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(vq, iova, perm);
60 if (vhost_user_iotlb_miss(dev, iova, perm)) {
62 "IOTLB miss req failed for IOVA 0x%" PRIx64 "\n",
64 vhost_user_iotlb_pending_remove(vq, iova, 1, perm);
67 vhost_user_iotlb_rd_lock(vq);
73 #define VHOST_LOG_PAGE 4096
76 * Atomically set a bit in memory.
78 static __rte_always_inline void
79 vhost_set_bit(unsigned int nr, volatile uint8_t *addr)
81 #if defined(RTE_TOOLCHAIN_GCC) && (GCC_VERSION < 70100)
83 * __sync_ built-ins are deprecated, but __atomic_ ones
84 * are sub-optimized in older GCC versions.
86 __sync_fetch_and_or_1(addr, (1U << nr));
88 __atomic_fetch_or(addr, (1U << nr), __ATOMIC_RELAXED);
92 static __rte_always_inline void
93 vhost_log_page(uint8_t *log_base, uint64_t page)
95 vhost_set_bit(page % 8, &log_base[page / 8]);
99 __vhost_log_write(struct virtio_net *dev, uint64_t addr, uint64_t len)
103 if (unlikely(!dev->log_base || !len))
106 if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
109 /* To make sure guest memory updates are committed before logging */
110 rte_atomic_thread_fence(__ATOMIC_RELEASE);
112 page = addr / VHOST_LOG_PAGE;
113 while (page * VHOST_LOG_PAGE < addr + len) {
114 vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
120 __vhost_log_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
121 uint64_t iova, uint64_t len)
123 uint64_t hva, gpa, map_len;
126 hva = __vhost_iova_to_vva(dev, vq, iova, &map_len, VHOST_ACCESS_RW);
127 if (map_len != len) {
129 "Failed to write log for IOVA 0x%" PRIx64 ". No IOTLB entry found\n",
134 gpa = hva_to_gpa(dev, hva, len);
136 __vhost_log_write(dev, gpa, len);
140 __vhost_log_cache_sync(struct virtio_net *dev, struct vhost_virtqueue *vq)
142 unsigned long *log_base;
145 if (unlikely(!dev->log_base))
148 /* No cache, nothing to sync */
149 if (unlikely(!vq->log_cache))
152 rte_atomic_thread_fence(__ATOMIC_RELEASE);
154 log_base = (unsigned long *)(uintptr_t)dev->log_base;
156 for (i = 0; i < vq->log_cache_nb_elem; i++) {
157 struct log_cache_entry *elem = vq->log_cache + i;
159 #if defined(RTE_TOOLCHAIN_GCC) && (GCC_VERSION < 70100)
161 * '__sync' builtins are deprecated, but '__atomic' ones
162 * are sub-optimized in older GCC versions.
164 __sync_fetch_and_or(log_base + elem->offset, elem->val);
166 __atomic_fetch_or(log_base + elem->offset, elem->val,
171 rte_atomic_thread_fence(__ATOMIC_RELEASE);
173 vq->log_cache_nb_elem = 0;
176 static __rte_always_inline void
177 vhost_log_cache_page(struct virtio_net *dev, struct vhost_virtqueue *vq,
180 uint32_t bit_nr = page % (sizeof(unsigned long) << 3);
181 uint32_t offset = page / (sizeof(unsigned long) << 3);
184 if (unlikely(!vq->log_cache)) {
185 /* No logging cache allocated, write dirty log map directly */
186 rte_atomic_thread_fence(__ATOMIC_RELEASE);
187 vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
192 for (i = 0; i < vq->log_cache_nb_elem; i++) {
193 struct log_cache_entry *elem = vq->log_cache + i;
195 if (elem->offset == offset) {
196 elem->val |= (1UL << bit_nr);
201 if (unlikely(i >= VHOST_LOG_CACHE_NR)) {
203 * No more room for a new log cache entry,
204 * so write the dirty log map directly.
206 rte_atomic_thread_fence(__ATOMIC_RELEASE);
207 vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
212 vq->log_cache[i].offset = offset;
213 vq->log_cache[i].val = (1UL << bit_nr);
214 vq->log_cache_nb_elem++;
218 __vhost_log_cache_write(struct virtio_net *dev, struct vhost_virtqueue *vq,
219 uint64_t addr, uint64_t len)
223 if (unlikely(!dev->log_base || !len))
226 if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
229 page = addr / VHOST_LOG_PAGE;
230 while (page * VHOST_LOG_PAGE < addr + len) {
231 vhost_log_cache_page(dev, vq, page);
237 __vhost_log_cache_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
238 uint64_t iova, uint64_t len)
240 uint64_t hva, gpa, map_len;
243 hva = __vhost_iova_to_vva(dev, vq, iova, &map_len, VHOST_ACCESS_RW);
244 if (map_len != len) {
246 "Failed to write log for IOVA 0x%" PRIx64 ". No IOTLB entry found\n",
251 gpa = hva_to_gpa(dev, hva, len);
253 __vhost_log_cache_write(dev, vq, gpa, len);
257 vhost_alloc_copy_ind_table(struct virtio_net *dev, struct vhost_virtqueue *vq,
258 uint64_t desc_addr, uint64_t desc_len)
262 uint64_t len, remain = desc_len;
264 idesc = rte_malloc(__func__, desc_len, 0);
265 if (unlikely(!idesc))
268 dst = (uint64_t)(uintptr_t)idesc;
272 src = vhost_iova_to_vva(dev, vq, desc_addr, &len,
274 if (unlikely(!src || !len)) {
279 rte_memcpy((void *)(uintptr_t)dst, (void *)(uintptr_t)src, len);
290 cleanup_vq(struct vhost_virtqueue *vq, int destroy)
292 if ((vq->callfd >= 0) && (destroy != 0))
299 cleanup_vq_inflight(struct virtio_net *dev, struct vhost_virtqueue *vq)
301 if (!(dev->protocol_features &
302 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
305 if (vq_is_packed(dev)) {
306 if (vq->inflight_packed)
307 vq->inflight_packed = NULL;
309 if (vq->inflight_split)
310 vq->inflight_split = NULL;
313 if (vq->resubmit_inflight) {
314 if (vq->resubmit_inflight->resubmit_list) {
315 free(vq->resubmit_inflight->resubmit_list);
316 vq->resubmit_inflight->resubmit_list = NULL;
318 free(vq->resubmit_inflight);
319 vq->resubmit_inflight = NULL;
324 * Unmap any memory, close any file descriptors and
325 * free any memory owned by a device.
328 cleanup_device(struct virtio_net *dev, int destroy)
332 vhost_backend_cleanup(dev);
334 for (i = 0; i < dev->nr_vring; i++) {
335 cleanup_vq(dev->virtqueue[i], destroy);
336 cleanup_vq_inflight(dev, dev->virtqueue[i]);
341 vhost_free_async_mem(struct vhost_virtqueue *vq)
343 rte_free(vq->async_pkts_info);
345 rte_free(vq->async_buffers_packed);
346 vq->async_buffers_packed = NULL;
347 rte_free(vq->async_descs_split);
348 vq->async_descs_split = NULL;
350 rte_free(vq->it_pool);
351 rte_free(vq->vec_pool);
353 vq->async_pkts_info = NULL;
359 free_vq(struct virtio_net *dev, struct vhost_virtqueue *vq)
361 if (vq_is_packed(dev))
362 rte_free(vq->shadow_used_packed);
364 rte_free(vq->shadow_used_split);
366 vhost_free_async_mem(vq);
367 rte_free(vq->batch_copy_elems);
369 rte_mempool_free(vq->iotlb_pool);
370 rte_free(vq->log_cache);
375 * Release virtqueues and device memory.
378 free_device(struct virtio_net *dev)
382 for (i = 0; i < dev->nr_vring; i++)
383 free_vq(dev, dev->virtqueue[i]);
388 static __rte_always_inline int
389 log_translate(struct virtio_net *dev, struct vhost_virtqueue *vq)
391 if (likely(!(vq->ring_addrs.flags & (1 << VHOST_VRING_F_LOG))))
394 vq->log_guest_addr = translate_log_addr(dev, vq,
395 vq->ring_addrs.log_guest_addr);
396 if (vq->log_guest_addr == 0)
403 * Converts vring log address to GPA
404 * If IOMMU is enabled, the log address is IOVA
405 * If IOMMU not enabled, the log address is already GPA
407 * Caller should have iotlb_lock read-locked
410 translate_log_addr(struct virtio_net *dev, struct vhost_virtqueue *vq,
413 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
414 const uint64_t exp_size = sizeof(uint64_t);
416 uint64_t size = exp_size;
418 hva = vhost_iova_to_vva(dev, vq, log_addr,
419 &size, VHOST_ACCESS_RW);
421 if (size != exp_size)
424 gpa = hva_to_gpa(dev, hva, exp_size);
426 VHOST_LOG_CONFIG(ERR,
427 "VQ: Failed to find GPA for log_addr: 0x%"
428 PRIx64 " hva: 0x%" PRIx64 "\n",
438 /* Caller should have iotlb_lock read-locked */
440 vring_translate_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
442 uint64_t req_size, size;
444 req_size = sizeof(struct vring_desc) * vq->size;
446 vq->desc = (struct vring_desc *)(uintptr_t)vhost_iova_to_vva(dev, vq,
447 vq->ring_addrs.desc_user_addr,
448 &size, VHOST_ACCESS_RW);
449 if (!vq->desc || size != req_size)
452 req_size = sizeof(struct vring_avail);
453 req_size += sizeof(uint16_t) * vq->size;
454 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
455 req_size += sizeof(uint16_t);
457 vq->avail = (struct vring_avail *)(uintptr_t)vhost_iova_to_vva(dev, vq,
458 vq->ring_addrs.avail_user_addr,
459 &size, VHOST_ACCESS_RW);
460 if (!vq->avail || size != req_size)
463 req_size = sizeof(struct vring_used);
464 req_size += sizeof(struct vring_used_elem) * vq->size;
465 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
466 req_size += sizeof(uint16_t);
468 vq->used = (struct vring_used *)(uintptr_t)vhost_iova_to_vva(dev, vq,
469 vq->ring_addrs.used_user_addr,
470 &size, VHOST_ACCESS_RW);
471 if (!vq->used || size != req_size)
477 /* Caller should have iotlb_lock read-locked */
479 vring_translate_packed(struct virtio_net *dev, struct vhost_virtqueue *vq)
481 uint64_t req_size, size;
483 req_size = sizeof(struct vring_packed_desc) * vq->size;
485 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
486 vhost_iova_to_vva(dev, vq, vq->ring_addrs.desc_user_addr,
487 &size, VHOST_ACCESS_RW);
488 if (!vq->desc_packed || size != req_size)
491 req_size = sizeof(struct vring_packed_desc_event);
493 vq->driver_event = (struct vring_packed_desc_event *)(uintptr_t)
494 vhost_iova_to_vva(dev, vq, vq->ring_addrs.avail_user_addr,
495 &size, VHOST_ACCESS_RW);
496 if (!vq->driver_event || size != req_size)
499 req_size = sizeof(struct vring_packed_desc_event);
501 vq->device_event = (struct vring_packed_desc_event *)(uintptr_t)
502 vhost_iova_to_vva(dev, vq, vq->ring_addrs.used_user_addr,
503 &size, VHOST_ACCESS_RW);
504 if (!vq->device_event || size != req_size)
511 vring_translate(struct virtio_net *dev, struct vhost_virtqueue *vq)
514 if (!(dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
517 if (vq_is_packed(dev)) {
518 if (vring_translate_packed(dev, vq) < 0)
521 if (vring_translate_split(dev, vq) < 0)
525 if (log_translate(dev, vq) < 0)
528 vq->access_ok = true;
534 vring_invalidate(struct virtio_net *dev, struct vhost_virtqueue *vq)
536 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
537 vhost_user_iotlb_wr_lock(vq);
539 vq->access_ok = false;
543 vq->log_guest_addr = 0;
545 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
546 vhost_user_iotlb_wr_unlock(vq);
550 init_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
552 struct vhost_virtqueue *vq;
554 if (vring_idx >= VHOST_MAX_VRING) {
555 VHOST_LOG_CONFIG(ERR,
556 "Failed not init vring, out of bound (%d)\n",
561 vq = dev->virtqueue[vring_idx];
563 VHOST_LOG_CONFIG(ERR, "Virtqueue not allocated (%d)\n",
568 memset(vq, 0, sizeof(struct vhost_virtqueue));
570 vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
571 vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
572 vq->notif_enable = VIRTIO_UNINITIALIZED_NOTIF;
576 reset_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
578 struct vhost_virtqueue *vq;
581 if (vring_idx >= VHOST_MAX_VRING) {
582 VHOST_LOG_CONFIG(ERR,
583 "Failed not init vring, out of bound (%d)\n",
588 vq = dev->virtqueue[vring_idx];
590 VHOST_LOG_CONFIG(ERR, "Virtqueue not allocated (%d)\n",
596 init_vring_queue(dev, vring_idx);
601 alloc_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
603 struct vhost_virtqueue *vq;
606 /* Also allocate holes, if any, up to requested vring index. */
607 for (i = 0; i <= vring_idx; i++) {
608 if (dev->virtqueue[i])
611 vq = rte_zmalloc(NULL, sizeof(struct vhost_virtqueue), 0);
613 VHOST_LOG_CONFIG(ERR,
614 "Failed to allocate memory for vring:%u.\n", i);
618 dev->virtqueue[i] = vq;
619 init_vring_queue(dev, i);
620 rte_spinlock_init(&vq->access_lock);
621 vq->avail_wrap_counter = 1;
622 vq->used_wrap_counter = 1;
623 vq->signalled_used_valid = false;
626 dev->nr_vring = RTE_MAX(dev->nr_vring, vring_idx + 1);
632 * Reset some variables in device structure, while keeping few
633 * others untouched, such as vid, ifname, nr_vring: they
634 * should be same unless the device is removed.
637 reset_device(struct virtio_net *dev)
642 dev->protocol_features = 0;
643 dev->flags &= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
645 for (i = 0; i < dev->nr_vring; i++)
646 reset_vring_queue(dev, i);
650 * Invoked when there is a new vhost-user connection established (when
651 * there is a new virtio device being attached).
654 vhost_new_device(void)
656 struct virtio_net *dev;
659 pthread_mutex_lock(&vhost_dev_lock);
660 for (i = 0; i < MAX_VHOST_DEVICE; i++) {
661 if (vhost_devices[i] == NULL)
665 if (i == MAX_VHOST_DEVICE) {
666 VHOST_LOG_CONFIG(ERR,
667 "Failed to find a free slot for new device.\n");
668 pthread_mutex_unlock(&vhost_dev_lock);
672 dev = rte_zmalloc(NULL, sizeof(struct virtio_net), 0);
674 VHOST_LOG_CONFIG(ERR,
675 "Failed to allocate memory for new dev.\n");
676 pthread_mutex_unlock(&vhost_dev_lock);
680 vhost_devices[i] = dev;
681 pthread_mutex_unlock(&vhost_dev_lock);
684 dev->flags = VIRTIO_DEV_BUILTIN_VIRTIO_NET;
685 dev->slave_req_fd = -1;
686 dev->postcopy_ufd = -1;
687 rte_spinlock_init(&dev->slave_req_lock);
693 vhost_destroy_device_notify(struct virtio_net *dev)
695 struct rte_vdpa_device *vdpa_dev;
697 if (dev->flags & VIRTIO_DEV_RUNNING) {
698 vdpa_dev = dev->vdpa_dev;
700 vdpa_dev->ops->dev_close(dev->vid);
701 dev->flags &= ~VIRTIO_DEV_RUNNING;
702 dev->notify_ops->destroy_device(dev->vid);
707 * Invoked when there is the vhost-user connection is broken (when
708 * the virtio device is being detached).
711 vhost_destroy_device(int vid)
713 struct virtio_net *dev = get_device(vid);
718 vhost_destroy_device_notify(dev);
720 cleanup_device(dev, 1);
723 vhost_devices[vid] = NULL;
727 vhost_attach_vdpa_device(int vid, struct rte_vdpa_device *vdpa_dev)
729 struct virtio_net *dev = get_device(vid);
734 dev->vdpa_dev = vdpa_dev;
738 vhost_set_ifname(int vid, const char *if_name, unsigned int if_len)
740 struct virtio_net *dev;
743 dev = get_device(vid);
747 len = if_len > sizeof(dev->ifname) ?
748 sizeof(dev->ifname) : if_len;
750 strncpy(dev->ifname, if_name, len);
751 dev->ifname[sizeof(dev->ifname) - 1] = '\0';
755 vhost_setup_virtio_net(int vid, bool enable, bool compliant_ol_flags)
757 struct virtio_net *dev = get_device(vid);
763 dev->flags |= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
765 dev->flags &= ~VIRTIO_DEV_BUILTIN_VIRTIO_NET;
766 if (!compliant_ol_flags)
767 dev->flags |= VIRTIO_DEV_LEGACY_OL_FLAGS;
769 dev->flags &= ~VIRTIO_DEV_LEGACY_OL_FLAGS;
773 vhost_enable_extbuf(int vid)
775 struct virtio_net *dev = get_device(vid);
784 vhost_enable_linearbuf(int vid)
786 struct virtio_net *dev = get_device(vid);
795 rte_vhost_get_mtu(int vid, uint16_t *mtu)
797 struct virtio_net *dev = get_device(vid);
799 if (dev == NULL || mtu == NULL)
802 if (!(dev->flags & VIRTIO_DEV_READY))
805 if (!(dev->features & (1ULL << VIRTIO_NET_F_MTU)))
814 rte_vhost_get_numa_node(int vid)
816 #ifdef RTE_LIBRTE_VHOST_NUMA
817 struct virtio_net *dev = get_device(vid);
821 if (dev == NULL || numa_available() != 0)
824 ret = get_mempolicy(&numa_node, NULL, 0, dev,
825 MPOL_F_NODE | MPOL_F_ADDR);
827 VHOST_LOG_CONFIG(ERR,
828 "(%d) failed to query numa node: %s\n",
829 vid, rte_strerror(errno));
841 rte_vhost_get_queue_num(int vid)
843 struct virtio_net *dev = get_device(vid);
848 return dev->nr_vring / 2;
852 rte_vhost_get_vring_num(int vid)
854 struct virtio_net *dev = get_device(vid);
859 return dev->nr_vring;
863 rte_vhost_get_ifname(int vid, char *buf, size_t len)
865 struct virtio_net *dev = get_device(vid);
867 if (dev == NULL || buf == NULL)
870 len = RTE_MIN(len, sizeof(dev->ifname));
872 strncpy(buf, dev->ifname, len);
879 rte_vhost_get_negotiated_features(int vid, uint64_t *features)
881 struct virtio_net *dev;
883 dev = get_device(vid);
884 if (dev == NULL || features == NULL)
887 *features = dev->features;
892 rte_vhost_get_negotiated_protocol_features(int vid,
893 uint64_t *protocol_features)
895 struct virtio_net *dev;
897 dev = get_device(vid);
898 if (dev == NULL || protocol_features == NULL)
901 *protocol_features = dev->protocol_features;
906 rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem)
908 struct virtio_net *dev;
909 struct rte_vhost_memory *m;
912 dev = get_device(vid);
913 if (dev == NULL || mem == NULL)
916 size = dev->mem->nregions * sizeof(struct rte_vhost_mem_region);
917 m = malloc(sizeof(struct rte_vhost_memory) + size);
921 m->nregions = dev->mem->nregions;
922 memcpy(m->regions, dev->mem->regions, size);
929 rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
930 struct rte_vhost_vring *vring)
932 struct virtio_net *dev;
933 struct vhost_virtqueue *vq;
935 dev = get_device(vid);
936 if (dev == NULL || vring == NULL)
939 if (vring_idx >= VHOST_MAX_VRING)
942 vq = dev->virtqueue[vring_idx];
946 if (vq_is_packed(dev)) {
947 vring->desc_packed = vq->desc_packed;
948 vring->driver_event = vq->driver_event;
949 vring->device_event = vq->device_event;
951 vring->desc = vq->desc;
952 vring->avail = vq->avail;
953 vring->used = vq->used;
955 vring->log_guest_addr = vq->log_guest_addr;
957 vring->callfd = vq->callfd;
958 vring->kickfd = vq->kickfd;
959 vring->size = vq->size;
965 rte_vhost_get_vhost_ring_inflight(int vid, uint16_t vring_idx,
966 struct rte_vhost_ring_inflight *vring)
968 struct virtio_net *dev;
969 struct vhost_virtqueue *vq;
971 dev = get_device(vid);
975 if (vring_idx >= VHOST_MAX_VRING)
978 vq = dev->virtqueue[vring_idx];
982 if (vq_is_packed(dev)) {
983 if (unlikely(!vq->inflight_packed))
986 vring->inflight_packed = vq->inflight_packed;
988 if (unlikely(!vq->inflight_split))
991 vring->inflight_split = vq->inflight_split;
994 vring->resubmit_inflight = vq->resubmit_inflight;
1000 rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx,
1003 struct vhost_virtqueue *vq;
1004 struct virtio_net *dev;
1006 dev = get_device(vid);
1010 if (unlikely(!(dev->protocol_features &
1011 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1014 if (unlikely(vq_is_packed(dev)))
1017 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1020 vq = dev->virtqueue[vring_idx];
1024 if (unlikely(!vq->inflight_split))
1027 if (unlikely(idx >= vq->size))
1030 vq->inflight_split->desc[idx].counter = vq->global_counter++;
1031 vq->inflight_split->desc[idx].inflight = 1;
1036 rte_vhost_set_inflight_desc_packed(int vid, uint16_t vring_idx,
1037 uint16_t head, uint16_t last,
1038 uint16_t *inflight_entry)
1040 struct rte_vhost_inflight_info_packed *inflight_info;
1041 struct virtio_net *dev;
1042 struct vhost_virtqueue *vq;
1043 struct vring_packed_desc *desc;
1044 uint16_t old_free_head, free_head;
1046 dev = get_device(vid);
1050 if (unlikely(!(dev->protocol_features &
1051 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1054 if (unlikely(!vq_is_packed(dev)))
1057 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1060 vq = dev->virtqueue[vring_idx];
1064 inflight_info = vq->inflight_packed;
1065 if (unlikely(!inflight_info))
1068 if (unlikely(head >= vq->size))
1071 desc = vq->desc_packed;
1072 old_free_head = inflight_info->old_free_head;
1073 if (unlikely(old_free_head >= vq->size))
1076 free_head = old_free_head;
1078 /* init header descriptor */
1079 inflight_info->desc[old_free_head].num = 0;
1080 inflight_info->desc[old_free_head].counter = vq->global_counter++;
1081 inflight_info->desc[old_free_head].inflight = 1;
1083 /* save desc entry in flight entry */
1084 while (head != ((last + 1) % vq->size)) {
1085 inflight_info->desc[old_free_head].num++;
1086 inflight_info->desc[free_head].addr = desc[head].addr;
1087 inflight_info->desc[free_head].len = desc[head].len;
1088 inflight_info->desc[free_head].flags = desc[head].flags;
1089 inflight_info->desc[free_head].id = desc[head].id;
1091 inflight_info->desc[old_free_head].last = free_head;
1092 free_head = inflight_info->desc[free_head].next;
1093 inflight_info->free_head = free_head;
1094 head = (head + 1) % vq->size;
1097 inflight_info->old_free_head = free_head;
1098 *inflight_entry = old_free_head;
1104 rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
1105 uint16_t last_used_idx, uint16_t idx)
1107 struct virtio_net *dev;
1108 struct vhost_virtqueue *vq;
1110 dev = get_device(vid);
1114 if (unlikely(!(dev->protocol_features &
1115 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1118 if (unlikely(vq_is_packed(dev)))
1121 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1124 vq = dev->virtqueue[vring_idx];
1128 if (unlikely(!vq->inflight_split))
1131 if (unlikely(idx >= vq->size))
1134 rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1136 vq->inflight_split->desc[idx].inflight = 0;
1138 rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1140 vq->inflight_split->used_idx = last_used_idx;
1145 rte_vhost_clr_inflight_desc_packed(int vid, uint16_t vring_idx,
1148 struct rte_vhost_inflight_info_packed *inflight_info;
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 inflight_info = vq->inflight_packed;
1171 if (unlikely(!inflight_info))
1174 if (unlikely(head >= vq->size))
1177 rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1179 inflight_info->desc[head].inflight = 0;
1181 rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1183 inflight_info->old_free_head = inflight_info->free_head;
1184 inflight_info->old_used_idx = inflight_info->used_idx;
1185 inflight_info->old_used_wrap_counter = inflight_info->used_wrap_counter;
1191 rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx,
1194 struct virtio_net *dev;
1195 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 if (unlikely(!vq->inflight_split))
1218 vq->inflight_split->last_inflight_io = idx;
1223 rte_vhost_set_last_inflight_io_packed(int vid, uint16_t vring_idx,
1226 struct rte_vhost_inflight_info_packed *inflight_info;
1227 struct virtio_net *dev;
1228 struct vhost_virtqueue *vq;
1231 dev = get_device(vid);
1235 if (unlikely(!(dev->protocol_features &
1236 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1239 if (unlikely(!vq_is_packed(dev)))
1242 if (unlikely(vring_idx >= VHOST_MAX_VRING))
1245 vq = dev->virtqueue[vring_idx];
1249 inflight_info = vq->inflight_packed;
1250 if (unlikely(!inflight_info))
1253 if (unlikely(head >= vq->size))
1256 last = inflight_info->desc[head].last;
1257 if (unlikely(last >= vq->size))
1260 inflight_info->desc[last].next = inflight_info->free_head;
1261 inflight_info->free_head = head;
1262 inflight_info->used_idx += inflight_info->desc[head].num;
1263 if (inflight_info->used_idx >= inflight_info->desc_num) {
1264 inflight_info->used_idx -= inflight_info->desc_num;
1265 inflight_info->used_wrap_counter =
1266 !inflight_info->used_wrap_counter;
1273 rte_vhost_vring_call(int vid, uint16_t vring_idx)
1275 struct virtio_net *dev;
1276 struct vhost_virtqueue *vq;
1278 dev = get_device(vid);
1282 if (vring_idx >= VHOST_MAX_VRING)
1285 vq = dev->virtqueue[vring_idx];
1289 if (vq_is_packed(dev))
1290 vhost_vring_call_packed(dev, vq);
1292 vhost_vring_call_split(dev, vq);
1298 rte_vhost_avail_entries(int vid, uint16_t queue_id)
1300 struct virtio_net *dev;
1301 struct vhost_virtqueue *vq;
1304 dev = get_device(vid);
1308 if (queue_id >= VHOST_MAX_VRING)
1311 vq = dev->virtqueue[queue_id];
1315 rte_spinlock_lock(&vq->access_lock);
1317 if (unlikely(!vq->enabled || vq->avail == NULL))
1320 ret = *(volatile uint16_t *)&vq->avail->idx - vq->last_used_idx;
1323 rte_spinlock_unlock(&vq->access_lock);
1328 vhost_enable_notify_split(struct virtio_net *dev,
1329 struct vhost_virtqueue *vq, int enable)
1331 if (vq->used == NULL)
1334 if (!(dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))) {
1336 vq->used->flags &= ~VRING_USED_F_NO_NOTIFY;
1338 vq->used->flags |= VRING_USED_F_NO_NOTIFY;
1341 vhost_avail_event(vq) = vq->last_avail_idx;
1347 vhost_enable_notify_packed(struct virtio_net *dev,
1348 struct vhost_virtqueue *vq, int enable)
1352 if (vq->device_event == NULL)
1356 vq->device_event->flags = VRING_EVENT_F_DISABLE;
1360 flags = VRING_EVENT_F_ENABLE;
1361 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX)) {
1362 flags = VRING_EVENT_F_DESC;
1363 vq->device_event->off_wrap = vq->last_avail_idx |
1364 vq->avail_wrap_counter << 15;
1367 rte_atomic_thread_fence(__ATOMIC_RELEASE);
1369 vq->device_event->flags = flags;
1374 vhost_enable_guest_notification(struct virtio_net *dev,
1375 struct vhost_virtqueue *vq, int enable)
1378 * If the virtqueue is not ready yet, it will be applied
1379 * when it will become ready.
1384 if (vq_is_packed(dev))
1385 return vhost_enable_notify_packed(dev, vq, enable);
1387 return vhost_enable_notify_split(dev, vq, enable);
1391 rte_vhost_enable_guest_notification(int vid, uint16_t queue_id, int enable)
1393 struct virtio_net *dev = get_device(vid);
1394 struct vhost_virtqueue *vq;
1400 if (queue_id >= VHOST_MAX_VRING)
1403 vq = dev->virtqueue[queue_id];
1407 rte_spinlock_lock(&vq->access_lock);
1409 vq->notif_enable = enable;
1410 ret = vhost_enable_guest_notification(dev, vq, enable);
1412 rte_spinlock_unlock(&vq->access_lock);
1418 rte_vhost_log_write(int vid, uint64_t addr, uint64_t len)
1420 struct virtio_net *dev = get_device(vid);
1425 vhost_log_write(dev, addr, len);
1429 rte_vhost_log_used_vring(int vid, uint16_t vring_idx,
1430 uint64_t offset, uint64_t len)
1432 struct virtio_net *dev;
1433 struct vhost_virtqueue *vq;
1435 dev = get_device(vid);
1439 if (vring_idx >= VHOST_MAX_VRING)
1441 vq = dev->virtqueue[vring_idx];
1445 vhost_log_used_vring(dev, vq, offset, len);
1449 rte_vhost_rx_queue_count(int vid, uint16_t qid)
1451 struct virtio_net *dev;
1452 struct vhost_virtqueue *vq;
1455 dev = get_device(vid);
1459 if (unlikely(qid >= dev->nr_vring || (qid & 1) == 0)) {
1460 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1461 dev->vid, __func__, qid);
1465 vq = dev->virtqueue[qid];
1469 rte_spinlock_lock(&vq->access_lock);
1471 if (unlikely(!vq->enabled || vq->avail == NULL))
1474 ret = *((volatile uint16_t *)&vq->avail->idx) - vq->last_avail_idx;
1477 rte_spinlock_unlock(&vq->access_lock);
1481 struct rte_vdpa_device *
1482 rte_vhost_get_vdpa_device(int vid)
1484 struct virtio_net *dev = get_device(vid);
1489 return dev->vdpa_dev;
1492 int rte_vhost_get_log_base(int vid, uint64_t *log_base,
1495 struct virtio_net *dev = get_device(vid);
1497 if (dev == NULL || log_base == NULL || log_size == NULL)
1500 *log_base = dev->log_base;
1501 *log_size = dev->log_size;
1506 int rte_vhost_get_vring_base(int vid, uint16_t queue_id,
1507 uint16_t *last_avail_idx, uint16_t *last_used_idx)
1509 struct vhost_virtqueue *vq;
1510 struct virtio_net *dev = get_device(vid);
1512 if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1515 if (queue_id >= VHOST_MAX_VRING)
1518 vq = dev->virtqueue[queue_id];
1522 if (vq_is_packed(dev)) {
1523 *last_avail_idx = (vq->avail_wrap_counter << 15) |
1525 *last_used_idx = (vq->used_wrap_counter << 15) |
1528 *last_avail_idx = vq->last_avail_idx;
1529 *last_used_idx = vq->last_used_idx;
1535 int rte_vhost_set_vring_base(int vid, uint16_t queue_id,
1536 uint16_t last_avail_idx, uint16_t last_used_idx)
1538 struct vhost_virtqueue *vq;
1539 struct virtio_net *dev = get_device(vid);
1544 if (queue_id >= VHOST_MAX_VRING)
1547 vq = dev->virtqueue[queue_id];
1551 if (vq_is_packed(dev)) {
1552 vq->last_avail_idx = last_avail_idx & 0x7fff;
1553 vq->avail_wrap_counter = !!(last_avail_idx & (1 << 15));
1554 vq->last_used_idx = last_used_idx & 0x7fff;
1555 vq->used_wrap_counter = !!(last_used_idx & (1 << 15));
1557 vq->last_avail_idx = last_avail_idx;
1558 vq->last_used_idx = last_used_idx;
1565 rte_vhost_get_vring_base_from_inflight(int vid,
1567 uint16_t *last_avail_idx,
1568 uint16_t *last_used_idx)
1570 struct rte_vhost_inflight_info_packed *inflight_info;
1571 struct vhost_virtqueue *vq;
1572 struct virtio_net *dev = get_device(vid);
1574 if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1577 if (queue_id >= VHOST_MAX_VRING)
1580 vq = dev->virtqueue[queue_id];
1584 if (!vq_is_packed(dev))
1587 inflight_info = vq->inflight_packed;
1591 *last_avail_idx = (inflight_info->old_used_wrap_counter << 15) |
1592 inflight_info->old_used_idx;
1593 *last_used_idx = *last_avail_idx;
1598 int rte_vhost_extern_callback_register(int vid,
1599 struct rte_vhost_user_extern_ops const * const ops, void *ctx)
1601 struct virtio_net *dev = get_device(vid);
1603 if (dev == NULL || ops == NULL)
1606 dev->extern_ops = *ops;
1607 dev->extern_data = ctx;
1611 int rte_vhost_async_channel_register(int vid, uint16_t queue_id,
1613 struct rte_vhost_async_channel_ops *ops)
1615 struct vhost_virtqueue *vq;
1616 struct virtio_net *dev = get_device(vid);
1617 struct rte_vhost_async_features f;
1620 if (dev == NULL || ops == NULL)
1623 f.intval = features;
1625 if (queue_id >= VHOST_MAX_VRING)
1628 vq = dev->virtqueue[queue_id];
1630 if (unlikely(vq == NULL || !dev->async_copy))
1633 if (unlikely(!f.async_inorder)) {
1634 VHOST_LOG_CONFIG(ERR,
1635 "async copy is not supported on non-inorder mode "
1636 "(vid %d, qid: %d)\n", vid, queue_id);
1640 if (unlikely(ops->check_completed_copies == NULL ||
1641 ops->transfer_data == NULL))
1644 rte_spinlock_lock(&vq->access_lock);
1646 if (unlikely(vq->async_registered)) {
1647 VHOST_LOG_CONFIG(ERR,
1648 "async register failed: channel already registered "
1649 "(vid %d, qid: %d)\n", vid, queue_id);
1653 #ifdef RTE_LIBRTE_VHOST_NUMA
1654 if (get_mempolicy(&node, NULL, 0, vq, MPOL_F_NODE | MPOL_F_ADDR)) {
1655 VHOST_LOG_CONFIG(ERR,
1656 "unable to get numa information in async register. "
1657 "allocating async buffer memory on the caller thread node\n");
1658 node = SOCKET_ID_ANY;
1661 node = SOCKET_ID_ANY;
1664 vq->async_pkts_info = rte_malloc_socket(NULL,
1665 vq->size * sizeof(struct async_inflight_info),
1666 RTE_CACHE_LINE_SIZE, node);
1667 if (!vq->async_pkts_info) {
1668 vhost_free_async_mem(vq);
1669 VHOST_LOG_CONFIG(ERR,
1670 "async register failed: cannot allocate memory for async_pkts_info "
1671 "(vid %d, qid: %d)\n", vid, queue_id);
1675 vq->it_pool = rte_malloc_socket(NULL,
1676 VHOST_MAX_ASYNC_IT * sizeof(struct rte_vhost_iov_iter),
1677 RTE_CACHE_LINE_SIZE, node);
1679 vhost_free_async_mem(vq);
1680 VHOST_LOG_CONFIG(ERR,
1681 "async register failed: cannot allocate memory for it_pool "
1682 "(vid %d, qid: %d)\n", vid, queue_id);
1686 vq->vec_pool = rte_malloc_socket(NULL,
1687 VHOST_MAX_ASYNC_VEC * sizeof(struct iovec),
1688 RTE_CACHE_LINE_SIZE, node);
1689 if (!vq->vec_pool) {
1690 vhost_free_async_mem(vq);
1691 VHOST_LOG_CONFIG(ERR,
1692 "async register failed: cannot allocate memory for vec_pool "
1693 "(vid %d, qid: %d)\n", vid, queue_id);
1697 if (vq_is_packed(dev)) {
1698 vq->async_buffers_packed = rte_malloc_socket(NULL,
1699 vq->size * sizeof(struct vring_used_elem_packed),
1700 RTE_CACHE_LINE_SIZE, node);
1701 if (!vq->async_buffers_packed) {
1702 vhost_free_async_mem(vq);
1703 VHOST_LOG_CONFIG(ERR,
1704 "async register failed: cannot allocate memory for async buffers "
1705 "(vid %d, qid: %d)\n", vid, queue_id);
1709 vq->async_descs_split = rte_malloc_socket(NULL,
1710 vq->size * sizeof(struct vring_used_elem),
1711 RTE_CACHE_LINE_SIZE, node);
1712 if (!vq->async_descs_split) {
1713 vhost_free_async_mem(vq);
1714 VHOST_LOG_CONFIG(ERR,
1715 "async register failed: cannot allocate memory for async descs "
1716 "(vid %d, qid: %d)\n", vid, queue_id);
1721 vq->async_ops.check_completed_copies = ops->check_completed_copies;
1722 vq->async_ops.transfer_data = ops->transfer_data;
1724 vq->async_inorder = f.async_inorder;
1725 vq->async_threshold = f.async_threshold;
1727 vq->async_registered = true;
1730 rte_spinlock_unlock(&vq->access_lock);
1735 int rte_vhost_async_channel_unregister(int vid, uint16_t queue_id)
1737 struct vhost_virtqueue *vq;
1738 struct virtio_net *dev = get_device(vid);
1744 if (queue_id >= VHOST_MAX_VRING)
1747 vq = dev->virtqueue[queue_id];
1754 if (!vq->async_registered)
1757 if (!rte_spinlock_trylock(&vq->access_lock)) {
1758 VHOST_LOG_CONFIG(ERR, "Failed to unregister async channel. "
1759 "virt queue busy.\n");
1763 if (vq->async_pkts_inflight_n) {
1764 VHOST_LOG_CONFIG(ERR, "Failed to unregister async channel. "
1765 "async inflight packets must be completed before unregistration.\n");
1770 vhost_free_async_mem(vq);
1772 vq->async_ops.transfer_data = NULL;
1773 vq->async_ops.check_completed_copies = NULL;
1774 vq->async_registered = false;
1777 rte_spinlock_unlock(&vq->access_lock);
1782 RTE_LOG_REGISTER(vhost_config_log_level, lib.vhost.config, INFO);
1783 RTE_LOG_REGISTER(vhost_data_log_level, lib.vhost.data, WARNING);