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;
584 /* Also allocate holes, if any, up to requested vring index. */
585 for (i = 0; i <= vring_idx; i++) {
586 if (dev->virtqueue[i])
589 vq = rte_malloc(NULL, sizeof(struct vhost_virtqueue), 0);
591 VHOST_LOG_CONFIG(ERR,
592 "Failed to allocate memory for vring:%u.\n", i);
596 dev->virtqueue[i] = vq;
597 init_vring_queue(dev, vring_idx);
598 rte_spinlock_init(&vq->access_lock);
599 vq->avail_wrap_counter = 1;
600 vq->used_wrap_counter = 1;
601 vq->signalled_used_valid = false;
604 dev->nr_vring = RTE_MAX(dev->nr_vring, vring_idx + 1);
610 * Reset some variables in device structure, while keeping few
611 * others untouched, such as vid, ifname, nr_vring: they
612 * should be same unless the device is removed.
615 reset_device(struct virtio_net *dev)
620 dev->protocol_features = 0;
621 dev->flags &= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
623 for (i = 0; i < dev->nr_vring; i++)
624 reset_vring_queue(dev, i);
628 * Invoked when there is a new vhost-user connection established (when
629 * there is a new virtio device being attached).
632 vhost_new_device(void)
634 struct virtio_net *dev;
637 for (i = 0; i < MAX_VHOST_DEVICE; i++) {
638 if (vhost_devices[i] == NULL)
642 if (i == MAX_VHOST_DEVICE) {
643 VHOST_LOG_CONFIG(ERR,
644 "Failed to find a free slot for new device.\n");
648 dev = rte_zmalloc(NULL, sizeof(struct virtio_net), 0);
650 VHOST_LOG_CONFIG(ERR,
651 "Failed to allocate memory for new dev.\n");
655 vhost_devices[i] = dev;
657 dev->flags = VIRTIO_DEV_BUILTIN_VIRTIO_NET;
658 dev->slave_req_fd = -1;
659 dev->postcopy_ufd = -1;
660 rte_spinlock_init(&dev->slave_req_lock);
666 vhost_destroy_device_notify(struct virtio_net *dev)
668 struct rte_vdpa_device *vdpa_dev;
670 if (dev->flags & VIRTIO_DEV_RUNNING) {
671 vdpa_dev = dev->vdpa_dev;
673 vdpa_dev->ops->dev_close(dev->vid);
674 dev->flags &= ~VIRTIO_DEV_RUNNING;
675 dev->notify_ops->destroy_device(dev->vid);
680 * Invoked when there is the vhost-user connection is broken (when
681 * the virtio device is being detached).
684 vhost_destroy_device(int vid)
686 struct virtio_net *dev = get_device(vid);
691 vhost_destroy_device_notify(dev);
693 cleanup_device(dev, 1);
696 vhost_devices[vid] = NULL;
700 vhost_attach_vdpa_device(int vid, struct rte_vdpa_device *vdpa_dev)
702 struct virtio_net *dev = get_device(vid);
707 dev->vdpa_dev = vdpa_dev;
711 vhost_set_ifname(int vid, const char *if_name, unsigned int if_len)
713 struct virtio_net *dev;
716 dev = get_device(vid);
720 len = if_len > sizeof(dev->ifname) ?
721 sizeof(dev->ifname) : if_len;
723 strncpy(dev->ifname, if_name, len);
724 dev->ifname[sizeof(dev->ifname) - 1] = '\0';
728 vhost_set_builtin_virtio_net(int vid, bool enable)
730 struct virtio_net *dev = get_device(vid);
736 dev->flags |= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
738 dev->flags &= ~VIRTIO_DEV_BUILTIN_VIRTIO_NET;
742 vhost_enable_extbuf(int vid)
744 struct virtio_net *dev = get_device(vid);
753 vhost_enable_linearbuf(int vid)
755 struct virtio_net *dev = get_device(vid);
764 rte_vhost_get_mtu(int vid, uint16_t *mtu)
766 struct virtio_net *dev = get_device(vid);
768 if (dev == NULL || mtu == NULL)
771 if (!(dev->flags & VIRTIO_DEV_READY))
774 if (!(dev->features & (1ULL << VIRTIO_NET_F_MTU)))
783 rte_vhost_get_numa_node(int vid)
785 #ifdef RTE_LIBRTE_VHOST_NUMA
786 struct virtio_net *dev = get_device(vid);
790 if (dev == NULL || numa_available() != 0)
793 ret = get_mempolicy(&numa_node, NULL, 0, dev,
794 MPOL_F_NODE | MPOL_F_ADDR);
796 VHOST_LOG_CONFIG(ERR,
797 "(%d) failed to query numa node: %s\n",
798 vid, rte_strerror(errno));
810 rte_vhost_get_queue_num(int vid)
812 struct virtio_net *dev = get_device(vid);
817 return dev->nr_vring / 2;
821 rte_vhost_get_vring_num(int vid)
823 struct virtio_net *dev = get_device(vid);
828 return dev->nr_vring;
832 rte_vhost_get_ifname(int vid, char *buf, size_t len)
834 struct virtio_net *dev = get_device(vid);
836 if (dev == NULL || buf == NULL)
839 len = RTE_MIN(len, sizeof(dev->ifname));
841 strncpy(buf, dev->ifname, len);
848 rte_vhost_get_negotiated_features(int vid, uint64_t *features)
850 struct virtio_net *dev;
852 dev = get_device(vid);
853 if (dev == NULL || features == NULL)
856 *features = dev->features;
861 rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem)
863 struct virtio_net *dev;
864 struct rte_vhost_memory *m;
867 dev = get_device(vid);
868 if (dev == NULL || mem == NULL)
871 size = dev->mem->nregions * sizeof(struct rte_vhost_mem_region);
872 m = malloc(sizeof(struct rte_vhost_memory) + size);
876 m->nregions = dev->mem->nregions;
877 memcpy(m->regions, dev->mem->regions, size);
884 rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
885 struct rte_vhost_vring *vring)
887 struct virtio_net *dev;
888 struct vhost_virtqueue *vq;
890 dev = get_device(vid);
891 if (dev == NULL || vring == NULL)
894 if (vring_idx >= VHOST_MAX_VRING)
897 vq = dev->virtqueue[vring_idx];
901 if (vq_is_packed(dev)) {
902 vring->desc_packed = vq->desc_packed;
903 vring->driver_event = vq->driver_event;
904 vring->device_event = vq->device_event;
906 vring->desc = vq->desc;
907 vring->avail = vq->avail;
908 vring->used = vq->used;
910 vring->log_guest_addr = vq->log_guest_addr;
912 vring->callfd = vq->callfd;
913 vring->kickfd = vq->kickfd;
914 vring->size = vq->size;
920 rte_vhost_get_vhost_ring_inflight(int vid, uint16_t vring_idx,
921 struct rte_vhost_ring_inflight *vring)
923 struct virtio_net *dev;
924 struct vhost_virtqueue *vq;
926 dev = get_device(vid);
930 if (vring_idx >= VHOST_MAX_VRING)
933 vq = dev->virtqueue[vring_idx];
937 if (vq_is_packed(dev)) {
938 if (unlikely(!vq->inflight_packed))
941 vring->inflight_packed = vq->inflight_packed;
943 if (unlikely(!vq->inflight_split))
946 vring->inflight_split = vq->inflight_split;
949 vring->resubmit_inflight = vq->resubmit_inflight;
955 rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx,
958 struct vhost_virtqueue *vq;
959 struct virtio_net *dev;
961 dev = get_device(vid);
965 if (unlikely(!(dev->protocol_features &
966 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
969 if (unlikely(vq_is_packed(dev)))
972 if (unlikely(vring_idx >= VHOST_MAX_VRING))
975 vq = dev->virtqueue[vring_idx];
979 if (unlikely(!vq->inflight_split))
982 if (unlikely(idx >= vq->size))
985 vq->inflight_split->desc[idx].counter = vq->global_counter++;
986 vq->inflight_split->desc[idx].inflight = 1;
991 rte_vhost_set_inflight_desc_packed(int vid, uint16_t vring_idx,
992 uint16_t head, uint16_t last,
993 uint16_t *inflight_entry)
995 struct rte_vhost_inflight_info_packed *inflight_info;
996 struct virtio_net *dev;
997 struct vhost_virtqueue *vq;
998 struct vring_packed_desc *desc;
999 uint16_t old_free_head, free_head;
1001 dev = get_device(vid);
1005 if (unlikely(!(dev->protocol_features &
1006 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1009 if (unlikely(!vq_is_packed(dev)))
1012 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1015 vq = dev->virtqueue[vring_idx];
1019 inflight_info = vq->inflight_packed;
1020 if (unlikely(!inflight_info))
1023 if (unlikely(head >= vq->size))
1026 desc = vq->desc_packed;
1027 old_free_head = inflight_info->old_free_head;
1028 if (unlikely(old_free_head >= vq->size))
1031 free_head = old_free_head;
1033 /* init header descriptor */
1034 inflight_info->desc[old_free_head].num = 0;
1035 inflight_info->desc[old_free_head].counter = vq->global_counter++;
1036 inflight_info->desc[old_free_head].inflight = 1;
1038 /* save desc entry in flight entry */
1039 while (head != ((last + 1) % vq->size)) {
1040 inflight_info->desc[old_free_head].num++;
1041 inflight_info->desc[free_head].addr = desc[head].addr;
1042 inflight_info->desc[free_head].len = desc[head].len;
1043 inflight_info->desc[free_head].flags = desc[head].flags;
1044 inflight_info->desc[free_head].id = desc[head].id;
1046 inflight_info->desc[old_free_head].last = free_head;
1047 free_head = inflight_info->desc[free_head].next;
1048 inflight_info->free_head = free_head;
1049 head = (head + 1) % vq->size;
1052 inflight_info->old_free_head = free_head;
1053 *inflight_entry = old_free_head;
1059 rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
1060 uint16_t last_used_idx, uint16_t idx)
1062 struct virtio_net *dev;
1063 struct vhost_virtqueue *vq;
1065 dev = get_device(vid);
1069 if (unlikely(!(dev->protocol_features &
1070 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1073 if (unlikely(vq_is_packed(dev)))
1076 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1079 vq = dev->virtqueue[vring_idx];
1083 if (unlikely(!vq->inflight_split))
1086 if (unlikely(idx >= vq->size))
1091 vq->inflight_split->desc[idx].inflight = 0;
1095 vq->inflight_split->used_idx = last_used_idx;
1100 rte_vhost_clr_inflight_desc_packed(int vid, uint16_t vring_idx,
1103 struct rte_vhost_inflight_info_packed *inflight_info;
1104 struct virtio_net *dev;
1105 struct vhost_virtqueue *vq;
1107 dev = get_device(vid);
1111 if (unlikely(!(dev->protocol_features &
1112 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1115 if (unlikely(!vq_is_packed(dev)))
1118 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1121 vq = dev->virtqueue[vring_idx];
1125 inflight_info = vq->inflight_packed;
1126 if (unlikely(!inflight_info))
1129 if (unlikely(head >= vq->size))
1134 inflight_info->desc[head].inflight = 0;
1138 inflight_info->old_free_head = inflight_info->free_head;
1139 inflight_info->old_used_idx = inflight_info->used_idx;
1140 inflight_info->old_used_wrap_counter = inflight_info->used_wrap_counter;
1146 rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx,
1149 struct virtio_net *dev;
1150 struct vhost_virtqueue *vq;
1152 dev = get_device(vid);
1156 if (unlikely(!(dev->protocol_features &
1157 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1160 if (unlikely(vq_is_packed(dev)))
1163 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1166 vq = dev->virtqueue[vring_idx];
1170 if (unlikely(!vq->inflight_split))
1173 vq->inflight_split->last_inflight_io = idx;
1178 rte_vhost_set_last_inflight_io_packed(int vid, uint16_t vring_idx,
1181 struct rte_vhost_inflight_info_packed *inflight_info;
1182 struct virtio_net *dev;
1183 struct vhost_virtqueue *vq;
1186 dev = get_device(vid);
1190 if (unlikely(!(dev->protocol_features &
1191 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1194 if (unlikely(!vq_is_packed(dev)))
1197 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1200 vq = dev->virtqueue[vring_idx];
1204 inflight_info = vq->inflight_packed;
1205 if (unlikely(!inflight_info))
1208 if (unlikely(head >= vq->size))
1211 last = inflight_info->desc[head].last;
1212 if (unlikely(last >= vq->size))
1215 inflight_info->desc[last].next = inflight_info->free_head;
1216 inflight_info->free_head = head;
1217 inflight_info->used_idx += inflight_info->desc[head].num;
1218 if (inflight_info->used_idx >= inflight_info->desc_num) {
1219 inflight_info->used_idx -= inflight_info->desc_num;
1220 inflight_info->used_wrap_counter =
1221 !inflight_info->used_wrap_counter;
1228 rte_vhost_vring_call(int vid, uint16_t vring_idx)
1230 struct virtio_net *dev;
1231 struct vhost_virtqueue *vq;
1233 dev = get_device(vid);
1237 if (vring_idx >= VHOST_MAX_VRING)
1240 vq = dev->virtqueue[vring_idx];
1244 if (vq_is_packed(dev))
1245 vhost_vring_call_packed(dev, vq);
1247 vhost_vring_call_split(dev, vq);
1253 rte_vhost_avail_entries(int vid, uint16_t queue_id)
1255 struct virtio_net *dev;
1256 struct vhost_virtqueue *vq;
1259 dev = get_device(vid);
1263 if (queue_id >= VHOST_MAX_VRING)
1266 vq = dev->virtqueue[queue_id];
1270 rte_spinlock_lock(&vq->access_lock);
1272 if (unlikely(!vq->enabled || vq->avail == NULL))
1275 ret = *(volatile uint16_t *)&vq->avail->idx - vq->last_used_idx;
1278 rte_spinlock_unlock(&vq->access_lock);
1283 vhost_enable_notify_split(struct virtio_net *dev,
1284 struct vhost_virtqueue *vq, int enable)
1286 if (vq->used == NULL)
1289 if (!(dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))) {
1291 vq->used->flags &= ~VRING_USED_F_NO_NOTIFY;
1293 vq->used->flags |= VRING_USED_F_NO_NOTIFY;
1296 vhost_avail_event(vq) = vq->last_avail_idx;
1302 vhost_enable_notify_packed(struct virtio_net *dev,
1303 struct vhost_virtqueue *vq, int enable)
1307 if (vq->device_event == NULL)
1311 vq->device_event->flags = VRING_EVENT_F_DISABLE;
1315 flags = VRING_EVENT_F_ENABLE;
1316 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX)) {
1317 flags = VRING_EVENT_F_DESC;
1318 vq->device_event->off_wrap = vq->last_avail_idx |
1319 vq->avail_wrap_counter << 15;
1324 vq->device_event->flags = flags;
1329 vhost_enable_guest_notification(struct virtio_net *dev,
1330 struct vhost_virtqueue *vq, int enable)
1333 * If the virtqueue is not ready yet, it will be applied
1334 * when it will become ready.
1339 if (vq_is_packed(dev))
1340 return vhost_enable_notify_packed(dev, vq, enable);
1342 return vhost_enable_notify_split(dev, vq, enable);
1346 rte_vhost_enable_guest_notification(int vid, uint16_t queue_id, int enable)
1348 struct virtio_net *dev = get_device(vid);
1349 struct vhost_virtqueue *vq;
1355 if (queue_id >= VHOST_MAX_VRING)
1358 vq = dev->virtqueue[queue_id];
1362 rte_spinlock_lock(&vq->access_lock);
1364 vq->notif_enable = enable;
1365 ret = vhost_enable_guest_notification(dev, vq, enable);
1367 rte_spinlock_unlock(&vq->access_lock);
1373 rte_vhost_log_write(int vid, uint64_t addr, uint64_t len)
1375 struct virtio_net *dev = get_device(vid);
1380 vhost_log_write(dev, addr, len);
1384 rte_vhost_log_used_vring(int vid, uint16_t vring_idx,
1385 uint64_t offset, uint64_t len)
1387 struct virtio_net *dev;
1388 struct vhost_virtqueue *vq;
1390 dev = get_device(vid);
1394 if (vring_idx >= VHOST_MAX_VRING)
1396 vq = dev->virtqueue[vring_idx];
1400 vhost_log_used_vring(dev, vq, offset, len);
1404 rte_vhost_rx_queue_count(int vid, uint16_t qid)
1406 struct virtio_net *dev;
1407 struct vhost_virtqueue *vq;
1410 dev = get_device(vid);
1414 if (unlikely(qid >= dev->nr_vring || (qid & 1) == 0)) {
1415 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1416 dev->vid, __func__, qid);
1420 vq = dev->virtqueue[qid];
1424 rte_spinlock_lock(&vq->access_lock);
1426 if (unlikely(vq->enabled == 0 || vq->avail == NULL))
1429 ret = *((volatile uint16_t *)&vq->avail->idx) - vq->last_avail_idx;
1432 rte_spinlock_unlock(&vq->access_lock);
1436 struct rte_vdpa_device *
1437 rte_vhost_get_vdpa_device(int vid)
1439 struct virtio_net *dev = get_device(vid);
1444 return dev->vdpa_dev;
1447 int rte_vhost_get_log_base(int vid, uint64_t *log_base,
1450 struct virtio_net *dev = get_device(vid);
1452 if (dev == NULL || log_base == NULL || log_size == NULL)
1455 *log_base = dev->log_base;
1456 *log_size = dev->log_size;
1461 int rte_vhost_get_vring_base(int vid, uint16_t queue_id,
1462 uint16_t *last_avail_idx, uint16_t *last_used_idx)
1464 struct vhost_virtqueue *vq;
1465 struct virtio_net *dev = get_device(vid);
1467 if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1470 vq = dev->virtqueue[queue_id];
1474 if (vq_is_packed(dev)) {
1475 *last_avail_idx = (vq->avail_wrap_counter << 15) |
1477 *last_used_idx = (vq->used_wrap_counter << 15) |
1480 *last_avail_idx = vq->last_avail_idx;
1481 *last_used_idx = vq->last_used_idx;
1487 int rte_vhost_set_vring_base(int vid, uint16_t queue_id,
1488 uint16_t last_avail_idx, uint16_t last_used_idx)
1490 struct vhost_virtqueue *vq;
1491 struct virtio_net *dev = get_device(vid);
1496 vq = dev->virtqueue[queue_id];
1500 if (vq_is_packed(dev)) {
1501 vq->last_avail_idx = last_avail_idx & 0x7fff;
1502 vq->avail_wrap_counter = !!(last_avail_idx & (1 << 15));
1503 vq->last_used_idx = last_used_idx & 0x7fff;
1504 vq->used_wrap_counter = !!(last_used_idx & (1 << 15));
1506 vq->last_avail_idx = last_avail_idx;
1507 vq->last_used_idx = last_used_idx;
1514 rte_vhost_get_vring_base_from_inflight(int vid,
1516 uint16_t *last_avail_idx,
1517 uint16_t *last_used_idx)
1519 struct rte_vhost_inflight_info_packed *inflight_info;
1520 struct virtio_net *dev = get_device(vid);
1522 if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1525 if (!vq_is_packed(dev))
1528 inflight_info = dev->virtqueue[queue_id]->inflight_packed;
1532 *last_avail_idx = (inflight_info->old_used_wrap_counter << 15) |
1533 inflight_info->old_used_idx;
1534 *last_used_idx = *last_avail_idx;
1539 int rte_vhost_extern_callback_register(int vid,
1540 struct rte_vhost_user_extern_ops const * const ops, void *ctx)
1542 struct virtio_net *dev = get_device(vid);
1544 if (dev == NULL || ops == NULL)
1547 dev->extern_ops = *ops;
1548 dev->extern_data = ctx;
1552 int rte_vhost_async_channel_register(int vid, uint16_t queue_id,
1554 struct rte_vhost_async_channel_ops *ops)
1556 struct vhost_virtqueue *vq;
1557 struct virtio_net *dev = get_device(vid);
1558 struct rte_vhost_async_features f;
1561 if (dev == NULL || ops == NULL)
1564 f.intval = features;
1566 vq = dev->virtqueue[queue_id];
1568 if (unlikely(vq == NULL || !dev->async_copy))
1571 /* packed queue is not supported */
1572 if (unlikely(vq_is_packed(dev) || !f.async_inorder)) {
1573 VHOST_LOG_CONFIG(ERR,
1574 "async copy is not supported on packed queue or non-inorder mode "
1575 "(vid %d, qid: %d)\n", vid, queue_id);
1579 if (unlikely(ops->check_completed_copies == NULL ||
1580 ops->transfer_data == NULL))
1583 rte_spinlock_lock(&vq->access_lock);
1585 if (unlikely(vq->async_registered)) {
1586 VHOST_LOG_CONFIG(ERR,
1587 "async register failed: channel already registered "
1588 "(vid %d, qid: %d)\n", vid, queue_id);
1592 #ifdef RTE_LIBRTE_VHOST_NUMA
1593 if (get_mempolicy(&node, NULL, 0, vq, MPOL_F_NODE | MPOL_F_ADDR)) {
1594 VHOST_LOG_CONFIG(ERR,
1595 "unable to get numa information in async register. "
1596 "allocating async buffer memory on the caller thread node\n");
1597 node = SOCKET_ID_ANY;
1600 node = SOCKET_ID_ANY;
1603 vq->async_pkts_pending = rte_malloc_socket(NULL,
1604 vq->size * sizeof(uintptr_t),
1605 RTE_CACHE_LINE_SIZE, node);
1606 vq->async_pkts_info = rte_malloc_socket(NULL,
1607 vq->size * sizeof(struct async_inflight_info),
1608 RTE_CACHE_LINE_SIZE, node);
1609 vq->it_pool = rte_malloc_socket(NULL,
1610 VHOST_MAX_ASYNC_IT * sizeof(struct rte_vhost_iov_iter),
1611 RTE_CACHE_LINE_SIZE, node);
1612 vq->vec_pool = rte_malloc_socket(NULL,
1613 VHOST_MAX_ASYNC_VEC * sizeof(struct iovec),
1614 RTE_CACHE_LINE_SIZE, node);
1615 if (!vq->async_pkts_pending || !vq->async_pkts_info ||
1616 !vq->it_pool || !vq->vec_pool) {
1617 vhost_free_async_mem(vq);
1618 VHOST_LOG_CONFIG(ERR,
1619 "async register failed: cannot allocate memory for vq data "
1620 "(vid %d, qid: %d)\n", vid, queue_id);
1624 vq->async_ops.check_completed_copies = ops->check_completed_copies;
1625 vq->async_ops.transfer_data = ops->transfer_data;
1627 vq->async_inorder = f.async_inorder;
1628 vq->async_threshold = f.async_threshold;
1630 vq->async_registered = true;
1633 rte_spinlock_unlock(&vq->access_lock);
1638 int rte_vhost_async_channel_unregister(int vid, uint16_t queue_id)
1640 struct vhost_virtqueue *vq;
1641 struct virtio_net *dev = get_device(vid);
1647 vq = dev->virtqueue[queue_id];
1654 if (!vq->async_registered)
1657 if (!rte_spinlock_trylock(&vq->access_lock)) {
1658 VHOST_LOG_CONFIG(ERR, "Failed to unregister async channel. "
1659 "virt queue busy.\n");
1663 if (vq->async_pkts_inflight_n) {
1664 VHOST_LOG_CONFIG(ERR, "Failed to unregister async channel. "
1665 "async inflight packets must be completed before unregistration.\n");
1670 vhost_free_async_mem(vq);
1672 vq->async_ops.transfer_data = NULL;
1673 vq->async_ops.check_completed_copies = NULL;
1674 vq->async_registered = false;
1677 rte_spinlock_unlock(&vq->access_lock);
1682 RTE_LOG_REGISTER(vhost_config_log_level, lib.vhost.config, INFO);
1683 RTE_LOG_REGISTER(vhost_data_log_level, lib.vhost.data, WARNING);