xref: /dpdk/drivers/net/virtio/virtio_rxtx.c (revision 3e0ceb9f17fff027fc6c8f18de35e11719ffa61e)
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
5  *   All rights reserved.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of Intel Corporation nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 #include <stdint.h>
35 #include <stdio.h>
36 #include <stdlib.h>
37 #include <string.h>
38 #include <errno.h>
39 
40 #include <rte_cycles.h>
41 #include <rte_memory.h>
42 #include <rte_branch_prediction.h>
43 #include <rte_mempool.h>
44 #include <rte_malloc.h>
45 #include <rte_mbuf.h>
46 #include <rte_ether.h>
47 #include <rte_ethdev.h>
48 #include <rte_prefetch.h>
49 #include <rte_string_fns.h>
50 #include <rte_errno.h>
51 #include <rte_byteorder.h>
52 #include <rte_net.h>
53 #include <rte_ip.h>
54 #include <rte_udp.h>
55 #include <rte_tcp.h>
56 
57 #include "virtio_logs.h"
58 #include "virtio_ethdev.h"
59 #include "virtio_pci.h"
60 #include "virtqueue.h"
61 #include "virtio_rxtx.h"
62 
63 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
64 #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
65 #else
66 #define  VIRTIO_DUMP_PACKET(m, len) do { } while (0)
67 #endif
68 
69 
70 #define VIRTIO_SIMPLE_FLAGS ((uint32_t)ETH_TXQ_FLAGS_NOMULTSEGS | \
71 	ETH_TXQ_FLAGS_NOOFFLOADS)
72 
73 int
74 virtio_dev_rx_queue_done(void *rxq, uint16_t offset)
75 {
76 	struct virtnet_rx *rxvq = rxq;
77 	struct virtqueue *vq = rxvq->vq;
78 
79 	return VIRTQUEUE_NUSED(vq) >= offset;
80 }
81 
82 void
83 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
84 {
85 	struct vring_desc *dp, *dp_tail;
86 	struct vq_desc_extra *dxp;
87 	uint16_t desc_idx_last = desc_idx;
88 
89 	dp  = &vq->vq_ring.desc[desc_idx];
90 	dxp = &vq->vq_descx[desc_idx];
91 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
92 	if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
93 		while (dp->flags & VRING_DESC_F_NEXT) {
94 			desc_idx_last = dp->next;
95 			dp = &vq->vq_ring.desc[dp->next];
96 		}
97 	}
98 	dxp->ndescs = 0;
99 
100 	/*
101 	 * We must append the existing free chain, if any, to the end of
102 	 * newly freed chain. If the virtqueue was completely used, then
103 	 * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
104 	 */
105 	if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
106 		vq->vq_desc_head_idx = desc_idx;
107 	} else {
108 		dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
109 		dp_tail->next = desc_idx;
110 	}
111 
112 	vq->vq_desc_tail_idx = desc_idx_last;
113 	dp->next = VQ_RING_DESC_CHAIN_END;
114 }
115 
116 static uint16_t
117 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
118 			   uint32_t *len, uint16_t num)
119 {
120 	struct vring_used_elem *uep;
121 	struct rte_mbuf *cookie;
122 	uint16_t used_idx, desc_idx;
123 	uint16_t i;
124 
125 	/*  Caller does the check */
126 	for (i = 0; i < num ; i++) {
127 		used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
128 		uep = &vq->vq_ring.used->ring[used_idx];
129 		desc_idx = (uint16_t) uep->id;
130 		len[i] = uep->len;
131 		cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
132 
133 		if (unlikely(cookie == NULL)) {
134 			PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
135 				vq->vq_used_cons_idx);
136 			break;
137 		}
138 
139 		rte_prefetch0(cookie);
140 		rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
141 		rx_pkts[i]  = cookie;
142 		vq->vq_used_cons_idx++;
143 		vq_ring_free_chain(vq, desc_idx);
144 		vq->vq_descx[desc_idx].cookie = NULL;
145 	}
146 
147 	return i;
148 }
149 
150 #ifndef DEFAULT_TX_FREE_THRESH
151 #define DEFAULT_TX_FREE_THRESH 32
152 #endif
153 
154 /* Cleanup from completed transmits. */
155 static void
156 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
157 {
158 	uint16_t i, used_idx, desc_idx;
159 	for (i = 0; i < num; i++) {
160 		struct vring_used_elem *uep;
161 		struct vq_desc_extra *dxp;
162 
163 		used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
164 		uep = &vq->vq_ring.used->ring[used_idx];
165 
166 		desc_idx = (uint16_t) uep->id;
167 		dxp = &vq->vq_descx[desc_idx];
168 		vq->vq_used_cons_idx++;
169 		vq_ring_free_chain(vq, desc_idx);
170 
171 		if (dxp->cookie != NULL) {
172 			rte_pktmbuf_free(dxp->cookie);
173 			dxp->cookie = NULL;
174 		}
175 	}
176 }
177 
178 
179 static inline int
180 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
181 {
182 	struct vq_desc_extra *dxp;
183 	struct virtio_hw *hw = vq->hw;
184 	struct vring_desc *start_dp;
185 	uint16_t needed = 1;
186 	uint16_t head_idx, idx;
187 
188 	if (unlikely(vq->vq_free_cnt == 0))
189 		return -ENOSPC;
190 	if (unlikely(vq->vq_free_cnt < needed))
191 		return -EMSGSIZE;
192 
193 	head_idx = vq->vq_desc_head_idx;
194 	if (unlikely(head_idx >= vq->vq_nentries))
195 		return -EFAULT;
196 
197 	idx = head_idx;
198 	dxp = &vq->vq_descx[idx];
199 	dxp->cookie = (void *)cookie;
200 	dxp->ndescs = needed;
201 
202 	start_dp = vq->vq_ring.desc;
203 	start_dp[idx].addr =
204 		VIRTIO_MBUF_ADDR(cookie, vq) +
205 		RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
206 	start_dp[idx].len =
207 		cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
208 	start_dp[idx].flags =  VRING_DESC_F_WRITE;
209 	idx = start_dp[idx].next;
210 	vq->vq_desc_head_idx = idx;
211 	if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
212 		vq->vq_desc_tail_idx = idx;
213 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
214 	vq_update_avail_ring(vq, head_idx);
215 
216 	return 0;
217 }
218 
219 /* When doing TSO, the IP length is not included in the pseudo header
220  * checksum of the packet given to the PMD, but for virtio it is
221  * expected.
222  */
223 static void
224 virtio_tso_fix_cksum(struct rte_mbuf *m)
225 {
226 	/* common case: header is not fragmented */
227 	if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
228 			m->l4_len)) {
229 		struct ipv4_hdr *iph;
230 		struct ipv6_hdr *ip6h;
231 		struct tcp_hdr *th;
232 		uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
233 		uint32_t tmp;
234 
235 		iph = rte_pktmbuf_mtod_offset(m, struct ipv4_hdr *, m->l2_len);
236 		th = RTE_PTR_ADD(iph, m->l3_len);
237 		if ((iph->version_ihl >> 4) == 4) {
238 			iph->hdr_checksum = 0;
239 			iph->hdr_checksum = rte_ipv4_cksum(iph);
240 			ip_len = iph->total_length;
241 			ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
242 				m->l3_len);
243 		} else {
244 			ip6h = (struct ipv6_hdr *)iph;
245 			ip_paylen = ip6h->payload_len;
246 		}
247 
248 		/* calculate the new phdr checksum not including ip_paylen */
249 		prev_cksum = th->cksum;
250 		tmp = prev_cksum;
251 		tmp += ip_paylen;
252 		tmp = (tmp & 0xffff) + (tmp >> 16);
253 		new_cksum = tmp;
254 
255 		/* replace it in the packet */
256 		th->cksum = new_cksum;
257 	}
258 }
259 
260 static inline int
261 tx_offload_enabled(struct virtio_hw *hw)
262 {
263 	return vtpci_with_feature(hw, VIRTIO_NET_F_CSUM) ||
264 		vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO4) ||
265 		vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO6);
266 }
267 
268 /* avoid write operation when necessary, to lessen cache issues */
269 #define ASSIGN_UNLESS_EQUAL(var, val) do {	\
270 	if ((var) != (val))			\
271 		(var) = (val);			\
272 } while (0)
273 
274 static inline void
275 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
276 		       uint16_t needed, int use_indirect, int can_push)
277 {
278 	struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
279 	struct vq_desc_extra *dxp;
280 	struct virtqueue *vq = txvq->vq;
281 	struct vring_desc *start_dp;
282 	uint16_t seg_num = cookie->nb_segs;
283 	uint16_t head_idx, idx;
284 	uint16_t head_size = vq->hw->vtnet_hdr_size;
285 	struct virtio_net_hdr *hdr;
286 	int offload;
287 
288 	offload = tx_offload_enabled(vq->hw);
289 	head_idx = vq->vq_desc_head_idx;
290 	idx = head_idx;
291 	dxp = &vq->vq_descx[idx];
292 	dxp->cookie = (void *)cookie;
293 	dxp->ndescs = needed;
294 
295 	start_dp = vq->vq_ring.desc;
296 
297 	if (can_push) {
298 		/* prepend cannot fail, checked by caller */
299 		hdr = (struct virtio_net_hdr *)
300 			rte_pktmbuf_prepend(cookie, head_size);
301 		/* rte_pktmbuf_prepend() counts the hdr size to the pkt length,
302 		 * which is wrong. Below subtract restores correct pkt size.
303 		 */
304 		cookie->pkt_len -= head_size;
305 		/* if offload disabled, it is not zeroed below, do it now */
306 		if (offload == 0) {
307 			ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
308 			ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
309 			ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
310 			ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
311 			ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
312 			ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
313 		}
314 	} else if (use_indirect) {
315 		/* setup tx ring slot to point to indirect
316 		 * descriptor list stored in reserved region.
317 		 *
318 		 * the first slot in indirect ring is already preset
319 		 * to point to the header in reserved region
320 		 */
321 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
322 			RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
323 		start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
324 		start_dp[idx].flags = VRING_DESC_F_INDIRECT;
325 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
326 
327 		/* loop below will fill in rest of the indirect elements */
328 		start_dp = txr[idx].tx_indir;
329 		idx = 1;
330 	} else {
331 		/* setup first tx ring slot to point to header
332 		 * stored in reserved region.
333 		 */
334 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
335 			RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
336 		start_dp[idx].len   = vq->hw->vtnet_hdr_size;
337 		start_dp[idx].flags = VRING_DESC_F_NEXT;
338 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
339 
340 		idx = start_dp[idx].next;
341 	}
342 
343 	/* Checksum Offload / TSO */
344 	if (offload) {
345 		if (cookie->ol_flags & PKT_TX_TCP_SEG)
346 			cookie->ol_flags |= PKT_TX_TCP_CKSUM;
347 
348 		switch (cookie->ol_flags & PKT_TX_L4_MASK) {
349 		case PKT_TX_UDP_CKSUM:
350 			hdr->csum_start = cookie->l2_len + cookie->l3_len;
351 			hdr->csum_offset = offsetof(struct udp_hdr,
352 				dgram_cksum);
353 			hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
354 			break;
355 
356 		case PKT_TX_TCP_CKSUM:
357 			hdr->csum_start = cookie->l2_len + cookie->l3_len;
358 			hdr->csum_offset = offsetof(struct tcp_hdr, cksum);
359 			hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
360 			break;
361 
362 		default:
363 			ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
364 			ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
365 			ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
366 			break;
367 		}
368 
369 		/* TCP Segmentation Offload */
370 		if (cookie->ol_flags & PKT_TX_TCP_SEG) {
371 			virtio_tso_fix_cksum(cookie);
372 			hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
373 				VIRTIO_NET_HDR_GSO_TCPV6 :
374 				VIRTIO_NET_HDR_GSO_TCPV4;
375 			hdr->gso_size = cookie->tso_segsz;
376 			hdr->hdr_len =
377 				cookie->l2_len +
378 				cookie->l3_len +
379 				cookie->l4_len;
380 		} else {
381 			ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
382 			ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
383 			ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
384 		}
385 	}
386 
387 	do {
388 		start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
389 		start_dp[idx].len   = cookie->data_len;
390 		start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
391 		idx = start_dp[idx].next;
392 	} while ((cookie = cookie->next) != NULL);
393 
394 	if (use_indirect)
395 		idx = vq->vq_ring.desc[head_idx].next;
396 
397 	vq->vq_desc_head_idx = idx;
398 	if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
399 		vq->vq_desc_tail_idx = idx;
400 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
401 	vq_update_avail_ring(vq, head_idx);
402 }
403 
404 void
405 virtio_dev_cq_start(struct rte_eth_dev *dev)
406 {
407 	struct virtio_hw *hw = dev->data->dev_private;
408 
409 	if (hw->cvq && hw->cvq->vq) {
410 		VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
411 	}
412 }
413 
414 int
415 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
416 			uint16_t queue_idx,
417 			uint16_t nb_desc,
418 			unsigned int socket_id __rte_unused,
419 			__rte_unused const struct rte_eth_rxconf *rx_conf,
420 			struct rte_mempool *mp)
421 {
422 	uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
423 	struct virtio_hw *hw = dev->data->dev_private;
424 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
425 	struct virtnet_rx *rxvq;
426 
427 	PMD_INIT_FUNC_TRACE();
428 
429 	if (nb_desc == 0 || nb_desc > vq->vq_nentries)
430 		nb_desc = vq->vq_nentries;
431 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
432 
433 	rxvq = &vq->rxq;
434 	rxvq->queue_id = queue_idx;
435 	rxvq->mpool = mp;
436 	if (rxvq->mpool == NULL) {
437 		rte_exit(EXIT_FAILURE,
438 			"Cannot allocate mbufs for rx virtqueue");
439 	}
440 	dev->data->rx_queues[queue_idx] = rxvq;
441 
442 	return 0;
443 }
444 
445 int
446 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
447 {
448 	uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
449 	struct virtio_hw *hw = dev->data->dev_private;
450 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
451 	struct virtnet_rx *rxvq = &vq->rxq;
452 	struct rte_mbuf *m;
453 	uint16_t desc_idx;
454 	int error, nbufs;
455 
456 	PMD_INIT_FUNC_TRACE();
457 
458 	/* Allocate blank mbufs for the each rx descriptor */
459 	nbufs = 0;
460 
461 	if (hw->use_simple_rx) {
462 		for (desc_idx = 0; desc_idx < vq->vq_nentries;
463 		     desc_idx++) {
464 			vq->vq_ring.avail->ring[desc_idx] = desc_idx;
465 			vq->vq_ring.desc[desc_idx].flags =
466 				VRING_DESC_F_WRITE;
467 		}
468 	}
469 
470 	memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
471 	for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
472 	     desc_idx++) {
473 		vq->sw_ring[vq->vq_nentries + desc_idx] =
474 			&rxvq->fake_mbuf;
475 	}
476 
477 	while (!virtqueue_full(vq)) {
478 		m = rte_mbuf_raw_alloc(rxvq->mpool);
479 		if (m == NULL)
480 			break;
481 
482 		/* Enqueue allocated buffers */
483 		if (hw->use_simple_rx)
484 			error = virtqueue_enqueue_recv_refill_simple(vq, m);
485 		else
486 			error = virtqueue_enqueue_recv_refill(vq, m);
487 
488 		if (error) {
489 			rte_pktmbuf_free(m);
490 			break;
491 		}
492 		nbufs++;
493 	}
494 
495 	vq_update_avail_idx(vq);
496 
497 	PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
498 
499 	virtio_rxq_vec_setup(rxvq);
500 
501 	VIRTQUEUE_DUMP(vq);
502 
503 	return 0;
504 }
505 
506 /*
507  * struct rte_eth_dev *dev: Used to update dev
508  * uint16_t nb_desc: Defaults to values read from config space
509  * unsigned int socket_id: Used to allocate memzone
510  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
511  * uint16_t queue_idx: Just used as an index in dev txq list
512  */
513 int
514 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
515 			uint16_t queue_idx,
516 			uint16_t nb_desc,
517 			unsigned int socket_id __rte_unused,
518 			const struct rte_eth_txconf *tx_conf)
519 {
520 	uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
521 	struct virtio_hw *hw = dev->data->dev_private;
522 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
523 	struct virtnet_tx *txvq;
524 	uint16_t tx_free_thresh;
525 
526 	PMD_INIT_FUNC_TRACE();
527 
528 	/* cannot use simple rxtx funcs with multisegs or offloads */
529 	if ((tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) != VIRTIO_SIMPLE_FLAGS)
530 		hw->use_simple_tx = 0;
531 
532 	if (nb_desc == 0 || nb_desc > vq->vq_nentries)
533 		nb_desc = vq->vq_nentries;
534 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
535 
536 	txvq = &vq->txq;
537 	txvq->queue_id = queue_idx;
538 
539 	tx_free_thresh = tx_conf->tx_free_thresh;
540 	if (tx_free_thresh == 0)
541 		tx_free_thresh =
542 			RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
543 
544 	if (tx_free_thresh >= (vq->vq_nentries - 3)) {
545 		RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
546 			"number of TX entries minus 3 (%u)."
547 			" (tx_free_thresh=%u port=%u queue=%u)\n",
548 			vq->vq_nentries - 3,
549 			tx_free_thresh, dev->data->port_id, queue_idx);
550 		return -EINVAL;
551 	}
552 
553 	vq->vq_free_thresh = tx_free_thresh;
554 
555 	dev->data->tx_queues[queue_idx] = txvq;
556 	return 0;
557 }
558 
559 int
560 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
561 				uint16_t queue_idx)
562 {
563 	uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
564 	struct virtio_hw *hw = dev->data->dev_private;
565 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
566 	uint16_t mid_idx = vq->vq_nentries >> 1;
567 	struct virtnet_tx *txvq = &vq->txq;
568 	uint16_t desc_idx;
569 
570 	PMD_INIT_FUNC_TRACE();
571 
572 	if (hw->use_simple_tx) {
573 		for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
574 			vq->vq_ring.avail->ring[desc_idx] =
575 				desc_idx + mid_idx;
576 			vq->vq_ring.desc[desc_idx + mid_idx].next =
577 				desc_idx;
578 			vq->vq_ring.desc[desc_idx + mid_idx].addr =
579 				txvq->virtio_net_hdr_mem +
580 				offsetof(struct virtio_tx_region, tx_hdr);
581 			vq->vq_ring.desc[desc_idx + mid_idx].len =
582 				vq->hw->vtnet_hdr_size;
583 			vq->vq_ring.desc[desc_idx + mid_idx].flags =
584 				VRING_DESC_F_NEXT;
585 			vq->vq_ring.desc[desc_idx].flags = 0;
586 		}
587 		for (desc_idx = mid_idx; desc_idx < vq->vq_nentries;
588 		     desc_idx++)
589 			vq->vq_ring.avail->ring[desc_idx] = desc_idx;
590 	}
591 
592 	VIRTQUEUE_DUMP(vq);
593 
594 	return 0;
595 }
596 
597 static void
598 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
599 {
600 	int error;
601 	/*
602 	 * Requeue the discarded mbuf. This should always be
603 	 * successful since it was just dequeued.
604 	 */
605 	error = virtqueue_enqueue_recv_refill(vq, m);
606 	if (unlikely(error)) {
607 		RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
608 		rte_pktmbuf_free(m);
609 	}
610 }
611 
612 static void
613 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
614 {
615 	uint32_t s = mbuf->pkt_len;
616 	struct ether_addr *ea;
617 
618 	if (s == 64) {
619 		stats->size_bins[1]++;
620 	} else if (s > 64 && s < 1024) {
621 		uint32_t bin;
622 
623 		/* count zeros, and offset into correct bin */
624 		bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
625 		stats->size_bins[bin]++;
626 	} else {
627 		if (s < 64)
628 			stats->size_bins[0]++;
629 		else if (s < 1519)
630 			stats->size_bins[6]++;
631 		else if (s >= 1519)
632 			stats->size_bins[7]++;
633 	}
634 
635 	ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
636 	if (is_multicast_ether_addr(ea)) {
637 		if (is_broadcast_ether_addr(ea))
638 			stats->broadcast++;
639 		else
640 			stats->multicast++;
641 	}
642 }
643 
644 /* Optionally fill offload information in structure */
645 static int
646 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
647 {
648 	struct rte_net_hdr_lens hdr_lens;
649 	uint32_t hdrlen, ptype;
650 	int l4_supported = 0;
651 
652 	/* nothing to do */
653 	if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
654 		return 0;
655 
656 	m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
657 
658 	ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
659 	m->packet_type = ptype;
660 	if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
661 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
662 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
663 		l4_supported = 1;
664 
665 	if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
666 		hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
667 		if (hdr->csum_start <= hdrlen && l4_supported) {
668 			m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
669 		} else {
670 			/* Unknown proto or tunnel, do sw cksum. We can assume
671 			 * the cksum field is in the first segment since the
672 			 * buffers we provided to the host are large enough.
673 			 * In case of SCTP, this will be wrong since it's a CRC
674 			 * but there's nothing we can do.
675 			 */
676 			uint16_t csum = 0, off;
677 
678 			rte_raw_cksum_mbuf(m, hdr->csum_start,
679 				rte_pktmbuf_pkt_len(m) - hdr->csum_start,
680 				&csum);
681 			if (likely(csum != 0xffff))
682 				csum = ~csum;
683 			off = hdr->csum_offset + hdr->csum_start;
684 			if (rte_pktmbuf_data_len(m) >= off + 1)
685 				*rte_pktmbuf_mtod_offset(m, uint16_t *,
686 					off) = csum;
687 		}
688 	} else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
689 		m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
690 	}
691 
692 	/* GSO request, save required information in mbuf */
693 	if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
694 		/* Check unsupported modes */
695 		if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
696 		    (hdr->gso_size == 0)) {
697 			return -EINVAL;
698 		}
699 
700 		/* Update mss lengthes in mbuf */
701 		m->tso_segsz = hdr->gso_size;
702 		switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
703 			case VIRTIO_NET_HDR_GSO_TCPV4:
704 			case VIRTIO_NET_HDR_GSO_TCPV6:
705 				m->ol_flags |= PKT_RX_LRO | \
706 					PKT_RX_L4_CKSUM_NONE;
707 				break;
708 			default:
709 				return -EINVAL;
710 		}
711 	}
712 
713 	return 0;
714 }
715 
716 static inline int
717 rx_offload_enabled(struct virtio_hw *hw)
718 {
719 	return vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_CSUM) ||
720 		vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO4) ||
721 		vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO6);
722 }
723 
724 #define VIRTIO_MBUF_BURST_SZ 64
725 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
726 uint16_t
727 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
728 {
729 	struct virtnet_rx *rxvq = rx_queue;
730 	struct virtqueue *vq = rxvq->vq;
731 	struct virtio_hw *hw = vq->hw;
732 	struct rte_mbuf *rxm, *new_mbuf;
733 	uint16_t nb_used, num, nb_rx;
734 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
735 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
736 	int error;
737 	uint32_t i, nb_enqueued;
738 	uint32_t hdr_size;
739 	int offload;
740 	struct virtio_net_hdr *hdr;
741 
742 	nb_rx = 0;
743 	if (unlikely(hw->started == 0))
744 		return nb_rx;
745 
746 	nb_used = VIRTQUEUE_NUSED(vq);
747 
748 	virtio_rmb();
749 
750 	num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
751 	if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
752 		num = VIRTIO_MBUF_BURST_SZ;
753 	if (likely(num > DESC_PER_CACHELINE))
754 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
755 
756 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
757 	PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
758 
759 	nb_enqueued = 0;
760 	hdr_size = hw->vtnet_hdr_size;
761 	offload = rx_offload_enabled(hw);
762 
763 	for (i = 0; i < num ; i++) {
764 		rxm = rcv_pkts[i];
765 
766 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
767 
768 		if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
769 			PMD_RX_LOG(ERR, "Packet drop");
770 			nb_enqueued++;
771 			virtio_discard_rxbuf(vq, rxm);
772 			rxvq->stats.errors++;
773 			continue;
774 		}
775 
776 		rxm->port = rxvq->port_id;
777 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
778 		rxm->ol_flags = 0;
779 		rxm->vlan_tci = 0;
780 
781 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
782 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
783 
784 		hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
785 			RTE_PKTMBUF_HEADROOM - hdr_size);
786 
787 		if (hw->vlan_strip)
788 			rte_vlan_strip(rxm);
789 
790 		if (offload && virtio_rx_offload(rxm, hdr) < 0) {
791 			virtio_discard_rxbuf(vq, rxm);
792 			rxvq->stats.errors++;
793 			continue;
794 		}
795 
796 		VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
797 
798 		rx_pkts[nb_rx++] = rxm;
799 
800 		rxvq->stats.bytes += rxm->pkt_len;
801 		virtio_update_packet_stats(&rxvq->stats, rxm);
802 	}
803 
804 	rxvq->stats.packets += nb_rx;
805 
806 	/* Allocate new mbuf for the used descriptor */
807 	error = ENOSPC;
808 	while (likely(!virtqueue_full(vq))) {
809 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
810 		if (unlikely(new_mbuf == NULL)) {
811 			struct rte_eth_dev *dev
812 				= &rte_eth_devices[rxvq->port_id];
813 			dev->data->rx_mbuf_alloc_failed++;
814 			break;
815 		}
816 		error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
817 		if (unlikely(error)) {
818 			rte_pktmbuf_free(new_mbuf);
819 			break;
820 		}
821 		nb_enqueued++;
822 	}
823 
824 	if (likely(nb_enqueued)) {
825 		vq_update_avail_idx(vq);
826 
827 		if (unlikely(virtqueue_kick_prepare(vq))) {
828 			virtqueue_notify(vq);
829 			PMD_RX_LOG(DEBUG, "Notified");
830 		}
831 	}
832 
833 	return nb_rx;
834 }
835 
836 uint16_t
837 virtio_recv_mergeable_pkts(void *rx_queue,
838 			struct rte_mbuf **rx_pkts,
839 			uint16_t nb_pkts)
840 {
841 	struct virtnet_rx *rxvq = rx_queue;
842 	struct virtqueue *vq = rxvq->vq;
843 	struct virtio_hw *hw = vq->hw;
844 	struct rte_mbuf *rxm, *new_mbuf;
845 	uint16_t nb_used, num, nb_rx;
846 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
847 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
848 	struct rte_mbuf *prev;
849 	int error;
850 	uint32_t i, nb_enqueued;
851 	uint32_t seg_num;
852 	uint16_t extra_idx;
853 	uint32_t seg_res;
854 	uint32_t hdr_size;
855 	int offload;
856 
857 	nb_rx = 0;
858 	if (unlikely(hw->started == 0))
859 		return nb_rx;
860 
861 	nb_used = VIRTQUEUE_NUSED(vq);
862 
863 	virtio_rmb();
864 
865 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
866 
867 	i = 0;
868 	nb_enqueued = 0;
869 	seg_num = 0;
870 	extra_idx = 0;
871 	seg_res = 0;
872 	hdr_size = hw->vtnet_hdr_size;
873 	offload = rx_offload_enabled(hw);
874 
875 	while (i < nb_used) {
876 		struct virtio_net_hdr_mrg_rxbuf *header;
877 
878 		if (nb_rx == nb_pkts)
879 			break;
880 
881 		num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
882 		if (num != 1)
883 			continue;
884 
885 		i++;
886 
887 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
888 		PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
889 
890 		rxm = rcv_pkts[0];
891 
892 		if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
893 			PMD_RX_LOG(ERR, "Packet drop");
894 			nb_enqueued++;
895 			virtio_discard_rxbuf(vq, rxm);
896 			rxvq->stats.errors++;
897 			continue;
898 		}
899 
900 		header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
901 			RTE_PKTMBUF_HEADROOM - hdr_size);
902 		seg_num = header->num_buffers;
903 
904 		if (seg_num == 0)
905 			seg_num = 1;
906 
907 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
908 		rxm->nb_segs = seg_num;
909 		rxm->ol_flags = 0;
910 		rxm->vlan_tci = 0;
911 		rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
912 		rxm->data_len = (uint16_t)(len[0] - hdr_size);
913 
914 		rxm->port = rxvq->port_id;
915 		rx_pkts[nb_rx] = rxm;
916 		prev = rxm;
917 
918 		if (offload && virtio_rx_offload(rxm, &header->hdr) < 0) {
919 			virtio_discard_rxbuf(vq, rxm);
920 			rxvq->stats.errors++;
921 			continue;
922 		}
923 
924 		seg_res = seg_num - 1;
925 
926 		while (seg_res != 0) {
927 			/*
928 			 * Get extra segments for current uncompleted packet.
929 			 */
930 			uint16_t  rcv_cnt =
931 				RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
932 			if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
933 				uint32_t rx_num =
934 					virtqueue_dequeue_burst_rx(vq,
935 					rcv_pkts, len, rcv_cnt);
936 				i += rx_num;
937 				rcv_cnt = rx_num;
938 			} else {
939 				PMD_RX_LOG(ERR,
940 					   "No enough segments for packet.");
941 				nb_enqueued++;
942 				virtio_discard_rxbuf(vq, rxm);
943 				rxvq->stats.errors++;
944 				break;
945 			}
946 
947 			extra_idx = 0;
948 
949 			while (extra_idx < rcv_cnt) {
950 				rxm = rcv_pkts[extra_idx];
951 
952 				rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
953 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
954 				rxm->data_len = (uint16_t)(len[extra_idx]);
955 
956 				if (prev)
957 					prev->next = rxm;
958 
959 				prev = rxm;
960 				rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
961 				extra_idx++;
962 			};
963 			seg_res -= rcv_cnt;
964 		}
965 
966 		if (hw->vlan_strip)
967 			rte_vlan_strip(rx_pkts[nb_rx]);
968 
969 		VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
970 			rx_pkts[nb_rx]->data_len);
971 
972 		rxvq->stats.bytes += rx_pkts[nb_rx]->pkt_len;
973 		virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
974 		nb_rx++;
975 	}
976 
977 	rxvq->stats.packets += nb_rx;
978 
979 	/* Allocate new mbuf for the used descriptor */
980 	error = ENOSPC;
981 	while (likely(!virtqueue_full(vq))) {
982 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
983 		if (unlikely(new_mbuf == NULL)) {
984 			struct rte_eth_dev *dev
985 				= &rte_eth_devices[rxvq->port_id];
986 			dev->data->rx_mbuf_alloc_failed++;
987 			break;
988 		}
989 		error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
990 		if (unlikely(error)) {
991 			rte_pktmbuf_free(new_mbuf);
992 			break;
993 		}
994 		nb_enqueued++;
995 	}
996 
997 	if (likely(nb_enqueued)) {
998 		vq_update_avail_idx(vq);
999 
1000 		if (unlikely(virtqueue_kick_prepare(vq))) {
1001 			virtqueue_notify(vq);
1002 			PMD_RX_LOG(DEBUG, "Notified");
1003 		}
1004 	}
1005 
1006 	return nb_rx;
1007 }
1008 
1009 uint16_t
1010 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1011 {
1012 	struct virtnet_tx *txvq = tx_queue;
1013 	struct virtqueue *vq = txvq->vq;
1014 	struct virtio_hw *hw = vq->hw;
1015 	uint16_t hdr_size = hw->vtnet_hdr_size;
1016 	uint16_t nb_used, nb_tx = 0;
1017 	int error;
1018 
1019 	if (unlikely(hw->started == 0))
1020 		return nb_tx;
1021 
1022 	if (unlikely(nb_pkts < 1))
1023 		return nb_pkts;
1024 
1025 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1026 	nb_used = VIRTQUEUE_NUSED(vq);
1027 
1028 	virtio_rmb();
1029 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1030 		virtio_xmit_cleanup(vq, nb_used);
1031 
1032 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1033 		struct rte_mbuf *txm = tx_pkts[nb_tx];
1034 		int can_push = 0, use_indirect = 0, slots, need;
1035 
1036 		/* Do VLAN tag insertion */
1037 		if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1038 			error = rte_vlan_insert(&txm);
1039 			if (unlikely(error)) {
1040 				rte_pktmbuf_free(txm);
1041 				continue;
1042 			}
1043 		}
1044 
1045 		/* optimize ring usage */
1046 		if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1047 		      vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1048 		    rte_mbuf_refcnt_read(txm) == 1 &&
1049 		    RTE_MBUF_DIRECT(txm) &&
1050 		    txm->nb_segs == 1 &&
1051 		    rte_pktmbuf_headroom(txm) >= hdr_size &&
1052 		    rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1053 				   __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1054 			can_push = 1;
1055 		else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1056 			 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1057 			use_indirect = 1;
1058 
1059 		/* How many main ring entries are needed to this Tx?
1060 		 * any_layout => number of segments
1061 		 * indirect   => 1
1062 		 * default    => number of segments + 1
1063 		 */
1064 		slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1065 		need = slots - vq->vq_free_cnt;
1066 
1067 		/* Positive value indicates it need free vring descriptors */
1068 		if (unlikely(need > 0)) {
1069 			nb_used = VIRTQUEUE_NUSED(vq);
1070 			virtio_rmb();
1071 			need = RTE_MIN(need, (int)nb_used);
1072 
1073 			virtio_xmit_cleanup(vq, need);
1074 			need = slots - vq->vq_free_cnt;
1075 			if (unlikely(need > 0)) {
1076 				PMD_TX_LOG(ERR,
1077 					   "No free tx descriptors to transmit");
1078 				break;
1079 			}
1080 		}
1081 
1082 		/* Enqueue Packet buffers */
1083 		virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
1084 
1085 		txvq->stats.bytes += txm->pkt_len;
1086 		virtio_update_packet_stats(&txvq->stats, txm);
1087 	}
1088 
1089 	txvq->stats.packets += nb_tx;
1090 
1091 	if (likely(nb_tx)) {
1092 		vq_update_avail_idx(vq);
1093 
1094 		if (unlikely(virtqueue_kick_prepare(vq))) {
1095 			virtqueue_notify(vq);
1096 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1097 		}
1098 	}
1099 
1100 	return nb_tx;
1101 }
1102