xref: /dpdk/lib/vhost/vhost.c (revision 97b914f4e715565d53d38ac6e04815b9be5e58a9)
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 <stdint.h>
8 #include <stdlib.h>
9 #ifdef RTE_LIBRTE_VHOST_NUMA
10 #include <numa.h>
11 #include <numaif.h>
12 #endif
13 
14 #include <rte_errno.h>
15 #include <rte_log.h>
16 #include <rte_memory.h>
17 #include <rte_malloc.h>
18 #include <rte_vhost.h>
19 
20 #include "iotlb.h"
21 #include "vhost.h"
22 #include "vhost_user.h"
23 
24 struct virtio_net *vhost_devices[RTE_MAX_VHOST_DEVICE];
25 pthread_mutex_t vhost_dev_lock = PTHREAD_MUTEX_INITIALIZER;
26 
27 /* Called with iotlb_lock read-locked */
28 uint64_t
29 __vhost_iova_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
30 		    uint64_t iova, uint64_t *size, uint8_t perm)
31 {
32 	uint64_t vva, tmp_size;
33 
34 	if (unlikely(!*size))
35 		return 0;
36 
37 	tmp_size = *size;
38 
39 	vva = vhost_user_iotlb_cache_find(vq, iova, &tmp_size, perm);
40 	if (tmp_size == *size)
41 		return vva;
42 
43 	iova += tmp_size;
44 
45 	if (!vhost_user_iotlb_pending_miss(vq, iova, perm)) {
46 		/*
47 		 * iotlb_lock is read-locked for a full burst,
48 		 * but it only protects the iotlb cache.
49 		 * In case of IOTLB miss, we might block on the socket,
50 		 * which could cause a deadlock with QEMU if an IOTLB update
51 		 * is being handled. We can safely unlock here to avoid it.
52 		 */
53 		vhost_user_iotlb_rd_unlock(vq);
54 
55 		vhost_user_iotlb_pending_insert(dev, vq, iova, perm);
56 		if (vhost_user_iotlb_miss(dev, iova, perm)) {
57 			VHOST_LOG_DATA(ERR, "(%s) IOTLB miss req failed for IOVA 0x%" PRIx64 "\n",
58 				dev->ifname, iova);
59 			vhost_user_iotlb_pending_remove(vq, iova, 1, perm);
60 		}
61 
62 		vhost_user_iotlb_rd_lock(vq);
63 	}
64 
65 	return 0;
66 }
67 
68 #define VHOST_LOG_PAGE	4096
69 
70 /*
71  * Atomically set a bit in memory.
72  */
73 static __rte_always_inline void
74 vhost_set_bit(unsigned int nr, volatile uint8_t *addr)
75 {
76 #if defined(RTE_TOOLCHAIN_GCC) && (GCC_VERSION < 70100)
77 	/*
78 	 * __sync_ built-ins are deprecated, but __atomic_ ones
79 	 * are sub-optimized in older GCC versions.
80 	 */
81 	__sync_fetch_and_or_1(addr, (1U << nr));
82 #else
83 	__atomic_fetch_or(addr, (1U << nr), __ATOMIC_RELAXED);
84 #endif
85 }
86 
87 static __rte_always_inline void
88 vhost_log_page(uint8_t *log_base, uint64_t page)
89 {
90 	vhost_set_bit(page % 8, &log_base[page / 8]);
91 }
92 
93 void
94 __vhost_log_write(struct virtio_net *dev, uint64_t addr, uint64_t len)
95 {
96 	uint64_t page;
97 
98 	if (unlikely(!dev->log_base || !len))
99 		return;
100 
101 	if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
102 		return;
103 
104 	/* To make sure guest memory updates are committed before logging */
105 	rte_atomic_thread_fence(__ATOMIC_RELEASE);
106 
107 	page = addr / VHOST_LOG_PAGE;
108 	while (page * VHOST_LOG_PAGE < addr + len) {
109 		vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
110 		page += 1;
111 	}
112 }
113 
114 void
115 __vhost_log_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
116 			     uint64_t iova, uint64_t len)
117 {
118 	uint64_t hva, gpa, map_len;
119 	map_len = len;
120 
121 	hva = __vhost_iova_to_vva(dev, vq, iova, &map_len, VHOST_ACCESS_RW);
122 	if (map_len != len) {
123 		VHOST_LOG_DATA(ERR,
124 			"(%s) failed to write log for IOVA 0x%" PRIx64 ". No IOTLB entry found\n",
125 			dev->ifname, iova);
126 		return;
127 	}
128 
129 	gpa = hva_to_gpa(dev, hva, len);
130 	if (gpa)
131 		__vhost_log_write(dev, gpa, len);
132 }
133 
134 void
135 __vhost_log_cache_sync(struct virtio_net *dev, struct vhost_virtqueue *vq)
136 {
137 	unsigned long *log_base;
138 	int i;
139 
140 	if (unlikely(!dev->log_base))
141 		return;
142 
143 	/* No cache, nothing to sync */
144 	if (unlikely(!vq->log_cache))
145 		return;
146 
147 	rte_atomic_thread_fence(__ATOMIC_RELEASE);
148 
149 	log_base = (unsigned long *)(uintptr_t)dev->log_base;
150 
151 	for (i = 0; i < vq->log_cache_nb_elem; i++) {
152 		struct log_cache_entry *elem = vq->log_cache + i;
153 
154 #if defined(RTE_TOOLCHAIN_GCC) && (GCC_VERSION < 70100)
155 		/*
156 		 * '__sync' builtins are deprecated, but '__atomic' ones
157 		 * are sub-optimized in older GCC versions.
158 		 */
159 		__sync_fetch_and_or(log_base + elem->offset, elem->val);
160 #else
161 		__atomic_fetch_or(log_base + elem->offset, elem->val,
162 				__ATOMIC_RELAXED);
163 #endif
164 	}
165 
166 	rte_atomic_thread_fence(__ATOMIC_RELEASE);
167 
168 	vq->log_cache_nb_elem = 0;
169 }
170 
171 static __rte_always_inline void
172 vhost_log_cache_page(struct virtio_net *dev, struct vhost_virtqueue *vq,
173 			uint64_t page)
174 {
175 	uint32_t bit_nr = page % (sizeof(unsigned long) << 3);
176 	uint32_t offset = page / (sizeof(unsigned long) << 3);
177 	int i;
178 
179 	if (unlikely(!vq->log_cache)) {
180 		/* No logging cache allocated, write dirty log map directly */
181 		rte_atomic_thread_fence(__ATOMIC_RELEASE);
182 		vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
183 
184 		return;
185 	}
186 
187 	for (i = 0; i < vq->log_cache_nb_elem; i++) {
188 		struct log_cache_entry *elem = vq->log_cache + i;
189 
190 		if (elem->offset == offset) {
191 			elem->val |= (1UL << bit_nr);
192 			return;
193 		}
194 	}
195 
196 	if (unlikely(i >= VHOST_LOG_CACHE_NR)) {
197 		/*
198 		 * No more room for a new log cache entry,
199 		 * so write the dirty log map directly.
200 		 */
201 		rte_atomic_thread_fence(__ATOMIC_RELEASE);
202 		vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
203 
204 		return;
205 	}
206 
207 	vq->log_cache[i].offset = offset;
208 	vq->log_cache[i].val = (1UL << bit_nr);
209 	vq->log_cache_nb_elem++;
210 }
211 
212 void
213 __vhost_log_cache_write(struct virtio_net *dev, struct vhost_virtqueue *vq,
214 			uint64_t addr, uint64_t len)
215 {
216 	uint64_t page;
217 
218 	if (unlikely(!dev->log_base || !len))
219 		return;
220 
221 	if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
222 		return;
223 
224 	page = addr / VHOST_LOG_PAGE;
225 	while (page * VHOST_LOG_PAGE < addr + len) {
226 		vhost_log_cache_page(dev, vq, page);
227 		page += 1;
228 	}
229 }
230 
231 void
232 __vhost_log_cache_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
233 			     uint64_t iova, uint64_t len)
234 {
235 	uint64_t hva, gpa, map_len;
236 	map_len = len;
237 
238 	hva = __vhost_iova_to_vva(dev, vq, iova, &map_len, VHOST_ACCESS_RW);
239 	if (map_len != len) {
240 		VHOST_LOG_DATA(ERR,
241 			"(%s) failed to write log for IOVA 0x%" PRIx64 ". No IOTLB entry found\n",
242 			dev->ifname, iova);
243 		return;
244 	}
245 
246 	gpa = hva_to_gpa(dev, hva, len);
247 	if (gpa)
248 		__vhost_log_cache_write(dev, vq, gpa, len);
249 }
250 
251 void *
252 vhost_alloc_copy_ind_table(struct virtio_net *dev, struct vhost_virtqueue *vq,
253 		uint64_t desc_addr, uint64_t desc_len)
254 {
255 	void *idesc;
256 	uint64_t src, dst;
257 	uint64_t len, remain = desc_len;
258 
259 	idesc = rte_malloc_socket(__func__, desc_len, 0, vq->numa_node);
260 	if (unlikely(!idesc))
261 		return NULL;
262 
263 	dst = (uint64_t)(uintptr_t)idesc;
264 
265 	while (remain) {
266 		len = remain;
267 		src = vhost_iova_to_vva(dev, vq, desc_addr, &len,
268 				VHOST_ACCESS_RO);
269 		if (unlikely(!src || !len)) {
270 			rte_free(idesc);
271 			return NULL;
272 		}
273 
274 		rte_memcpy((void *)(uintptr_t)dst, (void *)(uintptr_t)src, len);
275 
276 		remain -= len;
277 		dst += len;
278 		desc_addr += len;
279 	}
280 
281 	return idesc;
282 }
283 
284 void
285 cleanup_vq(struct vhost_virtqueue *vq, int destroy)
286 {
287 	if ((vq->callfd >= 0) && (destroy != 0))
288 		close(vq->callfd);
289 	if (vq->kickfd >= 0)
290 		close(vq->kickfd);
291 }
292 
293 void
294 cleanup_vq_inflight(struct virtio_net *dev, struct vhost_virtqueue *vq)
295 {
296 	if (!(dev->protocol_features &
297 	    (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
298 		return;
299 
300 	if (vq_is_packed(dev)) {
301 		if (vq->inflight_packed)
302 			vq->inflight_packed = NULL;
303 	} else {
304 		if (vq->inflight_split)
305 			vq->inflight_split = NULL;
306 	}
307 
308 	if (vq->resubmit_inflight) {
309 		if (vq->resubmit_inflight->resubmit_list) {
310 			rte_free(vq->resubmit_inflight->resubmit_list);
311 			vq->resubmit_inflight->resubmit_list = NULL;
312 		}
313 		rte_free(vq->resubmit_inflight);
314 		vq->resubmit_inflight = NULL;
315 	}
316 }
317 
318 /*
319  * Unmap any memory, close any file descriptors and
320  * free any memory owned by a device.
321  */
322 void
323 cleanup_device(struct virtio_net *dev, int destroy)
324 {
325 	uint32_t i;
326 
327 	vhost_backend_cleanup(dev);
328 
329 	for (i = 0; i < dev->nr_vring; i++) {
330 		cleanup_vq(dev->virtqueue[i], destroy);
331 		cleanup_vq_inflight(dev, dev->virtqueue[i]);
332 	}
333 }
334 
335 static void
336 vhost_free_async_mem(struct vhost_virtqueue *vq)
337 {
338 	if (!vq->async)
339 		return;
340 
341 	rte_free(vq->async->pkts_info);
342 	rte_free(vq->async->pkts_cmpl_flag);
343 
344 	rte_free(vq->async->buffers_packed);
345 	vq->async->buffers_packed = NULL;
346 	rte_free(vq->async->descs_split);
347 	vq->async->descs_split = NULL;
348 
349 	rte_free(vq->async);
350 	vq->async = NULL;
351 }
352 
353 void
354 free_vq(struct virtio_net *dev, struct vhost_virtqueue *vq)
355 {
356 	if (vq_is_packed(dev))
357 		rte_free(vq->shadow_used_packed);
358 	else
359 		rte_free(vq->shadow_used_split);
360 
361 	vhost_free_async_mem(vq);
362 	rte_free(vq->batch_copy_elems);
363 	rte_mempool_free(vq->iotlb_pool);
364 	rte_free(vq->log_cache);
365 	rte_free(vq);
366 }
367 
368 /*
369  * Release virtqueues and device memory.
370  */
371 static void
372 free_device(struct virtio_net *dev)
373 {
374 	uint32_t i;
375 
376 	for (i = 0; i < dev->nr_vring; i++)
377 		free_vq(dev, dev->virtqueue[i]);
378 
379 	rte_free(dev);
380 }
381 
382 static __rte_always_inline int
383 log_translate(struct virtio_net *dev, struct vhost_virtqueue *vq)
384 {
385 	if (likely(!(vq->ring_addrs.flags & (1 << VHOST_VRING_F_LOG))))
386 		return 0;
387 
388 	vq->log_guest_addr = translate_log_addr(dev, vq,
389 						vq->ring_addrs.log_guest_addr);
390 	if (vq->log_guest_addr == 0)
391 		return -1;
392 
393 	return 0;
394 }
395 
396 /*
397  * Converts vring log address to GPA
398  * If IOMMU is enabled, the log address is IOVA
399  * If IOMMU not enabled, the log address is already GPA
400  *
401  * Caller should have iotlb_lock read-locked
402  */
403 uint64_t
404 translate_log_addr(struct virtio_net *dev, struct vhost_virtqueue *vq,
405 		uint64_t log_addr)
406 {
407 	if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
408 		const uint64_t exp_size = sizeof(uint64_t);
409 		uint64_t hva, gpa;
410 		uint64_t size = exp_size;
411 
412 		hva = vhost_iova_to_vva(dev, vq, log_addr,
413 					&size, VHOST_ACCESS_RW);
414 
415 		if (size != exp_size)
416 			return 0;
417 
418 		gpa = hva_to_gpa(dev, hva, exp_size);
419 		if (!gpa) {
420 			VHOST_LOG_DATA(ERR,
421 				"(%s) failed to find GPA for log_addr: 0x%"
422 				PRIx64 " hva: 0x%" PRIx64 "\n",
423 				dev->ifname, log_addr, hva);
424 			return 0;
425 		}
426 		return gpa;
427 
428 	} else
429 		return log_addr;
430 }
431 
432 /* Caller should have iotlb_lock read-locked */
433 static int
434 vring_translate_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
435 {
436 	uint64_t req_size, size;
437 
438 	req_size = sizeof(struct vring_desc) * vq->size;
439 	size = req_size;
440 	vq->desc = (struct vring_desc *)(uintptr_t)vhost_iova_to_vva(dev, vq,
441 						vq->ring_addrs.desc_user_addr,
442 						&size, VHOST_ACCESS_RW);
443 	if (!vq->desc || size != req_size)
444 		return -1;
445 
446 	req_size = sizeof(struct vring_avail);
447 	req_size += sizeof(uint16_t) * vq->size;
448 	if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
449 		req_size += sizeof(uint16_t);
450 	size = req_size;
451 	vq->avail = (struct vring_avail *)(uintptr_t)vhost_iova_to_vva(dev, vq,
452 						vq->ring_addrs.avail_user_addr,
453 						&size, VHOST_ACCESS_RW);
454 	if (!vq->avail || size != req_size)
455 		return -1;
456 
457 	req_size = sizeof(struct vring_used);
458 	req_size += sizeof(struct vring_used_elem) * vq->size;
459 	if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
460 		req_size += sizeof(uint16_t);
461 	size = req_size;
462 	vq->used = (struct vring_used *)(uintptr_t)vhost_iova_to_vva(dev, vq,
463 						vq->ring_addrs.used_user_addr,
464 						&size, VHOST_ACCESS_RW);
465 	if (!vq->used || size != req_size)
466 		return -1;
467 
468 	return 0;
469 }
470 
471 /* Caller should have iotlb_lock read-locked */
472 static int
473 vring_translate_packed(struct virtio_net *dev, struct vhost_virtqueue *vq)
474 {
475 	uint64_t req_size, size;
476 
477 	req_size = sizeof(struct vring_packed_desc) * vq->size;
478 	size = req_size;
479 	vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
480 		vhost_iova_to_vva(dev, vq, vq->ring_addrs.desc_user_addr,
481 				&size, VHOST_ACCESS_RW);
482 	if (!vq->desc_packed || size != req_size)
483 		return -1;
484 
485 	req_size = sizeof(struct vring_packed_desc_event);
486 	size = req_size;
487 	vq->driver_event = (struct vring_packed_desc_event *)(uintptr_t)
488 		vhost_iova_to_vva(dev, vq, vq->ring_addrs.avail_user_addr,
489 				&size, VHOST_ACCESS_RW);
490 	if (!vq->driver_event || size != req_size)
491 		return -1;
492 
493 	req_size = sizeof(struct vring_packed_desc_event);
494 	size = req_size;
495 	vq->device_event = (struct vring_packed_desc_event *)(uintptr_t)
496 		vhost_iova_to_vva(dev, vq, vq->ring_addrs.used_user_addr,
497 				&size, VHOST_ACCESS_RW);
498 	if (!vq->device_event || size != req_size)
499 		return -1;
500 
501 	return 0;
502 }
503 
504 int
505 vring_translate(struct virtio_net *dev, struct vhost_virtqueue *vq)
506 {
507 
508 	if (!(dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
509 		return -1;
510 
511 	if (vq_is_packed(dev)) {
512 		if (vring_translate_packed(dev, vq) < 0)
513 			return -1;
514 	} else {
515 		if (vring_translate_split(dev, vq) < 0)
516 			return -1;
517 	}
518 
519 	if (log_translate(dev, vq) < 0)
520 		return -1;
521 
522 	vq->access_ok = true;
523 
524 	return 0;
525 }
526 
527 void
528 vring_invalidate(struct virtio_net *dev, struct vhost_virtqueue *vq)
529 {
530 	if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
531 		vhost_user_iotlb_wr_lock(vq);
532 
533 	vq->access_ok = false;
534 	vq->desc = NULL;
535 	vq->avail = NULL;
536 	vq->used = NULL;
537 	vq->log_guest_addr = 0;
538 
539 	if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
540 		vhost_user_iotlb_wr_unlock(vq);
541 }
542 
543 static void
544 init_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
545 {
546 	struct vhost_virtqueue *vq;
547 	int numa_node = SOCKET_ID_ANY;
548 
549 	if (vring_idx >= VHOST_MAX_VRING) {
550 		VHOST_LOG_CONFIG(ERR, "(%s) failed to init vring, out of bound (%d)\n",
551 				dev->ifname, vring_idx);
552 		return;
553 	}
554 
555 	vq = dev->virtqueue[vring_idx];
556 	if (!vq) {
557 		VHOST_LOG_CONFIG(ERR, "(%s) virtqueue not allocated (%d)\n",
558 				dev->ifname, vring_idx);
559 		return;
560 	}
561 
562 	memset(vq, 0, sizeof(struct vhost_virtqueue));
563 
564 	vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
565 	vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
566 	vq->notif_enable = VIRTIO_UNINITIALIZED_NOTIF;
567 
568 #ifdef RTE_LIBRTE_VHOST_NUMA
569 	if (get_mempolicy(&numa_node, NULL, 0, vq, MPOL_F_NODE | MPOL_F_ADDR)) {
570 		VHOST_LOG_CONFIG(ERR, "(%s) failed to query numa node: %s\n",
571 			dev->ifname, rte_strerror(errno));
572 		numa_node = SOCKET_ID_ANY;
573 	}
574 #endif
575 	vq->numa_node = numa_node;
576 
577 	vhost_user_iotlb_init(dev, vring_idx);
578 }
579 
580 static void
581 reset_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
582 {
583 	struct vhost_virtqueue *vq;
584 	int callfd;
585 
586 	if (vring_idx >= VHOST_MAX_VRING) {
587 		VHOST_LOG_CONFIG(ERR,
588 				"(%s) failed to reset vring, out of bound (%d)\n",
589 				dev->ifname, vring_idx);
590 		return;
591 	}
592 
593 	vq = dev->virtqueue[vring_idx];
594 	if (!vq) {
595 		VHOST_LOG_CONFIG(ERR, "(%s) failed to reset vring, virtqueue not allocated (%d)\n",
596 				dev->ifname, vring_idx);
597 		return;
598 	}
599 
600 	callfd = vq->callfd;
601 	init_vring_queue(dev, vring_idx);
602 	vq->callfd = callfd;
603 }
604 
605 int
606 alloc_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
607 {
608 	struct vhost_virtqueue *vq;
609 	uint32_t i;
610 
611 	/* Also allocate holes, if any, up to requested vring index. */
612 	for (i = 0; i <= vring_idx; i++) {
613 		if (dev->virtqueue[i])
614 			continue;
615 
616 		vq = rte_zmalloc(NULL, sizeof(struct vhost_virtqueue), 0);
617 		if (vq == NULL) {
618 			VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate memory for vring %u.\n",
619 					dev->ifname, i);
620 			return -1;
621 		}
622 
623 		dev->virtqueue[i] = vq;
624 		init_vring_queue(dev, i);
625 		rte_spinlock_init(&vq->access_lock);
626 		vq->avail_wrap_counter = 1;
627 		vq->used_wrap_counter = 1;
628 		vq->signalled_used_valid = false;
629 	}
630 
631 	dev->nr_vring = RTE_MAX(dev->nr_vring, vring_idx + 1);
632 
633 	return 0;
634 }
635 
636 /*
637  * Reset some variables in device structure, while keeping few
638  * others untouched, such as vid, ifname, nr_vring: they
639  * should be same unless the device is removed.
640  */
641 void
642 reset_device(struct virtio_net *dev)
643 {
644 	uint32_t i;
645 
646 	dev->features = 0;
647 	dev->protocol_features = 0;
648 	dev->flags &= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
649 
650 	for (i = 0; i < dev->nr_vring; i++)
651 		reset_vring_queue(dev, i);
652 }
653 
654 /*
655  * Invoked when there is a new vhost-user connection established (when
656  * there is a new virtio device being attached).
657  */
658 int
659 vhost_new_device(void)
660 {
661 	struct virtio_net *dev;
662 	int i;
663 
664 	pthread_mutex_lock(&vhost_dev_lock);
665 	for (i = 0; i < RTE_MAX_VHOST_DEVICE; i++) {
666 		if (vhost_devices[i] == NULL)
667 			break;
668 	}
669 
670 	if (i == RTE_MAX_VHOST_DEVICE) {
671 		VHOST_LOG_CONFIG(ERR, "failed to find a free slot for new device.\n");
672 		pthread_mutex_unlock(&vhost_dev_lock);
673 		return -1;
674 	}
675 
676 	dev = rte_zmalloc(NULL, sizeof(struct virtio_net), 0);
677 	if (dev == NULL) {
678 		VHOST_LOG_CONFIG(ERR, "failed to allocate memory for new device.\n");
679 		pthread_mutex_unlock(&vhost_dev_lock);
680 		return -1;
681 	}
682 
683 	vhost_devices[i] = dev;
684 	pthread_mutex_unlock(&vhost_dev_lock);
685 
686 	dev->vid = i;
687 	dev->flags = VIRTIO_DEV_BUILTIN_VIRTIO_NET;
688 	dev->slave_req_fd = -1;
689 	dev->postcopy_ufd = -1;
690 	rte_spinlock_init(&dev->slave_req_lock);
691 
692 	return i;
693 }
694 
695 void
696 vhost_destroy_device_notify(struct virtio_net *dev)
697 {
698 	struct rte_vdpa_device *vdpa_dev;
699 
700 	if (dev->flags & VIRTIO_DEV_RUNNING) {
701 		vdpa_dev = dev->vdpa_dev;
702 		if (vdpa_dev)
703 			vdpa_dev->ops->dev_close(dev->vid);
704 		dev->flags &= ~VIRTIO_DEV_RUNNING;
705 		dev->notify_ops->destroy_device(dev->vid);
706 	}
707 }
708 
709 /*
710  * Invoked when there is the vhost-user connection is broken (when
711  * the virtio device is being detached).
712  */
713 void
714 vhost_destroy_device(int vid)
715 {
716 	struct virtio_net *dev = get_device(vid);
717 
718 	if (dev == NULL)
719 		return;
720 
721 	vhost_destroy_device_notify(dev);
722 
723 	cleanup_device(dev, 1);
724 	free_device(dev);
725 
726 	vhost_devices[vid] = NULL;
727 }
728 
729 void
730 vhost_attach_vdpa_device(int vid, struct rte_vdpa_device *vdpa_dev)
731 {
732 	struct virtio_net *dev = get_device(vid);
733 
734 	if (dev == NULL)
735 		return;
736 
737 	dev->vdpa_dev = vdpa_dev;
738 }
739 
740 void
741 vhost_set_ifname(int vid, const char *if_name, unsigned int if_len)
742 {
743 	struct virtio_net *dev;
744 	unsigned int len;
745 
746 	dev = get_device(vid);
747 	if (dev == NULL)
748 		return;
749 
750 	len = if_len > sizeof(dev->ifname) ?
751 		sizeof(dev->ifname) : if_len;
752 
753 	strncpy(dev->ifname, if_name, len);
754 	dev->ifname[sizeof(dev->ifname) - 1] = '\0';
755 }
756 
757 void
758 vhost_setup_virtio_net(int vid, bool enable, bool compliant_ol_flags)
759 {
760 	struct virtio_net *dev = get_device(vid);
761 
762 	if (dev == NULL)
763 		return;
764 
765 	if (enable)
766 		dev->flags |= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
767 	else
768 		dev->flags &= ~VIRTIO_DEV_BUILTIN_VIRTIO_NET;
769 	if (!compliant_ol_flags)
770 		dev->flags |= VIRTIO_DEV_LEGACY_OL_FLAGS;
771 	else
772 		dev->flags &= ~VIRTIO_DEV_LEGACY_OL_FLAGS;
773 }
774 
775 void
776 vhost_enable_extbuf(int vid)
777 {
778 	struct virtio_net *dev = get_device(vid);
779 
780 	if (dev == NULL)
781 		return;
782 
783 	dev->extbuf = 1;
784 }
785 
786 void
787 vhost_enable_linearbuf(int vid)
788 {
789 	struct virtio_net *dev = get_device(vid);
790 
791 	if (dev == NULL)
792 		return;
793 
794 	dev->linearbuf = 1;
795 }
796 
797 int
798 rte_vhost_get_mtu(int vid, uint16_t *mtu)
799 {
800 	struct virtio_net *dev = get_device(vid);
801 
802 	if (dev == NULL || mtu == NULL)
803 		return -ENODEV;
804 
805 	if (!(dev->flags & VIRTIO_DEV_READY))
806 		return -EAGAIN;
807 
808 	if (!(dev->features & (1ULL << VIRTIO_NET_F_MTU)))
809 		return -ENOTSUP;
810 
811 	*mtu = dev->mtu;
812 
813 	return 0;
814 }
815 
816 int
817 rte_vhost_get_numa_node(int vid)
818 {
819 #ifdef RTE_LIBRTE_VHOST_NUMA
820 	struct virtio_net *dev = get_device(vid);
821 	int numa_node;
822 	int ret;
823 
824 	if (dev == NULL || numa_available() != 0)
825 		return -1;
826 
827 	ret = get_mempolicy(&numa_node, NULL, 0, dev,
828 			    MPOL_F_NODE | MPOL_F_ADDR);
829 	if (ret < 0) {
830 		VHOST_LOG_CONFIG(ERR, "(%s) failed to query numa node: %s\n",
831 			dev->ifname, rte_strerror(errno));
832 		return -1;
833 	}
834 
835 	return numa_node;
836 #else
837 	RTE_SET_USED(vid);
838 	return -1;
839 #endif
840 }
841 
842 uint32_t
843 rte_vhost_get_queue_num(int vid)
844 {
845 	struct virtio_net *dev = get_device(vid);
846 
847 	if (dev == NULL)
848 		return 0;
849 
850 	return dev->nr_vring / 2;
851 }
852 
853 uint16_t
854 rte_vhost_get_vring_num(int vid)
855 {
856 	struct virtio_net *dev = get_device(vid);
857 
858 	if (dev == NULL)
859 		return 0;
860 
861 	return dev->nr_vring;
862 }
863 
864 int
865 rte_vhost_get_ifname(int vid, char *buf, size_t len)
866 {
867 	struct virtio_net *dev = get_device(vid);
868 
869 	if (dev == NULL || buf == NULL)
870 		return -1;
871 
872 	len = RTE_MIN(len, sizeof(dev->ifname));
873 
874 	strncpy(buf, dev->ifname, len);
875 	buf[len - 1] = '\0';
876 
877 	return 0;
878 }
879 
880 int
881 rte_vhost_get_negotiated_features(int vid, uint64_t *features)
882 {
883 	struct virtio_net *dev;
884 
885 	dev = get_device(vid);
886 	if (dev == NULL || features == NULL)
887 		return -1;
888 
889 	*features = dev->features;
890 	return 0;
891 }
892 
893 int
894 rte_vhost_get_negotiated_protocol_features(int vid,
895 					   uint64_t *protocol_features)
896 {
897 	struct virtio_net *dev;
898 
899 	dev = get_device(vid);
900 	if (dev == NULL || protocol_features == NULL)
901 		return -1;
902 
903 	*protocol_features = dev->protocol_features;
904 	return 0;
905 }
906 
907 int
908 rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem)
909 {
910 	struct virtio_net *dev;
911 	struct rte_vhost_memory *m;
912 	size_t size;
913 
914 	dev = get_device(vid);
915 	if (dev == NULL || mem == NULL)
916 		return -1;
917 
918 	size = dev->mem->nregions * sizeof(struct rte_vhost_mem_region);
919 	m = malloc(sizeof(struct rte_vhost_memory) + size);
920 	if (!m)
921 		return -1;
922 
923 	m->nregions = dev->mem->nregions;
924 	memcpy(m->regions, dev->mem->regions, size);
925 	*mem = m;
926 
927 	return 0;
928 }
929 
930 int
931 rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
932 			  struct rte_vhost_vring *vring)
933 {
934 	struct virtio_net *dev;
935 	struct vhost_virtqueue *vq;
936 
937 	dev = get_device(vid);
938 	if (dev == NULL || vring == NULL)
939 		return -1;
940 
941 	if (vring_idx >= VHOST_MAX_VRING)
942 		return -1;
943 
944 	vq = dev->virtqueue[vring_idx];
945 	if (!vq)
946 		return -1;
947 
948 	if (vq_is_packed(dev)) {
949 		vring->desc_packed = vq->desc_packed;
950 		vring->driver_event = vq->driver_event;
951 		vring->device_event = vq->device_event;
952 	} else {
953 		vring->desc = vq->desc;
954 		vring->avail = vq->avail;
955 		vring->used = vq->used;
956 	}
957 	vring->log_guest_addr  = vq->log_guest_addr;
958 
959 	vring->callfd  = vq->callfd;
960 	vring->kickfd  = vq->kickfd;
961 	vring->size    = vq->size;
962 
963 	return 0;
964 }
965 
966 int
967 rte_vhost_get_vhost_ring_inflight(int vid, uint16_t vring_idx,
968 				  struct rte_vhost_ring_inflight *vring)
969 {
970 	struct virtio_net *dev;
971 	struct vhost_virtqueue *vq;
972 
973 	dev = get_device(vid);
974 	if (unlikely(!dev))
975 		return -1;
976 
977 	if (vring_idx >= VHOST_MAX_VRING)
978 		return -1;
979 
980 	vq = dev->virtqueue[vring_idx];
981 	if (unlikely(!vq))
982 		return -1;
983 
984 	if (vq_is_packed(dev)) {
985 		if (unlikely(!vq->inflight_packed))
986 			return -1;
987 
988 		vring->inflight_packed = vq->inflight_packed;
989 	} else {
990 		if (unlikely(!vq->inflight_split))
991 			return -1;
992 
993 		vring->inflight_split = vq->inflight_split;
994 	}
995 
996 	vring->resubmit_inflight = vq->resubmit_inflight;
997 
998 	return 0;
999 }
1000 
1001 int
1002 rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx,
1003 				  uint16_t idx)
1004 {
1005 	struct vhost_virtqueue *vq;
1006 	struct virtio_net *dev;
1007 
1008 	dev = get_device(vid);
1009 	if (unlikely(!dev))
1010 		return -1;
1011 
1012 	if (unlikely(!(dev->protocol_features &
1013 	    (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1014 		return 0;
1015 
1016 	if (unlikely(vq_is_packed(dev)))
1017 		return -1;
1018 
1019 	if (unlikely(vring_idx >= VHOST_MAX_VRING))
1020 		return -1;
1021 
1022 	vq = dev->virtqueue[vring_idx];
1023 	if (unlikely(!vq))
1024 		return -1;
1025 
1026 	if (unlikely(!vq->inflight_split))
1027 		return -1;
1028 
1029 	if (unlikely(idx >= vq->size))
1030 		return -1;
1031 
1032 	vq->inflight_split->desc[idx].counter = vq->global_counter++;
1033 	vq->inflight_split->desc[idx].inflight = 1;
1034 	return 0;
1035 }
1036 
1037 int
1038 rte_vhost_set_inflight_desc_packed(int vid, uint16_t vring_idx,
1039 				   uint16_t head, uint16_t last,
1040 				   uint16_t *inflight_entry)
1041 {
1042 	struct rte_vhost_inflight_info_packed *inflight_info;
1043 	struct virtio_net *dev;
1044 	struct vhost_virtqueue *vq;
1045 	struct vring_packed_desc *desc;
1046 	uint16_t old_free_head, free_head;
1047 
1048 	dev = get_device(vid);
1049 	if (unlikely(!dev))
1050 		return -1;
1051 
1052 	if (unlikely(!(dev->protocol_features &
1053 	    (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1054 		return 0;
1055 
1056 	if (unlikely(!vq_is_packed(dev)))
1057 		return -1;
1058 
1059 	if (unlikely(vring_idx >= VHOST_MAX_VRING))
1060 		return -1;
1061 
1062 	vq = dev->virtqueue[vring_idx];
1063 	if (unlikely(!vq))
1064 		return -1;
1065 
1066 	inflight_info = vq->inflight_packed;
1067 	if (unlikely(!inflight_info))
1068 		return -1;
1069 
1070 	if (unlikely(head >= vq->size))
1071 		return -1;
1072 
1073 	desc = vq->desc_packed;
1074 	old_free_head = inflight_info->old_free_head;
1075 	if (unlikely(old_free_head >= vq->size))
1076 		return -1;
1077 
1078 	free_head = old_free_head;
1079 
1080 	/* init header descriptor */
1081 	inflight_info->desc[old_free_head].num = 0;
1082 	inflight_info->desc[old_free_head].counter = vq->global_counter++;
1083 	inflight_info->desc[old_free_head].inflight = 1;
1084 
1085 	/* save desc entry in flight entry */
1086 	while (head != ((last + 1) % vq->size)) {
1087 		inflight_info->desc[old_free_head].num++;
1088 		inflight_info->desc[free_head].addr = desc[head].addr;
1089 		inflight_info->desc[free_head].len = desc[head].len;
1090 		inflight_info->desc[free_head].flags = desc[head].flags;
1091 		inflight_info->desc[free_head].id = desc[head].id;
1092 
1093 		inflight_info->desc[old_free_head].last = free_head;
1094 		free_head = inflight_info->desc[free_head].next;
1095 		inflight_info->free_head = free_head;
1096 		head = (head + 1) % vq->size;
1097 	}
1098 
1099 	inflight_info->old_free_head = free_head;
1100 	*inflight_entry = old_free_head;
1101 
1102 	return 0;
1103 }
1104 
1105 int
1106 rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
1107 				  uint16_t last_used_idx, uint16_t idx)
1108 {
1109 	struct virtio_net *dev;
1110 	struct vhost_virtqueue *vq;
1111 
1112 	dev = get_device(vid);
1113 	if (unlikely(!dev))
1114 		return -1;
1115 
1116 	if (unlikely(!(dev->protocol_features &
1117 	    (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1118 		return 0;
1119 
1120 	if (unlikely(vq_is_packed(dev)))
1121 		return -1;
1122 
1123 	if (unlikely(vring_idx >= VHOST_MAX_VRING))
1124 		return -1;
1125 
1126 	vq = dev->virtqueue[vring_idx];
1127 	if (unlikely(!vq))
1128 		return -1;
1129 
1130 	if (unlikely(!vq->inflight_split))
1131 		return -1;
1132 
1133 	if (unlikely(idx >= vq->size))
1134 		return -1;
1135 
1136 	rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1137 
1138 	vq->inflight_split->desc[idx].inflight = 0;
1139 
1140 	rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1141 
1142 	vq->inflight_split->used_idx = last_used_idx;
1143 	return 0;
1144 }
1145 
1146 int
1147 rte_vhost_clr_inflight_desc_packed(int vid, uint16_t vring_idx,
1148 				   uint16_t head)
1149 {
1150 	struct rte_vhost_inflight_info_packed *inflight_info;
1151 	struct virtio_net *dev;
1152 	struct vhost_virtqueue *vq;
1153 
1154 	dev = get_device(vid);
1155 	if (unlikely(!dev))
1156 		return -1;
1157 
1158 	if (unlikely(!(dev->protocol_features &
1159 	    (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1160 		return 0;
1161 
1162 	if (unlikely(!vq_is_packed(dev)))
1163 		return -1;
1164 
1165 	if (unlikely(vring_idx >= VHOST_MAX_VRING))
1166 		return -1;
1167 
1168 	vq = dev->virtqueue[vring_idx];
1169 	if (unlikely(!vq))
1170 		return -1;
1171 
1172 	inflight_info = vq->inflight_packed;
1173 	if (unlikely(!inflight_info))
1174 		return -1;
1175 
1176 	if (unlikely(head >= vq->size))
1177 		return -1;
1178 
1179 	rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1180 
1181 	inflight_info->desc[head].inflight = 0;
1182 
1183 	rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1184 
1185 	inflight_info->old_free_head = inflight_info->free_head;
1186 	inflight_info->old_used_idx = inflight_info->used_idx;
1187 	inflight_info->old_used_wrap_counter = inflight_info->used_wrap_counter;
1188 
1189 	return 0;
1190 }
1191 
1192 int
1193 rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx,
1194 				     uint16_t idx)
1195 {
1196 	struct virtio_net *dev;
1197 	struct vhost_virtqueue *vq;
1198 
1199 	dev = get_device(vid);
1200 	if (unlikely(!dev))
1201 		return -1;
1202 
1203 	if (unlikely(!(dev->protocol_features &
1204 	    (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1205 		return 0;
1206 
1207 	if (unlikely(vq_is_packed(dev)))
1208 		return -1;
1209 
1210 	if (unlikely(vring_idx >= VHOST_MAX_VRING))
1211 		return -1;
1212 
1213 	vq = dev->virtqueue[vring_idx];
1214 	if (unlikely(!vq))
1215 		return -1;
1216 
1217 	if (unlikely(!vq->inflight_split))
1218 		return -1;
1219 
1220 	if (unlikely(idx >= vq->size))
1221 		return -1;
1222 
1223 	vq->inflight_split->last_inflight_io = idx;
1224 	return 0;
1225 }
1226 
1227 int
1228 rte_vhost_set_last_inflight_io_packed(int vid, uint16_t vring_idx,
1229 				      uint16_t head)
1230 {
1231 	struct rte_vhost_inflight_info_packed *inflight_info;
1232 	struct virtio_net *dev;
1233 	struct vhost_virtqueue *vq;
1234 	uint16_t last;
1235 
1236 	dev = get_device(vid);
1237 	if (unlikely(!dev))
1238 		return -1;
1239 
1240 	if (unlikely(!(dev->protocol_features &
1241 	    (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1242 		return 0;
1243 
1244 	if (unlikely(!vq_is_packed(dev)))
1245 		return -1;
1246 
1247 	if (unlikely(vring_idx >= VHOST_MAX_VRING))
1248 		return -1;
1249 
1250 	vq = dev->virtqueue[vring_idx];
1251 	if (unlikely(!vq))
1252 		return -1;
1253 
1254 	inflight_info = vq->inflight_packed;
1255 	if (unlikely(!inflight_info))
1256 		return -1;
1257 
1258 	if (unlikely(head >= vq->size))
1259 		return -1;
1260 
1261 	last = inflight_info->desc[head].last;
1262 	if (unlikely(last >= vq->size))
1263 		return -1;
1264 
1265 	inflight_info->desc[last].next = inflight_info->free_head;
1266 	inflight_info->free_head = head;
1267 	inflight_info->used_idx += inflight_info->desc[head].num;
1268 	if (inflight_info->used_idx >= inflight_info->desc_num) {
1269 		inflight_info->used_idx -= inflight_info->desc_num;
1270 		inflight_info->used_wrap_counter =
1271 			!inflight_info->used_wrap_counter;
1272 	}
1273 
1274 	return 0;
1275 }
1276 
1277 int
1278 rte_vhost_vring_call(int vid, uint16_t vring_idx)
1279 {
1280 	struct virtio_net *dev;
1281 	struct vhost_virtqueue *vq;
1282 
1283 	dev = get_device(vid);
1284 	if (!dev)
1285 		return -1;
1286 
1287 	if (vring_idx >= VHOST_MAX_VRING)
1288 		return -1;
1289 
1290 	vq = dev->virtqueue[vring_idx];
1291 	if (!vq)
1292 		return -1;
1293 
1294 	rte_spinlock_lock(&vq->access_lock);
1295 
1296 	if (vq_is_packed(dev))
1297 		vhost_vring_call_packed(dev, vq);
1298 	else
1299 		vhost_vring_call_split(dev, vq);
1300 
1301 	rte_spinlock_unlock(&vq->access_lock);
1302 
1303 	return 0;
1304 }
1305 
1306 uint16_t
1307 rte_vhost_avail_entries(int vid, uint16_t queue_id)
1308 {
1309 	struct virtio_net *dev;
1310 	struct vhost_virtqueue *vq;
1311 	uint16_t ret = 0;
1312 
1313 	dev = get_device(vid);
1314 	if (!dev)
1315 		return 0;
1316 
1317 	if (queue_id >= VHOST_MAX_VRING)
1318 		return 0;
1319 
1320 	vq = dev->virtqueue[queue_id];
1321 	if (!vq)
1322 		return 0;
1323 
1324 	rte_spinlock_lock(&vq->access_lock);
1325 
1326 	if (unlikely(!vq->enabled || vq->avail == NULL))
1327 		goto out;
1328 
1329 	ret = *(volatile uint16_t *)&vq->avail->idx - vq->last_used_idx;
1330 
1331 out:
1332 	rte_spinlock_unlock(&vq->access_lock);
1333 	return ret;
1334 }
1335 
1336 static inline int
1337 vhost_enable_notify_split(struct virtio_net *dev,
1338 		struct vhost_virtqueue *vq, int enable)
1339 {
1340 	if (vq->used == NULL)
1341 		return -1;
1342 
1343 	if (!(dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))) {
1344 		if (enable)
1345 			vq->used->flags &= ~VRING_USED_F_NO_NOTIFY;
1346 		else
1347 			vq->used->flags |= VRING_USED_F_NO_NOTIFY;
1348 	} else {
1349 		if (enable)
1350 			vhost_avail_event(vq) = vq->last_avail_idx;
1351 	}
1352 	return 0;
1353 }
1354 
1355 static inline int
1356 vhost_enable_notify_packed(struct virtio_net *dev,
1357 		struct vhost_virtqueue *vq, int enable)
1358 {
1359 	uint16_t flags;
1360 
1361 	if (vq->device_event == NULL)
1362 		return -1;
1363 
1364 	if (!enable) {
1365 		vq->device_event->flags = VRING_EVENT_F_DISABLE;
1366 		return 0;
1367 	}
1368 
1369 	flags = VRING_EVENT_F_ENABLE;
1370 	if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX)) {
1371 		flags = VRING_EVENT_F_DESC;
1372 		vq->device_event->off_wrap = vq->last_avail_idx |
1373 			vq->avail_wrap_counter << 15;
1374 	}
1375 
1376 	rte_atomic_thread_fence(__ATOMIC_RELEASE);
1377 
1378 	vq->device_event->flags = flags;
1379 	return 0;
1380 }
1381 
1382 int
1383 vhost_enable_guest_notification(struct virtio_net *dev,
1384 		struct vhost_virtqueue *vq, int enable)
1385 {
1386 	/*
1387 	 * If the virtqueue is not ready yet, it will be applied
1388 	 * when it will become ready.
1389 	 */
1390 	if (!vq->ready)
1391 		return 0;
1392 
1393 	if (vq_is_packed(dev))
1394 		return vhost_enable_notify_packed(dev, vq, enable);
1395 	else
1396 		return vhost_enable_notify_split(dev, vq, enable);
1397 }
1398 
1399 int
1400 rte_vhost_enable_guest_notification(int vid, uint16_t queue_id, int enable)
1401 {
1402 	struct virtio_net *dev = get_device(vid);
1403 	struct vhost_virtqueue *vq;
1404 	int ret;
1405 
1406 	if (!dev)
1407 		return -1;
1408 
1409 	if (queue_id >= VHOST_MAX_VRING)
1410 		return -1;
1411 
1412 	vq = dev->virtqueue[queue_id];
1413 	if (!vq)
1414 		return -1;
1415 
1416 	rte_spinlock_lock(&vq->access_lock);
1417 
1418 	vq->notif_enable = enable;
1419 	ret = vhost_enable_guest_notification(dev, vq, enable);
1420 
1421 	rte_spinlock_unlock(&vq->access_lock);
1422 
1423 	return ret;
1424 }
1425 
1426 void
1427 rte_vhost_log_write(int vid, uint64_t addr, uint64_t len)
1428 {
1429 	struct virtio_net *dev = get_device(vid);
1430 
1431 	if (dev == NULL)
1432 		return;
1433 
1434 	vhost_log_write(dev, addr, len);
1435 }
1436 
1437 void
1438 rte_vhost_log_used_vring(int vid, uint16_t vring_idx,
1439 			 uint64_t offset, uint64_t len)
1440 {
1441 	struct virtio_net *dev;
1442 	struct vhost_virtqueue *vq;
1443 
1444 	dev = get_device(vid);
1445 	if (dev == NULL)
1446 		return;
1447 
1448 	if (vring_idx >= VHOST_MAX_VRING)
1449 		return;
1450 	vq = dev->virtqueue[vring_idx];
1451 	if (!vq)
1452 		return;
1453 
1454 	vhost_log_used_vring(dev, vq, offset, len);
1455 }
1456 
1457 uint32_t
1458 rte_vhost_rx_queue_count(int vid, uint16_t qid)
1459 {
1460 	struct virtio_net *dev;
1461 	struct vhost_virtqueue *vq;
1462 	uint32_t ret = 0;
1463 
1464 	dev = get_device(vid);
1465 	if (dev == NULL)
1466 		return 0;
1467 
1468 	if (unlikely(qid >= dev->nr_vring || (qid & 1) == 0)) {
1469 		VHOST_LOG_DATA(ERR, "(%s) %s: invalid virtqueue idx %d.\n",
1470 			dev->ifname, __func__, qid);
1471 		return 0;
1472 	}
1473 
1474 	vq = dev->virtqueue[qid];
1475 	if (vq == NULL)
1476 		return 0;
1477 
1478 	rte_spinlock_lock(&vq->access_lock);
1479 
1480 	if (unlikely(!vq->enabled || vq->avail == NULL))
1481 		goto out;
1482 
1483 	ret = *((volatile uint16_t *)&vq->avail->idx) - vq->last_avail_idx;
1484 
1485 out:
1486 	rte_spinlock_unlock(&vq->access_lock);
1487 	return ret;
1488 }
1489 
1490 struct rte_vdpa_device *
1491 rte_vhost_get_vdpa_device(int vid)
1492 {
1493 	struct virtio_net *dev = get_device(vid);
1494 
1495 	if (dev == NULL)
1496 		return NULL;
1497 
1498 	return dev->vdpa_dev;
1499 }
1500 
1501 int
1502 rte_vhost_get_log_base(int vid, uint64_t *log_base,
1503 		uint64_t *log_size)
1504 {
1505 	struct virtio_net *dev = get_device(vid);
1506 
1507 	if (dev == NULL || log_base == NULL || log_size == NULL)
1508 		return -1;
1509 
1510 	*log_base = dev->log_base;
1511 	*log_size = dev->log_size;
1512 
1513 	return 0;
1514 }
1515 
1516 int
1517 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
1547 rte_vhost_set_vring_base(int vid, uint16_t queue_id,
1548 		uint16_t last_avail_idx, uint16_t last_used_idx)
1549 {
1550 	struct vhost_virtqueue *vq;
1551 	struct virtio_net *dev = get_device(vid);
1552 
1553 	if (!dev)
1554 		return -1;
1555 
1556 	if (queue_id >= VHOST_MAX_VRING)
1557 		return -1;
1558 
1559 	vq = dev->virtqueue[queue_id];
1560 	if (!vq)
1561 		return -1;
1562 
1563 	if (vq_is_packed(dev)) {
1564 		vq->last_avail_idx = last_avail_idx & 0x7fff;
1565 		vq->avail_wrap_counter = !!(last_avail_idx & (1 << 15));
1566 		vq->last_used_idx = last_used_idx & 0x7fff;
1567 		vq->used_wrap_counter = !!(last_used_idx & (1 << 15));
1568 	} else {
1569 		vq->last_avail_idx = last_avail_idx;
1570 		vq->last_used_idx = last_used_idx;
1571 	}
1572 
1573 	return 0;
1574 }
1575 
1576 int
1577 rte_vhost_get_vring_base_from_inflight(int vid,
1578 				       uint16_t queue_id,
1579 				       uint16_t *last_avail_idx,
1580 				       uint16_t *last_used_idx)
1581 {
1582 	struct rte_vhost_inflight_info_packed *inflight_info;
1583 	struct vhost_virtqueue *vq;
1584 	struct virtio_net *dev = get_device(vid);
1585 
1586 	if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1587 		return -1;
1588 
1589 	if (queue_id >= VHOST_MAX_VRING)
1590 		return -1;
1591 
1592 	vq = dev->virtqueue[queue_id];
1593 	if (!vq)
1594 		return -1;
1595 
1596 	if (!vq_is_packed(dev))
1597 		return -1;
1598 
1599 	inflight_info = vq->inflight_packed;
1600 	if (!inflight_info)
1601 		return -1;
1602 
1603 	*last_avail_idx = (inflight_info->old_used_wrap_counter << 15) |
1604 			  inflight_info->old_used_idx;
1605 	*last_used_idx = *last_avail_idx;
1606 
1607 	return 0;
1608 }
1609 
1610 int
1611 rte_vhost_extern_callback_register(int vid,
1612 		struct rte_vhost_user_extern_ops const * const ops, void *ctx)
1613 {
1614 	struct virtio_net *dev = get_device(vid);
1615 
1616 	if (dev == NULL || ops == NULL)
1617 		return -1;
1618 
1619 	dev->extern_ops = *ops;
1620 	dev->extern_data = ctx;
1621 	return 0;
1622 }
1623 
1624 static __rte_always_inline int
1625 async_channel_register(int vid, uint16_t queue_id)
1626 {
1627 	struct virtio_net *dev = get_device(vid);
1628 	struct vhost_virtqueue *vq = dev->virtqueue[queue_id];
1629 	struct vhost_async *async;
1630 	int node = vq->numa_node;
1631 
1632 	if (unlikely(vq->async)) {
1633 		VHOST_LOG_CONFIG(ERR,
1634 				"(%s) async register failed: already registered (qid: %d)\n",
1635 				dev->ifname, queue_id);
1636 		return -1;
1637 	}
1638 
1639 	async = rte_zmalloc_socket(NULL, sizeof(struct vhost_async), 0, node);
1640 	if (!async) {
1641 		VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate async metadata (qid: %d)\n",
1642 				dev->ifname, queue_id);
1643 		return -1;
1644 	}
1645 
1646 	async->pkts_info = rte_malloc_socket(NULL, vq->size * sizeof(struct async_inflight_info),
1647 			RTE_CACHE_LINE_SIZE, node);
1648 	if (!async->pkts_info) {
1649 		VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate async_pkts_info (qid: %d)\n",
1650 				dev->ifname, queue_id);
1651 		goto out_free_async;
1652 	}
1653 
1654 	async->pkts_cmpl_flag = rte_zmalloc_socket(NULL, vq->size * sizeof(bool),
1655 			RTE_CACHE_LINE_SIZE, node);
1656 	if (!async->pkts_cmpl_flag) {
1657 		VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate async pkts_cmpl_flag (qid: %d)\n",
1658 				dev->ifname, queue_id);
1659 		goto out_free_async;
1660 	}
1661 
1662 	if (vq_is_packed(dev)) {
1663 		async->buffers_packed = rte_malloc_socket(NULL,
1664 				vq->size * sizeof(struct vring_used_elem_packed),
1665 				RTE_CACHE_LINE_SIZE, node);
1666 		if (!async->buffers_packed) {
1667 			VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate async buffers (qid: %d)\n",
1668 					dev->ifname, queue_id);
1669 			goto out_free_inflight;
1670 		}
1671 	} else {
1672 		async->descs_split = rte_malloc_socket(NULL,
1673 				vq->size * sizeof(struct vring_used_elem),
1674 				RTE_CACHE_LINE_SIZE, node);
1675 		if (!async->descs_split) {
1676 			VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate async descs (qid: %d)\n",
1677 					dev->ifname, queue_id);
1678 			goto out_free_inflight;
1679 		}
1680 	}
1681 
1682 	vq->async = async;
1683 
1684 	return 0;
1685 out_free_inflight:
1686 	rte_free(async->pkts_info);
1687 out_free_async:
1688 	rte_free(async);
1689 
1690 	return -1;
1691 }
1692 
1693 int
1694 rte_vhost_async_channel_register(int vid, uint16_t queue_id)
1695 {
1696 	struct vhost_virtqueue *vq;
1697 	struct virtio_net *dev = get_device(vid);
1698 	int ret;
1699 
1700 	if (dev == NULL)
1701 		return -1;
1702 
1703 	if (queue_id >= VHOST_MAX_VRING)
1704 		return -1;
1705 
1706 	vq = dev->virtqueue[queue_id];
1707 
1708 	if (unlikely(vq == NULL || !dev->async_copy))
1709 		return -1;
1710 
1711 	rte_spinlock_lock(&vq->access_lock);
1712 	ret = async_channel_register(vid, queue_id);
1713 	rte_spinlock_unlock(&vq->access_lock);
1714 
1715 	return ret;
1716 }
1717 
1718 int
1719 rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t queue_id)
1720 {
1721 	struct vhost_virtqueue *vq;
1722 	struct virtio_net *dev = get_device(vid);
1723 
1724 	if (dev == NULL)
1725 		return -1;
1726 
1727 	if (queue_id >= VHOST_MAX_VRING)
1728 		return -1;
1729 
1730 	vq = dev->virtqueue[queue_id];
1731 
1732 	if (unlikely(vq == NULL || !dev->async_copy))
1733 		return -1;
1734 
1735 	return async_channel_register(vid, queue_id);
1736 }
1737 
1738 int
1739 rte_vhost_async_channel_unregister(int vid, uint16_t queue_id)
1740 {
1741 	struct vhost_virtqueue *vq;
1742 	struct virtio_net *dev = get_device(vid);
1743 	int ret = -1;
1744 
1745 	if (dev == NULL)
1746 		return ret;
1747 
1748 	if (queue_id >= VHOST_MAX_VRING)
1749 		return ret;
1750 
1751 	vq = dev->virtqueue[queue_id];
1752 
1753 	if (vq == NULL)
1754 		return ret;
1755 
1756 	ret = 0;
1757 
1758 	if (!vq->async)
1759 		return ret;
1760 
1761 	if (!rte_spinlock_trylock(&vq->access_lock)) {
1762 		VHOST_LOG_CONFIG(ERR, "(%s) failed to unregister async channel, virtqueue busy.\n",
1763 				dev->ifname);
1764 		return -1;
1765 	}
1766 
1767 	if (vq->async->pkts_inflight_n) {
1768 		VHOST_LOG_CONFIG(ERR, "(%s) failed to unregister async channel.\n", dev->ifname);
1769 		VHOST_LOG_CONFIG(ERR, "(%s) inflight packets must be completed before unregistration.\n",
1770 			dev->ifname);
1771 		ret = -1;
1772 		goto out;
1773 	}
1774 
1775 	vhost_free_async_mem(vq);
1776 out:
1777 	rte_spinlock_unlock(&vq->access_lock);
1778 
1779 	return ret;
1780 }
1781 
1782 int
1783 rte_vhost_async_channel_unregister_thread_unsafe(int vid, uint16_t queue_id)
1784 {
1785 	struct vhost_virtqueue *vq;
1786 	struct virtio_net *dev = get_device(vid);
1787 
1788 	if (dev == NULL)
1789 		return -1;
1790 
1791 	if (queue_id >= VHOST_MAX_VRING)
1792 		return -1;
1793 
1794 	vq = dev->virtqueue[queue_id];
1795 
1796 	if (vq == NULL)
1797 		return -1;
1798 
1799 	if (!vq->async)
1800 		return 0;
1801 
1802 	if (vq->async->pkts_inflight_n) {
1803 		VHOST_LOG_CONFIG(ERR, "(%s) failed to unregister async channel.\n", dev->ifname);
1804 		VHOST_LOG_CONFIG(ERR, "(%s) inflight packets must be completed before unregistration.\n",
1805 			dev->ifname);
1806 		return -1;
1807 	}
1808 
1809 	vhost_free_async_mem(vq);
1810 
1811 	return 0;
1812 }
1813 
1814 int
1815 rte_vhost_async_dma_configure(int16_t dma_id, uint16_t vchan_id)
1816 {
1817 	struct rte_dma_info info;
1818 	void *pkts_cmpl_flag_addr;
1819 	uint16_t max_desc;
1820 
1821 	if (!rte_dma_is_valid(dma_id)) {
1822 		VHOST_LOG_CONFIG(ERR, "DMA %d is not found.\n", dma_id);
1823 		return -1;
1824 	}
1825 
1826 	rte_dma_info_get(dma_id, &info);
1827 	if (vchan_id >= info.max_vchans) {
1828 		VHOST_LOG_CONFIG(ERR, "Invalid DMA %d vChannel %u.\n", dma_id, vchan_id);
1829 		return -1;
1830 	}
1831 
1832 	if (!dma_copy_track[dma_id].vchans) {
1833 		struct async_dma_vchan_info *vchans;
1834 
1835 		vchans = rte_zmalloc(NULL, sizeof(struct async_dma_vchan_info) * info.max_vchans,
1836 				RTE_CACHE_LINE_SIZE);
1837 		if (vchans == NULL) {
1838 			VHOST_LOG_CONFIG(ERR, "Failed to allocate vchans for DMA %d vChannel %u.\n",
1839 					dma_id, vchan_id);
1840 			return -1;
1841 		}
1842 
1843 		dma_copy_track[dma_id].vchans = vchans;
1844 	}
1845 
1846 	if (dma_copy_track[dma_id].vchans[vchan_id].pkts_cmpl_flag_addr) {
1847 		VHOST_LOG_CONFIG(INFO, "DMA %d vChannel %u already registered.\n", dma_id,
1848 				vchan_id);
1849 		return 0;
1850 	}
1851 
1852 	max_desc = info.max_desc;
1853 	if (!rte_is_power_of_2(max_desc))
1854 		max_desc = rte_align32pow2(max_desc);
1855 
1856 	pkts_cmpl_flag_addr = rte_zmalloc(NULL, sizeof(bool *) * max_desc, RTE_CACHE_LINE_SIZE);
1857 	if (!pkts_cmpl_flag_addr) {
1858 		VHOST_LOG_CONFIG(ERR, "Failed to allocate pkts_cmpl_flag_addr for DMA %d "
1859 				"vChannel %u.\n", dma_id, vchan_id);
1860 
1861 		if (dma_copy_track[dma_id].nr_vchans == 0) {
1862 			rte_free(dma_copy_track[dma_id].vchans);
1863 			dma_copy_track[dma_id].vchans = NULL;
1864 		}
1865 		return -1;
1866 	}
1867 
1868 	dma_copy_track[dma_id].vchans[vchan_id].pkts_cmpl_flag_addr = pkts_cmpl_flag_addr;
1869 	dma_copy_track[dma_id].vchans[vchan_id].ring_size = max_desc;
1870 	dma_copy_track[dma_id].vchans[vchan_id].ring_mask = max_desc - 1;
1871 	dma_copy_track[dma_id].nr_vchans++;
1872 
1873 	return 0;
1874 }
1875 
1876 int
1877 rte_vhost_async_get_inflight(int vid, uint16_t queue_id)
1878 {
1879 	struct vhost_virtqueue *vq;
1880 	struct virtio_net *dev = get_device(vid);
1881 	int ret = -1;
1882 
1883 	if (dev == NULL)
1884 		return ret;
1885 
1886 	if (queue_id >= VHOST_MAX_VRING)
1887 		return ret;
1888 
1889 	vq = dev->virtqueue[queue_id];
1890 
1891 	if (vq == NULL)
1892 		return ret;
1893 
1894 	if (!vq->async)
1895 		return ret;
1896 
1897 	if (!rte_spinlock_trylock(&vq->access_lock)) {
1898 		VHOST_LOG_CONFIG(DEBUG,
1899 			"(%s) failed to check in-flight packets. virtqueue busy.\n",
1900 			dev->ifname);
1901 		return ret;
1902 	}
1903 
1904 	ret = vq->async->pkts_inflight_n;
1905 	rte_spinlock_unlock(&vq->access_lock);
1906 
1907 	return ret;
1908 }
1909 
1910 int
1911 rte_vhost_async_get_inflight_thread_unsafe(int vid, uint16_t queue_id)
1912 {
1913 	struct vhost_virtqueue *vq;
1914 	struct virtio_net *dev = get_device(vid);
1915 	int ret = -1;
1916 
1917 	if (dev == NULL)
1918 		return ret;
1919 
1920 	if (queue_id >= VHOST_MAX_VRING)
1921 		return ret;
1922 
1923 	vq = dev->virtqueue[queue_id];
1924 
1925 	if (vq == NULL)
1926 		return ret;
1927 
1928 	if (!vq->async)
1929 		return ret;
1930 
1931 	ret = vq->async->pkts_inflight_n;
1932 
1933 	return ret;
1934 }
1935 
1936 int
1937 rte_vhost_get_monitor_addr(int vid, uint16_t queue_id,
1938 		struct rte_vhost_power_monitor_cond *pmc)
1939 {
1940 	struct virtio_net *dev = get_device(vid);
1941 	struct vhost_virtqueue *vq;
1942 
1943 	if (dev == NULL)
1944 		return -1;
1945 	if (queue_id >= VHOST_MAX_VRING)
1946 		return -1;
1947 
1948 	vq = dev->virtqueue[queue_id];
1949 	if (vq == NULL)
1950 		return -1;
1951 
1952 	if (vq_is_packed(dev)) {
1953 		struct vring_packed_desc *desc;
1954 		desc = vq->desc_packed;
1955 		pmc->addr = &desc[vq->last_avail_idx].flags;
1956 		if (vq->avail_wrap_counter)
1957 			pmc->val = VRING_DESC_F_AVAIL;
1958 		else
1959 			pmc->val = VRING_DESC_F_USED;
1960 		pmc->mask = VRING_DESC_F_AVAIL | VRING_DESC_F_USED;
1961 		pmc->size = sizeof(desc[vq->last_avail_idx].flags);
1962 		pmc->match = 1;
1963 	} else {
1964 		pmc->addr = &vq->avail->idx;
1965 		pmc->val = vq->last_avail_idx & (vq->size - 1);
1966 		pmc->mask = vq->size - 1;
1967 		pmc->size = sizeof(vq->avail->idx);
1968 		pmc->match = 0;
1969 	}
1970 
1971 	return 0;
1972 }
1973 
1974 RTE_LOG_REGISTER_SUFFIX(vhost_config_log_level, config, INFO);
1975 RTE_LOG_REGISTER_SUFFIX(vhost_data_log_level, data, WARNING);
1976