xref: /dpdk/drivers/net/virtio/virtio_rxtx.c (revision 0857b942113874c69dc3db5df11a828ee3cc9b6b)
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_cpuflags.h>
54 #include <rte_net.h>
55 #include <rte_ip.h>
56 #include <rte_udp.h>
57 #include <rte_tcp.h>
58 
59 #include "virtio_logs.h"
60 #include "virtio_ethdev.h"
61 #include "virtio_pci.h"
62 #include "virtqueue.h"
63 #include "virtio_rxtx.h"
64 
65 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
66 #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
67 #else
68 #define  VIRTIO_DUMP_PACKET(m, len) do { } while (0)
69 #endif
70 
71 
72 #define VIRTIO_SIMPLE_FLAGS ((uint32_t)ETH_TXQ_FLAGS_NOMULTSEGS | \
73 	ETH_TXQ_FLAGS_NOOFFLOADS)
74 
75 int
76 virtio_dev_rx_queue_done(void *rxq, uint16_t offset)
77 {
78 	struct virtnet_rx *rxvq = rxq;
79 	struct virtqueue *vq = rxvq->vq;
80 
81 	return VIRTQUEUE_NUSED(vq) >= offset;
82 }
83 
84 static void
85 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
86 {
87 	struct vring_desc *dp, *dp_tail;
88 	struct vq_desc_extra *dxp;
89 	uint16_t desc_idx_last = desc_idx;
90 
91 	dp  = &vq->vq_ring.desc[desc_idx];
92 	dxp = &vq->vq_descx[desc_idx];
93 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
94 	if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
95 		while (dp->flags & VRING_DESC_F_NEXT) {
96 			desc_idx_last = dp->next;
97 			dp = &vq->vq_ring.desc[dp->next];
98 		}
99 	}
100 	dxp->ndescs = 0;
101 
102 	/*
103 	 * We must append the existing free chain, if any, to the end of
104 	 * newly freed chain. If the virtqueue was completely used, then
105 	 * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
106 	 */
107 	if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
108 		vq->vq_desc_head_idx = desc_idx;
109 	} else {
110 		dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
111 		dp_tail->next = desc_idx;
112 	}
113 
114 	vq->vq_desc_tail_idx = desc_idx_last;
115 	dp->next = VQ_RING_DESC_CHAIN_END;
116 }
117 
118 static uint16_t
119 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
120 			   uint32_t *len, uint16_t num)
121 {
122 	struct vring_used_elem *uep;
123 	struct rte_mbuf *cookie;
124 	uint16_t used_idx, desc_idx;
125 	uint16_t i;
126 
127 	/*  Caller does the check */
128 	for (i = 0; i < num ; i++) {
129 		used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
130 		uep = &vq->vq_ring.used->ring[used_idx];
131 		desc_idx = (uint16_t) uep->id;
132 		len[i] = uep->len;
133 		cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
134 
135 		if (unlikely(cookie == NULL)) {
136 			PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
137 				vq->vq_used_cons_idx);
138 			break;
139 		}
140 
141 		rte_prefetch0(cookie);
142 		rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
143 		rx_pkts[i]  = cookie;
144 		vq->vq_used_cons_idx++;
145 		vq_ring_free_chain(vq, desc_idx);
146 		vq->vq_descx[desc_idx].cookie = NULL;
147 	}
148 
149 	return i;
150 }
151 
152 #ifndef DEFAULT_TX_FREE_THRESH
153 #define DEFAULT_TX_FREE_THRESH 32
154 #endif
155 
156 /* Cleanup from completed transmits. */
157 static void
158 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
159 {
160 	uint16_t i, used_idx, desc_idx;
161 	for (i = 0; i < num; i++) {
162 		struct vring_used_elem *uep;
163 		struct vq_desc_extra *dxp;
164 
165 		used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
166 		uep = &vq->vq_ring.used->ring[used_idx];
167 
168 		desc_idx = (uint16_t) uep->id;
169 		dxp = &vq->vq_descx[desc_idx];
170 		vq->vq_used_cons_idx++;
171 		vq_ring_free_chain(vq, desc_idx);
172 
173 		if (dxp->cookie != NULL) {
174 			rte_pktmbuf_free(dxp->cookie);
175 			dxp->cookie = NULL;
176 		}
177 	}
178 }
179 
180 
181 static inline int
182 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
183 {
184 	struct vq_desc_extra *dxp;
185 	struct virtio_hw *hw = vq->hw;
186 	struct vring_desc *start_dp;
187 	uint16_t needed = 1;
188 	uint16_t head_idx, idx;
189 
190 	if (unlikely(vq->vq_free_cnt == 0))
191 		return -ENOSPC;
192 	if (unlikely(vq->vq_free_cnt < needed))
193 		return -EMSGSIZE;
194 
195 	head_idx = vq->vq_desc_head_idx;
196 	if (unlikely(head_idx >= vq->vq_nentries))
197 		return -EFAULT;
198 
199 	idx = head_idx;
200 	dxp = &vq->vq_descx[idx];
201 	dxp->cookie = (void *)cookie;
202 	dxp->ndescs = needed;
203 
204 	start_dp = vq->vq_ring.desc;
205 	start_dp[idx].addr =
206 		VIRTIO_MBUF_ADDR(cookie, vq) +
207 		RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
208 	start_dp[idx].len =
209 		cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
210 	start_dp[idx].flags =  VRING_DESC_F_WRITE;
211 	idx = start_dp[idx].next;
212 	vq->vq_desc_head_idx = idx;
213 	if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
214 		vq->vq_desc_tail_idx = idx;
215 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
216 	vq_update_avail_ring(vq, head_idx);
217 
218 	return 0;
219 }
220 
221 /* When doing TSO, the IP length is not included in the pseudo header
222  * checksum of the packet given to the PMD, but for virtio it is
223  * expected.
224  */
225 static void
226 virtio_tso_fix_cksum(struct rte_mbuf *m)
227 {
228 	/* common case: header is not fragmented */
229 	if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
230 			m->l4_len)) {
231 		struct ipv4_hdr *iph;
232 		struct ipv6_hdr *ip6h;
233 		struct tcp_hdr *th;
234 		uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
235 		uint32_t tmp;
236 
237 		iph = rte_pktmbuf_mtod_offset(m, struct ipv4_hdr *, m->l2_len);
238 		th = RTE_PTR_ADD(iph, m->l3_len);
239 		if ((iph->version_ihl >> 4) == 4) {
240 			iph->hdr_checksum = 0;
241 			iph->hdr_checksum = rte_ipv4_cksum(iph);
242 			ip_len = iph->total_length;
243 			ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
244 				m->l3_len);
245 		} else {
246 			ip6h = (struct ipv6_hdr *)iph;
247 			ip_paylen = ip6h->payload_len;
248 		}
249 
250 		/* calculate the new phdr checksum not including ip_paylen */
251 		prev_cksum = th->cksum;
252 		tmp = prev_cksum;
253 		tmp += ip_paylen;
254 		tmp = (tmp & 0xffff) + (tmp >> 16);
255 		new_cksum = tmp;
256 
257 		/* replace it in the packet */
258 		th->cksum = new_cksum;
259 	}
260 }
261 
262 static inline int
263 tx_offload_enabled(struct virtio_hw *hw)
264 {
265 	return vtpci_with_feature(hw, VIRTIO_NET_F_CSUM) ||
266 		vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO4) ||
267 		vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO6);
268 }
269 
270 /* avoid write operation when necessary, to lessen cache issues */
271 #define ASSIGN_UNLESS_EQUAL(var, val) do {	\
272 	if ((var) != (val))			\
273 		(var) = (val);			\
274 } while (0)
275 
276 static inline void
277 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
278 		       uint16_t needed, int use_indirect, int can_push)
279 {
280 	struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
281 	struct vq_desc_extra *dxp;
282 	struct virtqueue *vq = txvq->vq;
283 	struct vring_desc *start_dp;
284 	uint16_t seg_num = cookie->nb_segs;
285 	uint16_t head_idx, idx;
286 	uint16_t head_size = vq->hw->vtnet_hdr_size;
287 	struct virtio_net_hdr *hdr;
288 	int offload;
289 
290 	offload = tx_offload_enabled(vq->hw);
291 	head_idx = vq->vq_desc_head_idx;
292 	idx = head_idx;
293 	dxp = &vq->vq_descx[idx];
294 	dxp->cookie = (void *)cookie;
295 	dxp->ndescs = needed;
296 
297 	start_dp = vq->vq_ring.desc;
298 
299 	if (can_push) {
300 		/* prepend cannot fail, checked by caller */
301 		hdr = (struct virtio_net_hdr *)
302 			rte_pktmbuf_prepend(cookie, head_size);
303 		/* if offload disabled, it is not zeroed below, do it now */
304 		if (offload == 0) {
305 			ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
306 			ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
307 			ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
308 			ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
309 			ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
310 			ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
311 		}
312 	} else if (use_indirect) {
313 		/* setup tx ring slot to point to indirect
314 		 * descriptor list stored in reserved region.
315 		 *
316 		 * the first slot in indirect ring is already preset
317 		 * to point to the header in reserved region
318 		 */
319 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
320 			RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
321 		start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
322 		start_dp[idx].flags = VRING_DESC_F_INDIRECT;
323 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
324 
325 		/* loop below will fill in rest of the indirect elements */
326 		start_dp = txr[idx].tx_indir;
327 		idx = 1;
328 	} else {
329 		/* setup first tx ring slot to point to header
330 		 * stored in reserved region.
331 		 */
332 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
333 			RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
334 		start_dp[idx].len   = vq->hw->vtnet_hdr_size;
335 		start_dp[idx].flags = VRING_DESC_F_NEXT;
336 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
337 
338 		idx = start_dp[idx].next;
339 	}
340 
341 	/* Checksum Offload / TSO */
342 	if (offload) {
343 		if (cookie->ol_flags & PKT_TX_TCP_SEG)
344 			cookie->ol_flags |= PKT_TX_TCP_CKSUM;
345 
346 		switch (cookie->ol_flags & PKT_TX_L4_MASK) {
347 		case PKT_TX_UDP_CKSUM:
348 			hdr->csum_start = cookie->l2_len + cookie->l3_len;
349 			hdr->csum_offset = offsetof(struct udp_hdr,
350 				dgram_cksum);
351 			hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
352 			break;
353 
354 		case PKT_TX_TCP_CKSUM:
355 			hdr->csum_start = cookie->l2_len + cookie->l3_len;
356 			hdr->csum_offset = offsetof(struct tcp_hdr, cksum);
357 			hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
358 			break;
359 
360 		default:
361 			ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
362 			ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
363 			ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
364 			break;
365 		}
366 
367 		/* TCP Segmentation Offload */
368 		if (cookie->ol_flags & PKT_TX_TCP_SEG) {
369 			virtio_tso_fix_cksum(cookie);
370 			hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
371 				VIRTIO_NET_HDR_GSO_TCPV6 :
372 				VIRTIO_NET_HDR_GSO_TCPV4;
373 			hdr->gso_size = cookie->tso_segsz;
374 			hdr->hdr_len =
375 				cookie->l2_len +
376 				cookie->l3_len +
377 				cookie->l4_len;
378 		} else {
379 			ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
380 			ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
381 			ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
382 		}
383 	}
384 
385 	do {
386 		start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
387 		start_dp[idx].len   = cookie->data_len;
388 		start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
389 		idx = start_dp[idx].next;
390 	} while ((cookie = cookie->next) != NULL);
391 
392 	if (use_indirect)
393 		idx = vq->vq_ring.desc[head_idx].next;
394 
395 	vq->vq_desc_head_idx = idx;
396 	if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
397 		vq->vq_desc_tail_idx = idx;
398 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
399 	vq_update_avail_ring(vq, head_idx);
400 }
401 
402 void
403 virtio_dev_cq_start(struct rte_eth_dev *dev)
404 {
405 	struct virtio_hw *hw = dev->data->dev_private;
406 
407 	if (hw->cvq && hw->cvq->vq) {
408 		VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
409 	}
410 }
411 
412 int
413 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
414 			uint16_t queue_idx,
415 			uint16_t nb_desc,
416 			unsigned int socket_id __rte_unused,
417 			__rte_unused const struct rte_eth_rxconf *rx_conf,
418 			struct rte_mempool *mp)
419 {
420 	uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
421 	struct virtio_hw *hw = dev->data->dev_private;
422 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
423 	struct virtnet_rx *rxvq;
424 	int error, nbufs;
425 	struct rte_mbuf *m;
426 	uint16_t desc_idx;
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 
444 	/* Allocate blank mbufs for the each rx descriptor */
445 	nbufs = 0;
446 	error = ENOSPC;
447 
448 	if (hw->use_simple_rxtx) {
449 		for (desc_idx = 0; desc_idx < vq->vq_nentries;
450 		     desc_idx++) {
451 			vq->vq_ring.avail->ring[desc_idx] = desc_idx;
452 			vq->vq_ring.desc[desc_idx].flags =
453 				VRING_DESC_F_WRITE;
454 		}
455 	}
456 
457 	memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
458 	for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
459 	     desc_idx++) {
460 		vq->sw_ring[vq->vq_nentries + desc_idx] =
461 			&rxvq->fake_mbuf;
462 	}
463 
464 	while (!virtqueue_full(vq)) {
465 		m = rte_mbuf_raw_alloc(rxvq->mpool);
466 		if (m == NULL)
467 			break;
468 
469 		/* Enqueue allocated buffers */
470 		if (hw->use_simple_rxtx)
471 			error = virtqueue_enqueue_recv_refill_simple(vq, m);
472 		else
473 			error = virtqueue_enqueue_recv_refill(vq, m);
474 
475 		if (error) {
476 			rte_pktmbuf_free(m);
477 			break;
478 		}
479 		nbufs++;
480 	}
481 
482 	vq_update_avail_idx(vq);
483 
484 	PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
485 
486 	virtio_rxq_vec_setup(rxvq);
487 
488 	VIRTQUEUE_DUMP(vq);
489 
490 	return 0;
491 }
492 
493 static void
494 virtio_update_rxtx_handler(struct rte_eth_dev *dev,
495 			   const struct rte_eth_txconf *tx_conf)
496 {
497 	uint8_t use_simple_rxtx = 0;
498 	struct virtio_hw *hw = dev->data->dev_private;
499 
500 #if defined RTE_ARCH_X86
501 	if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_SSE3))
502 		use_simple_rxtx = 1;
503 #elif defined RTE_ARCH_ARM64 || defined CONFIG_RTE_ARCH_ARM
504 	if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_NEON))
505 		use_simple_rxtx = 1;
506 #endif
507 	/* Use simple rx/tx func if single segment and no offloads */
508 	if (use_simple_rxtx &&
509 	    (tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) == VIRTIO_SIMPLE_FLAGS &&
510 	    !vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
511 		PMD_INIT_LOG(INFO, "Using simple rx/tx path");
512 		dev->tx_pkt_burst = virtio_xmit_pkts_simple;
513 		dev->rx_pkt_burst = virtio_recv_pkts_vec;
514 		hw->use_simple_rxtx = use_simple_rxtx;
515 	}
516 }
517 
518 /*
519  * struct rte_eth_dev *dev: Used to update dev
520  * uint16_t nb_desc: Defaults to values read from config space
521  * unsigned int socket_id: Used to allocate memzone
522  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
523  * uint16_t queue_idx: Just used as an index in dev txq list
524  */
525 int
526 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
527 			uint16_t queue_idx,
528 			uint16_t nb_desc,
529 			unsigned int socket_id __rte_unused,
530 			const struct rte_eth_txconf *tx_conf)
531 {
532 	uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
533 	struct virtio_hw *hw = dev->data->dev_private;
534 	struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
535 	struct virtnet_tx *txvq;
536 	uint16_t tx_free_thresh;
537 	uint16_t desc_idx;
538 
539 	PMD_INIT_FUNC_TRACE();
540 
541 	virtio_update_rxtx_handler(dev, tx_conf);
542 
543 	if (nb_desc == 0 || nb_desc > vq->vq_nentries)
544 		nb_desc = vq->vq_nentries;
545 	vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
546 
547 	txvq = &vq->txq;
548 	txvq->queue_id = queue_idx;
549 
550 	tx_free_thresh = tx_conf->tx_free_thresh;
551 	if (tx_free_thresh == 0)
552 		tx_free_thresh =
553 			RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
554 
555 	if (tx_free_thresh >= (vq->vq_nentries - 3)) {
556 		RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
557 			"number of TX entries minus 3 (%u)."
558 			" (tx_free_thresh=%u port=%u queue=%u)\n",
559 			vq->vq_nentries - 3,
560 			tx_free_thresh, dev->data->port_id, queue_idx);
561 		return -EINVAL;
562 	}
563 
564 	vq->vq_free_thresh = tx_free_thresh;
565 
566 	if (hw->use_simple_rxtx) {
567 		uint16_t mid_idx  = vq->vq_nentries >> 1;
568 
569 		for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
570 			vq->vq_ring.avail->ring[desc_idx] =
571 				desc_idx + mid_idx;
572 			vq->vq_ring.desc[desc_idx + mid_idx].next =
573 				desc_idx;
574 			vq->vq_ring.desc[desc_idx + mid_idx].addr =
575 				txvq->virtio_net_hdr_mem +
576 				offsetof(struct virtio_tx_region, tx_hdr);
577 			vq->vq_ring.desc[desc_idx + mid_idx].len =
578 				vq->hw->vtnet_hdr_size;
579 			vq->vq_ring.desc[desc_idx + mid_idx].flags =
580 				VRING_DESC_F_NEXT;
581 			vq->vq_ring.desc[desc_idx].flags = 0;
582 		}
583 		for (desc_idx = mid_idx; desc_idx < vq->vq_nentries;
584 		     desc_idx++)
585 			vq->vq_ring.avail->ring[desc_idx] = desc_idx;
586 	}
587 
588 	VIRTQUEUE_DUMP(vq);
589 
590 	dev->data->tx_queues[queue_idx] = txvq;
591 	return 0;
592 }
593 
594 static void
595 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
596 {
597 	int error;
598 	/*
599 	 * Requeue the discarded mbuf. This should always be
600 	 * successful since it was just dequeued.
601 	 */
602 	error = virtqueue_enqueue_recv_refill(vq, m);
603 	if (unlikely(error)) {
604 		RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
605 		rte_pktmbuf_free(m);
606 	}
607 }
608 
609 static void
610 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
611 {
612 	uint32_t s = mbuf->pkt_len;
613 	struct ether_addr *ea;
614 
615 	if (s == 64) {
616 		stats->size_bins[1]++;
617 	} else if (s > 64 && s < 1024) {
618 		uint32_t bin;
619 
620 		/* count zeros, and offset into correct bin */
621 		bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
622 		stats->size_bins[bin]++;
623 	} else {
624 		if (s < 64)
625 			stats->size_bins[0]++;
626 		else if (s < 1519)
627 			stats->size_bins[6]++;
628 		else if (s >= 1519)
629 			stats->size_bins[7]++;
630 	}
631 
632 	ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
633 	if (is_multicast_ether_addr(ea)) {
634 		if (is_broadcast_ether_addr(ea))
635 			stats->broadcast++;
636 		else
637 			stats->multicast++;
638 	}
639 }
640 
641 /* Optionally fill offload information in structure */
642 static int
643 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
644 {
645 	struct rte_net_hdr_lens hdr_lens;
646 	uint32_t hdrlen, ptype;
647 	int l4_supported = 0;
648 
649 	/* nothing to do */
650 	if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
651 		return 0;
652 
653 	m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
654 
655 	ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
656 	m->packet_type = ptype;
657 	if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
658 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
659 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
660 		l4_supported = 1;
661 
662 	if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
663 		hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
664 		if (hdr->csum_start <= hdrlen && l4_supported) {
665 			m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
666 		} else {
667 			/* Unknown proto or tunnel, do sw cksum. We can assume
668 			 * the cksum field is in the first segment since the
669 			 * buffers we provided to the host are large enough.
670 			 * In case of SCTP, this will be wrong since it's a CRC
671 			 * but there's nothing we can do.
672 			 */
673 			uint16_t csum, off;
674 
675 			rte_raw_cksum_mbuf(m, hdr->csum_start,
676 				rte_pktmbuf_pkt_len(m) - hdr->csum_start,
677 				&csum);
678 			if (likely(csum != 0xffff))
679 				csum = ~csum;
680 			off = hdr->csum_offset + hdr->csum_start;
681 			if (rte_pktmbuf_data_len(m) >= off + 1)
682 				*rte_pktmbuf_mtod_offset(m, uint16_t *,
683 					off) = csum;
684 		}
685 	} else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
686 		m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
687 	}
688 
689 	/* GSO request, save required information in mbuf */
690 	if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
691 		/* Check unsupported modes */
692 		if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
693 		    (hdr->gso_size == 0)) {
694 			return -EINVAL;
695 		}
696 
697 		/* Update mss lengthes in mbuf */
698 		m->tso_segsz = hdr->gso_size;
699 		switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
700 			case VIRTIO_NET_HDR_GSO_TCPV4:
701 			case VIRTIO_NET_HDR_GSO_TCPV6:
702 				m->ol_flags |= PKT_RX_LRO | \
703 					PKT_RX_L4_CKSUM_NONE;
704 				break;
705 			default:
706 				return -EINVAL;
707 		}
708 	}
709 
710 	return 0;
711 }
712 
713 static inline int
714 rx_offload_enabled(struct virtio_hw *hw)
715 {
716 	return vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_CSUM) ||
717 		vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO4) ||
718 		vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO6);
719 }
720 
721 #define VIRTIO_MBUF_BURST_SZ 64
722 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
723 uint16_t
724 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
725 {
726 	struct virtnet_rx *rxvq = rx_queue;
727 	struct virtqueue *vq = rxvq->vq;
728 	struct virtio_hw *hw;
729 	struct rte_mbuf *rxm, *new_mbuf;
730 	uint16_t nb_used, num, nb_rx;
731 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
732 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
733 	int error;
734 	uint32_t i, nb_enqueued;
735 	uint32_t hdr_size;
736 	int offload;
737 	struct virtio_net_hdr *hdr;
738 
739 	nb_used = VIRTQUEUE_NUSED(vq);
740 
741 	virtio_rmb();
742 
743 	num = (uint16_t)(likely(nb_used <= nb_pkts) ? nb_used : nb_pkts);
744 	num = (uint16_t)(likely(num <= VIRTIO_MBUF_BURST_SZ) ? num : VIRTIO_MBUF_BURST_SZ);
745 	if (likely(num > DESC_PER_CACHELINE))
746 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
747 
748 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
749 	PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
750 
751 	hw = vq->hw;
752 	nb_rx = 0;
753 	nb_enqueued = 0;
754 	hdr_size = hw->vtnet_hdr_size;
755 	offload = rx_offload_enabled(hw);
756 
757 	for (i = 0; i < num ; i++) {
758 		rxm = rcv_pkts[i];
759 
760 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
761 
762 		if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
763 			PMD_RX_LOG(ERR, "Packet drop");
764 			nb_enqueued++;
765 			virtio_discard_rxbuf(vq, rxm);
766 			rxvq->stats.errors++;
767 			continue;
768 		}
769 
770 		rxm->port = rxvq->port_id;
771 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
772 		rxm->ol_flags = 0;
773 		rxm->vlan_tci = 0;
774 
775 		rxm->nb_segs = 1;
776 		rxm->next = NULL;
777 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
778 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
779 
780 		hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
781 			RTE_PKTMBUF_HEADROOM - hdr_size);
782 
783 		if (hw->vlan_strip)
784 			rte_vlan_strip(rxm);
785 
786 		if (offload && virtio_rx_offload(rxm, hdr) < 0) {
787 			virtio_discard_rxbuf(vq, rxm);
788 			rxvq->stats.errors++;
789 			continue;
790 		}
791 
792 		VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
793 
794 		rx_pkts[nb_rx++] = rxm;
795 
796 		rxvq->stats.bytes += rxm->pkt_len;
797 		virtio_update_packet_stats(&rxvq->stats, rxm);
798 	}
799 
800 	rxvq->stats.packets += nb_rx;
801 
802 	/* Allocate new mbuf for the used descriptor */
803 	error = ENOSPC;
804 	while (likely(!virtqueue_full(vq))) {
805 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
806 		if (unlikely(new_mbuf == NULL)) {
807 			struct rte_eth_dev *dev
808 				= &rte_eth_devices[rxvq->port_id];
809 			dev->data->rx_mbuf_alloc_failed++;
810 			break;
811 		}
812 		error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
813 		if (unlikely(error)) {
814 			rte_pktmbuf_free(new_mbuf);
815 			break;
816 		}
817 		nb_enqueued++;
818 	}
819 
820 	if (likely(nb_enqueued)) {
821 		vq_update_avail_idx(vq);
822 
823 		if (unlikely(virtqueue_kick_prepare(vq))) {
824 			virtqueue_notify(vq);
825 			PMD_RX_LOG(DEBUG, "Notified");
826 		}
827 	}
828 
829 	return nb_rx;
830 }
831 
832 uint16_t
833 virtio_recv_mergeable_pkts(void *rx_queue,
834 			struct rte_mbuf **rx_pkts,
835 			uint16_t nb_pkts)
836 {
837 	struct virtnet_rx *rxvq = rx_queue;
838 	struct virtqueue *vq = rxvq->vq;
839 	struct virtio_hw *hw;
840 	struct rte_mbuf *rxm, *new_mbuf;
841 	uint16_t nb_used, num, nb_rx;
842 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
843 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
844 	struct rte_mbuf *prev;
845 	int error;
846 	uint32_t i, nb_enqueued;
847 	uint32_t seg_num;
848 	uint16_t extra_idx;
849 	uint32_t seg_res;
850 	uint32_t hdr_size;
851 	int offload;
852 
853 	nb_used = VIRTQUEUE_NUSED(vq);
854 
855 	virtio_rmb();
856 
857 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
858 
859 	hw = vq->hw;
860 	nb_rx = 0;
861 	i = 0;
862 	nb_enqueued = 0;
863 	seg_num = 0;
864 	extra_idx = 0;
865 	seg_res = 0;
866 	hdr_size = hw->vtnet_hdr_size;
867 	offload = rx_offload_enabled(hw);
868 
869 	while (i < nb_used) {
870 		struct virtio_net_hdr_mrg_rxbuf *header;
871 
872 		if (nb_rx == nb_pkts)
873 			break;
874 
875 		num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
876 		if (num != 1)
877 			continue;
878 
879 		i++;
880 
881 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
882 		PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
883 
884 		rxm = rcv_pkts[0];
885 
886 		if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
887 			PMD_RX_LOG(ERR, "Packet drop");
888 			nb_enqueued++;
889 			virtio_discard_rxbuf(vq, rxm);
890 			rxvq->stats.errors++;
891 			continue;
892 		}
893 
894 		header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
895 			RTE_PKTMBUF_HEADROOM - hdr_size);
896 		seg_num = header->num_buffers;
897 
898 		if (seg_num == 0)
899 			seg_num = 1;
900 
901 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
902 		rxm->nb_segs = seg_num;
903 		rxm->next = NULL;
904 		rxm->ol_flags = 0;
905 		rxm->vlan_tci = 0;
906 		rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
907 		rxm->data_len = (uint16_t)(len[0] - hdr_size);
908 
909 		rxm->port = rxvq->port_id;
910 		rx_pkts[nb_rx] = rxm;
911 		prev = rxm;
912 
913 		if (offload && virtio_rx_offload(rxm, &header->hdr) < 0) {
914 			virtio_discard_rxbuf(vq, rxm);
915 			rxvq->stats.errors++;
916 			continue;
917 		}
918 
919 		seg_res = seg_num - 1;
920 
921 		while (seg_res != 0) {
922 			/*
923 			 * Get extra segments for current uncompleted packet.
924 			 */
925 			uint16_t  rcv_cnt =
926 				RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
927 			if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
928 				uint32_t rx_num =
929 					virtqueue_dequeue_burst_rx(vq,
930 					rcv_pkts, len, rcv_cnt);
931 				i += rx_num;
932 				rcv_cnt = rx_num;
933 			} else {
934 				PMD_RX_LOG(ERR,
935 					   "No enough segments for packet.");
936 				nb_enqueued++;
937 				virtio_discard_rxbuf(vq, rxm);
938 				rxvq->stats.errors++;
939 				break;
940 			}
941 
942 			extra_idx = 0;
943 
944 			while (extra_idx < rcv_cnt) {
945 				rxm = rcv_pkts[extra_idx];
946 
947 				rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
948 				rxm->next = NULL;
949 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
950 				rxm->data_len = (uint16_t)(len[extra_idx]);
951 
952 				if (prev)
953 					prev->next = rxm;
954 
955 				prev = rxm;
956 				rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
957 				extra_idx++;
958 			};
959 			seg_res -= rcv_cnt;
960 		}
961 
962 		if (hw->vlan_strip)
963 			rte_vlan_strip(rx_pkts[nb_rx]);
964 
965 		VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
966 			rx_pkts[nb_rx]->data_len);
967 
968 		rxvq->stats.bytes += rx_pkts[nb_rx]->pkt_len;
969 		virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
970 		nb_rx++;
971 	}
972 
973 	rxvq->stats.packets += nb_rx;
974 
975 	/* Allocate new mbuf for the used descriptor */
976 	error = ENOSPC;
977 	while (likely(!virtqueue_full(vq))) {
978 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
979 		if (unlikely(new_mbuf == NULL)) {
980 			struct rte_eth_dev *dev
981 				= &rte_eth_devices[rxvq->port_id];
982 			dev->data->rx_mbuf_alloc_failed++;
983 			break;
984 		}
985 		error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
986 		if (unlikely(error)) {
987 			rte_pktmbuf_free(new_mbuf);
988 			break;
989 		}
990 		nb_enqueued++;
991 	}
992 
993 	if (likely(nb_enqueued)) {
994 		vq_update_avail_idx(vq);
995 
996 		if (unlikely(virtqueue_kick_prepare(vq))) {
997 			virtqueue_notify(vq);
998 			PMD_RX_LOG(DEBUG, "Notified");
999 		}
1000 	}
1001 
1002 	return nb_rx;
1003 }
1004 
1005 uint16_t
1006 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1007 {
1008 	struct virtnet_tx *txvq = tx_queue;
1009 	struct virtqueue *vq = txvq->vq;
1010 	struct virtio_hw *hw = vq->hw;
1011 	uint16_t hdr_size = hw->vtnet_hdr_size;
1012 	uint16_t nb_used, nb_tx;
1013 	int error;
1014 
1015 	if (unlikely(nb_pkts < 1))
1016 		return nb_pkts;
1017 
1018 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1019 	nb_used = VIRTQUEUE_NUSED(vq);
1020 
1021 	virtio_rmb();
1022 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1023 		virtio_xmit_cleanup(vq, nb_used);
1024 
1025 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1026 		struct rte_mbuf *txm = tx_pkts[nb_tx];
1027 		int can_push = 0, use_indirect = 0, slots, need;
1028 
1029 		/* Do VLAN tag insertion */
1030 		if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1031 			error = rte_vlan_insert(&txm);
1032 			if (unlikely(error)) {
1033 				rte_pktmbuf_free(txm);
1034 				continue;
1035 			}
1036 		}
1037 
1038 		/* optimize ring usage */
1039 		if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1040 		      vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1041 		    rte_mbuf_refcnt_read(txm) == 1 &&
1042 		    RTE_MBUF_DIRECT(txm) &&
1043 		    txm->nb_segs == 1 &&
1044 		    rte_pktmbuf_headroom(txm) >= hdr_size &&
1045 		    rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1046 				   __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1047 			can_push = 1;
1048 		else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1049 			 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1050 			use_indirect = 1;
1051 
1052 		/* How many main ring entries are needed to this Tx?
1053 		 * any_layout => number of segments
1054 		 * indirect   => 1
1055 		 * default    => number of segments + 1
1056 		 */
1057 		slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1058 		need = slots - vq->vq_free_cnt;
1059 
1060 		/* Positive value indicates it need free vring descriptors */
1061 		if (unlikely(need > 0)) {
1062 			nb_used = VIRTQUEUE_NUSED(vq);
1063 			virtio_rmb();
1064 			need = RTE_MIN(need, (int)nb_used);
1065 
1066 			virtio_xmit_cleanup(vq, need);
1067 			need = slots - vq->vq_free_cnt;
1068 			if (unlikely(need > 0)) {
1069 				PMD_TX_LOG(ERR,
1070 					   "No free tx descriptors to transmit");
1071 				break;
1072 			}
1073 		}
1074 
1075 		/* Enqueue Packet buffers */
1076 		virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
1077 
1078 		txvq->stats.bytes += txm->pkt_len;
1079 		virtio_update_packet_stats(&txvq->stats, txm);
1080 	}
1081 
1082 	txvq->stats.packets += nb_tx;
1083 
1084 	if (likely(nb_tx)) {
1085 		vq_update_avail_idx(vq);
1086 
1087 		if (unlikely(virtqueue_kick_prepare(vq))) {
1088 			virtqueue_notify(vq);
1089 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1090 		}
1091 	}
1092 
1093 	return nb_tx;
1094 }
1095