xref: /dpdk/drivers/net/virtio/virtio_rxtx.c (revision f1216c1eca5a54706af9f13133d06b3477099f8b)
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_memzone.h>
43 #include <rte_branch_prediction.h>
44 #include <rte_mempool.h>
45 #include <rte_malloc.h>
46 #include <rte_mbuf.h>
47 #include <rte_ether.h>
48 #include <rte_ethdev.h>
49 #include <rte_prefetch.h>
50 #include <rte_string_fns.h>
51 #include <rte_errno.h>
52 #include <rte_byteorder.h>
53 #include <rte_net.h>
54 #include <rte_ip.h>
55 #include <rte_udp.h>
56 #include <rte_tcp.h>
57 
58 #include "virtio_logs.h"
59 #include "virtio_ethdev.h"
60 #include "virtio_pci.h"
61 #include "virtqueue.h"
62 #include "virtio_rxtx.h"
63 
64 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
65 #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
66 #else
67 #define  VIRTIO_DUMP_PACKET(m, len) do { } while (0)
68 #endif
69 
70 
71 #define VIRTIO_SIMPLE_FLAGS ((uint32_t)ETH_TXQ_FLAGS_NOMULTSEGS | \
72 	ETH_TXQ_FLAGS_NOOFFLOADS)
73 
74 int
75 virtio_dev_rx_queue_done(void *rxq, uint16_t offset)
76 {
77 	struct virtnet_rx *rxvq = rxq;
78 	struct virtqueue *vq = rxvq->vq;
79 
80 	return VIRTQUEUE_NUSED(vq) >= offset;
81 }
82 
83 void
84 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
85 {
86 	struct vring_desc *dp, *dp_tail;
87 	struct vq_desc_extra *dxp;
88 	uint16_t desc_idx_last = desc_idx;
89 
90 	dp  = &vq->vq_ring.desc[desc_idx];
91 	dxp = &vq->vq_descx[desc_idx];
92 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
93 	if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
94 		while (dp->flags & VRING_DESC_F_NEXT) {
95 			desc_idx_last = dp->next;
96 			dp = &vq->vq_ring.desc[dp->next];
97 		}
98 	}
99 	dxp->ndescs = 0;
100 
101 	/*
102 	 * We must append the existing free chain, if any, to the end of
103 	 * newly freed chain. If the virtqueue was completely used, then
104 	 * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
105 	 */
106 	if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
107 		vq->vq_desc_head_idx = desc_idx;
108 	} else {
109 		dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
110 		dp_tail->next = desc_idx;
111 	}
112 
113 	vq->vq_desc_tail_idx = desc_idx_last;
114 	dp->next = VQ_RING_DESC_CHAIN_END;
115 }
116 
117 static uint16_t
118 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
119 			   uint32_t *len, uint16_t num)
120 {
121 	struct vring_used_elem *uep;
122 	struct rte_mbuf *cookie;
123 	uint16_t used_idx, desc_idx;
124 	uint16_t i;
125 
126 	/*  Caller does the check */
127 	for (i = 0; i < num ; i++) {
128 		used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
129 		uep = &vq->vq_ring.used->ring[used_idx];
130 		desc_idx = (uint16_t) uep->id;
131 		len[i] = uep->len;
132 		cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
133 
134 		if (unlikely(cookie == NULL)) {
135 			PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
136 				vq->vq_used_cons_idx);
137 			break;
138 		}
139 
140 		rte_prefetch0(cookie);
141 		rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
142 		rx_pkts[i]  = cookie;
143 		vq->vq_used_cons_idx++;
144 		vq_ring_free_chain(vq, desc_idx);
145 		vq->vq_descx[desc_idx].cookie = NULL;
146 	}
147 
148 	return i;
149 }
150 
151 #ifndef DEFAULT_TX_FREE_THRESH
152 #define DEFAULT_TX_FREE_THRESH 32
153 #endif
154 
155 /* Cleanup from completed transmits. */
156 static void
157 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
158 {
159 	uint16_t i, used_idx, desc_idx;
160 	for (i = 0; i < num; i++) {
161 		struct vring_used_elem *uep;
162 		struct vq_desc_extra *dxp;
163 
164 		used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
165 		uep = &vq->vq_ring.used->ring[used_idx];
166 
167 		desc_idx = (uint16_t) uep->id;
168 		dxp = &vq->vq_descx[desc_idx];
169 		vq->vq_used_cons_idx++;
170 		vq_ring_free_chain(vq, desc_idx);
171 
172 		if (dxp->cookie != NULL) {
173 			rte_pktmbuf_free(dxp->cookie);
174 			dxp->cookie = NULL;
175 		}
176 	}
177 }
178 
179 
180 static inline int
181 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
182 {
183 	struct vq_desc_extra *dxp;
184 	struct virtio_hw *hw = vq->hw;
185 	struct vring_desc *start_dp;
186 	uint16_t needed = 1;
187 	uint16_t head_idx, idx;
188 
189 	if (unlikely(vq->vq_free_cnt == 0))
190 		return -ENOSPC;
191 	if (unlikely(vq->vq_free_cnt < needed))
192 		return -EMSGSIZE;
193 
194 	head_idx = vq->vq_desc_head_idx;
195 	if (unlikely(head_idx >= vq->vq_nentries))
196 		return -EFAULT;
197 
198 	idx = head_idx;
199 	dxp = &vq->vq_descx[idx];
200 	dxp->cookie = (void *)cookie;
201 	dxp->ndescs = needed;
202 
203 	start_dp = vq->vq_ring.desc;
204 	start_dp[idx].addr =
205 		VIRTIO_MBUF_ADDR(cookie, vq) +
206 		RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
207 	start_dp[idx].len =
208 		cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
209 	start_dp[idx].flags =  VRING_DESC_F_WRITE;
210 	idx = start_dp[idx].next;
211 	vq->vq_desc_head_idx = idx;
212 	if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
213 		vq->vq_desc_tail_idx = idx;
214 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
215 	vq_update_avail_ring(vq, head_idx);
216 
217 	return 0;
218 }
219 
220 /* When doing TSO, the IP length is not included in the pseudo header
221  * checksum of the packet given to the PMD, but for virtio it is
222  * expected.
223  */
224 static void
225 virtio_tso_fix_cksum(struct rte_mbuf *m)
226 {
227 	/* common case: header is not fragmented */
228 	if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
229 			m->l4_len)) {
230 		struct ipv4_hdr *iph;
231 		struct ipv6_hdr *ip6h;
232 		struct tcp_hdr *th;
233 		uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
234 		uint32_t tmp;
235 
236 		iph = rte_pktmbuf_mtod_offset(m, struct ipv4_hdr *, m->l2_len);
237 		th = RTE_PTR_ADD(iph, m->l3_len);
238 		if ((iph->version_ihl >> 4) == 4) {
239 			iph->hdr_checksum = 0;
240 			iph->hdr_checksum = rte_ipv4_cksum(iph);
241 			ip_len = iph->total_length;
242 			ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
243 				m->l3_len);
244 		} else {
245 			ip6h = (struct ipv6_hdr *)iph;
246 			ip_paylen = ip6h->payload_len;
247 		}
248 
249 		/* calculate the new phdr checksum not including ip_paylen */
250 		prev_cksum = th->cksum;
251 		tmp = prev_cksum;
252 		tmp += ip_paylen;
253 		tmp = (tmp & 0xffff) + (tmp >> 16);
254 		new_cksum = tmp;
255 
256 		/* replace it in the packet */
257 		th->cksum = new_cksum;
258 	}
259 }
260 
261 static inline int
262 tx_offload_enabled(struct virtio_hw *hw)
263 {
264 	return vtpci_with_feature(hw, VIRTIO_NET_F_CSUM) ||
265 		vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO4) ||
266 		vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO6);
267 }
268 
269 /* avoid write operation when necessary, to lessen cache issues */
270 #define ASSIGN_UNLESS_EQUAL(var, val) do {	\
271 	if ((var) != (val))			\
272 		(var) = (val);			\
273 } while (0)
274 
275 static inline void
276 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
277 		       uint16_t needed, int use_indirect, int can_push)
278 {
279 	struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
280 	struct vq_desc_extra *dxp;
281 	struct virtqueue *vq = txvq->vq;
282 	struct vring_desc *start_dp;
283 	uint16_t seg_num = cookie->nb_segs;
284 	uint16_t head_idx, idx;
285 	uint16_t head_size = vq->hw->vtnet_hdr_size;
286 	struct virtio_net_hdr *hdr;
287 	int offload;
288 
289 	offload = tx_offload_enabled(vq->hw);
290 	head_idx = vq->vq_desc_head_idx;
291 	idx = head_idx;
292 	dxp = &vq->vq_descx[idx];
293 	dxp->cookie = (void *)cookie;
294 	dxp->ndescs = needed;
295 
296 	start_dp = vq->vq_ring.desc;
297 
298 	if (can_push) {
299 		/* prepend cannot fail, checked by caller */
300 		hdr = (struct virtio_net_hdr *)
301 			rte_pktmbuf_prepend(cookie, head_size);
302 		/* rte_pktmbuf_prepend() counts the hdr size to the pkt length,
303 		 * which is wrong. Below subtract restores correct pkt size.
304 		 */
305 		cookie->pkt_len -= head_size;
306 		/* if offload disabled, it is not zeroed below, do it now */
307 		if (offload == 0) {
308 			ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
309 			ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
310 			ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
311 			ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
312 			ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
313 			ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
314 		}
315 	} else if (use_indirect) {
316 		/* setup tx ring slot to point to indirect
317 		 * descriptor list stored in reserved region.
318 		 *
319 		 * the first slot in indirect ring is already preset
320 		 * to point to the header in reserved region
321 		 */
322 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
323 			RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
324 		start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
325 		start_dp[idx].flags = VRING_DESC_F_INDIRECT;
326 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
327 
328 		/* loop below will fill in rest of the indirect elements */
329 		start_dp = txr[idx].tx_indir;
330 		idx = 1;
331 	} else {
332 		/* setup first tx ring slot to point to header
333 		 * stored in reserved region.
334 		 */
335 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
336 			RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
337 		start_dp[idx].len   = vq->hw->vtnet_hdr_size;
338 		start_dp[idx].flags = VRING_DESC_F_NEXT;
339 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
340 
341 		idx = start_dp[idx].next;
342 	}
343 
344 	/* Checksum Offload / TSO */
345 	if (offload) {
346 		if (cookie->ol_flags & PKT_TX_TCP_SEG)
347 			cookie->ol_flags |= PKT_TX_TCP_CKSUM;
348 
349 		switch (cookie->ol_flags & PKT_TX_L4_MASK) {
350 		case PKT_TX_UDP_CKSUM:
351 			hdr->csum_start = cookie->l2_len + cookie->l3_len;
352 			hdr->csum_offset = offsetof(struct udp_hdr,
353 				dgram_cksum);
354 			hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
355 			break;
356 
357 		case PKT_TX_TCP_CKSUM:
358 			hdr->csum_start = cookie->l2_len + cookie->l3_len;
359 			hdr->csum_offset = offsetof(struct tcp_hdr, cksum);
360 			hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
361 			break;
362 
363 		default:
364 			ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
365 			ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
366 			ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
367 			break;
368 		}
369 
370 		/* TCP Segmentation Offload */
371 		if (cookie->ol_flags & PKT_TX_TCP_SEG) {
372 			virtio_tso_fix_cksum(cookie);
373 			hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
374 				VIRTIO_NET_HDR_GSO_TCPV6 :
375 				VIRTIO_NET_HDR_GSO_TCPV4;
376 			hdr->gso_size = cookie->tso_segsz;
377 			hdr->hdr_len =
378 				cookie->l2_len +
379 				cookie->l3_len +
380 				cookie->l4_len;
381 		} else {
382 			ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
383 			ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
384 			ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
385 		}
386 	}
387 
388 	do {
389 		start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
390 		start_dp[idx].len   = cookie->data_len;
391 		start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
392 		idx = start_dp[idx].next;
393 	} while ((cookie = cookie->next) != NULL);
394 
395 	if (use_indirect)
396 		idx = vq->vq_ring.desc[head_idx].next;
397 
398 	vq->vq_desc_head_idx = idx;
399 	if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
400 		vq->vq_desc_tail_idx = idx;
401 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
402 	vq_update_avail_ring(vq, head_idx);
403 }
404 
405 void
406 virtio_dev_cq_start(struct rte_eth_dev *dev)
407 {
408 	struct virtio_hw *hw = dev->data->dev_private;
409 
410 	if (hw->cvq && hw->cvq->vq) {
411 		VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
412 	}
413 }
414 
415 int
416 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
417 			uint16_t queue_idx,
418 			uint16_t nb_desc,
419 			unsigned int socket_id __rte_unused,
420 			__rte_unused const struct rte_eth_rxconf *rx_conf,
421 			struct rte_mempool *mp)
422 {
423 	uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
424 	struct virtio_hw *hw = dev->data->dev_private;
425 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
426 	struct virtnet_rx *rxvq;
427 
428 	PMD_INIT_FUNC_TRACE();
429 
430 	if (nb_desc == 0 || nb_desc > vq->vq_nentries)
431 		nb_desc = vq->vq_nentries;
432 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
433 
434 	rxvq = &vq->rxq;
435 	rxvq->queue_id = queue_idx;
436 	rxvq->mpool = mp;
437 	if (rxvq->mpool == NULL) {
438 		rte_exit(EXIT_FAILURE,
439 			"Cannot allocate mbufs for rx virtqueue");
440 	}
441 	dev->data->rx_queues[queue_idx] = rxvq;
442 
443 	return 0;
444 }
445 
446 int
447 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
448 {
449 	uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
450 	struct virtio_hw *hw = dev->data->dev_private;
451 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
452 	struct virtnet_rx *rxvq = &vq->rxq;
453 	struct rte_mbuf *m;
454 	uint16_t desc_idx;
455 	int error, nbufs;
456 
457 	PMD_INIT_FUNC_TRACE();
458 
459 	/* Allocate blank mbufs for the each rx descriptor */
460 	nbufs = 0;
461 
462 	if (hw->use_simple_rx) {
463 		for (desc_idx = 0; desc_idx < vq->vq_nentries;
464 		     desc_idx++) {
465 			vq->vq_ring.avail->ring[desc_idx] = desc_idx;
466 			vq->vq_ring.desc[desc_idx].flags =
467 				VRING_DESC_F_WRITE;
468 		}
469 	}
470 
471 	memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
472 	for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
473 	     desc_idx++) {
474 		vq->sw_ring[vq->vq_nentries + desc_idx] =
475 			&rxvq->fake_mbuf;
476 	}
477 
478 	while (!virtqueue_full(vq)) {
479 		m = rte_mbuf_raw_alloc(rxvq->mpool);
480 		if (m == NULL)
481 			break;
482 
483 		/* Enqueue allocated buffers */
484 		if (hw->use_simple_rx)
485 			error = virtqueue_enqueue_recv_refill_simple(vq, m);
486 		else
487 			error = virtqueue_enqueue_recv_refill(vq, m);
488 
489 		if (error) {
490 			rte_pktmbuf_free(m);
491 			break;
492 		}
493 		nbufs++;
494 	}
495 
496 	vq_update_avail_idx(vq);
497 
498 	PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
499 
500 	virtio_rxq_vec_setup(rxvq);
501 
502 	VIRTQUEUE_DUMP(vq);
503 
504 	return 0;
505 }
506 
507 /*
508  * struct rte_eth_dev *dev: Used to update dev
509  * uint16_t nb_desc: Defaults to values read from config space
510  * unsigned int socket_id: Used to allocate memzone
511  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
512  * uint16_t queue_idx: Just used as an index in dev txq list
513  */
514 int
515 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
516 			uint16_t queue_idx,
517 			uint16_t nb_desc,
518 			unsigned int socket_id __rte_unused,
519 			const struct rte_eth_txconf *tx_conf)
520 {
521 	uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
522 	struct virtio_hw *hw = dev->data->dev_private;
523 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
524 	struct virtnet_tx *txvq;
525 	uint16_t tx_free_thresh;
526 
527 	PMD_INIT_FUNC_TRACE();
528 
529 	/* cannot use simple rxtx funcs with multisegs or offloads */
530 	if ((tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) != VIRTIO_SIMPLE_FLAGS)
531 		hw->use_simple_tx = 0;
532 
533 	if (nb_desc == 0 || nb_desc > vq->vq_nentries)
534 		nb_desc = vq->vq_nentries;
535 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
536 
537 	txvq = &vq->txq;
538 	txvq->queue_id = queue_idx;
539 
540 	tx_free_thresh = tx_conf->tx_free_thresh;
541 	if (tx_free_thresh == 0)
542 		tx_free_thresh =
543 			RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
544 
545 	if (tx_free_thresh >= (vq->vq_nentries - 3)) {
546 		RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
547 			"number of TX entries minus 3 (%u)."
548 			" (tx_free_thresh=%u port=%u queue=%u)\n",
549 			vq->vq_nentries - 3,
550 			tx_free_thresh, dev->data->port_id, queue_idx);
551 		return -EINVAL;
552 	}
553 
554 	vq->vq_free_thresh = tx_free_thresh;
555 
556 	dev->data->tx_queues[queue_idx] = txvq;
557 	return 0;
558 }
559 
560 int
561 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
562 				uint16_t queue_idx)
563 {
564 	uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
565 	struct virtio_hw *hw = dev->data->dev_private;
566 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
567 	uint16_t mid_idx = vq->vq_nentries >> 1;
568 	struct virtnet_tx *txvq = &vq->txq;
569 	uint16_t desc_idx;
570 
571 	PMD_INIT_FUNC_TRACE();
572 
573 	if (hw->use_simple_tx) {
574 		for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
575 			vq->vq_ring.avail->ring[desc_idx] =
576 				desc_idx + mid_idx;
577 			vq->vq_ring.desc[desc_idx + mid_idx].next =
578 				desc_idx;
579 			vq->vq_ring.desc[desc_idx + mid_idx].addr =
580 				txvq->virtio_net_hdr_mem +
581 				offsetof(struct virtio_tx_region, tx_hdr);
582 			vq->vq_ring.desc[desc_idx + mid_idx].len =
583 				vq->hw->vtnet_hdr_size;
584 			vq->vq_ring.desc[desc_idx + mid_idx].flags =
585 				VRING_DESC_F_NEXT;
586 			vq->vq_ring.desc[desc_idx].flags = 0;
587 		}
588 		for (desc_idx = mid_idx; desc_idx < vq->vq_nentries;
589 		     desc_idx++)
590 			vq->vq_ring.avail->ring[desc_idx] = desc_idx;
591 	}
592 
593 	VIRTQUEUE_DUMP(vq);
594 
595 	return 0;
596 }
597 
598 static void
599 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
600 {
601 	int error;
602 	/*
603 	 * Requeue the discarded mbuf. This should always be
604 	 * successful since it was just dequeued.
605 	 */
606 	error = virtqueue_enqueue_recv_refill(vq, m);
607 	if (unlikely(error)) {
608 		RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
609 		rte_pktmbuf_free(m);
610 	}
611 }
612 
613 static void
614 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
615 {
616 	uint32_t s = mbuf->pkt_len;
617 	struct ether_addr *ea;
618 
619 	if (s == 64) {
620 		stats->size_bins[1]++;
621 	} else if (s > 64 && s < 1024) {
622 		uint32_t bin;
623 
624 		/* count zeros, and offset into correct bin */
625 		bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
626 		stats->size_bins[bin]++;
627 	} else {
628 		if (s < 64)
629 			stats->size_bins[0]++;
630 		else if (s < 1519)
631 			stats->size_bins[6]++;
632 		else if (s >= 1519)
633 			stats->size_bins[7]++;
634 	}
635 
636 	ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
637 	if (is_multicast_ether_addr(ea)) {
638 		if (is_broadcast_ether_addr(ea))
639 			stats->broadcast++;
640 		else
641 			stats->multicast++;
642 	}
643 }
644 
645 /* Optionally fill offload information in structure */
646 static int
647 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
648 {
649 	struct rte_net_hdr_lens hdr_lens;
650 	uint32_t hdrlen, ptype;
651 	int l4_supported = 0;
652 
653 	/* nothing to do */
654 	if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
655 		return 0;
656 
657 	m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
658 
659 	ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
660 	m->packet_type = ptype;
661 	if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
662 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
663 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
664 		l4_supported = 1;
665 
666 	if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
667 		hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
668 		if (hdr->csum_start <= hdrlen && l4_supported) {
669 			m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
670 		} else {
671 			/* Unknown proto or tunnel, do sw cksum. We can assume
672 			 * the cksum field is in the first segment since the
673 			 * buffers we provided to the host are large enough.
674 			 * In case of SCTP, this will be wrong since it's a CRC
675 			 * but there's nothing we can do.
676 			 */
677 			uint16_t csum = 0, off;
678 
679 			rte_raw_cksum_mbuf(m, hdr->csum_start,
680 				rte_pktmbuf_pkt_len(m) - hdr->csum_start,
681 				&csum);
682 			if (likely(csum != 0xffff))
683 				csum = ~csum;
684 			off = hdr->csum_offset + hdr->csum_start;
685 			if (rte_pktmbuf_data_len(m) >= off + 1)
686 				*rte_pktmbuf_mtod_offset(m, uint16_t *,
687 					off) = csum;
688 		}
689 	} else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
690 		m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
691 	}
692 
693 	/* GSO request, save required information in mbuf */
694 	if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
695 		/* Check unsupported modes */
696 		if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
697 		    (hdr->gso_size == 0)) {
698 			return -EINVAL;
699 		}
700 
701 		/* Update mss lengthes in mbuf */
702 		m->tso_segsz = hdr->gso_size;
703 		switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
704 			case VIRTIO_NET_HDR_GSO_TCPV4:
705 			case VIRTIO_NET_HDR_GSO_TCPV6:
706 				m->ol_flags |= PKT_RX_LRO | \
707 					PKT_RX_L4_CKSUM_NONE;
708 				break;
709 			default:
710 				return -EINVAL;
711 		}
712 	}
713 
714 	return 0;
715 }
716 
717 static inline int
718 rx_offload_enabled(struct virtio_hw *hw)
719 {
720 	return vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_CSUM) ||
721 		vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO4) ||
722 		vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO6);
723 }
724 
725 #define VIRTIO_MBUF_BURST_SZ 64
726 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
727 uint16_t
728 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
729 {
730 	struct virtnet_rx *rxvq = rx_queue;
731 	struct virtqueue *vq = rxvq->vq;
732 	struct virtio_hw *hw = vq->hw;
733 	struct rte_mbuf *rxm, *new_mbuf;
734 	uint16_t nb_used, num, nb_rx;
735 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
736 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
737 	int error;
738 	uint32_t i, nb_enqueued;
739 	uint32_t hdr_size;
740 	int offload;
741 	struct virtio_net_hdr *hdr;
742 
743 	nb_rx = 0;
744 	if (unlikely(hw->started == 0))
745 		return nb_rx;
746 
747 	nb_used = VIRTQUEUE_NUSED(vq);
748 
749 	virtio_rmb();
750 
751 	num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
752 	if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
753 		num = VIRTIO_MBUF_BURST_SZ;
754 	if (likely(num > DESC_PER_CACHELINE))
755 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
756 
757 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
758 	PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
759 
760 	nb_enqueued = 0;
761 	hdr_size = hw->vtnet_hdr_size;
762 	offload = rx_offload_enabled(hw);
763 
764 	for (i = 0; i < num ; i++) {
765 		rxm = rcv_pkts[i];
766 
767 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
768 
769 		if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
770 			PMD_RX_LOG(ERR, "Packet drop");
771 			nb_enqueued++;
772 			virtio_discard_rxbuf(vq, rxm);
773 			rxvq->stats.errors++;
774 			continue;
775 		}
776 
777 		rxm->port = rxvq->port_id;
778 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
779 		rxm->ol_flags = 0;
780 		rxm->vlan_tci = 0;
781 
782 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
783 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
784 
785 		hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
786 			RTE_PKTMBUF_HEADROOM - hdr_size);
787 
788 		if (hw->vlan_strip)
789 			rte_vlan_strip(rxm);
790 
791 		if (offload && virtio_rx_offload(rxm, hdr) < 0) {
792 			virtio_discard_rxbuf(vq, rxm);
793 			rxvq->stats.errors++;
794 			continue;
795 		}
796 
797 		VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
798 
799 		rx_pkts[nb_rx++] = rxm;
800 
801 		rxvq->stats.bytes += rxm->pkt_len;
802 		virtio_update_packet_stats(&rxvq->stats, rxm);
803 	}
804 
805 	rxvq->stats.packets += nb_rx;
806 
807 	/* Allocate new mbuf for the used descriptor */
808 	error = ENOSPC;
809 	while (likely(!virtqueue_full(vq))) {
810 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
811 		if (unlikely(new_mbuf == NULL)) {
812 			struct rte_eth_dev *dev
813 				= &rte_eth_devices[rxvq->port_id];
814 			dev->data->rx_mbuf_alloc_failed++;
815 			break;
816 		}
817 		error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
818 		if (unlikely(error)) {
819 			rte_pktmbuf_free(new_mbuf);
820 			break;
821 		}
822 		nb_enqueued++;
823 	}
824 
825 	if (likely(nb_enqueued)) {
826 		vq_update_avail_idx(vq);
827 
828 		if (unlikely(virtqueue_kick_prepare(vq))) {
829 			virtqueue_notify(vq);
830 			PMD_RX_LOG(DEBUG, "Notified");
831 		}
832 	}
833 
834 	return nb_rx;
835 }
836 
837 uint16_t
838 virtio_recv_mergeable_pkts(void *rx_queue,
839 			struct rte_mbuf **rx_pkts,
840 			uint16_t nb_pkts)
841 {
842 	struct virtnet_rx *rxvq = rx_queue;
843 	struct virtqueue *vq = rxvq->vq;
844 	struct virtio_hw *hw = vq->hw;
845 	struct rte_mbuf *rxm, *new_mbuf;
846 	uint16_t nb_used, num, nb_rx;
847 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
848 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
849 	struct rte_mbuf *prev;
850 	int error;
851 	uint32_t i, nb_enqueued;
852 	uint32_t seg_num;
853 	uint16_t extra_idx;
854 	uint32_t seg_res;
855 	uint32_t hdr_size;
856 	int offload;
857 
858 	nb_rx = 0;
859 	if (unlikely(hw->started == 0))
860 		return nb_rx;
861 
862 	nb_used = VIRTQUEUE_NUSED(vq);
863 
864 	virtio_rmb();
865 
866 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
867 
868 	i = 0;
869 	nb_enqueued = 0;
870 	seg_num = 0;
871 	extra_idx = 0;
872 	seg_res = 0;
873 	hdr_size = hw->vtnet_hdr_size;
874 	offload = rx_offload_enabled(hw);
875 
876 	while (i < nb_used) {
877 		struct virtio_net_hdr_mrg_rxbuf *header;
878 
879 		if (nb_rx == nb_pkts)
880 			break;
881 
882 		num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
883 		if (num != 1)
884 			continue;
885 
886 		i++;
887 
888 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
889 		PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
890 
891 		rxm = rcv_pkts[0];
892 
893 		if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
894 			PMD_RX_LOG(ERR, "Packet drop");
895 			nb_enqueued++;
896 			virtio_discard_rxbuf(vq, rxm);
897 			rxvq->stats.errors++;
898 			continue;
899 		}
900 
901 		header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
902 			RTE_PKTMBUF_HEADROOM - hdr_size);
903 		seg_num = header->num_buffers;
904 
905 		if (seg_num == 0)
906 			seg_num = 1;
907 
908 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
909 		rxm->nb_segs = seg_num;
910 		rxm->ol_flags = 0;
911 		rxm->vlan_tci = 0;
912 		rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
913 		rxm->data_len = (uint16_t)(len[0] - hdr_size);
914 
915 		rxm->port = rxvq->port_id;
916 		rx_pkts[nb_rx] = rxm;
917 		prev = rxm;
918 
919 		if (offload && virtio_rx_offload(rxm, &header->hdr) < 0) {
920 			virtio_discard_rxbuf(vq, rxm);
921 			rxvq->stats.errors++;
922 			continue;
923 		}
924 
925 		seg_res = seg_num - 1;
926 
927 		while (seg_res != 0) {
928 			/*
929 			 * Get extra segments for current uncompleted packet.
930 			 */
931 			uint16_t  rcv_cnt =
932 				RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
933 			if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
934 				uint32_t rx_num =
935 					virtqueue_dequeue_burst_rx(vq,
936 					rcv_pkts, len, rcv_cnt);
937 				i += rx_num;
938 				rcv_cnt = rx_num;
939 			} else {
940 				PMD_RX_LOG(ERR,
941 					   "No enough segments for packet.");
942 				nb_enqueued++;
943 				virtio_discard_rxbuf(vq, rxm);
944 				rxvq->stats.errors++;
945 				break;
946 			}
947 
948 			extra_idx = 0;
949 
950 			while (extra_idx < rcv_cnt) {
951 				rxm = rcv_pkts[extra_idx];
952 
953 				rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
954 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
955 				rxm->data_len = (uint16_t)(len[extra_idx]);
956 
957 				if (prev)
958 					prev->next = rxm;
959 
960 				prev = rxm;
961 				rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
962 				extra_idx++;
963 			};
964 			seg_res -= rcv_cnt;
965 		}
966 
967 		if (hw->vlan_strip)
968 			rte_vlan_strip(rx_pkts[nb_rx]);
969 
970 		VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
971 			rx_pkts[nb_rx]->data_len);
972 
973 		rxvq->stats.bytes += rx_pkts[nb_rx]->pkt_len;
974 		virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
975 		nb_rx++;
976 	}
977 
978 	rxvq->stats.packets += nb_rx;
979 
980 	/* Allocate new mbuf for the used descriptor */
981 	error = ENOSPC;
982 	while (likely(!virtqueue_full(vq))) {
983 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
984 		if (unlikely(new_mbuf == NULL)) {
985 			struct rte_eth_dev *dev
986 				= &rte_eth_devices[rxvq->port_id];
987 			dev->data->rx_mbuf_alloc_failed++;
988 			break;
989 		}
990 		error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
991 		if (unlikely(error)) {
992 			rte_pktmbuf_free(new_mbuf);
993 			break;
994 		}
995 		nb_enqueued++;
996 	}
997 
998 	if (likely(nb_enqueued)) {
999 		vq_update_avail_idx(vq);
1000 
1001 		if (unlikely(virtqueue_kick_prepare(vq))) {
1002 			virtqueue_notify(vq);
1003 			PMD_RX_LOG(DEBUG, "Notified");
1004 		}
1005 	}
1006 
1007 	return nb_rx;
1008 }
1009 
1010 uint16_t
1011 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1012 {
1013 	struct virtnet_tx *txvq = tx_queue;
1014 	struct virtqueue *vq = txvq->vq;
1015 	struct virtio_hw *hw = vq->hw;
1016 	uint16_t hdr_size = hw->vtnet_hdr_size;
1017 	uint16_t nb_used, nb_tx = 0;
1018 	int error;
1019 
1020 	if (unlikely(hw->started == 0))
1021 		return nb_tx;
1022 
1023 	if (unlikely(nb_pkts < 1))
1024 		return nb_pkts;
1025 
1026 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1027 	nb_used = VIRTQUEUE_NUSED(vq);
1028 
1029 	virtio_rmb();
1030 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1031 		virtio_xmit_cleanup(vq, nb_used);
1032 
1033 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1034 		struct rte_mbuf *txm = tx_pkts[nb_tx];
1035 		int can_push = 0, use_indirect = 0, slots, need;
1036 
1037 		/* Do VLAN tag insertion */
1038 		if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1039 			error = rte_vlan_insert(&txm);
1040 			if (unlikely(error)) {
1041 				rte_pktmbuf_free(txm);
1042 				continue;
1043 			}
1044 		}
1045 
1046 		/* optimize ring usage */
1047 		if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1048 		      vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1049 		    rte_mbuf_refcnt_read(txm) == 1 &&
1050 		    RTE_MBUF_DIRECT(txm) &&
1051 		    txm->nb_segs == 1 &&
1052 		    rte_pktmbuf_headroom(txm) >= hdr_size &&
1053 		    rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1054 				   __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1055 			can_push = 1;
1056 		else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1057 			 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1058 			use_indirect = 1;
1059 
1060 		/* How many main ring entries are needed to this Tx?
1061 		 * any_layout => number of segments
1062 		 * indirect   => 1
1063 		 * default    => number of segments + 1
1064 		 */
1065 		slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1066 		need = slots - vq->vq_free_cnt;
1067 
1068 		/* Positive value indicates it need free vring descriptors */
1069 		if (unlikely(need > 0)) {
1070 			nb_used = VIRTQUEUE_NUSED(vq);
1071 			virtio_rmb();
1072 			need = RTE_MIN(need, (int)nb_used);
1073 
1074 			virtio_xmit_cleanup(vq, need);
1075 			need = slots - vq->vq_free_cnt;
1076 			if (unlikely(need > 0)) {
1077 				PMD_TX_LOG(ERR,
1078 					   "No free tx descriptors to transmit");
1079 				break;
1080 			}
1081 		}
1082 
1083 		/* Enqueue Packet buffers */
1084 		virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
1085 
1086 		txvq->stats.bytes += txm->pkt_len;
1087 		virtio_update_packet_stats(&txvq->stats, txm);
1088 	}
1089 
1090 	txvq->stats.packets += nb_tx;
1091 
1092 	if (likely(nb_tx)) {
1093 		vq_update_avail_idx(vq);
1094 
1095 		if (unlikely(virtqueue_kick_prepare(vq))) {
1096 			virtqueue_notify(vq);
1097 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1098 		}
1099 	}
1100 
1101 	return nb_tx;
1102 }
1103