xref: /dpdk/drivers/net/virtio/virtio_rxtx.c (revision 89f0711f9ddfb5822da9d34f384b92f72a61c4dc)
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 
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 		rte_spinlock_init(&hw->cvq->lock);
382 		VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
383 	}
384 }
385 
386 int
387 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
388 			uint16_t queue_idx,
389 			uint16_t nb_desc,
390 			unsigned int socket_id __rte_unused,
391 			__rte_unused const struct rte_eth_rxconf *rx_conf,
392 			struct rte_mempool *mp)
393 {
394 	uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
395 	struct virtio_hw *hw = dev->data->dev_private;
396 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
397 	struct virtnet_rx *rxvq;
398 
399 	PMD_INIT_FUNC_TRACE();
400 
401 	if (nb_desc == 0 || nb_desc > vq->vq_nentries)
402 		nb_desc = vq->vq_nentries;
403 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
404 
405 	rxvq = &vq->rxq;
406 	rxvq->queue_id = queue_idx;
407 	rxvq->mpool = mp;
408 	if (rxvq->mpool == NULL) {
409 		rte_exit(EXIT_FAILURE,
410 			"Cannot allocate mbufs for rx virtqueue");
411 	}
412 	dev->data->rx_queues[queue_idx] = rxvq;
413 
414 	return 0;
415 }
416 
417 int
418 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
419 {
420 	uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
421 	struct virtio_hw *hw = dev->data->dev_private;
422 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
423 	struct virtnet_rx *rxvq = &vq->rxq;
424 	struct rte_mbuf *m;
425 	uint16_t desc_idx;
426 	int error, nbufs;
427 
428 	PMD_INIT_FUNC_TRACE();
429 
430 	/* Allocate blank mbufs for the each rx descriptor */
431 	nbufs = 0;
432 
433 	if (hw->use_simple_rx) {
434 		for (desc_idx = 0; desc_idx < vq->vq_nentries;
435 		     desc_idx++) {
436 			vq->vq_ring.avail->ring[desc_idx] = desc_idx;
437 			vq->vq_ring.desc[desc_idx].flags =
438 				VRING_DESC_F_WRITE;
439 		}
440 	}
441 
442 	memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
443 	for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
444 	     desc_idx++) {
445 		vq->sw_ring[vq->vq_nentries + desc_idx] =
446 			&rxvq->fake_mbuf;
447 	}
448 
449 	while (!virtqueue_full(vq)) {
450 		m = rte_mbuf_raw_alloc(rxvq->mpool);
451 		if (m == NULL)
452 			break;
453 
454 		/* Enqueue allocated buffers */
455 		if (hw->use_simple_rx)
456 			error = virtqueue_enqueue_recv_refill_simple(vq, m);
457 		else
458 			error = virtqueue_enqueue_recv_refill(vq, m);
459 
460 		if (error) {
461 			rte_pktmbuf_free(m);
462 			break;
463 		}
464 		nbufs++;
465 	}
466 
467 	vq_update_avail_idx(vq);
468 
469 	PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
470 
471 	virtio_rxq_vec_setup(rxvq);
472 
473 	VIRTQUEUE_DUMP(vq);
474 
475 	return 0;
476 }
477 
478 /*
479  * struct rte_eth_dev *dev: Used to update dev
480  * uint16_t nb_desc: Defaults to values read from config space
481  * unsigned int socket_id: Used to allocate memzone
482  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
483  * uint16_t queue_idx: Just used as an index in dev txq list
484  */
485 int
486 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
487 			uint16_t queue_idx,
488 			uint16_t nb_desc,
489 			unsigned int socket_id __rte_unused,
490 			const struct rte_eth_txconf *tx_conf)
491 {
492 	uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
493 	struct virtio_hw *hw = dev->data->dev_private;
494 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
495 	struct virtnet_tx *txvq;
496 	uint16_t tx_free_thresh;
497 
498 	PMD_INIT_FUNC_TRACE();
499 
500 	/* cannot use simple rxtx funcs with multisegs or offloads */
501 	if ((tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) != VIRTIO_SIMPLE_FLAGS)
502 		hw->use_simple_tx = 0;
503 
504 	if (nb_desc == 0 || nb_desc > vq->vq_nentries)
505 		nb_desc = vq->vq_nentries;
506 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
507 
508 	txvq = &vq->txq;
509 	txvq->queue_id = queue_idx;
510 
511 	tx_free_thresh = tx_conf->tx_free_thresh;
512 	if (tx_free_thresh == 0)
513 		tx_free_thresh =
514 			RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
515 
516 	if (tx_free_thresh >= (vq->vq_nentries - 3)) {
517 		RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
518 			"number of TX entries minus 3 (%u)."
519 			" (tx_free_thresh=%u port=%u queue=%u)\n",
520 			vq->vq_nentries - 3,
521 			tx_free_thresh, dev->data->port_id, queue_idx);
522 		return -EINVAL;
523 	}
524 
525 	vq->vq_free_thresh = tx_free_thresh;
526 
527 	dev->data->tx_queues[queue_idx] = txvq;
528 	return 0;
529 }
530 
531 int
532 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
533 				uint16_t queue_idx)
534 {
535 	uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
536 	struct virtio_hw *hw = dev->data->dev_private;
537 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
538 	uint16_t mid_idx = vq->vq_nentries >> 1;
539 	struct virtnet_tx *txvq = &vq->txq;
540 	uint16_t desc_idx;
541 
542 	PMD_INIT_FUNC_TRACE();
543 
544 	if (hw->use_simple_tx) {
545 		for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
546 			vq->vq_ring.avail->ring[desc_idx] =
547 				desc_idx + mid_idx;
548 			vq->vq_ring.desc[desc_idx + mid_idx].next =
549 				desc_idx;
550 			vq->vq_ring.desc[desc_idx + mid_idx].addr =
551 				txvq->virtio_net_hdr_mem +
552 				offsetof(struct virtio_tx_region, tx_hdr);
553 			vq->vq_ring.desc[desc_idx + mid_idx].len =
554 				vq->hw->vtnet_hdr_size;
555 			vq->vq_ring.desc[desc_idx + mid_idx].flags =
556 				VRING_DESC_F_NEXT;
557 			vq->vq_ring.desc[desc_idx].flags = 0;
558 		}
559 		for (desc_idx = mid_idx; desc_idx < vq->vq_nentries;
560 		     desc_idx++)
561 			vq->vq_ring.avail->ring[desc_idx] = desc_idx;
562 	}
563 
564 	VIRTQUEUE_DUMP(vq);
565 
566 	return 0;
567 }
568 
569 static void
570 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
571 {
572 	int error;
573 	/*
574 	 * Requeue the discarded mbuf. This should always be
575 	 * successful since it was just dequeued.
576 	 */
577 	error = virtqueue_enqueue_recv_refill(vq, m);
578 	if (unlikely(error)) {
579 		RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
580 		rte_pktmbuf_free(m);
581 	}
582 }
583 
584 static void
585 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
586 {
587 	uint32_t s = mbuf->pkt_len;
588 	struct ether_addr *ea;
589 
590 	if (s == 64) {
591 		stats->size_bins[1]++;
592 	} else if (s > 64 && s < 1024) {
593 		uint32_t bin;
594 
595 		/* count zeros, and offset into correct bin */
596 		bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
597 		stats->size_bins[bin]++;
598 	} else {
599 		if (s < 64)
600 			stats->size_bins[0]++;
601 		else if (s < 1519)
602 			stats->size_bins[6]++;
603 		else if (s >= 1519)
604 			stats->size_bins[7]++;
605 	}
606 
607 	ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
608 	if (is_multicast_ether_addr(ea)) {
609 		if (is_broadcast_ether_addr(ea))
610 			stats->broadcast++;
611 		else
612 			stats->multicast++;
613 	}
614 }
615 
616 /* Optionally fill offload information in structure */
617 static int
618 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
619 {
620 	struct rte_net_hdr_lens hdr_lens;
621 	uint32_t hdrlen, ptype;
622 	int l4_supported = 0;
623 
624 	/* nothing to do */
625 	if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
626 		return 0;
627 
628 	m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
629 
630 	ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
631 	m->packet_type = ptype;
632 	if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
633 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
634 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
635 		l4_supported = 1;
636 
637 	if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
638 		hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
639 		if (hdr->csum_start <= hdrlen && l4_supported) {
640 			m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
641 		} else {
642 			/* Unknown proto or tunnel, do sw cksum. We can assume
643 			 * the cksum field is in the first segment since the
644 			 * buffers we provided to the host are large enough.
645 			 * In case of SCTP, this will be wrong since it's a CRC
646 			 * but there's nothing we can do.
647 			 */
648 			uint16_t csum = 0, off;
649 
650 			rte_raw_cksum_mbuf(m, hdr->csum_start,
651 				rte_pktmbuf_pkt_len(m) - hdr->csum_start,
652 				&csum);
653 			if (likely(csum != 0xffff))
654 				csum = ~csum;
655 			off = hdr->csum_offset + hdr->csum_start;
656 			if (rte_pktmbuf_data_len(m) >= off + 1)
657 				*rte_pktmbuf_mtod_offset(m, uint16_t *,
658 					off) = csum;
659 		}
660 	} else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
661 		m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
662 	}
663 
664 	/* GSO request, save required information in mbuf */
665 	if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
666 		/* Check unsupported modes */
667 		if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
668 		    (hdr->gso_size == 0)) {
669 			return -EINVAL;
670 		}
671 
672 		/* Update mss lengthes in mbuf */
673 		m->tso_segsz = hdr->gso_size;
674 		switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
675 			case VIRTIO_NET_HDR_GSO_TCPV4:
676 			case VIRTIO_NET_HDR_GSO_TCPV6:
677 				m->ol_flags |= PKT_RX_LRO | \
678 					PKT_RX_L4_CKSUM_NONE;
679 				break;
680 			default:
681 				return -EINVAL;
682 		}
683 	}
684 
685 	return 0;
686 }
687 
688 static inline int
689 rx_offload_enabled(struct virtio_hw *hw)
690 {
691 	return vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_CSUM) ||
692 		vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO4) ||
693 		vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO6);
694 }
695 
696 #define VIRTIO_MBUF_BURST_SZ 64
697 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
698 uint16_t
699 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
700 {
701 	struct virtnet_rx *rxvq = rx_queue;
702 	struct virtqueue *vq = rxvq->vq;
703 	struct virtio_hw *hw = vq->hw;
704 	struct rte_mbuf *rxm, *new_mbuf;
705 	uint16_t nb_used, num, nb_rx;
706 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
707 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
708 	int error;
709 	uint32_t i, nb_enqueued;
710 	uint32_t hdr_size;
711 	int offload;
712 	struct virtio_net_hdr *hdr;
713 
714 	nb_rx = 0;
715 	if (unlikely(hw->started == 0))
716 		return nb_rx;
717 
718 	nb_used = VIRTQUEUE_NUSED(vq);
719 
720 	virtio_rmb();
721 
722 	num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
723 	if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
724 		num = VIRTIO_MBUF_BURST_SZ;
725 	if (likely(num > DESC_PER_CACHELINE))
726 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
727 
728 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
729 	PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
730 
731 	nb_enqueued = 0;
732 	hdr_size = hw->vtnet_hdr_size;
733 	offload = rx_offload_enabled(hw);
734 
735 	for (i = 0; i < num ; i++) {
736 		rxm = rcv_pkts[i];
737 
738 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
739 
740 		if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
741 			PMD_RX_LOG(ERR, "Packet drop");
742 			nb_enqueued++;
743 			virtio_discard_rxbuf(vq, rxm);
744 			rxvq->stats.errors++;
745 			continue;
746 		}
747 
748 		rxm->port = rxvq->port_id;
749 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
750 		rxm->ol_flags = 0;
751 		rxm->vlan_tci = 0;
752 
753 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
754 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
755 
756 		hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
757 			RTE_PKTMBUF_HEADROOM - hdr_size);
758 
759 		if (hw->vlan_strip)
760 			rte_vlan_strip(rxm);
761 
762 		if (offload && virtio_rx_offload(rxm, hdr) < 0) {
763 			virtio_discard_rxbuf(vq, rxm);
764 			rxvq->stats.errors++;
765 			continue;
766 		}
767 
768 		VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
769 
770 		rx_pkts[nb_rx++] = rxm;
771 
772 		rxvq->stats.bytes += rxm->pkt_len;
773 		virtio_update_packet_stats(&rxvq->stats, rxm);
774 	}
775 
776 	rxvq->stats.packets += nb_rx;
777 
778 	/* Allocate new mbuf for the used descriptor */
779 	error = ENOSPC;
780 	while (likely(!virtqueue_full(vq))) {
781 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
782 		if (unlikely(new_mbuf == NULL)) {
783 			struct rte_eth_dev *dev
784 				= &rte_eth_devices[rxvq->port_id];
785 			dev->data->rx_mbuf_alloc_failed++;
786 			break;
787 		}
788 		error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
789 		if (unlikely(error)) {
790 			rte_pktmbuf_free(new_mbuf);
791 			break;
792 		}
793 		nb_enqueued++;
794 	}
795 
796 	if (likely(nb_enqueued)) {
797 		vq_update_avail_idx(vq);
798 
799 		if (unlikely(virtqueue_kick_prepare(vq))) {
800 			virtqueue_notify(vq);
801 			PMD_RX_LOG(DEBUG, "Notified");
802 		}
803 	}
804 
805 	return nb_rx;
806 }
807 
808 uint16_t
809 virtio_recv_mergeable_pkts(void *rx_queue,
810 			struct rte_mbuf **rx_pkts,
811 			uint16_t nb_pkts)
812 {
813 	struct virtnet_rx *rxvq = rx_queue;
814 	struct virtqueue *vq = rxvq->vq;
815 	struct virtio_hw *hw = vq->hw;
816 	struct rte_mbuf *rxm, *new_mbuf;
817 	uint16_t nb_used, num, nb_rx;
818 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
819 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
820 	struct rte_mbuf *prev;
821 	int error;
822 	uint32_t i, nb_enqueued;
823 	uint32_t seg_num;
824 	uint16_t extra_idx;
825 	uint32_t seg_res;
826 	uint32_t hdr_size;
827 	int offload;
828 
829 	nb_rx = 0;
830 	if (unlikely(hw->started == 0))
831 		return nb_rx;
832 
833 	nb_used = VIRTQUEUE_NUSED(vq);
834 
835 	virtio_rmb();
836 
837 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
838 
839 	i = 0;
840 	nb_enqueued = 0;
841 	seg_num = 0;
842 	extra_idx = 0;
843 	seg_res = 0;
844 	hdr_size = hw->vtnet_hdr_size;
845 	offload = rx_offload_enabled(hw);
846 
847 	while (i < nb_used) {
848 		struct virtio_net_hdr_mrg_rxbuf *header;
849 
850 		if (nb_rx == nb_pkts)
851 			break;
852 
853 		num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
854 		if (num != 1)
855 			continue;
856 
857 		i++;
858 
859 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
860 		PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
861 
862 		rxm = rcv_pkts[0];
863 
864 		if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
865 			PMD_RX_LOG(ERR, "Packet drop");
866 			nb_enqueued++;
867 			virtio_discard_rxbuf(vq, rxm);
868 			rxvq->stats.errors++;
869 			continue;
870 		}
871 
872 		header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
873 			RTE_PKTMBUF_HEADROOM - hdr_size);
874 		seg_num = header->num_buffers;
875 
876 		if (seg_num == 0)
877 			seg_num = 1;
878 
879 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
880 		rxm->nb_segs = seg_num;
881 		rxm->ol_flags = 0;
882 		rxm->vlan_tci = 0;
883 		rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
884 		rxm->data_len = (uint16_t)(len[0] - hdr_size);
885 
886 		rxm->port = rxvq->port_id;
887 		rx_pkts[nb_rx] = rxm;
888 		prev = rxm;
889 
890 		if (offload && virtio_rx_offload(rxm, &header->hdr) < 0) {
891 			virtio_discard_rxbuf(vq, rxm);
892 			rxvq->stats.errors++;
893 			continue;
894 		}
895 
896 		seg_res = seg_num - 1;
897 
898 		while (seg_res != 0) {
899 			/*
900 			 * Get extra segments for current uncompleted packet.
901 			 */
902 			uint16_t  rcv_cnt =
903 				RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
904 			if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
905 				uint32_t rx_num =
906 					virtqueue_dequeue_burst_rx(vq,
907 					rcv_pkts, len, rcv_cnt);
908 				i += rx_num;
909 				rcv_cnt = rx_num;
910 			} else {
911 				PMD_RX_LOG(ERR,
912 					   "No enough segments for packet.");
913 				nb_enqueued++;
914 				virtio_discard_rxbuf(vq, rxm);
915 				rxvq->stats.errors++;
916 				break;
917 			}
918 
919 			extra_idx = 0;
920 
921 			while (extra_idx < rcv_cnt) {
922 				rxm = rcv_pkts[extra_idx];
923 
924 				rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
925 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
926 				rxm->data_len = (uint16_t)(len[extra_idx]);
927 
928 				if (prev)
929 					prev->next = rxm;
930 
931 				prev = rxm;
932 				rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
933 				extra_idx++;
934 			};
935 			seg_res -= rcv_cnt;
936 		}
937 
938 		if (hw->vlan_strip)
939 			rte_vlan_strip(rx_pkts[nb_rx]);
940 
941 		VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
942 			rx_pkts[nb_rx]->data_len);
943 
944 		rxvq->stats.bytes += rx_pkts[nb_rx]->pkt_len;
945 		virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
946 		nb_rx++;
947 	}
948 
949 	rxvq->stats.packets += nb_rx;
950 
951 	/* Allocate new mbuf for the used descriptor */
952 	error = ENOSPC;
953 	while (likely(!virtqueue_full(vq))) {
954 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
955 		if (unlikely(new_mbuf == NULL)) {
956 			struct rte_eth_dev *dev
957 				= &rte_eth_devices[rxvq->port_id];
958 			dev->data->rx_mbuf_alloc_failed++;
959 			break;
960 		}
961 		error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
962 		if (unlikely(error)) {
963 			rte_pktmbuf_free(new_mbuf);
964 			break;
965 		}
966 		nb_enqueued++;
967 	}
968 
969 	if (likely(nb_enqueued)) {
970 		vq_update_avail_idx(vq);
971 
972 		if (unlikely(virtqueue_kick_prepare(vq))) {
973 			virtqueue_notify(vq);
974 			PMD_RX_LOG(DEBUG, "Notified");
975 		}
976 	}
977 
978 	return nb_rx;
979 }
980 
981 uint16_t
982 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
983 {
984 	struct virtnet_tx *txvq = tx_queue;
985 	struct virtqueue *vq = txvq->vq;
986 	struct virtio_hw *hw = vq->hw;
987 	uint16_t hdr_size = hw->vtnet_hdr_size;
988 	uint16_t nb_used, nb_tx = 0;
989 	int error;
990 
991 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
992 		return nb_tx;
993 
994 	if (unlikely(nb_pkts < 1))
995 		return nb_pkts;
996 
997 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
998 	nb_used = VIRTQUEUE_NUSED(vq);
999 
1000 	virtio_rmb();
1001 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1002 		virtio_xmit_cleanup(vq, nb_used);
1003 
1004 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1005 		struct rte_mbuf *txm = tx_pkts[nb_tx];
1006 		int can_push = 0, use_indirect = 0, slots, need;
1007 
1008 		/* Do VLAN tag insertion */
1009 		if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1010 			error = rte_vlan_insert(&txm);
1011 			if (unlikely(error)) {
1012 				rte_pktmbuf_free(txm);
1013 				continue;
1014 			}
1015 		}
1016 
1017 		/* optimize ring usage */
1018 		if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1019 		      vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1020 		    rte_mbuf_refcnt_read(txm) == 1 &&
1021 		    RTE_MBUF_DIRECT(txm) &&
1022 		    txm->nb_segs == 1 &&
1023 		    rte_pktmbuf_headroom(txm) >= hdr_size &&
1024 		    rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1025 				   __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1026 			can_push = 1;
1027 		else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1028 			 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1029 			use_indirect = 1;
1030 
1031 		/* How many main ring entries are needed to this Tx?
1032 		 * any_layout => number of segments
1033 		 * indirect   => 1
1034 		 * default    => number of segments + 1
1035 		 */
1036 		slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1037 		need = slots - vq->vq_free_cnt;
1038 
1039 		/* Positive value indicates it need free vring descriptors */
1040 		if (unlikely(need > 0)) {
1041 			nb_used = VIRTQUEUE_NUSED(vq);
1042 			virtio_rmb();
1043 			need = RTE_MIN(need, (int)nb_used);
1044 
1045 			virtio_xmit_cleanup(vq, need);
1046 			need = slots - vq->vq_free_cnt;
1047 			if (unlikely(need > 0)) {
1048 				PMD_TX_LOG(ERR,
1049 					   "No free tx descriptors to transmit");
1050 				break;
1051 			}
1052 		}
1053 
1054 		/* Enqueue Packet buffers */
1055 		virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
1056 
1057 		txvq->stats.bytes += txm->pkt_len;
1058 		virtio_update_packet_stats(&txvq->stats, txm);
1059 	}
1060 
1061 	txvq->stats.packets += nb_tx;
1062 
1063 	if (likely(nb_tx)) {
1064 		vq_update_avail_idx(vq);
1065 
1066 		if (unlikely(virtqueue_kick_prepare(vq))) {
1067 			virtqueue_notify(vq);
1068 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1069 		}
1070 	}
1071 
1072 	return nb_tx;
1073 }
1074