xref: /dpdk/drivers/net/virtio/virtio_rxtx.c (revision ebee5594a35e2af743ff03cde7d4bac67ac772f2)
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 		MBUF_DATA_DMA_ADDR(cookie, vq->offset) - hw->vtnet_hdr_size;
197 	start_dp[idx].len =
198 		cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
199 	start_dp[idx].flags =  VRING_DESC_F_WRITE;
200 	idx = start_dp[idx].next;
201 	vq->vq_desc_head_idx = idx;
202 	if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
203 		vq->vq_desc_tail_idx = idx;
204 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
205 	vq_update_avail_ring(vq, head_idx);
206 
207 	return 0;
208 }
209 
210 static inline void
211 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
212 		       uint16_t needed, int use_indirect, int can_push)
213 {
214 	struct vq_desc_extra *dxp;
215 	struct virtqueue *vq = txvq->vq;
216 	struct vring_desc *start_dp;
217 	uint16_t seg_num = cookie->nb_segs;
218 	uint16_t head_idx, idx;
219 	uint16_t head_size = vq->hw->vtnet_hdr_size;
220 	unsigned long offs;
221 
222 	head_idx = vq->vq_desc_head_idx;
223 	idx = head_idx;
224 	dxp = &vq->vq_descx[idx];
225 	dxp->cookie = (void *)cookie;
226 	dxp->ndescs = needed;
227 
228 	start_dp = vq->vq_ring.desc;
229 
230 	if (can_push) {
231 		/* put on zero'd transmit header (no offloads) */
232 		void *hdr = rte_pktmbuf_prepend(cookie, head_size);
233 
234 		memset(hdr, 0, head_size);
235 	} else if (use_indirect) {
236 		/* setup tx ring slot to point to indirect
237 		 * descriptor list stored in reserved region.
238 		 *
239 		 * the first slot in indirect ring is already preset
240 		 * to point to the header in reserved region
241 		 */
242 		struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
243 
244 		offs = idx * sizeof(struct virtio_tx_region)
245 			+ offsetof(struct virtio_tx_region, tx_indir);
246 
247 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem + offs;
248 		start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
249 		start_dp[idx].flags = VRING_DESC_F_INDIRECT;
250 
251 		/* loop below will fill in rest of the indirect elements */
252 		start_dp = txr[idx].tx_indir;
253 		idx = 1;
254 	} else {
255 		/* setup first tx ring slot to point to header
256 		 * stored in reserved region.
257 		 */
258 		offs = idx * sizeof(struct virtio_tx_region)
259 			+ offsetof(struct virtio_tx_region, tx_hdr);
260 
261 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem + offs;
262 		start_dp[idx].len   = vq->hw->vtnet_hdr_size;
263 		start_dp[idx].flags = VRING_DESC_F_NEXT;
264 		idx = start_dp[idx].next;
265 	}
266 
267 	do {
268 		start_dp[idx].addr  = MBUF_DATA_DMA_ADDR(cookie, vq->offset);
269 		start_dp[idx].len   = cookie->data_len;
270 		start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
271 		idx = start_dp[idx].next;
272 	} while ((cookie = cookie->next) != NULL);
273 
274 	if (use_indirect)
275 		idx = vq->vq_ring.desc[head_idx].next;
276 
277 	vq->vq_desc_head_idx = idx;
278 	if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
279 		vq->vq_desc_tail_idx = idx;
280 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
281 	vq_update_avail_ring(vq, head_idx);
282 }
283 
284 static void
285 virtio_dev_vring_start(struct virtqueue *vq)
286 {
287 	int size = vq->vq_nentries;
288 	struct vring *vr = &vq->vq_ring;
289 	uint8_t *ring_mem = vq->vq_ring_virt_mem;
290 
291 	PMD_INIT_FUNC_TRACE();
292 
293 	/*
294 	 * Reinitialise since virtio port might have been stopped and restarted
295 	 */
296 	memset(vq->vq_ring_virt_mem, 0, vq->vq_ring_size);
297 	vring_init(vr, size, ring_mem, VIRTIO_PCI_VRING_ALIGN);
298 	vq->vq_used_cons_idx = 0;
299 	vq->vq_desc_head_idx = 0;
300 	vq->vq_avail_idx = 0;
301 	vq->vq_desc_tail_idx = (uint16_t)(vq->vq_nentries - 1);
302 	vq->vq_free_cnt = vq->vq_nentries;
303 	memset(vq->vq_descx, 0, sizeof(struct vq_desc_extra) * vq->vq_nentries);
304 
305 	vring_desc_init(vr->desc, size);
306 
307 	/*
308 	 * Disable device(host) interrupting guest
309 	 */
310 	virtqueue_disable_intr(vq);
311 }
312 
313 void
314 virtio_dev_cq_start(struct rte_eth_dev *dev)
315 {
316 	struct virtio_hw *hw = dev->data->dev_private;
317 
318 	if (hw->cvq && hw->cvq->vq) {
319 		virtio_dev_vring_start(hw->cvq->vq);
320 		VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
321 	}
322 }
323 
324 void
325 virtio_dev_rxtx_start(struct rte_eth_dev *dev)
326 {
327 	/*
328 	 * Start receive and transmit vrings
329 	 * -	Setup vring structure for all queues
330 	 * -	Initialize descriptor for the rx vring
331 	 * -	Allocate blank mbufs for the each rx descriptor
332 	 *
333 	 */
334 	uint16_t i;
335 	uint16_t desc_idx;
336 
337 	PMD_INIT_FUNC_TRACE();
338 
339 	/* Start rx vring. */
340 	for (i = 0; i < dev->data->nb_rx_queues; i++) {
341 		struct virtnet_rx *rxvq = dev->data->rx_queues[i];
342 		struct virtqueue *vq = rxvq->vq;
343 		int error, nbufs;
344 		struct rte_mbuf *m;
345 
346 		virtio_dev_vring_start(vq);
347 		if (rxvq->mpool == NULL) {
348 			rte_exit(EXIT_FAILURE,
349 				"Cannot allocate mbufs for rx virtqueue");
350 		}
351 
352 		/* Allocate blank mbufs for the each rx descriptor */
353 		nbufs = 0;
354 		error = ENOSPC;
355 
356 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
357 		if (use_simple_rxtx) {
358 			for (desc_idx = 0; desc_idx < vq->vq_nentries;
359 			     desc_idx++) {
360 				vq->vq_ring.avail->ring[desc_idx] = desc_idx;
361 				vq->vq_ring.desc[desc_idx].flags =
362 					VRING_DESC_F_WRITE;
363 			}
364 		}
365 #endif
366 		memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
367 		for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
368 		     desc_idx++) {
369 			vq->sw_ring[vq->vq_nentries + desc_idx] =
370 				&rxvq->fake_mbuf;
371 		}
372 
373 		while (!virtqueue_full(vq)) {
374 			m = rte_mbuf_raw_alloc(rxvq->mpool);
375 			if (m == NULL)
376 				break;
377 
378 			/******************************************
379 			*         Enqueue allocated buffers        *
380 			*******************************************/
381 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
382 			if (use_simple_rxtx)
383 				error = virtqueue_enqueue_recv_refill_simple(vq, m);
384 			else
385 #endif
386 				error = virtqueue_enqueue_recv_refill(vq, m);
387 			if (error) {
388 				rte_pktmbuf_free(m);
389 				break;
390 			}
391 			nbufs++;
392 		}
393 
394 		vq_update_avail_idx(vq);
395 
396 		PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
397 
398 		VIRTQUEUE_DUMP(vq);
399 	}
400 
401 	/* Start tx vring. */
402 	for (i = 0; i < dev->data->nb_tx_queues; i++) {
403 		struct virtnet_tx *txvq = dev->data->tx_queues[i];
404 		struct virtqueue *vq = txvq->vq;
405 
406 		virtio_dev_vring_start(vq);
407 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
408 		if (use_simple_rxtx) {
409 			uint16_t mid_idx  = vq->vq_nentries >> 1;
410 
411 			for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
412 				vq->vq_ring.avail->ring[desc_idx] =
413 					desc_idx + mid_idx;
414 				vq->vq_ring.desc[desc_idx + mid_idx].next =
415 					desc_idx;
416 				vq->vq_ring.desc[desc_idx + mid_idx].addr =
417 					txvq->virtio_net_hdr_mem +
418 					offsetof(struct virtio_tx_region, tx_hdr);
419 				vq->vq_ring.desc[desc_idx + mid_idx].len =
420 					vq->hw->vtnet_hdr_size;
421 				vq->vq_ring.desc[desc_idx + mid_idx].flags =
422 					VRING_DESC_F_NEXT;
423 				vq->vq_ring.desc[desc_idx].flags = 0;
424 			}
425 			for (desc_idx = mid_idx; desc_idx < vq->vq_nentries;
426 			     desc_idx++)
427 				vq->vq_ring.avail->ring[desc_idx] = desc_idx;
428 		}
429 #endif
430 		VIRTQUEUE_DUMP(vq);
431 	}
432 }
433 
434 int
435 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
436 			uint16_t queue_idx,
437 			uint16_t nb_desc,
438 			unsigned int socket_id,
439 			__rte_unused const struct rte_eth_rxconf *rx_conf,
440 			struct rte_mempool *mp)
441 {
442 	uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
443 	struct virtnet_rx *rxvq;
444 	int ret;
445 
446 	PMD_INIT_FUNC_TRACE();
447 	ret = virtio_dev_queue_setup(dev, VTNET_RQ, queue_idx, vtpci_queue_idx,
448 			nb_desc, socket_id, (void **)&rxvq);
449 	if (ret < 0) {
450 		PMD_INIT_LOG(ERR, "rvq initialization failed");
451 		return ret;
452 	}
453 
454 	/* Create mempool for rx mbuf allocation */
455 	rxvq->mpool = mp;
456 
457 	dev->data->rx_queues[queue_idx] = rxvq;
458 
459 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
460 	virtio_rxq_vec_setup(rxvq);
461 #endif
462 
463 	return 0;
464 }
465 
466 void
467 virtio_dev_rx_queue_release(void *rxq)
468 {
469 	struct virtnet_rx *rxvq = rxq;
470 	struct virtqueue *vq = rxvq->vq;
471 	/* rxvq is freed when vq is freed, and as mz should be freed after the
472 	 * del_queue, so we reserve the mz pointer first.
473 	 */
474 	const struct rte_memzone *mz = rxvq->mz;
475 
476 	/* no need to free rxq as vq and rxq are allocated together */
477 	virtio_dev_queue_release(vq);
478 	rte_memzone_free(mz);
479 }
480 
481 /*
482  * struct rte_eth_dev *dev: Used to update dev
483  * uint16_t nb_desc: Defaults to values read from config space
484  * unsigned int socket_id: Used to allocate memzone
485  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
486  * uint16_t queue_idx: Just used as an index in dev txq list
487  */
488 int
489 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
490 			uint16_t queue_idx,
491 			uint16_t nb_desc,
492 			unsigned int socket_id,
493 			const struct rte_eth_txconf *tx_conf)
494 {
495 	uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
496 
497 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
498 	struct virtio_hw *hw = dev->data->dev_private;
499 #endif
500 	struct virtnet_tx *txvq;
501 	struct virtqueue *vq;
502 	uint16_t tx_free_thresh;
503 	int ret;
504 
505 	PMD_INIT_FUNC_TRACE();
506 
507 	if ((tx_conf->txq_flags & ETH_TXQ_FLAGS_NOXSUMS)
508 	    != ETH_TXQ_FLAGS_NOXSUMS) {
509 		PMD_INIT_LOG(ERR, "TX checksum offload not supported\n");
510 		return -EINVAL;
511 	}
512 
513 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
514 	/* Use simple rx/tx func if single segment and no offloads */
515 	if ((tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) == VIRTIO_SIMPLE_FLAGS &&
516 	     !vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
517 		PMD_INIT_LOG(INFO, "Using simple rx/tx path");
518 		dev->tx_pkt_burst = virtio_xmit_pkts_simple;
519 		dev->rx_pkt_burst = virtio_recv_pkts_vec;
520 		use_simple_rxtx = 1;
521 	}
522 #endif
523 
524 	ret = virtio_dev_queue_setup(dev, VTNET_TQ, queue_idx, vtpci_queue_idx,
525 			nb_desc, socket_id, (void **)&txvq);
526 	if (ret < 0) {
527 		PMD_INIT_LOG(ERR, "tvq initialization failed");
528 		return ret;
529 	}
530 	vq = txvq->vq;
531 
532 	tx_free_thresh = tx_conf->tx_free_thresh;
533 	if (tx_free_thresh == 0)
534 		tx_free_thresh =
535 			RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
536 
537 	if (tx_free_thresh >= (vq->vq_nentries - 3)) {
538 		RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
539 			"number of TX entries minus 3 (%u)."
540 			" (tx_free_thresh=%u port=%u queue=%u)\n",
541 			vq->vq_nentries - 3,
542 			tx_free_thresh, dev->data->port_id, queue_idx);
543 		return -EINVAL;
544 	}
545 
546 	vq->vq_free_thresh = tx_free_thresh;
547 
548 	dev->data->tx_queues[queue_idx] = txvq;
549 	return 0;
550 }
551 
552 void
553 virtio_dev_tx_queue_release(void *txq)
554 {
555 	struct virtnet_tx *txvq = txq;
556 	struct virtqueue *vq = txvq->vq;
557 	/* txvq is freed when vq is freed, and as mz should be freed after the
558 	 * del_queue, so we reserve the mz pointer first.
559 	 */
560 	const struct rte_memzone *hdr_mz = txvq->virtio_net_hdr_mz;
561 	const struct rte_memzone *mz = txvq->mz;
562 
563 	virtio_dev_queue_release(vq);
564 	rte_memzone_free(mz);
565 	rte_memzone_free(hdr_mz);
566 }
567 
568 static void
569 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
570 {
571 	int error;
572 	/*
573 	 * Requeue the discarded mbuf. This should always be
574 	 * successful since it was just dequeued.
575 	 */
576 	error = virtqueue_enqueue_recv_refill(vq, m);
577 	if (unlikely(error)) {
578 		RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
579 		rte_pktmbuf_free(m);
580 	}
581 }
582 
583 static void
584 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
585 {
586 	uint32_t s = mbuf->pkt_len;
587 	struct ether_addr *ea;
588 
589 	if (s == 64) {
590 		stats->size_bins[1]++;
591 	} else if (s > 64 && s < 1024) {
592 		uint32_t bin;
593 
594 		/* count zeros, and offset into correct bin */
595 		bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
596 		stats->size_bins[bin]++;
597 	} else {
598 		if (s < 64)
599 			stats->size_bins[0]++;
600 		else if (s < 1519)
601 			stats->size_bins[6]++;
602 		else if (s >= 1519)
603 			stats->size_bins[7]++;
604 	}
605 
606 	ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
607 	if (is_multicast_ether_addr(ea)) {
608 		if (is_broadcast_ether_addr(ea))
609 			stats->broadcast++;
610 		else
611 			stats->multicast++;
612 	}
613 }
614 
615 #define VIRTIO_MBUF_BURST_SZ 64
616 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
617 uint16_t
618 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
619 {
620 	struct virtnet_rx *rxvq = rx_queue;
621 	struct virtqueue *vq = rxvq->vq;
622 	struct virtio_hw *hw;
623 	struct rte_mbuf *rxm, *new_mbuf;
624 	uint16_t nb_used, num, nb_rx;
625 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
626 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
627 	int error;
628 	uint32_t i, nb_enqueued;
629 	uint32_t hdr_size;
630 
631 	nb_used = VIRTQUEUE_NUSED(vq);
632 
633 	virtio_rmb();
634 
635 	num = (uint16_t)(likely(nb_used <= nb_pkts) ? nb_used : nb_pkts);
636 	num = (uint16_t)(likely(num <= VIRTIO_MBUF_BURST_SZ) ? num : VIRTIO_MBUF_BURST_SZ);
637 	if (likely(num > DESC_PER_CACHELINE))
638 		num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
639 
640 	num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
641 	PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
642 
643 	hw = vq->hw;
644 	nb_rx = 0;
645 	nb_enqueued = 0;
646 	hdr_size = hw->vtnet_hdr_size;
647 
648 	for (i = 0; i < num ; i++) {
649 		rxm = rcv_pkts[i];
650 
651 		PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
652 
653 		if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
654 			PMD_RX_LOG(ERR, "Packet drop");
655 			nb_enqueued++;
656 			virtio_discard_rxbuf(vq, rxm);
657 			rxvq->stats.errors++;
658 			continue;
659 		}
660 
661 		rxm->port = rxvq->port_id;
662 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
663 		rxm->ol_flags = 0;
664 		rxm->vlan_tci = 0;
665 
666 		rxm->nb_segs = 1;
667 		rxm->next = NULL;
668 		rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
669 		rxm->data_len = (uint16_t)(len[i] - hdr_size);
670 
671 		if (hw->vlan_strip)
672 			rte_vlan_strip(rxm);
673 
674 		VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
675 
676 		rx_pkts[nb_rx++] = rxm;
677 
678 		rxvq->stats.bytes += rx_pkts[nb_rx - 1]->pkt_len;
679 		virtio_update_packet_stats(&rxvq->stats, rxm);
680 	}
681 
682 	rxvq->stats.packets += nb_rx;
683 
684 	/* Allocate new mbuf for the used descriptor */
685 	error = ENOSPC;
686 	while (likely(!virtqueue_full(vq))) {
687 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
688 		if (unlikely(new_mbuf == NULL)) {
689 			struct rte_eth_dev *dev
690 				= &rte_eth_devices[rxvq->port_id];
691 			dev->data->rx_mbuf_alloc_failed++;
692 			break;
693 		}
694 		error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
695 		if (unlikely(error)) {
696 			rte_pktmbuf_free(new_mbuf);
697 			break;
698 		}
699 		nb_enqueued++;
700 	}
701 
702 	if (likely(nb_enqueued)) {
703 		vq_update_avail_idx(vq);
704 
705 		if (unlikely(virtqueue_kick_prepare(vq))) {
706 			virtqueue_notify(vq);
707 			PMD_RX_LOG(DEBUG, "Notified");
708 		}
709 	}
710 
711 	return nb_rx;
712 }
713 
714 uint16_t
715 virtio_recv_mergeable_pkts(void *rx_queue,
716 			struct rte_mbuf **rx_pkts,
717 			uint16_t nb_pkts)
718 {
719 	struct virtnet_rx *rxvq = rx_queue;
720 	struct virtqueue *vq = rxvq->vq;
721 	struct virtio_hw *hw;
722 	struct rte_mbuf *rxm, *new_mbuf;
723 	uint16_t nb_used, num, nb_rx;
724 	uint32_t len[VIRTIO_MBUF_BURST_SZ];
725 	struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
726 	struct rte_mbuf *prev;
727 	int error;
728 	uint32_t i, nb_enqueued;
729 	uint32_t seg_num;
730 	uint16_t extra_idx;
731 	uint32_t seg_res;
732 	uint32_t hdr_size;
733 
734 	nb_used = VIRTQUEUE_NUSED(vq);
735 
736 	virtio_rmb();
737 
738 	PMD_RX_LOG(DEBUG, "used:%d", nb_used);
739 
740 	hw = vq->hw;
741 	nb_rx = 0;
742 	i = 0;
743 	nb_enqueued = 0;
744 	seg_num = 0;
745 	extra_idx = 0;
746 	seg_res = 0;
747 	hdr_size = hw->vtnet_hdr_size;
748 
749 	while (i < nb_used) {
750 		struct virtio_net_hdr_mrg_rxbuf *header;
751 
752 		if (nb_rx == nb_pkts)
753 			break;
754 
755 		num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
756 		if (num != 1)
757 			continue;
758 
759 		i++;
760 
761 		PMD_RX_LOG(DEBUG, "dequeue:%d", num);
762 		PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
763 
764 		rxm = rcv_pkts[0];
765 
766 		if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
767 			PMD_RX_LOG(ERR, "Packet drop");
768 			nb_enqueued++;
769 			virtio_discard_rxbuf(vq, rxm);
770 			rxvq->stats.errors++;
771 			continue;
772 		}
773 
774 		header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
775 			RTE_PKTMBUF_HEADROOM - hdr_size);
776 		seg_num = header->num_buffers;
777 
778 		if (seg_num == 0)
779 			seg_num = 1;
780 
781 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
782 		rxm->nb_segs = seg_num;
783 		rxm->next = NULL;
784 		rxm->ol_flags = 0;
785 		rxm->vlan_tci = 0;
786 		rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
787 		rxm->data_len = (uint16_t)(len[0] - hdr_size);
788 
789 		rxm->port = rxvq->port_id;
790 		rx_pkts[nb_rx] = rxm;
791 		prev = rxm;
792 
793 		seg_res = seg_num - 1;
794 
795 		while (seg_res != 0) {
796 			/*
797 			 * Get extra segments for current uncompleted packet.
798 			 */
799 			uint16_t  rcv_cnt =
800 				RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
801 			if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
802 				uint32_t rx_num =
803 					virtqueue_dequeue_burst_rx(vq,
804 					rcv_pkts, len, rcv_cnt);
805 				i += rx_num;
806 				rcv_cnt = rx_num;
807 			} else {
808 				PMD_RX_LOG(ERR,
809 					   "No enough segments for packet.");
810 				nb_enqueued++;
811 				virtio_discard_rxbuf(vq, rxm);
812 				rxvq->stats.errors++;
813 				break;
814 			}
815 
816 			extra_idx = 0;
817 
818 			while (extra_idx < rcv_cnt) {
819 				rxm = rcv_pkts[extra_idx];
820 
821 				rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
822 				rxm->next = NULL;
823 				rxm->pkt_len = (uint32_t)(len[extra_idx]);
824 				rxm->data_len = (uint16_t)(len[extra_idx]);
825 
826 				if (prev)
827 					prev->next = rxm;
828 
829 				prev = rxm;
830 				rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
831 				extra_idx++;
832 			};
833 			seg_res -= rcv_cnt;
834 		}
835 
836 		if (hw->vlan_strip)
837 			rte_vlan_strip(rx_pkts[nb_rx]);
838 
839 		VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
840 			rx_pkts[nb_rx]->data_len);
841 
842 		rxvq->stats.bytes += rx_pkts[nb_rx]->pkt_len;
843 		virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
844 		nb_rx++;
845 	}
846 
847 	rxvq->stats.packets += nb_rx;
848 
849 	/* Allocate new mbuf for the used descriptor */
850 	error = ENOSPC;
851 	while (likely(!virtqueue_full(vq))) {
852 		new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
853 		if (unlikely(new_mbuf == NULL)) {
854 			struct rte_eth_dev *dev
855 				= &rte_eth_devices[rxvq->port_id];
856 			dev->data->rx_mbuf_alloc_failed++;
857 			break;
858 		}
859 		error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
860 		if (unlikely(error)) {
861 			rte_pktmbuf_free(new_mbuf);
862 			break;
863 		}
864 		nb_enqueued++;
865 	}
866 
867 	if (likely(nb_enqueued)) {
868 		vq_update_avail_idx(vq);
869 
870 		if (unlikely(virtqueue_kick_prepare(vq))) {
871 			virtqueue_notify(vq);
872 			PMD_RX_LOG(DEBUG, "Notified");
873 		}
874 	}
875 
876 	return nb_rx;
877 }
878 
879 uint16_t
880 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
881 {
882 	struct virtnet_tx *txvq = tx_queue;
883 	struct virtqueue *vq = txvq->vq;
884 	struct virtio_hw *hw = vq->hw;
885 	uint16_t hdr_size = hw->vtnet_hdr_size;
886 	uint16_t nb_used, nb_tx;
887 	int error;
888 
889 	if (unlikely(nb_pkts < 1))
890 		return nb_pkts;
891 
892 	PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
893 	nb_used = VIRTQUEUE_NUSED(vq);
894 
895 	virtio_rmb();
896 	if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
897 		virtio_xmit_cleanup(vq, nb_used);
898 
899 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
900 		struct rte_mbuf *txm = tx_pkts[nb_tx];
901 		int can_push = 0, use_indirect = 0, slots, need;
902 
903 		/* Do VLAN tag insertion */
904 		if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
905 			error = rte_vlan_insert(&txm);
906 			if (unlikely(error)) {
907 				rte_pktmbuf_free(txm);
908 				continue;
909 			}
910 		}
911 
912 		/* optimize ring usage */
913 		if (vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) &&
914 		    rte_mbuf_refcnt_read(txm) == 1 &&
915 		    RTE_MBUF_DIRECT(txm) &&
916 		    txm->nb_segs == 1 &&
917 		    rte_pktmbuf_headroom(txm) >= hdr_size &&
918 		    rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
919 				   __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
920 			can_push = 1;
921 		else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
922 			 txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
923 			use_indirect = 1;
924 
925 		/* How many main ring entries are needed to this Tx?
926 		 * any_layout => number of segments
927 		 * indirect   => 1
928 		 * default    => number of segments + 1
929 		 */
930 		slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
931 		need = slots - vq->vq_free_cnt;
932 
933 		/* Positive value indicates it need free vring descriptors */
934 		if (unlikely(need > 0)) {
935 			nb_used = VIRTQUEUE_NUSED(vq);
936 			virtio_rmb();
937 			need = RTE_MIN(need, (int)nb_used);
938 
939 			virtio_xmit_cleanup(vq, need);
940 			need = slots - vq->vq_free_cnt;
941 			if (unlikely(need > 0)) {
942 				PMD_TX_LOG(ERR,
943 					   "No free tx descriptors to transmit");
944 				break;
945 			}
946 		}
947 
948 		/* Enqueue Packet buffers */
949 		virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
950 
951 		txvq->stats.bytes += txm->pkt_len;
952 		virtio_update_packet_stats(&txvq->stats, txm);
953 	}
954 
955 	txvq->stats.packets += nb_tx;
956 
957 	if (likely(nb_tx)) {
958 		vq_update_avail_idx(vq);
959 
960 		if (unlikely(virtqueue_kick_prepare(vq))) {
961 			virtqueue_notify(vq);
962 			PMD_TX_LOG(DEBUG, "Notified backend after xmit");
963 		}
964 	}
965 
966 	return nb_tx;
967 }
968