xref: /dpdk/lib/vhost/vhost.c (revision c9902a15bd005b6d4fe072cf7b60fe4ee679155f)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2010-2017 Intel Corporation
3  */
4 
5 #include <linux/vhost.h>
6 #include <linux/virtio_net.h>
7 #include <stddef.h>
8 #include <stdint.h>
9 #include <stdlib.h>
10 #ifdef RTE_LIBRTE_VHOST_NUMA
11 #include <numa.h>
12 #include <numaif.h>
13 #endif
14 
15 #include <rte_errno.h>
16 #include <rte_ethdev.h>
17 #include <rte_log.h>
18 #include <rte_string_fns.h>
19 #include <rte_memory.h>
20 #include <rte_malloc.h>
21 #include <rte_vhost.h>
22 #include <rte_rwlock.h>
23 
24 #include "iotlb.h"
25 #include "vhost.h"
26 #include "vhost_user.h"
27 
28 struct virtio_net *vhost_devices[MAX_VHOST_DEVICE];
29 pthread_mutex_t vhost_dev_lock = PTHREAD_MUTEX_INITIALIZER;
30 
31 /* Called with iotlb_lock read-locked */
32 uint64_t
33 __vhost_iova_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
34 		    uint64_t iova, uint64_t *size, uint8_t perm)
35 {
36 	uint64_t vva, tmp_size;
37 
38 	if (unlikely(!*size))
39 		return 0;
40 
41 	tmp_size = *size;
42 
43 	vva = vhost_user_iotlb_cache_find(vq, iova, &tmp_size, perm);
44 	if (tmp_size == *size)
45 		return vva;
46 
47 	iova += tmp_size;
48 
49 	if (!vhost_user_iotlb_pending_miss(vq, iova, perm)) {
50 		/*
51 		 * iotlb_lock is read-locked for a full burst,
52 		 * but it only protects the iotlb cache.
53 		 * In case of IOTLB miss, we might block on the socket,
54 		 * which could cause a deadlock with QEMU if an IOTLB update
55 		 * is being handled. We can safely unlock here to avoid it.
56 		 */
57 		vhost_user_iotlb_rd_unlock(vq);
58 
59 		vhost_user_iotlb_pending_insert(vq, iova, perm);
60 		if (vhost_user_iotlb_miss(dev, iova, perm)) {
61 			VHOST_LOG_CONFIG(ERR,
62 				"IOTLB miss req failed for IOVA 0x%" PRIx64 "\n",
63 				iova);
64 			vhost_user_iotlb_pending_remove(vq, iova, 1, perm);
65 		}
66 
67 		vhost_user_iotlb_rd_lock(vq);
68 	}
69 
70 	return 0;
71 }
72 
73 #define VHOST_LOG_PAGE	4096
74 
75 /*
76  * Atomically set a bit in memory.
77  */
78 static __rte_always_inline void
79 vhost_set_bit(unsigned int nr, volatile uint8_t *addr)
80 {
81 #if defined(RTE_TOOLCHAIN_GCC) && (GCC_VERSION < 70100)
82 	/*
83 	 * __sync_ built-ins are deprecated, but __atomic_ ones
84 	 * are sub-optimized in older GCC versions.
85 	 */
86 	__sync_fetch_and_or_1(addr, (1U << nr));
87 #else
88 	__atomic_fetch_or(addr, (1U << nr), __ATOMIC_RELAXED);
89 #endif
90 }
91 
92 static __rte_always_inline void
93 vhost_log_page(uint8_t *log_base, uint64_t page)
94 {
95 	vhost_set_bit(page % 8, &log_base[page / 8]);
96 }
97 
98 void
99 __vhost_log_write(struct virtio_net *dev, uint64_t addr, uint64_t len)
100 {
101 	uint64_t page;
102 
103 	if (unlikely(!dev->log_base || !len))
104 		return;
105 
106 	if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
107 		return;
108 
109 	/* To make sure guest memory updates are committed before logging */
110 	rte_atomic_thread_fence(__ATOMIC_RELEASE);
111 
112 	page = addr / VHOST_LOG_PAGE;
113 	while (page * VHOST_LOG_PAGE < addr + len) {
114 		vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
115 		page += 1;
116 	}
117 }
118 
119 void
120 __vhost_log_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
121 			     uint64_t iova, uint64_t len)
122 {
123 	uint64_t hva, gpa, map_len;
124 	map_len = len;
125 
126 	hva = __vhost_iova_to_vva(dev, vq, iova, &map_len, VHOST_ACCESS_RW);
127 	if (map_len != len) {
128 		VHOST_LOG_DATA(ERR,
129 			"Failed to write log for IOVA 0x%" PRIx64 ". No IOTLB entry found\n",
130 			iova);
131 		return;
132 	}
133 
134 	gpa = hva_to_gpa(dev, hva, len);
135 	if (gpa)
136 		__vhost_log_write(dev, gpa, len);
137 }
138 
139 void
140 __vhost_log_cache_sync(struct virtio_net *dev, struct vhost_virtqueue *vq)
141 {
142 	unsigned long *log_base;
143 	int i;
144 
145 	if (unlikely(!dev->log_base))
146 		return;
147 
148 	/* No cache, nothing to sync */
149 	if (unlikely(!vq->log_cache))
150 		return;
151 
152 	rte_atomic_thread_fence(__ATOMIC_RELEASE);
153 
154 	log_base = (unsigned long *)(uintptr_t)dev->log_base;
155 
156 	for (i = 0; i < vq->log_cache_nb_elem; i++) {
157 		struct log_cache_entry *elem = vq->log_cache + i;
158 
159 #if defined(RTE_TOOLCHAIN_GCC) && (GCC_VERSION < 70100)
160 		/*
161 		 * '__sync' builtins are deprecated, but '__atomic' ones
162 		 * are sub-optimized in older GCC versions.
163 		 */
164 		__sync_fetch_and_or(log_base + elem->offset, elem->val);
165 #else
166 		__atomic_fetch_or(log_base + elem->offset, elem->val,
167 				__ATOMIC_RELAXED);
168 #endif
169 	}
170 
171 	rte_atomic_thread_fence(__ATOMIC_RELEASE);
172 
173 	vq->log_cache_nb_elem = 0;
174 }
175 
176 static __rte_always_inline void
177 vhost_log_cache_page(struct virtio_net *dev, struct vhost_virtqueue *vq,
178 			uint64_t page)
179 {
180 	uint32_t bit_nr = page % (sizeof(unsigned long) << 3);
181 	uint32_t offset = page / (sizeof(unsigned long) << 3);
182 	int i;
183 
184 	if (unlikely(!vq->log_cache)) {
185 		/* No logging cache allocated, write dirty log map directly */
186 		rte_atomic_thread_fence(__ATOMIC_RELEASE);
187 		vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
188 
189 		return;
190 	}
191 
192 	for (i = 0; i < vq->log_cache_nb_elem; i++) {
193 		struct log_cache_entry *elem = vq->log_cache + i;
194 
195 		if (elem->offset == offset) {
196 			elem->val |= (1UL << bit_nr);
197 			return;
198 		}
199 	}
200 
201 	if (unlikely(i >= VHOST_LOG_CACHE_NR)) {
202 		/*
203 		 * No more room for a new log cache entry,
204 		 * so write the dirty log map directly.
205 		 */
206 		rte_atomic_thread_fence(__ATOMIC_RELEASE);
207 		vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
208 
209 		return;
210 	}
211 
212 	vq->log_cache[i].offset = offset;
213 	vq->log_cache[i].val = (1UL << bit_nr);
214 	vq->log_cache_nb_elem++;
215 }
216 
217 void
218 __vhost_log_cache_write(struct virtio_net *dev, struct vhost_virtqueue *vq,
219 			uint64_t addr, uint64_t len)
220 {
221 	uint64_t page;
222 
223 	if (unlikely(!dev->log_base || !len))
224 		return;
225 
226 	if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
227 		return;
228 
229 	page = addr / VHOST_LOG_PAGE;
230 	while (page * VHOST_LOG_PAGE < addr + len) {
231 		vhost_log_cache_page(dev, vq, page);
232 		page += 1;
233 	}
234 }
235 
236 void
237 __vhost_log_cache_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
238 			     uint64_t iova, uint64_t len)
239 {
240 	uint64_t hva, gpa, map_len;
241 	map_len = len;
242 
243 	hva = __vhost_iova_to_vva(dev, vq, iova, &map_len, VHOST_ACCESS_RW);
244 	if (map_len != len) {
245 		VHOST_LOG_DATA(ERR,
246 			"Failed to write log for IOVA 0x%" PRIx64 ". No IOTLB entry found\n",
247 			iova);
248 		return;
249 	}
250 
251 	gpa = hva_to_gpa(dev, hva, len);
252 	if (gpa)
253 		__vhost_log_cache_write(dev, vq, gpa, len);
254 }
255 
256 void *
257 vhost_alloc_copy_ind_table(struct virtio_net *dev, struct vhost_virtqueue *vq,
258 		uint64_t desc_addr, uint64_t desc_len)
259 {
260 	void *idesc;
261 	uint64_t src, dst;
262 	uint64_t len, remain = desc_len;
263 
264 	idesc = rte_malloc_socket(__func__, desc_len, 0, vq->numa_node);
265 	if (unlikely(!idesc))
266 		return NULL;
267 
268 	dst = (uint64_t)(uintptr_t)idesc;
269 
270 	while (remain) {
271 		len = remain;
272 		src = vhost_iova_to_vva(dev, vq, desc_addr, &len,
273 				VHOST_ACCESS_RO);
274 		if (unlikely(!src || !len)) {
275 			rte_free(idesc);
276 			return NULL;
277 		}
278 
279 		rte_memcpy((void *)(uintptr_t)dst, (void *)(uintptr_t)src, len);
280 
281 		remain -= len;
282 		dst += len;
283 		desc_addr += len;
284 	}
285 
286 	return idesc;
287 }
288 
289 void
290 cleanup_vq(struct vhost_virtqueue *vq, int destroy)
291 {
292 	if ((vq->callfd >= 0) && (destroy != 0))
293 		close(vq->callfd);
294 	if (vq->kickfd >= 0)
295 		close(vq->kickfd);
296 }
297 
298 void
299 cleanup_vq_inflight(struct virtio_net *dev, struct vhost_virtqueue *vq)
300 {
301 	if (!(dev->protocol_features &
302 	    (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
303 		return;
304 
305 	if (vq_is_packed(dev)) {
306 		if (vq->inflight_packed)
307 			vq->inflight_packed = NULL;
308 	} else {
309 		if (vq->inflight_split)
310 			vq->inflight_split = NULL;
311 	}
312 
313 	if (vq->resubmit_inflight) {
314 		if (vq->resubmit_inflight->resubmit_list) {
315 			rte_free(vq->resubmit_inflight->resubmit_list);
316 			vq->resubmit_inflight->resubmit_list = NULL;
317 		}
318 		rte_free(vq->resubmit_inflight);
319 		vq->resubmit_inflight = NULL;
320 	}
321 }
322 
323 /*
324  * Unmap any memory, close any file descriptors and
325  * free any memory owned by a device.
326  */
327 void
328 cleanup_device(struct virtio_net *dev, int destroy)
329 {
330 	uint32_t i;
331 
332 	vhost_backend_cleanup(dev);
333 
334 	for (i = 0; i < dev->nr_vring; i++) {
335 		cleanup_vq(dev->virtqueue[i], destroy);
336 		cleanup_vq_inflight(dev, dev->virtqueue[i]);
337 	}
338 }
339 
340 static void
341 vhost_free_async_mem(struct vhost_virtqueue *vq)
342 {
343 	rte_free(vq->async_pkts_info);
344 
345 	rte_free(vq->async_buffers_packed);
346 	vq->async_buffers_packed = NULL;
347 	rte_free(vq->async_descs_split);
348 	vq->async_descs_split = NULL;
349 
350 	rte_free(vq->it_pool);
351 	rte_free(vq->vec_pool);
352 
353 	vq->async_pkts_info = NULL;
354 	vq->it_pool = NULL;
355 	vq->vec_pool = NULL;
356 }
357 
358 void
359 free_vq(struct virtio_net *dev, struct vhost_virtqueue *vq)
360 {
361 	if (vq_is_packed(dev))
362 		rte_free(vq->shadow_used_packed);
363 	else
364 		rte_free(vq->shadow_used_split);
365 
366 	vhost_free_async_mem(vq);
367 	rte_free(vq->batch_copy_elems);
368 	rte_mempool_free(vq->iotlb_pool);
369 	rte_free(vq->log_cache);
370 	rte_free(vq);
371 }
372 
373 /*
374  * Release virtqueues and device memory.
375  */
376 static void
377 free_device(struct virtio_net *dev)
378 {
379 	uint32_t i;
380 
381 	for (i = 0; i < dev->nr_vring; i++)
382 		free_vq(dev, dev->virtqueue[i]);
383 
384 	rte_free(dev);
385 }
386 
387 static __rte_always_inline int
388 log_translate(struct virtio_net *dev, struct vhost_virtqueue *vq)
389 {
390 	if (likely(!(vq->ring_addrs.flags & (1 << VHOST_VRING_F_LOG))))
391 		return 0;
392 
393 	vq->log_guest_addr = translate_log_addr(dev, vq,
394 						vq->ring_addrs.log_guest_addr);
395 	if (vq->log_guest_addr == 0)
396 		return -1;
397 
398 	return 0;
399 }
400 
401 /*
402  * Converts vring log address to GPA
403  * If IOMMU is enabled, the log address is IOVA
404  * If IOMMU not enabled, the log address is already GPA
405  *
406  * Caller should have iotlb_lock read-locked
407  */
408 uint64_t
409 translate_log_addr(struct virtio_net *dev, struct vhost_virtqueue *vq,
410 		uint64_t log_addr)
411 {
412 	if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
413 		const uint64_t exp_size = sizeof(uint64_t);
414 		uint64_t hva, gpa;
415 		uint64_t size = exp_size;
416 
417 		hva = vhost_iova_to_vva(dev, vq, log_addr,
418 					&size, VHOST_ACCESS_RW);
419 
420 		if (size != exp_size)
421 			return 0;
422 
423 		gpa = hva_to_gpa(dev, hva, exp_size);
424 		if (!gpa) {
425 			VHOST_LOG_CONFIG(ERR,
426 				"VQ: Failed to find GPA for log_addr: 0x%"
427 				PRIx64 " hva: 0x%" PRIx64 "\n",
428 				log_addr, hva);
429 			return 0;
430 		}
431 		return gpa;
432 
433 	} else
434 		return log_addr;
435 }
436 
437 /* Caller should have iotlb_lock read-locked */
438 static int
439 vring_translate_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
440 {
441 	uint64_t req_size, size;
442 
443 	req_size = sizeof(struct vring_desc) * vq->size;
444 	size = req_size;
445 	vq->desc = (struct vring_desc *)(uintptr_t)vhost_iova_to_vva(dev, vq,
446 						vq->ring_addrs.desc_user_addr,
447 						&size, VHOST_ACCESS_RW);
448 	if (!vq->desc || size != req_size)
449 		return -1;
450 
451 	req_size = sizeof(struct vring_avail);
452 	req_size += sizeof(uint16_t) * vq->size;
453 	if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
454 		req_size += sizeof(uint16_t);
455 	size = req_size;
456 	vq->avail = (struct vring_avail *)(uintptr_t)vhost_iova_to_vva(dev, vq,
457 						vq->ring_addrs.avail_user_addr,
458 						&size, VHOST_ACCESS_RW);
459 	if (!vq->avail || size != req_size)
460 		return -1;
461 
462 	req_size = sizeof(struct vring_used);
463 	req_size += sizeof(struct vring_used_elem) * vq->size;
464 	if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
465 		req_size += sizeof(uint16_t);
466 	size = req_size;
467 	vq->used = (struct vring_used *)(uintptr_t)vhost_iova_to_vva(dev, vq,
468 						vq->ring_addrs.used_user_addr,
469 						&size, VHOST_ACCESS_RW);
470 	if (!vq->used || size != req_size)
471 		return -1;
472 
473 	return 0;
474 }
475 
476 /* Caller should have iotlb_lock read-locked */
477 static int
478 vring_translate_packed(struct virtio_net *dev, struct vhost_virtqueue *vq)
479 {
480 	uint64_t req_size, size;
481 
482 	req_size = sizeof(struct vring_packed_desc) * vq->size;
483 	size = req_size;
484 	vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
485 		vhost_iova_to_vva(dev, vq, vq->ring_addrs.desc_user_addr,
486 				&size, VHOST_ACCESS_RW);
487 	if (!vq->desc_packed || size != req_size)
488 		return -1;
489 
490 	req_size = sizeof(struct vring_packed_desc_event);
491 	size = req_size;
492 	vq->driver_event = (struct vring_packed_desc_event *)(uintptr_t)
493 		vhost_iova_to_vva(dev, vq, vq->ring_addrs.avail_user_addr,
494 				&size, VHOST_ACCESS_RW);
495 	if (!vq->driver_event || size != req_size)
496 		return -1;
497 
498 	req_size = sizeof(struct vring_packed_desc_event);
499 	size = req_size;
500 	vq->device_event = (struct vring_packed_desc_event *)(uintptr_t)
501 		vhost_iova_to_vva(dev, vq, vq->ring_addrs.used_user_addr,
502 				&size, VHOST_ACCESS_RW);
503 	if (!vq->device_event || size != req_size)
504 		return -1;
505 
506 	return 0;
507 }
508 
509 int
510 vring_translate(struct virtio_net *dev, struct vhost_virtqueue *vq)
511 {
512 
513 	if (!(dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
514 		return -1;
515 
516 	if (vq_is_packed(dev)) {
517 		if (vring_translate_packed(dev, vq) < 0)
518 			return -1;
519 	} else {
520 		if (vring_translate_split(dev, vq) < 0)
521 			return -1;
522 	}
523 
524 	if (log_translate(dev, vq) < 0)
525 		return -1;
526 
527 	vq->access_ok = true;
528 
529 	return 0;
530 }
531 
532 void
533 vring_invalidate(struct virtio_net *dev, struct vhost_virtqueue *vq)
534 {
535 	if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
536 		vhost_user_iotlb_wr_lock(vq);
537 
538 	vq->access_ok = false;
539 	vq->desc = NULL;
540 	vq->avail = NULL;
541 	vq->used = NULL;
542 	vq->log_guest_addr = 0;
543 
544 	if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
545 		vhost_user_iotlb_wr_unlock(vq);
546 }
547 
548 static void
549 init_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
550 {
551 	struct vhost_virtqueue *vq;
552 	int numa_node = SOCKET_ID_ANY;
553 
554 	if (vring_idx >= VHOST_MAX_VRING) {
555 		VHOST_LOG_CONFIG(ERR,
556 				"Failed not init vring, out of bound (%d)\n",
557 				vring_idx);
558 		return;
559 	}
560 
561 	vq = dev->virtqueue[vring_idx];
562 	if (!vq) {
563 		VHOST_LOG_CONFIG(ERR, "Virtqueue not allocated (%d)\n",
564 				vring_idx);
565 		return;
566 	}
567 
568 	memset(vq, 0, sizeof(struct vhost_virtqueue));
569 
570 	vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
571 	vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
572 	vq->notif_enable = VIRTIO_UNINITIALIZED_NOTIF;
573 
574 #ifdef RTE_LIBRTE_VHOST_NUMA
575 	if (get_mempolicy(&numa_node, NULL, 0, vq, MPOL_F_NODE | MPOL_F_ADDR)) {
576 		VHOST_LOG_CONFIG(ERR, "(%d) failed to query numa node: %s\n",
577 			dev->vid, rte_strerror(errno));
578 		numa_node = SOCKET_ID_ANY;
579 	}
580 #endif
581 	vq->numa_node = numa_node;
582 
583 	vhost_user_iotlb_init(dev, vring_idx);
584 }
585 
586 static void
587 reset_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
588 {
589 	struct vhost_virtqueue *vq;
590 	int callfd;
591 
592 	if (vring_idx >= VHOST_MAX_VRING) {
593 		VHOST_LOG_CONFIG(ERR,
594 				"Failed not init vring, out of bound (%d)\n",
595 				vring_idx);
596 		return;
597 	}
598 
599 	vq = dev->virtqueue[vring_idx];
600 	if (!vq) {
601 		VHOST_LOG_CONFIG(ERR, "Virtqueue not allocated (%d)\n",
602 				vring_idx);
603 		return;
604 	}
605 
606 	callfd = vq->callfd;
607 	init_vring_queue(dev, vring_idx);
608 	vq->callfd = callfd;
609 }
610 
611 int
612 alloc_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
613 {
614 	struct vhost_virtqueue *vq;
615 	uint32_t i;
616 
617 	/* Also allocate holes, if any, up to requested vring index. */
618 	for (i = 0; i <= vring_idx; i++) {
619 		if (dev->virtqueue[i])
620 			continue;
621 
622 		vq = rte_zmalloc(NULL, sizeof(struct vhost_virtqueue), 0);
623 		if (vq == NULL) {
624 			VHOST_LOG_CONFIG(ERR,
625 				"Failed to allocate memory for vring:%u.\n", i);
626 			return -1;
627 		}
628 
629 		dev->virtqueue[i] = vq;
630 		init_vring_queue(dev, i);
631 		rte_spinlock_init(&vq->access_lock);
632 		vq->avail_wrap_counter = 1;
633 		vq->used_wrap_counter = 1;
634 		vq->signalled_used_valid = false;
635 	}
636 
637 	dev->nr_vring = RTE_MAX(dev->nr_vring, vring_idx + 1);
638 
639 	return 0;
640 }
641 
642 /*
643  * Reset some variables in device structure, while keeping few
644  * others untouched, such as vid, ifname, nr_vring: they
645  * should be same unless the device is removed.
646  */
647 void
648 reset_device(struct virtio_net *dev)
649 {
650 	uint32_t i;
651 
652 	dev->features = 0;
653 	dev->protocol_features = 0;
654 	dev->flags &= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
655 
656 	for (i = 0; i < dev->nr_vring; i++)
657 		reset_vring_queue(dev, i);
658 }
659 
660 /*
661  * Invoked when there is a new vhost-user connection established (when
662  * there is a new virtio device being attached).
663  */
664 int
665 vhost_new_device(void)
666 {
667 	struct virtio_net *dev;
668 	int i;
669 
670 	pthread_mutex_lock(&vhost_dev_lock);
671 	for (i = 0; i < MAX_VHOST_DEVICE; i++) {
672 		if (vhost_devices[i] == NULL)
673 			break;
674 	}
675 
676 	if (i == MAX_VHOST_DEVICE) {
677 		VHOST_LOG_CONFIG(ERR,
678 			"Failed to find a free slot for new device.\n");
679 		pthread_mutex_unlock(&vhost_dev_lock);
680 		return -1;
681 	}
682 
683 	dev = rte_zmalloc(NULL, sizeof(struct virtio_net), 0);
684 	if (dev == NULL) {
685 		VHOST_LOG_CONFIG(ERR,
686 			"Failed to allocate memory for new dev.\n");
687 		pthread_mutex_unlock(&vhost_dev_lock);
688 		return -1;
689 	}
690 
691 	vhost_devices[i] = dev;
692 	pthread_mutex_unlock(&vhost_dev_lock);
693 
694 	dev->vid = i;
695 	dev->flags = VIRTIO_DEV_BUILTIN_VIRTIO_NET;
696 	dev->slave_req_fd = -1;
697 	dev->postcopy_ufd = -1;
698 	rte_spinlock_init(&dev->slave_req_lock);
699 
700 	return i;
701 }
702 
703 void
704 vhost_destroy_device_notify(struct virtio_net *dev)
705 {
706 	struct rte_vdpa_device *vdpa_dev;
707 
708 	if (dev->flags & VIRTIO_DEV_RUNNING) {
709 		vdpa_dev = dev->vdpa_dev;
710 		if (vdpa_dev)
711 			vdpa_dev->ops->dev_close(dev->vid);
712 		dev->flags &= ~VIRTIO_DEV_RUNNING;
713 		dev->notify_ops->destroy_device(dev->vid);
714 	}
715 }
716 
717 /*
718  * Invoked when there is the vhost-user connection is broken (when
719  * the virtio device is being detached).
720  */
721 void
722 vhost_destroy_device(int vid)
723 {
724 	struct virtio_net *dev = get_device(vid);
725 
726 	if (dev == NULL)
727 		return;
728 
729 	vhost_destroy_device_notify(dev);
730 
731 	cleanup_device(dev, 1);
732 	free_device(dev);
733 
734 	vhost_devices[vid] = NULL;
735 }
736 
737 void
738 vhost_attach_vdpa_device(int vid, struct rte_vdpa_device *vdpa_dev)
739 {
740 	struct virtio_net *dev = get_device(vid);
741 
742 	if (dev == NULL)
743 		return;
744 
745 	dev->vdpa_dev = vdpa_dev;
746 }
747 
748 void
749 vhost_set_ifname(int vid, const char *if_name, unsigned int if_len)
750 {
751 	struct virtio_net *dev;
752 	unsigned int len;
753 
754 	dev = get_device(vid);
755 	if (dev == NULL)
756 		return;
757 
758 	len = if_len > sizeof(dev->ifname) ?
759 		sizeof(dev->ifname) : if_len;
760 
761 	strncpy(dev->ifname, if_name, len);
762 	dev->ifname[sizeof(dev->ifname) - 1] = '\0';
763 }
764 
765 void
766 vhost_setup_virtio_net(int vid, bool enable, bool compliant_ol_flags)
767 {
768 	struct virtio_net *dev = get_device(vid);
769 
770 	if (dev == NULL)
771 		return;
772 
773 	if (enable)
774 		dev->flags |= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
775 	else
776 		dev->flags &= ~VIRTIO_DEV_BUILTIN_VIRTIO_NET;
777 	if (!compliant_ol_flags)
778 		dev->flags |= VIRTIO_DEV_LEGACY_OL_FLAGS;
779 	else
780 		dev->flags &= ~VIRTIO_DEV_LEGACY_OL_FLAGS;
781 }
782 
783 void
784 vhost_enable_extbuf(int vid)
785 {
786 	struct virtio_net *dev = get_device(vid);
787 
788 	if (dev == NULL)
789 		return;
790 
791 	dev->extbuf = 1;
792 }
793 
794 void
795 vhost_enable_linearbuf(int vid)
796 {
797 	struct virtio_net *dev = get_device(vid);
798 
799 	if (dev == NULL)
800 		return;
801 
802 	dev->linearbuf = 1;
803 }
804 
805 int
806 rte_vhost_get_mtu(int vid, uint16_t *mtu)
807 {
808 	struct virtio_net *dev = get_device(vid);
809 
810 	if (dev == NULL || mtu == NULL)
811 		return -ENODEV;
812 
813 	if (!(dev->flags & VIRTIO_DEV_READY))
814 		return -EAGAIN;
815 
816 	if (!(dev->features & (1ULL << VIRTIO_NET_F_MTU)))
817 		return -ENOTSUP;
818 
819 	*mtu = dev->mtu;
820 
821 	return 0;
822 }
823 
824 int
825 rte_vhost_get_numa_node(int vid)
826 {
827 #ifdef RTE_LIBRTE_VHOST_NUMA
828 	struct virtio_net *dev = get_device(vid);
829 	int numa_node;
830 	int ret;
831 
832 	if (dev == NULL || numa_available() != 0)
833 		return -1;
834 
835 	ret = get_mempolicy(&numa_node, NULL, 0, dev,
836 			    MPOL_F_NODE | MPOL_F_ADDR);
837 	if (ret < 0) {
838 		VHOST_LOG_CONFIG(ERR,
839 			"(%d) failed to query numa node: %s\n",
840 			vid, rte_strerror(errno));
841 		return -1;
842 	}
843 
844 	return numa_node;
845 #else
846 	RTE_SET_USED(vid);
847 	return -1;
848 #endif
849 }
850 
851 uint32_t
852 rte_vhost_get_queue_num(int vid)
853 {
854 	struct virtio_net *dev = get_device(vid);
855 
856 	if (dev == NULL)
857 		return 0;
858 
859 	return dev->nr_vring / 2;
860 }
861 
862 uint16_t
863 rte_vhost_get_vring_num(int vid)
864 {
865 	struct virtio_net *dev = get_device(vid);
866 
867 	if (dev == NULL)
868 		return 0;
869 
870 	return dev->nr_vring;
871 }
872 
873 int
874 rte_vhost_get_ifname(int vid, char *buf, size_t len)
875 {
876 	struct virtio_net *dev = get_device(vid);
877 
878 	if (dev == NULL || buf == NULL)
879 		return -1;
880 
881 	len = RTE_MIN(len, sizeof(dev->ifname));
882 
883 	strncpy(buf, dev->ifname, len);
884 	buf[len - 1] = '\0';
885 
886 	return 0;
887 }
888 
889 int
890 rte_vhost_get_negotiated_features(int vid, uint64_t *features)
891 {
892 	struct virtio_net *dev;
893 
894 	dev = get_device(vid);
895 	if (dev == NULL || features == NULL)
896 		return -1;
897 
898 	*features = dev->features;
899 	return 0;
900 }
901 
902 int
903 rte_vhost_get_negotiated_protocol_features(int vid,
904 					   uint64_t *protocol_features)
905 {
906 	struct virtio_net *dev;
907 
908 	dev = get_device(vid);
909 	if (dev == NULL || protocol_features == NULL)
910 		return -1;
911 
912 	*protocol_features = dev->protocol_features;
913 	return 0;
914 }
915 
916 int
917 rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem)
918 {
919 	struct virtio_net *dev;
920 	struct rte_vhost_memory *m;
921 	size_t size;
922 
923 	dev = get_device(vid);
924 	if (dev == NULL || mem == NULL)
925 		return -1;
926 
927 	size = dev->mem->nregions * sizeof(struct rte_vhost_mem_region);
928 	m = malloc(sizeof(struct rte_vhost_memory) + size);
929 	if (!m)
930 		return -1;
931 
932 	m->nregions = dev->mem->nregions;
933 	memcpy(m->regions, dev->mem->regions, size);
934 	*mem = m;
935 
936 	return 0;
937 }
938 
939 int
940 rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
941 			  struct rte_vhost_vring *vring)
942 {
943 	struct virtio_net *dev;
944 	struct vhost_virtqueue *vq;
945 
946 	dev = get_device(vid);
947 	if (dev == NULL || vring == NULL)
948 		return -1;
949 
950 	if (vring_idx >= VHOST_MAX_VRING)
951 		return -1;
952 
953 	vq = dev->virtqueue[vring_idx];
954 	if (!vq)
955 		return -1;
956 
957 	if (vq_is_packed(dev)) {
958 		vring->desc_packed = vq->desc_packed;
959 		vring->driver_event = vq->driver_event;
960 		vring->device_event = vq->device_event;
961 	} else {
962 		vring->desc = vq->desc;
963 		vring->avail = vq->avail;
964 		vring->used = vq->used;
965 	}
966 	vring->log_guest_addr  = vq->log_guest_addr;
967 
968 	vring->callfd  = vq->callfd;
969 	vring->kickfd  = vq->kickfd;
970 	vring->size    = vq->size;
971 
972 	return 0;
973 }
974 
975 int
976 rte_vhost_get_vhost_ring_inflight(int vid, uint16_t vring_idx,
977 				  struct rte_vhost_ring_inflight *vring)
978 {
979 	struct virtio_net *dev;
980 	struct vhost_virtqueue *vq;
981 
982 	dev = get_device(vid);
983 	if (unlikely(!dev))
984 		return -1;
985 
986 	if (vring_idx >= VHOST_MAX_VRING)
987 		return -1;
988 
989 	vq = dev->virtqueue[vring_idx];
990 	if (unlikely(!vq))
991 		return -1;
992 
993 	if (vq_is_packed(dev)) {
994 		if (unlikely(!vq->inflight_packed))
995 			return -1;
996 
997 		vring->inflight_packed = vq->inflight_packed;
998 	} else {
999 		if (unlikely(!vq->inflight_split))
1000 			return -1;
1001 
1002 		vring->inflight_split = vq->inflight_split;
1003 	}
1004 
1005 	vring->resubmit_inflight = vq->resubmit_inflight;
1006 
1007 	return 0;
1008 }
1009 
1010 int
1011 rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx,
1012 				  uint16_t idx)
1013 {
1014 	struct vhost_virtqueue *vq;
1015 	struct virtio_net *dev;
1016 
1017 	dev = get_device(vid);
1018 	if (unlikely(!dev))
1019 		return -1;
1020 
1021 	if (unlikely(!(dev->protocol_features &
1022 	    (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1023 		return 0;
1024 
1025 	if (unlikely(vq_is_packed(dev)))
1026 		return -1;
1027 
1028 	if (unlikely(vring_idx >= VHOST_MAX_VRING))
1029 		return -1;
1030 
1031 	vq = dev->virtqueue[vring_idx];
1032 	if (unlikely(!vq))
1033 		return -1;
1034 
1035 	if (unlikely(!vq->inflight_split))
1036 		return -1;
1037 
1038 	if (unlikely(idx >= vq->size))
1039 		return -1;
1040 
1041 	vq->inflight_split->desc[idx].counter = vq->global_counter++;
1042 	vq->inflight_split->desc[idx].inflight = 1;
1043 	return 0;
1044 }
1045 
1046 int
1047 rte_vhost_set_inflight_desc_packed(int vid, uint16_t vring_idx,
1048 				   uint16_t head, uint16_t last,
1049 				   uint16_t *inflight_entry)
1050 {
1051 	struct rte_vhost_inflight_info_packed *inflight_info;
1052 	struct virtio_net *dev;
1053 	struct vhost_virtqueue *vq;
1054 	struct vring_packed_desc *desc;
1055 	uint16_t old_free_head, free_head;
1056 
1057 	dev = get_device(vid);
1058 	if (unlikely(!dev))
1059 		return -1;
1060 
1061 	if (unlikely(!(dev->protocol_features &
1062 	    (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1063 		return 0;
1064 
1065 	if (unlikely(!vq_is_packed(dev)))
1066 		return -1;
1067 
1068 	if (unlikely(vring_idx >= VHOST_MAX_VRING))
1069 		return -1;
1070 
1071 	vq = dev->virtqueue[vring_idx];
1072 	if (unlikely(!vq))
1073 		return -1;
1074 
1075 	inflight_info = vq->inflight_packed;
1076 	if (unlikely(!inflight_info))
1077 		return -1;
1078 
1079 	if (unlikely(head >= vq->size))
1080 		return -1;
1081 
1082 	desc = vq->desc_packed;
1083 	old_free_head = inflight_info->old_free_head;
1084 	if (unlikely(old_free_head >= vq->size))
1085 		return -1;
1086 
1087 	free_head = old_free_head;
1088 
1089 	/* init header descriptor */
1090 	inflight_info->desc[old_free_head].num = 0;
1091 	inflight_info->desc[old_free_head].counter = vq->global_counter++;
1092 	inflight_info->desc[old_free_head].inflight = 1;
1093 
1094 	/* save desc entry in flight entry */
1095 	while (head != ((last + 1) % vq->size)) {
1096 		inflight_info->desc[old_free_head].num++;
1097 		inflight_info->desc[free_head].addr = desc[head].addr;
1098 		inflight_info->desc[free_head].len = desc[head].len;
1099 		inflight_info->desc[free_head].flags = desc[head].flags;
1100 		inflight_info->desc[free_head].id = desc[head].id;
1101 
1102 		inflight_info->desc[old_free_head].last = free_head;
1103 		free_head = inflight_info->desc[free_head].next;
1104 		inflight_info->free_head = free_head;
1105 		head = (head + 1) % vq->size;
1106 	}
1107 
1108 	inflight_info->old_free_head = free_head;
1109 	*inflight_entry = old_free_head;
1110 
1111 	return 0;
1112 }
1113 
1114 int
1115 rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
1116 				  uint16_t last_used_idx, uint16_t idx)
1117 {
1118 	struct virtio_net *dev;
1119 	struct vhost_virtqueue *vq;
1120 
1121 	dev = get_device(vid);
1122 	if (unlikely(!dev))
1123 		return -1;
1124 
1125 	if (unlikely(!(dev->protocol_features &
1126 	    (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1127 		return 0;
1128 
1129 	if (unlikely(vq_is_packed(dev)))
1130 		return -1;
1131 
1132 	if (unlikely(vring_idx >= VHOST_MAX_VRING))
1133 		return -1;
1134 
1135 	vq = dev->virtqueue[vring_idx];
1136 	if (unlikely(!vq))
1137 		return -1;
1138 
1139 	if (unlikely(!vq->inflight_split))
1140 		return -1;
1141 
1142 	if (unlikely(idx >= vq->size))
1143 		return -1;
1144 
1145 	rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1146 
1147 	vq->inflight_split->desc[idx].inflight = 0;
1148 
1149 	rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1150 
1151 	vq->inflight_split->used_idx = last_used_idx;
1152 	return 0;
1153 }
1154 
1155 int
1156 rte_vhost_clr_inflight_desc_packed(int vid, uint16_t vring_idx,
1157 				   uint16_t head)
1158 {
1159 	struct rte_vhost_inflight_info_packed *inflight_info;
1160 	struct virtio_net *dev;
1161 	struct vhost_virtqueue *vq;
1162 
1163 	dev = get_device(vid);
1164 	if (unlikely(!dev))
1165 		return -1;
1166 
1167 	if (unlikely(!(dev->protocol_features &
1168 	    (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1169 		return 0;
1170 
1171 	if (unlikely(!vq_is_packed(dev)))
1172 		return -1;
1173 
1174 	if (unlikely(vring_idx >= VHOST_MAX_VRING))
1175 		return -1;
1176 
1177 	vq = dev->virtqueue[vring_idx];
1178 	if (unlikely(!vq))
1179 		return -1;
1180 
1181 	inflight_info = vq->inflight_packed;
1182 	if (unlikely(!inflight_info))
1183 		return -1;
1184 
1185 	if (unlikely(head >= vq->size))
1186 		return -1;
1187 
1188 	rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1189 
1190 	inflight_info->desc[head].inflight = 0;
1191 
1192 	rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1193 
1194 	inflight_info->old_free_head = inflight_info->free_head;
1195 	inflight_info->old_used_idx = inflight_info->used_idx;
1196 	inflight_info->old_used_wrap_counter = inflight_info->used_wrap_counter;
1197 
1198 	return 0;
1199 }
1200 
1201 int
1202 rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx,
1203 				     uint16_t idx)
1204 {
1205 	struct virtio_net *dev;
1206 	struct vhost_virtqueue *vq;
1207 
1208 	dev = get_device(vid);
1209 	if (unlikely(!dev))
1210 		return -1;
1211 
1212 	if (unlikely(!(dev->protocol_features &
1213 	    (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1214 		return 0;
1215 
1216 	if (unlikely(vq_is_packed(dev)))
1217 		return -1;
1218 
1219 	if (unlikely(vring_idx >= VHOST_MAX_VRING))
1220 		return -1;
1221 
1222 	vq = dev->virtqueue[vring_idx];
1223 	if (unlikely(!vq))
1224 		return -1;
1225 
1226 	if (unlikely(!vq->inflight_split))
1227 		return -1;
1228 
1229 	vq->inflight_split->last_inflight_io = idx;
1230 	return 0;
1231 }
1232 
1233 int
1234 rte_vhost_set_last_inflight_io_packed(int vid, uint16_t vring_idx,
1235 				      uint16_t head)
1236 {
1237 	struct rte_vhost_inflight_info_packed *inflight_info;
1238 	struct virtio_net *dev;
1239 	struct vhost_virtqueue *vq;
1240 	uint16_t last;
1241 
1242 	dev = get_device(vid);
1243 	if (unlikely(!dev))
1244 		return -1;
1245 
1246 	if (unlikely(!(dev->protocol_features &
1247 	    (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1248 		return 0;
1249 
1250 	if (unlikely(!vq_is_packed(dev)))
1251 		return -1;
1252 
1253 	if (unlikely(vring_idx >= VHOST_MAX_VRING))
1254 		return -1;
1255 
1256 	vq = dev->virtqueue[vring_idx];
1257 	if (unlikely(!vq))
1258 		return -1;
1259 
1260 	inflight_info = vq->inflight_packed;
1261 	if (unlikely(!inflight_info))
1262 		return -1;
1263 
1264 	if (unlikely(head >= vq->size))
1265 		return -1;
1266 
1267 	last = inflight_info->desc[head].last;
1268 	if (unlikely(last >= vq->size))
1269 		return -1;
1270 
1271 	inflight_info->desc[last].next = inflight_info->free_head;
1272 	inflight_info->free_head = head;
1273 	inflight_info->used_idx += inflight_info->desc[head].num;
1274 	if (inflight_info->used_idx >= inflight_info->desc_num) {
1275 		inflight_info->used_idx -= inflight_info->desc_num;
1276 		inflight_info->used_wrap_counter =
1277 			!inflight_info->used_wrap_counter;
1278 	}
1279 
1280 	return 0;
1281 }
1282 
1283 int
1284 rte_vhost_vring_call(int vid, uint16_t vring_idx)
1285 {
1286 	struct virtio_net *dev;
1287 	struct vhost_virtqueue *vq;
1288 
1289 	dev = get_device(vid);
1290 	if (!dev)
1291 		return -1;
1292 
1293 	if (vring_idx >= VHOST_MAX_VRING)
1294 		return -1;
1295 
1296 	vq = dev->virtqueue[vring_idx];
1297 	if (!vq)
1298 		return -1;
1299 
1300 	if (vq_is_packed(dev))
1301 		vhost_vring_call_packed(dev, vq);
1302 	else
1303 		vhost_vring_call_split(dev, vq);
1304 
1305 	return 0;
1306 }
1307 
1308 uint16_t
1309 rte_vhost_avail_entries(int vid, uint16_t queue_id)
1310 {
1311 	struct virtio_net *dev;
1312 	struct vhost_virtqueue *vq;
1313 	uint16_t ret = 0;
1314 
1315 	dev = get_device(vid);
1316 	if (!dev)
1317 		return 0;
1318 
1319 	if (queue_id >= VHOST_MAX_VRING)
1320 		return 0;
1321 
1322 	vq = dev->virtqueue[queue_id];
1323 	if (!vq)
1324 		return 0;
1325 
1326 	rte_spinlock_lock(&vq->access_lock);
1327 
1328 	if (unlikely(!vq->enabled || vq->avail == NULL))
1329 		goto out;
1330 
1331 	ret = *(volatile uint16_t *)&vq->avail->idx - vq->last_used_idx;
1332 
1333 out:
1334 	rte_spinlock_unlock(&vq->access_lock);
1335 	return ret;
1336 }
1337 
1338 static inline int
1339 vhost_enable_notify_split(struct virtio_net *dev,
1340 		struct vhost_virtqueue *vq, int enable)
1341 {
1342 	if (vq->used == NULL)
1343 		return -1;
1344 
1345 	if (!(dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))) {
1346 		if (enable)
1347 			vq->used->flags &= ~VRING_USED_F_NO_NOTIFY;
1348 		else
1349 			vq->used->flags |= VRING_USED_F_NO_NOTIFY;
1350 	} else {
1351 		if (enable)
1352 			vhost_avail_event(vq) = vq->last_avail_idx;
1353 	}
1354 	return 0;
1355 }
1356 
1357 static inline int
1358 vhost_enable_notify_packed(struct virtio_net *dev,
1359 		struct vhost_virtqueue *vq, int enable)
1360 {
1361 	uint16_t flags;
1362 
1363 	if (vq->device_event == NULL)
1364 		return -1;
1365 
1366 	if (!enable) {
1367 		vq->device_event->flags = VRING_EVENT_F_DISABLE;
1368 		return 0;
1369 	}
1370 
1371 	flags = VRING_EVENT_F_ENABLE;
1372 	if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX)) {
1373 		flags = VRING_EVENT_F_DESC;
1374 		vq->device_event->off_wrap = vq->last_avail_idx |
1375 			vq->avail_wrap_counter << 15;
1376 	}
1377 
1378 	rte_atomic_thread_fence(__ATOMIC_RELEASE);
1379 
1380 	vq->device_event->flags = flags;
1381 	return 0;
1382 }
1383 
1384 int
1385 vhost_enable_guest_notification(struct virtio_net *dev,
1386 		struct vhost_virtqueue *vq, int enable)
1387 {
1388 	/*
1389 	 * If the virtqueue is not ready yet, it will be applied
1390 	 * when it will become ready.
1391 	 */
1392 	if (!vq->ready)
1393 		return 0;
1394 
1395 	if (vq_is_packed(dev))
1396 		return vhost_enable_notify_packed(dev, vq, enable);
1397 	else
1398 		return vhost_enable_notify_split(dev, vq, enable);
1399 }
1400 
1401 int
1402 rte_vhost_enable_guest_notification(int vid, uint16_t queue_id, int enable)
1403 {
1404 	struct virtio_net *dev = get_device(vid);
1405 	struct vhost_virtqueue *vq;
1406 	int ret;
1407 
1408 	if (!dev)
1409 		return -1;
1410 
1411 	if (queue_id >= VHOST_MAX_VRING)
1412 		return -1;
1413 
1414 	vq = dev->virtqueue[queue_id];
1415 	if (!vq)
1416 		return -1;
1417 
1418 	rte_spinlock_lock(&vq->access_lock);
1419 
1420 	vq->notif_enable = enable;
1421 	ret = vhost_enable_guest_notification(dev, vq, enable);
1422 
1423 	rte_spinlock_unlock(&vq->access_lock);
1424 
1425 	return ret;
1426 }
1427 
1428 void
1429 rte_vhost_log_write(int vid, uint64_t addr, uint64_t len)
1430 {
1431 	struct virtio_net *dev = get_device(vid);
1432 
1433 	if (dev == NULL)
1434 		return;
1435 
1436 	vhost_log_write(dev, addr, len);
1437 }
1438 
1439 void
1440 rte_vhost_log_used_vring(int vid, uint16_t vring_idx,
1441 			 uint64_t offset, uint64_t len)
1442 {
1443 	struct virtio_net *dev;
1444 	struct vhost_virtqueue *vq;
1445 
1446 	dev = get_device(vid);
1447 	if (dev == NULL)
1448 		return;
1449 
1450 	if (vring_idx >= VHOST_MAX_VRING)
1451 		return;
1452 	vq = dev->virtqueue[vring_idx];
1453 	if (!vq)
1454 		return;
1455 
1456 	vhost_log_used_vring(dev, vq, offset, len);
1457 }
1458 
1459 uint32_t
1460 rte_vhost_rx_queue_count(int vid, uint16_t qid)
1461 {
1462 	struct virtio_net *dev;
1463 	struct vhost_virtqueue *vq;
1464 	uint32_t ret = 0;
1465 
1466 	dev = get_device(vid);
1467 	if (dev == NULL)
1468 		return 0;
1469 
1470 	if (unlikely(qid >= dev->nr_vring || (qid & 1) == 0)) {
1471 		VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1472 			dev->vid, __func__, qid);
1473 		return 0;
1474 	}
1475 
1476 	vq = dev->virtqueue[qid];
1477 	if (vq == NULL)
1478 		return 0;
1479 
1480 	rte_spinlock_lock(&vq->access_lock);
1481 
1482 	if (unlikely(!vq->enabled || vq->avail == NULL))
1483 		goto out;
1484 
1485 	ret = *((volatile uint16_t *)&vq->avail->idx) - vq->last_avail_idx;
1486 
1487 out:
1488 	rte_spinlock_unlock(&vq->access_lock);
1489 	return ret;
1490 }
1491 
1492 struct rte_vdpa_device *
1493 rte_vhost_get_vdpa_device(int vid)
1494 {
1495 	struct virtio_net *dev = get_device(vid);
1496 
1497 	if (dev == NULL)
1498 		return NULL;
1499 
1500 	return dev->vdpa_dev;
1501 }
1502 
1503 int rte_vhost_get_log_base(int vid, uint64_t *log_base,
1504 		uint64_t *log_size)
1505 {
1506 	struct virtio_net *dev = get_device(vid);
1507 
1508 	if (dev == NULL || log_base == NULL || log_size == NULL)
1509 		return -1;
1510 
1511 	*log_base = dev->log_base;
1512 	*log_size = dev->log_size;
1513 
1514 	return 0;
1515 }
1516 
1517 int rte_vhost_get_vring_base(int vid, uint16_t queue_id,
1518 		uint16_t *last_avail_idx, uint16_t *last_used_idx)
1519 {
1520 	struct vhost_virtqueue *vq;
1521 	struct virtio_net *dev = get_device(vid);
1522 
1523 	if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1524 		return -1;
1525 
1526 	if (queue_id >= VHOST_MAX_VRING)
1527 		return -1;
1528 
1529 	vq = dev->virtqueue[queue_id];
1530 	if (!vq)
1531 		return -1;
1532 
1533 	if (vq_is_packed(dev)) {
1534 		*last_avail_idx = (vq->avail_wrap_counter << 15) |
1535 				  vq->last_avail_idx;
1536 		*last_used_idx = (vq->used_wrap_counter << 15) |
1537 				 vq->last_used_idx;
1538 	} else {
1539 		*last_avail_idx = vq->last_avail_idx;
1540 		*last_used_idx = vq->last_used_idx;
1541 	}
1542 
1543 	return 0;
1544 }
1545 
1546 int rte_vhost_set_vring_base(int vid, uint16_t queue_id,
1547 		uint16_t last_avail_idx, uint16_t last_used_idx)
1548 {
1549 	struct vhost_virtqueue *vq;
1550 	struct virtio_net *dev = get_device(vid);
1551 
1552 	if (!dev)
1553 		return -1;
1554 
1555 	if (queue_id >= VHOST_MAX_VRING)
1556 		return -1;
1557 
1558 	vq = dev->virtqueue[queue_id];
1559 	if (!vq)
1560 		return -1;
1561 
1562 	if (vq_is_packed(dev)) {
1563 		vq->last_avail_idx = last_avail_idx & 0x7fff;
1564 		vq->avail_wrap_counter = !!(last_avail_idx & (1 << 15));
1565 		vq->last_used_idx = last_used_idx & 0x7fff;
1566 		vq->used_wrap_counter = !!(last_used_idx & (1 << 15));
1567 	} else {
1568 		vq->last_avail_idx = last_avail_idx;
1569 		vq->last_used_idx = last_used_idx;
1570 	}
1571 
1572 	return 0;
1573 }
1574 
1575 int
1576 rte_vhost_get_vring_base_from_inflight(int vid,
1577 				       uint16_t queue_id,
1578 				       uint16_t *last_avail_idx,
1579 				       uint16_t *last_used_idx)
1580 {
1581 	struct rte_vhost_inflight_info_packed *inflight_info;
1582 	struct vhost_virtqueue *vq;
1583 	struct virtio_net *dev = get_device(vid);
1584 
1585 	if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1586 		return -1;
1587 
1588 	if (queue_id >= VHOST_MAX_VRING)
1589 		return -1;
1590 
1591 	vq = dev->virtqueue[queue_id];
1592 	if (!vq)
1593 		return -1;
1594 
1595 	if (!vq_is_packed(dev))
1596 		return -1;
1597 
1598 	inflight_info = vq->inflight_packed;
1599 	if (!inflight_info)
1600 		return -1;
1601 
1602 	*last_avail_idx = (inflight_info->old_used_wrap_counter << 15) |
1603 			  inflight_info->old_used_idx;
1604 	*last_used_idx = *last_avail_idx;
1605 
1606 	return 0;
1607 }
1608 
1609 int rte_vhost_extern_callback_register(int vid,
1610 		struct rte_vhost_user_extern_ops const * const ops, void *ctx)
1611 {
1612 	struct virtio_net *dev = get_device(vid);
1613 
1614 	if (dev == NULL || ops == NULL)
1615 		return -1;
1616 
1617 	dev->extern_ops = *ops;
1618 	dev->extern_data = ctx;
1619 	return 0;
1620 }
1621 
1622 static __rte_always_inline int
1623 async_channel_register(int vid, uint16_t queue_id,
1624 		struct rte_vhost_async_channel_ops *ops)
1625 {
1626 	struct virtio_net *dev = get_device(vid);
1627 	struct vhost_virtqueue *vq = dev->virtqueue[queue_id];
1628 
1629 	if (unlikely(vq->async_registered)) {
1630 		VHOST_LOG_CONFIG(ERR,
1631 			"async register failed: channel already registered "
1632 			"(vid %d, qid: %d)\n", vid, queue_id);
1633 		return -1;
1634 	}
1635 
1636 	vq->async_pkts_info = rte_malloc_socket(NULL,
1637 			vq->size * sizeof(struct async_inflight_info),
1638 			RTE_CACHE_LINE_SIZE, vq->numa_node);
1639 	if (!vq->async_pkts_info) {
1640 		vhost_free_async_mem(vq);
1641 		VHOST_LOG_CONFIG(ERR,
1642 			"async register failed: cannot allocate memory for async_pkts_info "
1643 			"(vid %d, qid: %d)\n", vid, queue_id);
1644 		return -1;
1645 	}
1646 
1647 	vq->it_pool = rte_malloc_socket(NULL,
1648 			VHOST_MAX_ASYNC_IT * sizeof(struct rte_vhost_iov_iter),
1649 			RTE_CACHE_LINE_SIZE, vq->numa_node);
1650 	if (!vq->it_pool) {
1651 		vhost_free_async_mem(vq);
1652 		VHOST_LOG_CONFIG(ERR,
1653 			"async register failed: cannot allocate memory for it_pool "
1654 			"(vid %d, qid: %d)\n", vid, queue_id);
1655 		return -1;
1656 	}
1657 
1658 	vq->vec_pool = rte_malloc_socket(NULL,
1659 			VHOST_MAX_ASYNC_VEC * sizeof(struct iovec),
1660 			RTE_CACHE_LINE_SIZE, vq->numa_node);
1661 	if (!vq->vec_pool) {
1662 		vhost_free_async_mem(vq);
1663 		VHOST_LOG_CONFIG(ERR,
1664 			"async register failed: cannot allocate memory for vec_pool "
1665 			"(vid %d, qid: %d)\n", vid, queue_id);
1666 		return -1;
1667 	}
1668 
1669 	if (vq_is_packed(dev)) {
1670 		vq->async_buffers_packed = rte_malloc_socket(NULL,
1671 			vq->size * sizeof(struct vring_used_elem_packed),
1672 			RTE_CACHE_LINE_SIZE, vq->numa_node);
1673 		if (!vq->async_buffers_packed) {
1674 			vhost_free_async_mem(vq);
1675 			VHOST_LOG_CONFIG(ERR,
1676 				"async register failed: cannot allocate memory for async buffers "
1677 				"(vid %d, qid: %d)\n", vid, queue_id);
1678 			return -1;
1679 		}
1680 	} else {
1681 		vq->async_descs_split = rte_malloc_socket(NULL,
1682 			vq->size * sizeof(struct vring_used_elem),
1683 			RTE_CACHE_LINE_SIZE, vq->numa_node);
1684 		if (!vq->async_descs_split) {
1685 			vhost_free_async_mem(vq);
1686 			VHOST_LOG_CONFIG(ERR,
1687 				"async register failed: cannot allocate memory for async descs "
1688 				"(vid %d, qid: %d)\n", vid, queue_id);
1689 			return -1;
1690 		}
1691 	}
1692 
1693 	vq->async_ops.check_completed_copies = ops->check_completed_copies;
1694 	vq->async_ops.transfer_data = ops->transfer_data;
1695 
1696 	vq->async_registered = true;
1697 
1698 	return 0;
1699 }
1700 
1701 int
1702 rte_vhost_async_channel_register(int vid, uint16_t queue_id,
1703 		struct rte_vhost_async_config config,
1704 		struct rte_vhost_async_channel_ops *ops)
1705 {
1706 	struct vhost_virtqueue *vq;
1707 	struct virtio_net *dev = get_device(vid);
1708 	int ret;
1709 
1710 	if (dev == NULL || ops == NULL)
1711 		return -1;
1712 
1713 	if (queue_id >= VHOST_MAX_VRING)
1714 		return -1;
1715 
1716 	vq = dev->virtqueue[queue_id];
1717 
1718 	if (unlikely(vq == NULL || !dev->async_copy))
1719 		return -1;
1720 
1721 	if (unlikely(!(config.features & RTE_VHOST_ASYNC_INORDER))) {
1722 		VHOST_LOG_CONFIG(ERR,
1723 			"async copy is not supported on non-inorder mode "
1724 			"(vid %d, qid: %d)\n", vid, queue_id);
1725 		return -1;
1726 	}
1727 
1728 	if (unlikely(ops->check_completed_copies == NULL ||
1729 		ops->transfer_data == NULL))
1730 		return -1;
1731 
1732 	rte_spinlock_lock(&vq->access_lock);
1733 	ret = async_channel_register(vid, queue_id, ops);
1734 	rte_spinlock_unlock(&vq->access_lock);
1735 
1736 	return ret;
1737 }
1738 
1739 int
1740 rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t queue_id,
1741 		struct rte_vhost_async_config config,
1742 		struct rte_vhost_async_channel_ops *ops)
1743 {
1744 	struct vhost_virtqueue *vq;
1745 	struct virtio_net *dev = get_device(vid);
1746 
1747 	if (dev == NULL || ops == NULL)
1748 		return -1;
1749 
1750 	if (queue_id >= VHOST_MAX_VRING)
1751 		return -1;
1752 
1753 	vq = dev->virtqueue[queue_id];
1754 
1755 	if (unlikely(vq == NULL || !dev->async_copy))
1756 		return -1;
1757 
1758 	if (unlikely(!(config.features & RTE_VHOST_ASYNC_INORDER))) {
1759 		VHOST_LOG_CONFIG(ERR,
1760 			"async copy is not supported on non-inorder mode "
1761 			"(vid %d, qid: %d)\n", vid, queue_id);
1762 		return -1;
1763 	}
1764 
1765 	if (unlikely(ops->check_completed_copies == NULL ||
1766 		ops->transfer_data == NULL))
1767 		return -1;
1768 
1769 	return async_channel_register(vid, queue_id, ops);
1770 }
1771 
1772 int
1773 rte_vhost_async_channel_unregister(int vid, uint16_t queue_id)
1774 {
1775 	struct vhost_virtqueue *vq;
1776 	struct virtio_net *dev = get_device(vid);
1777 	int ret = -1;
1778 
1779 	if (dev == NULL)
1780 		return ret;
1781 
1782 	if (queue_id >= VHOST_MAX_VRING)
1783 		return ret;
1784 
1785 	vq = dev->virtqueue[queue_id];
1786 
1787 	if (vq == NULL)
1788 		return ret;
1789 
1790 	ret = 0;
1791 
1792 	if (!vq->async_registered)
1793 		return ret;
1794 
1795 	if (!rte_spinlock_trylock(&vq->access_lock)) {
1796 		VHOST_LOG_CONFIG(ERR, "Failed to unregister async channel. "
1797 			"virt queue busy.\n");
1798 		return -1;
1799 	}
1800 
1801 	if (vq->async_pkts_inflight_n) {
1802 		VHOST_LOG_CONFIG(ERR, "Failed to unregister async channel. "
1803 			"async inflight packets must be completed before unregistration.\n");
1804 		ret = -1;
1805 		goto out;
1806 	}
1807 
1808 	vhost_free_async_mem(vq);
1809 
1810 	vq->async_ops.transfer_data = NULL;
1811 	vq->async_ops.check_completed_copies = NULL;
1812 	vq->async_registered = false;
1813 
1814 out:
1815 	rte_spinlock_unlock(&vq->access_lock);
1816 
1817 	return ret;
1818 }
1819 
1820 int
1821 rte_vhost_async_channel_unregister_thread_unsafe(int vid, uint16_t queue_id)
1822 {
1823 	struct vhost_virtqueue *vq;
1824 	struct virtio_net *dev = get_device(vid);
1825 
1826 	if (dev == NULL)
1827 		return -1;
1828 
1829 	if (queue_id >= VHOST_MAX_VRING)
1830 		return -1;
1831 
1832 	vq = dev->virtqueue[queue_id];
1833 
1834 	if (vq == NULL)
1835 		return -1;
1836 
1837 	if (!vq->async_registered)
1838 		return 0;
1839 
1840 	if (vq->async_pkts_inflight_n) {
1841 		VHOST_LOG_CONFIG(ERR, "Failed to unregister async channel. "
1842 			"async inflight packets must be completed before unregistration.\n");
1843 		return -1;
1844 	}
1845 
1846 	vhost_free_async_mem(vq);
1847 
1848 	vq->async_ops.transfer_data = NULL;
1849 	vq->async_ops.check_completed_copies = NULL;
1850 	vq->async_registered = false;
1851 
1852 	return 0;
1853 }
1854 
1855 int rte_vhost_async_get_inflight(int vid, uint16_t queue_id)
1856 {
1857 	struct vhost_virtqueue *vq;
1858 	struct virtio_net *dev = get_device(vid);
1859 	int ret = -1;
1860 
1861 	if (dev == NULL)
1862 		return ret;
1863 
1864 	if (queue_id >= VHOST_MAX_VRING)
1865 		return ret;
1866 
1867 	vq = dev->virtqueue[queue_id];
1868 
1869 	if (vq == NULL)
1870 		return ret;
1871 
1872 	if (!vq->async_registered)
1873 		return ret;
1874 
1875 	if (!rte_spinlock_trylock(&vq->access_lock)) {
1876 		VHOST_LOG_CONFIG(DEBUG, "Failed to check in-flight packets. "
1877 			"virt queue busy.\n");
1878 		return ret;
1879 	}
1880 
1881 	ret = vq->async_pkts_inflight_n;
1882 	rte_spinlock_unlock(&vq->access_lock);
1883 
1884 	return ret;
1885 }
1886 
1887 RTE_LOG_REGISTER_SUFFIX(vhost_config_log_level, config, INFO);
1888 RTE_LOG_REGISTER_SUFFIX(vhost_data_log_level, data, WARNING);
1889