xref: /dpdk/drivers/net/virtio/virtio_rxtx.c (revision d429cc0b53735cc7b1e304ec1d0f35ae06ace7d0)
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 
35 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
36 #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
37 #else
38 #define  VIRTIO_DUMP_PACKET(m, len) do { } while (0)
39 #endif
40 
41 
42 #define VIRTIO_SIMPLE_FLAGS ((uint32_t)ETH_TXQ_FLAGS_NOMULTSEGS | \
43 	ETH_TXQ_FLAGS_NOOFFLOADS)
44 
45 int
46 virtio_dev_rx_queue_done(void *rxq, uint16_t offset)
47 {
48 	struct virtnet_rx *rxvq = rxq;
49 	struct virtqueue *vq = rxvq->vq;
50 
51 	return VIRTQUEUE_NUSED(vq) >= offset;
52 }
53 
54 void
55 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
56 {
57 	struct vring_desc *dp, *dp_tail;
58 	struct vq_desc_extra *dxp;
59 	uint16_t desc_idx_last = desc_idx;
60 
61 	dp  = &vq->vq_ring.desc[desc_idx];
62 	dxp = &vq->vq_descx[desc_idx];
63 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
64 	if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
65 		while (dp->flags & VRING_DESC_F_NEXT) {
66 			desc_idx_last = dp->next;
67 			dp = &vq->vq_ring.desc[dp->next];
68 		}
69 	}
70 	dxp->ndescs = 0;
71 
72 	/*
73 	 * We must append the existing free chain, if any, to the end of
74 	 * newly freed chain. If the virtqueue was completely used, then
75 	 * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
76 	 */
77 	if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
78 		vq->vq_desc_head_idx = desc_idx;
79 	} else {
80 		dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
81 		dp_tail->next = desc_idx;
82 	}
83 
84 	vq->vq_desc_tail_idx = desc_idx_last;
85 	dp->next = VQ_RING_DESC_CHAIN_END;
86 }
87 
88 static uint16_t
89 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
90 			   uint32_t *len, uint16_t num)
91 {
92 	struct vring_used_elem *uep;
93 	struct rte_mbuf *cookie;
94 	uint16_t used_idx, desc_idx;
95 	uint16_t i;
96 
97 	/*  Caller does the check */
98 	for (i = 0; i < num ; i++) {
99 		used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
100 		uep = &vq->vq_ring.used->ring[used_idx];
101 		desc_idx = (uint16_t) uep->id;
102 		len[i] = uep->len;
103 		cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
104 
105 		if (unlikely(cookie == NULL)) {
106 			PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
107 				vq->vq_used_cons_idx);
108 			break;
109 		}
110 
111 		rte_prefetch0(cookie);
112 		rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
113 		rx_pkts[i]  = cookie;
114 		vq->vq_used_cons_idx++;
115 		vq_ring_free_chain(vq, desc_idx);
116 		vq->vq_descx[desc_idx].cookie = NULL;
117 	}
118 
119 	return i;
120 }
121 
122 #ifndef DEFAULT_TX_FREE_THRESH
123 #define DEFAULT_TX_FREE_THRESH 32
124 #endif
125 
126 /* Cleanup from completed transmits. */
127 static void
128 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
129 {
130 	uint16_t i, used_idx, desc_idx;
131 	for (i = 0; i < num; i++) {
132 		struct vring_used_elem *uep;
133 		struct vq_desc_extra *dxp;
134 
135 		used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
136 		uep = &vq->vq_ring.used->ring[used_idx];
137 
138 		desc_idx = (uint16_t) uep->id;
139 		dxp = &vq->vq_descx[desc_idx];
140 		vq->vq_used_cons_idx++;
141 		vq_ring_free_chain(vq, desc_idx);
142 
143 		if (dxp->cookie != NULL) {
144 			rte_pktmbuf_free(dxp->cookie);
145 			dxp->cookie = NULL;
146 		}
147 	}
148 }
149 
150 
151 static inline int
152 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
153 {
154 	struct vq_desc_extra *dxp;
155 	struct virtio_hw *hw = vq->hw;
156 	struct vring_desc *start_dp;
157 	uint16_t needed = 1;
158 	uint16_t head_idx, idx;
159 
160 	if (unlikely(vq->vq_free_cnt == 0))
161 		return -ENOSPC;
162 	if (unlikely(vq->vq_free_cnt < needed))
163 		return -EMSGSIZE;
164 
165 	head_idx = vq->vq_desc_head_idx;
166 	if (unlikely(head_idx >= vq->vq_nentries))
167 		return -EFAULT;
168 
169 	idx = head_idx;
170 	dxp = &vq->vq_descx[idx];
171 	dxp->cookie = (void *)cookie;
172 	dxp->ndescs = needed;
173 
174 	start_dp = vq->vq_ring.desc;
175 	start_dp[idx].addr =
176 		VIRTIO_MBUF_ADDR(cookie, vq) +
177 		RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
178 	start_dp[idx].len =
179 		cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
180 	start_dp[idx].flags =  VRING_DESC_F_WRITE;
181 	idx = start_dp[idx].next;
182 	vq->vq_desc_head_idx = idx;
183 	if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
184 		vq->vq_desc_tail_idx = idx;
185 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
186 	vq_update_avail_ring(vq, head_idx);
187 
188 	return 0;
189 }
190 
191 /* When doing TSO, the IP length is not included in the pseudo header
192  * checksum of the packet given to the PMD, but for virtio it is
193  * expected.
194  */
195 static void
196 virtio_tso_fix_cksum(struct rte_mbuf *m)
197 {
198 	/* common case: header is not fragmented */
199 	if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
200 			m->l4_len)) {
201 		struct ipv4_hdr *iph;
202 		struct ipv6_hdr *ip6h;
203 		struct tcp_hdr *th;
204 		uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
205 		uint32_t tmp;
206 
207 		iph = rte_pktmbuf_mtod_offset(m, struct ipv4_hdr *, m->l2_len);
208 		th = RTE_PTR_ADD(iph, m->l3_len);
209 		if ((iph->version_ihl >> 4) == 4) {
210 			iph->hdr_checksum = 0;
211 			iph->hdr_checksum = rte_ipv4_cksum(iph);
212 			ip_len = iph->total_length;
213 			ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
214 				m->l3_len);
215 		} else {
216 			ip6h = (struct ipv6_hdr *)iph;
217 			ip_paylen = ip6h->payload_len;
218 		}
219 
220 		/* calculate the new phdr checksum not including ip_paylen */
221 		prev_cksum = th->cksum;
222 		tmp = prev_cksum;
223 		tmp += ip_paylen;
224 		tmp = (tmp & 0xffff) + (tmp >> 16);
225 		new_cksum = tmp;
226 
227 		/* replace it in the packet */
228 		th->cksum = new_cksum;
229 	}
230 }
231 
232 static inline int
233 tx_offload_enabled(struct virtio_hw *hw)
234 {
235 	return vtpci_with_feature(hw, VIRTIO_NET_F_CSUM) ||
236 		vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO4) ||
237 		vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO6);
238 }
239 
240 /* avoid write operation when necessary, to lessen cache issues */
241 #define ASSIGN_UNLESS_EQUAL(var, val) do {	\
242 	if ((var) != (val))			\
243 		(var) = (val);			\
244 } while (0)
245 
246 static inline void
247 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
248 		       uint16_t needed, int use_indirect, int can_push)
249 {
250 	struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
251 	struct vq_desc_extra *dxp;
252 	struct virtqueue *vq = txvq->vq;
253 	struct vring_desc *start_dp;
254 	uint16_t seg_num = cookie->nb_segs;
255 	uint16_t head_idx, idx;
256 	uint16_t head_size = vq->hw->vtnet_hdr_size;
257 	struct virtio_net_hdr *hdr;
258 	int offload;
259 
260 	offload = tx_offload_enabled(vq->hw);
261 	head_idx = vq->vq_desc_head_idx;
262 	idx = head_idx;
263 	dxp = &vq->vq_descx[idx];
264 	dxp->cookie = (void *)cookie;
265 	dxp->ndescs = needed;
266 
267 	start_dp = vq->vq_ring.desc;
268 
269 	if (can_push) {
270 		/* prepend cannot fail, checked by caller */
271 		hdr = (struct virtio_net_hdr *)
272 			rte_pktmbuf_prepend(cookie, head_size);
273 		/* rte_pktmbuf_prepend() counts the hdr size to the pkt length,
274 		 * which is wrong. Below subtract restores correct pkt size.
275 		 */
276 		cookie->pkt_len -= head_size;
277 		/* if offload disabled, it is not zeroed below, do it now */
278 		if (offload == 0) {
279 			ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
280 			ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
281 			ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
282 			ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
283 			ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
284 			ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
285 		}
286 	} else if (use_indirect) {
287 		/* setup tx ring slot to point to indirect
288 		 * descriptor list stored in reserved region.
289 		 *
290 		 * the first slot in indirect ring is already preset
291 		 * to point to the header in reserved region
292 		 */
293 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
294 			RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
295 		start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
296 		start_dp[idx].flags = VRING_DESC_F_INDIRECT;
297 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
298 
299 		/* loop below will fill in rest of the indirect elements */
300 		start_dp = txr[idx].tx_indir;
301 		idx = 1;
302 	} else {
303 		/* setup first tx ring slot to point to header
304 		 * stored in reserved region.
305 		 */
306 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
307 			RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
308 		start_dp[idx].len   = vq->hw->vtnet_hdr_size;
309 		start_dp[idx].flags = VRING_DESC_F_NEXT;
310 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
311 
312 		idx = start_dp[idx].next;
313 	}
314 
315 	/* Checksum Offload / TSO */
316 	if (offload) {
317 		if (cookie->ol_flags & PKT_TX_TCP_SEG)
318 			cookie->ol_flags |= PKT_TX_TCP_CKSUM;
319 
320 		switch (cookie->ol_flags & PKT_TX_L4_MASK) {
321 		case PKT_TX_UDP_CKSUM:
322 			hdr->csum_start = cookie->l2_len + cookie->l3_len;
323 			hdr->csum_offset = offsetof(struct udp_hdr,
324 				dgram_cksum);
325 			hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
326 			break;
327 
328 		case PKT_TX_TCP_CKSUM:
329 			hdr->csum_start = cookie->l2_len + cookie->l3_len;
330 			hdr->csum_offset = offsetof(struct tcp_hdr, cksum);
331 			hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
332 			break;
333 
334 		default:
335 			ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
336 			ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
337 			ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
338 			break;
339 		}
340 
341 		/* TCP Segmentation Offload */
342 		if (cookie->ol_flags & PKT_TX_TCP_SEG) {
343 			virtio_tso_fix_cksum(cookie);
344 			hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
345 				VIRTIO_NET_HDR_GSO_TCPV6 :
346 				VIRTIO_NET_HDR_GSO_TCPV4;
347 			hdr->gso_size = cookie->tso_segsz;
348 			hdr->hdr_len =
349 				cookie->l2_len +
350 				cookie->l3_len +
351 				cookie->l4_len;
352 		} else {
353 			ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
354 			ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
355 			ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
356 		}
357 	}
358 
359 	do {
360 		start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
361 		start_dp[idx].len   = cookie->data_len;
362 		start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
363 		idx = start_dp[idx].next;
364 	} while ((cookie = cookie->next) != NULL);
365 
366 	if (use_indirect)
367 		idx = vq->vq_ring.desc[head_idx].next;
368 
369 	vq->vq_desc_head_idx = idx;
370 	if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
371 		vq->vq_desc_tail_idx = idx;
372 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
373 	vq_update_avail_ring(vq, head_idx);
374 }
375 
376 void
377 virtio_dev_cq_start(struct rte_eth_dev *dev)
378 {
379 	struct virtio_hw *hw = dev->data->dev_private;
380 
381 	if (hw->cvq && hw->cvq->vq) {
382 		rte_spinlock_init(&hw->cvq->lock);
383 		VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
384 	}
385 }
386 
387 int
388 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
389 			uint16_t queue_idx,
390 			uint16_t nb_desc,
391 			unsigned int socket_id __rte_unused,
392 			__rte_unused const struct rte_eth_rxconf *rx_conf,
393 			struct rte_mempool *mp)
394 {
395 	uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
396 	struct virtio_hw *hw = dev->data->dev_private;
397 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
398 	struct virtnet_rx *rxvq;
399 
400 	PMD_INIT_FUNC_TRACE();
401 
402 	if (nb_desc == 0 || nb_desc > vq->vq_nentries)
403 		nb_desc = vq->vq_nentries;
404 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
405 
406 	rxvq = &vq->rxq;
407 	rxvq->queue_id = queue_idx;
408 	rxvq->mpool = mp;
409 	if (rxvq->mpool == NULL) {
410 		rte_exit(EXIT_FAILURE,
411 			"Cannot allocate mbufs for rx virtqueue");
412 	}
413 	dev->data->rx_queues[queue_idx] = rxvq;
414 
415 	return 0;
416 }
417 
418 int
419 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
420 {
421 	uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
422 	struct virtio_hw *hw = dev->data->dev_private;
423 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
424 	struct virtnet_rx *rxvq = &vq->rxq;
425 	struct rte_mbuf *m;
426 	uint16_t desc_idx;
427 	int error, nbufs;
428 
429 	PMD_INIT_FUNC_TRACE();
430 
431 	/* Allocate blank mbufs for the each rx descriptor */
432 	nbufs = 0;
433 
434 	if (hw->use_simple_rx) {
435 		for (desc_idx = 0; desc_idx < vq->vq_nentries;
436 		     desc_idx++) {
437 			vq->vq_ring.avail->ring[desc_idx] = desc_idx;
438 			vq->vq_ring.desc[desc_idx].flags =
439 				VRING_DESC_F_WRITE;
440 		}
441 
442 		virtio_rxq_vec_setup(rxvq);
443 	}
444 
445 	memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
446 	for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
447 	     desc_idx++) {
448 		vq->sw_ring[vq->vq_nentries + desc_idx] =
449 			&rxvq->fake_mbuf;
450 	}
451 
452 	if (hw->use_simple_rx) {
453 		while (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
454 			virtio_rxq_rearm_vec(rxvq);
455 			nbufs += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
456 		}
457 	} else {
458 		while (!virtqueue_full(vq)) {
459 			m = rte_mbuf_raw_alloc(rxvq->mpool);
460 			if (m == NULL)
461 				break;
462 
463 			/* Enqueue allocated buffers */
464 			error = virtqueue_enqueue_recv_refill(vq, m);
465 			if (error) {
466 				rte_pktmbuf_free(m);
467 				break;
468 			}
469 			nbufs++;
470 		}
471 
472 		vq_update_avail_idx(vq);
473 	}
474 
475 	PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
476 
477 	VIRTQUEUE_DUMP(vq);
478 
479 	return 0;
480 }
481 
482 /*
483  * struct rte_eth_dev *dev: Used to update dev
484  * uint16_t nb_desc: Defaults to values read from config space
485  * unsigned int socket_id: Used to allocate memzone
486  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
487  * uint16_t queue_idx: Just used as an index in dev txq list
488  */
489 int
490 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
491 			uint16_t queue_idx,
492 			uint16_t nb_desc,
493 			unsigned int socket_id __rte_unused,
494 			const struct rte_eth_txconf *tx_conf)
495 {
496 	uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
497 	struct virtio_hw *hw = dev->data->dev_private;
498 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
499 	struct virtnet_tx *txvq;
500 	uint16_t tx_free_thresh;
501 
502 	PMD_INIT_FUNC_TRACE();
503 
504 	/* cannot use simple rxtx funcs with multisegs or offloads */
505 	if ((tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) != VIRTIO_SIMPLE_FLAGS)
506 		hw->use_simple_tx = 0;
507 
508 	if (nb_desc == 0 || nb_desc > vq->vq_nentries)
509 		nb_desc = vq->vq_nentries;
510 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
511 
512 	txvq = &vq->txq;
513 	txvq->queue_id = queue_idx;
514 
515 	tx_free_thresh = tx_conf->tx_free_thresh;
516 	if (tx_free_thresh == 0)
517 		tx_free_thresh =
518 			RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
519 
520 	if (tx_free_thresh >= (vq->vq_nentries - 3)) {
521 		RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
522 			"number of TX entries minus 3 (%u)."
523 			" (tx_free_thresh=%u port=%u queue=%u)\n",
524 			vq->vq_nentries - 3,
525 			tx_free_thresh, dev->data->port_id, queue_idx);
526 		return -EINVAL;
527 	}
528 
529 	vq->vq_free_thresh = tx_free_thresh;
530 
531 	dev->data->tx_queues[queue_idx] = txvq;
532 	return 0;
533 }
534 
535 int
536 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
537 				uint16_t queue_idx)
538 {
539 	uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
540 	struct virtio_hw *hw = dev->data->dev_private;
541 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
542 	uint16_t mid_idx = vq->vq_nentries >> 1;
543 	struct virtnet_tx *txvq = &vq->txq;
544 	uint16_t desc_idx;
545 
546 	PMD_INIT_FUNC_TRACE();
547 
548 	if (hw->use_simple_tx) {
549 		for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
550 			vq->vq_ring.avail->ring[desc_idx] =
551 				desc_idx + mid_idx;
552 			vq->vq_ring.desc[desc_idx + mid_idx].next =
553 				desc_idx;
554 			vq->vq_ring.desc[desc_idx + mid_idx].addr =
555 				txvq->virtio_net_hdr_mem +
556 				offsetof(struct virtio_tx_region, tx_hdr);
557 			vq->vq_ring.desc[desc_idx + mid_idx].len =
558 				vq->hw->vtnet_hdr_size;
559 			vq->vq_ring.desc[desc_idx + mid_idx].flags =
560 				VRING_DESC_F_NEXT;
561 			vq->vq_ring.desc[desc_idx].flags = 0;
562 		}
563 		for (desc_idx = mid_idx; desc_idx < vq->vq_nentries;
564 		     desc_idx++)
565 			vq->vq_ring.avail->ring[desc_idx] = desc_idx;
566 	}
567 
568 	VIRTQUEUE_DUMP(vq);
569 
570 	return 0;
571 }
572 
573 static void
574 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
575 {
576 	int error;
577 	/*
578 	 * Requeue the discarded mbuf. This should always be
579 	 * successful since it was just dequeued.
580 	 */
581 	error = virtqueue_enqueue_recv_refill(vq, m);
582 	if (unlikely(error)) {
583 		RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
584 		rte_pktmbuf_free(m);
585 	}
586 }
587 
588 static void
589 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
590 {
591 	uint32_t s = mbuf->pkt_len;
592 	struct ether_addr *ea;
593 
594 	if (s == 64) {
595 		stats->size_bins[1]++;
596 	} else if (s > 64 && s < 1024) {
597 		uint32_t bin;
598 
599 		/* count zeros, and offset into correct bin */
600 		bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
601 		stats->size_bins[bin]++;
602 	} else {
603 		if (s < 64)
604 			stats->size_bins[0]++;
605 		else if (s < 1519)
606 			stats->size_bins[6]++;
607 		else if (s >= 1519)
608 			stats->size_bins[7]++;
609 	}
610 
611 	ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
612 	if (is_multicast_ether_addr(ea)) {
613 		if (is_broadcast_ether_addr(ea))
614 			stats->broadcast++;
615 		else
616 			stats->multicast++;
617 	}
618 }
619 
620 /* Optionally fill offload information in structure */
621 static int
622 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
623 {
624 	struct rte_net_hdr_lens hdr_lens;
625 	uint32_t hdrlen, ptype;
626 	int l4_supported = 0;
627 
628 	/* nothing to do */
629 	if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
630 		return 0;
631 
632 	m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
633 
634 	ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
635 	m->packet_type = ptype;
636 	if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
637 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
638 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
639 		l4_supported = 1;
640 
641 	if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
642 		hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
643 		if (hdr->csum_start <= hdrlen && l4_supported) {
644 			m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
645 		} else {
646 			/* Unknown proto or tunnel, do sw cksum. We can assume
647 			 * the cksum field is in the first segment since the
648 			 * buffers we provided to the host are large enough.
649 			 * In case of SCTP, this will be wrong since it's a CRC
650 			 * but there's nothing we can do.
651 			 */
652 			uint16_t csum = 0, off;
653 
654 			rte_raw_cksum_mbuf(m, hdr->csum_start,
655 				rte_pktmbuf_pkt_len(m) - hdr->csum_start,
656 				&csum);
657 			if (likely(csum != 0xffff))
658 				csum = ~csum;
659 			off = hdr->csum_offset + hdr->csum_start;
660 			if (rte_pktmbuf_data_len(m) >= off + 1)
661 				*rte_pktmbuf_mtod_offset(m, uint16_t *,
662 					off) = csum;
663 		}
664 	} else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
665 		m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
666 	}
667 
668 	/* GSO request, save required information in mbuf */
669 	if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
670 		/* Check unsupported modes */
671 		if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
672 		    (hdr->gso_size == 0)) {
673 			return -EINVAL;
674 		}
675 
676 		/* Update mss lengthes in mbuf */
677 		m->tso_segsz = hdr->gso_size;
678 		switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
679 			case VIRTIO_NET_HDR_GSO_TCPV4:
680 			case VIRTIO_NET_HDR_GSO_TCPV6:
681 				m->ol_flags |= PKT_RX_LRO | \
682 					PKT_RX_L4_CKSUM_NONE;
683 				break;
684 			default:
685 				return -EINVAL;
686 		}
687 	}
688 
689 	return 0;
690 }
691 
692 static inline int
693 rx_offload_enabled(struct virtio_hw *hw)
694 {
695 	return vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_CSUM) ||
696 		vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO4) ||
697 		vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO6);
698 }
699 
700 #define VIRTIO_MBUF_BURST_SZ 64
701 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
702 uint16_t
703 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
704 {
705 	struct virtnet_rx *rxvq = rx_queue;
706 	struct virtqueue *vq = rxvq->vq;
707 	struct virtio_hw *hw = vq->hw;
708 	struct rte_mbuf *rxm, *new_mbuf;
709 	uint16_t nb_used, num, nb_rx;
710 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
711 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
712 	int error;
713 	uint32_t i, nb_enqueued;
714 	uint32_t hdr_size;
715 	int offload;
716 	struct virtio_net_hdr *hdr;
717 
718 	nb_rx = 0;
719 	if (unlikely(hw->started == 0))
720 		return nb_rx;
721 
722 	nb_used = VIRTQUEUE_NUSED(vq);
723 
724 	virtio_rmb();
725 
726 	num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
727 	if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
728 		num = VIRTIO_MBUF_BURST_SZ;
729 	if (likely(num > DESC_PER_CACHELINE))
730 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
731 
732 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
733 	PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
734 
735 	nb_enqueued = 0;
736 	hdr_size = hw->vtnet_hdr_size;
737 	offload = rx_offload_enabled(hw);
738 
739 	for (i = 0; i < num ; i++) {
740 		rxm = rcv_pkts[i];
741 
742 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
743 
744 		if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
745 			PMD_RX_LOG(ERR, "Packet drop");
746 			nb_enqueued++;
747 			virtio_discard_rxbuf(vq, rxm);
748 			rxvq->stats.errors++;
749 			continue;
750 		}
751 
752 		rxm->port = rxvq->port_id;
753 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
754 		rxm->ol_flags = 0;
755 		rxm->vlan_tci = 0;
756 
757 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
758 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
759 
760 		hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
761 			RTE_PKTMBUF_HEADROOM - hdr_size);
762 
763 		if (hw->vlan_strip)
764 			rte_vlan_strip(rxm);
765 
766 		if (offload && virtio_rx_offload(rxm, hdr) < 0) {
767 			virtio_discard_rxbuf(vq, rxm);
768 			rxvq->stats.errors++;
769 			continue;
770 		}
771 
772 		VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
773 
774 		rx_pkts[nb_rx++] = rxm;
775 
776 		rxvq->stats.bytes += rxm->pkt_len;
777 		virtio_update_packet_stats(&rxvq->stats, rxm);
778 	}
779 
780 	rxvq->stats.packets += nb_rx;
781 
782 	/* Allocate new mbuf for the used descriptor */
783 	error = ENOSPC;
784 	while (likely(!virtqueue_full(vq))) {
785 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
786 		if (unlikely(new_mbuf == NULL)) {
787 			struct rte_eth_dev *dev
788 				= &rte_eth_devices[rxvq->port_id];
789 			dev->data->rx_mbuf_alloc_failed++;
790 			break;
791 		}
792 		error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
793 		if (unlikely(error)) {
794 			rte_pktmbuf_free(new_mbuf);
795 			break;
796 		}
797 		nb_enqueued++;
798 	}
799 
800 	if (likely(nb_enqueued)) {
801 		vq_update_avail_idx(vq);
802 
803 		if (unlikely(virtqueue_kick_prepare(vq))) {
804 			virtqueue_notify(vq);
805 			PMD_RX_LOG(DEBUG, "Notified");
806 		}
807 	}
808 
809 	return nb_rx;
810 }
811 
812 uint16_t
813 virtio_recv_mergeable_pkts(void *rx_queue,
814 			struct rte_mbuf **rx_pkts,
815 			uint16_t nb_pkts)
816 {
817 	struct virtnet_rx *rxvq = rx_queue;
818 	struct virtqueue *vq = rxvq->vq;
819 	struct virtio_hw *hw = vq->hw;
820 	struct rte_mbuf *rxm, *new_mbuf;
821 	uint16_t nb_used, num, nb_rx;
822 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
823 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
824 	struct rte_mbuf *prev;
825 	int error;
826 	uint32_t i, nb_enqueued;
827 	uint32_t seg_num;
828 	uint16_t extra_idx;
829 	uint32_t seg_res;
830 	uint32_t hdr_size;
831 	int offload;
832 
833 	nb_rx = 0;
834 	if (unlikely(hw->started == 0))
835 		return nb_rx;
836 
837 	nb_used = VIRTQUEUE_NUSED(vq);
838 
839 	virtio_rmb();
840 
841 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
842 
843 	i = 0;
844 	nb_enqueued = 0;
845 	seg_num = 0;
846 	extra_idx = 0;
847 	seg_res = 0;
848 	hdr_size = hw->vtnet_hdr_size;
849 	offload = rx_offload_enabled(hw);
850 
851 	while (i < nb_used) {
852 		struct virtio_net_hdr_mrg_rxbuf *header;
853 
854 		if (nb_rx == nb_pkts)
855 			break;
856 
857 		num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
858 		if (num != 1)
859 			continue;
860 
861 		i++;
862 
863 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
864 		PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
865 
866 		rxm = rcv_pkts[0];
867 
868 		if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
869 			PMD_RX_LOG(ERR, "Packet drop");
870 			nb_enqueued++;
871 			virtio_discard_rxbuf(vq, rxm);
872 			rxvq->stats.errors++;
873 			continue;
874 		}
875 
876 		header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
877 			RTE_PKTMBUF_HEADROOM - hdr_size);
878 		seg_num = header->num_buffers;
879 
880 		if (seg_num == 0)
881 			seg_num = 1;
882 
883 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
884 		rxm->nb_segs = seg_num;
885 		rxm->ol_flags = 0;
886 		rxm->vlan_tci = 0;
887 		rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
888 		rxm->data_len = (uint16_t)(len[0] - hdr_size);
889 
890 		rxm->port = rxvq->port_id;
891 		rx_pkts[nb_rx] = rxm;
892 		prev = rxm;
893 
894 		if (offload && virtio_rx_offload(rxm, &header->hdr) < 0) {
895 			virtio_discard_rxbuf(vq, rxm);
896 			rxvq->stats.errors++;
897 			continue;
898 		}
899 
900 		seg_res = seg_num - 1;
901 
902 		while (seg_res != 0) {
903 			/*
904 			 * Get extra segments for current uncompleted packet.
905 			 */
906 			uint16_t  rcv_cnt =
907 				RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
908 			if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
909 				uint32_t rx_num =
910 					virtqueue_dequeue_burst_rx(vq,
911 					rcv_pkts, len, rcv_cnt);
912 				i += rx_num;
913 				rcv_cnt = rx_num;
914 			} else {
915 				PMD_RX_LOG(ERR,
916 					   "No enough segments for packet.");
917 				nb_enqueued++;
918 				virtio_discard_rxbuf(vq, rxm);
919 				rxvq->stats.errors++;
920 				break;
921 			}
922 
923 			extra_idx = 0;
924 
925 			while (extra_idx < rcv_cnt) {
926 				rxm = rcv_pkts[extra_idx];
927 
928 				rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
929 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
930 				rxm->data_len = (uint16_t)(len[extra_idx]);
931 
932 				if (prev)
933 					prev->next = rxm;
934 
935 				prev = rxm;
936 				rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
937 				extra_idx++;
938 			};
939 			seg_res -= rcv_cnt;
940 		}
941 
942 		if (hw->vlan_strip)
943 			rte_vlan_strip(rx_pkts[nb_rx]);
944 
945 		VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
946 			rx_pkts[nb_rx]->data_len);
947 
948 		rxvq->stats.bytes += rx_pkts[nb_rx]->pkt_len;
949 		virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
950 		nb_rx++;
951 	}
952 
953 	rxvq->stats.packets += nb_rx;
954 
955 	/* Allocate new mbuf for the used descriptor */
956 	error = ENOSPC;
957 	while (likely(!virtqueue_full(vq))) {
958 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
959 		if (unlikely(new_mbuf == NULL)) {
960 			struct rte_eth_dev *dev
961 				= &rte_eth_devices[rxvq->port_id];
962 			dev->data->rx_mbuf_alloc_failed++;
963 			break;
964 		}
965 		error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
966 		if (unlikely(error)) {
967 			rte_pktmbuf_free(new_mbuf);
968 			break;
969 		}
970 		nb_enqueued++;
971 	}
972 
973 	if (likely(nb_enqueued)) {
974 		vq_update_avail_idx(vq);
975 
976 		if (unlikely(virtqueue_kick_prepare(vq))) {
977 			virtqueue_notify(vq);
978 			PMD_RX_LOG(DEBUG, "Notified");
979 		}
980 	}
981 
982 	return nb_rx;
983 }
984 
985 uint16_t
986 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
987 {
988 	struct virtnet_tx *txvq = tx_queue;
989 	struct virtqueue *vq = txvq->vq;
990 	struct virtio_hw *hw = vq->hw;
991 	uint16_t hdr_size = hw->vtnet_hdr_size;
992 	uint16_t nb_used, nb_tx = 0;
993 	int error;
994 
995 	if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
996 		return nb_tx;
997 
998 	if (unlikely(nb_pkts < 1))
999 		return nb_pkts;
1000 
1001 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1002 	nb_used = VIRTQUEUE_NUSED(vq);
1003 
1004 	virtio_rmb();
1005 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1006 		virtio_xmit_cleanup(vq, nb_used);
1007 
1008 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1009 		struct rte_mbuf *txm = tx_pkts[nb_tx];
1010 		int can_push = 0, use_indirect = 0, slots, need;
1011 
1012 		/* Do VLAN tag insertion */
1013 		if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1014 			error = rte_vlan_insert(&txm);
1015 			if (unlikely(error)) {
1016 				rte_pktmbuf_free(txm);
1017 				continue;
1018 			}
1019 		}
1020 
1021 		/* optimize ring usage */
1022 		if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1023 		      vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1024 		    rte_mbuf_refcnt_read(txm) == 1 &&
1025 		    RTE_MBUF_DIRECT(txm) &&
1026 		    txm->nb_segs == 1 &&
1027 		    rte_pktmbuf_headroom(txm) >= hdr_size &&
1028 		    rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1029 				   __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1030 			can_push = 1;
1031 		else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1032 			 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1033 			use_indirect = 1;
1034 
1035 		/* How many main ring entries are needed to this Tx?
1036 		 * any_layout => number of segments
1037 		 * indirect   => 1
1038 		 * default    => number of segments + 1
1039 		 */
1040 		slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1041 		need = slots - vq->vq_free_cnt;
1042 
1043 		/* Positive value indicates it need free vring descriptors */
1044 		if (unlikely(need > 0)) {
1045 			nb_used = VIRTQUEUE_NUSED(vq);
1046 			virtio_rmb();
1047 			need = RTE_MIN(need, (int)nb_used);
1048 
1049 			virtio_xmit_cleanup(vq, need);
1050 			need = slots - vq->vq_free_cnt;
1051 			if (unlikely(need > 0)) {
1052 				PMD_TX_LOG(ERR,
1053 					   "No free tx descriptors to transmit");
1054 				break;
1055 			}
1056 		}
1057 
1058 		/* Enqueue Packet buffers */
1059 		virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
1060 
1061 		txvq->stats.bytes += txm->pkt_len;
1062 		virtio_update_packet_stats(&txvq->stats, txm);
1063 	}
1064 
1065 	txvq->stats.packets += nb_tx;
1066 
1067 	if (likely(nb_tx)) {
1068 		vq_update_avail_idx(vq);
1069 
1070 		if (unlikely(virtqueue_kick_prepare(vq))) {
1071 			virtqueue_notify(vq);
1072 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1073 		}
1074 	}
1075 
1076 	return nb_tx;
1077 }
1078