xref: /dpdk/drivers/net/virtio/virtio_rxtx.c (revision 01ad44fd374fb8ecb9ecc80b9bfd1f45dbaa4a1f)
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 		(uint64_t)(cookie->buf_physaddr + RTE_PKTMBUF_HEADROOM
197 		- 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  = rte_mbuf_data_dma_addr(cookie);
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 = rxvq->vq;
472 	/* rxvq is freed when vq is freed, and as mz should be freed after the
473 	 * del_queue, so we reserve the mz pointer first.
474 	 */
475 	const struct rte_memzone *mz = rxvq->mz;
476 
477 	/* no need to free rxq as vq and rxq are allocated together */
478 	virtio_dev_queue_release(vq);
479 	rte_memzone_free(mz);
480 }
481 
482 /*
483  * struct rte_eth_dev *dev: Used to update dev
484  * uint16_t nb_desc: Defaults to values read from config space
485  * unsigned int socket_id: Used to allocate memzone
486  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
487  * uint16_t queue_idx: Just used as an index in dev txq list
488  */
489 int
490 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
491 			uint16_t queue_idx,
492 			uint16_t nb_desc,
493 			unsigned int socket_id,
494 			const struct rte_eth_txconf *tx_conf)
495 {
496 	uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
497 
498 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
499 	struct virtio_hw *hw = dev->data->dev_private;
500 #endif
501 	struct virtnet_tx *txvq;
502 	struct virtqueue *vq;
503 	uint16_t tx_free_thresh;
504 	int ret;
505 
506 	PMD_INIT_FUNC_TRACE();
507 
508 	if ((tx_conf->txq_flags & ETH_TXQ_FLAGS_NOXSUMS)
509 	    != ETH_TXQ_FLAGS_NOXSUMS) {
510 		PMD_INIT_LOG(ERR, "TX checksum offload not supported\n");
511 		return -EINVAL;
512 	}
513 
514 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
515 	/* Use simple rx/tx func if single segment and no offloads */
516 	if ((tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) == VIRTIO_SIMPLE_FLAGS &&
517 	     !vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
518 		PMD_INIT_LOG(INFO, "Using simple rx/tx path");
519 		dev->tx_pkt_burst = virtio_xmit_pkts_simple;
520 		dev->rx_pkt_burst = virtio_recv_pkts_vec;
521 		use_simple_rxtx = 1;
522 	}
523 #endif
524 
525 	ret = virtio_dev_queue_setup(dev, VTNET_TQ, queue_idx, vtpci_queue_idx,
526 			nb_desc, socket_id, (void **)&txvq);
527 	if (ret < 0) {
528 		PMD_INIT_LOG(ERR, "tvq initialization failed");
529 		return ret;
530 	}
531 	vq = txvq->vq;
532 
533 	tx_free_thresh = tx_conf->tx_free_thresh;
534 	if (tx_free_thresh == 0)
535 		tx_free_thresh =
536 			RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
537 
538 	if (tx_free_thresh >= (vq->vq_nentries - 3)) {
539 		RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
540 			"number of TX entries minus 3 (%u)."
541 			" (tx_free_thresh=%u port=%u queue=%u)\n",
542 			vq->vq_nentries - 3,
543 			tx_free_thresh, dev->data->port_id, queue_idx);
544 		return -EINVAL;
545 	}
546 
547 	vq->vq_free_thresh = tx_free_thresh;
548 
549 	dev->data->tx_queues[queue_idx] = txvq;
550 	return 0;
551 }
552 
553 void
554 virtio_dev_tx_queue_release(void *txq)
555 {
556 	struct virtnet_tx *txvq = txq;
557 	struct virtqueue *vq = txvq->vq;
558 	/* txvq is freed when vq is freed, and as mz should be freed after the
559 	 * del_queue, so we reserve the mz pointer first.
560 	 */
561 	const struct rte_memzone *hdr_mz = txvq->virtio_net_hdr_mz;
562 	const struct rte_memzone *mz = txvq->mz;
563 
564 	virtio_dev_queue_release(vq);
565 	rte_memzone_free(mz);
566 	rte_memzone_free(hdr_mz);
567 }
568 
569 static void
570 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
571 {
572 	int error;
573 	/*
574 	 * Requeue the discarded mbuf. This should always be
575 	 * successful since it was just dequeued.
576 	 */
577 	error = virtqueue_enqueue_recv_refill(vq, m);
578 	if (unlikely(error)) {
579 		RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
580 		rte_pktmbuf_free(m);
581 	}
582 }
583 
584 static void
585 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
586 {
587 	uint32_t s = mbuf->pkt_len;
588 	struct ether_addr *ea;
589 
590 	if (s == 64) {
591 		stats->size_bins[1]++;
592 	} else if (s > 64 && s < 1024) {
593 		uint32_t bin;
594 
595 		/* count zeros, and offset into correct bin */
596 		bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
597 		stats->size_bins[bin]++;
598 	} else {
599 		if (s < 64)
600 			stats->size_bins[0]++;
601 		else if (s < 1519)
602 			stats->size_bins[6]++;
603 		else if (s >= 1519)
604 			stats->size_bins[7]++;
605 	}
606 
607 	ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
608 	if (is_multicast_ether_addr(ea)) {
609 		if (is_broadcast_ether_addr(ea))
610 			stats->broadcast++;
611 		else
612 			stats->multicast++;
613 	}
614 }
615 
616 #define VIRTIO_MBUF_BURST_SZ 64
617 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
618 uint16_t
619 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
620 {
621 	struct virtnet_rx *rxvq = rx_queue;
622 	struct virtqueue *vq = rxvq->vq;
623 	struct virtio_hw *hw;
624 	struct rte_mbuf *rxm, *new_mbuf;
625 	uint16_t nb_used, num, nb_rx;
626 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
627 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
628 	int error;
629 	uint32_t i, nb_enqueued;
630 	uint32_t hdr_size;
631 
632 	nb_used = VIRTQUEUE_NUSED(vq);
633 
634 	virtio_rmb();
635 
636 	num = (uint16_t)(likely(nb_used <= nb_pkts) ? nb_used : nb_pkts);
637 	num = (uint16_t)(likely(num <= VIRTIO_MBUF_BURST_SZ) ? num : VIRTIO_MBUF_BURST_SZ);
638 	if (likely(num > DESC_PER_CACHELINE))
639 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
640 
641 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
642 	PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
643 
644 	hw = vq->hw;
645 	nb_rx = 0;
646 	nb_enqueued = 0;
647 	hdr_size = hw->vtnet_hdr_size;
648 
649 	for (i = 0; i < num ; i++) {
650 		rxm = rcv_pkts[i];
651 
652 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
653 
654 		if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
655 			PMD_RX_LOG(ERR, "Packet drop");
656 			nb_enqueued++;
657 			virtio_discard_rxbuf(vq, rxm);
658 			rxvq->stats.errors++;
659 			continue;
660 		}
661 
662 		rxm->port = rxvq->port_id;
663 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
664 		rxm->ol_flags = 0;
665 		rxm->vlan_tci = 0;
666 
667 		rxm->nb_segs = 1;
668 		rxm->next = NULL;
669 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
670 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
671 
672 		if (hw->vlan_strip)
673 			rte_vlan_strip(rxm);
674 
675 		VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
676 
677 		rx_pkts[nb_rx++] = rxm;
678 
679 		rxvq->stats.bytes += rx_pkts[nb_rx - 1]->pkt_len;
680 		virtio_update_packet_stats(&rxvq->stats, rxm);
681 	}
682 
683 	rxvq->stats.packets += nb_rx;
684 
685 	/* Allocate new mbuf for the used descriptor */
686 	error = ENOSPC;
687 	while (likely(!virtqueue_full(vq))) {
688 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
689 		if (unlikely(new_mbuf == NULL)) {
690 			struct rte_eth_dev *dev
691 				= &rte_eth_devices[rxvq->port_id];
692 			dev->data->rx_mbuf_alloc_failed++;
693 			break;
694 		}
695 		error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
696 		if (unlikely(error)) {
697 			rte_pktmbuf_free(new_mbuf);
698 			break;
699 		}
700 		nb_enqueued++;
701 	}
702 
703 	if (likely(nb_enqueued)) {
704 		vq_update_avail_idx(vq);
705 
706 		if (unlikely(virtqueue_kick_prepare(vq))) {
707 			virtqueue_notify(vq);
708 			PMD_RX_LOG(DEBUG, "Notified");
709 		}
710 	}
711 
712 	return nb_rx;
713 }
714 
715 uint16_t
716 virtio_recv_mergeable_pkts(void *rx_queue,
717 			struct rte_mbuf **rx_pkts,
718 			uint16_t nb_pkts)
719 {
720 	struct virtnet_rx *rxvq = rx_queue;
721 	struct virtqueue *vq = rxvq->vq;
722 	struct virtio_hw *hw;
723 	struct rte_mbuf *rxm, *new_mbuf;
724 	uint16_t nb_used, num, nb_rx;
725 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
726 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
727 	struct rte_mbuf *prev;
728 	int error;
729 	uint32_t i, nb_enqueued;
730 	uint32_t seg_num;
731 	uint16_t extra_idx;
732 	uint32_t seg_res;
733 	uint32_t hdr_size;
734 
735 	nb_used = VIRTQUEUE_NUSED(vq);
736 
737 	virtio_rmb();
738 
739 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
740 
741 	hw = vq->hw;
742 	nb_rx = 0;
743 	i = 0;
744 	nb_enqueued = 0;
745 	seg_num = 0;
746 	extra_idx = 0;
747 	seg_res = 0;
748 	hdr_size = hw->vtnet_hdr_size;
749 
750 	while (i < nb_used) {
751 		struct virtio_net_hdr_mrg_rxbuf *header;
752 
753 		if (nb_rx == nb_pkts)
754 			break;
755 
756 		num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
757 		if (num != 1)
758 			continue;
759 
760 		i++;
761 
762 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
763 		PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
764 
765 		rxm = rcv_pkts[0];
766 
767 		if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
768 			PMD_RX_LOG(ERR, "Packet drop");
769 			nb_enqueued++;
770 			virtio_discard_rxbuf(vq, rxm);
771 			rxvq->stats.errors++;
772 			continue;
773 		}
774 
775 		header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
776 			RTE_PKTMBUF_HEADROOM - hdr_size);
777 		seg_num = header->num_buffers;
778 
779 		if (seg_num == 0)
780 			seg_num = 1;
781 
782 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
783 		rxm->nb_segs = seg_num;
784 		rxm->next = NULL;
785 		rxm->ol_flags = 0;
786 		rxm->vlan_tci = 0;
787 		rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
788 		rxm->data_len = (uint16_t)(len[0] - hdr_size);
789 
790 		rxm->port = rxvq->port_id;
791 		rx_pkts[nb_rx] = rxm;
792 		prev = rxm;
793 
794 		seg_res = seg_num - 1;
795 
796 		while (seg_res != 0) {
797 			/*
798 			 * Get extra segments for current uncompleted packet.
799 			 */
800 			uint16_t  rcv_cnt =
801 				RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
802 			if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
803 				uint32_t rx_num =
804 					virtqueue_dequeue_burst_rx(vq,
805 					rcv_pkts, len, rcv_cnt);
806 				i += rx_num;
807 				rcv_cnt = rx_num;
808 			} else {
809 				PMD_RX_LOG(ERR,
810 					   "No enough segments for packet.");
811 				nb_enqueued++;
812 				virtio_discard_rxbuf(vq, rxm);
813 				rxvq->stats.errors++;
814 				break;
815 			}
816 
817 			extra_idx = 0;
818 
819 			while (extra_idx < rcv_cnt) {
820 				rxm = rcv_pkts[extra_idx];
821 
822 				rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
823 				rxm->next = NULL;
824 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
825 				rxm->data_len = (uint16_t)(len[extra_idx]);
826 
827 				if (prev)
828 					prev->next = rxm;
829 
830 				prev = rxm;
831 				rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
832 				extra_idx++;
833 			};
834 			seg_res -= rcv_cnt;
835 		}
836 
837 		if (hw->vlan_strip)
838 			rte_vlan_strip(rx_pkts[nb_rx]);
839 
840 		VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
841 			rx_pkts[nb_rx]->data_len);
842 
843 		rxvq->stats.bytes += rx_pkts[nb_rx]->pkt_len;
844 		virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
845 		nb_rx++;
846 	}
847 
848 	rxvq->stats.packets += nb_rx;
849 
850 	/* Allocate new mbuf for the used descriptor */
851 	error = ENOSPC;
852 	while (likely(!virtqueue_full(vq))) {
853 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
854 		if (unlikely(new_mbuf == NULL)) {
855 			struct rte_eth_dev *dev
856 				= &rte_eth_devices[rxvq->port_id];
857 			dev->data->rx_mbuf_alloc_failed++;
858 			break;
859 		}
860 		error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
861 		if (unlikely(error)) {
862 			rte_pktmbuf_free(new_mbuf);
863 			break;
864 		}
865 		nb_enqueued++;
866 	}
867 
868 	if (likely(nb_enqueued)) {
869 		vq_update_avail_idx(vq);
870 
871 		if (unlikely(virtqueue_kick_prepare(vq))) {
872 			virtqueue_notify(vq);
873 			PMD_RX_LOG(DEBUG, "Notified");
874 		}
875 	}
876 
877 	return nb_rx;
878 }
879 
880 uint16_t
881 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
882 {
883 	struct virtnet_tx *txvq = tx_queue;
884 	struct virtqueue *vq = txvq->vq;
885 	struct virtio_hw *hw = vq->hw;
886 	uint16_t hdr_size = hw->vtnet_hdr_size;
887 	uint16_t nb_used, nb_tx;
888 	int error;
889 
890 	if (unlikely(nb_pkts < 1))
891 		return nb_pkts;
892 
893 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
894 	nb_used = VIRTQUEUE_NUSED(vq);
895 
896 	virtio_rmb();
897 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
898 		virtio_xmit_cleanup(vq, nb_used);
899 
900 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
901 		struct rte_mbuf *txm = tx_pkts[nb_tx];
902 		int can_push = 0, use_indirect = 0, slots, need;
903 
904 		/* Do VLAN tag insertion */
905 		if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
906 			error = rte_vlan_insert(&txm);
907 			if (unlikely(error)) {
908 				rte_pktmbuf_free(txm);
909 				continue;
910 			}
911 		}
912 
913 		/* optimize ring usage */
914 		if (vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) &&
915 		    rte_mbuf_refcnt_read(txm) == 1 &&
916 		    RTE_MBUF_DIRECT(txm) &&
917 		    txm->nb_segs == 1 &&
918 		    rte_pktmbuf_headroom(txm) >= hdr_size &&
919 		    rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
920 				   __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
921 			can_push = 1;
922 		else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
923 			 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
924 			use_indirect = 1;
925 
926 		/* How many main ring entries are needed to this Tx?
927 		 * any_layout => number of segments
928 		 * indirect   => 1
929 		 * default    => number of segments + 1
930 		 */
931 		slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
932 		need = slots - vq->vq_free_cnt;
933 
934 		/* Positive value indicates it need free vring descriptors */
935 		if (unlikely(need > 0)) {
936 			nb_used = VIRTQUEUE_NUSED(vq);
937 			virtio_rmb();
938 			need = RTE_MIN(need, (int)nb_used);
939 
940 			virtio_xmit_cleanup(vq, need);
941 			need = slots - vq->vq_free_cnt;
942 			if (unlikely(need > 0)) {
943 				PMD_TX_LOG(ERR,
944 					   "No free tx descriptors to transmit");
945 				break;
946 			}
947 		}
948 
949 		/* Enqueue Packet buffers */
950 		virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
951 
952 		txvq->stats.bytes += txm->pkt_len;
953 		virtio_update_packet_stats(&txvq->stats, txm);
954 	}
955 
956 	txvq->stats.packets += nb_tx;
957 
958 	if (likely(nb_tx)) {
959 		vq_update_avail_idx(vq);
960 
961 		if (unlikely(virtqueue_kick_prepare(vq))) {
962 			virtqueue_notify(vq);
963 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
964 		}
965 	}
966 
967 	return nb_tx;
968 }
969