1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2017 Intel Corporation
5 #include <linux/vhost.h>
6 #include <linux/virtio_net.h>
10 #ifdef RTE_LIBRTE_VHOST_NUMA
15 #include <rte_errno.h>
16 #include <rte_ethdev.h>
18 #include <rte_string_fns.h>
19 #include <rte_memory.h>
20 #include <rte_malloc.h>
21 #include <rte_vhost.h>
22 #include <rte_rwlock.h>
26 #include "vhost_user.h"
28 struct virtio_net *vhost_devices[MAX_VHOST_DEVICE];
30 /* Called with iotlb_lock read-locked */
32 __vhost_iova_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
33 uint64_t iova, uint64_t *size, uint8_t perm)
35 uint64_t vva, tmp_size;
42 vva = vhost_user_iotlb_cache_find(vq, iova, &tmp_size, perm);
43 if (tmp_size == *size)
48 if (!vhost_user_iotlb_pending_miss(vq, iova, perm)) {
50 * iotlb_lock is read-locked for a full burst,
51 * but it only protects the iotlb cache.
52 * In case of IOTLB miss, we might block on the socket,
53 * which could cause a deadlock with QEMU if an IOTLB update
54 * is being handled. We can safely unlock here to avoid it.
56 vhost_user_iotlb_rd_unlock(vq);
58 vhost_user_iotlb_pending_insert(vq, iova, perm);
59 if (vhost_user_iotlb_miss(dev, iova, perm)) {
60 RTE_LOG(ERR, VHOST_CONFIG,
61 "IOTLB miss req failed for IOVA 0x%" PRIx64 "\n",
63 vhost_user_iotlb_pending_remove(vq, iova, 1, perm);
66 vhost_user_iotlb_rd_lock(vq);
72 #define VHOST_LOG_PAGE 4096
75 * Atomically set a bit in memory.
77 static __rte_always_inline void
78 vhost_set_bit(unsigned int nr, volatile uint8_t *addr)
80 #if defined(RTE_TOOLCHAIN_GCC) && (GCC_VERSION < 70100)
82 * __sync_ built-ins are deprecated, but __atomic_ ones
83 * are sub-optimized in older GCC versions.
85 __sync_fetch_and_or_1(addr, (1U << nr));
87 __atomic_fetch_or(addr, (1U << nr), __ATOMIC_RELAXED);
91 static __rte_always_inline void
92 vhost_log_page(uint8_t *log_base, uint64_t page)
94 vhost_set_bit(page % 8, &log_base[page / 8]);
98 __vhost_log_write(struct virtio_net *dev, uint64_t addr, uint64_t len)
102 if (unlikely(!dev->log_base || !len))
105 if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
108 /* To make sure guest memory updates are committed before logging */
111 page = addr / VHOST_LOG_PAGE;
112 while (page * VHOST_LOG_PAGE < addr + len) {
113 vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
119 __vhost_log_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
120 uint64_t iova, uint64_t len)
122 uint64_t hva, gpa, map_len;
125 hva = __vhost_iova_to_vva(dev, vq, iova, &map_len, VHOST_ACCESS_RW);
126 if (map_len != len) {
127 RTE_LOG(ERR, VHOST_CONFIG,
128 "Failed to write log for IOVA 0x%" PRIx64 ". No IOTLB entry found\n",
133 gpa = hva_to_gpa(dev, hva, len);
135 __vhost_log_write(dev, gpa, len);
139 __vhost_log_cache_sync(struct virtio_net *dev, struct vhost_virtqueue *vq)
141 unsigned long *log_base;
144 if (unlikely(!dev->log_base))
149 log_base = (unsigned long *)(uintptr_t)dev->log_base;
151 for (i = 0; i < vq->log_cache_nb_elem; i++) {
152 struct log_cache_entry *elem = vq->log_cache + i;
154 #if defined(RTE_TOOLCHAIN_GCC) && (GCC_VERSION < 70100)
156 * '__sync' builtins are deprecated, but '__atomic' ones
157 * are sub-optimized in older GCC versions.
159 __sync_fetch_and_or(log_base + elem->offset, elem->val);
161 __atomic_fetch_or(log_base + elem->offset, elem->val,
168 vq->log_cache_nb_elem = 0;
171 static __rte_always_inline void
172 vhost_log_cache_page(struct virtio_net *dev, struct vhost_virtqueue *vq,
175 uint32_t bit_nr = page % (sizeof(unsigned long) << 3);
176 uint32_t offset = page / (sizeof(unsigned long) << 3);
179 for (i = 0; i < vq->log_cache_nb_elem; i++) {
180 struct log_cache_entry *elem = vq->log_cache + i;
182 if (elem->offset == offset) {
183 elem->val |= (1UL << bit_nr);
188 if (unlikely(i >= VHOST_LOG_CACHE_NR)) {
190 * No more room for a new log cache entry,
191 * so write the dirty log map directly.
194 vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
199 vq->log_cache[i].offset = offset;
200 vq->log_cache[i].val = (1UL << bit_nr);
201 vq->log_cache_nb_elem++;
205 __vhost_log_cache_write(struct virtio_net *dev, struct vhost_virtqueue *vq,
206 uint64_t addr, uint64_t len)
210 if (unlikely(!dev->log_base || !len))
213 if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
216 page = addr / VHOST_LOG_PAGE;
217 while (page * VHOST_LOG_PAGE < addr + len) {
218 vhost_log_cache_page(dev, vq, page);
224 __vhost_log_cache_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
225 uint64_t iova, uint64_t len)
227 uint64_t hva, gpa, map_len;
230 hva = __vhost_iova_to_vva(dev, vq, iova, &map_len, VHOST_ACCESS_RW);
231 if (map_len != len) {
232 RTE_LOG(ERR, VHOST_CONFIG,
233 "Failed to write log for IOVA 0x%" PRIx64 ". No IOTLB entry found\n",
238 gpa = hva_to_gpa(dev, hva, len);
240 __vhost_log_cache_write(dev, vq, gpa, len);
244 vhost_alloc_copy_ind_table(struct virtio_net *dev, struct vhost_virtqueue *vq,
245 uint64_t desc_addr, uint64_t desc_len)
249 uint64_t len, remain = desc_len;
251 idesc = rte_malloc(__func__, desc_len, 0);
252 if (unlikely(!idesc))
255 dst = (uint64_t)(uintptr_t)idesc;
259 src = vhost_iova_to_vva(dev, vq, desc_addr, &len,
261 if (unlikely(!src || !len)) {
266 rte_memcpy((void *)(uintptr_t)dst, (void *)(uintptr_t)src, len);
277 cleanup_vq(struct vhost_virtqueue *vq, int destroy)
279 if ((vq->callfd >= 0) && (destroy != 0))
286 * Unmap any memory, close any file descriptors and
287 * free any memory owned by a device.
290 cleanup_device(struct virtio_net *dev, int destroy)
294 vhost_backend_cleanup(dev);
296 for (i = 0; i < dev->nr_vring; i++)
297 cleanup_vq(dev->virtqueue[i], destroy);
301 free_vq(struct virtio_net *dev, struct vhost_virtqueue *vq)
303 if (vq_is_packed(dev))
304 rte_free(vq->shadow_used_packed);
306 rte_free(vq->shadow_used_split);
307 rte_free(vq->batch_copy_elems);
308 rte_mempool_free(vq->iotlb_pool);
313 * Release virtqueues and device memory.
316 free_device(struct virtio_net *dev)
320 for (i = 0; i < dev->nr_vring; i++)
321 free_vq(dev, dev->virtqueue[i]);
327 vring_translate_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
329 uint64_t req_size, size;
331 req_size = sizeof(struct vring_desc) * vq->size;
333 vq->desc = (struct vring_desc *)(uintptr_t)vhost_iova_to_vva(dev, vq,
334 vq->ring_addrs.desc_user_addr,
335 &size, VHOST_ACCESS_RW);
336 if (!vq->desc || size != req_size)
339 req_size = sizeof(struct vring_avail);
340 req_size += sizeof(uint16_t) * vq->size;
341 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
342 req_size += sizeof(uint16_t);
344 vq->avail = (struct vring_avail *)(uintptr_t)vhost_iova_to_vva(dev, vq,
345 vq->ring_addrs.avail_user_addr,
346 &size, VHOST_ACCESS_RW);
347 if (!vq->avail || size != req_size)
350 req_size = sizeof(struct vring_used);
351 req_size += sizeof(struct vring_used_elem) * vq->size;
352 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
353 req_size += sizeof(uint16_t);
355 vq->used = (struct vring_used *)(uintptr_t)vhost_iova_to_vva(dev, vq,
356 vq->ring_addrs.used_user_addr,
357 &size, VHOST_ACCESS_RW);
358 if (!vq->used || size != req_size)
365 vring_translate_packed(struct virtio_net *dev, struct vhost_virtqueue *vq)
367 uint64_t req_size, size;
369 req_size = sizeof(struct vring_packed_desc) * vq->size;
371 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
372 vhost_iova_to_vva(dev, vq, vq->ring_addrs.desc_user_addr,
373 &size, VHOST_ACCESS_RW);
374 if (!vq->desc_packed || size != req_size)
377 req_size = sizeof(struct vring_packed_desc_event);
379 vq->driver_event = (struct vring_packed_desc_event *)(uintptr_t)
380 vhost_iova_to_vva(dev, vq, vq->ring_addrs.avail_user_addr,
381 &size, VHOST_ACCESS_RW);
382 if (!vq->driver_event || size != req_size)
385 req_size = sizeof(struct vring_packed_desc_event);
387 vq->device_event = (struct vring_packed_desc_event *)(uintptr_t)
388 vhost_iova_to_vva(dev, vq, vq->ring_addrs.used_user_addr,
389 &size, VHOST_ACCESS_RW);
390 if (!vq->device_event || size != req_size)
397 vring_translate(struct virtio_net *dev, struct vhost_virtqueue *vq)
400 if (!(dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
403 if (vq_is_packed(dev)) {
404 if (vring_translate_packed(dev, vq) < 0)
407 if (vring_translate_split(dev, vq) < 0)
416 vring_invalidate(struct virtio_net *dev, struct vhost_virtqueue *vq)
418 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
419 vhost_user_iotlb_wr_lock(vq);
425 vq->log_guest_addr = 0;
427 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
428 vhost_user_iotlb_wr_unlock(vq);
432 init_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
434 struct vhost_virtqueue *vq;
436 if (vring_idx >= VHOST_MAX_VRING) {
437 RTE_LOG(ERR, VHOST_CONFIG,
438 "Failed not init vring, out of bound (%d)\n",
443 vq = dev->virtqueue[vring_idx];
445 memset(vq, 0, sizeof(struct vhost_virtqueue));
447 vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
448 vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
450 vhost_user_iotlb_init(dev, vring_idx);
451 /* Backends are set to -1 indicating an inactive device. */
454 TAILQ_INIT(&vq->zmbuf_list);
458 reset_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
460 struct vhost_virtqueue *vq;
463 if (vring_idx >= VHOST_MAX_VRING) {
464 RTE_LOG(ERR, VHOST_CONFIG,
465 "Failed not init vring, out of bound (%d)\n",
470 vq = dev->virtqueue[vring_idx];
472 init_vring_queue(dev, vring_idx);
477 alloc_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
479 struct vhost_virtqueue *vq;
481 vq = rte_malloc(NULL, sizeof(struct vhost_virtqueue), 0);
483 RTE_LOG(ERR, VHOST_CONFIG,
484 "Failed to allocate memory for vring:%u.\n", vring_idx);
488 dev->virtqueue[vring_idx] = vq;
489 init_vring_queue(dev, vring_idx);
490 rte_spinlock_init(&vq->access_lock);
491 vq->avail_wrap_counter = 1;
492 vq->used_wrap_counter = 1;
493 vq->signalled_used_valid = false;
501 * Reset some variables in device structure, while keeping few
502 * others untouched, such as vid, ifname, nr_vring: they
503 * should be same unless the device is removed.
506 reset_device(struct virtio_net *dev)
511 dev->protocol_features = 0;
512 dev->flags &= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
514 for (i = 0; i < dev->nr_vring; i++)
515 reset_vring_queue(dev, i);
519 * Invoked when there is a new vhost-user connection established (when
520 * there is a new virtio device being attached).
523 vhost_new_device(void)
525 struct virtio_net *dev;
528 for (i = 0; i < MAX_VHOST_DEVICE; i++) {
529 if (vhost_devices[i] == NULL)
533 if (i == MAX_VHOST_DEVICE) {
534 RTE_LOG(ERR, VHOST_CONFIG,
535 "Failed to find a free slot for new device.\n");
539 dev = rte_zmalloc(NULL, sizeof(struct virtio_net), 0);
541 RTE_LOG(ERR, VHOST_CONFIG,
542 "Failed to allocate memory for new dev.\n");
546 vhost_devices[i] = dev;
548 dev->flags = VIRTIO_DEV_BUILTIN_VIRTIO_NET;
549 dev->slave_req_fd = -1;
550 dev->vdpa_dev_id = -1;
551 dev->postcopy_ufd = -1;
552 rte_spinlock_init(&dev->slave_req_lock);
558 vhost_destroy_device_notify(struct virtio_net *dev)
560 struct rte_vdpa_device *vdpa_dev;
563 if (dev->flags & VIRTIO_DEV_RUNNING) {
564 did = dev->vdpa_dev_id;
565 vdpa_dev = rte_vdpa_get_device(did);
566 if (vdpa_dev && vdpa_dev->ops->dev_close)
567 vdpa_dev->ops->dev_close(dev->vid);
568 dev->flags &= ~VIRTIO_DEV_RUNNING;
569 dev->notify_ops->destroy_device(dev->vid);
574 * Invoked when there is the vhost-user connection is broken (when
575 * the virtio device is being detached).
578 vhost_destroy_device(int vid)
580 struct virtio_net *dev = get_device(vid);
585 vhost_destroy_device_notify(dev);
587 cleanup_device(dev, 1);
590 vhost_devices[vid] = NULL;
594 vhost_attach_vdpa_device(int vid, int did)
596 struct virtio_net *dev = get_device(vid);
601 if (rte_vdpa_get_device(did) == NULL)
604 dev->vdpa_dev_id = did;
608 vhost_set_ifname(int vid, const char *if_name, unsigned int if_len)
610 struct virtio_net *dev;
613 dev = get_device(vid);
617 len = if_len > sizeof(dev->ifname) ?
618 sizeof(dev->ifname) : if_len;
620 strncpy(dev->ifname, if_name, len);
621 dev->ifname[sizeof(dev->ifname) - 1] = '\0';
625 vhost_enable_dequeue_zero_copy(int vid)
627 struct virtio_net *dev = get_device(vid);
632 dev->dequeue_zero_copy = 1;
636 vhost_set_builtin_virtio_net(int vid, bool enable)
638 struct virtio_net *dev = get_device(vid);
644 dev->flags |= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
646 dev->flags &= ~VIRTIO_DEV_BUILTIN_VIRTIO_NET;
650 rte_vhost_get_mtu(int vid, uint16_t *mtu)
652 struct virtio_net *dev = get_device(vid);
654 if (dev == NULL || mtu == NULL)
657 if (!(dev->flags & VIRTIO_DEV_READY))
660 if (!(dev->features & (1ULL << VIRTIO_NET_F_MTU)))
669 rte_vhost_get_numa_node(int vid)
671 #ifdef RTE_LIBRTE_VHOST_NUMA
672 struct virtio_net *dev = get_device(vid);
676 if (dev == NULL || numa_available() != 0)
679 ret = get_mempolicy(&numa_node, NULL, 0, dev,
680 MPOL_F_NODE | MPOL_F_ADDR);
682 RTE_LOG(ERR, VHOST_CONFIG,
683 "(%d) failed to query numa node: %s\n",
684 vid, rte_strerror(errno));
696 rte_vhost_get_queue_num(int vid)
698 struct virtio_net *dev = get_device(vid);
703 return dev->nr_vring / 2;
707 rte_vhost_get_vring_num(int vid)
709 struct virtio_net *dev = get_device(vid);
714 return dev->nr_vring;
718 rte_vhost_get_ifname(int vid, char *buf, size_t len)
720 struct virtio_net *dev = get_device(vid);
722 if (dev == NULL || buf == NULL)
725 len = RTE_MIN(len, sizeof(dev->ifname));
727 strncpy(buf, dev->ifname, len);
734 rte_vhost_get_negotiated_features(int vid, uint64_t *features)
736 struct virtio_net *dev;
738 dev = get_device(vid);
739 if (dev == NULL || features == NULL)
742 *features = dev->features;
747 rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem)
749 struct virtio_net *dev;
750 struct rte_vhost_memory *m;
753 dev = get_device(vid);
754 if (dev == NULL || mem == NULL)
757 size = dev->mem->nregions * sizeof(struct rte_vhost_mem_region);
758 m = malloc(sizeof(struct rte_vhost_memory) + size);
762 m->nregions = dev->mem->nregions;
763 memcpy(m->regions, dev->mem->regions, size);
770 rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
771 struct rte_vhost_vring *vring)
773 struct virtio_net *dev;
774 struct vhost_virtqueue *vq;
776 dev = get_device(vid);
777 if (dev == NULL || vring == NULL)
780 if (vring_idx >= VHOST_MAX_VRING)
783 vq = dev->virtqueue[vring_idx];
787 vring->desc = vq->desc;
788 vring->avail = vq->avail;
789 vring->used = vq->used;
790 vring->log_guest_addr = vq->log_guest_addr;
792 vring->callfd = vq->callfd;
793 vring->kickfd = vq->kickfd;
794 vring->size = vq->size;
800 rte_vhost_vring_call(int vid, uint16_t vring_idx)
802 struct virtio_net *dev;
803 struct vhost_virtqueue *vq;
805 dev = get_device(vid);
809 if (vring_idx >= VHOST_MAX_VRING)
812 vq = dev->virtqueue[vring_idx];
816 if (vq_is_packed(dev))
817 vhost_vring_call_packed(dev, vq);
819 vhost_vring_call_split(dev, vq);
825 rte_vhost_avail_entries(int vid, uint16_t queue_id)
827 struct virtio_net *dev;
828 struct vhost_virtqueue *vq;
831 dev = get_device(vid);
835 vq = dev->virtqueue[queue_id];
837 rte_spinlock_lock(&vq->access_lock);
839 if (unlikely(!vq->enabled || vq->avail == NULL))
842 ret = *(volatile uint16_t *)&vq->avail->idx - vq->last_used_idx;
845 rte_spinlock_unlock(&vq->access_lock);
850 vhost_enable_notify_split(struct virtio_net *dev,
851 struct vhost_virtqueue *vq, int enable)
853 if (vq->used == NULL)
856 if (!(dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))) {
858 vq->used->flags &= ~VRING_USED_F_NO_NOTIFY;
860 vq->used->flags |= VRING_USED_F_NO_NOTIFY;
863 vhost_avail_event(vq) = vq->last_avail_idx;
869 vhost_enable_notify_packed(struct virtio_net *dev,
870 struct vhost_virtqueue *vq, int enable)
874 if (vq->device_event == NULL)
878 vq->device_event->flags = VRING_EVENT_F_DISABLE;
882 flags = VRING_EVENT_F_ENABLE;
883 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX)) {
884 flags = VRING_EVENT_F_DESC;
885 vq->device_event->off_wrap = vq->last_avail_idx |
886 vq->avail_wrap_counter << 15;
891 vq->device_event->flags = flags;
896 rte_vhost_enable_guest_notification(int vid, uint16_t queue_id, int enable)
898 struct virtio_net *dev = get_device(vid);
899 struct vhost_virtqueue *vq;
905 vq = dev->virtqueue[queue_id];
907 rte_spinlock_lock(&vq->access_lock);
909 if (vq_is_packed(dev))
910 ret = vhost_enable_notify_packed(dev, vq, enable);
912 ret = vhost_enable_notify_split(dev, vq, enable);
914 rte_spinlock_unlock(&vq->access_lock);
920 rte_vhost_log_write(int vid, uint64_t addr, uint64_t len)
922 struct virtio_net *dev = get_device(vid);
927 vhost_log_write(dev, addr, len);
931 rte_vhost_log_used_vring(int vid, uint16_t vring_idx,
932 uint64_t offset, uint64_t len)
934 struct virtio_net *dev;
935 struct vhost_virtqueue *vq;
937 dev = get_device(vid);
941 if (vring_idx >= VHOST_MAX_VRING)
943 vq = dev->virtqueue[vring_idx];
947 vhost_log_used_vring(dev, vq, offset, len);
951 rte_vhost_rx_queue_count(int vid, uint16_t qid)
953 struct virtio_net *dev;
954 struct vhost_virtqueue *vq;
957 dev = get_device(vid);
961 if (unlikely(qid >= dev->nr_vring || (qid & 1) == 0)) {
962 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
963 dev->vid, __func__, qid);
967 vq = dev->virtqueue[qid];
971 rte_spinlock_lock(&vq->access_lock);
973 if (unlikely(vq->enabled == 0 || vq->avail == NULL))
976 ret = *((volatile uint16_t *)&vq->avail->idx) - vq->last_avail_idx;
979 rte_spinlock_unlock(&vq->access_lock);
983 int rte_vhost_get_vdpa_device_id(int vid)
985 struct virtio_net *dev = get_device(vid);
990 return dev->vdpa_dev_id;
993 int rte_vhost_get_log_base(int vid, uint64_t *log_base,
996 struct virtio_net *dev = get_device(vid);
998 if (dev == NULL || log_base == NULL || log_size == NULL)
1001 *log_base = dev->log_base;
1002 *log_size = dev->log_size;
1007 int rte_vhost_get_vring_base(int vid, uint16_t queue_id,
1008 uint16_t *last_avail_idx, uint16_t *last_used_idx)
1010 struct virtio_net *dev = get_device(vid);
1012 if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1015 *last_avail_idx = dev->virtqueue[queue_id]->last_avail_idx;
1016 *last_used_idx = dev->virtqueue[queue_id]->last_used_idx;
1021 int rte_vhost_set_vring_base(int vid, uint16_t queue_id,
1022 uint16_t last_avail_idx, uint16_t last_used_idx)
1024 struct virtio_net *dev = get_device(vid);
1029 dev->virtqueue[queue_id]->last_avail_idx = last_avail_idx;
1030 dev->virtqueue[queue_id]->last_used_idx = last_used_idx;
1035 int rte_vhost_extern_callback_register(int vid,
1036 struct rte_vhost_user_extern_ops const * const ops, void *ctx)
1038 struct virtio_net *dev = get_device(vid);
1040 if (dev == NULL || ops == NULL)
1043 dev->extern_ops = *ops;
1044 dev->extern_data = ctx;