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];
29 pthread_mutex_t vhost_dev_lock = PTHREAD_MUTEX_INITIALIZER;
31 /* Called with iotlb_lock read-locked */
33 __vhost_iova_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
34 uint64_t iova, uint64_t *size, uint8_t perm)
36 uint64_t vva, tmp_size;
43 vva = vhost_user_iotlb_cache_find(vq, iova, &tmp_size, perm);
44 if (tmp_size == *size)
49 if (!vhost_user_iotlb_pending_miss(vq, iova, perm)) {
51 * iotlb_lock is read-locked for a full burst,
52 * but it only protects the iotlb cache.
53 * In case of IOTLB miss, we might block on the socket,
54 * which could cause a deadlock with QEMU if an IOTLB update
55 * is being handled. We can safely unlock here to avoid it.
57 vhost_user_iotlb_rd_unlock(vq);
59 vhost_user_iotlb_pending_insert(dev, vq, iova, perm);
60 if (vhost_user_iotlb_miss(dev, iova, perm)) {
61 VHOST_LOG_CONFIG(ERR, "(%s) 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 */
109 rte_atomic_thread_fence(__ATOMIC_RELEASE);
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 "(%s) 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))
147 /* No cache, nothing to sync */
148 if (unlikely(!vq->log_cache))
151 rte_atomic_thread_fence(__ATOMIC_RELEASE);
153 log_base = (unsigned long *)(uintptr_t)dev->log_base;
155 for (i = 0; i < vq->log_cache_nb_elem; i++) {
156 struct log_cache_entry *elem = vq->log_cache + i;
158 #if defined(RTE_TOOLCHAIN_GCC) && (GCC_VERSION < 70100)
160 * '__sync' builtins are deprecated, but '__atomic' ones
161 * are sub-optimized in older GCC versions.
163 __sync_fetch_and_or(log_base + elem->offset, elem->val);
165 __atomic_fetch_or(log_base + elem->offset, elem->val,
170 rte_atomic_thread_fence(__ATOMIC_RELEASE);
172 vq->log_cache_nb_elem = 0;
175 static __rte_always_inline void
176 vhost_log_cache_page(struct virtio_net *dev, struct vhost_virtqueue *vq,
179 uint32_t bit_nr = page % (sizeof(unsigned long) << 3);
180 uint32_t offset = page / (sizeof(unsigned long) << 3);
183 if (unlikely(!vq->log_cache)) {
184 /* No logging cache allocated, write dirty log map directly */
185 rte_atomic_thread_fence(__ATOMIC_RELEASE);
186 vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
191 for (i = 0; i < vq->log_cache_nb_elem; i++) {
192 struct log_cache_entry *elem = vq->log_cache + i;
194 if (elem->offset == offset) {
195 elem->val |= (1UL << bit_nr);
200 if (unlikely(i >= VHOST_LOG_CACHE_NR)) {
202 * No more room for a new log cache entry,
203 * so write the dirty log map directly.
205 rte_atomic_thread_fence(__ATOMIC_RELEASE);
206 vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
211 vq->log_cache[i].offset = offset;
212 vq->log_cache[i].val = (1UL << bit_nr);
213 vq->log_cache_nb_elem++;
217 __vhost_log_cache_write(struct virtio_net *dev, struct vhost_virtqueue *vq,
218 uint64_t addr, uint64_t len)
222 if (unlikely(!dev->log_base || !len))
225 if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
228 page = addr / VHOST_LOG_PAGE;
229 while (page * VHOST_LOG_PAGE < addr + len) {
230 vhost_log_cache_page(dev, vq, page);
236 __vhost_log_cache_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
237 uint64_t iova, uint64_t len)
239 uint64_t hva, gpa, map_len;
242 hva = __vhost_iova_to_vva(dev, vq, iova, &map_len, VHOST_ACCESS_RW);
243 if (map_len != len) {
245 "(%s) failed to write log for IOVA 0x%" PRIx64 ". No IOTLB entry found\n",
250 gpa = hva_to_gpa(dev, hva, len);
252 __vhost_log_cache_write(dev, vq, gpa, len);
256 vhost_alloc_copy_ind_table(struct virtio_net *dev, struct vhost_virtqueue *vq,
257 uint64_t desc_addr, uint64_t desc_len)
261 uint64_t len, remain = desc_len;
263 idesc = rte_malloc_socket(__func__, desc_len, 0, vq->numa_node);
264 if (unlikely(!idesc))
267 dst = (uint64_t)(uintptr_t)idesc;
271 src = vhost_iova_to_vva(dev, vq, desc_addr, &len,
273 if (unlikely(!src || !len)) {
278 rte_memcpy((void *)(uintptr_t)dst, (void *)(uintptr_t)src, len);
289 cleanup_vq(struct vhost_virtqueue *vq, int destroy)
291 if ((vq->callfd >= 0) && (destroy != 0))
298 cleanup_vq_inflight(struct virtio_net *dev, struct vhost_virtqueue *vq)
300 if (!(dev->protocol_features &
301 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
304 if (vq_is_packed(dev)) {
305 if (vq->inflight_packed)
306 vq->inflight_packed = NULL;
308 if (vq->inflight_split)
309 vq->inflight_split = NULL;
312 if (vq->resubmit_inflight) {
313 if (vq->resubmit_inflight->resubmit_list) {
314 rte_free(vq->resubmit_inflight->resubmit_list);
315 vq->resubmit_inflight->resubmit_list = NULL;
317 rte_free(vq->resubmit_inflight);
318 vq->resubmit_inflight = NULL;
323 * Unmap any memory, close any file descriptors and
324 * free any memory owned by a device.
327 cleanup_device(struct virtio_net *dev, int destroy)
331 vhost_backend_cleanup(dev);
333 for (i = 0; i < dev->nr_vring; i++) {
334 cleanup_vq(dev->virtqueue[i], destroy);
335 cleanup_vq_inflight(dev, dev->virtqueue[i]);
340 vhost_free_async_mem(struct vhost_virtqueue *vq)
345 rte_free(vq->async->pkts_info);
347 rte_free(vq->async->buffers_packed);
348 vq->async->buffers_packed = NULL;
349 rte_free(vq->async->descs_split);
350 vq->async->descs_split = NULL;
357 free_vq(struct virtio_net *dev, struct vhost_virtqueue *vq)
359 if (vq_is_packed(dev))
360 rte_free(vq->shadow_used_packed);
362 rte_free(vq->shadow_used_split);
364 vhost_free_async_mem(vq);
365 rte_free(vq->batch_copy_elems);
366 rte_mempool_free(vq->iotlb_pool);
367 rte_free(vq->log_cache);
372 * Release virtqueues and device memory.
375 free_device(struct virtio_net *dev)
379 for (i = 0; i < dev->nr_vring; i++)
380 free_vq(dev, dev->virtqueue[i]);
385 static __rte_always_inline int
386 log_translate(struct virtio_net *dev, struct vhost_virtqueue *vq)
388 if (likely(!(vq->ring_addrs.flags & (1 << VHOST_VRING_F_LOG))))
391 vq->log_guest_addr = translate_log_addr(dev, vq,
392 vq->ring_addrs.log_guest_addr);
393 if (vq->log_guest_addr == 0)
400 * Converts vring log address to GPA
401 * If IOMMU is enabled, the log address is IOVA
402 * If IOMMU not enabled, the log address is already GPA
404 * Caller should have iotlb_lock read-locked
407 translate_log_addr(struct virtio_net *dev, struct vhost_virtqueue *vq,
410 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
411 const uint64_t exp_size = sizeof(uint64_t);
413 uint64_t size = exp_size;
415 hva = vhost_iova_to_vva(dev, vq, log_addr,
416 &size, VHOST_ACCESS_RW);
418 if (size != exp_size)
421 gpa = hva_to_gpa(dev, hva, exp_size);
423 VHOST_LOG_CONFIG(ERR,
424 "(%s) failed to find GPA for log_addr: 0x%"
425 PRIx64 " hva: 0x%" PRIx64 "\n",
426 dev->ifname, log_addr, hva);
435 /* Caller should have iotlb_lock read-locked */
437 vring_translate_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
439 uint64_t req_size, size;
441 req_size = sizeof(struct vring_desc) * vq->size;
443 vq->desc = (struct vring_desc *)(uintptr_t)vhost_iova_to_vva(dev, vq,
444 vq->ring_addrs.desc_user_addr,
445 &size, VHOST_ACCESS_RW);
446 if (!vq->desc || size != req_size)
449 req_size = sizeof(struct vring_avail);
450 req_size += sizeof(uint16_t) * vq->size;
451 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
452 req_size += sizeof(uint16_t);
454 vq->avail = (struct vring_avail *)(uintptr_t)vhost_iova_to_vva(dev, vq,
455 vq->ring_addrs.avail_user_addr,
456 &size, VHOST_ACCESS_RW);
457 if (!vq->avail || size != req_size)
460 req_size = sizeof(struct vring_used);
461 req_size += sizeof(struct vring_used_elem) * vq->size;
462 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
463 req_size += sizeof(uint16_t);
465 vq->used = (struct vring_used *)(uintptr_t)vhost_iova_to_vva(dev, vq,
466 vq->ring_addrs.used_user_addr,
467 &size, VHOST_ACCESS_RW);
468 if (!vq->used || size != req_size)
474 /* Caller should have iotlb_lock read-locked */
476 vring_translate_packed(struct virtio_net *dev, struct vhost_virtqueue *vq)
478 uint64_t req_size, size;
480 req_size = sizeof(struct vring_packed_desc) * vq->size;
482 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
483 vhost_iova_to_vva(dev, vq, vq->ring_addrs.desc_user_addr,
484 &size, VHOST_ACCESS_RW);
485 if (!vq->desc_packed || size != req_size)
488 req_size = sizeof(struct vring_packed_desc_event);
490 vq->driver_event = (struct vring_packed_desc_event *)(uintptr_t)
491 vhost_iova_to_vva(dev, vq, vq->ring_addrs.avail_user_addr,
492 &size, VHOST_ACCESS_RW);
493 if (!vq->driver_event || size != req_size)
496 req_size = sizeof(struct vring_packed_desc_event);
498 vq->device_event = (struct vring_packed_desc_event *)(uintptr_t)
499 vhost_iova_to_vva(dev, vq, vq->ring_addrs.used_user_addr,
500 &size, VHOST_ACCESS_RW);
501 if (!vq->device_event || size != req_size)
508 vring_translate(struct virtio_net *dev, struct vhost_virtqueue *vq)
511 if (!(dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
514 if (vq_is_packed(dev)) {
515 if (vring_translate_packed(dev, vq) < 0)
518 if (vring_translate_split(dev, vq) < 0)
522 if (log_translate(dev, vq) < 0)
525 vq->access_ok = true;
531 vring_invalidate(struct virtio_net *dev, struct vhost_virtqueue *vq)
533 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
534 vhost_user_iotlb_wr_lock(vq);
536 vq->access_ok = false;
540 vq->log_guest_addr = 0;
542 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
543 vhost_user_iotlb_wr_unlock(vq);
547 init_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
549 struct vhost_virtqueue *vq;
550 int numa_node = SOCKET_ID_ANY;
552 if (vring_idx >= VHOST_MAX_VRING) {
553 VHOST_LOG_CONFIG(ERR, "(%s) failed to init vring, out of bound (%d)\n",
554 dev->ifname, vring_idx);
558 vq = dev->virtqueue[vring_idx];
560 VHOST_LOG_CONFIG(ERR, "(%s) virtqueue not allocated (%d)\n",
561 dev->ifname, vring_idx);
565 memset(vq, 0, sizeof(struct vhost_virtqueue));
567 vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
568 vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
569 vq->notif_enable = VIRTIO_UNINITIALIZED_NOTIF;
571 #ifdef RTE_LIBRTE_VHOST_NUMA
572 if (get_mempolicy(&numa_node, NULL, 0, vq, MPOL_F_NODE | MPOL_F_ADDR)) {
573 VHOST_LOG_CONFIG(ERR, "(%s) failed to query numa node: %s\n",
574 dev->ifname, rte_strerror(errno));
575 numa_node = SOCKET_ID_ANY;
578 vq->numa_node = numa_node;
580 vhost_user_iotlb_init(dev, vring_idx);
584 reset_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
586 struct vhost_virtqueue *vq;
589 if (vring_idx >= VHOST_MAX_VRING) {
590 VHOST_LOG_CONFIG(ERR,
591 "(%s) failed to reset vring, out of bound (%d)\n",
592 dev->ifname, vring_idx);
596 vq = dev->virtqueue[vring_idx];
598 VHOST_LOG_CONFIG(ERR, "(%s) failed to reset vring, virtqueue not allocated (%d)\n",
599 dev->ifname, vring_idx);
604 init_vring_queue(dev, vring_idx);
609 alloc_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
611 struct vhost_virtqueue *vq;
614 /* Also allocate holes, if any, up to requested vring index. */
615 for (i = 0; i <= vring_idx; i++) {
616 if (dev->virtqueue[i])
619 vq = rte_zmalloc(NULL, sizeof(struct vhost_virtqueue), 0);
621 VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate memory for vring %u.\n",
626 dev->virtqueue[i] = vq;
627 init_vring_queue(dev, i);
628 rte_spinlock_init(&vq->access_lock);
629 vq->avail_wrap_counter = 1;
630 vq->used_wrap_counter = 1;
631 vq->signalled_used_valid = false;
634 dev->nr_vring = RTE_MAX(dev->nr_vring, vring_idx + 1);
640 * Reset some variables in device structure, while keeping few
641 * others untouched, such as vid, ifname, nr_vring: they
642 * should be same unless the device is removed.
645 reset_device(struct virtio_net *dev)
650 dev->protocol_features = 0;
651 dev->flags &= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
653 for (i = 0; i < dev->nr_vring; i++)
654 reset_vring_queue(dev, i);
658 * Invoked when there is a new vhost-user connection established (when
659 * there is a new virtio device being attached).
662 vhost_new_device(void)
664 struct virtio_net *dev;
667 pthread_mutex_lock(&vhost_dev_lock);
668 for (i = 0; i < MAX_VHOST_DEVICE; i++) {
669 if (vhost_devices[i] == NULL)
673 if (i == MAX_VHOST_DEVICE) {
674 VHOST_LOG_CONFIG(ERR, "failed to find a free slot for new device.\n");
675 pthread_mutex_unlock(&vhost_dev_lock);
679 dev = rte_zmalloc(NULL, sizeof(struct virtio_net), 0);
681 VHOST_LOG_CONFIG(ERR, "failed to allocate memory for new device.\n");
682 pthread_mutex_unlock(&vhost_dev_lock);
686 vhost_devices[i] = dev;
687 pthread_mutex_unlock(&vhost_dev_lock);
690 dev->flags = VIRTIO_DEV_BUILTIN_VIRTIO_NET;
691 dev->slave_req_fd = -1;
692 dev->postcopy_ufd = -1;
693 rte_spinlock_init(&dev->slave_req_lock);
699 vhost_destroy_device_notify(struct virtio_net *dev)
701 struct rte_vdpa_device *vdpa_dev;
703 if (dev->flags & VIRTIO_DEV_RUNNING) {
704 vdpa_dev = dev->vdpa_dev;
706 vdpa_dev->ops->dev_close(dev->vid);
707 dev->flags &= ~VIRTIO_DEV_RUNNING;
708 dev->notify_ops->destroy_device(dev->vid);
713 * Invoked when there is the vhost-user connection is broken (when
714 * the virtio device is being detached).
717 vhost_destroy_device(int vid)
719 struct virtio_net *dev = get_device(vid);
724 vhost_destroy_device_notify(dev);
726 cleanup_device(dev, 1);
729 vhost_devices[vid] = NULL;
733 vhost_attach_vdpa_device(int vid, struct rte_vdpa_device *vdpa_dev)
735 struct virtio_net *dev = get_device(vid);
740 dev->vdpa_dev = vdpa_dev;
744 vhost_set_ifname(int vid, const char *if_name, unsigned int if_len)
746 struct virtio_net *dev;
749 dev = get_device(vid);
753 len = if_len > sizeof(dev->ifname) ?
754 sizeof(dev->ifname) : if_len;
756 strncpy(dev->ifname, if_name, len);
757 dev->ifname[sizeof(dev->ifname) - 1] = '\0';
761 vhost_setup_virtio_net(int vid, bool enable, bool compliant_ol_flags)
763 struct virtio_net *dev = get_device(vid);
769 dev->flags |= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
771 dev->flags &= ~VIRTIO_DEV_BUILTIN_VIRTIO_NET;
772 if (!compliant_ol_flags)
773 dev->flags |= VIRTIO_DEV_LEGACY_OL_FLAGS;
775 dev->flags &= ~VIRTIO_DEV_LEGACY_OL_FLAGS;
779 vhost_enable_extbuf(int vid)
781 struct virtio_net *dev = get_device(vid);
790 vhost_enable_linearbuf(int vid)
792 struct virtio_net *dev = get_device(vid);
801 rte_vhost_get_mtu(int vid, uint16_t *mtu)
803 struct virtio_net *dev = get_device(vid);
805 if (dev == NULL || mtu == NULL)
808 if (!(dev->flags & VIRTIO_DEV_READY))
811 if (!(dev->features & (1ULL << VIRTIO_NET_F_MTU)))
820 rte_vhost_get_numa_node(int vid)
822 #ifdef RTE_LIBRTE_VHOST_NUMA
823 struct virtio_net *dev = get_device(vid);
827 if (dev == NULL || numa_available() != 0)
830 ret = get_mempolicy(&numa_node, NULL, 0, dev,
831 MPOL_F_NODE | MPOL_F_ADDR);
833 VHOST_LOG_CONFIG(ERR, "(%s) failed to query numa node: %s\n",
834 dev->ifname, rte_strerror(errno));
846 rte_vhost_get_queue_num(int vid)
848 struct virtio_net *dev = get_device(vid);
853 return dev->nr_vring / 2;
857 rte_vhost_get_vring_num(int vid)
859 struct virtio_net *dev = get_device(vid);
864 return dev->nr_vring;
868 rte_vhost_get_ifname(int vid, char *buf, size_t len)
870 struct virtio_net *dev = get_device(vid);
872 if (dev == NULL || buf == NULL)
875 len = RTE_MIN(len, sizeof(dev->ifname));
877 strncpy(buf, dev->ifname, len);
884 rte_vhost_get_negotiated_features(int vid, uint64_t *features)
886 struct virtio_net *dev;
888 dev = get_device(vid);
889 if (dev == NULL || features == NULL)
892 *features = dev->features;
897 rte_vhost_get_negotiated_protocol_features(int vid,
898 uint64_t *protocol_features)
900 struct virtio_net *dev;
902 dev = get_device(vid);
903 if (dev == NULL || protocol_features == NULL)
906 *protocol_features = dev->protocol_features;
911 rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem)
913 struct virtio_net *dev;
914 struct rte_vhost_memory *m;
917 dev = get_device(vid);
918 if (dev == NULL || mem == NULL)
921 size = dev->mem->nregions * sizeof(struct rte_vhost_mem_region);
922 m = malloc(sizeof(struct rte_vhost_memory) + size);
926 m->nregions = dev->mem->nregions;
927 memcpy(m->regions, dev->mem->regions, size);
934 rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
935 struct rte_vhost_vring *vring)
937 struct virtio_net *dev;
938 struct vhost_virtqueue *vq;
940 dev = get_device(vid);
941 if (dev == NULL || vring == NULL)
944 if (vring_idx >= VHOST_MAX_VRING)
947 vq = dev->virtqueue[vring_idx];
951 if (vq_is_packed(dev)) {
952 vring->desc_packed = vq->desc_packed;
953 vring->driver_event = vq->driver_event;
954 vring->device_event = vq->device_event;
956 vring->desc = vq->desc;
957 vring->avail = vq->avail;
958 vring->used = vq->used;
960 vring->log_guest_addr = vq->log_guest_addr;
962 vring->callfd = vq->callfd;
963 vring->kickfd = vq->kickfd;
964 vring->size = vq->size;
970 rte_vhost_get_vhost_ring_inflight(int vid, uint16_t vring_idx,
971 struct rte_vhost_ring_inflight *vring)
973 struct virtio_net *dev;
974 struct vhost_virtqueue *vq;
976 dev = get_device(vid);
980 if (vring_idx >= VHOST_MAX_VRING)
983 vq = dev->virtqueue[vring_idx];
987 if (vq_is_packed(dev)) {
988 if (unlikely(!vq->inflight_packed))
991 vring->inflight_packed = vq->inflight_packed;
993 if (unlikely(!vq->inflight_split))
996 vring->inflight_split = vq->inflight_split;
999 vring->resubmit_inflight = vq->resubmit_inflight;
1005 rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx,
1008 struct vhost_virtqueue *vq;
1009 struct virtio_net *dev;
1011 dev = get_device(vid);
1015 if (unlikely(!(dev->protocol_features &
1016 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1019 if (unlikely(vq_is_packed(dev)))
1022 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1025 vq = dev->virtqueue[vring_idx];
1029 if (unlikely(!vq->inflight_split))
1032 if (unlikely(idx >= vq->size))
1035 vq->inflight_split->desc[idx].counter = vq->global_counter++;
1036 vq->inflight_split->desc[idx].inflight = 1;
1041 rte_vhost_set_inflight_desc_packed(int vid, uint16_t vring_idx,
1042 uint16_t head, uint16_t last,
1043 uint16_t *inflight_entry)
1045 struct rte_vhost_inflight_info_packed *inflight_info;
1046 struct virtio_net *dev;
1047 struct vhost_virtqueue *vq;
1048 struct vring_packed_desc *desc;
1049 uint16_t old_free_head, free_head;
1051 dev = get_device(vid);
1055 if (unlikely(!(dev->protocol_features &
1056 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1059 if (unlikely(!vq_is_packed(dev)))
1062 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1065 vq = dev->virtqueue[vring_idx];
1069 inflight_info = vq->inflight_packed;
1070 if (unlikely(!inflight_info))
1073 if (unlikely(head >= vq->size))
1076 desc = vq->desc_packed;
1077 old_free_head = inflight_info->old_free_head;
1078 if (unlikely(old_free_head >= vq->size))
1081 free_head = old_free_head;
1083 /* init header descriptor */
1084 inflight_info->desc[old_free_head].num = 0;
1085 inflight_info->desc[old_free_head].counter = vq->global_counter++;
1086 inflight_info->desc[old_free_head].inflight = 1;
1088 /* save desc entry in flight entry */
1089 while (head != ((last + 1) % vq->size)) {
1090 inflight_info->desc[old_free_head].num++;
1091 inflight_info->desc[free_head].addr = desc[head].addr;
1092 inflight_info->desc[free_head].len = desc[head].len;
1093 inflight_info->desc[free_head].flags = desc[head].flags;
1094 inflight_info->desc[free_head].id = desc[head].id;
1096 inflight_info->desc[old_free_head].last = free_head;
1097 free_head = inflight_info->desc[free_head].next;
1098 inflight_info->free_head = free_head;
1099 head = (head + 1) % vq->size;
1102 inflight_info->old_free_head = free_head;
1103 *inflight_entry = old_free_head;
1109 rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
1110 uint16_t last_used_idx, uint16_t idx)
1112 struct virtio_net *dev;
1113 struct vhost_virtqueue *vq;
1115 dev = get_device(vid);
1119 if (unlikely(!(dev->protocol_features &
1120 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1123 if (unlikely(vq_is_packed(dev)))
1126 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1129 vq = dev->virtqueue[vring_idx];
1133 if (unlikely(!vq->inflight_split))
1136 if (unlikely(idx >= vq->size))
1139 rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1141 vq->inflight_split->desc[idx].inflight = 0;
1143 rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1145 vq->inflight_split->used_idx = last_used_idx;
1150 rte_vhost_clr_inflight_desc_packed(int vid, uint16_t vring_idx,
1153 struct rte_vhost_inflight_info_packed *inflight_info;
1154 struct virtio_net *dev;
1155 struct vhost_virtqueue *vq;
1157 dev = get_device(vid);
1161 if (unlikely(!(dev->protocol_features &
1162 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1165 if (unlikely(!vq_is_packed(dev)))
1168 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1171 vq = dev->virtqueue[vring_idx];
1175 inflight_info = vq->inflight_packed;
1176 if (unlikely(!inflight_info))
1179 if (unlikely(head >= vq->size))
1182 rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1184 inflight_info->desc[head].inflight = 0;
1186 rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1188 inflight_info->old_free_head = inflight_info->free_head;
1189 inflight_info->old_used_idx = inflight_info->used_idx;
1190 inflight_info->old_used_wrap_counter = inflight_info->used_wrap_counter;
1196 rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx,
1199 struct virtio_net *dev;
1200 struct vhost_virtqueue *vq;
1202 dev = get_device(vid);
1206 if (unlikely(!(dev->protocol_features &
1207 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1210 if (unlikely(vq_is_packed(dev)))
1213 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1216 vq = dev->virtqueue[vring_idx];
1220 if (unlikely(!vq->inflight_split))
1223 if (unlikely(idx >= vq->size))
1226 vq->inflight_split->last_inflight_io = idx;
1231 rte_vhost_set_last_inflight_io_packed(int vid, uint16_t vring_idx,
1234 struct rte_vhost_inflight_info_packed *inflight_info;
1235 struct virtio_net *dev;
1236 struct vhost_virtqueue *vq;
1239 dev = get_device(vid);
1243 if (unlikely(!(dev->protocol_features &
1244 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1247 if (unlikely(!vq_is_packed(dev)))
1250 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1253 vq = dev->virtqueue[vring_idx];
1257 inflight_info = vq->inflight_packed;
1258 if (unlikely(!inflight_info))
1261 if (unlikely(head >= vq->size))
1264 last = inflight_info->desc[head].last;
1265 if (unlikely(last >= vq->size))
1268 inflight_info->desc[last].next = inflight_info->free_head;
1269 inflight_info->free_head = head;
1270 inflight_info->used_idx += inflight_info->desc[head].num;
1271 if (inflight_info->used_idx >= inflight_info->desc_num) {
1272 inflight_info->used_idx -= inflight_info->desc_num;
1273 inflight_info->used_wrap_counter =
1274 !inflight_info->used_wrap_counter;
1281 rte_vhost_vring_call(int vid, uint16_t vring_idx)
1283 struct virtio_net *dev;
1284 struct vhost_virtqueue *vq;
1286 dev = get_device(vid);
1290 if (vring_idx >= VHOST_MAX_VRING)
1293 vq = dev->virtqueue[vring_idx];
1297 if (vq_is_packed(dev))
1298 vhost_vring_call_packed(dev, vq);
1300 vhost_vring_call_split(dev, vq);
1306 rte_vhost_avail_entries(int vid, uint16_t queue_id)
1308 struct virtio_net *dev;
1309 struct vhost_virtqueue *vq;
1312 dev = get_device(vid);
1316 if (queue_id >= VHOST_MAX_VRING)
1319 vq = dev->virtqueue[queue_id];
1323 rte_spinlock_lock(&vq->access_lock);
1325 if (unlikely(!vq->enabled || vq->avail == NULL))
1328 ret = *(volatile uint16_t *)&vq->avail->idx - vq->last_used_idx;
1331 rte_spinlock_unlock(&vq->access_lock);
1336 vhost_enable_notify_split(struct virtio_net *dev,
1337 struct vhost_virtqueue *vq, int enable)
1339 if (vq->used == NULL)
1342 if (!(dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))) {
1344 vq->used->flags &= ~VRING_USED_F_NO_NOTIFY;
1346 vq->used->flags |= VRING_USED_F_NO_NOTIFY;
1349 vhost_avail_event(vq) = vq->last_avail_idx;
1355 vhost_enable_notify_packed(struct virtio_net *dev,
1356 struct vhost_virtqueue *vq, int enable)
1360 if (vq->device_event == NULL)
1364 vq->device_event->flags = VRING_EVENT_F_DISABLE;
1368 flags = VRING_EVENT_F_ENABLE;
1369 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX)) {
1370 flags = VRING_EVENT_F_DESC;
1371 vq->device_event->off_wrap = vq->last_avail_idx |
1372 vq->avail_wrap_counter << 15;
1375 rte_atomic_thread_fence(__ATOMIC_RELEASE);
1377 vq->device_event->flags = flags;
1382 vhost_enable_guest_notification(struct virtio_net *dev,
1383 struct vhost_virtqueue *vq, int enable)
1386 * If the virtqueue is not ready yet, it will be applied
1387 * when it will become ready.
1392 if (vq_is_packed(dev))
1393 return vhost_enable_notify_packed(dev, vq, enable);
1395 return vhost_enable_notify_split(dev, vq, enable);
1399 rte_vhost_enable_guest_notification(int vid, uint16_t queue_id, int enable)
1401 struct virtio_net *dev = get_device(vid);
1402 struct vhost_virtqueue *vq;
1408 if (queue_id >= VHOST_MAX_VRING)
1411 vq = dev->virtqueue[queue_id];
1415 rte_spinlock_lock(&vq->access_lock);
1417 vq->notif_enable = enable;
1418 ret = vhost_enable_guest_notification(dev, vq, enable);
1420 rte_spinlock_unlock(&vq->access_lock);
1426 rte_vhost_log_write(int vid, uint64_t addr, uint64_t len)
1428 struct virtio_net *dev = get_device(vid);
1433 vhost_log_write(dev, addr, len);
1437 rte_vhost_log_used_vring(int vid, uint16_t vring_idx,
1438 uint64_t offset, uint64_t len)
1440 struct virtio_net *dev;
1441 struct vhost_virtqueue *vq;
1443 dev = get_device(vid);
1447 if (vring_idx >= VHOST_MAX_VRING)
1449 vq = dev->virtqueue[vring_idx];
1453 vhost_log_used_vring(dev, vq, offset, len);
1457 rte_vhost_rx_queue_count(int vid, uint16_t qid)
1459 struct virtio_net *dev;
1460 struct vhost_virtqueue *vq;
1463 dev = get_device(vid);
1467 if (unlikely(qid >= dev->nr_vring || (qid & 1) == 0)) {
1468 VHOST_LOG_DATA(ERR, "(%s) %s: invalid virtqueue idx %d.\n",
1469 dev->ifname, __func__, qid);
1473 vq = dev->virtqueue[qid];
1477 rte_spinlock_lock(&vq->access_lock);
1479 if (unlikely(!vq->enabled || vq->avail == NULL))
1482 ret = *((volatile uint16_t *)&vq->avail->idx) - vq->last_avail_idx;
1485 rte_spinlock_unlock(&vq->access_lock);
1489 struct rte_vdpa_device *
1490 rte_vhost_get_vdpa_device(int vid)
1492 struct virtio_net *dev = get_device(vid);
1497 return dev->vdpa_dev;
1501 rte_vhost_get_log_base(int vid, uint64_t *log_base,
1504 struct virtio_net *dev = get_device(vid);
1506 if (dev == NULL || log_base == NULL || log_size == NULL)
1509 *log_base = dev->log_base;
1510 *log_size = dev->log_size;
1516 rte_vhost_get_vring_base(int vid, uint16_t queue_id,
1517 uint16_t *last_avail_idx, uint16_t *last_used_idx)
1519 struct vhost_virtqueue *vq;
1520 struct virtio_net *dev = get_device(vid);
1522 if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1525 if (queue_id >= VHOST_MAX_VRING)
1528 vq = dev->virtqueue[queue_id];
1532 if (vq_is_packed(dev)) {
1533 *last_avail_idx = (vq->avail_wrap_counter << 15) |
1535 *last_used_idx = (vq->used_wrap_counter << 15) |
1538 *last_avail_idx = vq->last_avail_idx;
1539 *last_used_idx = vq->last_used_idx;
1546 rte_vhost_set_vring_base(int vid, uint16_t queue_id,
1547 uint16_t last_avail_idx, uint16_t last_used_idx)
1549 struct vhost_virtqueue *vq;
1550 struct virtio_net *dev = get_device(vid);
1555 if (queue_id >= VHOST_MAX_VRING)
1558 vq = dev->virtqueue[queue_id];
1562 if (vq_is_packed(dev)) {
1563 vq->last_avail_idx = last_avail_idx & 0x7fff;
1564 vq->avail_wrap_counter = !!(last_avail_idx & (1 << 15));
1565 vq->last_used_idx = last_used_idx & 0x7fff;
1566 vq->used_wrap_counter = !!(last_used_idx & (1 << 15));
1568 vq->last_avail_idx = last_avail_idx;
1569 vq->last_used_idx = last_used_idx;
1576 rte_vhost_get_vring_base_from_inflight(int vid,
1578 uint16_t *last_avail_idx,
1579 uint16_t *last_used_idx)
1581 struct rte_vhost_inflight_info_packed *inflight_info;
1582 struct vhost_virtqueue *vq;
1583 struct virtio_net *dev = get_device(vid);
1585 if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1588 if (queue_id >= VHOST_MAX_VRING)
1591 vq = dev->virtqueue[queue_id];
1595 if (!vq_is_packed(dev))
1598 inflight_info = vq->inflight_packed;
1602 *last_avail_idx = (inflight_info->old_used_wrap_counter << 15) |
1603 inflight_info->old_used_idx;
1604 *last_used_idx = *last_avail_idx;
1610 rte_vhost_extern_callback_register(int vid,
1611 struct rte_vhost_user_extern_ops const * const ops, void *ctx)
1613 struct virtio_net *dev = get_device(vid);
1615 if (dev == NULL || ops == NULL)
1618 dev->extern_ops = *ops;
1619 dev->extern_data = ctx;
1623 static __rte_always_inline int
1624 async_channel_register(int vid, uint16_t queue_id,
1625 struct rte_vhost_async_channel_ops *ops)
1627 struct virtio_net *dev = get_device(vid);
1628 struct vhost_virtqueue *vq = dev->virtqueue[queue_id];
1629 struct vhost_async *async;
1630 int node = vq->numa_node;
1632 if (unlikely(vq->async)) {
1633 VHOST_LOG_CONFIG(ERR,
1634 "(%s) async register failed: already registered (qid: %d)\n",
1635 dev->ifname, queue_id);
1639 async = rte_zmalloc_socket(NULL, sizeof(struct vhost_async), 0, node);
1641 VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate async metadata (qid: %d)\n",
1642 dev->ifname, queue_id);
1646 async->pkts_info = rte_malloc_socket(NULL, vq->size * sizeof(struct async_inflight_info),
1647 RTE_CACHE_LINE_SIZE, node);
1648 if (!async->pkts_info) {
1649 VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate async_pkts_info (qid: %d)\n",
1650 dev->ifname, queue_id);
1651 goto out_free_async;
1654 if (vq_is_packed(dev)) {
1655 async->buffers_packed = rte_malloc_socket(NULL,
1656 vq->size * sizeof(struct vring_used_elem_packed),
1657 RTE_CACHE_LINE_SIZE, node);
1658 if (!async->buffers_packed) {
1659 VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate async buffers (qid: %d)\n",
1660 dev->ifname, queue_id);
1661 goto out_free_inflight;
1664 async->descs_split = rte_malloc_socket(NULL,
1665 vq->size * sizeof(struct vring_used_elem),
1666 RTE_CACHE_LINE_SIZE, node);
1667 if (!async->descs_split) {
1668 VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate async descs (qid: %d)\n",
1669 dev->ifname, queue_id);
1670 goto out_free_inflight;
1674 async->ops.check_completed_copies = ops->check_completed_copies;
1675 async->ops.transfer_data = ops->transfer_data;
1681 rte_free(async->pkts_info);
1689 rte_vhost_async_channel_register(int vid, uint16_t queue_id,
1690 struct rte_vhost_async_config config,
1691 struct rte_vhost_async_channel_ops *ops)
1693 struct vhost_virtqueue *vq;
1694 struct virtio_net *dev = get_device(vid);
1697 if (dev == NULL || ops == NULL)
1700 if (queue_id >= VHOST_MAX_VRING)
1703 vq = dev->virtqueue[queue_id];
1705 if (unlikely(vq == NULL || !dev->async_copy))
1708 if (unlikely(!(config.features & RTE_VHOST_ASYNC_INORDER))) {
1709 VHOST_LOG_CONFIG(ERR,
1710 "(%s) async copy is not supported on non-inorder mode (qid: %d)\n",
1711 dev->ifname, queue_id);
1715 if (unlikely(ops->check_completed_copies == NULL ||
1716 ops->transfer_data == NULL))
1719 rte_spinlock_lock(&vq->access_lock);
1720 ret = async_channel_register(vid, queue_id, ops);
1721 rte_spinlock_unlock(&vq->access_lock);
1727 rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t queue_id,
1728 struct rte_vhost_async_config config,
1729 struct rte_vhost_async_channel_ops *ops)
1731 struct vhost_virtqueue *vq;
1732 struct virtio_net *dev = get_device(vid);
1734 if (dev == NULL || ops == NULL)
1737 if (queue_id >= VHOST_MAX_VRING)
1740 vq = dev->virtqueue[queue_id];
1742 if (unlikely(vq == NULL || !dev->async_copy))
1745 if (unlikely(!(config.features & RTE_VHOST_ASYNC_INORDER))) {
1746 VHOST_LOG_CONFIG(ERR,
1747 "(%s) async copy is not supported on non-inorder mode (qid: %d)\n",
1748 dev->ifname, queue_id);
1752 if (unlikely(ops->check_completed_copies == NULL ||
1753 ops->transfer_data == NULL))
1756 return async_channel_register(vid, queue_id, ops);
1760 rte_vhost_async_channel_unregister(int vid, uint16_t queue_id)
1762 struct vhost_virtqueue *vq;
1763 struct virtio_net *dev = get_device(vid);
1769 if (queue_id >= VHOST_MAX_VRING)
1772 vq = dev->virtqueue[queue_id];
1782 if (!rte_spinlock_trylock(&vq->access_lock)) {
1783 VHOST_LOG_CONFIG(ERR, "(%s) failed to unregister async channel, virtqueue busy.\n",
1788 if (vq->async->pkts_inflight_n) {
1789 VHOST_LOG_CONFIG(ERR, "(%s) failed to unregister async channel.\n", dev->ifname);
1790 VHOST_LOG_CONFIG(ERR, "(%s) inflight packets must be completed before unregistration.\n",
1796 vhost_free_async_mem(vq);
1798 rte_spinlock_unlock(&vq->access_lock);
1804 rte_vhost_async_channel_unregister_thread_unsafe(int vid, uint16_t queue_id)
1806 struct vhost_virtqueue *vq;
1807 struct virtio_net *dev = get_device(vid);
1812 if (queue_id >= VHOST_MAX_VRING)
1815 vq = dev->virtqueue[queue_id];
1823 if (vq->async->pkts_inflight_n) {
1824 VHOST_LOG_CONFIG(ERR, "(%s) failed to unregister async channel.\n", dev->ifname);
1825 VHOST_LOG_CONFIG(ERR, "(%s) inflight packets must be completed before unregistration.\n",
1830 vhost_free_async_mem(vq);
1836 rte_vhost_async_get_inflight(int vid, uint16_t queue_id)
1838 struct vhost_virtqueue *vq;
1839 struct virtio_net *dev = get_device(vid);
1845 if (queue_id >= VHOST_MAX_VRING)
1848 vq = dev->virtqueue[queue_id];
1856 if (!rte_spinlock_trylock(&vq->access_lock)) {
1857 VHOST_LOG_CONFIG(DEBUG,
1858 "(%s) failed to check in-flight packets. virtqueue busy.\n",
1863 ret = vq->async->pkts_inflight_n;
1864 rte_spinlock_unlock(&vq->access_lock);
1870 rte_vhost_get_monitor_addr(int vid, uint16_t queue_id,
1871 struct rte_vhost_power_monitor_cond *pmc)
1873 struct virtio_net *dev = get_device(vid);
1874 struct vhost_virtqueue *vq;
1878 if (queue_id >= VHOST_MAX_VRING)
1881 vq = dev->virtqueue[queue_id];
1885 if (vq_is_packed(dev)) {
1886 struct vring_packed_desc *desc;
1887 desc = vq->desc_packed;
1888 pmc->addr = &desc[vq->last_avail_idx].flags;
1889 if (vq->avail_wrap_counter)
1890 pmc->val = VRING_DESC_F_AVAIL;
1892 pmc->val = VRING_DESC_F_USED;
1893 pmc->mask = VRING_DESC_F_AVAIL | VRING_DESC_F_USED;
1894 pmc->size = sizeof(desc[vq->last_avail_idx].flags);
1897 pmc->addr = &vq->avail->idx;
1898 pmc->val = vq->last_avail_idx & (vq->size - 1);
1899 pmc->mask = vq->size - 1;
1900 pmc->size = sizeof(vq->avail->idx);
1907 RTE_LOG_REGISTER_SUFFIX(vhost_config_log_level, config, INFO);
1908 RTE_LOG_REGISTER_SUFFIX(vhost_data_log_level, data, WARNING);