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)) {
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) {
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) {
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 cleanup_vq_inflight(struct virtio_net *dev, struct vhost_virtqueue *vq)
288 if (!(dev->protocol_features &
289 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
292 if (vq_is_packed(dev)) {
293 if (vq->inflight_packed)
294 vq->inflight_packed = NULL;
296 if (vq->inflight_split)
297 vq->inflight_split = NULL;
300 if (vq->resubmit_inflight) {
301 if (vq->resubmit_inflight->resubmit_list) {
302 free(vq->resubmit_inflight->resubmit_list);
303 vq->resubmit_inflight->resubmit_list = NULL;
305 free(vq->resubmit_inflight);
306 vq->resubmit_inflight = NULL;
311 * Unmap any memory, close any file descriptors and
312 * free any memory owned by a device.
315 cleanup_device(struct virtio_net *dev, int destroy)
319 vhost_backend_cleanup(dev);
321 for (i = 0; i < dev->nr_vring; i++) {
322 cleanup_vq(dev->virtqueue[i], destroy);
323 cleanup_vq_inflight(dev, dev->virtqueue[i]);
328 vhost_free_async_mem(struct vhost_virtqueue *vq)
330 if (vq->async_pkts_pending)
331 rte_free(vq->async_pkts_pending);
332 if (vq->async_pkts_info)
333 rte_free(vq->async_pkts_info);
335 rte_free(vq->it_pool);
337 rte_free(vq->vec_pool);
339 vq->async_pkts_pending = NULL;
340 vq->async_pkts_info = NULL;
346 free_vq(struct virtio_net *dev, struct vhost_virtqueue *vq)
348 if (vq_is_packed(dev))
349 rte_free(vq->shadow_used_packed);
351 rte_free(vq->shadow_used_split);
352 vhost_free_async_mem(vq);
354 rte_free(vq->batch_copy_elems);
355 rte_mempool_free(vq->iotlb_pool);
360 * Release virtqueues and device memory.
363 free_device(struct virtio_net *dev)
367 for (i = 0; i < dev->nr_vring; i++)
368 free_vq(dev, dev->virtqueue[i]);
373 static __rte_always_inline int
374 log_translate(struct virtio_net *dev, struct vhost_virtqueue *vq)
376 if (likely(!(vq->ring_addrs.flags & (1 << VHOST_VRING_F_LOG))))
379 vq->log_guest_addr = translate_log_addr(dev, vq,
380 vq->ring_addrs.log_guest_addr);
381 if (vq->log_guest_addr == 0)
388 * Converts vring log address to GPA
389 * If IOMMU is enabled, the log address is IOVA
390 * If IOMMU not enabled, the log address is already GPA
392 * Caller should have iotlb_lock read-locked
395 translate_log_addr(struct virtio_net *dev, struct vhost_virtqueue *vq,
398 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
399 const uint64_t exp_size = sizeof(uint64_t);
401 uint64_t size = exp_size;
403 hva = vhost_iova_to_vva(dev, vq, log_addr,
404 &size, VHOST_ACCESS_RW);
406 if (size != exp_size)
409 gpa = hva_to_gpa(dev, hva, exp_size);
411 VHOST_LOG_CONFIG(ERR,
412 "VQ: Failed to find GPA for log_addr: 0x%"
413 PRIx64 " hva: 0x%" PRIx64 "\n",
423 /* Caller should have iotlb_lock read-locked */
425 vring_translate_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
427 uint64_t req_size, size;
429 req_size = sizeof(struct vring_desc) * vq->size;
431 vq->desc = (struct vring_desc *)(uintptr_t)vhost_iova_to_vva(dev, vq,
432 vq->ring_addrs.desc_user_addr,
433 &size, VHOST_ACCESS_RW);
434 if (!vq->desc || size != req_size)
437 req_size = sizeof(struct vring_avail);
438 req_size += sizeof(uint16_t) * vq->size;
439 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
440 req_size += sizeof(uint16_t);
442 vq->avail = (struct vring_avail *)(uintptr_t)vhost_iova_to_vva(dev, vq,
443 vq->ring_addrs.avail_user_addr,
444 &size, VHOST_ACCESS_RW);
445 if (!vq->avail || size != req_size)
448 req_size = sizeof(struct vring_used);
449 req_size += sizeof(struct vring_used_elem) * vq->size;
450 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
451 req_size += sizeof(uint16_t);
453 vq->used = (struct vring_used *)(uintptr_t)vhost_iova_to_vva(dev, vq,
454 vq->ring_addrs.used_user_addr,
455 &size, VHOST_ACCESS_RW);
456 if (!vq->used || size != req_size)
462 /* Caller should have iotlb_lock read-locked */
464 vring_translate_packed(struct virtio_net *dev, struct vhost_virtqueue *vq)
466 uint64_t req_size, size;
468 req_size = sizeof(struct vring_packed_desc) * vq->size;
470 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
471 vhost_iova_to_vva(dev, vq, vq->ring_addrs.desc_user_addr,
472 &size, VHOST_ACCESS_RW);
473 if (!vq->desc_packed || size != req_size)
476 req_size = sizeof(struct vring_packed_desc_event);
478 vq->driver_event = (struct vring_packed_desc_event *)(uintptr_t)
479 vhost_iova_to_vva(dev, vq, vq->ring_addrs.avail_user_addr,
480 &size, VHOST_ACCESS_RW);
481 if (!vq->driver_event || size != req_size)
484 req_size = sizeof(struct vring_packed_desc_event);
486 vq->device_event = (struct vring_packed_desc_event *)(uintptr_t)
487 vhost_iova_to_vva(dev, vq, vq->ring_addrs.used_user_addr,
488 &size, VHOST_ACCESS_RW);
489 if (!vq->device_event || size != req_size)
496 vring_translate(struct virtio_net *dev, struct vhost_virtqueue *vq)
499 if (!(dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
502 if (vq_is_packed(dev)) {
503 if (vring_translate_packed(dev, vq) < 0)
506 if (vring_translate_split(dev, vq) < 0)
510 if (log_translate(dev, vq) < 0)
519 vring_invalidate(struct virtio_net *dev, struct vhost_virtqueue *vq)
521 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
522 vhost_user_iotlb_wr_lock(vq);
528 vq->log_guest_addr = 0;
530 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
531 vhost_user_iotlb_wr_unlock(vq);
535 init_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
537 struct vhost_virtqueue *vq;
539 if (vring_idx >= VHOST_MAX_VRING) {
540 VHOST_LOG_CONFIG(ERR,
541 "Failed not init vring, out of bound (%d)\n",
546 vq = dev->virtqueue[vring_idx];
548 memset(vq, 0, sizeof(struct vhost_virtqueue));
550 vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
551 vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
552 vq->notif_enable = VIRTIO_UNINITIALIZED_NOTIF;
554 vhost_user_iotlb_init(dev, vring_idx);
555 /* Backends are set to -1 indicating an inactive device. */
560 reset_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
562 struct vhost_virtqueue *vq;
565 if (vring_idx >= VHOST_MAX_VRING) {
566 VHOST_LOG_CONFIG(ERR,
567 "Failed not init vring, out of bound (%d)\n",
572 vq = dev->virtqueue[vring_idx];
574 init_vring_queue(dev, vring_idx);
579 alloc_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
581 struct vhost_virtqueue *vq;
583 vq = rte_malloc(NULL, sizeof(struct vhost_virtqueue), 0);
585 VHOST_LOG_CONFIG(ERR,
586 "Failed to allocate memory for vring:%u.\n", vring_idx);
590 dev->virtqueue[vring_idx] = vq;
591 init_vring_queue(dev, vring_idx);
592 rte_spinlock_init(&vq->access_lock);
593 vq->avail_wrap_counter = 1;
594 vq->used_wrap_counter = 1;
595 vq->signalled_used_valid = false;
603 * Reset some variables in device structure, while keeping few
604 * others untouched, such as vid, ifname, nr_vring: they
605 * should be same unless the device is removed.
608 reset_device(struct virtio_net *dev)
613 dev->protocol_features = 0;
614 dev->flags &= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
616 for (i = 0; i < dev->nr_vring; i++)
617 reset_vring_queue(dev, i);
621 * Invoked when there is a new vhost-user connection established (when
622 * there is a new virtio device being attached).
625 vhost_new_device(void)
627 struct virtio_net *dev;
630 for (i = 0; i < MAX_VHOST_DEVICE; i++) {
631 if (vhost_devices[i] == NULL)
635 if (i == MAX_VHOST_DEVICE) {
636 VHOST_LOG_CONFIG(ERR,
637 "Failed to find a free slot for new device.\n");
641 dev = rte_zmalloc(NULL, sizeof(struct virtio_net), 0);
643 VHOST_LOG_CONFIG(ERR,
644 "Failed to allocate memory for new dev.\n");
648 vhost_devices[i] = dev;
650 dev->flags = VIRTIO_DEV_BUILTIN_VIRTIO_NET;
651 dev->slave_req_fd = -1;
652 dev->postcopy_ufd = -1;
653 rte_spinlock_init(&dev->slave_req_lock);
659 vhost_destroy_device_notify(struct virtio_net *dev)
661 struct rte_vdpa_device *vdpa_dev;
663 if (dev->flags & VIRTIO_DEV_RUNNING) {
664 vdpa_dev = dev->vdpa_dev;
666 vdpa_dev->ops->dev_close(dev->vid);
667 dev->flags &= ~VIRTIO_DEV_RUNNING;
668 dev->notify_ops->destroy_device(dev->vid);
673 * Invoked when there is the vhost-user connection is broken (when
674 * the virtio device is being detached).
677 vhost_destroy_device(int vid)
679 struct virtio_net *dev = get_device(vid);
684 vhost_destroy_device_notify(dev);
686 cleanup_device(dev, 1);
689 vhost_devices[vid] = NULL;
693 vhost_attach_vdpa_device(int vid, struct rte_vdpa_device *vdpa_dev)
695 struct virtio_net *dev = get_device(vid);
700 dev->vdpa_dev = vdpa_dev;
704 vhost_set_ifname(int vid, const char *if_name, unsigned int if_len)
706 struct virtio_net *dev;
709 dev = get_device(vid);
713 len = if_len > sizeof(dev->ifname) ?
714 sizeof(dev->ifname) : if_len;
716 strncpy(dev->ifname, if_name, len);
717 dev->ifname[sizeof(dev->ifname) - 1] = '\0';
721 vhost_set_builtin_virtio_net(int vid, bool enable)
723 struct virtio_net *dev = get_device(vid);
729 dev->flags |= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
731 dev->flags &= ~VIRTIO_DEV_BUILTIN_VIRTIO_NET;
735 vhost_enable_extbuf(int vid)
737 struct virtio_net *dev = get_device(vid);
746 vhost_enable_linearbuf(int vid)
748 struct virtio_net *dev = get_device(vid);
757 rte_vhost_get_mtu(int vid, uint16_t *mtu)
759 struct virtio_net *dev = get_device(vid);
761 if (dev == NULL || mtu == NULL)
764 if (!(dev->flags & VIRTIO_DEV_READY))
767 if (!(dev->features & (1ULL << VIRTIO_NET_F_MTU)))
776 rte_vhost_get_numa_node(int vid)
778 #ifdef RTE_LIBRTE_VHOST_NUMA
779 struct virtio_net *dev = get_device(vid);
783 if (dev == NULL || numa_available() != 0)
786 ret = get_mempolicy(&numa_node, NULL, 0, dev,
787 MPOL_F_NODE | MPOL_F_ADDR);
789 VHOST_LOG_CONFIG(ERR,
790 "(%d) failed to query numa node: %s\n",
791 vid, rte_strerror(errno));
803 rte_vhost_get_queue_num(int vid)
805 struct virtio_net *dev = get_device(vid);
810 return dev->nr_vring / 2;
814 rte_vhost_get_vring_num(int vid)
816 struct virtio_net *dev = get_device(vid);
821 return dev->nr_vring;
825 rte_vhost_get_ifname(int vid, char *buf, size_t len)
827 struct virtio_net *dev = get_device(vid);
829 if (dev == NULL || buf == NULL)
832 len = RTE_MIN(len, sizeof(dev->ifname));
834 strncpy(buf, dev->ifname, len);
841 rte_vhost_get_negotiated_features(int vid, uint64_t *features)
843 struct virtio_net *dev;
845 dev = get_device(vid);
846 if (dev == NULL || features == NULL)
849 *features = dev->features;
854 rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem)
856 struct virtio_net *dev;
857 struct rte_vhost_memory *m;
860 dev = get_device(vid);
861 if (dev == NULL || mem == NULL)
864 size = dev->mem->nregions * sizeof(struct rte_vhost_mem_region);
865 m = malloc(sizeof(struct rte_vhost_memory) + size);
869 m->nregions = dev->mem->nregions;
870 memcpy(m->regions, dev->mem->regions, size);
877 rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
878 struct rte_vhost_vring *vring)
880 struct virtio_net *dev;
881 struct vhost_virtqueue *vq;
883 dev = get_device(vid);
884 if (dev == NULL || vring == NULL)
887 if (vring_idx >= VHOST_MAX_VRING)
890 vq = dev->virtqueue[vring_idx];
894 if (vq_is_packed(dev)) {
895 vring->desc_packed = vq->desc_packed;
896 vring->driver_event = vq->driver_event;
897 vring->device_event = vq->device_event;
899 vring->desc = vq->desc;
900 vring->avail = vq->avail;
901 vring->used = vq->used;
903 vring->log_guest_addr = vq->log_guest_addr;
905 vring->callfd = vq->callfd;
906 vring->kickfd = vq->kickfd;
907 vring->size = vq->size;
913 rte_vhost_get_vhost_ring_inflight(int vid, uint16_t vring_idx,
914 struct rte_vhost_ring_inflight *vring)
916 struct virtio_net *dev;
917 struct vhost_virtqueue *vq;
919 dev = get_device(vid);
923 if (vring_idx >= VHOST_MAX_VRING)
926 vq = dev->virtqueue[vring_idx];
930 if (vq_is_packed(dev)) {
931 if (unlikely(!vq->inflight_packed))
934 vring->inflight_packed = vq->inflight_packed;
936 if (unlikely(!vq->inflight_split))
939 vring->inflight_split = vq->inflight_split;
942 vring->resubmit_inflight = vq->resubmit_inflight;
948 rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx,
951 struct vhost_virtqueue *vq;
952 struct virtio_net *dev;
954 dev = get_device(vid);
958 if (unlikely(!(dev->protocol_features &
959 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
962 if (unlikely(vq_is_packed(dev)))
965 if (unlikely(vring_idx >= VHOST_MAX_VRING))
968 vq = dev->virtqueue[vring_idx];
972 if (unlikely(!vq->inflight_split))
975 if (unlikely(idx >= vq->size))
978 vq->inflight_split->desc[idx].counter = vq->global_counter++;
979 vq->inflight_split->desc[idx].inflight = 1;
984 rte_vhost_set_inflight_desc_packed(int vid, uint16_t vring_idx,
985 uint16_t head, uint16_t last,
986 uint16_t *inflight_entry)
988 struct rte_vhost_inflight_info_packed *inflight_info;
989 struct virtio_net *dev;
990 struct vhost_virtqueue *vq;
991 struct vring_packed_desc *desc;
992 uint16_t old_free_head, free_head;
994 dev = get_device(vid);
998 if (unlikely(!(dev->protocol_features &
999 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1002 if (unlikely(!vq_is_packed(dev)))
1005 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1008 vq = dev->virtqueue[vring_idx];
1012 inflight_info = vq->inflight_packed;
1013 if (unlikely(!inflight_info))
1016 if (unlikely(head >= vq->size))
1019 desc = vq->desc_packed;
1020 old_free_head = inflight_info->old_free_head;
1021 if (unlikely(old_free_head >= vq->size))
1024 free_head = old_free_head;
1026 /* init header descriptor */
1027 inflight_info->desc[old_free_head].num = 0;
1028 inflight_info->desc[old_free_head].counter = vq->global_counter++;
1029 inflight_info->desc[old_free_head].inflight = 1;
1031 /* save desc entry in flight entry */
1032 while (head != ((last + 1) % vq->size)) {
1033 inflight_info->desc[old_free_head].num++;
1034 inflight_info->desc[free_head].addr = desc[head].addr;
1035 inflight_info->desc[free_head].len = desc[head].len;
1036 inflight_info->desc[free_head].flags = desc[head].flags;
1037 inflight_info->desc[free_head].id = desc[head].id;
1039 inflight_info->desc[old_free_head].last = free_head;
1040 free_head = inflight_info->desc[free_head].next;
1041 inflight_info->free_head = free_head;
1042 head = (head + 1) % vq->size;
1045 inflight_info->old_free_head = free_head;
1046 *inflight_entry = old_free_head;
1052 rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
1053 uint16_t last_used_idx, uint16_t idx)
1055 struct virtio_net *dev;
1056 struct vhost_virtqueue *vq;
1058 dev = get_device(vid);
1062 if (unlikely(!(dev->protocol_features &
1063 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1066 if (unlikely(vq_is_packed(dev)))
1069 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1072 vq = dev->virtqueue[vring_idx];
1076 if (unlikely(!vq->inflight_split))
1079 if (unlikely(idx >= vq->size))
1084 vq->inflight_split->desc[idx].inflight = 0;
1088 vq->inflight_split->used_idx = last_used_idx;
1093 rte_vhost_clr_inflight_desc_packed(int vid, uint16_t vring_idx,
1096 struct rte_vhost_inflight_info_packed *inflight_info;
1097 struct virtio_net *dev;
1098 struct vhost_virtqueue *vq;
1100 dev = get_device(vid);
1104 if (unlikely(!(dev->protocol_features &
1105 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1108 if (unlikely(!vq_is_packed(dev)))
1111 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1114 vq = dev->virtqueue[vring_idx];
1118 inflight_info = vq->inflight_packed;
1119 if (unlikely(!inflight_info))
1122 if (unlikely(head >= vq->size))
1127 inflight_info->desc[head].inflight = 0;
1131 inflight_info->old_free_head = inflight_info->free_head;
1132 inflight_info->old_used_idx = inflight_info->used_idx;
1133 inflight_info->old_used_wrap_counter = inflight_info->used_wrap_counter;
1139 rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx,
1142 struct virtio_net *dev;
1143 struct vhost_virtqueue *vq;
1145 dev = get_device(vid);
1149 if (unlikely(!(dev->protocol_features &
1150 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1153 if (unlikely(vq_is_packed(dev)))
1156 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1159 vq = dev->virtqueue[vring_idx];
1163 if (unlikely(!vq->inflight_split))
1166 vq->inflight_split->last_inflight_io = idx;
1171 rte_vhost_set_last_inflight_io_packed(int vid, uint16_t vring_idx,
1174 struct rte_vhost_inflight_info_packed *inflight_info;
1175 struct virtio_net *dev;
1176 struct vhost_virtqueue *vq;
1179 dev = get_device(vid);
1183 if (unlikely(!(dev->protocol_features &
1184 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1187 if (unlikely(!vq_is_packed(dev)))
1190 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1193 vq = dev->virtqueue[vring_idx];
1197 inflight_info = vq->inflight_packed;
1198 if (unlikely(!inflight_info))
1201 if (unlikely(head >= vq->size))
1204 last = inflight_info->desc[head].last;
1205 if (unlikely(last >= vq->size))
1208 inflight_info->desc[last].next = inflight_info->free_head;
1209 inflight_info->free_head = head;
1210 inflight_info->used_idx += inflight_info->desc[head].num;
1211 if (inflight_info->used_idx >= inflight_info->desc_num) {
1212 inflight_info->used_idx -= inflight_info->desc_num;
1213 inflight_info->used_wrap_counter =
1214 !inflight_info->used_wrap_counter;
1221 rte_vhost_vring_call(int vid, uint16_t vring_idx)
1223 struct virtio_net *dev;
1224 struct vhost_virtqueue *vq;
1226 dev = get_device(vid);
1230 if (vring_idx >= VHOST_MAX_VRING)
1233 vq = dev->virtqueue[vring_idx];
1237 if (vq_is_packed(dev))
1238 vhost_vring_call_packed(dev, vq);
1240 vhost_vring_call_split(dev, vq);
1246 rte_vhost_avail_entries(int vid, uint16_t queue_id)
1248 struct virtio_net *dev;
1249 struct vhost_virtqueue *vq;
1252 dev = get_device(vid);
1256 vq = dev->virtqueue[queue_id];
1258 rte_spinlock_lock(&vq->access_lock);
1260 if (unlikely(!vq->enabled || vq->avail == NULL))
1263 ret = *(volatile uint16_t *)&vq->avail->idx - vq->last_used_idx;
1266 rte_spinlock_unlock(&vq->access_lock);
1271 vhost_enable_notify_split(struct virtio_net *dev,
1272 struct vhost_virtqueue *vq, int enable)
1274 if (vq->used == NULL)
1277 if (!(dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))) {
1279 vq->used->flags &= ~VRING_USED_F_NO_NOTIFY;
1281 vq->used->flags |= VRING_USED_F_NO_NOTIFY;
1284 vhost_avail_event(vq) = vq->last_avail_idx;
1290 vhost_enable_notify_packed(struct virtio_net *dev,
1291 struct vhost_virtqueue *vq, int enable)
1295 if (vq->device_event == NULL)
1299 vq->device_event->flags = VRING_EVENT_F_DISABLE;
1303 flags = VRING_EVENT_F_ENABLE;
1304 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX)) {
1305 flags = VRING_EVENT_F_DESC;
1306 vq->device_event->off_wrap = vq->last_avail_idx |
1307 vq->avail_wrap_counter << 15;
1312 vq->device_event->flags = flags;
1317 vhost_enable_guest_notification(struct virtio_net *dev,
1318 struct vhost_virtqueue *vq, int enable)
1321 * If the virtqueue is not ready yet, it will be applied
1322 * when it will become ready.
1327 if (vq_is_packed(dev))
1328 return vhost_enable_notify_packed(dev, vq, enable);
1330 return vhost_enable_notify_split(dev, vq, enable);
1334 rte_vhost_enable_guest_notification(int vid, uint16_t queue_id, int enable)
1336 struct virtio_net *dev = get_device(vid);
1337 struct vhost_virtqueue *vq;
1343 vq = dev->virtqueue[queue_id];
1345 rte_spinlock_lock(&vq->access_lock);
1347 vq->notif_enable = enable;
1348 ret = vhost_enable_guest_notification(dev, vq, enable);
1350 rte_spinlock_unlock(&vq->access_lock);
1356 rte_vhost_log_write(int vid, uint64_t addr, uint64_t len)
1358 struct virtio_net *dev = get_device(vid);
1363 vhost_log_write(dev, addr, len);
1367 rte_vhost_log_used_vring(int vid, uint16_t vring_idx,
1368 uint64_t offset, uint64_t len)
1370 struct virtio_net *dev;
1371 struct vhost_virtqueue *vq;
1373 dev = get_device(vid);
1377 if (vring_idx >= VHOST_MAX_VRING)
1379 vq = dev->virtqueue[vring_idx];
1383 vhost_log_used_vring(dev, vq, offset, len);
1387 rte_vhost_rx_queue_count(int vid, uint16_t qid)
1389 struct virtio_net *dev;
1390 struct vhost_virtqueue *vq;
1393 dev = get_device(vid);
1397 if (unlikely(qid >= dev->nr_vring || (qid & 1) == 0)) {
1398 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1399 dev->vid, __func__, qid);
1403 vq = dev->virtqueue[qid];
1407 rte_spinlock_lock(&vq->access_lock);
1409 if (unlikely(vq->enabled == 0 || vq->avail == NULL))
1412 ret = *((volatile uint16_t *)&vq->avail->idx) - vq->last_avail_idx;
1415 rte_spinlock_unlock(&vq->access_lock);
1419 struct rte_vdpa_device *
1420 rte_vhost_get_vdpa_device(int vid)
1422 struct virtio_net *dev = get_device(vid);
1427 return dev->vdpa_dev;
1430 int rte_vhost_get_log_base(int vid, uint64_t *log_base,
1433 struct virtio_net *dev = get_device(vid);
1435 if (dev == NULL || log_base == NULL || log_size == NULL)
1438 *log_base = dev->log_base;
1439 *log_size = dev->log_size;
1444 int rte_vhost_get_vring_base(int vid, uint16_t queue_id,
1445 uint16_t *last_avail_idx, uint16_t *last_used_idx)
1447 struct vhost_virtqueue *vq;
1448 struct virtio_net *dev = get_device(vid);
1450 if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1453 vq = dev->virtqueue[queue_id];
1457 if (vq_is_packed(dev)) {
1458 *last_avail_idx = (vq->avail_wrap_counter << 15) |
1460 *last_used_idx = (vq->used_wrap_counter << 15) |
1463 *last_avail_idx = vq->last_avail_idx;
1464 *last_used_idx = vq->last_used_idx;
1470 int rte_vhost_set_vring_base(int vid, uint16_t queue_id,
1471 uint16_t last_avail_idx, uint16_t last_used_idx)
1473 struct vhost_virtqueue *vq;
1474 struct virtio_net *dev = get_device(vid);
1479 vq = dev->virtqueue[queue_id];
1483 if (vq_is_packed(dev)) {
1484 vq->last_avail_idx = last_avail_idx & 0x7fff;
1485 vq->avail_wrap_counter = !!(last_avail_idx & (1 << 15));
1486 vq->last_used_idx = last_used_idx & 0x7fff;
1487 vq->used_wrap_counter = !!(last_used_idx & (1 << 15));
1489 vq->last_avail_idx = last_avail_idx;
1490 vq->last_used_idx = last_used_idx;
1497 rte_vhost_get_vring_base_from_inflight(int vid,
1499 uint16_t *last_avail_idx,
1500 uint16_t *last_used_idx)
1502 struct rte_vhost_inflight_info_packed *inflight_info;
1503 struct virtio_net *dev = get_device(vid);
1505 if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1508 if (!vq_is_packed(dev))
1511 inflight_info = dev->virtqueue[queue_id]->inflight_packed;
1515 *last_avail_idx = (inflight_info->old_used_wrap_counter << 15) |
1516 inflight_info->old_used_idx;
1517 *last_used_idx = *last_avail_idx;
1522 int rte_vhost_extern_callback_register(int vid,
1523 struct rte_vhost_user_extern_ops const * const ops, void *ctx)
1525 struct virtio_net *dev = get_device(vid);
1527 if (dev == NULL || ops == NULL)
1530 dev->extern_ops = *ops;
1531 dev->extern_data = ctx;
1535 int rte_vhost_async_channel_register(int vid, uint16_t queue_id,
1537 struct rte_vhost_async_channel_ops *ops)
1539 struct vhost_virtqueue *vq;
1540 struct virtio_net *dev = get_device(vid);
1541 struct rte_vhost_async_features f;
1544 if (dev == NULL || ops == NULL)
1547 f.intval = features;
1549 vq = dev->virtqueue[queue_id];
1551 if (unlikely(vq == NULL || !dev->async_copy))
1554 /* packed queue is not supported */
1555 if (unlikely(vq_is_packed(dev) || !f.async_inorder)) {
1556 VHOST_LOG_CONFIG(ERR,
1557 "async copy is not supported on packed queue or non-inorder mode "
1558 "(vid %d, qid: %d)\n", vid, queue_id);
1562 if (unlikely(ops->check_completed_copies == NULL ||
1563 ops->transfer_data == NULL))
1566 rte_spinlock_lock(&vq->access_lock);
1568 if (unlikely(vq->async_registered)) {
1569 VHOST_LOG_CONFIG(ERR,
1570 "async register failed: channel already registered "
1571 "(vid %d, qid: %d)\n", vid, queue_id);
1575 #ifdef RTE_LIBRTE_VHOST_NUMA
1576 if (get_mempolicy(&node, NULL, 0, vq, MPOL_F_NODE | MPOL_F_ADDR)) {
1577 VHOST_LOG_CONFIG(ERR,
1578 "unable to get numa information in async register. "
1579 "allocating async buffer memory on the caller thread node\n");
1580 node = SOCKET_ID_ANY;
1583 node = SOCKET_ID_ANY;
1586 vq->async_pkts_pending = rte_malloc_socket(NULL,
1587 vq->size * sizeof(uintptr_t),
1588 RTE_CACHE_LINE_SIZE, node);
1589 vq->async_pkts_info = rte_malloc_socket(NULL,
1590 vq->size * sizeof(struct async_inflight_info),
1591 RTE_CACHE_LINE_SIZE, node);
1592 vq->it_pool = rte_malloc_socket(NULL,
1593 VHOST_MAX_ASYNC_IT * sizeof(struct rte_vhost_iov_iter),
1594 RTE_CACHE_LINE_SIZE, node);
1595 vq->vec_pool = rte_malloc_socket(NULL,
1596 VHOST_MAX_ASYNC_VEC * sizeof(struct iovec),
1597 RTE_CACHE_LINE_SIZE, node);
1598 if (!vq->async_pkts_pending || !vq->async_pkts_info ||
1599 !vq->it_pool || !vq->vec_pool) {
1600 vhost_free_async_mem(vq);
1601 VHOST_LOG_CONFIG(ERR,
1602 "async register failed: cannot allocate memory for vq data "
1603 "(vid %d, qid: %d)\n", vid, queue_id);
1607 vq->async_ops.check_completed_copies = ops->check_completed_copies;
1608 vq->async_ops.transfer_data = ops->transfer_data;
1610 vq->async_inorder = f.async_inorder;
1611 vq->async_threshold = f.async_threshold;
1613 vq->async_registered = true;
1616 rte_spinlock_unlock(&vq->access_lock);
1621 int rte_vhost_async_channel_unregister(int vid, uint16_t queue_id)
1623 struct vhost_virtqueue *vq;
1624 struct virtio_net *dev = get_device(vid);
1630 vq = dev->virtqueue[queue_id];
1637 if (!vq->async_registered)
1640 if (!rte_spinlock_trylock(&vq->access_lock)) {
1641 VHOST_LOG_CONFIG(ERR, "Failed to unregister async channel. "
1642 "virt queue busy.\n");
1646 if (vq->async_pkts_inflight_n) {
1647 VHOST_LOG_CONFIG(ERR, "Failed to unregister async channel. "
1648 "async inflight packets must be completed before unregistration.\n");
1653 vhost_free_async_mem(vq);
1655 vq->async_ops.transfer_data = NULL;
1656 vq->async_ops.check_completed_copies = NULL;
1657 vq->async_registered = false;
1660 rte_spinlock_unlock(&vq->access_lock);
1665 RTE_LOG_REGISTER(vhost_config_log_level, lib.vhost.config, INFO);
1666 RTE_LOG_REGISTER(vhost_data_log_level, lib.vhost.data, WARNING);