xref: /dpdk/drivers/net/virtio/virtio_user/virtio_user_dev.c (revision 10b71caecbe1cddcbb65c050ca775fba575e88db)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2010-2016 Intel Corporation
3  */
4 
5 #include <stdint.h>
6 #include <stdio.h>
7 #include <fcntl.h>
8 #include <string.h>
9 #include <errno.h>
10 #include <sys/mman.h>
11 #include <unistd.h>
12 #include <sys/eventfd.h>
13 #include <sys/types.h>
14 #include <sys/stat.h>
15 
16 #include <rte_string_fns.h>
17 #include <rte_eal_memconfig.h>
18 
19 #include "vhost.h"
20 #include "virtio_user_dev.h"
21 #include "../virtio_ethdev.h"
22 
23 #define VIRTIO_USER_MEM_EVENT_CLB_NAME "virtio_user_mem_event_clb"
24 
25 static int
26 virtio_user_create_queue(struct virtio_user_dev *dev, uint32_t queue_sel)
27 {
28 	/* Of all per virtqueue MSGs, make sure VHOST_SET_VRING_CALL come
29 	 * firstly because vhost depends on this msg to allocate virtqueue
30 	 * pair.
31 	 */
32 	struct vhost_vring_file file;
33 
34 	file.index = queue_sel;
35 	file.fd = dev->callfds[queue_sel];
36 	dev->ops->send_request(dev, VHOST_USER_SET_VRING_CALL, &file);
37 
38 	return 0;
39 }
40 
41 static int
42 virtio_user_kick_queue(struct virtio_user_dev *dev, uint32_t queue_sel)
43 {
44 	struct vhost_vring_file file;
45 	struct vhost_vring_state state;
46 	struct vring *vring = &dev->vrings[queue_sel];
47 	struct vring_packed *pq_vring = &dev->packed_vrings[queue_sel];
48 	struct vhost_vring_addr addr = {
49 		.index = queue_sel,
50 		.log_guest_addr = 0,
51 		.flags = 0, /* disable log */
52 	};
53 
54 	if (dev->features & (1ULL << VIRTIO_F_RING_PACKED)) {
55 		addr.desc_user_addr =
56 			(uint64_t)(uintptr_t)pq_vring->desc;
57 		addr.avail_user_addr =
58 			(uint64_t)(uintptr_t)pq_vring->driver;
59 		addr.used_user_addr =
60 			(uint64_t)(uintptr_t)pq_vring->device;
61 	} else {
62 		addr.desc_user_addr = (uint64_t)(uintptr_t)vring->desc;
63 		addr.avail_user_addr = (uint64_t)(uintptr_t)vring->avail;
64 		addr.used_user_addr = (uint64_t)(uintptr_t)vring->used;
65 	}
66 
67 	state.index = queue_sel;
68 	state.num = vring->num;
69 	dev->ops->send_request(dev, VHOST_USER_SET_VRING_NUM, &state);
70 
71 	state.index = queue_sel;
72 	state.num = 0; /* no reservation */
73 	if (dev->features & (1ULL << VIRTIO_F_RING_PACKED))
74 		state.num |= (1 << 15);
75 	dev->ops->send_request(dev, VHOST_USER_SET_VRING_BASE, &state);
76 
77 	dev->ops->send_request(dev, VHOST_USER_SET_VRING_ADDR, &addr);
78 
79 	/* Of all per virtqueue MSGs, make sure VHOST_USER_SET_VRING_KICK comes
80 	 * lastly because vhost depends on this msg to judge if
81 	 * virtio is ready.
82 	 */
83 	file.index = queue_sel;
84 	file.fd = dev->kickfds[queue_sel];
85 	dev->ops->send_request(dev, VHOST_USER_SET_VRING_KICK, &file);
86 
87 	return 0;
88 }
89 
90 static int
91 virtio_user_queue_setup(struct virtio_user_dev *dev,
92 			int (*fn)(struct virtio_user_dev *, uint32_t))
93 {
94 	uint32_t i, queue_sel;
95 
96 	for (i = 0; i < dev->max_queue_pairs; ++i) {
97 		queue_sel = 2 * i + VTNET_SQ_RQ_QUEUE_IDX;
98 		if (fn(dev, queue_sel) < 0) {
99 			PMD_DRV_LOG(INFO, "setup rx vq fails: %u", i);
100 			return -1;
101 		}
102 	}
103 	for (i = 0; i < dev->max_queue_pairs; ++i) {
104 		queue_sel = 2 * i + VTNET_SQ_TQ_QUEUE_IDX;
105 		if (fn(dev, queue_sel) < 0) {
106 			PMD_DRV_LOG(INFO, "setup tx vq fails: %u", i);
107 			return -1;
108 		}
109 	}
110 
111 	return 0;
112 }
113 
114 int
115 is_vhost_user_by_type(const char *path)
116 {
117 	struct stat sb;
118 
119 	if (stat(path, &sb) == -1)
120 		return 0;
121 
122 	return S_ISSOCK(sb.st_mode);
123 }
124 
125 int
126 virtio_user_start_device(struct virtio_user_dev *dev)
127 {
128 	uint64_t features;
129 	int ret;
130 
131 	/*
132 	 * XXX workaround!
133 	 *
134 	 * We need to make sure that the locks will be
135 	 * taken in the correct order to avoid deadlocks.
136 	 *
137 	 * Before releasing this lock, this thread should
138 	 * not trigger any memory hotplug events.
139 	 *
140 	 * This is a temporary workaround, and should be
141 	 * replaced when we get proper supports from the
142 	 * memory subsystem in the future.
143 	 */
144 	rte_mcfg_mem_read_lock();
145 	pthread_mutex_lock(&dev->mutex);
146 
147 	if (is_vhost_user_by_type(dev->path) && dev->vhostfd < 0)
148 		goto error;
149 
150 	/* Step 0: tell vhost to create queues */
151 	if (virtio_user_queue_setup(dev, virtio_user_create_queue) < 0)
152 		goto error;
153 
154 	/* Step 1: negotiate protocol features & set features */
155 	features = dev->features;
156 
157 
158 	/* Strip VIRTIO_NET_F_MAC, as MAC address is handled in vdev init */
159 	features &= ~(1ull << VIRTIO_NET_F_MAC);
160 	/* Strip VIRTIO_NET_F_CTRL_VQ, as devices do not really need to know */
161 	features &= ~(1ull << VIRTIO_NET_F_CTRL_VQ);
162 	features &= ~(1ull << VIRTIO_NET_F_STATUS);
163 	ret = dev->ops->send_request(dev, VHOST_USER_SET_FEATURES, &features);
164 	if (ret < 0)
165 		goto error;
166 	PMD_DRV_LOG(INFO, "set features: %" PRIx64, features);
167 
168 	/* Step 2: share memory regions */
169 	ret = dev->ops->send_request(dev, VHOST_USER_SET_MEM_TABLE, NULL);
170 	if (ret < 0)
171 		goto error;
172 
173 	/* Step 3: kick queues */
174 	if (virtio_user_queue_setup(dev, virtio_user_kick_queue) < 0)
175 		goto error;
176 
177 	/* Step 4: enable queues
178 	 * we enable the 1st queue pair by default.
179 	 */
180 	dev->ops->enable_qp(dev, 0, 1);
181 
182 	dev->started = true;
183 	pthread_mutex_unlock(&dev->mutex);
184 	rte_mcfg_mem_read_unlock();
185 
186 	return 0;
187 error:
188 	pthread_mutex_unlock(&dev->mutex);
189 	rte_mcfg_mem_read_unlock();
190 	/* TODO: free resource here or caller to check */
191 	return -1;
192 }
193 
194 int virtio_user_stop_device(struct virtio_user_dev *dev)
195 {
196 	struct vhost_vring_state state;
197 	uint32_t i;
198 	int error = 0;
199 
200 	pthread_mutex_lock(&dev->mutex);
201 	if (!dev->started)
202 		goto out;
203 
204 	for (i = 0; i < dev->max_queue_pairs; ++i)
205 		dev->ops->enable_qp(dev, i, 0);
206 
207 	/* Stop the backend. */
208 	for (i = 0; i < dev->max_queue_pairs * 2; ++i) {
209 		state.index = i;
210 		if (dev->ops->send_request(dev, VHOST_USER_GET_VRING_BASE,
211 					   &state) < 0) {
212 			PMD_DRV_LOG(ERR, "get_vring_base failed, index=%u\n",
213 				    i);
214 			error = -1;
215 			goto out;
216 		}
217 	}
218 
219 	dev->started = false;
220 out:
221 	pthread_mutex_unlock(&dev->mutex);
222 
223 	return error;
224 }
225 
226 static inline void
227 parse_mac(struct virtio_user_dev *dev, const char *mac)
228 {
229 	struct rte_ether_addr tmp;
230 
231 	if (!mac)
232 		return;
233 
234 	if (rte_ether_unformat_addr(mac, &tmp) == 0) {
235 		memcpy(dev->mac_addr, &tmp, RTE_ETHER_ADDR_LEN);
236 		dev->mac_specified = 1;
237 	} else {
238 		/* ignore the wrong mac, use random mac */
239 		PMD_DRV_LOG(ERR, "wrong format of mac: %s", mac);
240 	}
241 }
242 
243 static int
244 virtio_user_dev_init_notify(struct virtio_user_dev *dev)
245 {
246 	uint32_t i, j;
247 	int callfd;
248 	int kickfd;
249 
250 	for (i = 0; i < VIRTIO_MAX_VIRTQUEUES; ++i) {
251 		if (i >= dev->max_queue_pairs * 2) {
252 			dev->kickfds[i] = -1;
253 			dev->callfds[i] = -1;
254 			continue;
255 		}
256 
257 		/* May use invalid flag, but some backend uses kickfd and
258 		 * callfd as criteria to judge if dev is alive. so finally we
259 		 * use real event_fd.
260 		 */
261 		callfd = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK);
262 		if (callfd < 0) {
263 			PMD_DRV_LOG(ERR, "callfd error, %s", strerror(errno));
264 			break;
265 		}
266 		kickfd = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK);
267 		if (kickfd < 0) {
268 			PMD_DRV_LOG(ERR, "kickfd error, %s", strerror(errno));
269 			break;
270 		}
271 		dev->callfds[i] = callfd;
272 		dev->kickfds[i] = kickfd;
273 	}
274 
275 	if (i < VIRTIO_MAX_VIRTQUEUES) {
276 		for (j = 0; j <= i; ++j) {
277 			close(dev->callfds[j]);
278 			close(dev->kickfds[j]);
279 		}
280 
281 		return -1;
282 	}
283 
284 	return 0;
285 }
286 
287 static int
288 virtio_user_fill_intr_handle(struct virtio_user_dev *dev)
289 {
290 	uint32_t i;
291 	struct rte_eth_dev *eth_dev = &rte_eth_devices[dev->port_id];
292 
293 	if (!eth_dev->intr_handle) {
294 		eth_dev->intr_handle = malloc(sizeof(*eth_dev->intr_handle));
295 		if (!eth_dev->intr_handle) {
296 			PMD_DRV_LOG(ERR, "fail to allocate intr_handle");
297 			return -1;
298 		}
299 		memset(eth_dev->intr_handle, 0, sizeof(*eth_dev->intr_handle));
300 	}
301 
302 	for (i = 0; i < dev->max_queue_pairs; ++i)
303 		eth_dev->intr_handle->efds[i] = dev->callfds[i];
304 	eth_dev->intr_handle->nb_efd = dev->max_queue_pairs;
305 	eth_dev->intr_handle->max_intr = dev->max_queue_pairs + 1;
306 	eth_dev->intr_handle->type = RTE_INTR_HANDLE_VDEV;
307 	/* For virtio vdev, no need to read counter for clean */
308 	eth_dev->intr_handle->efd_counter_size = 0;
309 	eth_dev->intr_handle->fd = -1;
310 	if (dev->vhostfd >= 0)
311 		eth_dev->intr_handle->fd = dev->vhostfd;
312 	else if (dev->is_server)
313 		eth_dev->intr_handle->fd = dev->listenfd;
314 
315 	return 0;
316 }
317 
318 static void
319 virtio_user_mem_event_cb(enum rte_mem_event type __rte_unused,
320 			 const void *addr,
321 			 size_t len __rte_unused,
322 			 void *arg)
323 {
324 	struct virtio_user_dev *dev = arg;
325 	struct rte_memseg_list *msl;
326 	uint16_t i;
327 
328 	/* ignore externally allocated memory */
329 	msl = rte_mem_virt2memseg_list(addr);
330 	if (msl->external)
331 		return;
332 
333 	pthread_mutex_lock(&dev->mutex);
334 
335 	if (dev->started == false)
336 		goto exit;
337 
338 	/* Step 1: pause the active queues */
339 	for (i = 0; i < dev->queue_pairs; i++)
340 		dev->ops->enable_qp(dev, i, 0);
341 
342 	/* Step 2: update memory regions */
343 	dev->ops->send_request(dev, VHOST_USER_SET_MEM_TABLE, NULL);
344 
345 	/* Step 3: resume the active queues */
346 	for (i = 0; i < dev->queue_pairs; i++)
347 		dev->ops->enable_qp(dev, i, 1);
348 
349 exit:
350 	pthread_mutex_unlock(&dev->mutex);
351 }
352 
353 static int
354 virtio_user_dev_setup(struct virtio_user_dev *dev)
355 {
356 	uint32_t q;
357 
358 	dev->vhostfd = -1;
359 	dev->vhostfds = NULL;
360 	dev->tapfds = NULL;
361 
362 	if (dev->is_server) {
363 		if (access(dev->path, F_OK) == 0 &&
364 		    !is_vhost_user_by_type(dev->path)) {
365 			PMD_DRV_LOG(ERR, "Server mode doesn't support vhost-kernel!");
366 			return -1;
367 		}
368 		dev->ops = &virtio_ops_user;
369 	} else {
370 		if (is_vhost_user_by_type(dev->path)) {
371 			dev->ops = &virtio_ops_user;
372 		} else {
373 			dev->ops = &virtio_ops_kernel;
374 
375 			dev->vhostfds = malloc(dev->max_queue_pairs *
376 					       sizeof(int));
377 			dev->tapfds = malloc(dev->max_queue_pairs *
378 					     sizeof(int));
379 			if (!dev->vhostfds || !dev->tapfds) {
380 				PMD_INIT_LOG(ERR, "Failed to malloc");
381 				return -1;
382 			}
383 
384 			for (q = 0; q < dev->max_queue_pairs; ++q) {
385 				dev->vhostfds[q] = -1;
386 				dev->tapfds[q] = -1;
387 			}
388 		}
389 	}
390 
391 	if (dev->ops->setup(dev) < 0)
392 		return -1;
393 
394 	if (virtio_user_dev_init_notify(dev) < 0)
395 		return -1;
396 
397 	if (virtio_user_fill_intr_handle(dev) < 0)
398 		return -1;
399 
400 	return 0;
401 }
402 
403 /* Use below macro to filter features from vhost backend */
404 #define VIRTIO_USER_SUPPORTED_FEATURES			\
405 	(1ULL << VIRTIO_NET_F_MAC		|	\
406 	 1ULL << VIRTIO_NET_F_STATUS		|	\
407 	 1ULL << VIRTIO_NET_F_MQ		|	\
408 	 1ULL << VIRTIO_NET_F_CTRL_MAC_ADDR	|	\
409 	 1ULL << VIRTIO_NET_F_CTRL_VQ		|	\
410 	 1ULL << VIRTIO_NET_F_CTRL_RX		|	\
411 	 1ULL << VIRTIO_NET_F_CTRL_VLAN		|	\
412 	 1ULL << VIRTIO_NET_F_CSUM		|	\
413 	 1ULL << VIRTIO_NET_F_HOST_TSO4		|	\
414 	 1ULL << VIRTIO_NET_F_HOST_TSO6		|	\
415 	 1ULL << VIRTIO_NET_F_MRG_RXBUF		|	\
416 	 1ULL << VIRTIO_RING_F_INDIRECT_DESC	|	\
417 	 1ULL << VIRTIO_NET_F_GUEST_CSUM	|	\
418 	 1ULL << VIRTIO_NET_F_GUEST_TSO4	|	\
419 	 1ULL << VIRTIO_NET_F_GUEST_TSO6	|	\
420 	 1ULL << VIRTIO_F_IN_ORDER		|	\
421 	 1ULL << VIRTIO_F_VERSION_1		|	\
422 	 1ULL << VIRTIO_F_RING_PACKED		|	\
423 	 1ULL << VHOST_USER_F_PROTOCOL_FEATURES)
424 
425 #define VIRTIO_USER_SUPPORTED_PROTOCOL_FEATURES		\
426 	(1ULL << VHOST_USER_PROTOCOL_F_MQ |		\
427 	 1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK)
428 
429 int
430 virtio_user_dev_init(struct virtio_user_dev *dev, char *path, int queues,
431 		     int cq, int queue_size, const char *mac, char **ifname,
432 		     int server, int mrg_rxbuf, int in_order, int packed_vq)
433 {
434 	uint64_t protocol_features = 0;
435 
436 	pthread_mutex_init(&dev->mutex, NULL);
437 	strlcpy(dev->path, path, PATH_MAX);
438 	dev->started = 0;
439 	dev->max_queue_pairs = queues;
440 	dev->queue_pairs = 1; /* mq disabled by default */
441 	dev->queue_size = queue_size;
442 	dev->is_server = server;
443 	dev->mac_specified = 0;
444 	dev->frontend_features = 0;
445 	dev->unsupported_features = ~VIRTIO_USER_SUPPORTED_FEATURES;
446 	dev->protocol_features = VIRTIO_USER_SUPPORTED_PROTOCOL_FEATURES;
447 	parse_mac(dev, mac);
448 
449 	if (*ifname) {
450 		dev->ifname = *ifname;
451 		*ifname = NULL;
452 	}
453 
454 	if (virtio_user_dev_setup(dev) < 0) {
455 		PMD_INIT_LOG(ERR, "backend set up fails");
456 		return -1;
457 	}
458 
459 	if (!is_vhost_user_by_type(dev->path))
460 		dev->unsupported_features |=
461 			(1ULL << VHOST_USER_F_PROTOCOL_FEATURES);
462 
463 	if (!dev->is_server) {
464 		if (dev->ops->send_request(dev, VHOST_USER_SET_OWNER,
465 					   NULL) < 0) {
466 			PMD_INIT_LOG(ERR, "set_owner fails: %s",
467 				     strerror(errno));
468 			return -1;
469 		}
470 
471 		if (dev->ops->send_request(dev, VHOST_USER_GET_FEATURES,
472 					   &dev->device_features) < 0) {
473 			PMD_INIT_LOG(ERR, "get_features failed: %s",
474 				     strerror(errno));
475 			return -1;
476 		}
477 
478 
479 		if (dev->device_features &
480 				(1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) {
481 			if (dev->ops->send_request(dev,
482 					VHOST_USER_GET_PROTOCOL_FEATURES,
483 					&protocol_features))
484 				return -1;
485 
486 			dev->protocol_features &= protocol_features;
487 
488 			if (dev->ops->send_request(dev,
489 					VHOST_USER_SET_PROTOCOL_FEATURES,
490 					&dev->protocol_features))
491 				return -1;
492 
493 			if (!(dev->protocol_features &
494 					(1ULL << VHOST_USER_PROTOCOL_F_MQ)))
495 				dev->unsupported_features |=
496 					(1ull << VIRTIO_NET_F_MQ);
497 		}
498 	} else {
499 		/* We just pretend vhost-user can support all these features.
500 		 * Note that this could be problematic that if some feature is
501 		 * negotiated but not supported by the vhost-user which comes
502 		 * later.
503 		 */
504 		dev->device_features = VIRTIO_USER_SUPPORTED_FEATURES;
505 	}
506 
507 
508 
509 	if (!mrg_rxbuf)
510 		dev->unsupported_features |= (1ull << VIRTIO_NET_F_MRG_RXBUF);
511 
512 	if (!in_order)
513 		dev->unsupported_features |= (1ull << VIRTIO_F_IN_ORDER);
514 
515 	if (!packed_vq)
516 		dev->unsupported_features |= (1ull << VIRTIO_F_RING_PACKED);
517 
518 	if (dev->mac_specified)
519 		dev->frontend_features |= (1ull << VIRTIO_NET_F_MAC);
520 	else
521 		dev->unsupported_features |= (1ull << VIRTIO_NET_F_MAC);
522 
523 	if (cq) {
524 		/* device does not really need to know anything about CQ,
525 		 * so if necessary, we just claim to support CQ
526 		 */
527 		dev->frontend_features |= (1ull << VIRTIO_NET_F_CTRL_VQ);
528 	} else {
529 		dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_VQ);
530 		/* Also disable features that depend on VIRTIO_NET_F_CTRL_VQ */
531 		dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_RX);
532 		dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_VLAN);
533 		dev->unsupported_features |=
534 			(1ull << VIRTIO_NET_F_GUEST_ANNOUNCE);
535 		dev->unsupported_features |= (1ull << VIRTIO_NET_F_MQ);
536 		dev->unsupported_features |=
537 			(1ull << VIRTIO_NET_F_CTRL_MAC_ADDR);
538 	}
539 
540 	/* The backend will not report this feature, we add it explicitly */
541 	if (is_vhost_user_by_type(dev->path))
542 		dev->frontend_features |= (1ull << VIRTIO_NET_F_STATUS);
543 
544 	/*
545 	 * Device features =
546 	 *     (frontend_features | backend_features) & ~unsupported_features;
547 	 */
548 	dev->device_features |= dev->frontend_features;
549 	dev->device_features &= ~dev->unsupported_features;
550 
551 	if (rte_mem_event_callback_register(VIRTIO_USER_MEM_EVENT_CLB_NAME,
552 				virtio_user_mem_event_cb, dev)) {
553 		if (rte_errno != ENOTSUP) {
554 			PMD_INIT_LOG(ERR, "Failed to register mem event"
555 					" callback\n");
556 			return -1;
557 		}
558 	}
559 
560 	return 0;
561 }
562 
563 void
564 virtio_user_dev_uninit(struct virtio_user_dev *dev)
565 {
566 	uint32_t i;
567 
568 	virtio_user_stop_device(dev);
569 
570 	rte_mem_event_callback_unregister(VIRTIO_USER_MEM_EVENT_CLB_NAME, dev);
571 
572 	for (i = 0; i < dev->max_queue_pairs * 2; ++i) {
573 		close(dev->callfds[i]);
574 		close(dev->kickfds[i]);
575 	}
576 
577 	if (dev->vhostfd >= 0)
578 		close(dev->vhostfd);
579 
580 	if (dev->is_server && dev->listenfd >= 0) {
581 		close(dev->listenfd);
582 		dev->listenfd = -1;
583 	}
584 
585 	if (dev->vhostfds) {
586 		for (i = 0; i < dev->max_queue_pairs; ++i) {
587 			close(dev->vhostfds[i]);
588 			if (dev->tapfds[i] >= 0)
589 				close(dev->tapfds[i]);
590 		}
591 		free(dev->vhostfds);
592 		free(dev->tapfds);
593 	}
594 
595 	free(dev->ifname);
596 
597 	if (dev->is_server)
598 		unlink(dev->path);
599 }
600 
601 uint8_t
602 virtio_user_handle_mq(struct virtio_user_dev *dev, uint16_t q_pairs)
603 {
604 	uint16_t i;
605 	uint8_t ret = 0;
606 
607 	if (q_pairs > dev->max_queue_pairs) {
608 		PMD_INIT_LOG(ERR, "multi-q config %u, but only %u supported",
609 			     q_pairs, dev->max_queue_pairs);
610 		return -1;
611 	}
612 
613 	/* Server mode can't enable queue pairs if vhostfd is invalid,
614 	 * always return 0 in this case.
615 	 */
616 	if (!dev->is_server || dev->vhostfd >= 0) {
617 		for (i = 0; i < q_pairs; ++i)
618 			ret |= dev->ops->enable_qp(dev, i, 1);
619 		for (i = q_pairs; i < dev->max_queue_pairs; ++i)
620 			ret |= dev->ops->enable_qp(dev, i, 0);
621 	}
622 	dev->queue_pairs = q_pairs;
623 
624 	return ret;
625 }
626 
627 static uint32_t
628 virtio_user_handle_ctrl_msg(struct virtio_user_dev *dev, struct vring *vring,
629 			    uint16_t idx_hdr)
630 {
631 	struct virtio_net_ctrl_hdr *hdr;
632 	virtio_net_ctrl_ack status = ~0;
633 	uint16_t i, idx_data, idx_status;
634 	uint32_t n_descs = 0;
635 
636 	/* locate desc for header, data, and status */
637 	idx_data = vring->desc[idx_hdr].next;
638 	n_descs++;
639 
640 	i = idx_data;
641 	while (vring->desc[i].flags == VRING_DESC_F_NEXT) {
642 		i = vring->desc[i].next;
643 		n_descs++;
644 	}
645 
646 	/* locate desc for status */
647 	idx_status = i;
648 	n_descs++;
649 
650 	hdr = (void *)(uintptr_t)vring->desc[idx_hdr].addr;
651 	if (hdr->class == VIRTIO_NET_CTRL_MQ &&
652 	    hdr->cmd == VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET) {
653 		uint16_t queues;
654 
655 		queues = *(uint16_t *)(uintptr_t)vring->desc[idx_data].addr;
656 		status = virtio_user_handle_mq(dev, queues);
657 	} else if (hdr->class == VIRTIO_NET_CTRL_RX  ||
658 		   hdr->class == VIRTIO_NET_CTRL_MAC ||
659 		   hdr->class == VIRTIO_NET_CTRL_VLAN) {
660 		status = 0;
661 	}
662 
663 	/* Update status */
664 	*(virtio_net_ctrl_ack *)(uintptr_t)vring->desc[idx_status].addr = status;
665 
666 	return n_descs;
667 }
668 
669 static inline int
670 desc_is_avail(struct vring_packed_desc *desc, bool wrap_counter)
671 {
672 	uint16_t flags = __atomic_load_n(&desc->flags, __ATOMIC_ACQUIRE);
673 
674 	return wrap_counter == !!(flags & VRING_PACKED_DESC_F_AVAIL) &&
675 		wrap_counter != !!(flags & VRING_PACKED_DESC_F_USED);
676 }
677 
678 static uint32_t
679 virtio_user_handle_ctrl_msg_packed(struct virtio_user_dev *dev,
680 				   struct vring_packed *vring,
681 				   uint16_t idx_hdr)
682 {
683 	struct virtio_net_ctrl_hdr *hdr;
684 	virtio_net_ctrl_ack status = ~0;
685 	uint16_t idx_data, idx_status;
686 	/* initialize to one, header is first */
687 	uint32_t n_descs = 1;
688 
689 	/* locate desc for header, data, and status */
690 	idx_data = idx_hdr + 1;
691 	if (idx_data >= dev->queue_size)
692 		idx_data -= dev->queue_size;
693 
694 	n_descs++;
695 
696 	idx_status = idx_data;
697 	while (vring->desc[idx_status].flags & VRING_DESC_F_NEXT) {
698 		idx_status++;
699 		if (idx_status >= dev->queue_size)
700 			idx_status -= dev->queue_size;
701 		n_descs++;
702 	}
703 
704 	hdr = (void *)(uintptr_t)vring->desc[idx_hdr].addr;
705 	if (hdr->class == VIRTIO_NET_CTRL_MQ &&
706 	    hdr->cmd == VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET) {
707 		uint16_t queues;
708 
709 		queues = *(uint16_t *)(uintptr_t)
710 				vring->desc[idx_data].addr;
711 		status = virtio_user_handle_mq(dev, queues);
712 	} else if (hdr->class == VIRTIO_NET_CTRL_RX  ||
713 		   hdr->class == VIRTIO_NET_CTRL_MAC ||
714 		   hdr->class == VIRTIO_NET_CTRL_VLAN) {
715 		status = 0;
716 	}
717 
718 	/* Update status */
719 	*(virtio_net_ctrl_ack *)(uintptr_t)
720 		vring->desc[idx_status].addr = status;
721 
722 	/* Update used descriptor */
723 	vring->desc[idx_hdr].id = vring->desc[idx_status].id;
724 	vring->desc[idx_hdr].len = sizeof(status);
725 
726 	return n_descs;
727 }
728 
729 void
730 virtio_user_handle_cq_packed(struct virtio_user_dev *dev, uint16_t queue_idx)
731 {
732 	struct virtio_user_queue *vq = &dev->packed_queues[queue_idx];
733 	struct vring_packed *vring = &dev->packed_vrings[queue_idx];
734 	uint16_t n_descs, flags;
735 
736 	/* Perform a load-acquire barrier in desc_is_avail to
737 	 * enforce the ordering between desc flags and desc
738 	 * content.
739 	 */
740 	while (desc_is_avail(&vring->desc[vq->used_idx],
741 			     vq->used_wrap_counter)) {
742 
743 		n_descs = virtio_user_handle_ctrl_msg_packed(dev, vring,
744 				vq->used_idx);
745 
746 		flags = VRING_DESC_F_WRITE;
747 		if (vq->used_wrap_counter)
748 			flags |= VRING_PACKED_DESC_F_AVAIL_USED;
749 
750 		__atomic_store_n(&vring->desc[vq->used_idx].flags, flags,
751 				 __ATOMIC_RELEASE);
752 
753 		vq->used_idx += n_descs;
754 		if (vq->used_idx >= dev->queue_size) {
755 			vq->used_idx -= dev->queue_size;
756 			vq->used_wrap_counter ^= 1;
757 		}
758 	}
759 }
760 
761 void
762 virtio_user_handle_cq(struct virtio_user_dev *dev, uint16_t queue_idx)
763 {
764 	uint16_t avail_idx, desc_idx;
765 	struct vring_used_elem *uep;
766 	uint32_t n_descs;
767 	struct vring *vring = &dev->vrings[queue_idx];
768 
769 	/* Consume avail ring, using used ring idx as first one */
770 	while (__atomic_load_n(&vring->used->idx, __ATOMIC_RELAXED)
771 	       != vring->avail->idx) {
772 		avail_idx = __atomic_load_n(&vring->used->idx, __ATOMIC_RELAXED)
773 			    & (vring->num - 1);
774 		desc_idx = vring->avail->ring[avail_idx];
775 
776 		n_descs = virtio_user_handle_ctrl_msg(dev, vring, desc_idx);
777 
778 		/* Update used ring */
779 		uep = &vring->used->ring[avail_idx];
780 		uep->id = desc_idx;
781 		uep->len = n_descs;
782 
783 		__atomic_add_fetch(&vring->used->idx, 1, __ATOMIC_RELAXED);
784 	}
785 }
786