xref: /dpdk/drivers/net/virtio/virtio_user/virtio_user_dev.c (revision e11bdd37745229bf26b557305c07d118c3dbaad7)
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: set features */
155 	features = dev->features;
156 	/* Strip VIRTIO_NET_F_MAC, as MAC address is handled in vdev init */
157 	features &= ~(1ull << VIRTIO_NET_F_MAC);
158 	/* Strip VIRTIO_NET_F_CTRL_VQ, as devices do not really need to know */
159 	features &= ~(1ull << VIRTIO_NET_F_CTRL_VQ);
160 	features &= ~(1ull << VIRTIO_NET_F_STATUS);
161 	ret = dev->ops->send_request(dev, VHOST_USER_SET_FEATURES, &features);
162 	if (ret < 0)
163 		goto error;
164 	PMD_DRV_LOG(INFO, "set features: %" PRIx64, features);
165 
166 	/* Step 2: share memory regions */
167 	ret = dev->ops->send_request(dev, VHOST_USER_SET_MEM_TABLE, NULL);
168 	if (ret < 0)
169 		goto error;
170 
171 	/* Step 3: kick queues */
172 	if (virtio_user_queue_setup(dev, virtio_user_kick_queue) < 0)
173 		goto error;
174 
175 	/* Step 4: enable queues
176 	 * we enable the 1st queue pair by default.
177 	 */
178 	dev->ops->enable_qp(dev, 0, 1);
179 
180 	dev->started = true;
181 	pthread_mutex_unlock(&dev->mutex);
182 	rte_mcfg_mem_read_unlock();
183 
184 	return 0;
185 error:
186 	pthread_mutex_unlock(&dev->mutex);
187 	rte_mcfg_mem_read_unlock();
188 	/* TODO: free resource here or caller to check */
189 	return -1;
190 }
191 
192 int virtio_user_stop_device(struct virtio_user_dev *dev)
193 {
194 	struct vhost_vring_state state;
195 	uint32_t i;
196 	int error = 0;
197 
198 	pthread_mutex_lock(&dev->mutex);
199 	if (!dev->started)
200 		goto out;
201 
202 	for (i = 0; i < dev->max_queue_pairs; ++i)
203 		dev->ops->enable_qp(dev, i, 0);
204 
205 	/* Stop the backend. */
206 	for (i = 0; i < dev->max_queue_pairs * 2; ++i) {
207 		state.index = i;
208 		if (dev->ops->send_request(dev, VHOST_USER_GET_VRING_BASE,
209 					   &state) < 0) {
210 			PMD_DRV_LOG(ERR, "get_vring_base failed, index=%u\n",
211 				    i);
212 			error = -1;
213 			goto out;
214 		}
215 	}
216 
217 	dev->started = false;
218 out:
219 	pthread_mutex_unlock(&dev->mutex);
220 
221 	return error;
222 }
223 
224 static inline void
225 parse_mac(struct virtio_user_dev *dev, const char *mac)
226 {
227 	struct rte_ether_addr tmp;
228 
229 	if (!mac)
230 		return;
231 
232 	if (rte_ether_unformat_addr(mac, &tmp) == 0) {
233 		memcpy(dev->mac_addr, &tmp, RTE_ETHER_ADDR_LEN);
234 		dev->mac_specified = 1;
235 	} else {
236 		/* ignore the wrong mac, use random mac */
237 		PMD_DRV_LOG(ERR, "wrong format of mac: %s", mac);
238 	}
239 }
240 
241 static int
242 virtio_user_dev_init_notify(struct virtio_user_dev *dev)
243 {
244 	uint32_t i, j;
245 	int callfd;
246 	int kickfd;
247 
248 	for (i = 0; i < VIRTIO_MAX_VIRTQUEUES; ++i) {
249 		if (i >= dev->max_queue_pairs * 2) {
250 			dev->kickfds[i] = -1;
251 			dev->callfds[i] = -1;
252 			continue;
253 		}
254 
255 		/* May use invalid flag, but some backend uses kickfd and
256 		 * callfd as criteria to judge if dev is alive. so finally we
257 		 * use real event_fd.
258 		 */
259 		callfd = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK);
260 		if (callfd < 0) {
261 			PMD_DRV_LOG(ERR, "callfd error, %s", strerror(errno));
262 			break;
263 		}
264 		kickfd = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK);
265 		if (kickfd < 0) {
266 			PMD_DRV_LOG(ERR, "kickfd error, %s", strerror(errno));
267 			break;
268 		}
269 		dev->callfds[i] = callfd;
270 		dev->kickfds[i] = kickfd;
271 	}
272 
273 	if (i < VIRTIO_MAX_VIRTQUEUES) {
274 		for (j = 0; j <= i; ++j) {
275 			close(dev->callfds[j]);
276 			close(dev->kickfds[j]);
277 		}
278 
279 		return -1;
280 	}
281 
282 	return 0;
283 }
284 
285 static int
286 virtio_user_fill_intr_handle(struct virtio_user_dev *dev)
287 {
288 	uint32_t i;
289 	struct rte_eth_dev *eth_dev = &rte_eth_devices[dev->port_id];
290 
291 	if (!eth_dev->intr_handle) {
292 		eth_dev->intr_handle = malloc(sizeof(*eth_dev->intr_handle));
293 		if (!eth_dev->intr_handle) {
294 			PMD_DRV_LOG(ERR, "fail to allocate intr_handle");
295 			return -1;
296 		}
297 		memset(eth_dev->intr_handle, 0, sizeof(*eth_dev->intr_handle));
298 	}
299 
300 	for (i = 0; i < dev->max_queue_pairs; ++i)
301 		eth_dev->intr_handle->efds[i] = dev->callfds[i];
302 	eth_dev->intr_handle->nb_efd = dev->max_queue_pairs;
303 	eth_dev->intr_handle->max_intr = dev->max_queue_pairs + 1;
304 	eth_dev->intr_handle->type = RTE_INTR_HANDLE_VDEV;
305 	/* For virtio vdev, no need to read counter for clean */
306 	eth_dev->intr_handle->efd_counter_size = 0;
307 	eth_dev->intr_handle->fd = -1;
308 	if (dev->vhostfd >= 0)
309 		eth_dev->intr_handle->fd = dev->vhostfd;
310 	else if (dev->is_server)
311 		eth_dev->intr_handle->fd = dev->listenfd;
312 
313 	return 0;
314 }
315 
316 static void
317 virtio_user_mem_event_cb(enum rte_mem_event type __rte_unused,
318 			 const void *addr,
319 			 size_t len __rte_unused,
320 			 void *arg)
321 {
322 	struct virtio_user_dev *dev = arg;
323 	struct rte_memseg_list *msl;
324 	uint16_t i;
325 
326 	/* ignore externally allocated memory */
327 	msl = rte_mem_virt2memseg_list(addr);
328 	if (msl->external)
329 		return;
330 
331 	pthread_mutex_lock(&dev->mutex);
332 
333 	if (dev->started == false)
334 		goto exit;
335 
336 	/* Step 1: pause the active queues */
337 	for (i = 0; i < dev->queue_pairs; i++)
338 		dev->ops->enable_qp(dev, i, 0);
339 
340 	/* Step 2: update memory regions */
341 	dev->ops->send_request(dev, VHOST_USER_SET_MEM_TABLE, NULL);
342 
343 	/* Step 3: resume the active queues */
344 	for (i = 0; i < dev->queue_pairs; i++)
345 		dev->ops->enable_qp(dev, i, 1);
346 
347 exit:
348 	pthread_mutex_unlock(&dev->mutex);
349 }
350 
351 static int
352 virtio_user_dev_setup(struct virtio_user_dev *dev)
353 {
354 	uint32_t q;
355 
356 	dev->vhostfd = -1;
357 	dev->vhostfds = NULL;
358 	dev->tapfds = NULL;
359 
360 	if (dev->is_server) {
361 		if (access(dev->path, F_OK) == 0 &&
362 		    !is_vhost_user_by_type(dev->path)) {
363 			PMD_DRV_LOG(ERR, "Server mode doesn't support vhost-kernel!");
364 			return -1;
365 		}
366 		dev->ops = &virtio_ops_user;
367 	} else {
368 		if (is_vhost_user_by_type(dev->path)) {
369 			dev->ops = &virtio_ops_user;
370 		} else {
371 			dev->ops = &virtio_ops_kernel;
372 
373 			dev->vhostfds = malloc(dev->max_queue_pairs *
374 					       sizeof(int));
375 			dev->tapfds = malloc(dev->max_queue_pairs *
376 					     sizeof(int));
377 			if (!dev->vhostfds || !dev->tapfds) {
378 				PMD_INIT_LOG(ERR, "Failed to malloc");
379 				return -1;
380 			}
381 
382 			for (q = 0; q < dev->max_queue_pairs; ++q) {
383 				dev->vhostfds[q] = -1;
384 				dev->tapfds[q] = -1;
385 			}
386 		}
387 	}
388 
389 	if (dev->ops->setup(dev) < 0)
390 		return -1;
391 
392 	if (virtio_user_dev_init_notify(dev) < 0)
393 		return -1;
394 
395 	if (virtio_user_fill_intr_handle(dev) < 0)
396 		return -1;
397 
398 	return 0;
399 }
400 
401 /* Use below macro to filter features from vhost backend */
402 #define VIRTIO_USER_SUPPORTED_FEATURES			\
403 	(1ULL << VIRTIO_NET_F_MAC		|	\
404 	 1ULL << VIRTIO_NET_F_STATUS		|	\
405 	 1ULL << VIRTIO_NET_F_MQ		|	\
406 	 1ULL << VIRTIO_NET_F_CTRL_MAC_ADDR	|	\
407 	 1ULL << VIRTIO_NET_F_CTRL_VQ		|	\
408 	 1ULL << VIRTIO_NET_F_CTRL_RX		|	\
409 	 1ULL << VIRTIO_NET_F_CTRL_VLAN		|	\
410 	 1ULL << VIRTIO_NET_F_CSUM		|	\
411 	 1ULL << VIRTIO_NET_F_HOST_TSO4		|	\
412 	 1ULL << VIRTIO_NET_F_HOST_TSO6		|	\
413 	 1ULL << VIRTIO_NET_F_MRG_RXBUF		|	\
414 	 1ULL << VIRTIO_RING_F_INDIRECT_DESC	|	\
415 	 1ULL << VIRTIO_NET_F_GUEST_CSUM	|	\
416 	 1ULL << VIRTIO_NET_F_GUEST_TSO4	|	\
417 	 1ULL << VIRTIO_NET_F_GUEST_TSO6	|	\
418 	 1ULL << VIRTIO_F_IN_ORDER		|	\
419 	 1ULL << VIRTIO_F_VERSION_1		|	\
420 	 1ULL << VIRTIO_F_RING_PACKED)
421 
422 int
423 virtio_user_dev_init(struct virtio_user_dev *dev, char *path, int queues,
424 		     int cq, int queue_size, const char *mac, char **ifname,
425 		     int server, int mrg_rxbuf, int in_order, int packed_vq)
426 {
427 	pthread_mutex_init(&dev->mutex, NULL);
428 	strlcpy(dev->path, path, PATH_MAX);
429 	dev->started = 0;
430 	dev->max_queue_pairs = queues;
431 	dev->queue_pairs = 1; /* mq disabled by default */
432 	dev->queue_size = queue_size;
433 	dev->is_server = server;
434 	dev->mac_specified = 0;
435 	dev->frontend_features = 0;
436 	dev->unsupported_features = ~VIRTIO_USER_SUPPORTED_FEATURES;
437 	parse_mac(dev, mac);
438 
439 	if (*ifname) {
440 		dev->ifname = *ifname;
441 		*ifname = NULL;
442 	}
443 
444 	if (virtio_user_dev_setup(dev) < 0) {
445 		PMD_INIT_LOG(ERR, "backend set up fails");
446 		return -1;
447 	}
448 
449 	if (!dev->is_server) {
450 		if (dev->ops->send_request(dev, VHOST_USER_SET_OWNER,
451 					   NULL) < 0) {
452 			PMD_INIT_LOG(ERR, "set_owner fails: %s",
453 				     strerror(errno));
454 			return -1;
455 		}
456 
457 		if (dev->ops->send_request(dev, VHOST_USER_GET_FEATURES,
458 					   &dev->device_features) < 0) {
459 			PMD_INIT_LOG(ERR, "get_features failed: %s",
460 				     strerror(errno));
461 			return -1;
462 		}
463 	} else {
464 		/* We just pretend vhost-user can support all these features.
465 		 * Note that this could be problematic that if some feature is
466 		 * negotiated but not supported by the vhost-user which comes
467 		 * later.
468 		 */
469 		dev->device_features = VIRTIO_USER_SUPPORTED_FEATURES;
470 	}
471 
472 	if (!mrg_rxbuf)
473 		dev->unsupported_features |= (1ull << VIRTIO_NET_F_MRG_RXBUF);
474 
475 	if (!in_order)
476 		dev->unsupported_features |= (1ull << VIRTIO_F_IN_ORDER);
477 
478 	if (!packed_vq)
479 		dev->unsupported_features |= (1ull << VIRTIO_F_RING_PACKED);
480 
481 	if (dev->mac_specified)
482 		dev->frontend_features |= (1ull << VIRTIO_NET_F_MAC);
483 	else
484 		dev->unsupported_features |= (1ull << VIRTIO_NET_F_MAC);
485 
486 	if (cq) {
487 		/* device does not really need to know anything about CQ,
488 		 * so if necessary, we just claim to support CQ
489 		 */
490 		dev->frontend_features |= (1ull << VIRTIO_NET_F_CTRL_VQ);
491 	} else {
492 		dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_VQ);
493 		/* Also disable features that depend on VIRTIO_NET_F_CTRL_VQ */
494 		dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_RX);
495 		dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_VLAN);
496 		dev->unsupported_features |=
497 			(1ull << VIRTIO_NET_F_GUEST_ANNOUNCE);
498 		dev->unsupported_features |= (1ull << VIRTIO_NET_F_MQ);
499 		dev->unsupported_features |=
500 			(1ull << VIRTIO_NET_F_CTRL_MAC_ADDR);
501 	}
502 
503 	/* The backend will not report this feature, we add it explicitly */
504 	if (is_vhost_user_by_type(dev->path))
505 		dev->frontend_features |= (1ull << VIRTIO_NET_F_STATUS);
506 
507 	/*
508 	 * Device features =
509 	 *     (frontend_features | backend_features) & ~unsupported_features;
510 	 */
511 	dev->device_features |= dev->frontend_features;
512 	dev->device_features &= ~dev->unsupported_features;
513 
514 	if (rte_mem_event_callback_register(VIRTIO_USER_MEM_EVENT_CLB_NAME,
515 				virtio_user_mem_event_cb, dev)) {
516 		if (rte_errno != ENOTSUP) {
517 			PMD_INIT_LOG(ERR, "Failed to register mem event"
518 					" callback\n");
519 			return -1;
520 		}
521 	}
522 
523 	return 0;
524 }
525 
526 void
527 virtio_user_dev_uninit(struct virtio_user_dev *dev)
528 {
529 	uint32_t i;
530 
531 	virtio_user_stop_device(dev);
532 
533 	rte_mem_event_callback_unregister(VIRTIO_USER_MEM_EVENT_CLB_NAME, dev);
534 
535 	for (i = 0; i < dev->max_queue_pairs * 2; ++i) {
536 		close(dev->callfds[i]);
537 		close(dev->kickfds[i]);
538 	}
539 
540 	if (dev->vhostfd >= 0)
541 		close(dev->vhostfd);
542 
543 	if (dev->is_server && dev->listenfd >= 0) {
544 		close(dev->listenfd);
545 		dev->listenfd = -1;
546 	}
547 
548 	if (dev->vhostfds) {
549 		for (i = 0; i < dev->max_queue_pairs; ++i) {
550 			close(dev->vhostfds[i]);
551 			if (dev->tapfds[i] >= 0)
552 				close(dev->tapfds[i]);
553 		}
554 		free(dev->vhostfds);
555 		free(dev->tapfds);
556 	}
557 
558 	free(dev->ifname);
559 
560 	if (dev->is_server)
561 		unlink(dev->path);
562 }
563 
564 uint8_t
565 virtio_user_handle_mq(struct virtio_user_dev *dev, uint16_t q_pairs)
566 {
567 	uint16_t i;
568 	uint8_t ret = 0;
569 
570 	if (q_pairs > dev->max_queue_pairs) {
571 		PMD_INIT_LOG(ERR, "multi-q config %u, but only %u supported",
572 			     q_pairs, dev->max_queue_pairs);
573 		return -1;
574 	}
575 
576 	/* Server mode can't enable queue pairs if vhostfd is invalid,
577 	 * always return 0 in this case.
578 	 */
579 	if (!dev->is_server || dev->vhostfd >= 0) {
580 		for (i = 0; i < q_pairs; ++i)
581 			ret |= dev->ops->enable_qp(dev, i, 1);
582 		for (i = q_pairs; i < dev->max_queue_pairs; ++i)
583 			ret |= dev->ops->enable_qp(dev, i, 0);
584 	}
585 	dev->queue_pairs = q_pairs;
586 
587 	return ret;
588 }
589 
590 static uint32_t
591 virtio_user_handle_ctrl_msg(struct virtio_user_dev *dev, struct vring *vring,
592 			    uint16_t idx_hdr)
593 {
594 	struct virtio_net_ctrl_hdr *hdr;
595 	virtio_net_ctrl_ack status = ~0;
596 	uint16_t i, idx_data, idx_status;
597 	uint32_t n_descs = 0;
598 
599 	/* locate desc for header, data, and status */
600 	idx_data = vring->desc[idx_hdr].next;
601 	n_descs++;
602 
603 	i = idx_data;
604 	while (vring->desc[i].flags == VRING_DESC_F_NEXT) {
605 		i = vring->desc[i].next;
606 		n_descs++;
607 	}
608 
609 	/* locate desc for status */
610 	idx_status = i;
611 	n_descs++;
612 
613 	hdr = (void *)(uintptr_t)vring->desc[idx_hdr].addr;
614 	if (hdr->class == VIRTIO_NET_CTRL_MQ &&
615 	    hdr->cmd == VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET) {
616 		uint16_t queues;
617 
618 		queues = *(uint16_t *)(uintptr_t)vring->desc[idx_data].addr;
619 		status = virtio_user_handle_mq(dev, queues);
620 	} else if (hdr->class == VIRTIO_NET_CTRL_RX  ||
621 		   hdr->class == VIRTIO_NET_CTRL_MAC ||
622 		   hdr->class == VIRTIO_NET_CTRL_VLAN) {
623 		status = 0;
624 	}
625 
626 	/* Update status */
627 	*(virtio_net_ctrl_ack *)(uintptr_t)vring->desc[idx_status].addr = status;
628 
629 	return n_descs;
630 }
631 
632 static inline int
633 desc_is_avail(struct vring_packed_desc *desc, bool wrap_counter)
634 {
635 	uint16_t flags = __atomic_load_n(&desc->flags, __ATOMIC_ACQUIRE);
636 
637 	return wrap_counter == !!(flags & VRING_PACKED_DESC_F_AVAIL) &&
638 		wrap_counter != !!(flags & VRING_PACKED_DESC_F_USED);
639 }
640 
641 static uint32_t
642 virtio_user_handle_ctrl_msg_packed(struct virtio_user_dev *dev,
643 				   struct vring_packed *vring,
644 				   uint16_t idx_hdr)
645 {
646 	struct virtio_net_ctrl_hdr *hdr;
647 	virtio_net_ctrl_ack status = ~0;
648 	uint16_t idx_data, idx_status;
649 	/* initialize to one, header is first */
650 	uint32_t n_descs = 1;
651 
652 	/* locate desc for header, data, and status */
653 	idx_data = idx_hdr + 1;
654 	if (idx_data >= dev->queue_size)
655 		idx_data -= dev->queue_size;
656 
657 	n_descs++;
658 
659 	idx_status = idx_data;
660 	while (vring->desc[idx_status].flags & VRING_DESC_F_NEXT) {
661 		idx_status++;
662 		if (idx_status >= dev->queue_size)
663 			idx_status -= dev->queue_size;
664 		n_descs++;
665 	}
666 
667 	hdr = (void *)(uintptr_t)vring->desc[idx_hdr].addr;
668 	if (hdr->class == VIRTIO_NET_CTRL_MQ &&
669 	    hdr->cmd == VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET) {
670 		uint16_t queues;
671 
672 		queues = *(uint16_t *)(uintptr_t)
673 				vring->desc[idx_data].addr;
674 		status = virtio_user_handle_mq(dev, queues);
675 	} else if (hdr->class == VIRTIO_NET_CTRL_RX  ||
676 		   hdr->class == VIRTIO_NET_CTRL_MAC ||
677 		   hdr->class == VIRTIO_NET_CTRL_VLAN) {
678 		status = 0;
679 	}
680 
681 	/* Update status */
682 	*(virtio_net_ctrl_ack *)(uintptr_t)
683 		vring->desc[idx_status].addr = status;
684 
685 	/* Update used descriptor */
686 	vring->desc[idx_hdr].id = vring->desc[idx_status].id;
687 	vring->desc[idx_hdr].len = sizeof(status);
688 
689 	return n_descs;
690 }
691 
692 void
693 virtio_user_handle_cq_packed(struct virtio_user_dev *dev, uint16_t queue_idx)
694 {
695 	struct virtio_user_queue *vq = &dev->packed_queues[queue_idx];
696 	struct vring_packed *vring = &dev->packed_vrings[queue_idx];
697 	uint16_t n_descs, flags;
698 
699 	/* Perform a load-acquire barrier in desc_is_avail to
700 	 * enforce the ordering between desc flags and desc
701 	 * content.
702 	 */
703 	while (desc_is_avail(&vring->desc[vq->used_idx],
704 			     vq->used_wrap_counter)) {
705 
706 		n_descs = virtio_user_handle_ctrl_msg_packed(dev, vring,
707 				vq->used_idx);
708 
709 		flags = VRING_DESC_F_WRITE;
710 		if (vq->used_wrap_counter)
711 			flags |= VRING_PACKED_DESC_F_AVAIL_USED;
712 
713 		__atomic_store_n(&vring->desc[vq->used_idx].flags, flags,
714 				 __ATOMIC_RELEASE);
715 
716 		vq->used_idx += n_descs;
717 		if (vq->used_idx >= dev->queue_size) {
718 			vq->used_idx -= dev->queue_size;
719 			vq->used_wrap_counter ^= 1;
720 		}
721 	}
722 }
723 
724 void
725 virtio_user_handle_cq(struct virtio_user_dev *dev, uint16_t queue_idx)
726 {
727 	uint16_t avail_idx, desc_idx;
728 	struct vring_used_elem *uep;
729 	uint32_t n_descs;
730 	struct vring *vring = &dev->vrings[queue_idx];
731 
732 	/* Consume avail ring, using used ring idx as first one */
733 	while (__atomic_load_n(&vring->used->idx, __ATOMIC_RELAXED)
734 	       != vring->avail->idx) {
735 		avail_idx = __atomic_load_n(&vring->used->idx, __ATOMIC_RELAXED)
736 			    & (vring->num - 1);
737 		desc_idx = vring->avail->ring[avail_idx];
738 
739 		n_descs = virtio_user_handle_ctrl_msg(dev, vring, desc_idx);
740 
741 		/* Update used ring */
742 		uep = &vring->used->ring[avail_idx];
743 		uep->id = desc_idx;
744 		uep->len = n_descs;
745 
746 		__atomic_add_fetch(&vring->used->idx, 1, __ATOMIC_RELAXED);
747 	}
748 }
749