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 */
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 "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 rte_atomic_thread_fence(__ATOMIC_RELEASE);
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,
166 rte_atomic_thread_fence(__ATOMIC_RELEASE);
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.
193 rte_atomic_thread_fence(__ATOMIC_RELEASE);
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_info)
331 rte_free(vq->async_pkts_info);
332 if (vq->async_descs_split)
333 rte_free(vq->async_descs_split);
335 rte_free(vq->it_pool);
337 rte_free(vq->vec_pool);
339 vq->async_pkts_info = NULL;
340 vq->async_descs_split = 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 VHOST_LOG_CONFIG(ERR, "Virtqueue not allocated (%d)\n",
553 memset(vq, 0, sizeof(struct vhost_virtqueue));
555 vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
556 vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
557 vq->notif_enable = VIRTIO_UNINITIALIZED_NOTIF;
559 vhost_user_iotlb_init(dev, vring_idx);
560 /* Backends are set to -1 indicating an inactive device. */
565 reset_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
567 struct vhost_virtqueue *vq;
570 if (vring_idx >= VHOST_MAX_VRING) {
571 VHOST_LOG_CONFIG(ERR,
572 "Failed not init vring, out of bound (%d)\n",
577 vq = dev->virtqueue[vring_idx];
579 VHOST_LOG_CONFIG(ERR, "Virtqueue not allocated (%d)\n",
585 init_vring_queue(dev, vring_idx);
590 alloc_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
592 struct vhost_virtqueue *vq;
595 /* Also allocate holes, if any, up to requested vring index. */
596 for (i = 0; i <= vring_idx; i++) {
597 if (dev->virtqueue[i])
600 vq = rte_malloc(NULL, sizeof(struct vhost_virtqueue), 0);
602 VHOST_LOG_CONFIG(ERR,
603 "Failed to allocate memory for vring:%u.\n", i);
607 dev->virtqueue[i] = vq;
608 init_vring_queue(dev, i);
609 rte_spinlock_init(&vq->access_lock);
610 vq->avail_wrap_counter = 1;
611 vq->used_wrap_counter = 1;
612 vq->signalled_used_valid = false;
615 dev->nr_vring = RTE_MAX(dev->nr_vring, vring_idx + 1);
621 * Reset some variables in device structure, while keeping few
622 * others untouched, such as vid, ifname, nr_vring: they
623 * should be same unless the device is removed.
626 reset_device(struct virtio_net *dev)
631 dev->protocol_features = 0;
632 dev->flags &= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
634 for (i = 0; i < dev->nr_vring; i++)
635 reset_vring_queue(dev, i);
639 * Invoked when there is a new vhost-user connection established (when
640 * there is a new virtio device being attached).
643 vhost_new_device(void)
645 struct virtio_net *dev;
648 for (i = 0; i < MAX_VHOST_DEVICE; i++) {
649 if (vhost_devices[i] == NULL)
653 if (i == MAX_VHOST_DEVICE) {
654 VHOST_LOG_CONFIG(ERR,
655 "Failed to find a free slot for new device.\n");
659 dev = rte_zmalloc(NULL, sizeof(struct virtio_net), 0);
661 VHOST_LOG_CONFIG(ERR,
662 "Failed to allocate memory for new dev.\n");
666 vhost_devices[i] = dev;
668 dev->flags = VIRTIO_DEV_BUILTIN_VIRTIO_NET;
669 dev->slave_req_fd = -1;
670 dev->postcopy_ufd = -1;
671 rte_spinlock_init(&dev->slave_req_lock);
677 vhost_destroy_device_notify(struct virtio_net *dev)
679 struct rte_vdpa_device *vdpa_dev;
681 if (dev->flags & VIRTIO_DEV_RUNNING) {
682 vdpa_dev = dev->vdpa_dev;
684 vdpa_dev->ops->dev_close(dev->vid);
685 dev->flags &= ~VIRTIO_DEV_RUNNING;
686 dev->notify_ops->destroy_device(dev->vid);
691 * Invoked when there is the vhost-user connection is broken (when
692 * the virtio device is being detached).
695 vhost_destroy_device(int vid)
697 struct virtio_net *dev = get_device(vid);
702 vhost_destroy_device_notify(dev);
704 cleanup_device(dev, 1);
707 vhost_devices[vid] = NULL;
711 vhost_attach_vdpa_device(int vid, struct rte_vdpa_device *vdpa_dev)
713 struct virtio_net *dev = get_device(vid);
718 dev->vdpa_dev = vdpa_dev;
722 vhost_set_ifname(int vid, const char *if_name, unsigned int if_len)
724 struct virtio_net *dev;
727 dev = get_device(vid);
731 len = if_len > sizeof(dev->ifname) ?
732 sizeof(dev->ifname) : if_len;
734 strncpy(dev->ifname, if_name, len);
735 dev->ifname[sizeof(dev->ifname) - 1] = '\0';
739 vhost_set_builtin_virtio_net(int vid, bool enable)
741 struct virtio_net *dev = get_device(vid);
747 dev->flags |= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
749 dev->flags &= ~VIRTIO_DEV_BUILTIN_VIRTIO_NET;
753 vhost_enable_extbuf(int vid)
755 struct virtio_net *dev = get_device(vid);
764 vhost_enable_linearbuf(int vid)
766 struct virtio_net *dev = get_device(vid);
775 rte_vhost_get_mtu(int vid, uint16_t *mtu)
777 struct virtio_net *dev = get_device(vid);
779 if (dev == NULL || mtu == NULL)
782 if (!(dev->flags & VIRTIO_DEV_READY))
785 if (!(dev->features & (1ULL << VIRTIO_NET_F_MTU)))
794 rte_vhost_get_numa_node(int vid)
796 #ifdef RTE_LIBRTE_VHOST_NUMA
797 struct virtio_net *dev = get_device(vid);
801 if (dev == NULL || numa_available() != 0)
804 ret = get_mempolicy(&numa_node, NULL, 0, dev,
805 MPOL_F_NODE | MPOL_F_ADDR);
807 VHOST_LOG_CONFIG(ERR,
808 "(%d) failed to query numa node: %s\n",
809 vid, rte_strerror(errno));
821 rte_vhost_get_queue_num(int vid)
823 struct virtio_net *dev = get_device(vid);
828 return dev->nr_vring / 2;
832 rte_vhost_get_vring_num(int vid)
834 struct virtio_net *dev = get_device(vid);
839 return dev->nr_vring;
843 rte_vhost_get_ifname(int vid, char *buf, size_t len)
845 struct virtio_net *dev = get_device(vid);
847 if (dev == NULL || buf == NULL)
850 len = RTE_MIN(len, sizeof(dev->ifname));
852 strncpy(buf, dev->ifname, len);
859 rte_vhost_get_negotiated_features(int vid, uint64_t *features)
861 struct virtio_net *dev;
863 dev = get_device(vid);
864 if (dev == NULL || features == NULL)
867 *features = dev->features;
872 rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem)
874 struct virtio_net *dev;
875 struct rte_vhost_memory *m;
878 dev = get_device(vid);
879 if (dev == NULL || mem == NULL)
882 size = dev->mem->nregions * sizeof(struct rte_vhost_mem_region);
883 m = malloc(sizeof(struct rte_vhost_memory) + size);
887 m->nregions = dev->mem->nregions;
888 memcpy(m->regions, dev->mem->regions, size);
895 rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
896 struct rte_vhost_vring *vring)
898 struct virtio_net *dev;
899 struct vhost_virtqueue *vq;
901 dev = get_device(vid);
902 if (dev == NULL || vring == NULL)
905 if (vring_idx >= VHOST_MAX_VRING)
908 vq = dev->virtqueue[vring_idx];
912 if (vq_is_packed(dev)) {
913 vring->desc_packed = vq->desc_packed;
914 vring->driver_event = vq->driver_event;
915 vring->device_event = vq->device_event;
917 vring->desc = vq->desc;
918 vring->avail = vq->avail;
919 vring->used = vq->used;
921 vring->log_guest_addr = vq->log_guest_addr;
923 vring->callfd = vq->callfd;
924 vring->kickfd = vq->kickfd;
925 vring->size = vq->size;
931 rte_vhost_get_vhost_ring_inflight(int vid, uint16_t vring_idx,
932 struct rte_vhost_ring_inflight *vring)
934 struct virtio_net *dev;
935 struct vhost_virtqueue *vq;
937 dev = get_device(vid);
941 if (vring_idx >= VHOST_MAX_VRING)
944 vq = dev->virtqueue[vring_idx];
948 if (vq_is_packed(dev)) {
949 if (unlikely(!vq->inflight_packed))
952 vring->inflight_packed = vq->inflight_packed;
954 if (unlikely(!vq->inflight_split))
957 vring->inflight_split = vq->inflight_split;
960 vring->resubmit_inflight = vq->resubmit_inflight;
966 rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx,
969 struct vhost_virtqueue *vq;
970 struct virtio_net *dev;
972 dev = get_device(vid);
976 if (unlikely(!(dev->protocol_features &
977 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
980 if (unlikely(vq_is_packed(dev)))
983 if (unlikely(vring_idx >= VHOST_MAX_VRING))
986 vq = dev->virtqueue[vring_idx];
990 if (unlikely(!vq->inflight_split))
993 if (unlikely(idx >= vq->size))
996 vq->inflight_split->desc[idx].counter = vq->global_counter++;
997 vq->inflight_split->desc[idx].inflight = 1;
1002 rte_vhost_set_inflight_desc_packed(int vid, uint16_t vring_idx,
1003 uint16_t head, uint16_t last,
1004 uint16_t *inflight_entry)
1006 struct rte_vhost_inflight_info_packed *inflight_info;
1007 struct virtio_net *dev;
1008 struct vhost_virtqueue *vq;
1009 struct vring_packed_desc *desc;
1010 uint16_t old_free_head, free_head;
1012 dev = get_device(vid);
1016 if (unlikely(!(dev->protocol_features &
1017 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1020 if (unlikely(!vq_is_packed(dev)))
1023 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1026 vq = dev->virtqueue[vring_idx];
1030 inflight_info = vq->inflight_packed;
1031 if (unlikely(!inflight_info))
1034 if (unlikely(head >= vq->size))
1037 desc = vq->desc_packed;
1038 old_free_head = inflight_info->old_free_head;
1039 if (unlikely(old_free_head >= vq->size))
1042 free_head = old_free_head;
1044 /* init header descriptor */
1045 inflight_info->desc[old_free_head].num = 0;
1046 inflight_info->desc[old_free_head].counter = vq->global_counter++;
1047 inflight_info->desc[old_free_head].inflight = 1;
1049 /* save desc entry in flight entry */
1050 while (head != ((last + 1) % vq->size)) {
1051 inflight_info->desc[old_free_head].num++;
1052 inflight_info->desc[free_head].addr = desc[head].addr;
1053 inflight_info->desc[free_head].len = desc[head].len;
1054 inflight_info->desc[free_head].flags = desc[head].flags;
1055 inflight_info->desc[free_head].id = desc[head].id;
1057 inflight_info->desc[old_free_head].last = free_head;
1058 free_head = inflight_info->desc[free_head].next;
1059 inflight_info->free_head = free_head;
1060 head = (head + 1) % vq->size;
1063 inflight_info->old_free_head = free_head;
1064 *inflight_entry = old_free_head;
1070 rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
1071 uint16_t last_used_idx, uint16_t idx)
1073 struct virtio_net *dev;
1074 struct vhost_virtqueue *vq;
1076 dev = get_device(vid);
1080 if (unlikely(!(dev->protocol_features &
1081 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1084 if (unlikely(vq_is_packed(dev)))
1087 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1090 vq = dev->virtqueue[vring_idx];
1094 if (unlikely(!vq->inflight_split))
1097 if (unlikely(idx >= vq->size))
1100 rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1102 vq->inflight_split->desc[idx].inflight = 0;
1104 rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1106 vq->inflight_split->used_idx = last_used_idx;
1111 rte_vhost_clr_inflight_desc_packed(int vid, uint16_t vring_idx,
1114 struct rte_vhost_inflight_info_packed *inflight_info;
1115 struct virtio_net *dev;
1116 struct vhost_virtqueue *vq;
1118 dev = get_device(vid);
1122 if (unlikely(!(dev->protocol_features &
1123 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1126 if (unlikely(!vq_is_packed(dev)))
1129 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1132 vq = dev->virtqueue[vring_idx];
1136 inflight_info = vq->inflight_packed;
1137 if (unlikely(!inflight_info))
1140 if (unlikely(head >= vq->size))
1143 rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1145 inflight_info->desc[head].inflight = 0;
1147 rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1149 inflight_info->old_free_head = inflight_info->free_head;
1150 inflight_info->old_used_idx = inflight_info->used_idx;
1151 inflight_info->old_used_wrap_counter = inflight_info->used_wrap_counter;
1157 rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx,
1160 struct virtio_net *dev;
1161 struct vhost_virtqueue *vq;
1163 dev = get_device(vid);
1167 if (unlikely(!(dev->protocol_features &
1168 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1171 if (unlikely(vq_is_packed(dev)))
1174 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1177 vq = dev->virtqueue[vring_idx];
1181 if (unlikely(!vq->inflight_split))
1184 vq->inflight_split->last_inflight_io = idx;
1189 rte_vhost_set_last_inflight_io_packed(int vid, uint16_t vring_idx,
1192 struct rte_vhost_inflight_info_packed *inflight_info;
1193 struct virtio_net *dev;
1194 struct vhost_virtqueue *vq;
1197 dev = get_device(vid);
1201 if (unlikely(!(dev->protocol_features &
1202 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1205 if (unlikely(!vq_is_packed(dev)))
1208 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1211 vq = dev->virtqueue[vring_idx];
1215 inflight_info = vq->inflight_packed;
1216 if (unlikely(!inflight_info))
1219 if (unlikely(head >= vq->size))
1222 last = inflight_info->desc[head].last;
1223 if (unlikely(last >= vq->size))
1226 inflight_info->desc[last].next = inflight_info->free_head;
1227 inflight_info->free_head = head;
1228 inflight_info->used_idx += inflight_info->desc[head].num;
1229 if (inflight_info->used_idx >= inflight_info->desc_num) {
1230 inflight_info->used_idx -= inflight_info->desc_num;
1231 inflight_info->used_wrap_counter =
1232 !inflight_info->used_wrap_counter;
1239 rte_vhost_vring_call(int vid, uint16_t vring_idx)
1241 struct virtio_net *dev;
1242 struct vhost_virtqueue *vq;
1244 dev = get_device(vid);
1248 if (vring_idx >= VHOST_MAX_VRING)
1251 vq = dev->virtqueue[vring_idx];
1255 if (vq_is_packed(dev))
1256 vhost_vring_call_packed(dev, vq);
1258 vhost_vring_call_split(dev, vq);
1264 rte_vhost_avail_entries(int vid, uint16_t queue_id)
1266 struct virtio_net *dev;
1267 struct vhost_virtqueue *vq;
1270 dev = get_device(vid);
1274 if (queue_id >= VHOST_MAX_VRING)
1277 vq = dev->virtqueue[queue_id];
1281 rte_spinlock_lock(&vq->access_lock);
1283 if (unlikely(!vq->enabled || vq->avail == NULL))
1286 ret = *(volatile uint16_t *)&vq->avail->idx - vq->last_used_idx;
1289 rte_spinlock_unlock(&vq->access_lock);
1294 vhost_enable_notify_split(struct virtio_net *dev,
1295 struct vhost_virtqueue *vq, int enable)
1297 if (vq->used == NULL)
1300 if (!(dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))) {
1302 vq->used->flags &= ~VRING_USED_F_NO_NOTIFY;
1304 vq->used->flags |= VRING_USED_F_NO_NOTIFY;
1307 vhost_avail_event(vq) = vq->last_avail_idx;
1313 vhost_enable_notify_packed(struct virtio_net *dev,
1314 struct vhost_virtqueue *vq, int enable)
1318 if (vq->device_event == NULL)
1322 vq->device_event->flags = VRING_EVENT_F_DISABLE;
1326 flags = VRING_EVENT_F_ENABLE;
1327 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX)) {
1328 flags = VRING_EVENT_F_DESC;
1329 vq->device_event->off_wrap = vq->last_avail_idx |
1330 vq->avail_wrap_counter << 15;
1333 rte_atomic_thread_fence(__ATOMIC_RELEASE);
1335 vq->device_event->flags = flags;
1340 vhost_enable_guest_notification(struct virtio_net *dev,
1341 struct vhost_virtqueue *vq, int enable)
1344 * If the virtqueue is not ready yet, it will be applied
1345 * when it will become ready.
1350 if (vq_is_packed(dev))
1351 return vhost_enable_notify_packed(dev, vq, enable);
1353 return vhost_enable_notify_split(dev, vq, enable);
1357 rte_vhost_enable_guest_notification(int vid, uint16_t queue_id, int enable)
1359 struct virtio_net *dev = get_device(vid);
1360 struct vhost_virtqueue *vq;
1366 if (queue_id >= VHOST_MAX_VRING)
1369 vq = dev->virtqueue[queue_id];
1373 rte_spinlock_lock(&vq->access_lock);
1375 vq->notif_enable = enable;
1376 ret = vhost_enable_guest_notification(dev, vq, enable);
1378 rte_spinlock_unlock(&vq->access_lock);
1384 rte_vhost_log_write(int vid, uint64_t addr, uint64_t len)
1386 struct virtio_net *dev = get_device(vid);
1391 vhost_log_write(dev, addr, len);
1395 rte_vhost_log_used_vring(int vid, uint16_t vring_idx,
1396 uint64_t offset, uint64_t len)
1398 struct virtio_net *dev;
1399 struct vhost_virtqueue *vq;
1401 dev = get_device(vid);
1405 if (vring_idx >= VHOST_MAX_VRING)
1407 vq = dev->virtqueue[vring_idx];
1411 vhost_log_used_vring(dev, vq, offset, len);
1415 rte_vhost_rx_queue_count(int vid, uint16_t qid)
1417 struct virtio_net *dev;
1418 struct vhost_virtqueue *vq;
1421 dev = get_device(vid);
1425 if (unlikely(qid >= dev->nr_vring || (qid & 1) == 0)) {
1426 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1427 dev->vid, __func__, qid);
1431 vq = dev->virtqueue[qid];
1435 rte_spinlock_lock(&vq->access_lock);
1437 if (unlikely(vq->enabled == 0 || vq->avail == NULL))
1440 ret = *((volatile uint16_t *)&vq->avail->idx) - vq->last_avail_idx;
1443 rte_spinlock_unlock(&vq->access_lock);
1447 struct rte_vdpa_device *
1448 rte_vhost_get_vdpa_device(int vid)
1450 struct virtio_net *dev = get_device(vid);
1455 return dev->vdpa_dev;
1458 int rte_vhost_get_log_base(int vid, uint64_t *log_base,
1461 struct virtio_net *dev = get_device(vid);
1463 if (dev == NULL || log_base == NULL || log_size == NULL)
1466 *log_base = dev->log_base;
1467 *log_size = dev->log_size;
1472 int rte_vhost_get_vring_base(int vid, uint16_t queue_id,
1473 uint16_t *last_avail_idx, uint16_t *last_used_idx)
1475 struct vhost_virtqueue *vq;
1476 struct virtio_net *dev = get_device(vid);
1478 if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1481 if (queue_id >= VHOST_MAX_VRING)
1484 vq = dev->virtqueue[queue_id];
1488 if (vq_is_packed(dev)) {
1489 *last_avail_idx = (vq->avail_wrap_counter << 15) |
1491 *last_used_idx = (vq->used_wrap_counter << 15) |
1494 *last_avail_idx = vq->last_avail_idx;
1495 *last_used_idx = vq->last_used_idx;
1501 int rte_vhost_set_vring_base(int vid, uint16_t queue_id,
1502 uint16_t last_avail_idx, uint16_t last_used_idx)
1504 struct vhost_virtqueue *vq;
1505 struct virtio_net *dev = get_device(vid);
1510 if (queue_id >= VHOST_MAX_VRING)
1513 vq = dev->virtqueue[queue_id];
1517 if (vq_is_packed(dev)) {
1518 vq->last_avail_idx = last_avail_idx & 0x7fff;
1519 vq->avail_wrap_counter = !!(last_avail_idx & (1 << 15));
1520 vq->last_used_idx = last_used_idx & 0x7fff;
1521 vq->used_wrap_counter = !!(last_used_idx & (1 << 15));
1523 vq->last_avail_idx = last_avail_idx;
1524 vq->last_used_idx = last_used_idx;
1531 rte_vhost_get_vring_base_from_inflight(int vid,
1533 uint16_t *last_avail_idx,
1534 uint16_t *last_used_idx)
1536 struct rte_vhost_inflight_info_packed *inflight_info;
1537 struct vhost_virtqueue *vq;
1538 struct virtio_net *dev = get_device(vid);
1540 if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1543 if (queue_id >= VHOST_MAX_VRING)
1546 vq = dev->virtqueue[queue_id];
1550 if (!vq_is_packed(dev))
1553 inflight_info = vq->inflight_packed;
1557 *last_avail_idx = (inflight_info->old_used_wrap_counter << 15) |
1558 inflight_info->old_used_idx;
1559 *last_used_idx = *last_avail_idx;
1564 int rte_vhost_extern_callback_register(int vid,
1565 struct rte_vhost_user_extern_ops const * const ops, void *ctx)
1567 struct virtio_net *dev = get_device(vid);
1569 if (dev == NULL || ops == NULL)
1572 dev->extern_ops = *ops;
1573 dev->extern_data = ctx;
1577 int rte_vhost_async_channel_register(int vid, uint16_t queue_id,
1579 struct rte_vhost_async_channel_ops *ops)
1581 struct vhost_virtqueue *vq;
1582 struct virtio_net *dev = get_device(vid);
1583 struct rte_vhost_async_features f;
1586 if (dev == NULL || ops == NULL)
1589 f.intval = features;
1591 if (queue_id >= VHOST_MAX_VRING)
1594 vq = dev->virtqueue[queue_id];
1596 if (unlikely(vq == NULL || !dev->async_copy))
1599 /* packed queue is not supported */
1600 if (unlikely(vq_is_packed(dev) || !f.async_inorder)) {
1601 VHOST_LOG_CONFIG(ERR,
1602 "async copy is not supported on packed queue or non-inorder mode "
1603 "(vid %d, qid: %d)\n", vid, queue_id);
1607 if (unlikely(ops->check_completed_copies == NULL ||
1608 ops->transfer_data == NULL))
1611 rte_spinlock_lock(&vq->access_lock);
1613 if (unlikely(vq->async_registered)) {
1614 VHOST_LOG_CONFIG(ERR,
1615 "async register failed: channel already registered "
1616 "(vid %d, qid: %d)\n", vid, queue_id);
1620 #ifdef RTE_LIBRTE_VHOST_NUMA
1621 if (get_mempolicy(&node, NULL, 0, vq, MPOL_F_NODE | MPOL_F_ADDR)) {
1622 VHOST_LOG_CONFIG(ERR,
1623 "unable to get numa information in async register. "
1624 "allocating async buffer memory on the caller thread node\n");
1625 node = SOCKET_ID_ANY;
1628 node = SOCKET_ID_ANY;
1631 vq->async_pkts_info = rte_malloc_socket(NULL,
1632 vq->size * sizeof(struct async_inflight_info),
1633 RTE_CACHE_LINE_SIZE, node);
1634 vq->it_pool = rte_malloc_socket(NULL,
1635 VHOST_MAX_ASYNC_IT * sizeof(struct rte_vhost_iov_iter),
1636 RTE_CACHE_LINE_SIZE, node);
1637 vq->vec_pool = rte_malloc_socket(NULL,
1638 VHOST_MAX_ASYNC_VEC * sizeof(struct iovec),
1639 RTE_CACHE_LINE_SIZE, node);
1640 vq->async_descs_split = rte_malloc_socket(NULL,
1641 vq->size * sizeof(struct vring_used_elem),
1642 RTE_CACHE_LINE_SIZE, node);
1643 if (!vq->async_descs_split || !vq->async_pkts_info ||
1644 !vq->it_pool || !vq->vec_pool) {
1645 vhost_free_async_mem(vq);
1646 VHOST_LOG_CONFIG(ERR,
1647 "async register failed: cannot allocate memory for vq data "
1648 "(vid %d, qid: %d)\n", vid, queue_id);
1652 vq->async_ops.check_completed_copies = ops->check_completed_copies;
1653 vq->async_ops.transfer_data = ops->transfer_data;
1655 vq->async_inorder = f.async_inorder;
1656 vq->async_threshold = f.async_threshold;
1658 vq->async_registered = true;
1661 rte_spinlock_unlock(&vq->access_lock);
1666 int rte_vhost_async_channel_unregister(int vid, uint16_t queue_id)
1668 struct vhost_virtqueue *vq;
1669 struct virtio_net *dev = get_device(vid);
1675 if (queue_id >= VHOST_MAX_VRING)
1678 vq = dev->virtqueue[queue_id];
1685 if (!vq->async_registered)
1688 if (!rte_spinlock_trylock(&vq->access_lock)) {
1689 VHOST_LOG_CONFIG(ERR, "Failed to unregister async channel. "
1690 "virt queue busy.\n");
1694 if (vq->async_pkts_inflight_n) {
1695 VHOST_LOG_CONFIG(ERR, "Failed to unregister async channel. "
1696 "async inflight packets must be completed before unregistration.\n");
1701 vhost_free_async_mem(vq);
1703 vq->async_ops.transfer_data = NULL;
1704 vq->async_ops.check_completed_copies = NULL;
1705 vq->async_registered = false;
1708 rte_spinlock_unlock(&vq->access_lock);
1713 RTE_LOG_REGISTER(vhost_config_log_level, lib.vhost.config, INFO);
1714 RTE_LOG_REGISTER(vhost_data_log_level, lib.vhost.data, WARNING);