xref: /dpdk/lib/vhost/virtio_net.c (revision 72206323a5dd3182b13f61b25a64abdddfee595c)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2010-2016 Intel Corporation
3  */
4 
5 #include <stdint.h>
6 #include <stdbool.h>
7 #include <linux/virtio_net.h>
8 
9 #include <rte_mbuf.h>
10 #include <rte_memcpy.h>
11 #include <rte_net.h>
12 #include <rte_ether.h>
13 #include <rte_ip.h>
14 #include <rte_dmadev.h>
15 #include <rte_vhost.h>
16 #include <rte_tcp.h>
17 #include <rte_udp.h>
18 #include <rte_sctp.h>
19 #include <rte_arp.h>
20 #include <rte_spinlock.h>
21 #include <rte_malloc.h>
22 #include <rte_vhost_async.h>
23 
24 #include "iotlb.h"
25 #include "vhost.h"
26 
27 #define MAX_BATCH_LEN 256
28 
29 static __rte_always_inline uint16_t
30 async_poll_dequeue_completed(struct virtio_net *dev, struct vhost_virtqueue *vq,
31 		struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
32 		uint16_t vchan_id, bool legacy_ol_flags);
33 
34 /* DMA device copy operation tracking array. */
35 struct async_dma_info dma_copy_track[RTE_DMADEV_DEFAULT_MAX];
36 
37 static  __rte_always_inline bool
38 rxvq_is_mergeable(struct virtio_net *dev)
39 {
40 	return dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF);
41 }
42 
43 static  __rte_always_inline bool
44 virtio_net_is_inorder(struct virtio_net *dev)
45 {
46 	return dev->features & (1ULL << VIRTIO_F_IN_ORDER);
47 }
48 
49 static bool
50 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
51 {
52 	return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
53 }
54 
55 /*
56  * This function must be called with virtqueue's access_lock taken.
57  */
58 static inline void
59 vhost_queue_stats_update(struct virtio_net *dev, struct vhost_virtqueue *vq,
60 		struct rte_mbuf **pkts, uint16_t count)
61 {
62 	struct virtqueue_stats *stats = &vq->stats;
63 	int i;
64 
65 	if (!(dev->flags & VIRTIO_DEV_STATS_ENABLED))
66 		return;
67 
68 	for (i = 0; i < count; i++) {
69 		struct rte_ether_addr *ea;
70 		struct rte_mbuf *pkt = pkts[i];
71 		uint32_t pkt_len = rte_pktmbuf_pkt_len(pkt);
72 
73 		stats->packets++;
74 		stats->bytes += pkt_len;
75 
76 		if (pkt_len == 64) {
77 			stats->size_bins[1]++;
78 		} else if (pkt_len > 64 && pkt_len < 1024) {
79 			uint32_t bin;
80 
81 			/* count zeros, and offset into correct bin */
82 			bin = (sizeof(pkt_len) * 8) - __builtin_clz(pkt_len) - 5;
83 			stats->size_bins[bin]++;
84 		} else {
85 			if (pkt_len < 64)
86 				stats->size_bins[0]++;
87 			else if (pkt_len < 1519)
88 				stats->size_bins[6]++;
89 			else
90 				stats->size_bins[7]++;
91 		}
92 
93 		ea = rte_pktmbuf_mtod(pkt, struct rte_ether_addr *);
94 		if (rte_is_multicast_ether_addr(ea)) {
95 			if (rte_is_broadcast_ether_addr(ea))
96 				stats->broadcast++;
97 			else
98 				stats->multicast++;
99 		}
100 	}
101 }
102 
103 static __rte_always_inline int64_t
104 vhost_async_dma_transfer_one(struct virtio_net *dev, struct vhost_virtqueue *vq,
105 		int16_t dma_id, uint16_t vchan_id, uint16_t flag_idx,
106 		struct vhost_iov_iter *pkt)
107 {
108 	struct async_dma_vchan_info *dma_info = &dma_copy_track[dma_id].vchans[vchan_id];
109 	uint16_t ring_mask = dma_info->ring_mask;
110 	static bool vhost_async_dma_copy_log;
111 
112 
113 	struct vhost_iovec *iov = pkt->iov;
114 	int copy_idx = 0;
115 	uint32_t nr_segs = pkt->nr_segs;
116 	uint16_t i;
117 
118 	if (rte_dma_burst_capacity(dma_id, vchan_id) < nr_segs)
119 		return -1;
120 
121 	for (i = 0; i < nr_segs; i++) {
122 		copy_idx = rte_dma_copy(dma_id, vchan_id, (rte_iova_t)iov[i].src_addr,
123 				(rte_iova_t)iov[i].dst_addr, iov[i].len, RTE_DMA_OP_FLAG_LLC);
124 		/**
125 		 * Since all memory is pinned and DMA vChannel
126 		 * ring has enough space, failure should be a
127 		 * rare case. If failure happens, it means DMA
128 		 * device encounters serious errors; in this
129 		 * case, please stop async data-path and check
130 		 * what has happened to DMA device.
131 		 */
132 		if (unlikely(copy_idx < 0)) {
133 			if (!vhost_async_dma_copy_log) {
134 				VHOST_LOG_DATA(dev->ifname, ERR,
135 					"DMA copy failed for channel %d:%u\n",
136 					dma_id, vchan_id);
137 				vhost_async_dma_copy_log = true;
138 			}
139 			return -1;
140 		}
141 	}
142 
143 	/**
144 	 * Only store packet completion flag address in the last copy's
145 	 * slot, and other slots are set to NULL.
146 	 */
147 	dma_info->pkts_cmpl_flag_addr[copy_idx & ring_mask] = &vq->async->pkts_cmpl_flag[flag_idx];
148 
149 	return nr_segs;
150 }
151 
152 static __rte_always_inline uint16_t
153 vhost_async_dma_transfer(struct virtio_net *dev, struct vhost_virtqueue *vq,
154 		int16_t dma_id, uint16_t vchan_id, uint16_t head_idx,
155 		struct vhost_iov_iter *pkts, uint16_t nr_pkts)
156 {
157 	struct async_dma_vchan_info *dma_info = &dma_copy_track[dma_id].vchans[vchan_id];
158 	int64_t ret, nr_copies = 0;
159 	uint16_t pkt_idx;
160 
161 	rte_spinlock_lock(&dma_info->dma_lock);
162 
163 	for (pkt_idx = 0; pkt_idx < nr_pkts; pkt_idx++) {
164 		ret = vhost_async_dma_transfer_one(dev, vq, dma_id, vchan_id, head_idx,
165 				&pkts[pkt_idx]);
166 		if (unlikely(ret < 0))
167 			break;
168 
169 		nr_copies += ret;
170 		head_idx++;
171 		if (head_idx >= vq->size)
172 			head_idx -= vq->size;
173 	}
174 
175 	if (likely(nr_copies > 0))
176 		rte_dma_submit(dma_id, vchan_id);
177 
178 	rte_spinlock_unlock(&dma_info->dma_lock);
179 
180 	return pkt_idx;
181 }
182 
183 static __rte_always_inline uint16_t
184 vhost_async_dma_check_completed(struct virtio_net *dev, int16_t dma_id, uint16_t vchan_id,
185 		uint16_t max_pkts)
186 {
187 	struct async_dma_vchan_info *dma_info = &dma_copy_track[dma_id].vchans[vchan_id];
188 	uint16_t ring_mask = dma_info->ring_mask;
189 	uint16_t last_idx = 0;
190 	uint16_t nr_copies;
191 	uint16_t copy_idx;
192 	uint16_t i;
193 	bool has_error = false;
194 	static bool vhost_async_dma_complete_log;
195 
196 	rte_spinlock_lock(&dma_info->dma_lock);
197 
198 	/**
199 	 * Print error log for debugging, if DMA reports error during
200 	 * DMA transfer. We do not handle error in vhost level.
201 	 */
202 	nr_copies = rte_dma_completed(dma_id, vchan_id, max_pkts, &last_idx, &has_error);
203 	if (unlikely(!vhost_async_dma_complete_log && has_error)) {
204 		VHOST_LOG_DATA(dev->ifname, ERR,
205 			"DMA completion failure on channel %d:%u\n",
206 			dma_id, vchan_id);
207 		vhost_async_dma_complete_log = true;
208 	} else if (nr_copies == 0) {
209 		goto out;
210 	}
211 
212 	copy_idx = last_idx - nr_copies + 1;
213 	for (i = 0; i < nr_copies; i++) {
214 		bool *flag;
215 
216 		flag = dma_info->pkts_cmpl_flag_addr[copy_idx & ring_mask];
217 		if (flag) {
218 			/**
219 			 * Mark the packet flag as received. The flag
220 			 * could belong to another virtqueue but write
221 			 * is atomic.
222 			 */
223 			*flag = true;
224 			dma_info->pkts_cmpl_flag_addr[copy_idx & ring_mask] = NULL;
225 		}
226 		copy_idx++;
227 	}
228 
229 out:
230 	rte_spinlock_unlock(&dma_info->dma_lock);
231 	return nr_copies;
232 }
233 
234 static inline void
235 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
236 {
237 	struct batch_copy_elem *elem = vq->batch_copy_elems;
238 	uint16_t count = vq->batch_copy_nb_elems;
239 	int i;
240 
241 	for (i = 0; i < count; i++) {
242 		rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
243 		vhost_log_cache_write_iova(dev, vq, elem[i].log_addr,
244 					   elem[i].len);
245 		PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
246 	}
247 
248 	vq->batch_copy_nb_elems = 0;
249 }
250 
251 static inline void
252 do_data_copy_dequeue(struct vhost_virtqueue *vq)
253 {
254 	struct batch_copy_elem *elem = vq->batch_copy_elems;
255 	uint16_t count = vq->batch_copy_nb_elems;
256 	int i;
257 
258 	for (i = 0; i < count; i++)
259 		rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
260 
261 	vq->batch_copy_nb_elems = 0;
262 }
263 
264 static __rte_always_inline void
265 do_flush_shadow_used_ring_split(struct virtio_net *dev,
266 			struct vhost_virtqueue *vq,
267 			uint16_t to, uint16_t from, uint16_t size)
268 {
269 	rte_memcpy(&vq->used->ring[to],
270 			&vq->shadow_used_split[from],
271 			size * sizeof(struct vring_used_elem));
272 	vhost_log_cache_used_vring(dev, vq,
273 			offsetof(struct vring_used, ring[to]),
274 			size * sizeof(struct vring_used_elem));
275 }
276 
277 static __rte_always_inline void
278 flush_shadow_used_ring_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
279 {
280 	uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
281 
282 	if (used_idx + vq->shadow_used_idx <= vq->size) {
283 		do_flush_shadow_used_ring_split(dev, vq, used_idx, 0,
284 					  vq->shadow_used_idx);
285 	} else {
286 		uint16_t size;
287 
288 		/* update used ring interval [used_idx, vq->size] */
289 		size = vq->size - used_idx;
290 		do_flush_shadow_used_ring_split(dev, vq, used_idx, 0, size);
291 
292 		/* update the left half used ring interval [0, left_size] */
293 		do_flush_shadow_used_ring_split(dev, vq, 0, size,
294 					  vq->shadow_used_idx - size);
295 	}
296 	vq->last_used_idx += vq->shadow_used_idx;
297 
298 	vhost_log_cache_sync(dev, vq);
299 
300 	__atomic_add_fetch(&vq->used->idx, vq->shadow_used_idx,
301 			   __ATOMIC_RELEASE);
302 	vq->shadow_used_idx = 0;
303 	vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
304 		sizeof(vq->used->idx));
305 }
306 
307 static __rte_always_inline void
308 update_shadow_used_ring_split(struct vhost_virtqueue *vq,
309 			 uint16_t desc_idx, uint32_t len)
310 {
311 	uint16_t i = vq->shadow_used_idx++;
312 
313 	vq->shadow_used_split[i].id  = desc_idx;
314 	vq->shadow_used_split[i].len = len;
315 }
316 
317 static __rte_always_inline void
318 vhost_flush_enqueue_shadow_packed(struct virtio_net *dev,
319 				  struct vhost_virtqueue *vq)
320 {
321 	int i;
322 	uint16_t used_idx = vq->last_used_idx;
323 	uint16_t head_idx = vq->last_used_idx;
324 	uint16_t head_flags = 0;
325 
326 	/* Split loop in two to save memory barriers */
327 	for (i = 0; i < vq->shadow_used_idx; i++) {
328 		vq->desc_packed[used_idx].id = vq->shadow_used_packed[i].id;
329 		vq->desc_packed[used_idx].len = vq->shadow_used_packed[i].len;
330 
331 		used_idx += vq->shadow_used_packed[i].count;
332 		if (used_idx >= vq->size)
333 			used_idx -= vq->size;
334 	}
335 
336 	/* The ordering for storing desc flags needs to be enforced. */
337 	rte_atomic_thread_fence(__ATOMIC_RELEASE);
338 
339 	for (i = 0; i < vq->shadow_used_idx; i++) {
340 		uint16_t flags;
341 
342 		if (vq->shadow_used_packed[i].len)
343 			flags = VRING_DESC_F_WRITE;
344 		else
345 			flags = 0;
346 
347 		if (vq->used_wrap_counter) {
348 			flags |= VRING_DESC_F_USED;
349 			flags |= VRING_DESC_F_AVAIL;
350 		} else {
351 			flags &= ~VRING_DESC_F_USED;
352 			flags &= ~VRING_DESC_F_AVAIL;
353 		}
354 
355 		if (i > 0) {
356 			vq->desc_packed[vq->last_used_idx].flags = flags;
357 
358 			vhost_log_cache_used_vring(dev, vq,
359 					vq->last_used_idx *
360 					sizeof(struct vring_packed_desc),
361 					sizeof(struct vring_packed_desc));
362 		} else {
363 			head_idx = vq->last_used_idx;
364 			head_flags = flags;
365 		}
366 
367 		vq_inc_last_used_packed(vq, vq->shadow_used_packed[i].count);
368 	}
369 
370 	vq->desc_packed[head_idx].flags = head_flags;
371 
372 	vhost_log_cache_used_vring(dev, vq,
373 				head_idx *
374 				sizeof(struct vring_packed_desc),
375 				sizeof(struct vring_packed_desc));
376 
377 	vq->shadow_used_idx = 0;
378 	vhost_log_cache_sync(dev, vq);
379 }
380 
381 static __rte_always_inline void
382 vhost_flush_dequeue_shadow_packed(struct virtio_net *dev,
383 				  struct vhost_virtqueue *vq)
384 {
385 	struct vring_used_elem_packed *used_elem = &vq->shadow_used_packed[0];
386 
387 	vq->desc_packed[vq->shadow_last_used_idx].id = used_elem->id;
388 	/* desc flags is the synchronization point for virtio packed vring */
389 	__atomic_store_n(&vq->desc_packed[vq->shadow_last_used_idx].flags,
390 			 used_elem->flags, __ATOMIC_RELEASE);
391 
392 	vhost_log_cache_used_vring(dev, vq, vq->shadow_last_used_idx *
393 				   sizeof(struct vring_packed_desc),
394 				   sizeof(struct vring_packed_desc));
395 	vq->shadow_used_idx = 0;
396 	vhost_log_cache_sync(dev, vq);
397 }
398 
399 static __rte_always_inline void
400 vhost_flush_enqueue_batch_packed(struct virtio_net *dev,
401 				 struct vhost_virtqueue *vq,
402 				 uint64_t *lens,
403 				 uint16_t *ids)
404 {
405 	uint16_t i;
406 	uint16_t flags;
407 	uint16_t last_used_idx;
408 	struct vring_packed_desc *desc_base;
409 
410 	last_used_idx = vq->last_used_idx;
411 	desc_base = &vq->desc_packed[last_used_idx];
412 
413 	flags = PACKED_DESC_ENQUEUE_USED_FLAG(vq->used_wrap_counter);
414 
415 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
416 		desc_base[i].id = ids[i];
417 		desc_base[i].len = lens[i];
418 	}
419 
420 	rte_atomic_thread_fence(__ATOMIC_RELEASE);
421 
422 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
423 		desc_base[i].flags = flags;
424 	}
425 
426 	vhost_log_cache_used_vring(dev, vq, last_used_idx *
427 				   sizeof(struct vring_packed_desc),
428 				   sizeof(struct vring_packed_desc) *
429 				   PACKED_BATCH_SIZE);
430 	vhost_log_cache_sync(dev, vq);
431 
432 	vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
433 }
434 
435 static __rte_always_inline void
436 vhost_shadow_dequeue_batch_packed_inorder(struct vhost_virtqueue *vq,
437 					  uint16_t id)
438 {
439 	vq->shadow_used_packed[0].id = id;
440 
441 	if (!vq->shadow_used_idx) {
442 		vq->shadow_last_used_idx = vq->last_used_idx;
443 		vq->shadow_used_packed[0].flags =
444 			PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
445 		vq->shadow_used_packed[0].len = 0;
446 		vq->shadow_used_packed[0].count = 1;
447 		vq->shadow_used_idx++;
448 	}
449 
450 	vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
451 }
452 
453 static __rte_always_inline void
454 vhost_shadow_dequeue_batch_packed(struct virtio_net *dev,
455 				  struct vhost_virtqueue *vq,
456 				  uint16_t *ids)
457 {
458 	uint16_t flags;
459 	uint16_t i;
460 	uint16_t begin;
461 
462 	flags = PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
463 
464 	if (!vq->shadow_used_idx) {
465 		vq->shadow_last_used_idx = vq->last_used_idx;
466 		vq->shadow_used_packed[0].id  = ids[0];
467 		vq->shadow_used_packed[0].len = 0;
468 		vq->shadow_used_packed[0].count = 1;
469 		vq->shadow_used_packed[0].flags = flags;
470 		vq->shadow_used_idx++;
471 		begin = 1;
472 	} else
473 		begin = 0;
474 
475 	vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE) {
476 		vq->desc_packed[vq->last_used_idx + i].id = ids[i];
477 		vq->desc_packed[vq->last_used_idx + i].len = 0;
478 	}
479 
480 	rte_atomic_thread_fence(__ATOMIC_RELEASE);
481 	vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE)
482 		vq->desc_packed[vq->last_used_idx + i].flags = flags;
483 
484 	vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
485 				   sizeof(struct vring_packed_desc),
486 				   sizeof(struct vring_packed_desc) *
487 				   PACKED_BATCH_SIZE);
488 	vhost_log_cache_sync(dev, vq);
489 
490 	vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
491 }
492 
493 static __rte_always_inline void
494 vhost_shadow_dequeue_single_packed(struct vhost_virtqueue *vq,
495 				   uint16_t buf_id,
496 				   uint16_t count)
497 {
498 	uint16_t flags;
499 
500 	flags = vq->desc_packed[vq->last_used_idx].flags;
501 	if (vq->used_wrap_counter) {
502 		flags |= VRING_DESC_F_USED;
503 		flags |= VRING_DESC_F_AVAIL;
504 	} else {
505 		flags &= ~VRING_DESC_F_USED;
506 		flags &= ~VRING_DESC_F_AVAIL;
507 	}
508 
509 	if (!vq->shadow_used_idx) {
510 		vq->shadow_last_used_idx = vq->last_used_idx;
511 
512 		vq->shadow_used_packed[0].id  = buf_id;
513 		vq->shadow_used_packed[0].len = 0;
514 		vq->shadow_used_packed[0].flags = flags;
515 		vq->shadow_used_idx++;
516 	} else {
517 		vq->desc_packed[vq->last_used_idx].id = buf_id;
518 		vq->desc_packed[vq->last_used_idx].len = 0;
519 		vq->desc_packed[vq->last_used_idx].flags = flags;
520 	}
521 
522 	vq_inc_last_used_packed(vq, count);
523 }
524 
525 static __rte_always_inline void
526 vhost_shadow_dequeue_single_packed_inorder(struct vhost_virtqueue *vq,
527 					   uint16_t buf_id,
528 					   uint16_t count)
529 {
530 	uint16_t flags;
531 
532 	vq->shadow_used_packed[0].id = buf_id;
533 
534 	flags = vq->desc_packed[vq->last_used_idx].flags;
535 	if (vq->used_wrap_counter) {
536 		flags |= VRING_DESC_F_USED;
537 		flags |= VRING_DESC_F_AVAIL;
538 	} else {
539 		flags &= ~VRING_DESC_F_USED;
540 		flags &= ~VRING_DESC_F_AVAIL;
541 	}
542 
543 	if (!vq->shadow_used_idx) {
544 		vq->shadow_last_used_idx = vq->last_used_idx;
545 		vq->shadow_used_packed[0].len = 0;
546 		vq->shadow_used_packed[0].flags = flags;
547 		vq->shadow_used_idx++;
548 	}
549 
550 	vq_inc_last_used_packed(vq, count);
551 }
552 
553 static __rte_always_inline void
554 vhost_shadow_enqueue_packed(struct vhost_virtqueue *vq,
555 				   uint32_t *len,
556 				   uint16_t *id,
557 				   uint16_t *count,
558 				   uint16_t num_buffers)
559 {
560 	uint16_t i;
561 
562 	for (i = 0; i < num_buffers; i++) {
563 		/* enqueue shadow flush action aligned with batch num */
564 		if (!vq->shadow_used_idx)
565 			vq->shadow_aligned_idx = vq->last_used_idx &
566 				PACKED_BATCH_MASK;
567 		vq->shadow_used_packed[vq->shadow_used_idx].id  = id[i];
568 		vq->shadow_used_packed[vq->shadow_used_idx].len = len[i];
569 		vq->shadow_used_packed[vq->shadow_used_idx].count = count[i];
570 		vq->shadow_aligned_idx += count[i];
571 		vq->shadow_used_idx++;
572 	}
573 }
574 
575 static __rte_always_inline void
576 vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
577 				   struct vhost_virtqueue *vq,
578 				   uint32_t *len,
579 				   uint16_t *id,
580 				   uint16_t *count,
581 				   uint16_t num_buffers)
582 {
583 	vhost_shadow_enqueue_packed(vq, len, id, count, num_buffers);
584 
585 	if (vq->shadow_aligned_idx >= PACKED_BATCH_SIZE) {
586 		do_data_copy_enqueue(dev, vq);
587 		vhost_flush_enqueue_shadow_packed(dev, vq);
588 	}
589 }
590 
591 /* avoid write operation when necessary, to lessen cache issues */
592 #define ASSIGN_UNLESS_EQUAL(var, val) do {	\
593 	if ((var) != (val))			\
594 		(var) = (val);			\
595 } while (0)
596 
597 static __rte_always_inline void
598 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
599 {
600 	uint64_t csum_l4 = m_buf->ol_flags & RTE_MBUF_F_TX_L4_MASK;
601 
602 	if (m_buf->ol_flags & RTE_MBUF_F_TX_TCP_SEG)
603 		csum_l4 |= RTE_MBUF_F_TX_TCP_CKSUM;
604 
605 	if (csum_l4) {
606 		/*
607 		 * Pseudo-header checksum must be set as per Virtio spec.
608 		 *
609 		 * Note: We don't propagate rte_net_intel_cksum_prepare()
610 		 * errors, as it would have an impact on performance, and an
611 		 * error would mean the packet is dropped by the guest instead
612 		 * of being dropped here.
613 		 */
614 		rte_net_intel_cksum_prepare(m_buf);
615 
616 		net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
617 		net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
618 
619 		switch (csum_l4) {
620 		case RTE_MBUF_F_TX_TCP_CKSUM:
621 			net_hdr->csum_offset = (offsetof(struct rte_tcp_hdr,
622 						cksum));
623 			break;
624 		case RTE_MBUF_F_TX_UDP_CKSUM:
625 			net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
626 						dgram_cksum));
627 			break;
628 		case RTE_MBUF_F_TX_SCTP_CKSUM:
629 			net_hdr->csum_offset = (offsetof(struct rte_sctp_hdr,
630 						cksum));
631 			break;
632 		}
633 	} else {
634 		ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
635 		ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
636 		ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
637 	}
638 
639 	/* IP cksum verification cannot be bypassed, then calculate here */
640 	if (m_buf->ol_flags & RTE_MBUF_F_TX_IP_CKSUM) {
641 		struct rte_ipv4_hdr *ipv4_hdr;
642 
643 		ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_ipv4_hdr *,
644 						   m_buf->l2_len);
645 		ipv4_hdr->hdr_checksum = 0;
646 		ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
647 	}
648 
649 	if (m_buf->ol_flags & RTE_MBUF_F_TX_TCP_SEG) {
650 		if (m_buf->ol_flags & RTE_MBUF_F_TX_IPV4)
651 			net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
652 		else
653 			net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
654 		net_hdr->gso_size = m_buf->tso_segsz;
655 		net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
656 					+ m_buf->l4_len;
657 	} else if (m_buf->ol_flags & RTE_MBUF_F_TX_UDP_SEG) {
658 		net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
659 		net_hdr->gso_size = m_buf->tso_segsz;
660 		net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
661 			m_buf->l4_len;
662 	} else {
663 		ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
664 		ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
665 		ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
666 	}
667 }
668 
669 static __rte_always_inline int
670 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
671 		struct buf_vector *buf_vec, uint16_t *vec_idx,
672 		uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
673 {
674 	uint16_t vec_id = *vec_idx;
675 
676 	while (desc_len) {
677 		uint64_t desc_addr;
678 		uint64_t desc_chunck_len = desc_len;
679 
680 		if (unlikely(vec_id >= BUF_VECTOR_MAX))
681 			return -1;
682 
683 		desc_addr = vhost_iova_to_vva(dev, vq,
684 				desc_iova,
685 				&desc_chunck_len,
686 				perm);
687 		if (unlikely(!desc_addr))
688 			return -1;
689 
690 		rte_prefetch0((void *)(uintptr_t)desc_addr);
691 
692 		buf_vec[vec_id].buf_iova = desc_iova;
693 		buf_vec[vec_id].buf_addr = desc_addr;
694 		buf_vec[vec_id].buf_len  = desc_chunck_len;
695 
696 		desc_len -= desc_chunck_len;
697 		desc_iova += desc_chunck_len;
698 		vec_id++;
699 	}
700 	*vec_idx = vec_id;
701 
702 	return 0;
703 }
704 
705 static __rte_always_inline int
706 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
707 			 uint32_t avail_idx, uint16_t *vec_idx,
708 			 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
709 			 uint32_t *desc_chain_len, uint8_t perm)
710 {
711 	uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
712 	uint16_t vec_id = *vec_idx;
713 	uint32_t len    = 0;
714 	uint64_t dlen;
715 	uint32_t nr_descs = vq->size;
716 	uint32_t cnt    = 0;
717 	struct vring_desc *descs = vq->desc;
718 	struct vring_desc *idesc = NULL;
719 
720 	if (unlikely(idx >= vq->size))
721 		return -1;
722 
723 	*desc_chain_head = idx;
724 
725 	if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
726 		dlen = vq->desc[idx].len;
727 		nr_descs = dlen / sizeof(struct vring_desc);
728 		if (unlikely(nr_descs > vq->size))
729 			return -1;
730 
731 		descs = (struct vring_desc *)(uintptr_t)
732 			vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
733 						&dlen,
734 						VHOST_ACCESS_RO);
735 		if (unlikely(!descs))
736 			return -1;
737 
738 		if (unlikely(dlen < vq->desc[idx].len)) {
739 			/*
740 			 * The indirect desc table is not contiguous
741 			 * in process VA space, we have to copy it.
742 			 */
743 			idesc = vhost_alloc_copy_ind_table(dev, vq,
744 					vq->desc[idx].addr, vq->desc[idx].len);
745 			if (unlikely(!idesc))
746 				return -1;
747 
748 			descs = idesc;
749 		}
750 
751 		idx = 0;
752 	}
753 
754 	while (1) {
755 		if (unlikely(idx >= nr_descs || cnt++ >= nr_descs)) {
756 			free_ind_table(idesc);
757 			return -1;
758 		}
759 
760 		dlen = descs[idx].len;
761 		len += dlen;
762 
763 		if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
764 						descs[idx].addr, dlen,
765 						perm))) {
766 			free_ind_table(idesc);
767 			return -1;
768 		}
769 
770 		if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
771 			break;
772 
773 		idx = descs[idx].next;
774 	}
775 
776 	*desc_chain_len = len;
777 	*vec_idx = vec_id;
778 
779 	if (unlikely(!!idesc))
780 		free_ind_table(idesc);
781 
782 	return 0;
783 }
784 
785 /*
786  * Returns -1 on fail, 0 on success
787  */
788 static inline int
789 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
790 				uint32_t size, struct buf_vector *buf_vec,
791 				uint16_t *num_buffers, uint16_t avail_head,
792 				uint16_t *nr_vec)
793 {
794 	uint16_t cur_idx;
795 	uint16_t vec_idx = 0;
796 	uint16_t max_tries, tries = 0;
797 
798 	uint16_t head_idx = 0;
799 	uint32_t len = 0;
800 
801 	*num_buffers = 0;
802 	cur_idx  = vq->last_avail_idx;
803 
804 	if (rxvq_is_mergeable(dev))
805 		max_tries = vq->size - 1;
806 	else
807 		max_tries = 1;
808 
809 	while (size > 0) {
810 		if (unlikely(cur_idx == avail_head))
811 			return -1;
812 		/*
813 		 * if we tried all available ring items, and still
814 		 * can't get enough buf, it means something abnormal
815 		 * happened.
816 		 */
817 		if (unlikely(++tries > max_tries))
818 			return -1;
819 
820 		if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
821 						&vec_idx, buf_vec,
822 						&head_idx, &len,
823 						VHOST_ACCESS_RW) < 0))
824 			return -1;
825 		len = RTE_MIN(len, size);
826 		update_shadow_used_ring_split(vq, head_idx, len);
827 		size -= len;
828 
829 		cur_idx++;
830 		*num_buffers += 1;
831 	}
832 
833 	*nr_vec = vec_idx;
834 
835 	return 0;
836 }
837 
838 static __rte_always_inline int
839 fill_vec_buf_packed_indirect(struct virtio_net *dev,
840 			struct vhost_virtqueue *vq,
841 			struct vring_packed_desc *desc, uint16_t *vec_idx,
842 			struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
843 {
844 	uint16_t i;
845 	uint32_t nr_descs;
846 	uint16_t vec_id = *vec_idx;
847 	uint64_t dlen;
848 	struct vring_packed_desc *descs, *idescs = NULL;
849 
850 	dlen = desc->len;
851 	descs = (struct vring_packed_desc *)(uintptr_t)
852 		vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
853 	if (unlikely(!descs))
854 		return -1;
855 
856 	if (unlikely(dlen < desc->len)) {
857 		/*
858 		 * The indirect desc table is not contiguous
859 		 * in process VA space, we have to copy it.
860 		 */
861 		idescs = vhost_alloc_copy_ind_table(dev,
862 				vq, desc->addr, desc->len);
863 		if (unlikely(!idescs))
864 			return -1;
865 
866 		descs = idescs;
867 	}
868 
869 	nr_descs =  desc->len / sizeof(struct vring_packed_desc);
870 	if (unlikely(nr_descs >= vq->size)) {
871 		free_ind_table(idescs);
872 		return -1;
873 	}
874 
875 	for (i = 0; i < nr_descs; i++) {
876 		if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
877 			free_ind_table(idescs);
878 			return -1;
879 		}
880 
881 		dlen = descs[i].len;
882 		*len += dlen;
883 		if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
884 						descs[i].addr, dlen,
885 						perm)))
886 			return -1;
887 	}
888 	*vec_idx = vec_id;
889 
890 	if (unlikely(!!idescs))
891 		free_ind_table(idescs);
892 
893 	return 0;
894 }
895 
896 static __rte_always_inline int
897 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
898 				uint16_t avail_idx, uint16_t *desc_count,
899 				struct buf_vector *buf_vec, uint16_t *vec_idx,
900 				uint16_t *buf_id, uint32_t *len, uint8_t perm)
901 {
902 	bool wrap_counter = vq->avail_wrap_counter;
903 	struct vring_packed_desc *descs = vq->desc_packed;
904 	uint16_t vec_id = *vec_idx;
905 	uint64_t dlen;
906 
907 	if (avail_idx < vq->last_avail_idx)
908 		wrap_counter ^= 1;
909 
910 	/*
911 	 * Perform a load-acquire barrier in desc_is_avail to
912 	 * enforce the ordering between desc flags and desc
913 	 * content.
914 	 */
915 	if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
916 		return -1;
917 
918 	*desc_count = 0;
919 	*len = 0;
920 
921 	while (1) {
922 		if (unlikely(vec_id >= BUF_VECTOR_MAX))
923 			return -1;
924 
925 		if (unlikely(*desc_count >= vq->size))
926 			return -1;
927 
928 		*desc_count += 1;
929 		*buf_id = descs[avail_idx].id;
930 
931 		if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
932 			if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
933 							&descs[avail_idx],
934 							&vec_id, buf_vec,
935 							len, perm) < 0))
936 				return -1;
937 		} else {
938 			dlen = descs[avail_idx].len;
939 			*len += dlen;
940 
941 			if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
942 							descs[avail_idx].addr,
943 							dlen,
944 							perm)))
945 				return -1;
946 		}
947 
948 		if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
949 			break;
950 
951 		if (++avail_idx >= vq->size) {
952 			avail_idx -= vq->size;
953 			wrap_counter ^= 1;
954 		}
955 	}
956 
957 	*vec_idx = vec_id;
958 
959 	return 0;
960 }
961 
962 static __rte_noinline void
963 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
964 		struct buf_vector *buf_vec,
965 		struct virtio_net_hdr_mrg_rxbuf *hdr)
966 {
967 	uint64_t len;
968 	uint64_t remain = dev->vhost_hlen;
969 	uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
970 	uint64_t iova = buf_vec->buf_iova;
971 
972 	while (remain) {
973 		len = RTE_MIN(remain,
974 				buf_vec->buf_len);
975 		dst = buf_vec->buf_addr;
976 		rte_memcpy((void *)(uintptr_t)dst,
977 				(void *)(uintptr_t)src,
978 				len);
979 
980 		PRINT_PACKET(dev, (uintptr_t)dst,
981 				(uint32_t)len, 0);
982 		vhost_log_cache_write_iova(dev, vq,
983 				iova, len);
984 
985 		remain -= len;
986 		iova += len;
987 		src += len;
988 		buf_vec++;
989 	}
990 }
991 
992 static __rte_always_inline int
993 async_iter_initialize(struct virtio_net *dev, struct vhost_async *async)
994 {
995 	struct vhost_iov_iter *iter;
996 
997 	if (unlikely(async->iovec_idx >= VHOST_MAX_ASYNC_VEC)) {
998 		VHOST_LOG_DATA(dev->ifname, ERR, "no more async iovec available\n");
999 		return -1;
1000 	}
1001 
1002 	iter = async->iov_iter + async->iter_idx;
1003 	iter->iov = async->iovec + async->iovec_idx;
1004 	iter->nr_segs = 0;
1005 
1006 	return 0;
1007 }
1008 
1009 static __rte_always_inline int
1010 async_iter_add_iovec(struct virtio_net *dev, struct vhost_async *async,
1011 		void *src, void *dst, size_t len)
1012 {
1013 	struct vhost_iov_iter *iter;
1014 	struct vhost_iovec *iovec;
1015 
1016 	if (unlikely(async->iovec_idx >= VHOST_MAX_ASYNC_VEC)) {
1017 		static bool vhost_max_async_vec_log;
1018 
1019 		if (!vhost_max_async_vec_log) {
1020 			VHOST_LOG_DATA(dev->ifname, ERR, "no more async iovec available\n");
1021 			vhost_max_async_vec_log = true;
1022 		}
1023 
1024 		return -1;
1025 	}
1026 
1027 	iter = async->iov_iter + async->iter_idx;
1028 	iovec = async->iovec + async->iovec_idx;
1029 
1030 	iovec->src_addr = src;
1031 	iovec->dst_addr = dst;
1032 	iovec->len = len;
1033 
1034 	iter->nr_segs++;
1035 	async->iovec_idx++;
1036 
1037 	return 0;
1038 }
1039 
1040 static __rte_always_inline void
1041 async_iter_finalize(struct vhost_async *async)
1042 {
1043 	async->iter_idx++;
1044 }
1045 
1046 static __rte_always_inline void
1047 async_iter_cancel(struct vhost_async *async)
1048 {
1049 	struct vhost_iov_iter *iter;
1050 
1051 	iter = async->iov_iter + async->iter_idx;
1052 	async->iovec_idx -= iter->nr_segs;
1053 	iter->nr_segs = 0;
1054 	iter->iov = NULL;
1055 }
1056 
1057 static __rte_always_inline void
1058 async_iter_reset(struct vhost_async *async)
1059 {
1060 	async->iter_idx = 0;
1061 	async->iovec_idx = 0;
1062 }
1063 
1064 static __rte_always_inline int
1065 async_fill_seg(struct virtio_net *dev, struct vhost_virtqueue *vq,
1066 		struct rte_mbuf *m, uint32_t mbuf_offset,
1067 		uint64_t buf_iova, uint32_t cpy_len, bool to_desc)
1068 {
1069 	struct vhost_async *async = vq->async;
1070 	uint64_t mapped_len;
1071 	uint32_t buf_offset = 0;
1072 	void *src, *dst;
1073 	void *host_iova;
1074 
1075 	while (cpy_len) {
1076 		host_iova = (void *)(uintptr_t)gpa_to_first_hpa(dev,
1077 				buf_iova + buf_offset, cpy_len, &mapped_len);
1078 		if (unlikely(!host_iova)) {
1079 			VHOST_LOG_DATA(dev->ifname, ERR,
1080 				"%s: failed to get host iova.\n",
1081 				__func__);
1082 			return -1;
1083 		}
1084 
1085 		if (to_desc) {
1086 			src = (void *)(uintptr_t)rte_pktmbuf_iova_offset(m, mbuf_offset);
1087 			dst = host_iova;
1088 		} else {
1089 			src = host_iova;
1090 			dst = (void *)(uintptr_t)rte_pktmbuf_iova_offset(m, mbuf_offset);
1091 		}
1092 
1093 		if (unlikely(async_iter_add_iovec(dev, async, src, dst, (size_t)mapped_len)))
1094 			return -1;
1095 
1096 		cpy_len -= (uint32_t)mapped_len;
1097 		mbuf_offset += (uint32_t)mapped_len;
1098 		buf_offset += (uint32_t)mapped_len;
1099 	}
1100 
1101 	return 0;
1102 }
1103 
1104 static __rte_always_inline void
1105 sync_fill_seg(struct virtio_net *dev, struct vhost_virtqueue *vq,
1106 		struct rte_mbuf *m, uint32_t mbuf_offset,
1107 		uint64_t buf_addr, uint64_t buf_iova, uint32_t cpy_len, bool to_desc)
1108 {
1109 	struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
1110 
1111 	if (likely(cpy_len > MAX_BATCH_LEN || vq->batch_copy_nb_elems >= vq->size)) {
1112 		if (to_desc) {
1113 			rte_memcpy((void *)((uintptr_t)(buf_addr)),
1114 				rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
1115 				cpy_len);
1116 			vhost_log_cache_write_iova(dev, vq, buf_iova, cpy_len);
1117 			PRINT_PACKET(dev, (uintptr_t)(buf_addr), cpy_len, 0);
1118 		} else {
1119 			rte_memcpy(rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
1120 				(void *)((uintptr_t)(buf_addr)),
1121 				cpy_len);
1122 		}
1123 	} else {
1124 		if (to_desc) {
1125 			batch_copy[vq->batch_copy_nb_elems].dst =
1126 				(void *)((uintptr_t)(buf_addr));
1127 			batch_copy[vq->batch_copy_nb_elems].src =
1128 				rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
1129 			batch_copy[vq->batch_copy_nb_elems].log_addr = buf_iova;
1130 		} else {
1131 			batch_copy[vq->batch_copy_nb_elems].dst =
1132 				rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
1133 			batch_copy[vq->batch_copy_nb_elems].src =
1134 				(void *)((uintptr_t)(buf_addr));
1135 		}
1136 		batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
1137 		vq->batch_copy_nb_elems++;
1138 	}
1139 }
1140 
1141 static __rte_always_inline int
1142 mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
1143 		struct rte_mbuf *m, struct buf_vector *buf_vec,
1144 		uint16_t nr_vec, uint16_t num_buffers, bool is_async)
1145 {
1146 	uint32_t vec_idx = 0;
1147 	uint32_t mbuf_offset, mbuf_avail;
1148 	uint32_t buf_offset, buf_avail;
1149 	uint64_t buf_addr, buf_iova, buf_len;
1150 	uint32_t cpy_len;
1151 	uint64_t hdr_addr;
1152 	struct rte_mbuf *hdr_mbuf;
1153 	struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
1154 	struct vhost_async *async = vq->async;
1155 
1156 	if (unlikely(m == NULL))
1157 		return -1;
1158 
1159 	buf_addr = buf_vec[vec_idx].buf_addr;
1160 	buf_iova = buf_vec[vec_idx].buf_iova;
1161 	buf_len = buf_vec[vec_idx].buf_len;
1162 
1163 	if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1))
1164 		return -1;
1165 
1166 	hdr_mbuf = m;
1167 	hdr_addr = buf_addr;
1168 	if (unlikely(buf_len < dev->vhost_hlen)) {
1169 		memset(&tmp_hdr, 0, sizeof(struct virtio_net_hdr_mrg_rxbuf));
1170 		hdr = &tmp_hdr;
1171 	} else
1172 		hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
1173 
1174 	VHOST_LOG_DATA(dev->ifname, DEBUG, "RX: num merge buffers %d\n", num_buffers);
1175 
1176 	if (unlikely(buf_len < dev->vhost_hlen)) {
1177 		buf_offset = dev->vhost_hlen - buf_len;
1178 		vec_idx++;
1179 		buf_addr = buf_vec[vec_idx].buf_addr;
1180 		buf_iova = buf_vec[vec_idx].buf_iova;
1181 		buf_len = buf_vec[vec_idx].buf_len;
1182 		buf_avail = buf_len - buf_offset;
1183 	} else {
1184 		buf_offset = dev->vhost_hlen;
1185 		buf_avail = buf_len - dev->vhost_hlen;
1186 	}
1187 
1188 	mbuf_avail  = rte_pktmbuf_data_len(m);
1189 	mbuf_offset = 0;
1190 
1191 	if (is_async) {
1192 		if (async_iter_initialize(dev, async))
1193 			return -1;
1194 	}
1195 
1196 	while (mbuf_avail != 0 || m->next != NULL) {
1197 		/* done with current buf, get the next one */
1198 		if (buf_avail == 0) {
1199 			vec_idx++;
1200 			if (unlikely(vec_idx >= nr_vec))
1201 				goto error;
1202 
1203 			buf_addr = buf_vec[vec_idx].buf_addr;
1204 			buf_iova = buf_vec[vec_idx].buf_iova;
1205 			buf_len = buf_vec[vec_idx].buf_len;
1206 
1207 			buf_offset = 0;
1208 			buf_avail  = buf_len;
1209 		}
1210 
1211 		/* done with current mbuf, get the next one */
1212 		if (mbuf_avail == 0) {
1213 			m = m->next;
1214 
1215 			mbuf_offset = 0;
1216 			mbuf_avail  = rte_pktmbuf_data_len(m);
1217 		}
1218 
1219 		if (hdr_addr) {
1220 			virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
1221 			if (rxvq_is_mergeable(dev))
1222 				ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
1223 						num_buffers);
1224 
1225 			if (unlikely(hdr == &tmp_hdr)) {
1226 				copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
1227 			} else {
1228 				PRINT_PACKET(dev, (uintptr_t)hdr_addr,
1229 						dev->vhost_hlen, 0);
1230 				vhost_log_cache_write_iova(dev, vq,
1231 						buf_vec[0].buf_iova,
1232 						dev->vhost_hlen);
1233 			}
1234 
1235 			hdr_addr = 0;
1236 		}
1237 
1238 		cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1239 
1240 		if (is_async) {
1241 			if (async_fill_seg(dev, vq, m, mbuf_offset,
1242 					   buf_iova + buf_offset, cpy_len, true) < 0)
1243 				goto error;
1244 		} else {
1245 			sync_fill_seg(dev, vq, m, mbuf_offset,
1246 				      buf_addr + buf_offset,
1247 				      buf_iova + buf_offset, cpy_len, true);
1248 		}
1249 
1250 		mbuf_avail  -= cpy_len;
1251 		mbuf_offset += cpy_len;
1252 		buf_avail  -= cpy_len;
1253 		buf_offset += cpy_len;
1254 	}
1255 
1256 	if (is_async)
1257 		async_iter_finalize(async);
1258 
1259 	return 0;
1260 error:
1261 	if (is_async)
1262 		async_iter_cancel(async);
1263 
1264 	return -1;
1265 }
1266 
1267 static __rte_always_inline int
1268 vhost_enqueue_single_packed(struct virtio_net *dev,
1269 			    struct vhost_virtqueue *vq,
1270 			    struct rte_mbuf *pkt,
1271 			    struct buf_vector *buf_vec,
1272 			    uint16_t *nr_descs)
1273 {
1274 	uint16_t nr_vec = 0;
1275 	uint16_t avail_idx = vq->last_avail_idx;
1276 	uint16_t max_tries, tries = 0;
1277 	uint16_t buf_id = 0;
1278 	uint32_t len = 0;
1279 	uint16_t desc_count;
1280 	uint32_t size = pkt->pkt_len + sizeof(struct virtio_net_hdr_mrg_rxbuf);
1281 	uint16_t num_buffers = 0;
1282 	uint32_t buffer_len[vq->size];
1283 	uint16_t buffer_buf_id[vq->size];
1284 	uint16_t buffer_desc_count[vq->size];
1285 
1286 	if (rxvq_is_mergeable(dev))
1287 		max_tries = vq->size - 1;
1288 	else
1289 		max_tries = 1;
1290 
1291 	while (size > 0) {
1292 		/*
1293 		 * if we tried all available ring items, and still
1294 		 * can't get enough buf, it means something abnormal
1295 		 * happened.
1296 		 */
1297 		if (unlikely(++tries > max_tries))
1298 			return -1;
1299 
1300 		if (unlikely(fill_vec_buf_packed(dev, vq,
1301 						avail_idx, &desc_count,
1302 						buf_vec, &nr_vec,
1303 						&buf_id, &len,
1304 						VHOST_ACCESS_RW) < 0))
1305 			return -1;
1306 
1307 		len = RTE_MIN(len, size);
1308 		size -= len;
1309 
1310 		buffer_len[num_buffers] = len;
1311 		buffer_buf_id[num_buffers] = buf_id;
1312 		buffer_desc_count[num_buffers] = desc_count;
1313 		num_buffers += 1;
1314 
1315 		*nr_descs += desc_count;
1316 		avail_idx += desc_count;
1317 		if (avail_idx >= vq->size)
1318 			avail_idx -= vq->size;
1319 	}
1320 
1321 	if (mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, num_buffers, false) < 0)
1322 		return -1;
1323 
1324 	vhost_shadow_enqueue_single_packed(dev, vq, buffer_len, buffer_buf_id,
1325 					   buffer_desc_count, num_buffers);
1326 
1327 	return 0;
1328 }
1329 
1330 static __rte_noinline uint32_t
1331 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1332 	struct rte_mbuf **pkts, uint32_t count)
1333 {
1334 	uint32_t pkt_idx = 0;
1335 	uint16_t num_buffers;
1336 	struct buf_vector buf_vec[BUF_VECTOR_MAX];
1337 	uint16_t avail_head;
1338 
1339 	/*
1340 	 * The ordering between avail index and
1341 	 * desc reads needs to be enforced.
1342 	 */
1343 	avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1344 
1345 	rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1346 
1347 	for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1348 		uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1349 		uint16_t nr_vec = 0;
1350 
1351 		if (unlikely(reserve_avail_buf_split(dev, vq,
1352 						pkt_len, buf_vec, &num_buffers,
1353 						avail_head, &nr_vec) < 0)) {
1354 			VHOST_LOG_DATA(dev->ifname, DEBUG,
1355 				"failed to get enough desc from vring\n");
1356 			vq->shadow_used_idx -= num_buffers;
1357 			break;
1358 		}
1359 
1360 		VHOST_LOG_DATA(dev->ifname, DEBUG,
1361 			"current index %d | end index %d\n",
1362 			vq->last_avail_idx, vq->last_avail_idx + num_buffers);
1363 
1364 		if (mbuf_to_desc(dev, vq, pkts[pkt_idx], buf_vec, nr_vec,
1365 					num_buffers, false) < 0) {
1366 			vq->shadow_used_idx -= num_buffers;
1367 			break;
1368 		}
1369 
1370 		vq->last_avail_idx += num_buffers;
1371 	}
1372 
1373 	do_data_copy_enqueue(dev, vq);
1374 
1375 	if (likely(vq->shadow_used_idx)) {
1376 		flush_shadow_used_ring_split(dev, vq);
1377 		vhost_vring_call_split(dev, vq);
1378 	}
1379 
1380 	return pkt_idx;
1381 }
1382 
1383 static __rte_always_inline int
1384 virtio_dev_rx_sync_batch_check(struct virtio_net *dev,
1385 			   struct vhost_virtqueue *vq,
1386 			   struct rte_mbuf **pkts,
1387 			   uint64_t *desc_addrs,
1388 			   uint64_t *lens)
1389 {
1390 	bool wrap_counter = vq->avail_wrap_counter;
1391 	struct vring_packed_desc *descs = vq->desc_packed;
1392 	uint16_t avail_idx = vq->last_avail_idx;
1393 	uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1394 	uint16_t i;
1395 
1396 	if (unlikely(avail_idx & PACKED_BATCH_MASK))
1397 		return -1;
1398 
1399 	if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1400 		return -1;
1401 
1402 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1403 		if (unlikely(pkts[i]->next != NULL))
1404 			return -1;
1405 		if (unlikely(!desc_is_avail(&descs[avail_idx + i],
1406 					    wrap_counter)))
1407 			return -1;
1408 	}
1409 
1410 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1411 		lens[i] = descs[avail_idx + i].len;
1412 
1413 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1414 		if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
1415 			return -1;
1416 	}
1417 
1418 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1419 		desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1420 						  descs[avail_idx + i].addr,
1421 						  &lens[i],
1422 						  VHOST_ACCESS_RW);
1423 
1424 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1425 		if (unlikely(!desc_addrs[i]))
1426 			return -1;
1427 		if (unlikely(lens[i] != descs[avail_idx + i].len))
1428 			return -1;
1429 	}
1430 
1431 	return 0;
1432 }
1433 
1434 static __rte_always_inline void
1435 virtio_dev_rx_batch_packed_copy(struct virtio_net *dev,
1436 			   struct vhost_virtqueue *vq,
1437 			   struct rte_mbuf **pkts,
1438 			   uint64_t *desc_addrs,
1439 			   uint64_t *lens)
1440 {
1441 	uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1442 	struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
1443 	struct vring_packed_desc *descs = vq->desc_packed;
1444 	uint16_t avail_idx = vq->last_avail_idx;
1445 	uint16_t ids[PACKED_BATCH_SIZE];
1446 	uint16_t i;
1447 
1448 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1449 		rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1450 		hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)
1451 					(uintptr_t)desc_addrs[i];
1452 		lens[i] = pkts[i]->pkt_len +
1453 			sizeof(struct virtio_net_hdr_mrg_rxbuf);
1454 	}
1455 
1456 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1457 		virtio_enqueue_offload(pkts[i], &hdrs[i]->hdr);
1458 
1459 	vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1460 
1461 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1462 		rte_memcpy((void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1463 			   rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1464 			   pkts[i]->pkt_len);
1465 	}
1466 
1467 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1468 		vhost_log_cache_write_iova(dev, vq, descs[avail_idx + i].addr,
1469 					   lens[i]);
1470 
1471 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1472 		ids[i] = descs[avail_idx + i].id;
1473 
1474 	vhost_flush_enqueue_batch_packed(dev, vq, lens, ids);
1475 }
1476 
1477 static __rte_always_inline int
1478 virtio_dev_rx_sync_batch_packed(struct virtio_net *dev,
1479 			   struct vhost_virtqueue *vq,
1480 			   struct rte_mbuf **pkts)
1481 {
1482 	uint64_t desc_addrs[PACKED_BATCH_SIZE];
1483 	uint64_t lens[PACKED_BATCH_SIZE];
1484 
1485 	if (virtio_dev_rx_sync_batch_check(dev, vq, pkts, desc_addrs, lens) == -1)
1486 		return -1;
1487 
1488 	if (vq->shadow_used_idx) {
1489 		do_data_copy_enqueue(dev, vq);
1490 		vhost_flush_enqueue_shadow_packed(dev, vq);
1491 	}
1492 
1493 	virtio_dev_rx_batch_packed_copy(dev, vq, pkts, desc_addrs, lens);
1494 
1495 	return 0;
1496 }
1497 
1498 static __rte_always_inline int16_t
1499 virtio_dev_rx_single_packed(struct virtio_net *dev,
1500 			    struct vhost_virtqueue *vq,
1501 			    struct rte_mbuf *pkt)
1502 {
1503 	struct buf_vector buf_vec[BUF_VECTOR_MAX];
1504 	uint16_t nr_descs = 0;
1505 
1506 	if (unlikely(vhost_enqueue_single_packed(dev, vq, pkt, buf_vec,
1507 						 &nr_descs) < 0)) {
1508 		VHOST_LOG_DATA(dev->ifname, DEBUG, "failed to get enough desc from vring\n");
1509 		return -1;
1510 	}
1511 
1512 	VHOST_LOG_DATA(dev->ifname, DEBUG,
1513 		"current index %d | end index %d\n",
1514 		vq->last_avail_idx, vq->last_avail_idx + nr_descs);
1515 
1516 	vq_inc_last_avail_packed(vq, nr_descs);
1517 
1518 	return 0;
1519 }
1520 
1521 static __rte_noinline uint32_t
1522 virtio_dev_rx_packed(struct virtio_net *dev,
1523 		     struct vhost_virtqueue *__rte_restrict vq,
1524 		     struct rte_mbuf **__rte_restrict pkts,
1525 		     uint32_t count)
1526 {
1527 	uint32_t pkt_idx = 0;
1528 
1529 	do {
1530 		rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1531 
1532 		if (count - pkt_idx >= PACKED_BATCH_SIZE) {
1533 			if (!virtio_dev_rx_sync_batch_packed(dev, vq,
1534 							&pkts[pkt_idx])) {
1535 				pkt_idx += PACKED_BATCH_SIZE;
1536 				continue;
1537 			}
1538 		}
1539 
1540 		if (virtio_dev_rx_single_packed(dev, vq, pkts[pkt_idx]))
1541 			break;
1542 		pkt_idx++;
1543 
1544 	} while (pkt_idx < count);
1545 
1546 	if (vq->shadow_used_idx) {
1547 		do_data_copy_enqueue(dev, vq);
1548 		vhost_flush_enqueue_shadow_packed(dev, vq);
1549 	}
1550 
1551 	if (pkt_idx)
1552 		vhost_vring_call_packed(dev, vq);
1553 
1554 	return pkt_idx;
1555 }
1556 
1557 static __rte_always_inline uint32_t
1558 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
1559 	struct rte_mbuf **pkts, uint32_t count)
1560 {
1561 	struct vhost_virtqueue *vq;
1562 	uint32_t nb_tx = 0;
1563 
1564 	VHOST_LOG_DATA(dev->ifname, DEBUG, "%s\n", __func__);
1565 	if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1566 		VHOST_LOG_DATA(dev->ifname, ERR,
1567 			"%s: invalid virtqueue idx %d.\n",
1568 			__func__, queue_id);
1569 		return 0;
1570 	}
1571 
1572 	vq = dev->virtqueue[queue_id];
1573 
1574 	rte_spinlock_lock(&vq->access_lock);
1575 
1576 	if (unlikely(!vq->enabled))
1577 		goto out_access_unlock;
1578 
1579 	if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1580 		vhost_user_iotlb_rd_lock(vq);
1581 
1582 	if (unlikely(!vq->access_ok))
1583 		if (unlikely(vring_translate(dev, vq) < 0))
1584 			goto out;
1585 
1586 	count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1587 	if (count == 0)
1588 		goto out;
1589 
1590 	if (vq_is_packed(dev))
1591 		nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
1592 	else
1593 		nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
1594 
1595 	vhost_queue_stats_update(dev, vq, pkts, nb_tx);
1596 
1597 out:
1598 	if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1599 		vhost_user_iotlb_rd_unlock(vq);
1600 
1601 out_access_unlock:
1602 	rte_spinlock_unlock(&vq->access_lock);
1603 
1604 	return nb_tx;
1605 }
1606 
1607 uint16_t
1608 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
1609 	struct rte_mbuf **__rte_restrict pkts, uint16_t count)
1610 {
1611 	struct virtio_net *dev = get_device(vid);
1612 
1613 	if (!dev)
1614 		return 0;
1615 
1616 	if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1617 		VHOST_LOG_DATA(dev->ifname, ERR,
1618 			"%s: built-in vhost net backend is disabled.\n",
1619 			__func__);
1620 		return 0;
1621 	}
1622 
1623 	return virtio_dev_rx(dev, queue_id, pkts, count);
1624 }
1625 
1626 static __rte_always_inline uint16_t
1627 async_get_first_inflight_pkt_idx(struct vhost_virtqueue *vq)
1628 {
1629 	struct vhost_async *async = vq->async;
1630 
1631 	if (async->pkts_idx >= async->pkts_inflight_n)
1632 		return async->pkts_idx - async->pkts_inflight_n;
1633 	else
1634 		return vq->size - async->pkts_inflight_n + async->pkts_idx;
1635 }
1636 
1637 static __rte_always_inline void
1638 store_dma_desc_info_split(struct vring_used_elem *s_ring, struct vring_used_elem *d_ring,
1639 		uint16_t ring_size, uint16_t s_idx, uint16_t d_idx, uint16_t count)
1640 {
1641 	size_t elem_size = sizeof(struct vring_used_elem);
1642 
1643 	if (d_idx + count <= ring_size) {
1644 		rte_memcpy(d_ring + d_idx, s_ring + s_idx, count * elem_size);
1645 	} else {
1646 		uint16_t size = ring_size - d_idx;
1647 
1648 		rte_memcpy(d_ring + d_idx, s_ring + s_idx, size * elem_size);
1649 		rte_memcpy(d_ring, s_ring + s_idx + size, (count - size) * elem_size);
1650 	}
1651 }
1652 
1653 static __rte_always_inline void
1654 store_dma_desc_info_packed(struct vring_used_elem_packed *s_ring,
1655 		struct vring_used_elem_packed *d_ring,
1656 		uint16_t ring_size, uint16_t s_idx, uint16_t d_idx, uint16_t count)
1657 {
1658 	size_t elem_size = sizeof(struct vring_used_elem_packed);
1659 
1660 	if (d_idx + count <= ring_size) {
1661 		rte_memcpy(d_ring + d_idx, s_ring + s_idx, count * elem_size);
1662 	} else {
1663 		uint16_t size = ring_size - d_idx;
1664 
1665 		rte_memcpy(d_ring + d_idx, s_ring + s_idx, size * elem_size);
1666 		rte_memcpy(d_ring, s_ring + s_idx + size, (count - size) * elem_size);
1667 	}
1668 }
1669 
1670 static __rte_noinline uint32_t
1671 virtio_dev_rx_async_submit_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1672 		uint16_t queue_id, struct rte_mbuf **pkts, uint32_t count,
1673 		int16_t dma_id, uint16_t vchan_id)
1674 {
1675 	struct buf_vector buf_vec[BUF_VECTOR_MAX];
1676 	uint32_t pkt_idx = 0;
1677 	uint16_t num_buffers;
1678 	uint16_t avail_head;
1679 
1680 	struct vhost_async *async = vq->async;
1681 	struct async_inflight_info *pkts_info = async->pkts_info;
1682 	uint32_t pkt_err = 0;
1683 	uint16_t n_xfer;
1684 	uint16_t slot_idx = 0;
1685 
1686 	/*
1687 	 * The ordering between avail index and desc reads need to be enforced.
1688 	 */
1689 	avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1690 
1691 	rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1692 
1693 	async_iter_reset(async);
1694 
1695 	for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1696 		uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1697 		uint16_t nr_vec = 0;
1698 
1699 		if (unlikely(reserve_avail_buf_split(dev, vq, pkt_len, buf_vec,
1700 						&num_buffers, avail_head, &nr_vec) < 0)) {
1701 			VHOST_LOG_DATA(dev->ifname, DEBUG,
1702 				"failed to get enough desc from vring\n");
1703 			vq->shadow_used_idx -= num_buffers;
1704 			break;
1705 		}
1706 
1707 		VHOST_LOG_DATA(dev->ifname, DEBUG,
1708 			"current index %d | end index %d\n",
1709 			vq->last_avail_idx, vq->last_avail_idx + num_buffers);
1710 
1711 		if (mbuf_to_desc(dev, vq, pkts[pkt_idx], buf_vec, nr_vec, num_buffers, true) < 0) {
1712 			vq->shadow_used_idx -= num_buffers;
1713 			break;
1714 		}
1715 
1716 		slot_idx = (async->pkts_idx + pkt_idx) & (vq->size - 1);
1717 		pkts_info[slot_idx].descs = num_buffers;
1718 		pkts_info[slot_idx].mbuf = pkts[pkt_idx];
1719 
1720 		vq->last_avail_idx += num_buffers;
1721 	}
1722 
1723 	if (unlikely(pkt_idx == 0))
1724 		return 0;
1725 
1726 	n_xfer = vhost_async_dma_transfer(dev, vq, dma_id, vchan_id, async->pkts_idx,
1727 			async->iov_iter, pkt_idx);
1728 
1729 	pkt_err = pkt_idx - n_xfer;
1730 	if (unlikely(pkt_err)) {
1731 		uint16_t num_descs = 0;
1732 
1733 		VHOST_LOG_DATA(dev->ifname, DEBUG,
1734 			"%s: failed to transfer %u packets for queue %u.\n",
1735 			__func__, pkt_err, queue_id);
1736 
1737 		/* update number of completed packets */
1738 		pkt_idx = n_xfer;
1739 
1740 		/* calculate the sum of descriptors to revert */
1741 		while (pkt_err-- > 0) {
1742 			num_descs += pkts_info[slot_idx & (vq->size - 1)].descs;
1743 			slot_idx--;
1744 		}
1745 
1746 		/* recover shadow used ring and available ring */
1747 		vq->shadow_used_idx -= num_descs;
1748 		vq->last_avail_idx -= num_descs;
1749 	}
1750 
1751 	/* keep used descriptors */
1752 	if (likely(vq->shadow_used_idx)) {
1753 		uint16_t to = async->desc_idx_split & (vq->size - 1);
1754 
1755 		store_dma_desc_info_split(vq->shadow_used_split,
1756 				async->descs_split, vq->size, 0, to,
1757 				vq->shadow_used_idx);
1758 
1759 		async->desc_idx_split += vq->shadow_used_idx;
1760 
1761 		async->pkts_idx += pkt_idx;
1762 		if (async->pkts_idx >= vq->size)
1763 			async->pkts_idx -= vq->size;
1764 
1765 		async->pkts_inflight_n += pkt_idx;
1766 		vq->shadow_used_idx = 0;
1767 	}
1768 
1769 	return pkt_idx;
1770 }
1771 
1772 
1773 static __rte_always_inline int
1774 vhost_enqueue_async_packed(struct virtio_net *dev,
1775 			    struct vhost_virtqueue *vq,
1776 			    struct rte_mbuf *pkt,
1777 			    struct buf_vector *buf_vec,
1778 			    uint16_t *nr_descs,
1779 			    uint16_t *nr_buffers)
1780 {
1781 	uint16_t nr_vec = 0;
1782 	uint16_t avail_idx = vq->last_avail_idx;
1783 	uint16_t max_tries, tries = 0;
1784 	uint16_t buf_id = 0;
1785 	uint32_t len = 0;
1786 	uint16_t desc_count = 0;
1787 	uint32_t size = pkt->pkt_len + sizeof(struct virtio_net_hdr_mrg_rxbuf);
1788 	uint32_t buffer_len[vq->size];
1789 	uint16_t buffer_buf_id[vq->size];
1790 	uint16_t buffer_desc_count[vq->size];
1791 
1792 	if (rxvq_is_mergeable(dev))
1793 		max_tries = vq->size - 1;
1794 	else
1795 		max_tries = 1;
1796 
1797 	while (size > 0) {
1798 		/*
1799 		 * if we tried all available ring items, and still
1800 		 * can't get enough buf, it means something abnormal
1801 		 * happened.
1802 		 */
1803 		if (unlikely(++tries > max_tries))
1804 			return -1;
1805 
1806 		if (unlikely(fill_vec_buf_packed(dev, vq,
1807 						avail_idx, &desc_count,
1808 						buf_vec, &nr_vec,
1809 						&buf_id, &len,
1810 						VHOST_ACCESS_RW) < 0))
1811 			return -1;
1812 
1813 		len = RTE_MIN(len, size);
1814 		size -= len;
1815 
1816 		buffer_len[*nr_buffers] = len;
1817 		buffer_buf_id[*nr_buffers] = buf_id;
1818 		buffer_desc_count[*nr_buffers] = desc_count;
1819 		*nr_buffers += 1;
1820 		*nr_descs += desc_count;
1821 		avail_idx += desc_count;
1822 		if (avail_idx >= vq->size)
1823 			avail_idx -= vq->size;
1824 	}
1825 
1826 	if (unlikely(mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, *nr_buffers, true) < 0))
1827 		return -1;
1828 
1829 	vhost_shadow_enqueue_packed(vq, buffer_len, buffer_buf_id, buffer_desc_count, *nr_buffers);
1830 
1831 	return 0;
1832 }
1833 
1834 static __rte_always_inline int16_t
1835 virtio_dev_rx_async_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
1836 			    struct rte_mbuf *pkt, uint16_t *nr_descs, uint16_t *nr_buffers)
1837 {
1838 	struct buf_vector buf_vec[BUF_VECTOR_MAX];
1839 
1840 	if (unlikely(vhost_enqueue_async_packed(dev, vq, pkt, buf_vec,
1841 					nr_descs, nr_buffers) < 0)) {
1842 		VHOST_LOG_DATA(dev->ifname, DEBUG, "failed to get enough desc from vring\n");
1843 		return -1;
1844 	}
1845 
1846 	VHOST_LOG_DATA(dev->ifname, DEBUG,
1847 		"current index %d | end index %d\n",
1848 		vq->last_avail_idx, vq->last_avail_idx + *nr_descs);
1849 
1850 	return 0;
1851 }
1852 
1853 static __rte_always_inline void
1854 dma_error_handler_packed(struct vhost_virtqueue *vq, uint16_t slot_idx,
1855 			uint32_t nr_err, uint32_t *pkt_idx)
1856 {
1857 	uint16_t descs_err = 0;
1858 	uint16_t buffers_err = 0;
1859 	struct async_inflight_info *pkts_info = vq->async->pkts_info;
1860 
1861 	*pkt_idx -= nr_err;
1862 	/* calculate the sum of buffers and descs of DMA-error packets. */
1863 	while (nr_err-- > 0) {
1864 		descs_err += pkts_info[slot_idx % vq->size].descs;
1865 		buffers_err += pkts_info[slot_idx % vq->size].nr_buffers;
1866 		slot_idx--;
1867 	}
1868 
1869 	if (vq->last_avail_idx >= descs_err) {
1870 		vq->last_avail_idx -= descs_err;
1871 	} else {
1872 		vq->last_avail_idx = vq->last_avail_idx + vq->size - descs_err;
1873 		vq->avail_wrap_counter ^= 1;
1874 	}
1875 
1876 	vq->shadow_used_idx -= buffers_err;
1877 }
1878 
1879 static __rte_noinline uint32_t
1880 virtio_dev_rx_async_submit_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
1881 		uint16_t queue_id, struct rte_mbuf **pkts, uint32_t count,
1882 		int16_t dma_id, uint16_t vchan_id)
1883 {
1884 	uint32_t pkt_idx = 0;
1885 	uint32_t remained = count;
1886 	uint16_t n_xfer;
1887 	uint16_t num_buffers;
1888 	uint16_t num_descs;
1889 
1890 	struct vhost_async *async = vq->async;
1891 	struct async_inflight_info *pkts_info = async->pkts_info;
1892 	uint32_t pkt_err = 0;
1893 	uint16_t slot_idx = 0;
1894 
1895 	do {
1896 		rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1897 
1898 		num_buffers = 0;
1899 		num_descs = 0;
1900 		if (unlikely(virtio_dev_rx_async_packed(dev, vq, pkts[pkt_idx],
1901 						&num_descs, &num_buffers) < 0))
1902 			break;
1903 
1904 		slot_idx = (async->pkts_idx + pkt_idx) % vq->size;
1905 
1906 		pkts_info[slot_idx].descs = num_descs;
1907 		pkts_info[slot_idx].nr_buffers = num_buffers;
1908 		pkts_info[slot_idx].mbuf = pkts[pkt_idx];
1909 
1910 		pkt_idx++;
1911 		remained--;
1912 		vq_inc_last_avail_packed(vq, num_descs);
1913 	} while (pkt_idx < count);
1914 
1915 	if (unlikely(pkt_idx == 0))
1916 		return 0;
1917 
1918 	n_xfer = vhost_async_dma_transfer(dev, vq, dma_id, vchan_id, async->pkts_idx,
1919 			async->iov_iter, pkt_idx);
1920 
1921 	async_iter_reset(async);
1922 
1923 	pkt_err = pkt_idx - n_xfer;
1924 	if (unlikely(pkt_err)) {
1925 		VHOST_LOG_DATA(dev->ifname, DEBUG,
1926 			"%s: failed to transfer %u packets for queue %u.\n",
1927 			__func__, pkt_err, queue_id);
1928 		dma_error_handler_packed(vq, slot_idx, pkt_err, &pkt_idx);
1929 	}
1930 
1931 	if (likely(vq->shadow_used_idx)) {
1932 		/* keep used descriptors. */
1933 		store_dma_desc_info_packed(vq->shadow_used_packed, async->buffers_packed,
1934 					vq->size, 0, async->buffer_idx_packed,
1935 					vq->shadow_used_idx);
1936 
1937 		async->buffer_idx_packed += vq->shadow_used_idx;
1938 		if (async->buffer_idx_packed >= vq->size)
1939 			async->buffer_idx_packed -= vq->size;
1940 
1941 		async->pkts_idx += pkt_idx;
1942 		if (async->pkts_idx >= vq->size)
1943 			async->pkts_idx -= vq->size;
1944 
1945 		vq->shadow_used_idx = 0;
1946 		async->pkts_inflight_n += pkt_idx;
1947 	}
1948 
1949 	return pkt_idx;
1950 }
1951 
1952 static __rte_always_inline void
1953 write_back_completed_descs_split(struct vhost_virtqueue *vq, uint16_t n_descs)
1954 {
1955 	struct vhost_async *async = vq->async;
1956 	uint16_t nr_left = n_descs;
1957 	uint16_t nr_copy;
1958 	uint16_t to, from;
1959 
1960 	do {
1961 		from = async->last_desc_idx_split & (vq->size - 1);
1962 		nr_copy = nr_left + from <= vq->size ? nr_left : vq->size - from;
1963 		to = vq->last_used_idx & (vq->size - 1);
1964 
1965 		if (to + nr_copy <= vq->size) {
1966 			rte_memcpy(&vq->used->ring[to], &async->descs_split[from],
1967 					nr_copy * sizeof(struct vring_used_elem));
1968 		} else {
1969 			uint16_t size = vq->size - to;
1970 
1971 			rte_memcpy(&vq->used->ring[to], &async->descs_split[from],
1972 					size * sizeof(struct vring_used_elem));
1973 			rte_memcpy(&vq->used->ring[0], &async->descs_split[from + size],
1974 					(nr_copy - size) * sizeof(struct vring_used_elem));
1975 		}
1976 
1977 		async->last_desc_idx_split += nr_copy;
1978 		vq->last_used_idx += nr_copy;
1979 		nr_left -= nr_copy;
1980 	} while (nr_left > 0);
1981 }
1982 
1983 static __rte_always_inline void
1984 write_back_completed_descs_packed(struct vhost_virtqueue *vq,
1985 				uint16_t n_buffers)
1986 {
1987 	struct vhost_async *async = vq->async;
1988 	uint16_t from = async->last_buffer_idx_packed;
1989 	uint16_t used_idx = vq->last_used_idx;
1990 	uint16_t head_idx = vq->last_used_idx;
1991 	uint16_t head_flags = 0;
1992 	uint16_t i;
1993 
1994 	/* Split loop in two to save memory barriers */
1995 	for (i = 0; i < n_buffers; i++) {
1996 		vq->desc_packed[used_idx].id = async->buffers_packed[from].id;
1997 		vq->desc_packed[used_idx].len = async->buffers_packed[from].len;
1998 
1999 		used_idx += async->buffers_packed[from].count;
2000 		if (used_idx >= vq->size)
2001 			used_idx -= vq->size;
2002 
2003 		from++;
2004 		if (from >= vq->size)
2005 			from = 0;
2006 	}
2007 
2008 	/* The ordering for storing desc flags needs to be enforced. */
2009 	rte_atomic_thread_fence(__ATOMIC_RELEASE);
2010 
2011 	from = async->last_buffer_idx_packed;
2012 
2013 	for (i = 0; i < n_buffers; i++) {
2014 		uint16_t flags;
2015 
2016 		if (async->buffers_packed[from].len)
2017 			flags = VRING_DESC_F_WRITE;
2018 		else
2019 			flags = 0;
2020 
2021 		if (vq->used_wrap_counter) {
2022 			flags |= VRING_DESC_F_USED;
2023 			flags |= VRING_DESC_F_AVAIL;
2024 		} else {
2025 			flags &= ~VRING_DESC_F_USED;
2026 			flags &= ~VRING_DESC_F_AVAIL;
2027 		}
2028 
2029 		if (i > 0) {
2030 			vq->desc_packed[vq->last_used_idx].flags = flags;
2031 		} else {
2032 			head_idx = vq->last_used_idx;
2033 			head_flags = flags;
2034 		}
2035 
2036 		vq_inc_last_used_packed(vq, async->buffers_packed[from].count);
2037 
2038 		from++;
2039 		if (from == vq->size)
2040 			from = 0;
2041 	}
2042 
2043 	vq->desc_packed[head_idx].flags = head_flags;
2044 	async->last_buffer_idx_packed = from;
2045 }
2046 
2047 static __rte_always_inline uint16_t
2048 vhost_poll_enqueue_completed(struct virtio_net *dev, uint16_t queue_id,
2049 		struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
2050 		uint16_t vchan_id)
2051 {
2052 	struct vhost_virtqueue *vq = dev->virtqueue[queue_id];
2053 	struct vhost_async *async = vq->async;
2054 	struct async_inflight_info *pkts_info = async->pkts_info;
2055 	uint16_t nr_cpl_pkts = 0;
2056 	uint16_t n_descs = 0, n_buffers = 0;
2057 	uint16_t start_idx, from, i;
2058 
2059 	/* Check completed copies for the given DMA vChannel */
2060 	vhost_async_dma_check_completed(dev, dma_id, vchan_id, VHOST_DMA_MAX_COPY_COMPLETE);
2061 
2062 	start_idx = async_get_first_inflight_pkt_idx(vq);
2063 	/**
2064 	 * Calculate the number of copy completed packets.
2065 	 * Note that there may be completed packets even if
2066 	 * no copies are reported done by the given DMA vChannel,
2067 	 * as it's possible that a virtqueue uses multiple DMA
2068 	 * vChannels.
2069 	 */
2070 	from = start_idx;
2071 	while (vq->async->pkts_cmpl_flag[from] && count--) {
2072 		vq->async->pkts_cmpl_flag[from] = false;
2073 		from++;
2074 		if (from >= vq->size)
2075 			from -= vq->size;
2076 		nr_cpl_pkts++;
2077 	}
2078 
2079 	if (nr_cpl_pkts == 0)
2080 		return 0;
2081 
2082 	for (i = 0; i < nr_cpl_pkts; i++) {
2083 		from = (start_idx + i) % vq->size;
2084 		/* Only used with packed ring */
2085 		n_buffers += pkts_info[from].nr_buffers;
2086 		/* Only used with split ring */
2087 		n_descs += pkts_info[from].descs;
2088 		pkts[i] = pkts_info[from].mbuf;
2089 	}
2090 
2091 	async->pkts_inflight_n -= nr_cpl_pkts;
2092 
2093 	if (likely(vq->enabled && vq->access_ok)) {
2094 		if (vq_is_packed(dev)) {
2095 			write_back_completed_descs_packed(vq, n_buffers);
2096 			vhost_vring_call_packed(dev, vq);
2097 		} else {
2098 			write_back_completed_descs_split(vq, n_descs);
2099 			__atomic_add_fetch(&vq->used->idx, n_descs, __ATOMIC_RELEASE);
2100 			vhost_vring_call_split(dev, vq);
2101 		}
2102 	} else {
2103 		if (vq_is_packed(dev)) {
2104 			async->last_buffer_idx_packed += n_buffers;
2105 			if (async->last_buffer_idx_packed >= vq->size)
2106 				async->last_buffer_idx_packed -= vq->size;
2107 		} else {
2108 			async->last_desc_idx_split += n_descs;
2109 		}
2110 	}
2111 
2112 	return nr_cpl_pkts;
2113 }
2114 
2115 uint16_t
2116 rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
2117 		struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
2118 		uint16_t vchan_id)
2119 {
2120 	struct virtio_net *dev = get_device(vid);
2121 	struct vhost_virtqueue *vq;
2122 	uint16_t n_pkts_cpl = 0;
2123 
2124 	if (unlikely(!dev))
2125 		return 0;
2126 
2127 	VHOST_LOG_DATA(dev->ifname, DEBUG, "%s\n", __func__);
2128 	if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
2129 		VHOST_LOG_DATA(dev->ifname, ERR,
2130 			"%s: invalid virtqueue idx %d.\n",
2131 			__func__, queue_id);
2132 		return 0;
2133 	}
2134 
2135 	if (unlikely(!dma_copy_track[dma_id].vchans ||
2136 				!dma_copy_track[dma_id].vchans[vchan_id].pkts_cmpl_flag_addr)) {
2137 		VHOST_LOG_DATA(dev->ifname, ERR,
2138 			"%s: invalid channel %d:%u.\n",
2139 			__func__, dma_id, vchan_id);
2140 		return 0;
2141 	}
2142 
2143 	vq = dev->virtqueue[queue_id];
2144 
2145 	if (!rte_spinlock_trylock(&vq->access_lock)) {
2146 		VHOST_LOG_DATA(dev->ifname, DEBUG,
2147 			"%s: virtqueue %u is busy.\n",
2148 			__func__, queue_id);
2149 		return 0;
2150 	}
2151 
2152 	if (unlikely(!vq->async)) {
2153 		VHOST_LOG_DATA(dev->ifname, ERR,
2154 			"%s: async not registered for virtqueue %d.\n",
2155 			__func__, queue_id);
2156 		goto out;
2157 	}
2158 
2159 	n_pkts_cpl = vhost_poll_enqueue_completed(dev, queue_id, pkts, count, dma_id, vchan_id);
2160 
2161 	vhost_queue_stats_update(dev, vq, pkts, n_pkts_cpl);
2162 	vq->stats.inflight_completed += n_pkts_cpl;
2163 
2164 out:
2165 	rte_spinlock_unlock(&vq->access_lock);
2166 
2167 	return n_pkts_cpl;
2168 }
2169 
2170 uint16_t
2171 rte_vhost_clear_queue_thread_unsafe(int vid, uint16_t queue_id,
2172 		struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
2173 		uint16_t vchan_id)
2174 {
2175 	struct virtio_net *dev = get_device(vid);
2176 	struct vhost_virtqueue *vq;
2177 	uint16_t n_pkts_cpl = 0;
2178 
2179 	if (!dev)
2180 		return 0;
2181 
2182 	VHOST_LOG_DATA(dev->ifname, DEBUG, "%s\n", __func__);
2183 	if (unlikely(queue_id >= dev->nr_vring)) {
2184 		VHOST_LOG_DATA(dev->ifname, ERR, "%s: invalid virtqueue idx %d.\n",
2185 			__func__, queue_id);
2186 		return 0;
2187 	}
2188 
2189 	if (unlikely(dma_id < 0 || dma_id >= RTE_DMADEV_DEFAULT_MAX)) {
2190 		VHOST_LOG_DATA(dev->ifname, ERR, "%s: invalid dma id %d.\n",
2191 			__func__, dma_id);
2192 		return 0;
2193 	}
2194 
2195 	vq = dev->virtqueue[queue_id];
2196 
2197 	if (unlikely(!rte_spinlock_is_locked(&vq->access_lock))) {
2198 		VHOST_LOG_DATA(dev->ifname, ERR, "%s() called without access lock taken.\n",
2199 			__func__);
2200 		return -1;
2201 	}
2202 
2203 	if (unlikely(!vq->async)) {
2204 		VHOST_LOG_DATA(dev->ifname, ERR,
2205 			"%s: async not registered for virtqueue %d.\n",
2206 			__func__, queue_id);
2207 		return 0;
2208 	}
2209 
2210 	if (unlikely(!dma_copy_track[dma_id].vchans ||
2211 				!dma_copy_track[dma_id].vchans[vchan_id].pkts_cmpl_flag_addr)) {
2212 		VHOST_LOG_DATA(dev->ifname, ERR,
2213 			"%s: invalid channel %d:%u.\n",
2214 			__func__, dma_id, vchan_id);
2215 		return 0;
2216 	}
2217 
2218 	if ((queue_id & 1) == 0)
2219 		n_pkts_cpl = vhost_poll_enqueue_completed(dev, queue_id,
2220 					pkts, count, dma_id, vchan_id);
2221 	else {
2222 		n_pkts_cpl = async_poll_dequeue_completed(dev, vq, pkts, count,
2223 					dma_id, vchan_id, dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS);
2224 	}
2225 
2226 	vhost_queue_stats_update(dev, vq, pkts, n_pkts_cpl);
2227 	vq->stats.inflight_completed += n_pkts_cpl;
2228 
2229 	return n_pkts_cpl;
2230 }
2231 
2232 uint16_t
2233 rte_vhost_clear_queue(int vid, uint16_t queue_id, struct rte_mbuf **pkts,
2234 		uint16_t count, int16_t dma_id, uint16_t vchan_id)
2235 {
2236 	struct virtio_net *dev = get_device(vid);
2237 	struct vhost_virtqueue *vq;
2238 	uint16_t n_pkts_cpl = 0;
2239 
2240 	if (!dev)
2241 		return 0;
2242 
2243 	VHOST_LOG_DATA(dev->ifname, DEBUG, "%s\n", __func__);
2244 	if (unlikely(queue_id >= dev->nr_vring)) {
2245 		VHOST_LOG_DATA(dev->ifname, ERR, "%s: invalid virtqueue idx %u.\n",
2246 			__func__, queue_id);
2247 		return 0;
2248 	}
2249 
2250 	if (unlikely(dma_id < 0 || dma_id >= RTE_DMADEV_DEFAULT_MAX)) {
2251 		VHOST_LOG_DATA(dev->ifname, ERR, "%s: invalid dma id %d.\n",
2252 			__func__, dma_id);
2253 		return 0;
2254 	}
2255 
2256 	vq = dev->virtqueue[queue_id];
2257 
2258 	if (!rte_spinlock_trylock(&vq->access_lock)) {
2259 		VHOST_LOG_DATA(dev->ifname, DEBUG, "%s: virtqueue %u is busy.\n",
2260 			__func__, queue_id);
2261 		return 0;
2262 	}
2263 
2264 	if (unlikely(!vq->async)) {
2265 		VHOST_LOG_DATA(dev->ifname, ERR, "%s: async not registered for queue id %u.\n",
2266 			__func__, queue_id);
2267 		goto out_access_unlock;
2268 	}
2269 
2270 	if (unlikely(!dma_copy_track[dma_id].vchans ||
2271 				!dma_copy_track[dma_id].vchans[vchan_id].pkts_cmpl_flag_addr)) {
2272 		VHOST_LOG_DATA(dev->ifname, ERR, "%s: invalid channel %d:%u.\n",
2273 			__func__, dma_id, vchan_id);
2274 		goto out_access_unlock;
2275 	}
2276 
2277 	if ((queue_id & 1) == 0)
2278 		n_pkts_cpl = vhost_poll_enqueue_completed(dev, queue_id,
2279 				pkts, count, dma_id, vchan_id);
2280 	else {
2281 		n_pkts_cpl = async_poll_dequeue_completed(dev, vq, pkts, count,
2282 					dma_id, vchan_id, dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS);
2283 	}
2284 
2285 	vhost_queue_stats_update(dev, vq, pkts, n_pkts_cpl);
2286 	vq->stats.inflight_completed += n_pkts_cpl;
2287 
2288 out_access_unlock:
2289 	rte_spinlock_unlock(&vq->access_lock);
2290 
2291 	return n_pkts_cpl;
2292 }
2293 
2294 static __rte_always_inline uint32_t
2295 virtio_dev_rx_async_submit(struct virtio_net *dev, uint16_t queue_id,
2296 	struct rte_mbuf **pkts, uint32_t count, int16_t dma_id, uint16_t vchan_id)
2297 {
2298 	struct vhost_virtqueue *vq;
2299 	uint32_t nb_tx = 0;
2300 
2301 	VHOST_LOG_DATA(dev->ifname, DEBUG, "%s\n", __func__);
2302 	if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
2303 		VHOST_LOG_DATA(dev->ifname, ERR,
2304 			"%s: invalid virtqueue idx %d.\n",
2305 			__func__, queue_id);
2306 		return 0;
2307 	}
2308 
2309 	if (unlikely(!dma_copy_track[dma_id].vchans ||
2310 				!dma_copy_track[dma_id].vchans[vchan_id].pkts_cmpl_flag_addr)) {
2311 		VHOST_LOG_DATA(dev->ifname, ERR,
2312 			"%s: invalid channel %d:%u.\n",
2313 			 __func__, dma_id, vchan_id);
2314 		return 0;
2315 	}
2316 
2317 	vq = dev->virtqueue[queue_id];
2318 
2319 	rte_spinlock_lock(&vq->access_lock);
2320 
2321 	if (unlikely(!vq->enabled || !vq->async))
2322 		goto out_access_unlock;
2323 
2324 	if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2325 		vhost_user_iotlb_rd_lock(vq);
2326 
2327 	if (unlikely(!vq->access_ok))
2328 		if (unlikely(vring_translate(dev, vq) < 0))
2329 			goto out;
2330 
2331 	count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
2332 	if (count == 0)
2333 		goto out;
2334 
2335 	if (vq_is_packed(dev))
2336 		nb_tx = virtio_dev_rx_async_submit_packed(dev, vq, queue_id,
2337 				pkts, count, dma_id, vchan_id);
2338 	else
2339 		nb_tx = virtio_dev_rx_async_submit_split(dev, vq, queue_id,
2340 				pkts, count, dma_id, vchan_id);
2341 
2342 	vq->stats.inflight_submitted += nb_tx;
2343 
2344 out:
2345 	if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2346 		vhost_user_iotlb_rd_unlock(vq);
2347 
2348 out_access_unlock:
2349 	rte_spinlock_unlock(&vq->access_lock);
2350 
2351 	return nb_tx;
2352 }
2353 
2354 uint16_t
2355 rte_vhost_submit_enqueue_burst(int vid, uint16_t queue_id,
2356 		struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
2357 		uint16_t vchan_id)
2358 {
2359 	struct virtio_net *dev = get_device(vid);
2360 
2361 	if (!dev)
2362 		return 0;
2363 
2364 	if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
2365 		VHOST_LOG_DATA(dev->ifname, ERR,
2366 			"%s: built-in vhost net backend is disabled.\n",
2367 			__func__);
2368 		return 0;
2369 	}
2370 
2371 	return virtio_dev_rx_async_submit(dev, queue_id, pkts, count, dma_id, vchan_id);
2372 }
2373 
2374 static inline bool
2375 virtio_net_with_host_offload(struct virtio_net *dev)
2376 {
2377 	if (dev->features &
2378 			((1ULL << VIRTIO_NET_F_CSUM) |
2379 			 (1ULL << VIRTIO_NET_F_HOST_ECN) |
2380 			 (1ULL << VIRTIO_NET_F_HOST_TSO4) |
2381 			 (1ULL << VIRTIO_NET_F_HOST_TSO6) |
2382 			 (1ULL << VIRTIO_NET_F_HOST_UFO)))
2383 		return true;
2384 
2385 	return false;
2386 }
2387 
2388 static int
2389 parse_headers(struct rte_mbuf *m, uint8_t *l4_proto)
2390 {
2391 	struct rte_ipv4_hdr *ipv4_hdr;
2392 	struct rte_ipv6_hdr *ipv6_hdr;
2393 	struct rte_ether_hdr *eth_hdr;
2394 	uint16_t ethertype;
2395 	uint16_t data_len = rte_pktmbuf_data_len(m);
2396 
2397 	if (data_len < sizeof(struct rte_ether_hdr))
2398 		return -EINVAL;
2399 
2400 	eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
2401 
2402 	m->l2_len = sizeof(struct rte_ether_hdr);
2403 	ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
2404 
2405 	if (ethertype == RTE_ETHER_TYPE_VLAN) {
2406 		if (data_len < sizeof(struct rte_ether_hdr) +
2407 				sizeof(struct rte_vlan_hdr))
2408 			goto error;
2409 
2410 		struct rte_vlan_hdr *vlan_hdr =
2411 			(struct rte_vlan_hdr *)(eth_hdr + 1);
2412 
2413 		m->l2_len += sizeof(struct rte_vlan_hdr);
2414 		ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
2415 	}
2416 
2417 	switch (ethertype) {
2418 	case RTE_ETHER_TYPE_IPV4:
2419 		if (data_len < m->l2_len + sizeof(struct rte_ipv4_hdr))
2420 			goto error;
2421 		ipv4_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv4_hdr *,
2422 				m->l2_len);
2423 		m->l3_len = rte_ipv4_hdr_len(ipv4_hdr);
2424 		if (data_len < m->l2_len + m->l3_len)
2425 			goto error;
2426 		m->ol_flags |= RTE_MBUF_F_TX_IPV4;
2427 		*l4_proto = ipv4_hdr->next_proto_id;
2428 		break;
2429 	case RTE_ETHER_TYPE_IPV6:
2430 		if (data_len < m->l2_len + sizeof(struct rte_ipv6_hdr))
2431 			goto error;
2432 		ipv6_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv6_hdr *,
2433 				m->l2_len);
2434 		m->l3_len = sizeof(struct rte_ipv6_hdr);
2435 		m->ol_flags |= RTE_MBUF_F_TX_IPV6;
2436 		*l4_proto = ipv6_hdr->proto;
2437 		break;
2438 	default:
2439 		/* a valid L3 header is needed for further L4 parsing */
2440 		goto error;
2441 	}
2442 
2443 	/* both CSUM and GSO need a valid L4 header */
2444 	switch (*l4_proto) {
2445 	case IPPROTO_TCP:
2446 		if (data_len < m->l2_len + m->l3_len +
2447 				sizeof(struct rte_tcp_hdr))
2448 			goto error;
2449 		break;
2450 	case IPPROTO_UDP:
2451 		if (data_len < m->l2_len + m->l3_len +
2452 				sizeof(struct rte_udp_hdr))
2453 			goto error;
2454 		break;
2455 	case IPPROTO_SCTP:
2456 		if (data_len < m->l2_len + m->l3_len +
2457 				sizeof(struct rte_sctp_hdr))
2458 			goto error;
2459 		break;
2460 	default:
2461 		goto error;
2462 	}
2463 
2464 	return 0;
2465 
2466 error:
2467 	m->l2_len = 0;
2468 	m->l3_len = 0;
2469 	m->ol_flags = 0;
2470 	return -EINVAL;
2471 }
2472 
2473 static __rte_always_inline void
2474 vhost_dequeue_offload_legacy(struct virtio_net *dev, struct virtio_net_hdr *hdr,
2475 		struct rte_mbuf *m)
2476 {
2477 	uint8_t l4_proto = 0;
2478 	struct rte_tcp_hdr *tcp_hdr = NULL;
2479 	uint16_t tcp_len;
2480 	uint16_t data_len = rte_pktmbuf_data_len(m);
2481 
2482 	if (parse_headers(m, &l4_proto) < 0)
2483 		return;
2484 
2485 	if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
2486 		if (hdr->csum_start == (m->l2_len + m->l3_len)) {
2487 			switch (hdr->csum_offset) {
2488 			case (offsetof(struct rte_tcp_hdr, cksum)):
2489 				if (l4_proto != IPPROTO_TCP)
2490 					goto error;
2491 				m->ol_flags |= RTE_MBUF_F_TX_TCP_CKSUM;
2492 				break;
2493 			case (offsetof(struct rte_udp_hdr, dgram_cksum)):
2494 				if (l4_proto != IPPROTO_UDP)
2495 					goto error;
2496 				m->ol_flags |= RTE_MBUF_F_TX_UDP_CKSUM;
2497 				break;
2498 			case (offsetof(struct rte_sctp_hdr, cksum)):
2499 				if (l4_proto != IPPROTO_SCTP)
2500 					goto error;
2501 				m->ol_flags |= RTE_MBUF_F_TX_SCTP_CKSUM;
2502 				break;
2503 			default:
2504 				goto error;
2505 			}
2506 		} else {
2507 			goto error;
2508 		}
2509 	}
2510 
2511 	if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
2512 		switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
2513 		case VIRTIO_NET_HDR_GSO_TCPV4:
2514 		case VIRTIO_NET_HDR_GSO_TCPV6:
2515 			if (l4_proto != IPPROTO_TCP)
2516 				goto error;
2517 			tcp_hdr = rte_pktmbuf_mtod_offset(m,
2518 					struct rte_tcp_hdr *,
2519 					m->l2_len + m->l3_len);
2520 			tcp_len = (tcp_hdr->data_off & 0xf0) >> 2;
2521 			if (data_len < m->l2_len + m->l3_len + tcp_len)
2522 				goto error;
2523 			m->ol_flags |= RTE_MBUF_F_TX_TCP_SEG;
2524 			m->tso_segsz = hdr->gso_size;
2525 			m->l4_len = tcp_len;
2526 			break;
2527 		case VIRTIO_NET_HDR_GSO_UDP:
2528 			if (l4_proto != IPPROTO_UDP)
2529 				goto error;
2530 			m->ol_flags |= RTE_MBUF_F_TX_UDP_SEG;
2531 			m->tso_segsz = hdr->gso_size;
2532 			m->l4_len = sizeof(struct rte_udp_hdr);
2533 			break;
2534 		default:
2535 			VHOST_LOG_DATA(dev->ifname, WARNING,
2536 				"unsupported gso type %u.\n",
2537 				hdr->gso_type);
2538 			goto error;
2539 		}
2540 	}
2541 	return;
2542 
2543 error:
2544 	m->l2_len = 0;
2545 	m->l3_len = 0;
2546 	m->ol_flags = 0;
2547 }
2548 
2549 static __rte_always_inline void
2550 vhost_dequeue_offload(struct virtio_net *dev, struct virtio_net_hdr *hdr,
2551 		struct rte_mbuf *m, bool legacy_ol_flags)
2552 {
2553 	struct rte_net_hdr_lens hdr_lens;
2554 	int l4_supported = 0;
2555 	uint32_t ptype;
2556 
2557 	if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
2558 		return;
2559 
2560 	if (legacy_ol_flags) {
2561 		vhost_dequeue_offload_legacy(dev, hdr, m);
2562 		return;
2563 	}
2564 
2565 	m->ol_flags |= RTE_MBUF_F_RX_IP_CKSUM_UNKNOWN;
2566 
2567 	ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
2568 	m->packet_type = ptype;
2569 	if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
2570 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
2571 	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
2572 		l4_supported = 1;
2573 
2574 	/* According to Virtio 1.1 spec, the device only needs to look at
2575 	 * VIRTIO_NET_HDR_F_NEEDS_CSUM in the packet transmission path.
2576 	 * This differs from the processing incoming packets path where the
2577 	 * driver could rely on VIRTIO_NET_HDR_F_DATA_VALID flag set by the
2578 	 * device.
2579 	 *
2580 	 * 5.1.6.2.1 Driver Requirements: Packet Transmission
2581 	 * The driver MUST NOT set the VIRTIO_NET_HDR_F_DATA_VALID and
2582 	 * VIRTIO_NET_HDR_F_RSC_INFO bits in flags.
2583 	 *
2584 	 * 5.1.6.2.2 Device Requirements: Packet Transmission
2585 	 * The device MUST ignore flag bits that it does not recognize.
2586 	 */
2587 	if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
2588 		uint32_t hdrlen;
2589 
2590 		hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
2591 		if (hdr->csum_start <= hdrlen && l4_supported != 0) {
2592 			m->ol_flags |= RTE_MBUF_F_RX_L4_CKSUM_NONE;
2593 		} else {
2594 			/* Unknown proto or tunnel, do sw cksum. We can assume
2595 			 * the cksum field is in the first segment since the
2596 			 * buffers we provided to the host are large enough.
2597 			 * In case of SCTP, this will be wrong since it's a CRC
2598 			 * but there's nothing we can do.
2599 			 */
2600 			uint16_t csum = 0, off;
2601 
2602 			if (rte_raw_cksum_mbuf(m, hdr->csum_start,
2603 					rte_pktmbuf_pkt_len(m) - hdr->csum_start, &csum) < 0)
2604 				return;
2605 			if (likely(csum != 0xffff))
2606 				csum = ~csum;
2607 			off = hdr->csum_offset + hdr->csum_start;
2608 			if (rte_pktmbuf_data_len(m) >= off + 1)
2609 				*rte_pktmbuf_mtod_offset(m, uint16_t *, off) = csum;
2610 		}
2611 	}
2612 
2613 	if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
2614 		if (hdr->gso_size == 0)
2615 			return;
2616 
2617 		switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
2618 		case VIRTIO_NET_HDR_GSO_TCPV4:
2619 		case VIRTIO_NET_HDR_GSO_TCPV6:
2620 			if ((ptype & RTE_PTYPE_L4_MASK) != RTE_PTYPE_L4_TCP)
2621 				break;
2622 			m->ol_flags |= RTE_MBUF_F_RX_LRO | RTE_MBUF_F_RX_L4_CKSUM_NONE;
2623 			m->tso_segsz = hdr->gso_size;
2624 			break;
2625 		case VIRTIO_NET_HDR_GSO_UDP:
2626 			if ((ptype & RTE_PTYPE_L4_MASK) != RTE_PTYPE_L4_UDP)
2627 				break;
2628 			m->ol_flags |= RTE_MBUF_F_RX_LRO | RTE_MBUF_F_RX_L4_CKSUM_NONE;
2629 			m->tso_segsz = hdr->gso_size;
2630 			break;
2631 		default:
2632 			break;
2633 		}
2634 	}
2635 }
2636 
2637 static __rte_noinline void
2638 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
2639 		struct buf_vector *buf_vec)
2640 {
2641 	uint64_t len;
2642 	uint64_t remain = sizeof(struct virtio_net_hdr);
2643 	uint64_t src;
2644 	uint64_t dst = (uint64_t)(uintptr_t)hdr;
2645 
2646 	while (remain) {
2647 		len = RTE_MIN(remain, buf_vec->buf_len);
2648 		src = buf_vec->buf_addr;
2649 		rte_memcpy((void *)(uintptr_t)dst,
2650 				(void *)(uintptr_t)src, len);
2651 
2652 		remain -= len;
2653 		dst += len;
2654 		buf_vec++;
2655 	}
2656 }
2657 
2658 static __rte_always_inline int
2659 desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
2660 		  struct buf_vector *buf_vec, uint16_t nr_vec,
2661 		  struct rte_mbuf *m, struct rte_mempool *mbuf_pool,
2662 		  bool legacy_ol_flags, uint16_t slot_idx, bool is_async)
2663 {
2664 	uint32_t buf_avail, buf_offset, buf_len;
2665 	uint64_t buf_addr, buf_iova;
2666 	uint32_t mbuf_avail, mbuf_offset;
2667 	uint32_t cpy_len;
2668 	struct rte_mbuf *cur = m, *prev = m;
2669 	struct virtio_net_hdr tmp_hdr;
2670 	struct virtio_net_hdr *hdr = NULL;
2671 	/* A counter to avoid desc dead loop chain */
2672 	uint16_t vec_idx = 0;
2673 	struct vhost_async *async = vq->async;
2674 	struct async_inflight_info *pkts_info;
2675 
2676 	buf_addr = buf_vec[vec_idx].buf_addr;
2677 	buf_iova = buf_vec[vec_idx].buf_iova;
2678 	buf_len = buf_vec[vec_idx].buf_len;
2679 
2680 	if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1))
2681 		return -1;
2682 
2683 	if (virtio_net_with_host_offload(dev)) {
2684 		if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
2685 			/*
2686 			 * No luck, the virtio-net header doesn't fit
2687 			 * in a contiguous virtual area.
2688 			 */
2689 			copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
2690 			hdr = &tmp_hdr;
2691 		} else {
2692 			hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
2693 		}
2694 	}
2695 
2696 	/*
2697 	 * A virtio driver normally uses at least 2 desc buffers
2698 	 * for Tx: the first for storing the header, and others
2699 	 * for storing the data.
2700 	 */
2701 	if (unlikely(buf_len < dev->vhost_hlen)) {
2702 		buf_offset = dev->vhost_hlen - buf_len;
2703 		vec_idx++;
2704 		buf_addr = buf_vec[vec_idx].buf_addr;
2705 		buf_iova = buf_vec[vec_idx].buf_iova;
2706 		buf_len = buf_vec[vec_idx].buf_len;
2707 		buf_avail  = buf_len - buf_offset;
2708 	} else if (buf_len == dev->vhost_hlen) {
2709 		if (unlikely(++vec_idx >= nr_vec))
2710 			goto error;
2711 		buf_addr = buf_vec[vec_idx].buf_addr;
2712 		buf_iova = buf_vec[vec_idx].buf_iova;
2713 		buf_len = buf_vec[vec_idx].buf_len;
2714 
2715 		buf_offset = 0;
2716 		buf_avail = buf_len;
2717 	} else {
2718 		buf_offset = dev->vhost_hlen;
2719 		buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
2720 	}
2721 
2722 	PRINT_PACKET(dev,
2723 			(uintptr_t)(buf_addr + buf_offset),
2724 			(uint32_t)buf_avail, 0);
2725 
2726 	mbuf_offset = 0;
2727 	mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
2728 
2729 	if (is_async) {
2730 		pkts_info = async->pkts_info;
2731 		if (async_iter_initialize(dev, async))
2732 			return -1;
2733 	}
2734 
2735 	while (1) {
2736 		cpy_len = RTE_MIN(buf_avail, mbuf_avail);
2737 
2738 		if (is_async) {
2739 			if (async_fill_seg(dev, vq, cur, mbuf_offset,
2740 					   buf_iova + buf_offset, cpy_len, false) < 0)
2741 				goto error;
2742 		} else if (likely(hdr && cur == m)) {
2743 			rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *, mbuf_offset),
2744 				(void *)((uintptr_t)(buf_addr + buf_offset)),
2745 				cpy_len);
2746 		} else {
2747 			sync_fill_seg(dev, vq, cur, mbuf_offset,
2748 				      buf_addr + buf_offset,
2749 				      buf_iova + buf_offset, cpy_len, false);
2750 		}
2751 
2752 		mbuf_avail  -= cpy_len;
2753 		mbuf_offset += cpy_len;
2754 		buf_avail -= cpy_len;
2755 		buf_offset += cpy_len;
2756 
2757 		/* This buf reaches to its end, get the next one */
2758 		if (buf_avail == 0) {
2759 			if (++vec_idx >= nr_vec)
2760 				break;
2761 
2762 			buf_addr = buf_vec[vec_idx].buf_addr;
2763 			buf_iova = buf_vec[vec_idx].buf_iova;
2764 			buf_len = buf_vec[vec_idx].buf_len;
2765 
2766 			buf_offset = 0;
2767 			buf_avail  = buf_len;
2768 
2769 			PRINT_PACKET(dev, (uintptr_t)buf_addr,
2770 					(uint32_t)buf_avail, 0);
2771 		}
2772 
2773 		/*
2774 		 * This mbuf reaches to its end, get a new one
2775 		 * to hold more data.
2776 		 */
2777 		if (mbuf_avail == 0) {
2778 			cur = rte_pktmbuf_alloc(mbuf_pool);
2779 			if (unlikely(cur == NULL)) {
2780 				VHOST_LOG_DATA(dev->ifname, ERR,
2781 					"failed to allocate memory for mbuf.\n");
2782 				goto error;
2783 			}
2784 
2785 			prev->next = cur;
2786 			prev->data_len = mbuf_offset;
2787 			m->nb_segs += 1;
2788 			m->pkt_len += mbuf_offset;
2789 			prev = cur;
2790 
2791 			mbuf_offset = 0;
2792 			mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
2793 		}
2794 	}
2795 
2796 	prev->data_len = mbuf_offset;
2797 	m->pkt_len    += mbuf_offset;
2798 
2799 	if (is_async) {
2800 		async_iter_finalize(async);
2801 		if (hdr)
2802 			pkts_info[slot_idx].nethdr = *hdr;
2803 	} else if (hdr) {
2804 		vhost_dequeue_offload(dev, hdr, m, legacy_ol_flags);
2805 	}
2806 
2807 	return 0;
2808 error:
2809 	if (is_async)
2810 		async_iter_cancel(async);
2811 
2812 	return -1;
2813 }
2814 
2815 static void
2816 virtio_dev_extbuf_free(void *addr __rte_unused, void *opaque)
2817 {
2818 	rte_free(opaque);
2819 }
2820 
2821 static int
2822 virtio_dev_extbuf_alloc(struct virtio_net *dev, struct rte_mbuf *pkt, uint32_t size)
2823 {
2824 	struct rte_mbuf_ext_shared_info *shinfo = NULL;
2825 	uint32_t total_len = RTE_PKTMBUF_HEADROOM + size;
2826 	uint16_t buf_len;
2827 	rte_iova_t iova;
2828 	void *buf;
2829 
2830 	total_len += sizeof(*shinfo) + sizeof(uintptr_t);
2831 	total_len = RTE_ALIGN_CEIL(total_len, sizeof(uintptr_t));
2832 
2833 	if (unlikely(total_len > UINT16_MAX))
2834 		return -ENOSPC;
2835 
2836 	buf_len = total_len;
2837 	buf = rte_malloc(NULL, buf_len, RTE_CACHE_LINE_SIZE);
2838 	if (unlikely(buf == NULL))
2839 		return -ENOMEM;
2840 
2841 	/* Initialize shinfo */
2842 	shinfo = rte_pktmbuf_ext_shinfo_init_helper(buf, &buf_len,
2843 						virtio_dev_extbuf_free, buf);
2844 	if (unlikely(shinfo == NULL)) {
2845 		rte_free(buf);
2846 		VHOST_LOG_DATA(dev->ifname, ERR, "failed to init shinfo\n");
2847 		return -1;
2848 	}
2849 
2850 	iova = rte_malloc_virt2iova(buf);
2851 	rte_pktmbuf_attach_extbuf(pkt, buf, iova, buf_len, shinfo);
2852 	rte_pktmbuf_reset_headroom(pkt);
2853 
2854 	return 0;
2855 }
2856 
2857 /*
2858  * Prepare a host supported pktmbuf.
2859  */
2860 static __rte_always_inline int
2861 virtio_dev_pktmbuf_prep(struct virtio_net *dev, struct rte_mbuf *pkt,
2862 			 uint32_t data_len)
2863 {
2864 	if (rte_pktmbuf_tailroom(pkt) >= data_len)
2865 		return 0;
2866 
2867 	/* attach an external buffer if supported */
2868 	if (dev->extbuf && !virtio_dev_extbuf_alloc(dev, pkt, data_len))
2869 		return 0;
2870 
2871 	/* check if chained buffers are allowed */
2872 	if (!dev->linearbuf)
2873 		return 0;
2874 
2875 	return -1;
2876 }
2877 
2878 __rte_always_inline
2879 static uint16_t
2880 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
2881 	struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count,
2882 	bool legacy_ol_flags)
2883 {
2884 	uint16_t i;
2885 	uint16_t avail_entries;
2886 	uint16_t dropped = 0;
2887 	static bool allocerr_warned;
2888 
2889 	/*
2890 	 * The ordering between avail index and
2891 	 * desc reads needs to be enforced.
2892 	 */
2893 	avail_entries = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE) -
2894 			vq->last_avail_idx;
2895 	if (avail_entries == 0)
2896 		return 0;
2897 
2898 	rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
2899 
2900 	VHOST_LOG_DATA(dev->ifname, DEBUG, "%s\n", __func__);
2901 
2902 	count = RTE_MIN(count, MAX_PKT_BURST);
2903 	count = RTE_MIN(count, avail_entries);
2904 	VHOST_LOG_DATA(dev->ifname, DEBUG, "about to dequeue %u buffers\n", count);
2905 
2906 	if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts, count))
2907 		return 0;
2908 
2909 	for (i = 0; i < count; i++) {
2910 		struct buf_vector buf_vec[BUF_VECTOR_MAX];
2911 		uint16_t head_idx;
2912 		uint32_t buf_len;
2913 		uint16_t nr_vec = 0;
2914 		int err;
2915 
2916 		if (unlikely(fill_vec_buf_split(dev, vq,
2917 						vq->last_avail_idx + i,
2918 						&nr_vec, buf_vec,
2919 						&head_idx, &buf_len,
2920 						VHOST_ACCESS_RO) < 0))
2921 			break;
2922 
2923 		update_shadow_used_ring_split(vq, head_idx, 0);
2924 
2925 		err = virtio_dev_pktmbuf_prep(dev, pkts[i], buf_len);
2926 		if (unlikely(err)) {
2927 			/*
2928 			 * mbuf allocation fails for jumbo packets when external
2929 			 * buffer allocation is not allowed and linear buffer
2930 			 * is required. Drop this packet.
2931 			 */
2932 			if (!allocerr_warned) {
2933 				VHOST_LOG_DATA(dev->ifname, ERR,
2934 					"failed mbuf alloc of size %d from %s.\n",
2935 					buf_len, mbuf_pool->name);
2936 				allocerr_warned = true;
2937 			}
2938 			dropped += 1;
2939 			i++;
2940 			break;
2941 		}
2942 
2943 		err = desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
2944 				   mbuf_pool, legacy_ol_flags, 0, false);
2945 		if (unlikely(err)) {
2946 			if (!allocerr_warned) {
2947 				VHOST_LOG_DATA(dev->ifname, ERR, "failed to copy desc to mbuf.\n");
2948 				allocerr_warned = true;
2949 			}
2950 			dropped += 1;
2951 			i++;
2952 			break;
2953 		}
2954 
2955 	}
2956 
2957 	if (dropped)
2958 		rte_pktmbuf_free_bulk(&pkts[i - 1], count - i + 1);
2959 
2960 	vq->last_avail_idx += i;
2961 
2962 	do_data_copy_dequeue(vq);
2963 	if (unlikely(i < count))
2964 		vq->shadow_used_idx = i;
2965 	if (likely(vq->shadow_used_idx)) {
2966 		flush_shadow_used_ring_split(dev, vq);
2967 		vhost_vring_call_split(dev, vq);
2968 	}
2969 
2970 	return (i - dropped);
2971 }
2972 
2973 __rte_noinline
2974 static uint16_t
2975 virtio_dev_tx_split_legacy(struct virtio_net *dev,
2976 	struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
2977 	struct rte_mbuf **pkts, uint16_t count)
2978 {
2979 	return virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count, true);
2980 }
2981 
2982 __rte_noinline
2983 static uint16_t
2984 virtio_dev_tx_split_compliant(struct virtio_net *dev,
2985 	struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
2986 	struct rte_mbuf **pkts, uint16_t count)
2987 {
2988 	return virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count, false);
2989 }
2990 
2991 static __rte_always_inline int
2992 vhost_reserve_avail_batch_packed(struct virtio_net *dev,
2993 				 struct vhost_virtqueue *vq,
2994 				 struct rte_mbuf **pkts,
2995 				 uint16_t avail_idx,
2996 				 uintptr_t *desc_addrs,
2997 				 uint16_t *ids)
2998 {
2999 	bool wrap = vq->avail_wrap_counter;
3000 	struct vring_packed_desc *descs = vq->desc_packed;
3001 	uint64_t lens[PACKED_BATCH_SIZE];
3002 	uint64_t buf_lens[PACKED_BATCH_SIZE];
3003 	uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3004 	uint16_t flags, i;
3005 
3006 	if (unlikely(avail_idx & PACKED_BATCH_MASK))
3007 		return -1;
3008 	if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
3009 		return -1;
3010 
3011 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
3012 		flags = descs[avail_idx + i].flags;
3013 		if (unlikely((wrap != !!(flags & VRING_DESC_F_AVAIL)) ||
3014 			     (wrap == !!(flags & VRING_DESC_F_USED))  ||
3015 			     (flags & PACKED_DESC_SINGLE_DEQUEUE_FLAG)))
3016 			return -1;
3017 	}
3018 
3019 	rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
3020 
3021 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
3022 		lens[i] = descs[avail_idx + i].len;
3023 
3024 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
3025 		desc_addrs[i] = vhost_iova_to_vva(dev, vq,
3026 						  descs[avail_idx + i].addr,
3027 						  &lens[i], VHOST_ACCESS_RW);
3028 	}
3029 
3030 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
3031 		if (unlikely(!desc_addrs[i]))
3032 			return -1;
3033 		if (unlikely((lens[i] != descs[avail_idx + i].len)))
3034 			return -1;
3035 	}
3036 
3037 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
3038 		if (virtio_dev_pktmbuf_prep(dev, pkts[i], lens[i]))
3039 			goto err;
3040 	}
3041 
3042 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
3043 		buf_lens[i] = pkts[i]->buf_len - pkts[i]->data_off;
3044 
3045 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
3046 		if (unlikely(buf_lens[i] < (lens[i] - buf_offset)))
3047 			goto err;
3048 	}
3049 
3050 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
3051 		pkts[i]->pkt_len = lens[i] - buf_offset;
3052 		pkts[i]->data_len = pkts[i]->pkt_len;
3053 		ids[i] = descs[avail_idx + i].id;
3054 	}
3055 
3056 	return 0;
3057 
3058 err:
3059 	return -1;
3060 }
3061 
3062 static __rte_always_inline int
3063 virtio_dev_tx_batch_packed(struct virtio_net *dev,
3064 			   struct vhost_virtqueue *vq,
3065 			   struct rte_mbuf **pkts,
3066 			   bool legacy_ol_flags)
3067 {
3068 	uint16_t avail_idx = vq->last_avail_idx;
3069 	uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3070 	struct virtio_net_hdr *hdr;
3071 	uintptr_t desc_addrs[PACKED_BATCH_SIZE];
3072 	uint16_t ids[PACKED_BATCH_SIZE];
3073 	uint16_t i;
3074 
3075 	if (vhost_reserve_avail_batch_packed(dev, vq, pkts, avail_idx,
3076 					     desc_addrs, ids))
3077 		return -1;
3078 
3079 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
3080 		rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
3081 
3082 	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
3083 		rte_memcpy(rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
3084 			   (void *)(uintptr_t)(desc_addrs[i] + buf_offset),
3085 			   pkts[i]->pkt_len);
3086 
3087 	if (virtio_net_with_host_offload(dev)) {
3088 		vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
3089 			hdr = (struct virtio_net_hdr *)(desc_addrs[i]);
3090 			vhost_dequeue_offload(dev, hdr, pkts[i], legacy_ol_flags);
3091 		}
3092 	}
3093 
3094 	if (virtio_net_is_inorder(dev))
3095 		vhost_shadow_dequeue_batch_packed_inorder(vq,
3096 			ids[PACKED_BATCH_SIZE - 1]);
3097 	else
3098 		vhost_shadow_dequeue_batch_packed(dev, vq, ids);
3099 
3100 	vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
3101 
3102 	return 0;
3103 }
3104 
3105 static __rte_always_inline int
3106 vhost_dequeue_single_packed(struct virtio_net *dev,
3107 			    struct vhost_virtqueue *vq,
3108 			    struct rte_mempool *mbuf_pool,
3109 			    struct rte_mbuf *pkts,
3110 			    uint16_t *buf_id,
3111 			    uint16_t *desc_count,
3112 			    bool legacy_ol_flags)
3113 {
3114 	struct buf_vector buf_vec[BUF_VECTOR_MAX];
3115 	uint32_t buf_len;
3116 	uint16_t nr_vec = 0;
3117 	int err;
3118 	static bool allocerr_warned;
3119 
3120 	if (unlikely(fill_vec_buf_packed(dev, vq,
3121 					 vq->last_avail_idx, desc_count,
3122 					 buf_vec, &nr_vec,
3123 					 buf_id, &buf_len,
3124 					 VHOST_ACCESS_RO) < 0))
3125 		return -1;
3126 
3127 	if (unlikely(virtio_dev_pktmbuf_prep(dev, pkts, buf_len))) {
3128 		if (!allocerr_warned) {
3129 			VHOST_LOG_DATA(dev->ifname, ERR,
3130 				"failed mbuf alloc of size %d from %s.\n",
3131 				buf_len, mbuf_pool->name);
3132 			allocerr_warned = true;
3133 		}
3134 		return -1;
3135 	}
3136 
3137 	err = desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts,
3138 			   mbuf_pool, legacy_ol_flags, 0, false);
3139 	if (unlikely(err)) {
3140 		if (!allocerr_warned) {
3141 			VHOST_LOG_DATA(dev->ifname, ERR, "failed to copy desc to mbuf.\n");
3142 			allocerr_warned = true;
3143 		}
3144 		return -1;
3145 	}
3146 
3147 	return 0;
3148 }
3149 
3150 static __rte_always_inline int
3151 virtio_dev_tx_single_packed(struct virtio_net *dev,
3152 			    struct vhost_virtqueue *vq,
3153 			    struct rte_mempool *mbuf_pool,
3154 			    struct rte_mbuf *pkts,
3155 			    bool legacy_ol_flags)
3156 {
3157 
3158 	uint16_t buf_id, desc_count = 0;
3159 	int ret;
3160 
3161 	ret = vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
3162 					&desc_count, legacy_ol_flags);
3163 
3164 	if (likely(desc_count > 0)) {
3165 		if (virtio_net_is_inorder(dev))
3166 			vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
3167 								   desc_count);
3168 		else
3169 			vhost_shadow_dequeue_single_packed(vq, buf_id,
3170 					desc_count);
3171 
3172 		vq_inc_last_avail_packed(vq, desc_count);
3173 	}
3174 
3175 	return ret;
3176 }
3177 
3178 __rte_always_inline
3179 static uint16_t
3180 virtio_dev_tx_packed(struct virtio_net *dev,
3181 		     struct vhost_virtqueue *__rte_restrict vq,
3182 		     struct rte_mempool *mbuf_pool,
3183 		     struct rte_mbuf **__rte_restrict pkts,
3184 		     uint32_t count,
3185 		     bool legacy_ol_flags)
3186 {
3187 	uint32_t pkt_idx = 0;
3188 
3189 	if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts, count))
3190 		return 0;
3191 
3192 	do {
3193 		rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
3194 
3195 		if (count - pkt_idx >= PACKED_BATCH_SIZE) {
3196 			if (!virtio_dev_tx_batch_packed(dev, vq,
3197 							&pkts[pkt_idx],
3198 							legacy_ol_flags)) {
3199 				pkt_idx += PACKED_BATCH_SIZE;
3200 				continue;
3201 			}
3202 		}
3203 
3204 		if (virtio_dev_tx_single_packed(dev, vq, mbuf_pool,
3205 						pkts[pkt_idx],
3206 						legacy_ol_flags))
3207 			break;
3208 		pkt_idx++;
3209 	} while (pkt_idx < count);
3210 
3211 	if (pkt_idx != count)
3212 		rte_pktmbuf_free_bulk(&pkts[pkt_idx], count - pkt_idx);
3213 
3214 	if (vq->shadow_used_idx) {
3215 		do_data_copy_dequeue(vq);
3216 
3217 		vhost_flush_dequeue_shadow_packed(dev, vq);
3218 		vhost_vring_call_packed(dev, vq);
3219 	}
3220 
3221 	return pkt_idx;
3222 }
3223 
3224 __rte_noinline
3225 static uint16_t
3226 virtio_dev_tx_packed_legacy(struct virtio_net *dev,
3227 	struct vhost_virtqueue *__rte_restrict vq, struct rte_mempool *mbuf_pool,
3228 	struct rte_mbuf **__rte_restrict pkts, uint32_t count)
3229 {
3230 	return virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count, true);
3231 }
3232 
3233 __rte_noinline
3234 static uint16_t
3235 virtio_dev_tx_packed_compliant(struct virtio_net *dev,
3236 	struct vhost_virtqueue *__rte_restrict vq, struct rte_mempool *mbuf_pool,
3237 	struct rte_mbuf **__rte_restrict pkts, uint32_t count)
3238 {
3239 	return virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count, false);
3240 }
3241 
3242 uint16_t
3243 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
3244 	struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
3245 {
3246 	struct virtio_net *dev;
3247 	struct rte_mbuf *rarp_mbuf = NULL;
3248 	struct vhost_virtqueue *vq;
3249 	int16_t success = 1;
3250 
3251 	dev = get_device(vid);
3252 	if (!dev)
3253 		return 0;
3254 
3255 	if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
3256 		VHOST_LOG_DATA(dev->ifname, ERR,
3257 			"%s: built-in vhost net backend is disabled.\n",
3258 			__func__);
3259 		return 0;
3260 	}
3261 
3262 	if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
3263 		VHOST_LOG_DATA(dev->ifname, ERR,
3264 			"%s: invalid virtqueue idx %d.\n",
3265 			__func__, queue_id);
3266 		return 0;
3267 	}
3268 
3269 	vq = dev->virtqueue[queue_id];
3270 
3271 	if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
3272 		return 0;
3273 
3274 	if (unlikely(!vq->enabled)) {
3275 		count = 0;
3276 		goto out_access_unlock;
3277 	}
3278 
3279 	if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
3280 		vhost_user_iotlb_rd_lock(vq);
3281 
3282 	if (unlikely(!vq->access_ok))
3283 		if (unlikely(vring_translate(dev, vq) < 0)) {
3284 			count = 0;
3285 			goto out;
3286 		}
3287 
3288 	/*
3289 	 * Construct a RARP broadcast packet, and inject it to the "pkts"
3290 	 * array, to looks like that guest actually send such packet.
3291 	 *
3292 	 * Check user_send_rarp() for more information.
3293 	 *
3294 	 * broadcast_rarp shares a cacheline in the virtio_net structure
3295 	 * with some fields that are accessed during enqueue and
3296 	 * __atomic_compare_exchange_n causes a write if performed compare
3297 	 * and exchange. This could result in false sharing between enqueue
3298 	 * and dequeue.
3299 	 *
3300 	 * Prevent unnecessary false sharing by reading broadcast_rarp first
3301 	 * and only performing compare and exchange if the read indicates it
3302 	 * is likely to be set.
3303 	 */
3304 	if (unlikely(__atomic_load_n(&dev->broadcast_rarp, __ATOMIC_ACQUIRE) &&
3305 			__atomic_compare_exchange_n(&dev->broadcast_rarp,
3306 			&success, 0, 0, __ATOMIC_RELEASE, __ATOMIC_RELAXED))) {
3307 
3308 		rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
3309 		if (rarp_mbuf == NULL) {
3310 			VHOST_LOG_DATA(dev->ifname, ERR, "failed to make RARP packet.\n");
3311 			count = 0;
3312 			goto out;
3313 		}
3314 		/*
3315 		 * Inject it to the head of "pkts" array, so that switch's mac
3316 		 * learning table will get updated first.
3317 		 */
3318 		pkts[0] = rarp_mbuf;
3319 		vhost_queue_stats_update(dev, vq, pkts, 1);
3320 		pkts++;
3321 		count -= 1;
3322 	}
3323 
3324 	if (vq_is_packed(dev)) {
3325 		if (dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS)
3326 			count = virtio_dev_tx_packed_legacy(dev, vq, mbuf_pool, pkts, count);
3327 		else
3328 			count = virtio_dev_tx_packed_compliant(dev, vq, mbuf_pool, pkts, count);
3329 	} else {
3330 		if (dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS)
3331 			count = virtio_dev_tx_split_legacy(dev, vq, mbuf_pool, pkts, count);
3332 		else
3333 			count = virtio_dev_tx_split_compliant(dev, vq, mbuf_pool, pkts, count);
3334 	}
3335 
3336 	vhost_queue_stats_update(dev, vq, pkts, count);
3337 
3338 out:
3339 	if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
3340 		vhost_user_iotlb_rd_unlock(vq);
3341 
3342 out_access_unlock:
3343 	rte_spinlock_unlock(&vq->access_lock);
3344 
3345 	if (unlikely(rarp_mbuf != NULL))
3346 		count += 1;
3347 
3348 	return count;
3349 }
3350 
3351 static __rte_always_inline uint16_t
3352 async_poll_dequeue_completed(struct virtio_net *dev, struct vhost_virtqueue *vq,
3353 		struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
3354 		uint16_t vchan_id, bool legacy_ol_flags)
3355 {
3356 	uint16_t start_idx, from, i;
3357 	uint16_t nr_cpl_pkts = 0;
3358 	struct async_inflight_info *pkts_info = vq->async->pkts_info;
3359 
3360 	vhost_async_dma_check_completed(dev, dma_id, vchan_id, VHOST_DMA_MAX_COPY_COMPLETE);
3361 
3362 	start_idx = async_get_first_inflight_pkt_idx(vq);
3363 
3364 	from = start_idx;
3365 	while (vq->async->pkts_cmpl_flag[from] && count--) {
3366 		vq->async->pkts_cmpl_flag[from] = false;
3367 		from = (from + 1) % vq->size;
3368 		nr_cpl_pkts++;
3369 	}
3370 
3371 	if (nr_cpl_pkts == 0)
3372 		return 0;
3373 
3374 	for (i = 0; i < nr_cpl_pkts; i++) {
3375 		from = (start_idx + i) % vq->size;
3376 		pkts[i] = pkts_info[from].mbuf;
3377 
3378 		if (virtio_net_with_host_offload(dev))
3379 			vhost_dequeue_offload(dev, &pkts_info[from].nethdr, pkts[i],
3380 					      legacy_ol_flags);
3381 	}
3382 
3383 	/* write back completed descs to used ring and update used idx */
3384 	if (vq_is_packed(dev)) {
3385 		write_back_completed_descs_packed(vq, nr_cpl_pkts);
3386 		vhost_vring_call_packed(dev, vq);
3387 	} else {
3388 		write_back_completed_descs_split(vq, nr_cpl_pkts);
3389 		__atomic_add_fetch(&vq->used->idx, nr_cpl_pkts, __ATOMIC_RELEASE);
3390 		vhost_vring_call_split(dev, vq);
3391 	}
3392 	vq->async->pkts_inflight_n -= nr_cpl_pkts;
3393 
3394 	return nr_cpl_pkts;
3395 }
3396 
3397 static __rte_always_inline uint16_t
3398 virtio_dev_tx_async_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
3399 		struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count,
3400 		int16_t dma_id, uint16_t vchan_id, bool legacy_ol_flags)
3401 {
3402 	static bool allocerr_warned;
3403 	bool dropped = false;
3404 	uint16_t avail_entries;
3405 	uint16_t pkt_idx, slot_idx = 0;
3406 	uint16_t nr_done_pkts = 0;
3407 	uint16_t pkt_err = 0;
3408 	uint16_t n_xfer;
3409 	struct vhost_async *async = vq->async;
3410 	struct async_inflight_info *pkts_info = async->pkts_info;
3411 	struct rte_mbuf *pkts_prealloc[MAX_PKT_BURST];
3412 	uint16_t pkts_size = count;
3413 
3414 	/**
3415 	 * The ordering between avail index and
3416 	 * desc reads needs to be enforced.
3417 	 */
3418 	avail_entries = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE) -
3419 			vq->last_avail_idx;
3420 	if (avail_entries == 0)
3421 		goto out;
3422 
3423 	rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
3424 
3425 	async_iter_reset(async);
3426 
3427 	count = RTE_MIN(count, MAX_PKT_BURST);
3428 	count = RTE_MIN(count, avail_entries);
3429 	VHOST_LOG_DATA(dev->ifname, DEBUG, "about to dequeue %u buffers\n", count);
3430 
3431 	if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts_prealloc, count))
3432 		goto out;
3433 
3434 	for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
3435 		uint16_t head_idx = 0;
3436 		uint16_t nr_vec = 0;
3437 		uint16_t to;
3438 		uint32_t buf_len;
3439 		int err;
3440 		struct buf_vector buf_vec[BUF_VECTOR_MAX];
3441 		struct rte_mbuf *pkt = pkts_prealloc[pkt_idx];
3442 
3443 		if (unlikely(fill_vec_buf_split(dev, vq, vq->last_avail_idx,
3444 						&nr_vec, buf_vec,
3445 						&head_idx, &buf_len,
3446 						VHOST_ACCESS_RO) < 0)) {
3447 			dropped = true;
3448 			break;
3449 		}
3450 
3451 		err = virtio_dev_pktmbuf_prep(dev, pkt, buf_len);
3452 		if (unlikely(err)) {
3453 			/**
3454 			 * mbuf allocation fails for jumbo packets when external
3455 			 * buffer allocation is not allowed and linear buffer
3456 			 * is required. Drop this packet.
3457 			 */
3458 			if (!allocerr_warned) {
3459 				VHOST_LOG_DATA(dev->ifname, ERR,
3460 					"%s: Failed mbuf alloc of size %d from %s\n",
3461 					__func__, buf_len, mbuf_pool->name);
3462 				allocerr_warned = true;
3463 			}
3464 			dropped = true;
3465 			break;
3466 		}
3467 
3468 		slot_idx = (async->pkts_idx + pkt_idx) & (vq->size - 1);
3469 		err = desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkt, mbuf_pool,
3470 					legacy_ol_flags, slot_idx, true);
3471 		if (unlikely(err)) {
3472 			if (!allocerr_warned) {
3473 				VHOST_LOG_DATA(dev->ifname, ERR,
3474 					"%s: Failed to offload copies to async channel.\n",
3475 					__func__);
3476 				allocerr_warned = true;
3477 			}
3478 			dropped = true;
3479 			break;
3480 		}
3481 
3482 		pkts_info[slot_idx].mbuf = pkt;
3483 
3484 		/* store used descs */
3485 		to = async->desc_idx_split & (vq->size - 1);
3486 		async->descs_split[to].id = head_idx;
3487 		async->descs_split[to].len = 0;
3488 		async->desc_idx_split++;
3489 
3490 		vq->last_avail_idx++;
3491 	}
3492 
3493 	if (unlikely(dropped))
3494 		rte_pktmbuf_free_bulk(&pkts_prealloc[pkt_idx], count - pkt_idx);
3495 
3496 	n_xfer = vhost_async_dma_transfer(dev, vq, dma_id, vchan_id, async->pkts_idx,
3497 					  async->iov_iter, pkt_idx);
3498 
3499 	async->pkts_inflight_n += n_xfer;
3500 
3501 	pkt_err = pkt_idx - n_xfer;
3502 	if (unlikely(pkt_err)) {
3503 		VHOST_LOG_DATA(dev->ifname, DEBUG, "%s: failed to transfer data.\n",
3504 			__func__);
3505 
3506 		pkt_idx = n_xfer;
3507 		/* recover available ring */
3508 		vq->last_avail_idx -= pkt_err;
3509 
3510 		/**
3511 		 * recover async channel copy related structures and free pktmbufs
3512 		 * for error pkts.
3513 		 */
3514 		async->desc_idx_split -= pkt_err;
3515 		while (pkt_err-- > 0) {
3516 			rte_pktmbuf_free(pkts_info[slot_idx & (vq->size - 1)].mbuf);
3517 			slot_idx--;
3518 		}
3519 	}
3520 
3521 	async->pkts_idx += pkt_idx;
3522 	if (async->pkts_idx >= vq->size)
3523 		async->pkts_idx -= vq->size;
3524 
3525 out:
3526 	/* DMA device may serve other queues, unconditionally check completed. */
3527 	nr_done_pkts = async_poll_dequeue_completed(dev, vq, pkts, pkts_size,
3528 							dma_id, vchan_id, legacy_ol_flags);
3529 
3530 	return nr_done_pkts;
3531 }
3532 
3533 __rte_noinline
3534 static uint16_t
3535 virtio_dev_tx_async_split_legacy(struct virtio_net *dev,
3536 		struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
3537 		struct rte_mbuf **pkts, uint16_t count,
3538 		int16_t dma_id, uint16_t vchan_id)
3539 {
3540 	return virtio_dev_tx_async_split(dev, vq, mbuf_pool,
3541 				pkts, count, dma_id, vchan_id, true);
3542 }
3543 
3544 __rte_noinline
3545 static uint16_t
3546 virtio_dev_tx_async_split_compliant(struct virtio_net *dev,
3547 		struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
3548 		struct rte_mbuf **pkts, uint16_t count,
3549 		int16_t dma_id, uint16_t vchan_id)
3550 {
3551 	return virtio_dev_tx_async_split(dev, vq, mbuf_pool,
3552 				pkts, count, dma_id, vchan_id, false);
3553 }
3554 
3555 static __rte_always_inline void
3556 vhost_async_shadow_dequeue_single_packed(struct vhost_virtqueue *vq, uint16_t buf_id)
3557 {
3558 	struct vhost_async *async = vq->async;
3559 	uint16_t idx = async->buffer_idx_packed;
3560 
3561 	async->buffers_packed[idx].id = buf_id;
3562 	async->buffers_packed[idx].len = 0;
3563 	async->buffers_packed[idx].count = 1;
3564 
3565 	async->buffer_idx_packed++;
3566 	if (async->buffer_idx_packed >= vq->size)
3567 		async->buffer_idx_packed -= vq->size;
3568 
3569 }
3570 
3571 static __rte_always_inline int
3572 virtio_dev_tx_async_single_packed(struct virtio_net *dev,
3573 			struct vhost_virtqueue *vq,
3574 			struct rte_mempool *mbuf_pool,
3575 			struct rte_mbuf *pkts,
3576 			uint16_t slot_idx,
3577 			bool legacy_ol_flags)
3578 {
3579 	int err;
3580 	uint16_t buf_id, desc_count = 0;
3581 	uint16_t nr_vec = 0;
3582 	uint32_t buf_len;
3583 	struct buf_vector buf_vec[BUF_VECTOR_MAX];
3584 	static bool allocerr_warned;
3585 
3586 	if (unlikely(fill_vec_buf_packed(dev, vq, vq->last_avail_idx, &desc_count,
3587 					 buf_vec, &nr_vec, &buf_id, &buf_len,
3588 					 VHOST_ACCESS_RO) < 0))
3589 		return -1;
3590 
3591 	if (unlikely(virtio_dev_pktmbuf_prep(dev, pkts, buf_len))) {
3592 		if (!allocerr_warned) {
3593 			VHOST_LOG_DATA(dev->ifname, ERR, "Failed mbuf alloc of size %d from %s.\n",
3594 				buf_len, mbuf_pool->name);
3595 
3596 			allocerr_warned = true;
3597 		}
3598 		return -1;
3599 	}
3600 
3601 	err = desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts, mbuf_pool,
3602 		legacy_ol_flags, slot_idx, true);
3603 	if (unlikely(err)) {
3604 		rte_pktmbuf_free(pkts);
3605 		if (!allocerr_warned) {
3606 			VHOST_LOG_DATA(dev->ifname, ERR, "Failed to copy desc to mbuf on.\n");
3607 			allocerr_warned = true;
3608 		}
3609 		return -1;
3610 	}
3611 
3612 	/* update async shadow packed ring */
3613 	vhost_async_shadow_dequeue_single_packed(vq, buf_id);
3614 
3615 	return err;
3616 }
3617 
3618 static __rte_always_inline uint16_t
3619 virtio_dev_tx_async_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
3620 		struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts,
3621 		uint16_t count, uint16_t dma_id, uint16_t vchan_id, bool legacy_ol_flags)
3622 {
3623 	uint16_t pkt_idx;
3624 	uint16_t slot_idx = 0;
3625 	uint16_t nr_done_pkts = 0;
3626 	uint16_t pkt_err = 0;
3627 	uint32_t n_xfer;
3628 	struct vhost_async *async = vq->async;
3629 	struct async_inflight_info *pkts_info = async->pkts_info;
3630 	struct rte_mbuf *pkts_prealloc[MAX_PKT_BURST];
3631 
3632 	VHOST_LOG_DATA(dev->ifname, DEBUG, "(%d) about to dequeue %u buffers\n", dev->vid, count);
3633 
3634 	async_iter_reset(async);
3635 
3636 	if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts_prealloc, count))
3637 		goto out;
3638 
3639 	for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
3640 		struct rte_mbuf *pkt = pkts_prealloc[pkt_idx];
3641 
3642 		rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
3643 
3644 		slot_idx = (async->pkts_idx + pkt_idx) % vq->size;
3645 		if (unlikely(virtio_dev_tx_async_single_packed(dev, vq, mbuf_pool, pkt,
3646 				slot_idx, legacy_ol_flags))) {
3647 			rte_pktmbuf_free_bulk(&pkts_prealloc[pkt_idx], count - pkt_idx);
3648 			break;
3649 		}
3650 
3651 		pkts_info[slot_idx].mbuf = pkt;
3652 
3653 		vq_inc_last_avail_packed(vq, 1);
3654 
3655 	}
3656 
3657 	n_xfer = vhost_async_dma_transfer(dev, vq, dma_id, vchan_id, async->pkts_idx,
3658 					async->iov_iter, pkt_idx);
3659 
3660 	async->pkts_inflight_n += n_xfer;
3661 
3662 	pkt_err = pkt_idx - n_xfer;
3663 
3664 	if (unlikely(pkt_err)) {
3665 		pkt_idx -= pkt_err;
3666 
3667 		/**
3668 		 * recover DMA-copy related structures and free pktmbuf for DMA-error pkts.
3669 		 */
3670 		if (async->buffer_idx_packed >= pkt_err)
3671 			async->buffer_idx_packed -= pkt_err;
3672 		else
3673 			async->buffer_idx_packed += vq->size - pkt_err;
3674 
3675 		while (pkt_err-- > 0) {
3676 			rte_pktmbuf_free(pkts_info[slot_idx % vq->size].mbuf);
3677 			slot_idx--;
3678 		}
3679 
3680 		/* recover available ring */
3681 		if (vq->last_avail_idx >= pkt_err) {
3682 			vq->last_avail_idx -= pkt_err;
3683 		} else {
3684 			vq->last_avail_idx += vq->size - pkt_err;
3685 			vq->avail_wrap_counter ^= 1;
3686 		}
3687 	}
3688 
3689 	async->pkts_idx += pkt_idx;
3690 	if (async->pkts_idx >= vq->size)
3691 		async->pkts_idx -= vq->size;
3692 
3693 out:
3694 	nr_done_pkts = async_poll_dequeue_completed(dev, vq, pkts, count,
3695 					dma_id, vchan_id, legacy_ol_flags);
3696 
3697 	return nr_done_pkts;
3698 }
3699 
3700 __rte_noinline
3701 static uint16_t
3702 virtio_dev_tx_async_packed_legacy(struct virtio_net *dev, struct vhost_virtqueue *vq,
3703 		struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts,
3704 		uint16_t count, uint16_t dma_id, uint16_t vchan_id)
3705 {
3706 	return virtio_dev_tx_async_packed(dev, vq, mbuf_pool,
3707 				pkts, count, dma_id, vchan_id, true);
3708 }
3709 
3710 __rte_noinline
3711 static uint16_t
3712 virtio_dev_tx_async_packed_compliant(struct virtio_net *dev, struct vhost_virtqueue *vq,
3713 		struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts,
3714 		uint16_t count, uint16_t dma_id, uint16_t vchan_id)
3715 {
3716 	return virtio_dev_tx_async_packed(dev, vq, mbuf_pool,
3717 				pkts, count, dma_id, vchan_id, false);
3718 }
3719 
3720 uint16_t
3721 rte_vhost_async_try_dequeue_burst(int vid, uint16_t queue_id,
3722 	struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count,
3723 	int *nr_inflight, int16_t dma_id, uint16_t vchan_id)
3724 {
3725 	struct virtio_net *dev;
3726 	struct rte_mbuf *rarp_mbuf = NULL;
3727 	struct vhost_virtqueue *vq;
3728 	int16_t success = 1;
3729 
3730 	dev = get_device(vid);
3731 	if (!dev || !nr_inflight)
3732 		return 0;
3733 
3734 	*nr_inflight = -1;
3735 
3736 	if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
3737 		VHOST_LOG_DATA(dev->ifname, ERR, "%s: built-in vhost net backend is disabled.\n",
3738 			__func__);
3739 		return 0;
3740 	}
3741 
3742 	if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
3743 		VHOST_LOG_DATA(dev->ifname, ERR, "%s: invalid virtqueue idx %d.\n",
3744 			__func__, queue_id);
3745 		return 0;
3746 	}
3747 
3748 	if (unlikely(dma_id < 0 || dma_id >= RTE_DMADEV_DEFAULT_MAX)) {
3749 		VHOST_LOG_DATA(dev->ifname, ERR, "%s: invalid dma id %d.\n",
3750 			__func__, dma_id);
3751 		return 0;
3752 	}
3753 
3754 	if (unlikely(!dma_copy_track[dma_id].vchans ||
3755 				!dma_copy_track[dma_id].vchans[vchan_id].pkts_cmpl_flag_addr)) {
3756 		VHOST_LOG_DATA(dev->ifname, ERR, "%s: invalid channel %d:%u.\n",
3757 			__func__, dma_id, vchan_id);
3758 		return 0;
3759 	}
3760 
3761 	vq = dev->virtqueue[queue_id];
3762 
3763 	if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
3764 		return 0;
3765 
3766 	if (unlikely(vq->enabled == 0)) {
3767 		count = 0;
3768 		goto out_access_unlock;
3769 	}
3770 
3771 	if (unlikely(!vq->async)) {
3772 		VHOST_LOG_DATA(dev->ifname, ERR, "%s: async not registered for queue id %d.\n",
3773 			__func__, queue_id);
3774 		count = 0;
3775 		goto out_access_unlock;
3776 	}
3777 
3778 	if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
3779 		vhost_user_iotlb_rd_lock(vq);
3780 
3781 	if (unlikely(vq->access_ok == 0))
3782 		if (unlikely(vring_translate(dev, vq) < 0)) {
3783 			count = 0;
3784 			goto out;
3785 		}
3786 
3787 	/*
3788 	 * Construct a RARP broadcast packet, and inject it to the "pkts"
3789 	 * array, to looks like that guest actually send such packet.
3790 	 *
3791 	 * Check user_send_rarp() for more information.
3792 	 *
3793 	 * broadcast_rarp shares a cacheline in the virtio_net structure
3794 	 * with some fields that are accessed during enqueue and
3795 	 * __atomic_compare_exchange_n causes a write if performed compare
3796 	 * and exchange. This could result in false sharing between enqueue
3797 	 * and dequeue.
3798 	 *
3799 	 * Prevent unnecessary false sharing by reading broadcast_rarp first
3800 	 * and only performing compare and exchange if the read indicates it
3801 	 * is likely to be set.
3802 	 */
3803 	if (unlikely(__atomic_load_n(&dev->broadcast_rarp, __ATOMIC_ACQUIRE) &&
3804 			__atomic_compare_exchange_n(&dev->broadcast_rarp,
3805 			&success, 0, 0, __ATOMIC_RELEASE, __ATOMIC_RELAXED))) {
3806 
3807 		rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
3808 		if (rarp_mbuf == NULL) {
3809 			VHOST_LOG_DATA(dev->ifname, ERR, "failed to make RARP packet.\n");
3810 			count = 0;
3811 			goto out;
3812 		}
3813 		/*
3814 		 * Inject it to the head of "pkts" array, so that switch's mac
3815 		 * learning table will get updated first.
3816 		 */
3817 		pkts[0] = rarp_mbuf;
3818 		vhost_queue_stats_update(dev, vq, pkts, 1);
3819 		pkts++;
3820 		count -= 1;
3821 	}
3822 
3823 	if (vq_is_packed(dev)) {
3824 		if (dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS)
3825 			count = virtio_dev_tx_async_packed_legacy(dev, vq, mbuf_pool,
3826 					pkts, count, dma_id, vchan_id);
3827 		else
3828 			count = virtio_dev_tx_async_packed_compliant(dev, vq, mbuf_pool,
3829 					pkts, count, dma_id, vchan_id);
3830 	} else {
3831 		if (dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS)
3832 			count = virtio_dev_tx_async_split_legacy(dev, vq, mbuf_pool,
3833 					pkts, count, dma_id, vchan_id);
3834 		else
3835 			count = virtio_dev_tx_async_split_compliant(dev, vq, mbuf_pool,
3836 					pkts, count, dma_id, vchan_id);
3837 	}
3838 
3839 	*nr_inflight = vq->async->pkts_inflight_n;
3840 	vhost_queue_stats_update(dev, vq, pkts, count);
3841 
3842 out:
3843 	if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
3844 		vhost_user_iotlb_rd_unlock(vq);
3845 
3846 out_access_unlock:
3847 	rte_spinlock_unlock(&vq->access_lock);
3848 
3849 	if (unlikely(rarp_mbuf != NULL))
3850 		count += 1;
3851 
3852 	return count;
3853 }
3854