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 vring_packed *pq_vring = &dev->packed_vrings[queue_sel];
47 struct vhost_vring_addr addr = {
50 .flags = 0, /* disable log */
53 if (dev->features & (1ULL << VIRTIO_F_RING_PACKED)) {
55 (uint64_t)(uintptr_t)pq_vring->desc;
56 addr.avail_user_addr =
57 (uint64_t)(uintptr_t)pq_vring->driver;
59 (uint64_t)(uintptr_t)pq_vring->device;
61 addr.desc_user_addr = (uint64_t)(uintptr_t)vring->desc;
62 addr.avail_user_addr = (uint64_t)(uintptr_t)vring->avail;
63 addr.used_user_addr = (uint64_t)(uintptr_t)vring->used;
66 state.index = queue_sel;
67 state.num = vring->num;
68 dev->ops->send_request(dev, VHOST_USER_SET_VRING_NUM, &state);
70 state.index = queue_sel;
71 state.num = 0; /* no reservation */
72 if (dev->features & (1ULL << VIRTIO_F_RING_PACKED))
73 state.num |= (1 << 15);
74 dev->ops->send_request(dev, VHOST_USER_SET_VRING_BASE, &state);
76 dev->ops->send_request(dev, VHOST_USER_SET_VRING_ADDR, &addr);
78 /* Of all per virtqueue MSGs, make sure VHOST_USER_SET_VRING_KICK comes
79 * lastly because vhost depends on this msg to judge if
82 file.index = queue_sel;
83 file.fd = dev->kickfds[queue_sel];
84 dev->ops->send_request(dev, VHOST_USER_SET_VRING_KICK, &file);
90 virtio_user_queue_setup(struct virtio_user_dev *dev,
91 int (*fn)(struct virtio_user_dev *, uint32_t))
93 uint32_t i, queue_sel;
95 for (i = 0; i < dev->max_queue_pairs; ++i) {
96 queue_sel = 2 * i + VTNET_SQ_RQ_QUEUE_IDX;
97 if (fn(dev, queue_sel) < 0) {
98 PMD_DRV_LOG(INFO, "setup rx vq fails: %u", i);
102 for (i = 0; i < dev->max_queue_pairs; ++i) {
103 queue_sel = 2 * i + VTNET_SQ_TQ_QUEUE_IDX;
104 if (fn(dev, queue_sel) < 0) {
105 PMD_DRV_LOG(INFO, "setup tx vq fails: %u", i);
114 is_vhost_user_by_type(const char *path)
118 if (stat(path, &sb) == -1)
121 return S_ISSOCK(sb.st_mode);
125 virtio_user_start_device(struct virtio_user_dev *dev)
127 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
134 * We need to make sure that the locks will be
135 * taken in the correct order to avoid deadlocks.
137 * Before releasing this lock, this thread should
138 * not trigger any memory hotplug events.
140 * This is a temporary workaround, and should be
141 * replaced when we get proper supports from the
142 * memory subsystem in the future.
144 rte_rwlock_read_lock(&mcfg->memory_hotplug_lock);
145 pthread_mutex_lock(&dev->mutex);
147 if (is_vhost_user_by_type(dev->path) && dev->vhostfd < 0)
150 /* Step 0: tell vhost to create queues */
151 if (virtio_user_queue_setup(dev, virtio_user_create_queue) < 0)
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);
164 PMD_DRV_LOG(INFO, "set features: %" PRIx64, features);
166 /* Step 2: share memory regions */
167 ret = dev->ops->send_request(dev, VHOST_USER_SET_MEM_TABLE, NULL);
171 /* Step 3: kick queues */
172 if (virtio_user_queue_setup(dev, virtio_user_kick_queue) < 0)
175 /* Step 4: enable queues
176 * we enable the 1st queue pair by default.
178 dev->ops->enable_qp(dev, 0, 1);
181 pthread_mutex_unlock(&dev->mutex);
182 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
186 pthread_mutex_unlock(&dev->mutex);
187 rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
188 /* TODO: free resource here or caller to check */
192 int virtio_user_stop_device(struct virtio_user_dev *dev)
194 struct vhost_vring_state state;
198 pthread_mutex_lock(&dev->mutex);
202 for (i = 0; i < dev->max_queue_pairs; ++i)
203 dev->ops->enable_qp(dev, i, 0);
205 /* Stop the backend. */
206 for (i = 0; i < dev->max_queue_pairs * 2; ++i) {
208 if (dev->ops->send_request(dev, VHOST_USER_GET_VRING_BASE,
210 PMD_DRV_LOG(ERR, "get_vring_base failed, index=%u\n",
217 dev->started = false;
219 pthread_mutex_unlock(&dev->mutex);
225 parse_mac(struct virtio_user_dev *dev, const char *mac)
228 uint32_t tmp[ETHER_ADDR_LEN];
233 r = sscanf(mac, "%x:%x:%x:%x:%x:%x", &tmp[0],
234 &tmp[1], &tmp[2], &tmp[3], &tmp[4], &tmp[5]);
235 if (r == ETHER_ADDR_LEN) {
236 for (i = 0; i < ETHER_ADDR_LEN; ++i)
237 dev->mac_addr[i] = (uint8_t)tmp[i];
238 dev->mac_specified = 1;
240 /* ignore the wrong mac, use random mac */
241 PMD_DRV_LOG(ERR, "wrong format of mac: %s", mac);
246 virtio_user_dev_init_notify(struct virtio_user_dev *dev)
252 for (i = 0; i < VIRTIO_MAX_VIRTQUEUES; ++i) {
253 if (i >= dev->max_queue_pairs * 2) {
254 dev->kickfds[i] = -1;
255 dev->callfds[i] = -1;
259 /* May use invalid flag, but some backend uses kickfd and
260 * callfd as criteria to judge if dev is alive. so finally we
263 callfd = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK);
265 PMD_DRV_LOG(ERR, "callfd error, %s", strerror(errno));
268 kickfd = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK);
270 PMD_DRV_LOG(ERR, "kickfd error, %s", strerror(errno));
273 dev->callfds[i] = callfd;
274 dev->kickfds[i] = kickfd;
277 if (i < VIRTIO_MAX_VIRTQUEUES) {
278 for (j = 0; j <= i; ++j) {
279 close(dev->callfds[j]);
280 close(dev->kickfds[j]);
290 virtio_user_fill_intr_handle(struct virtio_user_dev *dev)
293 struct rte_eth_dev *eth_dev = &rte_eth_devices[dev->port_id];
295 if (!eth_dev->intr_handle) {
296 eth_dev->intr_handle = malloc(sizeof(*eth_dev->intr_handle));
297 if (!eth_dev->intr_handle) {
298 PMD_DRV_LOG(ERR, "fail to allocate intr_handle");
301 memset(eth_dev->intr_handle, 0, sizeof(*eth_dev->intr_handle));
304 for (i = 0; i < dev->max_queue_pairs; ++i)
305 eth_dev->intr_handle->efds[i] = dev->callfds[i];
306 eth_dev->intr_handle->nb_efd = dev->max_queue_pairs;
307 eth_dev->intr_handle->max_intr = dev->max_queue_pairs + 1;
308 eth_dev->intr_handle->type = RTE_INTR_HANDLE_VDEV;
309 /* For virtio vdev, no need to read counter for clean */
310 eth_dev->intr_handle->efd_counter_size = 0;
311 eth_dev->intr_handle->fd = -1;
312 if (dev->vhostfd >= 0)
313 eth_dev->intr_handle->fd = dev->vhostfd;
314 else if (dev->is_server)
315 eth_dev->intr_handle->fd = dev->listenfd;
321 virtio_user_mem_event_cb(enum rte_mem_event type __rte_unused,
322 const void *addr __rte_unused,
323 size_t len __rte_unused,
326 struct virtio_user_dev *dev = arg;
327 struct rte_memseg_list *msl;
330 /* ignore externally allocated memory */
331 msl = rte_mem_virt2memseg_list(addr);
335 pthread_mutex_lock(&dev->mutex);
337 if (dev->started == false)
340 /* Step 1: pause the active queues */
341 for (i = 0; i < dev->queue_pairs; i++)
342 dev->ops->enable_qp(dev, i, 0);
344 /* Step 2: update memory regions */
345 dev->ops->send_request(dev, VHOST_USER_SET_MEM_TABLE, NULL);
347 /* Step 3: resume the active queues */
348 for (i = 0; i < dev->queue_pairs; i++)
349 dev->ops->enable_qp(dev, i, 1);
352 pthread_mutex_unlock(&dev->mutex);
356 virtio_user_dev_setup(struct virtio_user_dev *dev)
361 dev->vhostfds = NULL;
364 if (dev->is_server) {
365 if (access(dev->path, F_OK) == 0 &&
366 !is_vhost_user_by_type(dev->path)) {
367 PMD_DRV_LOG(ERR, "Server mode doesn't support vhost-kernel!");
370 dev->ops = &virtio_ops_user;
372 if (is_vhost_user_by_type(dev->path)) {
373 dev->ops = &virtio_ops_user;
375 dev->ops = &virtio_ops_kernel;
377 dev->vhostfds = malloc(dev->max_queue_pairs *
379 dev->tapfds = malloc(dev->max_queue_pairs *
381 if (!dev->vhostfds || !dev->tapfds) {
382 PMD_INIT_LOG(ERR, "Failed to malloc");
386 for (q = 0; q < dev->max_queue_pairs; ++q) {
387 dev->vhostfds[q] = -1;
393 if (dev->ops->setup(dev) < 0)
396 if (virtio_user_dev_init_notify(dev) < 0)
399 if (virtio_user_fill_intr_handle(dev) < 0)
405 /* Use below macro to filter features from vhost backend */
406 #define VIRTIO_USER_SUPPORTED_FEATURES \
407 (1ULL << VIRTIO_NET_F_MAC | \
408 1ULL << VIRTIO_NET_F_STATUS | \
409 1ULL << VIRTIO_NET_F_MQ | \
410 1ULL << VIRTIO_NET_F_CTRL_MAC_ADDR | \
411 1ULL << VIRTIO_NET_F_CTRL_VQ | \
412 1ULL << VIRTIO_NET_F_CTRL_RX | \
413 1ULL << VIRTIO_NET_F_CTRL_VLAN | \
414 1ULL << VIRTIO_NET_F_CSUM | \
415 1ULL << VIRTIO_NET_F_HOST_TSO4 | \
416 1ULL << VIRTIO_NET_F_HOST_TSO6 | \
417 1ULL << VIRTIO_NET_F_MRG_RXBUF | \
418 1ULL << VIRTIO_RING_F_INDIRECT_DESC | \
419 1ULL << VIRTIO_NET_F_GUEST_CSUM | \
420 1ULL << VIRTIO_NET_F_GUEST_TSO4 | \
421 1ULL << VIRTIO_NET_F_GUEST_TSO6 | \
422 1ULL << VIRTIO_F_IN_ORDER | \
423 1ULL << VIRTIO_F_VERSION_1 | \
424 1ULL << VIRTIO_F_RING_PACKED)
427 virtio_user_dev_init(struct virtio_user_dev *dev, char *path, int queues,
428 int cq, int queue_size, const char *mac, char **ifname,
429 int mrg_rxbuf, int in_order, int packed_vq)
431 pthread_mutex_init(&dev->mutex, NULL);
432 snprintf(dev->path, PATH_MAX, "%s", path);
434 dev->max_queue_pairs = queues;
435 dev->queue_pairs = 1; /* mq disabled by default */
436 dev->queue_size = queue_size;
437 dev->mac_specified = 0;
438 dev->frontend_features = 0;
439 dev->unsupported_features = ~VIRTIO_USER_SUPPORTED_FEATURES;
443 dev->ifname = *ifname;
447 if (virtio_user_dev_setup(dev) < 0) {
448 PMD_INIT_LOG(ERR, "backend set up fails");
452 if (!dev->is_server) {
453 if (dev->ops->send_request(dev, VHOST_USER_SET_OWNER,
455 PMD_INIT_LOG(ERR, "set_owner fails: %s",
460 if (dev->ops->send_request(dev, VHOST_USER_GET_FEATURES,
461 &dev->device_features) < 0) {
462 PMD_INIT_LOG(ERR, "get_features failed: %s",
467 /* We just pretend vhost-user can support all these features.
468 * Note that this could be problematic that if some feature is
469 * negotiated but not supported by the vhost-user which comes
472 dev->device_features = VIRTIO_USER_SUPPORTED_FEATURES;
476 dev->unsupported_features |= (1ull << VIRTIO_NET_F_MRG_RXBUF);
479 dev->unsupported_features |= (1ull << VIRTIO_F_IN_ORDER);
482 dev->unsupported_features |= (1ull << VIRTIO_F_RING_PACKED);
484 if (dev->mac_specified)
485 dev->frontend_features |= (1ull << VIRTIO_NET_F_MAC);
487 dev->unsupported_features |= (1ull << VIRTIO_NET_F_MAC);
490 /* device does not really need to know anything about CQ,
491 * so if necessary, we just claim to support CQ
493 dev->frontend_features |= (1ull << VIRTIO_NET_F_CTRL_VQ);
495 dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_VQ);
496 /* Also disable features that depend on VIRTIO_NET_F_CTRL_VQ */
497 dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_RX);
498 dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_VLAN);
499 dev->unsupported_features |=
500 (1ull << VIRTIO_NET_F_GUEST_ANNOUNCE);
501 dev->unsupported_features |= (1ull << VIRTIO_NET_F_MQ);
502 dev->unsupported_features |=
503 (1ull << VIRTIO_NET_F_CTRL_MAC_ADDR);
506 /* The backend will not report this feature, we add it explicitly */
507 if (is_vhost_user_by_type(dev->path))
508 dev->frontend_features |= (1ull << VIRTIO_NET_F_STATUS);
512 * (frontend_features | backend_features) & ~unsupported_features;
514 dev->device_features |= dev->frontend_features;
515 dev->device_features &= ~dev->unsupported_features;
517 if (rte_mem_event_callback_register(VIRTIO_USER_MEM_EVENT_CLB_NAME,
518 virtio_user_mem_event_cb, dev)) {
519 if (rte_errno != ENOTSUP) {
520 PMD_INIT_LOG(ERR, "Failed to register mem event"
530 virtio_user_dev_uninit(struct virtio_user_dev *dev)
534 virtio_user_stop_device(dev);
536 rte_mem_event_callback_unregister(VIRTIO_USER_MEM_EVENT_CLB_NAME, dev);
538 for (i = 0; i < dev->max_queue_pairs * 2; ++i) {
539 close(dev->callfds[i]);
540 close(dev->kickfds[i]);
545 if (dev->is_server && dev->listenfd >= 0) {
546 close(dev->listenfd);
551 for (i = 0; i < dev->max_queue_pairs; ++i)
552 close(dev->vhostfds[i]);
564 virtio_user_handle_mq(struct virtio_user_dev *dev, uint16_t q_pairs)
569 if (q_pairs > dev->max_queue_pairs) {
570 PMD_INIT_LOG(ERR, "multi-q config %u, but only %u supported",
571 q_pairs, dev->max_queue_pairs);
575 /* Server mode can't enable queue pairs if vhostfd is invalid,
576 * always return 0 in this case.
578 if (!dev->is_server || dev->vhostfd >= 0) {
579 for (i = 0; i < q_pairs; ++i)
580 ret |= dev->ops->enable_qp(dev, i, 1);
581 for (i = q_pairs; i < dev->max_queue_pairs; ++i)
582 ret |= dev->ops->enable_qp(dev, i, 0);
584 dev->queue_pairs = q_pairs;
590 virtio_user_handle_ctrl_msg(struct virtio_user_dev *dev, struct vring *vring,
593 struct virtio_net_ctrl_hdr *hdr;
594 virtio_net_ctrl_ack status = ~0;
595 uint16_t i, idx_data, idx_status;
596 uint32_t n_descs = 0;
598 /* locate desc for header, data, and status */
599 idx_data = vring->desc[idx_hdr].next;
603 while (vring->desc[i].flags == VRING_DESC_F_NEXT) {
604 i = vring->desc[i].next;
608 /* locate desc for status */
612 hdr = (void *)(uintptr_t)vring->desc[idx_hdr].addr;
613 if (hdr->class == VIRTIO_NET_CTRL_MQ &&
614 hdr->cmd == VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET) {
617 queues = *(uint16_t *)(uintptr_t)vring->desc[idx_data].addr;
618 status = virtio_user_handle_mq(dev, queues);
622 *(virtio_net_ctrl_ack *)(uintptr_t)vring->desc[idx_status].addr = status;
628 desc_is_avail(struct vring_packed_desc *desc, bool wrap_counter)
630 return wrap_counter == !!(desc->flags & VRING_DESC_F_AVAIL(1)) &&
631 wrap_counter != !!(desc->flags & VRING_DESC_F_USED(1));
635 virtio_user_handle_ctrl_msg_packed(struct virtio_user_dev *dev,
636 struct vring_packed *vring,
639 struct virtio_net_ctrl_hdr *hdr;
640 virtio_net_ctrl_ack status = ~0;
641 uint16_t idx_data, idx_status;
642 /* initialize to one, header is first */
643 uint32_t n_descs = 1;
645 /* locate desc for header, data, and status */
646 idx_data = idx_hdr + 1;
647 if (idx_data >= dev->queue_size)
648 idx_data -= dev->queue_size;
652 idx_status = idx_data;
653 while (vring->desc[idx_status].flags & VRING_DESC_F_NEXT) {
655 if (idx_status >= dev->queue_size)
656 idx_status -= dev->queue_size;
660 hdr = (void *)(uintptr_t)vring->desc[idx_hdr].addr;
661 if (hdr->class == VIRTIO_NET_CTRL_MQ &&
662 hdr->cmd == VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET) {
665 queues = *(uint16_t *)(uintptr_t)
666 vring->desc[idx_data].addr;
667 status = virtio_user_handle_mq(dev, queues);
671 *(virtio_net_ctrl_ack *)(uintptr_t)
672 vring->desc[idx_status].addr = status;
674 /* Update used descriptor */
675 vring->desc[idx_hdr].id = vring->desc[idx_status].id;
676 vring->desc[idx_hdr].len = sizeof(status);
682 virtio_user_handle_cq_packed(struct virtio_user_dev *dev, uint16_t queue_idx)
684 struct virtio_user_queue *vq = &dev->packed_queues[queue_idx];
685 struct vring_packed *vring = &dev->packed_vrings[queue_idx];
688 while (desc_is_avail(&vring->desc[vq->used_idx],
689 vq->used_wrap_counter)) {
691 n_descs = virtio_user_handle_ctrl_msg_packed(dev, vring,
695 vring->desc[vq->used_idx].flags =
697 VRING_DESC_F_AVAIL(vq->used_wrap_counter) |
698 VRING_DESC_F_USED(vq->used_wrap_counter);
700 vq->used_idx += n_descs;
701 if (vq->used_idx >= dev->queue_size) {
702 vq->used_idx -= dev->queue_size;
703 vq->used_wrap_counter ^= 1;
709 virtio_user_handle_cq(struct virtio_user_dev *dev, uint16_t queue_idx)
711 uint16_t avail_idx, desc_idx;
712 struct vring_used_elem *uep;
714 struct vring *vring = &dev->vrings[queue_idx];
716 /* Consume avail ring, using used ring idx as first one */
717 while (vring->used->idx != vring->avail->idx) {
718 avail_idx = (vring->used->idx) & (vring->num - 1);
719 desc_idx = vring->avail->ring[avail_idx];
721 n_descs = virtio_user_handle_ctrl_msg(dev, vring, desc_idx);
723 /* Update used ring */
724 uep = &vring->used->ring[avail_idx];