xref: /dpdk/drivers/net/virtio/virtio_rxtx.c (revision 2f45703c17acb943aaded9f79676fd56a72542b2)
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 
54 #include "virtio_logs.h"
55 #include "virtio_ethdev.h"
56 #include "virtio_pci.h"
57 #include "virtqueue.h"
58 #include "virtio_rxtx.h"
59 
60 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
61 #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
62 #else
63 #define  VIRTIO_DUMP_PACKET(m, len) do { } while (0)
64 #endif
65 
66 
67 #define VIRTIO_SIMPLE_FLAGS ((uint32_t)ETH_TXQ_FLAGS_NOMULTSEGS | \
68 	ETH_TXQ_FLAGS_NOOFFLOADS)
69 
70 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
71 static int use_simple_rxtx;
72 #endif
73 
74 static void
75 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
76 {
77 	struct vring_desc *dp, *dp_tail;
78 	struct vq_desc_extra *dxp;
79 	uint16_t desc_idx_last = desc_idx;
80 
81 	dp  = &vq->vq_ring.desc[desc_idx];
82 	dxp = &vq->vq_descx[desc_idx];
83 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
84 	if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
85 		while (dp->flags & VRING_DESC_F_NEXT) {
86 			desc_idx_last = dp->next;
87 			dp = &vq->vq_ring.desc[dp->next];
88 		}
89 	}
90 	dxp->ndescs = 0;
91 
92 	/*
93 	 * We must append the existing free chain, if any, to the end of
94 	 * newly freed chain. If the virtqueue was completely used, then
95 	 * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
96 	 */
97 	if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
98 		vq->vq_desc_head_idx = desc_idx;
99 	} else {
100 		dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
101 		dp_tail->next = desc_idx;
102 	}
103 
104 	vq->vq_desc_tail_idx = desc_idx_last;
105 	dp->next = VQ_RING_DESC_CHAIN_END;
106 }
107 
108 static uint16_t
109 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
110 			   uint32_t *len, uint16_t num)
111 {
112 	struct vring_used_elem *uep;
113 	struct rte_mbuf *cookie;
114 	uint16_t used_idx, desc_idx;
115 	uint16_t i;
116 
117 	/*  Caller does the check */
118 	for (i = 0; i < num ; i++) {
119 		used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
120 		uep = &vq->vq_ring.used->ring[used_idx];
121 		desc_idx = (uint16_t) uep->id;
122 		len[i] = uep->len;
123 		cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
124 
125 		if (unlikely(cookie == NULL)) {
126 			PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u\n",
127 				vq->vq_used_cons_idx);
128 			break;
129 		}
130 
131 		rte_prefetch0(cookie);
132 		rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
133 		rx_pkts[i]  = cookie;
134 		vq->vq_used_cons_idx++;
135 		vq_ring_free_chain(vq, desc_idx);
136 		vq->vq_descx[desc_idx].cookie = NULL;
137 	}
138 
139 	return i;
140 }
141 
142 #ifndef DEFAULT_TX_FREE_THRESH
143 #define DEFAULT_TX_FREE_THRESH 32
144 #endif
145 
146 /* Cleanup from completed transmits. */
147 static void
148 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
149 {
150 	uint16_t i, used_idx, desc_idx;
151 	for (i = 0; i < num; i++) {
152 		struct vring_used_elem *uep;
153 		struct vq_desc_extra *dxp;
154 
155 		used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
156 		uep = &vq->vq_ring.used->ring[used_idx];
157 
158 		desc_idx = (uint16_t) uep->id;
159 		dxp = &vq->vq_descx[desc_idx];
160 		vq->vq_used_cons_idx++;
161 		vq_ring_free_chain(vq, desc_idx);
162 
163 		if (dxp->cookie != NULL) {
164 			rte_pktmbuf_free(dxp->cookie);
165 			dxp->cookie = NULL;
166 		}
167 	}
168 }
169 
170 
171 static inline int
172 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
173 {
174 	struct vq_desc_extra *dxp;
175 	struct virtio_hw *hw = vq->hw;
176 	struct vring_desc *start_dp;
177 	uint16_t needed = 1;
178 	uint16_t head_idx, idx;
179 
180 	if (unlikely(vq->vq_free_cnt == 0))
181 		return -ENOSPC;
182 	if (unlikely(vq->vq_free_cnt < needed))
183 		return -EMSGSIZE;
184 
185 	head_idx = vq->vq_desc_head_idx;
186 	if (unlikely(head_idx >= vq->vq_nentries))
187 		return -EFAULT;
188 
189 	idx = head_idx;
190 	dxp = &vq->vq_descx[idx];
191 	dxp->cookie = (void *)cookie;
192 	dxp->ndescs = needed;
193 
194 	start_dp = vq->vq_ring.desc;
195 	start_dp[idx].addr =
196 		VIRTIO_MBUF_ADDR(cookie, vq) +
197 		RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
198 	start_dp[idx].len =
199 		cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
200 	start_dp[idx].flags =  VRING_DESC_F_WRITE;
201 	idx = start_dp[idx].next;
202 	vq->vq_desc_head_idx = idx;
203 	if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
204 		vq->vq_desc_tail_idx = idx;
205 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
206 	vq_update_avail_ring(vq, head_idx);
207 
208 	return 0;
209 }
210 
211 static inline void
212 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
213 		       uint16_t needed, int use_indirect, int can_push)
214 {
215 	struct vq_desc_extra *dxp;
216 	struct virtqueue *vq = txvq->vq;
217 	struct vring_desc *start_dp;
218 	uint16_t seg_num = cookie->nb_segs;
219 	uint16_t head_idx, idx;
220 	uint16_t head_size = vq->hw->vtnet_hdr_size;
221 	unsigned long offs;
222 
223 	head_idx = vq->vq_desc_head_idx;
224 	idx = head_idx;
225 	dxp = &vq->vq_descx[idx];
226 	dxp->cookie = (void *)cookie;
227 	dxp->ndescs = needed;
228 
229 	start_dp = vq->vq_ring.desc;
230 
231 	if (can_push) {
232 		/* put on zero'd transmit header (no offloads) */
233 		void *hdr = rte_pktmbuf_prepend(cookie, head_size);
234 
235 		memset(hdr, 0, head_size);
236 	} else if (use_indirect) {
237 		/* setup tx ring slot to point to indirect
238 		 * descriptor list stored in reserved region.
239 		 *
240 		 * the first slot in indirect ring is already preset
241 		 * to point to the header in reserved region
242 		 */
243 		struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
244 
245 		offs = idx * sizeof(struct virtio_tx_region)
246 			+ offsetof(struct virtio_tx_region, tx_indir);
247 
248 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem + offs;
249 		start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
250 		start_dp[idx].flags = VRING_DESC_F_INDIRECT;
251 
252 		/* loop below will fill in rest of the indirect elements */
253 		start_dp = txr[idx].tx_indir;
254 		idx = 1;
255 	} else {
256 		/* setup first tx ring slot to point to header
257 		 * stored in reserved region.
258 		 */
259 		offs = idx * sizeof(struct virtio_tx_region)
260 			+ offsetof(struct virtio_tx_region, tx_hdr);
261 
262 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem + offs;
263 		start_dp[idx].len   = vq->hw->vtnet_hdr_size;
264 		start_dp[idx].flags = VRING_DESC_F_NEXT;
265 		idx = start_dp[idx].next;
266 	}
267 
268 	do {
269 		start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
270 		start_dp[idx].len   = cookie->data_len;
271 		start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
272 		idx = start_dp[idx].next;
273 	} while ((cookie = cookie->next) != NULL);
274 
275 	if (use_indirect)
276 		idx = vq->vq_ring.desc[head_idx].next;
277 
278 	vq->vq_desc_head_idx = idx;
279 	if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
280 		vq->vq_desc_tail_idx = idx;
281 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
282 	vq_update_avail_ring(vq, head_idx);
283 }
284 
285 static void
286 virtio_dev_vring_start(struct virtqueue *vq)
287 {
288 	int size = vq->vq_nentries;
289 	struct vring *vr = &vq->vq_ring;
290 	uint8_t *ring_mem = vq->vq_ring_virt_mem;
291 
292 	PMD_INIT_FUNC_TRACE();
293 
294 	/*
295 	 * Reinitialise since virtio port might have been stopped and restarted
296 	 */
297 	memset(vq->vq_ring_virt_mem, 0, vq->vq_ring_size);
298 	vring_init(vr, size, ring_mem, VIRTIO_PCI_VRING_ALIGN);
299 	vq->vq_used_cons_idx = 0;
300 	vq->vq_desc_head_idx = 0;
301 	vq->vq_avail_idx = 0;
302 	vq->vq_desc_tail_idx = (uint16_t)(vq->vq_nentries - 1);
303 	vq->vq_free_cnt = vq->vq_nentries;
304 	memset(vq->vq_descx, 0, sizeof(struct vq_desc_extra) * vq->vq_nentries);
305 
306 	vring_desc_init(vr->desc, size);
307 
308 	/*
309 	 * Disable device(host) interrupting guest
310 	 */
311 	virtqueue_disable_intr(vq);
312 }
313 
314 void
315 virtio_dev_cq_start(struct rte_eth_dev *dev)
316 {
317 	struct virtio_hw *hw = dev->data->dev_private;
318 
319 	if (hw->cvq && hw->cvq->vq) {
320 		virtio_dev_vring_start(hw->cvq->vq);
321 		VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
322 	}
323 }
324 
325 void
326 virtio_dev_rxtx_start(struct rte_eth_dev *dev)
327 {
328 	/*
329 	 * Start receive and transmit vrings
330 	 * -	Setup vring structure for all queues
331 	 * -	Initialize descriptor for the rx vring
332 	 * -	Allocate blank mbufs for the each rx descriptor
333 	 *
334 	 */
335 	uint16_t i;
336 	uint16_t desc_idx;
337 
338 	PMD_INIT_FUNC_TRACE();
339 
340 	/* Start rx vring. */
341 	for (i = 0; i < dev->data->nb_rx_queues; i++) {
342 		struct virtnet_rx *rxvq = dev->data->rx_queues[i];
343 		struct virtqueue *vq = rxvq->vq;
344 		int error, nbufs;
345 		struct rte_mbuf *m;
346 
347 		virtio_dev_vring_start(vq);
348 		if (rxvq->mpool == NULL) {
349 			rte_exit(EXIT_FAILURE,
350 				"Cannot allocate mbufs for rx virtqueue");
351 		}
352 
353 		/* Allocate blank mbufs for the each rx descriptor */
354 		nbufs = 0;
355 		error = ENOSPC;
356 
357 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
358 		if (use_simple_rxtx) {
359 			for (desc_idx = 0; desc_idx < vq->vq_nentries;
360 			     desc_idx++) {
361 				vq->vq_ring.avail->ring[desc_idx] = desc_idx;
362 				vq->vq_ring.desc[desc_idx].flags =
363 					VRING_DESC_F_WRITE;
364 			}
365 		}
366 #endif
367 		memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
368 		for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
369 		     desc_idx++) {
370 			vq->sw_ring[vq->vq_nentries + desc_idx] =
371 				&rxvq->fake_mbuf;
372 		}
373 
374 		while (!virtqueue_full(vq)) {
375 			m = rte_mbuf_raw_alloc(rxvq->mpool);
376 			if (m == NULL)
377 				break;
378 
379 			/******************************************
380 			*         Enqueue allocated buffers        *
381 			*******************************************/
382 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
383 			if (use_simple_rxtx)
384 				error = virtqueue_enqueue_recv_refill_simple(vq, m);
385 			else
386 #endif
387 				error = virtqueue_enqueue_recv_refill(vq, m);
388 			if (error) {
389 				rte_pktmbuf_free(m);
390 				break;
391 			}
392 			nbufs++;
393 		}
394 
395 		vq_update_avail_idx(vq);
396 
397 		PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
398 
399 		VIRTQUEUE_DUMP(vq);
400 	}
401 
402 	/* Start tx vring. */
403 	for (i = 0; i < dev->data->nb_tx_queues; i++) {
404 		struct virtnet_tx *txvq = dev->data->tx_queues[i];
405 		struct virtqueue *vq = txvq->vq;
406 
407 		virtio_dev_vring_start(vq);
408 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
409 		if (use_simple_rxtx) {
410 			uint16_t mid_idx  = vq->vq_nentries >> 1;
411 
412 			for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
413 				vq->vq_ring.avail->ring[desc_idx] =
414 					desc_idx + mid_idx;
415 				vq->vq_ring.desc[desc_idx + mid_idx].next =
416 					desc_idx;
417 				vq->vq_ring.desc[desc_idx + mid_idx].addr =
418 					txvq->virtio_net_hdr_mem +
419 					offsetof(struct virtio_tx_region, tx_hdr);
420 				vq->vq_ring.desc[desc_idx + mid_idx].len =
421 					vq->hw->vtnet_hdr_size;
422 				vq->vq_ring.desc[desc_idx + mid_idx].flags =
423 					VRING_DESC_F_NEXT;
424 				vq->vq_ring.desc[desc_idx].flags = 0;
425 			}
426 			for (desc_idx = mid_idx; desc_idx < vq->vq_nentries;
427 			     desc_idx++)
428 				vq->vq_ring.avail->ring[desc_idx] = desc_idx;
429 		}
430 #endif
431 		VIRTQUEUE_DUMP(vq);
432 	}
433 }
434 
435 int
436 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
437 			uint16_t queue_idx,
438 			uint16_t nb_desc,
439 			unsigned int socket_id,
440 			__rte_unused const struct rte_eth_rxconf *rx_conf,
441 			struct rte_mempool *mp)
442 {
443 	uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
444 	struct virtnet_rx *rxvq;
445 	int ret;
446 
447 	PMD_INIT_FUNC_TRACE();
448 	ret = virtio_dev_queue_setup(dev, VTNET_RQ, queue_idx, vtpci_queue_idx,
449 			nb_desc, socket_id, (void **)&rxvq);
450 	if (ret < 0) {
451 		PMD_INIT_LOG(ERR, "rvq initialization failed");
452 		return ret;
453 	}
454 
455 	/* Create mempool for rx mbuf allocation */
456 	rxvq->mpool = mp;
457 
458 	dev->data->rx_queues[queue_idx] = rxvq;
459 
460 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
461 	virtio_rxq_vec_setup(rxvq);
462 #endif
463 
464 	return 0;
465 }
466 
467 void
468 virtio_dev_rx_queue_release(void *rxq)
469 {
470 	struct virtnet_rx *rxvq = rxq;
471 	struct virtqueue *vq;
472 	const struct rte_memzone *mz;
473 
474 	if (rxvq == NULL)
475 		return;
476 
477 	/*
478 	 * rxvq is freed when vq is freed, and as mz should be freed after the
479 	 * del_queue, so we reserve the mz pointer first.
480 	 */
481 	vq = rxvq->vq;
482 	mz = rxvq->mz;
483 
484 	virtio_dev_queue_release(vq);
485 	rte_memzone_free(mz);
486 }
487 
488 /*
489  * struct rte_eth_dev *dev: Used to update dev
490  * uint16_t nb_desc: Defaults to values read from config space
491  * unsigned int socket_id: Used to allocate memzone
492  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
493  * uint16_t queue_idx: Just used as an index in dev txq list
494  */
495 int
496 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
497 			uint16_t queue_idx,
498 			uint16_t nb_desc,
499 			unsigned int socket_id,
500 			const struct rte_eth_txconf *tx_conf)
501 {
502 	uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
503 
504 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
505 	struct virtio_hw *hw = dev->data->dev_private;
506 #endif
507 	struct virtnet_tx *txvq;
508 	struct virtqueue *vq;
509 	uint16_t tx_free_thresh;
510 	int ret;
511 
512 	PMD_INIT_FUNC_TRACE();
513 
514 	if ((tx_conf->txq_flags & ETH_TXQ_FLAGS_NOXSUMS)
515 	    != ETH_TXQ_FLAGS_NOXSUMS) {
516 		PMD_INIT_LOG(ERR, "TX checksum offload not supported\n");
517 		return -EINVAL;
518 	}
519 
520 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
521 	/* Use simple rx/tx func if single segment and no offloads */
522 	if ((tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) == VIRTIO_SIMPLE_FLAGS &&
523 	     !vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
524 		PMD_INIT_LOG(INFO, "Using simple rx/tx path");
525 		dev->tx_pkt_burst = virtio_xmit_pkts_simple;
526 		dev->rx_pkt_burst = virtio_recv_pkts_vec;
527 		use_simple_rxtx = 1;
528 	}
529 #endif
530 
531 	ret = virtio_dev_queue_setup(dev, VTNET_TQ, queue_idx, vtpci_queue_idx,
532 			nb_desc, socket_id, (void **)&txvq);
533 	if (ret < 0) {
534 		PMD_INIT_LOG(ERR, "tvq initialization failed");
535 		return ret;
536 	}
537 	vq = txvq->vq;
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 void
560 virtio_dev_tx_queue_release(void *txq)
561 {
562 	struct virtnet_tx *txvq = txq;
563 	struct virtqueue *vq;
564 	const struct rte_memzone *mz;
565 	const struct rte_memzone *hdr_mz;
566 
567 	if (txvq == NULL)
568 		return;
569 
570 	/*
571 	 * txvq is freed when vq is freed, and as mz should be freed after the
572 	 * del_queue, so we reserve the mz pointer first.
573 	 */
574 	vq = txvq->vq;
575 	mz = txvq->mz;
576 	hdr_mz = txvq->virtio_net_hdr_mz;
577 
578 	virtio_dev_queue_release(vq);
579 	rte_memzone_free(mz);
580 	rte_memzone_free(hdr_mz);
581 }
582 
583 static void
584 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
585 {
586 	int error;
587 	/*
588 	 * Requeue the discarded mbuf. This should always be
589 	 * successful since it was just dequeued.
590 	 */
591 	error = virtqueue_enqueue_recv_refill(vq, m);
592 	if (unlikely(error)) {
593 		RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
594 		rte_pktmbuf_free(m);
595 	}
596 }
597 
598 static void
599 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
600 {
601 	uint32_t s = mbuf->pkt_len;
602 	struct ether_addr *ea;
603 
604 	if (s == 64) {
605 		stats->size_bins[1]++;
606 	} else if (s > 64 && s < 1024) {
607 		uint32_t bin;
608 
609 		/* count zeros, and offset into correct bin */
610 		bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
611 		stats->size_bins[bin]++;
612 	} else {
613 		if (s < 64)
614 			stats->size_bins[0]++;
615 		else if (s < 1519)
616 			stats->size_bins[6]++;
617 		else if (s >= 1519)
618 			stats->size_bins[7]++;
619 	}
620 
621 	ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
622 	if (is_multicast_ether_addr(ea)) {
623 		if (is_broadcast_ether_addr(ea))
624 			stats->broadcast++;
625 		else
626 			stats->multicast++;
627 	}
628 }
629 
630 #define VIRTIO_MBUF_BURST_SZ 64
631 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
632 uint16_t
633 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
634 {
635 	struct virtnet_rx *rxvq = rx_queue;
636 	struct virtqueue *vq = rxvq->vq;
637 	struct virtio_hw *hw;
638 	struct rte_mbuf *rxm, *new_mbuf;
639 	uint16_t nb_used, num, nb_rx;
640 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
641 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
642 	int error;
643 	uint32_t i, nb_enqueued;
644 	uint32_t hdr_size;
645 
646 	nb_used = VIRTQUEUE_NUSED(vq);
647 
648 	virtio_rmb();
649 
650 	num = (uint16_t)(likely(nb_used <= nb_pkts) ? nb_used : nb_pkts);
651 	num = (uint16_t)(likely(num <= VIRTIO_MBUF_BURST_SZ) ? num : VIRTIO_MBUF_BURST_SZ);
652 	if (likely(num > DESC_PER_CACHELINE))
653 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
654 
655 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
656 	PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
657 
658 	hw = vq->hw;
659 	nb_rx = 0;
660 	nb_enqueued = 0;
661 	hdr_size = hw->vtnet_hdr_size;
662 
663 	for (i = 0; i < num ; i++) {
664 		rxm = rcv_pkts[i];
665 
666 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
667 
668 		if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
669 			PMD_RX_LOG(ERR, "Packet drop");
670 			nb_enqueued++;
671 			virtio_discard_rxbuf(vq, rxm);
672 			rxvq->stats.errors++;
673 			continue;
674 		}
675 
676 		rxm->port = rxvq->port_id;
677 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
678 		rxm->ol_flags = 0;
679 		rxm->vlan_tci = 0;
680 
681 		rxm->nb_segs = 1;
682 		rxm->next = NULL;
683 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
684 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
685 
686 		if (hw->vlan_strip)
687 			rte_vlan_strip(rxm);
688 
689 		VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
690 
691 		rx_pkts[nb_rx++] = rxm;
692 
693 		rxvq->stats.bytes += rx_pkts[nb_rx - 1]->pkt_len;
694 		virtio_update_packet_stats(&rxvq->stats, rxm);
695 	}
696 
697 	rxvq->stats.packets += nb_rx;
698 
699 	/* Allocate new mbuf for the used descriptor */
700 	error = ENOSPC;
701 	while (likely(!virtqueue_full(vq))) {
702 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
703 		if (unlikely(new_mbuf == NULL)) {
704 			struct rte_eth_dev *dev
705 				= &rte_eth_devices[rxvq->port_id];
706 			dev->data->rx_mbuf_alloc_failed++;
707 			break;
708 		}
709 		error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
710 		if (unlikely(error)) {
711 			rte_pktmbuf_free(new_mbuf);
712 			break;
713 		}
714 		nb_enqueued++;
715 	}
716 
717 	if (likely(nb_enqueued)) {
718 		vq_update_avail_idx(vq);
719 
720 		if (unlikely(virtqueue_kick_prepare(vq))) {
721 			virtqueue_notify(vq);
722 			PMD_RX_LOG(DEBUG, "Notified");
723 		}
724 	}
725 
726 	return nb_rx;
727 }
728 
729 uint16_t
730 virtio_recv_mergeable_pkts(void *rx_queue,
731 			struct rte_mbuf **rx_pkts,
732 			uint16_t nb_pkts)
733 {
734 	struct virtnet_rx *rxvq = rx_queue;
735 	struct virtqueue *vq = rxvq->vq;
736 	struct virtio_hw *hw;
737 	struct rte_mbuf *rxm, *new_mbuf;
738 	uint16_t nb_used, num, nb_rx;
739 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
740 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
741 	struct rte_mbuf *prev;
742 	int error;
743 	uint32_t i, nb_enqueued;
744 	uint32_t seg_num;
745 	uint16_t extra_idx;
746 	uint32_t seg_res;
747 	uint32_t hdr_size;
748 
749 	nb_used = VIRTQUEUE_NUSED(vq);
750 
751 	virtio_rmb();
752 
753 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
754 
755 	hw = vq->hw;
756 	nb_rx = 0;
757 	i = 0;
758 	nb_enqueued = 0;
759 	seg_num = 0;
760 	extra_idx = 0;
761 	seg_res = 0;
762 	hdr_size = hw->vtnet_hdr_size;
763 
764 	while (i < nb_used) {
765 		struct virtio_net_hdr_mrg_rxbuf *header;
766 
767 		if (nb_rx == nb_pkts)
768 			break;
769 
770 		num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
771 		if (num != 1)
772 			continue;
773 
774 		i++;
775 
776 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
777 		PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
778 
779 		rxm = rcv_pkts[0];
780 
781 		if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
782 			PMD_RX_LOG(ERR, "Packet drop");
783 			nb_enqueued++;
784 			virtio_discard_rxbuf(vq, rxm);
785 			rxvq->stats.errors++;
786 			continue;
787 		}
788 
789 		header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
790 			RTE_PKTMBUF_HEADROOM - hdr_size);
791 		seg_num = header->num_buffers;
792 
793 		if (seg_num == 0)
794 			seg_num = 1;
795 
796 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
797 		rxm->nb_segs = seg_num;
798 		rxm->next = NULL;
799 		rxm->ol_flags = 0;
800 		rxm->vlan_tci = 0;
801 		rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
802 		rxm->data_len = (uint16_t)(len[0] - hdr_size);
803 
804 		rxm->port = rxvq->port_id;
805 		rx_pkts[nb_rx] = rxm;
806 		prev = rxm;
807 
808 		seg_res = seg_num - 1;
809 
810 		while (seg_res != 0) {
811 			/*
812 			 * Get extra segments for current uncompleted packet.
813 			 */
814 			uint16_t  rcv_cnt =
815 				RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
816 			if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
817 				uint32_t rx_num =
818 					virtqueue_dequeue_burst_rx(vq,
819 					rcv_pkts, len, rcv_cnt);
820 				i += rx_num;
821 				rcv_cnt = rx_num;
822 			} else {
823 				PMD_RX_LOG(ERR,
824 					   "No enough segments for packet.");
825 				nb_enqueued++;
826 				virtio_discard_rxbuf(vq, rxm);
827 				rxvq->stats.errors++;
828 				break;
829 			}
830 
831 			extra_idx = 0;
832 
833 			while (extra_idx < rcv_cnt) {
834 				rxm = rcv_pkts[extra_idx];
835 
836 				rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
837 				rxm->next = NULL;
838 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
839 				rxm->data_len = (uint16_t)(len[extra_idx]);
840 
841 				if (prev)
842 					prev->next = rxm;
843 
844 				prev = rxm;
845 				rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
846 				extra_idx++;
847 			};
848 			seg_res -= rcv_cnt;
849 		}
850 
851 		if (hw->vlan_strip)
852 			rte_vlan_strip(rx_pkts[nb_rx]);
853 
854 		VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
855 			rx_pkts[nb_rx]->data_len);
856 
857 		rxvq->stats.bytes += rx_pkts[nb_rx]->pkt_len;
858 		virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
859 		nb_rx++;
860 	}
861 
862 	rxvq->stats.packets += nb_rx;
863 
864 	/* Allocate new mbuf for the used descriptor */
865 	error = ENOSPC;
866 	while (likely(!virtqueue_full(vq))) {
867 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
868 		if (unlikely(new_mbuf == NULL)) {
869 			struct rte_eth_dev *dev
870 				= &rte_eth_devices[rxvq->port_id];
871 			dev->data->rx_mbuf_alloc_failed++;
872 			break;
873 		}
874 		error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
875 		if (unlikely(error)) {
876 			rte_pktmbuf_free(new_mbuf);
877 			break;
878 		}
879 		nb_enqueued++;
880 	}
881 
882 	if (likely(nb_enqueued)) {
883 		vq_update_avail_idx(vq);
884 
885 		if (unlikely(virtqueue_kick_prepare(vq))) {
886 			virtqueue_notify(vq);
887 			PMD_RX_LOG(DEBUG, "Notified");
888 		}
889 	}
890 
891 	return nb_rx;
892 }
893 
894 uint16_t
895 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
896 {
897 	struct virtnet_tx *txvq = tx_queue;
898 	struct virtqueue *vq = txvq->vq;
899 	struct virtio_hw *hw = vq->hw;
900 	uint16_t hdr_size = hw->vtnet_hdr_size;
901 	uint16_t nb_used, nb_tx;
902 	int error;
903 
904 	if (unlikely(nb_pkts < 1))
905 		return nb_pkts;
906 
907 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
908 	nb_used = VIRTQUEUE_NUSED(vq);
909 
910 	virtio_rmb();
911 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
912 		virtio_xmit_cleanup(vq, nb_used);
913 
914 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
915 		struct rte_mbuf *txm = tx_pkts[nb_tx];
916 		int can_push = 0, use_indirect = 0, slots, need;
917 
918 		/* Do VLAN tag insertion */
919 		if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
920 			error = rte_vlan_insert(&txm);
921 			if (unlikely(error)) {
922 				rte_pktmbuf_free(txm);
923 				continue;
924 			}
925 		}
926 
927 		/* optimize ring usage */
928 		if (vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) &&
929 		    rte_mbuf_refcnt_read(txm) == 1 &&
930 		    RTE_MBUF_DIRECT(txm) &&
931 		    txm->nb_segs == 1 &&
932 		    rte_pktmbuf_headroom(txm) >= hdr_size &&
933 		    rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
934 				   __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
935 			can_push = 1;
936 		else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
937 			 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
938 			use_indirect = 1;
939 
940 		/* How many main ring entries are needed to this Tx?
941 		 * any_layout => number of segments
942 		 * indirect   => 1
943 		 * default    => number of segments + 1
944 		 */
945 		slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
946 		need = slots - vq->vq_free_cnt;
947 
948 		/* Positive value indicates it need free vring descriptors */
949 		if (unlikely(need > 0)) {
950 			nb_used = VIRTQUEUE_NUSED(vq);
951 			virtio_rmb();
952 			need = RTE_MIN(need, (int)nb_used);
953 
954 			virtio_xmit_cleanup(vq, need);
955 			need = slots - vq->vq_free_cnt;
956 			if (unlikely(need > 0)) {
957 				PMD_TX_LOG(ERR,
958 					   "No free tx descriptors to transmit");
959 				break;
960 			}
961 		}
962 
963 		/* Enqueue Packet buffers */
964 		virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
965 
966 		txvq->stats.bytes += txm->pkt_len;
967 		virtio_update_packet_stats(&txvq->stats, txm);
968 	}
969 
970 	txvq->stats.packets += nb_tx;
971 
972 	if (likely(nb_tx)) {
973 		vq_update_avail_idx(vq);
974 
975 		if (unlikely(virtqueue_kick_prepare(vq))) {
976 			virtqueue_notify(vq);
977 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
978 		}
979 	}
980 
981 	return nb_tx;
982 }
983