1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2016 Intel Corporation
12 #include <sys/eventfd.h>
13 #include <sys/types.h>
16 #include <rte_eal_memconfig.h>
19 #include "virtio_user_dev.h"
20 #include "../virtio_ethdev.h"
22 #define VIRTIO_USER_MEM_EVENT_CLB_NAME "virtio_user_mem_event_clb"
25 virtio_user_create_queue(struct virtio_user_dev *dev, uint32_t queue_sel)
27 /* Of all per virtqueue MSGs, make sure VHOST_SET_VRING_CALL come
28 * firstly because vhost depends on this msg to allocate virtqueue
31 struct vhost_vring_file file;
33 file.index = queue_sel;
34 file.fd = dev->callfds[queue_sel];
35 dev->ops->send_request(dev, VHOST_USER_SET_VRING_CALL, &file);
41 virtio_user_kick_queue(struct virtio_user_dev *dev, uint32_t queue_sel)
43 struct vhost_vring_file file;
44 struct vhost_vring_state state;
45 struct vring *vring = &dev->vrings[queue_sel];
46 struct vhost_vring_addr addr = {
48 .desc_user_addr = (uint64_t)(uintptr_t)vring->desc,
49 .avail_user_addr = (uint64_t)(uintptr_t)vring->avail,
50 .used_user_addr = (uint64_t)(uintptr_t)vring->used,
52 .flags = 0, /* disable log */
55 state.index = queue_sel;
56 state.num = vring->num;
57 dev->ops->send_request(dev, VHOST_USER_SET_VRING_NUM, &state);
59 state.index = queue_sel;
60 state.num = 0; /* no reservation */
61 dev->ops->send_request(dev, VHOST_USER_SET_VRING_BASE, &state);
63 dev->ops->send_request(dev, VHOST_USER_SET_VRING_ADDR, &addr);
65 /* Of all per virtqueue MSGs, make sure VHOST_USER_SET_VRING_KICK comes
66 * lastly because vhost depends on this msg to judge if
69 file.index = queue_sel;
70 file.fd = dev->kickfds[queue_sel];
71 dev->ops->send_request(dev, VHOST_USER_SET_VRING_KICK, &file);
77 virtio_user_queue_setup(struct virtio_user_dev *dev,
78 int (*fn)(struct virtio_user_dev *, uint32_t))
80 uint32_t i, queue_sel;
82 for (i = 0; i < dev->max_queue_pairs; ++i) {
83 queue_sel = 2 * i + VTNET_SQ_RQ_QUEUE_IDX;
84 if (fn(dev, queue_sel) < 0) {
85 PMD_DRV_LOG(INFO, "setup rx vq fails: %u", i);
89 for (i = 0; i < dev->max_queue_pairs; ++i) {
90 queue_sel = 2 * i + VTNET_SQ_TQ_QUEUE_IDX;
91 if (fn(dev, queue_sel) < 0) {
92 PMD_DRV_LOG(INFO, "setup tx vq fails: %u", i);
101 is_vhost_user_by_type(const char *path)
105 if (stat(path, &sb) == -1)
108 return S_ISSOCK(sb.st_mode);
112 virtio_user_start_device(struct virtio_user_dev *dev)
114 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
121 * We need to make sure that the locks will be
122 * taken in the correct order to avoid deadlocks.
124 * Before releasing this lock, this thread should
125 * not trigger any memory hotplug events.
127 * This is a temporary workaround, and should be
128 * replaced when we get proper supports from the
129 * memory subsystem in the future.
131 rte_rwlock_read_lock(&mcfg->memory_hotplug_lock);
132 pthread_mutex_lock(&dev->mutex);
134 if (is_vhost_user_by_type(dev->path) && dev->vhostfd < 0)
137 /* Step 0: tell vhost to create queues */
138 if (virtio_user_queue_setup(dev, virtio_user_create_queue) < 0)
141 /* Step 1: set features */
142 features = dev->features;
143 /* Strip VIRTIO_NET_F_MAC, as MAC address is handled in vdev init */
144 features &= ~(1ull << VIRTIO_NET_F_MAC);
145 /* Strip VIRTIO_NET_F_CTRL_VQ, as devices do not really need to know */
146 features &= ~(1ull << VIRTIO_NET_F_CTRL_VQ);
147 features &= ~(1ull << VIRTIO_NET_F_STATUS);
148 ret = dev->ops->send_request(dev, VHOST_USER_SET_FEATURES, &features);
151 PMD_DRV_LOG(INFO, "set features: %" PRIx64, features);
153 /* Step 2: share memory regions */
154 ret = dev->ops->send_request(dev, VHOST_USER_SET_MEM_TABLE, NULL);
158 /* Step 3: kick queues */
159 if (virtio_user_queue_setup(dev, virtio_user_kick_queue) < 0)
162 /* Step 4: enable queues
163 * we enable the 1st queue pair by default.
165 dev->ops->enable_qp(dev, 0, 1);
168 pthread_mutex_unlock(&dev->mutex);
169 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
173 pthread_mutex_unlock(&dev->mutex);
174 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
175 /* TODO: free resource here or caller to check */
179 int virtio_user_stop_device(struct virtio_user_dev *dev)
181 struct vhost_vring_state state;
185 pthread_mutex_lock(&dev->mutex);
189 for (i = 0; i < dev->max_queue_pairs; ++i)
190 dev->ops->enable_qp(dev, i, 0);
192 /* Stop the backend. */
193 for (i = 0; i < dev->max_queue_pairs * 2; ++i) {
195 if (dev->ops->send_request(dev, VHOST_USER_GET_VRING_BASE,
197 PMD_DRV_LOG(ERR, "get_vring_base failed, index=%u\n",
204 dev->started = false;
206 pthread_mutex_unlock(&dev->mutex);
212 parse_mac(struct virtio_user_dev *dev, const char *mac)
215 uint32_t tmp[ETHER_ADDR_LEN];
220 r = sscanf(mac, "%x:%x:%x:%x:%x:%x", &tmp[0],
221 &tmp[1], &tmp[2], &tmp[3], &tmp[4], &tmp[5]);
222 if (r == ETHER_ADDR_LEN) {
223 for (i = 0; i < ETHER_ADDR_LEN; ++i)
224 dev->mac_addr[i] = (uint8_t)tmp[i];
225 dev->mac_specified = 1;
227 /* ignore the wrong mac, use random mac */
228 PMD_DRV_LOG(ERR, "wrong format of mac: %s", mac);
233 virtio_user_dev_init_notify(struct virtio_user_dev *dev)
239 for (i = 0; i < VIRTIO_MAX_VIRTQUEUES; ++i) {
240 if (i >= dev->max_queue_pairs * 2) {
241 dev->kickfds[i] = -1;
242 dev->callfds[i] = -1;
246 /* May use invalid flag, but some backend uses kickfd and
247 * callfd as criteria to judge if dev is alive. so finally we
250 callfd = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK);
252 PMD_DRV_LOG(ERR, "callfd error, %s", strerror(errno));
255 kickfd = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK);
257 PMD_DRV_LOG(ERR, "kickfd error, %s", strerror(errno));
260 dev->callfds[i] = callfd;
261 dev->kickfds[i] = kickfd;
264 if (i < VIRTIO_MAX_VIRTQUEUES) {
265 for (j = 0; j <= i; ++j) {
266 close(dev->callfds[j]);
267 close(dev->kickfds[j]);
277 virtio_user_fill_intr_handle(struct virtio_user_dev *dev)
280 struct rte_eth_dev *eth_dev = &rte_eth_devices[dev->port_id];
282 if (!eth_dev->intr_handle) {
283 eth_dev->intr_handle = malloc(sizeof(*eth_dev->intr_handle));
284 if (!eth_dev->intr_handle) {
285 PMD_DRV_LOG(ERR, "fail to allocate intr_handle");
288 memset(eth_dev->intr_handle, 0, sizeof(*eth_dev->intr_handle));
291 for (i = 0; i < dev->max_queue_pairs; ++i)
292 eth_dev->intr_handle->efds[i] = dev->callfds[i];
293 eth_dev->intr_handle->nb_efd = dev->max_queue_pairs;
294 eth_dev->intr_handle->max_intr = dev->max_queue_pairs + 1;
295 eth_dev->intr_handle->type = RTE_INTR_HANDLE_VDEV;
296 /* For virtio vdev, no need to read counter for clean */
297 eth_dev->intr_handle->efd_counter_size = 0;
298 eth_dev->intr_handle->fd = -1;
299 if (dev->vhostfd >= 0)
300 eth_dev->intr_handle->fd = dev->vhostfd;
301 else if (dev->is_server)
302 eth_dev->intr_handle->fd = dev->listenfd;
308 virtio_user_mem_event_cb(enum rte_mem_event type __rte_unused,
309 const void *addr __rte_unused,
310 size_t len __rte_unused,
313 struct virtio_user_dev *dev = arg;
314 struct rte_memseg_list *msl;
317 /* ignore externally allocated memory */
318 msl = rte_mem_virt2memseg_list(addr);
322 pthread_mutex_lock(&dev->mutex);
324 if (dev->started == false)
327 /* Step 1: pause the active queues */
328 for (i = 0; i < dev->queue_pairs; i++)
329 dev->ops->enable_qp(dev, i, 0);
331 /* Step 2: update memory regions */
332 dev->ops->send_request(dev, VHOST_USER_SET_MEM_TABLE, NULL);
334 /* Step 3: resume the active queues */
335 for (i = 0; i < dev->queue_pairs; i++)
336 dev->ops->enable_qp(dev, i, 1);
339 pthread_mutex_unlock(&dev->mutex);
343 virtio_user_dev_setup(struct virtio_user_dev *dev)
348 dev->vhostfds = NULL;
351 if (dev->is_server) {
352 if (access(dev->path, F_OK) == 0 &&
353 !is_vhost_user_by_type(dev->path)) {
354 PMD_DRV_LOG(ERR, "Server mode doesn't support vhost-kernel!");
357 dev->ops = &virtio_ops_user;
359 if (is_vhost_user_by_type(dev->path)) {
360 dev->ops = &virtio_ops_user;
362 dev->ops = &virtio_ops_kernel;
364 dev->vhostfds = malloc(dev->max_queue_pairs *
366 dev->tapfds = malloc(dev->max_queue_pairs *
368 if (!dev->vhostfds || !dev->tapfds) {
369 PMD_INIT_LOG(ERR, "Failed to malloc");
373 for (q = 0; q < dev->max_queue_pairs; ++q) {
374 dev->vhostfds[q] = -1;
380 if (dev->ops->setup(dev) < 0)
383 if (virtio_user_dev_init_notify(dev) < 0)
386 if (virtio_user_fill_intr_handle(dev) < 0)
392 /* Use below macro to filter features from vhost backend */
393 #define VIRTIO_USER_SUPPORTED_FEATURES \
394 (1ULL << VIRTIO_NET_F_MAC | \
395 1ULL << VIRTIO_NET_F_STATUS | \
396 1ULL << VIRTIO_NET_F_MQ | \
397 1ULL << VIRTIO_NET_F_CTRL_MAC_ADDR | \
398 1ULL << VIRTIO_NET_F_CTRL_VQ | \
399 1ULL << VIRTIO_NET_F_CTRL_RX | \
400 1ULL << VIRTIO_NET_F_CTRL_VLAN | \
401 1ULL << VIRTIO_NET_F_CSUM | \
402 1ULL << VIRTIO_NET_F_HOST_TSO4 | \
403 1ULL << VIRTIO_NET_F_HOST_TSO6 | \
404 1ULL << VIRTIO_NET_F_MRG_RXBUF | \
405 1ULL << VIRTIO_RING_F_INDIRECT_DESC | \
406 1ULL << VIRTIO_NET_F_GUEST_CSUM | \
407 1ULL << VIRTIO_NET_F_GUEST_TSO4 | \
408 1ULL << VIRTIO_NET_F_GUEST_TSO6 | \
409 1ULL << VIRTIO_F_IN_ORDER | \
410 1ULL << VIRTIO_F_VERSION_1)
413 virtio_user_dev_init(struct virtio_user_dev *dev, char *path, int queues,
414 int cq, int queue_size, const char *mac, char **ifname,
415 int mrg_rxbuf, int in_order)
417 pthread_mutex_init(&dev->mutex, NULL);
418 snprintf(dev->path, PATH_MAX, "%s", path);
420 dev->max_queue_pairs = queues;
421 dev->queue_pairs = 1; /* mq disabled by default */
422 dev->queue_size = queue_size;
423 dev->mac_specified = 0;
424 dev->frontend_features = 0;
425 dev->unsupported_features = ~VIRTIO_USER_SUPPORTED_FEATURES;
429 dev->ifname = *ifname;
433 if (virtio_user_dev_setup(dev) < 0) {
434 PMD_INIT_LOG(ERR, "backend set up fails");
438 if (!dev->is_server) {
439 if (dev->ops->send_request(dev, VHOST_USER_SET_OWNER,
441 PMD_INIT_LOG(ERR, "set_owner fails: %s",
446 if (dev->ops->send_request(dev, VHOST_USER_GET_FEATURES,
447 &dev->device_features) < 0) {
448 PMD_INIT_LOG(ERR, "get_features failed: %s",
453 /* We just pretend vhost-user can support all these features.
454 * Note that this could be problematic that if some feature is
455 * negotiated but not supported by the vhost-user which comes
458 dev->device_features = VIRTIO_USER_SUPPORTED_FEATURES;
462 dev->unsupported_features |= (1ull << VIRTIO_NET_F_MRG_RXBUF);
465 dev->unsupported_features |= (1ull << VIRTIO_F_IN_ORDER);
467 if (dev->mac_specified)
468 dev->frontend_features |= (1ull << VIRTIO_NET_F_MAC);
470 dev->unsupported_features |= (1ull << VIRTIO_NET_F_MAC);
473 /* device does not really need to know anything about CQ,
474 * so if necessary, we just claim to support CQ
476 dev->frontend_features |= (1ull << VIRTIO_NET_F_CTRL_VQ);
478 dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_VQ);
479 /* Also disable features that depend on VIRTIO_NET_F_CTRL_VQ */
480 dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_RX);
481 dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_VLAN);
482 dev->unsupported_features |=
483 (1ull << VIRTIO_NET_F_GUEST_ANNOUNCE);
484 dev->unsupported_features |= (1ull << VIRTIO_NET_F_MQ);
485 dev->unsupported_features |=
486 (1ull << VIRTIO_NET_F_CTRL_MAC_ADDR);
489 /* The backend will not report this feature, we add it explicitly */
490 if (is_vhost_user_by_type(dev->path))
491 dev->frontend_features |= (1ull << VIRTIO_NET_F_STATUS);
495 * (frontend_features | backend_features) & ~unsupported_features;
497 dev->device_features |= dev->frontend_features;
498 dev->device_features &= ~dev->unsupported_features;
500 if (rte_mem_event_callback_register(VIRTIO_USER_MEM_EVENT_CLB_NAME,
501 virtio_user_mem_event_cb, dev)) {
502 if (rte_errno != ENOTSUP) {
503 PMD_INIT_LOG(ERR, "Failed to register mem event"
513 virtio_user_dev_uninit(struct virtio_user_dev *dev)
517 virtio_user_stop_device(dev);
519 rte_mem_event_callback_unregister(VIRTIO_USER_MEM_EVENT_CLB_NAME, dev);
521 for (i = 0; i < dev->max_queue_pairs * 2; ++i) {
522 close(dev->callfds[i]);
523 close(dev->kickfds[i]);
528 if (dev->is_server && dev->listenfd >= 0) {
529 close(dev->listenfd);
534 for (i = 0; i < dev->max_queue_pairs; ++i)
535 close(dev->vhostfds[i]);
547 virtio_user_handle_mq(struct virtio_user_dev *dev, uint16_t q_pairs)
552 if (q_pairs > dev->max_queue_pairs) {
553 PMD_INIT_LOG(ERR, "multi-q config %u, but only %u supported",
554 q_pairs, dev->max_queue_pairs);
558 /* Server mode can't enable queue pairs if vhostfd is invalid,
559 * always return 0 in this case.
561 if (!dev->is_server || dev->vhostfd >= 0) {
562 for (i = 0; i < q_pairs; ++i)
563 ret |= dev->ops->enable_qp(dev, i, 1);
564 for (i = q_pairs; i < dev->max_queue_pairs; ++i)
565 ret |= dev->ops->enable_qp(dev, i, 0);
567 dev->queue_pairs = q_pairs;
573 virtio_user_handle_ctrl_msg(struct virtio_user_dev *dev, struct vring *vring,
576 struct virtio_net_ctrl_hdr *hdr;
577 virtio_net_ctrl_ack status = ~0;
578 uint16_t i, idx_data, idx_status;
579 uint32_t n_descs = 0;
581 /* locate desc for header, data, and status */
582 idx_data = vring->desc[idx_hdr].next;
586 while (vring->desc[i].flags == VRING_DESC_F_NEXT) {
587 i = vring->desc[i].next;
591 /* locate desc for status */
595 hdr = (void *)(uintptr_t)vring->desc[idx_hdr].addr;
596 if (hdr->class == VIRTIO_NET_CTRL_MQ &&
597 hdr->cmd == VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET) {
600 queues = *(uint16_t *)(uintptr_t)vring->desc[idx_data].addr;
601 status = virtio_user_handle_mq(dev, queues);
605 *(virtio_net_ctrl_ack *)(uintptr_t)vring->desc[idx_status].addr = status;
611 virtio_user_handle_cq(struct virtio_user_dev *dev, uint16_t queue_idx)
613 uint16_t avail_idx, desc_idx;
614 struct vring_used_elem *uep;
616 struct vring *vring = &dev->vrings[queue_idx];
618 /* Consume avail ring, using used ring idx as first one */
619 while (vring->used->idx != vring->avail->idx) {
620 avail_idx = (vring->used->idx) & (vring->num - 1);
621 desc_idx = vring->avail->ring[avail_idx];
623 n_descs = virtio_user_handle_ctrl_msg(dev, vring, desc_idx);
625 /* Update used ring */
626 uep = &vring->used->ring[avail_idx];