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 free_vq(struct virtio_net *dev, struct vhost_virtqueue *vq)
330 if (vq_is_packed(dev))
331 rte_free(vq->shadow_used_packed);
333 rte_free(vq->shadow_used_split);
334 if (vq->async_pkts_pending)
335 rte_free(vq->async_pkts_pending);
336 if (vq->async_pending_info)
337 rte_free(vq->async_pending_info);
339 rte_free(vq->batch_copy_elems);
340 rte_mempool_free(vq->iotlb_pool);
345 * Release virtqueues and device memory.
348 free_device(struct virtio_net *dev)
352 for (i = 0; i < dev->nr_vring; i++)
353 free_vq(dev, dev->virtqueue[i]);
358 static __rte_always_inline int
359 log_translate(struct virtio_net *dev, struct vhost_virtqueue *vq)
361 if (likely(!(vq->ring_addrs.flags & (1 << VHOST_VRING_F_LOG))))
364 vq->log_guest_addr = translate_log_addr(dev, vq,
365 vq->ring_addrs.log_guest_addr);
366 if (vq->log_guest_addr == 0)
373 * Converts vring log address to GPA
374 * If IOMMU is enabled, the log address is IOVA
375 * If IOMMU not enabled, the log address is already GPA
377 * Caller should have iotlb_lock read-locked
380 translate_log_addr(struct virtio_net *dev, struct vhost_virtqueue *vq,
383 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
384 const uint64_t exp_size = sizeof(uint64_t);
386 uint64_t size = exp_size;
388 hva = vhost_iova_to_vva(dev, vq, log_addr,
389 &size, VHOST_ACCESS_RW);
391 if (size != exp_size)
394 gpa = hva_to_gpa(dev, hva, exp_size);
396 VHOST_LOG_CONFIG(ERR,
397 "VQ: Failed to find GPA for log_addr: 0x%"
398 PRIx64 " hva: 0x%" PRIx64 "\n",
408 /* Caller should have iotlb_lock read-locked */
410 vring_translate_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
412 uint64_t req_size, size;
414 req_size = sizeof(struct vring_desc) * vq->size;
416 vq->desc = (struct vring_desc *)(uintptr_t)vhost_iova_to_vva(dev, vq,
417 vq->ring_addrs.desc_user_addr,
418 &size, VHOST_ACCESS_RW);
419 if (!vq->desc || size != req_size)
422 req_size = sizeof(struct vring_avail);
423 req_size += sizeof(uint16_t) * vq->size;
424 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
425 req_size += sizeof(uint16_t);
427 vq->avail = (struct vring_avail *)(uintptr_t)vhost_iova_to_vva(dev, vq,
428 vq->ring_addrs.avail_user_addr,
429 &size, VHOST_ACCESS_RW);
430 if (!vq->avail || size != req_size)
433 req_size = sizeof(struct vring_used);
434 req_size += sizeof(struct vring_used_elem) * vq->size;
435 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
436 req_size += sizeof(uint16_t);
438 vq->used = (struct vring_used *)(uintptr_t)vhost_iova_to_vva(dev, vq,
439 vq->ring_addrs.used_user_addr,
440 &size, VHOST_ACCESS_RW);
441 if (!vq->used || size != req_size)
447 /* Caller should have iotlb_lock read-locked */
449 vring_translate_packed(struct virtio_net *dev, struct vhost_virtqueue *vq)
451 uint64_t req_size, size;
453 req_size = sizeof(struct vring_packed_desc) * vq->size;
455 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
456 vhost_iova_to_vva(dev, vq, vq->ring_addrs.desc_user_addr,
457 &size, VHOST_ACCESS_RW);
458 if (!vq->desc_packed || size != req_size)
461 req_size = sizeof(struct vring_packed_desc_event);
463 vq->driver_event = (struct vring_packed_desc_event *)(uintptr_t)
464 vhost_iova_to_vva(dev, vq, vq->ring_addrs.avail_user_addr,
465 &size, VHOST_ACCESS_RW);
466 if (!vq->driver_event || size != req_size)
469 req_size = sizeof(struct vring_packed_desc_event);
471 vq->device_event = (struct vring_packed_desc_event *)(uintptr_t)
472 vhost_iova_to_vva(dev, vq, vq->ring_addrs.used_user_addr,
473 &size, VHOST_ACCESS_RW);
474 if (!vq->device_event || size != req_size)
481 vring_translate(struct virtio_net *dev, struct vhost_virtqueue *vq)
484 if (!(dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
487 if (vq_is_packed(dev)) {
488 if (vring_translate_packed(dev, vq) < 0)
491 if (vring_translate_split(dev, vq) < 0)
495 if (log_translate(dev, vq) < 0)
504 vring_invalidate(struct virtio_net *dev, struct vhost_virtqueue *vq)
506 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
507 vhost_user_iotlb_wr_lock(vq);
513 vq->log_guest_addr = 0;
515 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
516 vhost_user_iotlb_wr_unlock(vq);
520 init_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
522 struct vhost_virtqueue *vq;
524 if (vring_idx >= VHOST_MAX_VRING) {
525 VHOST_LOG_CONFIG(ERR,
526 "Failed not init vring, out of bound (%d)\n",
531 vq = dev->virtqueue[vring_idx];
533 memset(vq, 0, sizeof(struct vhost_virtqueue));
535 vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
536 vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
538 vhost_user_iotlb_init(dev, vring_idx);
539 /* Backends are set to -1 indicating an inactive device. */
542 TAILQ_INIT(&vq->zmbuf_list);
546 reset_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
548 struct vhost_virtqueue *vq;
551 if (vring_idx >= VHOST_MAX_VRING) {
552 VHOST_LOG_CONFIG(ERR,
553 "Failed not init vring, out of bound (%d)\n",
558 vq = dev->virtqueue[vring_idx];
560 init_vring_queue(dev, vring_idx);
565 alloc_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
567 struct vhost_virtqueue *vq;
569 vq = rte_malloc(NULL, sizeof(struct vhost_virtqueue), 0);
571 VHOST_LOG_CONFIG(ERR,
572 "Failed to allocate memory for vring:%u.\n", vring_idx);
576 dev->virtqueue[vring_idx] = vq;
577 init_vring_queue(dev, vring_idx);
578 rte_spinlock_init(&vq->access_lock);
579 vq->avail_wrap_counter = 1;
580 vq->used_wrap_counter = 1;
581 vq->signalled_used_valid = false;
589 * Reset some variables in device structure, while keeping few
590 * others untouched, such as vid, ifname, nr_vring: they
591 * should be same unless the device is removed.
594 reset_device(struct virtio_net *dev)
599 dev->protocol_features = 0;
600 dev->flags &= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
602 for (i = 0; i < dev->nr_vring; i++)
603 reset_vring_queue(dev, i);
607 * Invoked when there is a new vhost-user connection established (when
608 * there is a new virtio device being attached).
611 vhost_new_device(void)
613 struct virtio_net *dev;
616 for (i = 0; i < MAX_VHOST_DEVICE; i++) {
617 if (vhost_devices[i] == NULL)
621 if (i == MAX_VHOST_DEVICE) {
622 VHOST_LOG_CONFIG(ERR,
623 "Failed to find a free slot for new device.\n");
627 dev = rte_zmalloc(NULL, sizeof(struct virtio_net), 0);
629 VHOST_LOG_CONFIG(ERR,
630 "Failed to allocate memory for new dev.\n");
634 vhost_devices[i] = dev;
636 dev->flags = VIRTIO_DEV_BUILTIN_VIRTIO_NET;
637 dev->slave_req_fd = -1;
638 dev->postcopy_ufd = -1;
639 rte_spinlock_init(&dev->slave_req_lock);
645 vhost_destroy_device_notify(struct virtio_net *dev)
647 struct rte_vdpa_device *vdpa_dev;
649 if (dev->flags & VIRTIO_DEV_RUNNING) {
650 vdpa_dev = dev->vdpa_dev;
652 vdpa_dev->ops->dev_close(dev->vid);
653 dev->flags &= ~VIRTIO_DEV_RUNNING;
654 dev->notify_ops->destroy_device(dev->vid);
659 * Invoked when there is the vhost-user connection is broken (when
660 * the virtio device is being detached).
663 vhost_destroy_device(int vid)
665 struct virtio_net *dev = get_device(vid);
670 vhost_destroy_device_notify(dev);
672 cleanup_device(dev, 1);
675 vhost_devices[vid] = NULL;
679 vhost_attach_vdpa_device(int vid, struct rte_vdpa_device *vdpa_dev)
681 struct virtio_net *dev = get_device(vid);
686 dev->vdpa_dev = vdpa_dev;
690 vhost_set_ifname(int vid, const char *if_name, unsigned int if_len)
692 struct virtio_net *dev;
695 dev = get_device(vid);
699 len = if_len > sizeof(dev->ifname) ?
700 sizeof(dev->ifname) : if_len;
702 strncpy(dev->ifname, if_name, len);
703 dev->ifname[sizeof(dev->ifname) - 1] = '\0';
707 vhost_enable_dequeue_zero_copy(int vid)
709 struct virtio_net *dev = get_device(vid);
714 dev->dequeue_zero_copy = 1;
715 VHOST_LOG_CONFIG(INFO, "dequeue zero copy is enabled\n");
719 vhost_set_builtin_virtio_net(int vid, bool enable)
721 struct virtio_net *dev = get_device(vid);
727 dev->flags |= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
729 dev->flags &= ~VIRTIO_DEV_BUILTIN_VIRTIO_NET;
733 vhost_enable_extbuf(int vid)
735 struct virtio_net *dev = get_device(vid);
744 vhost_enable_linearbuf(int vid)
746 struct virtio_net *dev = get_device(vid);
755 rte_vhost_get_mtu(int vid, uint16_t *mtu)
757 struct virtio_net *dev = get_device(vid);
759 if (dev == NULL || mtu == NULL)
762 if (!(dev->flags & VIRTIO_DEV_READY))
765 if (!(dev->features & (1ULL << VIRTIO_NET_F_MTU)))
774 rte_vhost_get_numa_node(int vid)
776 #ifdef RTE_LIBRTE_VHOST_NUMA
777 struct virtio_net *dev = get_device(vid);
781 if (dev == NULL || numa_available() != 0)
784 ret = get_mempolicy(&numa_node, NULL, 0, dev,
785 MPOL_F_NODE | MPOL_F_ADDR);
787 VHOST_LOG_CONFIG(ERR,
788 "(%d) failed to query numa node: %s\n",
789 vid, rte_strerror(errno));
801 rte_vhost_get_queue_num(int vid)
803 struct virtio_net *dev = get_device(vid);
808 return dev->nr_vring / 2;
812 rte_vhost_get_vring_num(int vid)
814 struct virtio_net *dev = get_device(vid);
819 return dev->nr_vring;
823 rte_vhost_get_ifname(int vid, char *buf, size_t len)
825 struct virtio_net *dev = get_device(vid);
827 if (dev == NULL || buf == NULL)
830 len = RTE_MIN(len, sizeof(dev->ifname));
832 strncpy(buf, dev->ifname, len);
839 rte_vhost_get_negotiated_features(int vid, uint64_t *features)
841 struct virtio_net *dev;
843 dev = get_device(vid);
844 if (dev == NULL || features == NULL)
847 *features = dev->features;
852 rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem)
854 struct virtio_net *dev;
855 struct rte_vhost_memory *m;
858 dev = get_device(vid);
859 if (dev == NULL || mem == NULL)
862 size = dev->mem->nregions * sizeof(struct rte_vhost_mem_region);
863 m = malloc(sizeof(struct rte_vhost_memory) + size);
867 m->nregions = dev->mem->nregions;
868 memcpy(m->regions, dev->mem->regions, size);
875 rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
876 struct rte_vhost_vring *vring)
878 struct virtio_net *dev;
879 struct vhost_virtqueue *vq;
881 dev = get_device(vid);
882 if (dev == NULL || vring == NULL)
885 if (vring_idx >= VHOST_MAX_VRING)
888 vq = dev->virtqueue[vring_idx];
892 if (vq_is_packed(dev)) {
893 vring->desc_packed = vq->desc_packed;
894 vring->driver_event = vq->driver_event;
895 vring->device_event = vq->device_event;
897 vring->desc = vq->desc;
898 vring->avail = vq->avail;
899 vring->used = vq->used;
901 vring->log_guest_addr = vq->log_guest_addr;
903 vring->callfd = vq->callfd;
904 vring->kickfd = vq->kickfd;
905 vring->size = vq->size;
911 rte_vhost_get_vhost_ring_inflight(int vid, uint16_t vring_idx,
912 struct rte_vhost_ring_inflight *vring)
914 struct virtio_net *dev;
915 struct vhost_virtqueue *vq;
917 dev = get_device(vid);
921 if (vring_idx >= VHOST_MAX_VRING)
924 vq = dev->virtqueue[vring_idx];
928 if (vq_is_packed(dev)) {
929 if (unlikely(!vq->inflight_packed))
932 vring->inflight_packed = vq->inflight_packed;
934 if (unlikely(!vq->inflight_split))
937 vring->inflight_split = vq->inflight_split;
940 vring->resubmit_inflight = vq->resubmit_inflight;
946 rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx,
949 struct vhost_virtqueue *vq;
950 struct virtio_net *dev;
952 dev = get_device(vid);
956 if (unlikely(!(dev->protocol_features &
957 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
960 if (unlikely(vq_is_packed(dev)))
963 if (unlikely(vring_idx >= VHOST_MAX_VRING))
966 vq = dev->virtqueue[vring_idx];
970 if (unlikely(!vq->inflight_split))
973 if (unlikely(idx >= vq->size))
976 vq->inflight_split->desc[idx].counter = vq->global_counter++;
977 vq->inflight_split->desc[idx].inflight = 1;
982 rte_vhost_set_inflight_desc_packed(int vid, uint16_t vring_idx,
983 uint16_t head, uint16_t last,
984 uint16_t *inflight_entry)
986 struct rte_vhost_inflight_info_packed *inflight_info;
987 struct virtio_net *dev;
988 struct vhost_virtqueue *vq;
989 struct vring_packed_desc *desc;
990 uint16_t old_free_head, free_head;
992 dev = get_device(vid);
996 if (unlikely(!(dev->protocol_features &
997 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1000 if (unlikely(!vq_is_packed(dev)))
1003 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1006 vq = dev->virtqueue[vring_idx];
1010 inflight_info = vq->inflight_packed;
1011 if (unlikely(!inflight_info))
1014 if (unlikely(head >= vq->size))
1017 desc = vq->desc_packed;
1018 old_free_head = inflight_info->old_free_head;
1019 if (unlikely(old_free_head >= vq->size))
1022 free_head = old_free_head;
1024 /* init header descriptor */
1025 inflight_info->desc[old_free_head].num = 0;
1026 inflight_info->desc[old_free_head].counter = vq->global_counter++;
1027 inflight_info->desc[old_free_head].inflight = 1;
1029 /* save desc entry in flight entry */
1030 while (head != ((last + 1) % vq->size)) {
1031 inflight_info->desc[old_free_head].num++;
1032 inflight_info->desc[free_head].addr = desc[head].addr;
1033 inflight_info->desc[free_head].len = desc[head].len;
1034 inflight_info->desc[free_head].flags = desc[head].flags;
1035 inflight_info->desc[free_head].id = desc[head].id;
1037 inflight_info->desc[old_free_head].last = free_head;
1038 free_head = inflight_info->desc[free_head].next;
1039 inflight_info->free_head = free_head;
1040 head = (head + 1) % vq->size;
1043 inflight_info->old_free_head = free_head;
1044 *inflight_entry = old_free_head;
1050 rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
1051 uint16_t last_used_idx, uint16_t idx)
1053 struct virtio_net *dev;
1054 struct vhost_virtqueue *vq;
1056 dev = get_device(vid);
1060 if (unlikely(!(dev->protocol_features &
1061 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1064 if (unlikely(vq_is_packed(dev)))
1067 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1070 vq = dev->virtqueue[vring_idx];
1074 if (unlikely(!vq->inflight_split))
1077 if (unlikely(idx >= vq->size))
1082 vq->inflight_split->desc[idx].inflight = 0;
1086 vq->inflight_split->used_idx = last_used_idx;
1091 rte_vhost_clr_inflight_desc_packed(int vid, uint16_t vring_idx,
1094 struct rte_vhost_inflight_info_packed *inflight_info;
1095 struct virtio_net *dev;
1096 struct vhost_virtqueue *vq;
1098 dev = get_device(vid);
1102 if (unlikely(!(dev->protocol_features &
1103 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1106 if (unlikely(!vq_is_packed(dev)))
1109 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1112 vq = dev->virtqueue[vring_idx];
1116 inflight_info = vq->inflight_packed;
1117 if (unlikely(!inflight_info))
1120 if (unlikely(head >= vq->size))
1125 inflight_info->desc[head].inflight = 0;
1129 inflight_info->old_free_head = inflight_info->free_head;
1130 inflight_info->old_used_idx = inflight_info->used_idx;
1131 inflight_info->old_used_wrap_counter = inflight_info->used_wrap_counter;
1137 rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx,
1140 struct virtio_net *dev;
1141 struct vhost_virtqueue *vq;
1143 dev = get_device(vid);
1147 if (unlikely(!(dev->protocol_features &
1148 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1151 if (unlikely(vq_is_packed(dev)))
1154 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1157 vq = dev->virtqueue[vring_idx];
1161 if (unlikely(!vq->inflight_split))
1164 vq->inflight_split->last_inflight_io = idx;
1169 rte_vhost_set_last_inflight_io_packed(int vid, uint16_t vring_idx,
1172 struct rte_vhost_inflight_info_packed *inflight_info;
1173 struct virtio_net *dev;
1174 struct vhost_virtqueue *vq;
1177 dev = get_device(vid);
1181 if (unlikely(!(dev->protocol_features &
1182 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1185 if (unlikely(!vq_is_packed(dev)))
1188 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1191 vq = dev->virtqueue[vring_idx];
1195 inflight_info = vq->inflight_packed;
1196 if (unlikely(!inflight_info))
1199 if (unlikely(head >= vq->size))
1202 last = inflight_info->desc[head].last;
1203 if (unlikely(last >= vq->size))
1206 inflight_info->desc[last].next = inflight_info->free_head;
1207 inflight_info->free_head = head;
1208 inflight_info->used_idx += inflight_info->desc[head].num;
1209 if (inflight_info->used_idx >= inflight_info->desc_num) {
1210 inflight_info->used_idx -= inflight_info->desc_num;
1211 inflight_info->used_wrap_counter =
1212 !inflight_info->used_wrap_counter;
1219 rte_vhost_vring_call(int vid, uint16_t vring_idx)
1221 struct virtio_net *dev;
1222 struct vhost_virtqueue *vq;
1224 dev = get_device(vid);
1228 if (vring_idx >= VHOST_MAX_VRING)
1231 vq = dev->virtqueue[vring_idx];
1235 if (vq_is_packed(dev))
1236 vhost_vring_call_packed(dev, vq);
1238 vhost_vring_call_split(dev, vq);
1244 rte_vhost_avail_entries(int vid, uint16_t queue_id)
1246 struct virtio_net *dev;
1247 struct vhost_virtqueue *vq;
1250 dev = get_device(vid);
1254 vq = dev->virtqueue[queue_id];
1256 rte_spinlock_lock(&vq->access_lock);
1258 if (unlikely(!vq->enabled || vq->avail == NULL))
1261 ret = *(volatile uint16_t *)&vq->avail->idx - vq->last_used_idx;
1264 rte_spinlock_unlock(&vq->access_lock);
1269 vhost_enable_notify_split(struct virtio_net *dev,
1270 struct vhost_virtqueue *vq, int enable)
1272 if (vq->used == NULL)
1275 if (!(dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))) {
1277 vq->used->flags &= ~VRING_USED_F_NO_NOTIFY;
1279 vq->used->flags |= VRING_USED_F_NO_NOTIFY;
1282 vhost_avail_event(vq) = vq->last_avail_idx;
1288 vhost_enable_notify_packed(struct virtio_net *dev,
1289 struct vhost_virtqueue *vq, int enable)
1293 if (vq->device_event == NULL)
1297 vq->device_event->flags = VRING_EVENT_F_DISABLE;
1301 flags = VRING_EVENT_F_ENABLE;
1302 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX)) {
1303 flags = VRING_EVENT_F_DESC;
1304 vq->device_event->off_wrap = vq->last_avail_idx |
1305 vq->avail_wrap_counter << 15;
1310 vq->device_event->flags = flags;
1315 rte_vhost_enable_guest_notification(int vid, uint16_t queue_id, int enable)
1317 struct virtio_net *dev = get_device(vid);
1318 struct vhost_virtqueue *vq;
1324 vq = dev->virtqueue[queue_id];
1326 rte_spinlock_lock(&vq->access_lock);
1328 if (vq_is_packed(dev))
1329 ret = vhost_enable_notify_packed(dev, vq, enable);
1331 ret = vhost_enable_notify_split(dev, vq, enable);
1333 rte_spinlock_unlock(&vq->access_lock);
1339 rte_vhost_log_write(int vid, uint64_t addr, uint64_t len)
1341 struct virtio_net *dev = get_device(vid);
1346 vhost_log_write(dev, addr, len);
1350 rte_vhost_log_used_vring(int vid, uint16_t vring_idx,
1351 uint64_t offset, uint64_t len)
1353 struct virtio_net *dev;
1354 struct vhost_virtqueue *vq;
1356 dev = get_device(vid);
1360 if (vring_idx >= VHOST_MAX_VRING)
1362 vq = dev->virtqueue[vring_idx];
1366 vhost_log_used_vring(dev, vq, offset, len);
1370 rte_vhost_rx_queue_count(int vid, uint16_t qid)
1372 struct virtio_net *dev;
1373 struct vhost_virtqueue *vq;
1376 dev = get_device(vid);
1380 if (unlikely(qid >= dev->nr_vring || (qid & 1) == 0)) {
1381 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1382 dev->vid, __func__, qid);
1386 vq = dev->virtqueue[qid];
1390 rte_spinlock_lock(&vq->access_lock);
1392 if (unlikely(vq->enabled == 0 || vq->avail == NULL))
1395 ret = *((volatile uint16_t *)&vq->avail->idx) - vq->last_avail_idx;
1398 rte_spinlock_unlock(&vq->access_lock);
1402 struct rte_vdpa_device *
1403 rte_vhost_get_vdpa_device(int vid)
1405 struct virtio_net *dev = get_device(vid);
1410 return dev->vdpa_dev;
1413 int rte_vhost_get_log_base(int vid, uint64_t *log_base,
1416 struct virtio_net *dev = get_device(vid);
1418 if (dev == NULL || log_base == NULL || log_size == NULL)
1421 *log_base = dev->log_base;
1422 *log_size = dev->log_size;
1427 int rte_vhost_get_vring_base(int vid, uint16_t queue_id,
1428 uint16_t *last_avail_idx, uint16_t *last_used_idx)
1430 struct vhost_virtqueue *vq;
1431 struct virtio_net *dev = get_device(vid);
1433 if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1436 vq = dev->virtqueue[queue_id];
1440 if (vq_is_packed(dev)) {
1441 *last_avail_idx = (vq->avail_wrap_counter << 15) |
1443 *last_used_idx = (vq->used_wrap_counter << 15) |
1446 *last_avail_idx = vq->last_avail_idx;
1447 *last_used_idx = vq->last_used_idx;
1453 int rte_vhost_set_vring_base(int vid, uint16_t queue_id,
1454 uint16_t last_avail_idx, uint16_t last_used_idx)
1456 struct vhost_virtqueue *vq;
1457 struct virtio_net *dev = get_device(vid);
1462 vq = dev->virtqueue[queue_id];
1466 if (vq_is_packed(dev)) {
1467 vq->last_avail_idx = last_avail_idx & 0x7fff;
1468 vq->avail_wrap_counter = !!(last_avail_idx & (1 << 15));
1469 vq->last_used_idx = last_used_idx & 0x7fff;
1470 vq->used_wrap_counter = !!(last_used_idx & (1 << 15));
1472 vq->last_avail_idx = last_avail_idx;
1473 vq->last_used_idx = last_used_idx;
1480 rte_vhost_get_vring_base_from_inflight(int vid,
1482 uint16_t *last_avail_idx,
1483 uint16_t *last_used_idx)
1485 struct rte_vhost_inflight_info_packed *inflight_info;
1486 struct virtio_net *dev = get_device(vid);
1488 if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1491 if (!vq_is_packed(dev))
1494 inflight_info = dev->virtqueue[queue_id]->inflight_packed;
1498 *last_avail_idx = (inflight_info->old_used_wrap_counter << 15) |
1499 inflight_info->old_used_idx;
1500 *last_used_idx = *last_avail_idx;
1505 int rte_vhost_extern_callback_register(int vid,
1506 struct rte_vhost_user_extern_ops const * const ops, void *ctx)
1508 struct virtio_net *dev = get_device(vid);
1510 if (dev == NULL || ops == NULL)
1513 dev->extern_ops = *ops;
1514 dev->extern_data = ctx;
1518 int rte_vhost_async_channel_register(int vid, uint16_t queue_id,
1520 struct rte_vhost_async_channel_ops *ops)
1522 struct vhost_virtqueue *vq;
1523 struct virtio_net *dev = get_device(vid);
1524 struct rte_vhost_async_features f;
1526 if (dev == NULL || ops == NULL)
1529 f.intval = features;
1531 vq = dev->virtqueue[queue_id];
1533 if (unlikely(vq == NULL || !dev->async_copy))
1536 /* packed queue is not supported */
1537 if (unlikely(vq_is_packed(dev) || !f.async_inorder)) {
1538 VHOST_LOG_CONFIG(ERR,
1539 "async copy is not supported on packed queue or non-inorder mode "
1540 "(vid %d, qid: %d)\n", vid, queue_id);
1544 if (unlikely(ops->check_completed_copies == NULL ||
1545 ops->transfer_data == NULL))
1548 rte_spinlock_lock(&vq->access_lock);
1550 if (unlikely(vq->async_registered)) {
1551 VHOST_LOG_CONFIG(ERR,
1552 "async register failed: channel already registered "
1553 "(vid %d, qid: %d)\n", vid, queue_id);
1557 vq->async_pkts_pending = rte_malloc(NULL,
1558 vq->size * sizeof(uintptr_t),
1559 RTE_CACHE_LINE_SIZE);
1560 vq->async_pending_info = rte_malloc(NULL,
1561 vq->size * sizeof(uint64_t),
1562 RTE_CACHE_LINE_SIZE);
1563 if (!vq->async_pkts_pending || !vq->async_pending_info) {
1564 if (vq->async_pkts_pending)
1565 rte_free(vq->async_pkts_pending);
1567 if (vq->async_pending_info)
1568 rte_free(vq->async_pending_info);
1570 VHOST_LOG_CONFIG(ERR,
1571 "async register failed: cannot allocate memory for vq data "
1572 "(vid %d, qid: %d)\n", vid, queue_id);
1576 vq->async_ops.check_completed_copies = ops->check_completed_copies;
1577 vq->async_ops.transfer_data = ops->transfer_data;
1579 vq->async_inorder = f.async_inorder;
1580 vq->async_threshold = f.async_threshold;
1582 vq->async_registered = true;
1585 rte_spinlock_unlock(&vq->access_lock);
1590 int rte_vhost_async_channel_unregister(int vid, uint16_t queue_id)
1592 struct vhost_virtqueue *vq;
1593 struct virtio_net *dev = get_device(vid);
1599 vq = dev->virtqueue[queue_id];
1605 rte_spinlock_lock(&vq->access_lock);
1607 if (!vq->async_registered)
1610 if (vq->async_pkts_inflight_n) {
1611 VHOST_LOG_CONFIG(ERR, "Failed to unregister async channel. "
1612 "async inflight packets must be completed before unregistration.\n");
1617 if (vq->async_pkts_pending) {
1618 rte_free(vq->async_pkts_pending);
1619 vq->async_pkts_pending = NULL;
1622 if (vq->async_pending_info) {
1623 rte_free(vq->async_pending_info);
1624 vq->async_pending_info = NULL;
1627 vq->async_ops.transfer_data = NULL;
1628 vq->async_ops.check_completed_copies = NULL;
1629 vq->async_registered = false;
1632 rte_spinlock_unlock(&vq->access_lock);
1637 RTE_LOG_REGISTER(vhost_config_log_level, lib.vhost.config, INFO);
1638 RTE_LOG_REGISTER(vhost_data_log_level, lib.vhost.data, WARNING);