vq->hw;
815 	struct rte_mbuf *rxm, *new_mbuf;
816 	uint16_t nb_used, num, nb_rx;
817 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
818 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
819 	struct rte_mbuf *prev;
820 	int error;
821 	uint32_t i, nb_enqueued;
822 	uint32_t seg_num;
823 	uint16_t extra_idx;
824 	uint32_t seg_res;
825 	uint32_t hdr_size;
826 	int offload;
827 
828 	nb_rx = 0;
829 	if (unlikely(hw->started == 0))
830 		return nb_rx;
831 
832 	nb_used = VIRTQUEUE_NUSED(vq);
833 
834 	virtio_rmb();
835 
836 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
837 
838 	i = 0;
839 	nb_enqueued = 0;
840 	seg_num = 0;
841 	extra_idx = 0;
842 	seg_res = 0;
843 	hdr_size = hw->vtnet_hdr_size;
844 	offload = rx_offload_enabled(hw);
845 
846 	while (i < nb_used) {
847 		struct virtio_net_hdr_mrg_rxbuf *header;
848 
849 		if (nb_rx == nb_pkts)
850 			break;
851 
852 		num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
853 		if (num != 1)
854 			continue;
855 
856 		i++;
857 
858 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
859 		PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
860 
861 		rxm = rcv_pkts[0];
862 
863 		if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
864 			PMD_RX_LOG(ERR, "Packet drop");
865 			nb_enqueued++;
866 			virtio_discard_rxbuf(vq, rxm);
867 			rxvq->stats.errors++;
868 			continue;
869 		}
870 
871 		header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
872 			RTE_PKTMBUF_HEADROOM - hdr_size);
873 		seg_num = header->num_buffers;
874 
875 		if (seg_num == 0)
876 			seg_num = 1;
877 
878 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
879 		rxm->nb_segs = seg_num;
880 		rxm->ol_flags = 0;
881 		rxm->vlan_tci = 0;
882 		rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
883 		rxm->data_len = (uint16_t)(len[0] - hdr_size);
884 
885 		rxm->port = rxvq->port_id;
886 		rx_pkts[nb_rx] = rxm;
887 		prev = rxm;
888 
889 		if (offload && virtio_rx_offload(rxm, &header->hdr) < 0) {
890 			virtio_discard_rxbuf(vq, rxm);
891 			rxvq->stats.errors++;
892 			continue;
893 		}
894 
895 		seg_res = seg_num - 1;
896 
897 		while (seg_res != 0) {
898 			/*
899 			 * Get extra segments for current uncompleted packet.
900 			 */
901 			uint16_t  rcv_cnt =
902 				RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
903 			if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
904 				uint32_t rx_num =
905 					virtqueue_dequeue_burst_rx(vq,
906 					rcv_pkts, len, rcv_cnt);
907 				i += rx_num;
908 				rcv_cnt = rx_num;
909 			} else {
910 				PMD_RX_LOG(ERR,
911 					   "No enough segments for packet.");
912 				nb_enqueued++;
913 				virtio_discard_rxbuf(vq, rxm);
914 				rxvq->stats.errors++;
915 				break;
916 			}
917 
918 			extra_idx = 0;
919 
920 			while (extra_idx < rcv_cnt) {
921 				rxm = rcv_pkts[extra_idx];
922 
923 				rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
924 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
925 				rxm->data_len = (uint16_t)(len[extra_idx]);
926 
927 				if (prev)
928 					prev->next = rxm;
929 
930 				prev = rxm;
931 				rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
932 				extra_idx++;
933 			};
934 			seg_res -= rcv_cnt;
935 		}
936 
937 		if (hw->vlan_strip)
938 			rte_vlan_strip(rx_pkts[nb_rx]);
939 
940 		VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
941 			rx_pkts[nb_rx]->data_len);
942 
943 		rxvq->stats.bytes += rx_pkts[nb_rx]->pkt_len;
944 		virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
945 		nb_rx++;
946 	}
947 
948 	rxvq->stats.packets += nb_rx;
949 
950 	/* Allocate new mbuf for the used descriptor */
951 	error = ENOSPC;
952 	while (likely(!virtqueue_full(vq))) {
953 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
954 		if (unlikely(new_mbuf == NULL)) {
955 			struct rte_eth_dev *dev
956 				= &rte_eth_devices[rxvq->port_id];
957 			dev->data->rx_mbuf_alloc_failed++;
958 			break;
959 		}
960 		error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
961 		if (unlikely(error)) {
962 			rte_pktmbuf_free(new_mbuf);
963 			break;
964 		}
965 		nb_enqueued++;
966 	}
967 
968 	if (likely(nb_enqueued)) {
969 		vq_update_avail_idx(vq);
970 
971 		if (unlikely(virtqueue_kick_prepare(vq))) {
972 			virtqueue_notify(vq);
973 			PMD_RX_LOG(DEBUG, "Notified");
974 		}
975 	}
976 
977 	return nb_rx;
978 }
979 
980 uint16_t
981 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
982 {
983 	struct virtnet_tx *txvq = tx_queue;
984 	struct virtqueue *vq = txvq->vq;
985 	struct virtio_hw *hw = vq->hw;
986 	uint16_t hdr_size = hw->vtnet_hdr_size;
987 	uint16_t nb_used, nb_tx = 0;
988 	int error;
989 
990 	if (unlikely(hw->started == 0))
991 		return nb_tx;
992 
993 	if (unlikely(nb_pkts < 1))
994 		return nb_pkts;
995 
996 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
997 	nb_used = VIRTQUEUE_NUSED(vq);
998 
999 	virtio_rmb();
1000 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1001 		virtio_xmit_cleanup(vq, nb_used);
1002 
1003 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1004 		struct rte_mbuf *txm = tx_pkts[nb_tx];
1005 		int can_push = 0, use_indirect = 0, slots, need;
1006 
1007 		/* Do VLAN tag insertion */
1008 		if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1009 			error = rte_vlan_insert(&txm);
1010 			if (unlikely(error)) {
1011 				rte_pktmbuf_free(txm);
1012 				continue;
1013 			}
1014 		}
1015 
1016 		/* optimize ring usage */
1017 		if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1018 		      vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1019 		    rte_mbuf_refcnt_read(txm) == 1 &&
1020 		    RTE_MBUF_DIRECT(txm) &&
1021 		    txm->nb_segs == 1 &&
1022 		    rte_pktmbuf_headroom(txm) >= hdr_size &&
1023 		    rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1024 				   __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1025 			can_push = 1;
1026 		else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1027 			 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1028 			use_indirect = 1;
1029 
1030 		/* How many main ring entries are needed to this Tx?
1031 		 * any_layout => number of segments
1032 		 * indirect   => 1
1033 		 * default    => number of segments + 1
1034 		 */
1035 		slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1036 		need = slots - vq->vq_free_cnt;
1037 
1038 		/* Positive value indicates it need free vring descriptors */
1039 		if (unlikely(need > 0)) {
1040 			nb_used = VIRTQUEUE_NUSED(vq);
1041 			virtio_rmb();
1042 			need = RTE_MIN(need, (int)nb_used);
1043 
1044 			virtio_xmit_cleanup(vq, need);
1045 			need = slots - vq->vq_free_cnt;
1046 			if (unlikely(need > 0)) {
1047 				PMD_TX_LOG(ERR,
1048 					   "No free tx descriptors to transmit");
1049 				break;
1050 			}
1051 		}
1052 
1053 		/* Enqueue Packet buffers */
1054 		virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
1055 
1056 		txvq->stats.bytes += txm->pkt_len;
1057 		virtio_update_packet_stats(&txvq->stats, txm);
1058 	}
1059 
1060 	txvq->stats.packets += nb_tx;
1061 
1062 	if (likely(nb_tx)) {
1063 		vq_update_avail_idx(vq);
1064 
1065 		if (unlikely(virtqueue_kick_prepare(vq))) {
1066 			virtqueue_notify(vq);
1067 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1068 		}
1069 	}
1070 
1071 	return nb_tx;
1072 }
1073