1 /* Copyright (C) 2009 Red Hat, Inc.
2 * Copyright (C) 2006 Rusty Russell IBM Corporation
4 * Author: Michael S. Tsirkin <mst@redhat.com>
6 * Inspiration, some code, and most witty comments come from
7 * Documentation/virtual/lguest/lguest.c, by Rusty Russell
9 * This work is licensed under the terms of the GNU GPL, version 2.
11 * Generic code for virtio server in host kernel.
14 #include <linux/eventfd.h>
15 #include <linux/vhost.h>
16 #include <linux/uio.h>
18 #include <linux/mmu_context.h>
19 #include <linux/miscdevice.h>
20 #include <linux/mutex.h>
21 #include <linux/poll.h>
22 #include <linux/file.h>
23 #include <linux/highmem.h>
24 #include <linux/slab.h>
25 #include <linux/vmalloc.h>
26 #include <linux/kthread.h>
27 #include <linux/cgroup.h>
28 #include <linux/module.h>
29 #include <linux/sort.h>
30 #include <linux/sched/mm.h>
31 #include <linux/sched/signal.h>
32 #include <linux/interval_tree_generic.h>
33 #include <linux/nospec.h>
37 static ushort max_mem_regions = 64;
38 module_param(max_mem_regions, ushort, 0444);
39 MODULE_PARM_DESC(max_mem_regions,
40 "Maximum number of memory regions in memory map. (default: 64)");
41 static int max_iotlb_entries = 2048;
42 module_param(max_iotlb_entries, int, 0444);
43 MODULE_PARM_DESC(max_iotlb_entries,
44 "Maximum number of iotlb entries. (default: 2048)");
47 VHOST_MEMORY_F_LOG = 0x1,
50 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
51 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
53 INTERVAL_TREE_DEFINE(struct vhost_umem_node,
54 rb, __u64, __subtree_last,
55 START, LAST, static inline, vhost_umem_interval_tree);
57 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
58 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
60 vq->user_be = !virtio_legacy_is_little_endian();
63 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
68 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
73 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
75 struct vhost_vring_state s;
80 if (copy_from_user(&s, argp, sizeof(s)))
83 if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
84 s.num != VHOST_VRING_BIG_ENDIAN)
87 if (s.num == VHOST_VRING_BIG_ENDIAN)
88 vhost_enable_cross_endian_big(vq);
90 vhost_enable_cross_endian_little(vq);
95 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
98 struct vhost_vring_state s = {
103 if (copy_to_user(argp, &s, sizeof(s)))
109 static void vhost_init_is_le(struct vhost_virtqueue *vq)
111 /* Note for legacy virtio: user_be is initialized at reset time
112 * according to the host endianness. If userspace does not set an
113 * explicit endianness, the default behavior is native endian, as
114 * expected by legacy virtio.
116 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
119 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
123 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
128 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
134 static void vhost_init_is_le(struct vhost_virtqueue *vq)
136 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
137 || virtio_legacy_is_little_endian();
139 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
141 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
143 vhost_init_is_le(vq);
146 struct vhost_flush_struct {
147 struct vhost_work work;
148 struct completion wait_event;
151 static void vhost_flush_work(struct vhost_work *work)
153 struct vhost_flush_struct *s;
155 s = container_of(work, struct vhost_flush_struct, work);
156 complete(&s->wait_event);
159 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
162 struct vhost_poll *poll;
164 poll = container_of(pt, struct vhost_poll, table);
166 add_wait_queue(wqh, &poll->wait);
169 static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
172 struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
174 if (!(key_to_poll(key) & poll->mask))
177 vhost_poll_queue(poll);
181 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
183 clear_bit(VHOST_WORK_QUEUED, &work->flags);
186 EXPORT_SYMBOL_GPL(vhost_work_init);
188 /* Init poll structure */
189 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
190 __poll_t mask, struct vhost_dev *dev)
192 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
193 init_poll_funcptr(&poll->table, vhost_poll_func);
198 vhost_work_init(&poll->work, fn);
200 EXPORT_SYMBOL_GPL(vhost_poll_init);
202 /* Start polling a file. We add ourselves to file's wait queue. The caller must
203 * keep a reference to a file until after vhost_poll_stop is called. */
204 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
212 mask = vfs_poll(file, &poll->table);
214 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
215 if (mask & EPOLLERR) {
216 vhost_poll_stop(poll);
222 EXPORT_SYMBOL_GPL(vhost_poll_start);
224 /* Stop polling a file. After this function returns, it becomes safe to drop the
225 * file reference. You must also flush afterwards. */
226 void vhost_poll_stop(struct vhost_poll *poll)
229 remove_wait_queue(poll->wqh, &poll->wait);
233 EXPORT_SYMBOL_GPL(vhost_poll_stop);
235 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
237 struct vhost_flush_struct flush;
240 init_completion(&flush.wait_event);
241 vhost_work_init(&flush.work, vhost_flush_work);
243 vhost_work_queue(dev, &flush.work);
244 wait_for_completion(&flush.wait_event);
247 EXPORT_SYMBOL_GPL(vhost_work_flush);
249 /* Flush any work that has been scheduled. When calling this, don't hold any
250 * locks that are also used by the callback. */
251 void vhost_poll_flush(struct vhost_poll *poll)
253 vhost_work_flush(poll->dev, &poll->work);
255 EXPORT_SYMBOL_GPL(vhost_poll_flush);
257 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
262 if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
263 /* We can only add the work to the list after we're
264 * sure it was not in the list.
265 * test_and_set_bit() implies a memory barrier.
267 llist_add(&work->node, &dev->work_list);
268 wake_up_process(dev->worker);
271 EXPORT_SYMBOL_GPL(vhost_work_queue);
273 /* A lockless hint for busy polling code to exit the loop */
274 bool vhost_has_work(struct vhost_dev *dev)
276 return !llist_empty(&dev->work_list);
278 EXPORT_SYMBOL_GPL(vhost_has_work);
280 void vhost_poll_queue(struct vhost_poll *poll)
282 vhost_work_queue(poll->dev, &poll->work);
284 EXPORT_SYMBOL_GPL(vhost_poll_queue);
286 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
290 for (j = 0; j < VHOST_NUM_ADDRS; j++)
291 vq->meta_iotlb[j] = NULL;
294 static void vhost_vq_meta_reset(struct vhost_dev *d)
298 for (i = 0; i < d->nvqs; ++i)
299 __vhost_vq_meta_reset(d->vqs[i]);
302 static void vhost_vq_reset(struct vhost_dev *dev,
303 struct vhost_virtqueue *vq)
309 vq->last_avail_idx = 0;
311 vq->last_used_idx = 0;
312 vq->signalled_used = 0;
313 vq->signalled_used_valid = false;
315 vq->log_used = false;
316 vq->log_addr = -1ull;
317 vq->private_data = NULL;
318 vq->acked_features = 0;
319 vq->acked_backend_features = 0;
321 vq->error_ctx = NULL;
325 vhost_reset_is_le(vq);
326 vhost_disable_cross_endian(vq);
327 vq->busyloop_timeout = 0;
330 __vhost_vq_meta_reset(vq);
333 static int vhost_worker(void *data)
335 struct vhost_dev *dev = data;
336 struct vhost_work *work, *work_next;
337 struct llist_node *node;
338 mm_segment_t oldfs = get_fs();
344 /* mb paired w/ kthread_stop */
345 set_current_state(TASK_INTERRUPTIBLE);
347 if (kthread_should_stop()) {
348 __set_current_state(TASK_RUNNING);
352 node = llist_del_all(&dev->work_list);
356 node = llist_reverse_order(node);
357 /* make sure flag is seen after deletion */
359 llist_for_each_entry_safe(work, work_next, node, node) {
360 clear_bit(VHOST_WORK_QUEUED, &work->flags);
361 __set_current_state(TASK_RUNNING);
372 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
382 /* Helper to allocate iovec buffers for all vqs. */
383 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
385 struct vhost_virtqueue *vq;
388 for (i = 0; i < dev->nvqs; ++i) {
390 vq->indirect = kmalloc_array(UIO_MAXIOV,
391 sizeof(*vq->indirect),
393 vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
395 vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
397 if (!vq->indirect || !vq->log || !vq->heads)
404 vhost_vq_free_iovecs(dev->vqs[i]);
408 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
412 for (i = 0; i < dev->nvqs; ++i)
413 vhost_vq_free_iovecs(dev->vqs[i]);
416 void vhost_dev_init(struct vhost_dev *dev,
417 struct vhost_virtqueue **vqs, int nvqs, int iov_limit)
419 struct vhost_virtqueue *vq;
424 mutex_init(&dev->mutex);
430 dev->iov_limit = iov_limit;
431 init_llist_head(&dev->work_list);
432 init_waitqueue_head(&dev->wait);
433 INIT_LIST_HEAD(&dev->read_list);
434 INIT_LIST_HEAD(&dev->pending_list);
435 spin_lock_init(&dev->iotlb_lock);
438 for (i = 0; i < dev->nvqs; ++i) {
444 mutex_init(&vq->mutex);
445 vhost_vq_reset(dev, vq);
447 vhost_poll_init(&vq->poll, vq->handle_kick,
451 EXPORT_SYMBOL_GPL(vhost_dev_init);
453 /* Caller should have device mutex */
454 long vhost_dev_check_owner(struct vhost_dev *dev)
456 /* Are you the owner? If not, I don't think you mean to do that */
457 return dev->mm == current->mm ? 0 : -EPERM;
459 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
461 struct vhost_attach_cgroups_struct {
462 struct vhost_work work;
463 struct task_struct *owner;
467 static void vhost_attach_cgroups_work(struct vhost_work *work)
469 struct vhost_attach_cgroups_struct *s;
471 s = container_of(work, struct vhost_attach_cgroups_struct, work);
472 s->ret = cgroup_attach_task_all(s->owner, current);
475 static int vhost_attach_cgroups(struct vhost_dev *dev)
477 struct vhost_attach_cgroups_struct attach;
479 attach.owner = current;
480 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
481 vhost_work_queue(dev, &attach.work);
482 vhost_work_flush(dev, &attach.work);
486 /* Caller should have device mutex */
487 bool vhost_dev_has_owner(struct vhost_dev *dev)
491 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
493 /* Caller should have device mutex */
494 long vhost_dev_set_owner(struct vhost_dev *dev)
496 struct task_struct *worker;
499 /* Is there an owner already? */
500 if (vhost_dev_has_owner(dev)) {
505 /* No owner, become one */
506 dev->mm = get_task_mm(current);
507 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
508 if (IS_ERR(worker)) {
509 err = PTR_ERR(worker);
513 dev->worker = worker;
514 wake_up_process(worker); /* avoid contributing to loadavg */
516 err = vhost_attach_cgroups(dev);
520 err = vhost_dev_alloc_iovecs(dev);
526 kthread_stop(worker);
535 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
537 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
539 return kvzalloc(sizeof(struct vhost_umem), GFP_KERNEL);
541 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
543 /* Caller should have device mutex */
544 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
548 vhost_dev_cleanup(dev);
550 /* Restore memory to default empty mapping. */
551 INIT_LIST_HEAD(&umem->umem_list);
553 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
554 * VQs aren't running.
556 for (i = 0; i < dev->nvqs; ++i)
557 dev->vqs[i]->umem = umem;
559 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
561 void vhost_dev_stop(struct vhost_dev *dev)
565 for (i = 0; i < dev->nvqs; ++i) {
566 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
567 vhost_poll_stop(&dev->vqs[i]->poll);
568 vhost_poll_flush(&dev->vqs[i]->poll);
572 EXPORT_SYMBOL_GPL(vhost_dev_stop);
574 static void vhost_umem_free(struct vhost_umem *umem,
575 struct vhost_umem_node *node)
577 vhost_umem_interval_tree_remove(node, &umem->umem_tree);
578 list_del(&node->link);
583 static void vhost_umem_clean(struct vhost_umem *umem)
585 struct vhost_umem_node *node, *tmp;
590 list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
591 vhost_umem_free(umem, node);
596 static void vhost_clear_msg(struct vhost_dev *dev)
598 struct vhost_msg_node *node, *n;
600 spin_lock(&dev->iotlb_lock);
602 list_for_each_entry_safe(node, n, &dev->read_list, node) {
603 list_del(&node->node);
607 list_for_each_entry_safe(node, n, &dev->pending_list, node) {
608 list_del(&node->node);
612 spin_unlock(&dev->iotlb_lock);
615 void vhost_dev_cleanup(struct vhost_dev *dev)
619 for (i = 0; i < dev->nvqs; ++i) {
620 if (dev->vqs[i]->error_ctx)
621 eventfd_ctx_put(dev->vqs[i]->error_ctx);
622 if (dev->vqs[i]->kick)
623 fput(dev->vqs[i]->kick);
624 if (dev->vqs[i]->call_ctx)
625 eventfd_ctx_put(dev->vqs[i]->call_ctx);
626 vhost_vq_reset(dev, dev->vqs[i]);
628 vhost_dev_free_iovecs(dev);
630 eventfd_ctx_put(dev->log_ctx);
632 /* No one will access memory at this point */
633 vhost_umem_clean(dev->umem);
635 vhost_umem_clean(dev->iotlb);
637 vhost_clear_msg(dev);
638 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
639 WARN_ON(!llist_empty(&dev->work_list));
641 kthread_stop(dev->worker);
648 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
650 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
652 u64 a = addr / VHOST_PAGE_SIZE / 8;
654 /* Make sure 64 bit math will not overflow. */
655 if (a > ULONG_MAX - (unsigned long)log_base ||
656 a + (unsigned long)log_base > ULONG_MAX)
659 return access_ok(log_base + a,
660 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
663 static bool vhost_overflow(u64 uaddr, u64 size)
665 /* Make sure 64 bit math will not overflow. */
666 return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
669 /* Caller should have vq mutex and device mutex. */
670 static bool vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
673 struct vhost_umem_node *node;
678 list_for_each_entry(node, &umem->umem_list, link) {
679 unsigned long a = node->userspace_addr;
681 if (vhost_overflow(node->userspace_addr, node->size))
685 if (!access_ok((void __user *)a,
688 else if (log_all && !log_access_ok(log_base,
696 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
697 u64 addr, unsigned int size,
700 const struct vhost_umem_node *node = vq->meta_iotlb[type];
705 return (void *)(uintptr_t)(node->userspace_addr + addr - node->start);
708 /* Can we switch to this memory table? */
709 /* Caller should have device mutex but not vq mutex */
710 static bool memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
715 for (i = 0; i < d->nvqs; ++i) {
719 mutex_lock(&d->vqs[i]->mutex);
720 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
721 /* If ring is inactive, will check when it's enabled. */
722 if (d->vqs[i]->private_data)
723 ok = vq_memory_access_ok(d->vqs[i]->log_base,
727 mutex_unlock(&d->vqs[i]->mutex);
734 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
735 struct iovec iov[], int iov_size, int access);
737 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
738 const void *from, unsigned size)
743 return __copy_to_user(to, from, size);
745 /* This function should be called after iotlb
746 * prefetch, which means we're sure that all vq
747 * could be access through iotlb. So -EAGAIN should
748 * not happen in this case.
751 void __user *uaddr = vhost_vq_meta_fetch(vq,
752 (u64)(uintptr_t)to, size,
756 return __copy_to_user(uaddr, from, size);
758 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
759 ARRAY_SIZE(vq->iotlb_iov),
763 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
764 ret = copy_to_iter(from, size, &t);
772 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
773 void __user *from, unsigned size)
778 return __copy_from_user(to, from, size);
780 /* This function should be called after iotlb
781 * prefetch, which means we're sure that vq
782 * could be access through iotlb. So -EAGAIN should
783 * not happen in this case.
785 void __user *uaddr = vhost_vq_meta_fetch(vq,
786 (u64)(uintptr_t)from, size,
791 return __copy_from_user(to, uaddr, size);
793 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
794 ARRAY_SIZE(vq->iotlb_iov),
797 vq_err(vq, "IOTLB translation failure: uaddr "
798 "%p size 0x%llx\n", from,
799 (unsigned long long) size);
802 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
803 ret = copy_from_iter(to, size, &f);
812 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
813 void __user *addr, unsigned int size,
818 ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
819 ARRAY_SIZE(vq->iotlb_iov),
822 vq_err(vq, "IOTLB translation failure: uaddr "
823 "%p size 0x%llx\n", addr,
824 (unsigned long long) size);
828 if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
829 vq_err(vq, "Non atomic userspace memory access: uaddr "
830 "%p size 0x%llx\n", addr,
831 (unsigned long long) size);
835 return vq->iotlb_iov[0].iov_base;
838 /* This function should be called after iotlb
839 * prefetch, which means we're sure that vq
840 * could be access through iotlb. So -EAGAIN should
841 * not happen in this case.
843 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
844 void *addr, unsigned int size,
847 void __user *uaddr = vhost_vq_meta_fetch(vq,
848 (u64)(uintptr_t)addr, size, type);
852 return __vhost_get_user_slow(vq, addr, size, type);
855 #define vhost_put_user(vq, x, ptr) \
859 ret = __put_user(x, ptr); \
861 __typeof__(ptr) to = \
862 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
863 sizeof(*ptr), VHOST_ADDR_USED); \
865 ret = __put_user(x, to); \
872 #define vhost_get_user(vq, x, ptr, type) \
876 ret = __get_user(x, ptr); \
878 __typeof__(ptr) from = \
879 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
883 ret = __get_user(x, from); \
890 #define vhost_get_avail(vq, x, ptr) \
891 vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
893 #define vhost_get_used(vq, x, ptr) \
894 vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
896 static void vhost_dev_lock_vqs(struct vhost_dev *d)
899 for (i = 0; i < d->nvqs; ++i)
900 mutex_lock_nested(&d->vqs[i]->mutex, i);
903 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
906 for (i = 0; i < d->nvqs; ++i)
907 mutex_unlock(&d->vqs[i]->mutex);
910 static int vhost_new_umem_range(struct vhost_umem *umem,
911 u64 start, u64 size, u64 end,
912 u64 userspace_addr, int perm)
914 struct vhost_umem_node *tmp, *node;
919 node = kmalloc(sizeof(*node), GFP_ATOMIC);
923 if (umem->numem == max_iotlb_entries) {
924 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
925 vhost_umem_free(umem, tmp);
931 node->userspace_addr = userspace_addr;
933 INIT_LIST_HEAD(&node->link);
934 list_add_tail(&node->link, &umem->umem_list);
935 vhost_umem_interval_tree_insert(node, &umem->umem_tree);
941 static void vhost_del_umem_range(struct vhost_umem *umem,
944 struct vhost_umem_node *node;
946 while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
948 vhost_umem_free(umem, node);
951 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
952 struct vhost_iotlb_msg *msg)
954 struct vhost_msg_node *node, *n;
956 spin_lock(&d->iotlb_lock);
958 list_for_each_entry_safe(node, n, &d->pending_list, node) {
959 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
960 if (msg->iova <= vq_msg->iova &&
961 msg->iova + msg->size - 1 >= vq_msg->iova &&
962 vq_msg->type == VHOST_IOTLB_MISS) {
963 vhost_poll_queue(&node->vq->poll);
964 list_del(&node->node);
969 spin_unlock(&d->iotlb_lock);
972 static bool umem_access_ok(u64 uaddr, u64 size, int access)
974 unsigned long a = uaddr;
976 /* Make sure 64 bit math will not overflow. */
977 if (vhost_overflow(uaddr, size))
980 if ((access & VHOST_ACCESS_RO) &&
981 !access_ok((void __user *)a, size))
983 if ((access & VHOST_ACCESS_WO) &&
984 !access_ok((void __user *)a, size))
989 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
990 struct vhost_iotlb_msg *msg)
994 mutex_lock(&dev->mutex);
995 vhost_dev_lock_vqs(dev);
997 case VHOST_IOTLB_UPDATE:
1002 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1006 vhost_vq_meta_reset(dev);
1007 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
1008 msg->iova + msg->size - 1,
1009 msg->uaddr, msg->perm)) {
1013 vhost_iotlb_notify_vq(dev, msg);
1015 case VHOST_IOTLB_INVALIDATE:
1020 vhost_vq_meta_reset(dev);
1021 vhost_del_umem_range(dev->iotlb, msg->iova,
1022 msg->iova + msg->size - 1);
1029 vhost_dev_unlock_vqs(dev);
1030 mutex_unlock(&dev->mutex);
1034 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1035 struct iov_iter *from)
1037 struct vhost_iotlb_msg msg;
1041 ret = copy_from_iter(&type, sizeof(type), from);
1042 if (ret != sizeof(type)) {
1048 case VHOST_IOTLB_MSG:
1049 /* There maybe a hole after type for V1 message type,
1052 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1054 case VHOST_IOTLB_MSG_V2:
1055 offset = sizeof(__u32);
1062 iov_iter_advance(from, offset);
1063 ret = copy_from_iter(&msg, sizeof(msg), from);
1064 if (ret != sizeof(msg)) {
1068 if (vhost_process_iotlb_msg(dev, &msg)) {
1073 ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1074 sizeof(struct vhost_msg_v2);
1078 EXPORT_SYMBOL(vhost_chr_write_iter);
1080 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1085 poll_wait(file, &dev->wait, wait);
1087 if (!list_empty(&dev->read_list))
1088 mask |= EPOLLIN | EPOLLRDNORM;
1092 EXPORT_SYMBOL(vhost_chr_poll);
1094 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1098 struct vhost_msg_node *node;
1100 unsigned size = sizeof(struct vhost_msg);
1102 if (iov_iter_count(to) < size)
1107 prepare_to_wait(&dev->wait, &wait,
1108 TASK_INTERRUPTIBLE);
1110 node = vhost_dequeue_msg(dev, &dev->read_list);
1117 if (signal_pending(current)) {
1130 finish_wait(&dev->wait, &wait);
1133 struct vhost_iotlb_msg *msg;
1134 void *start = &node->msg;
1136 switch (node->msg.type) {
1137 case VHOST_IOTLB_MSG:
1138 size = sizeof(node->msg);
1139 msg = &node->msg.iotlb;
1141 case VHOST_IOTLB_MSG_V2:
1142 size = sizeof(node->msg_v2);
1143 msg = &node->msg_v2.iotlb;
1150 ret = copy_to_iter(start, size, to);
1151 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1155 vhost_enqueue_msg(dev, &dev->pending_list, node);
1160 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1162 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1164 struct vhost_dev *dev = vq->dev;
1165 struct vhost_msg_node *node;
1166 struct vhost_iotlb_msg *msg;
1167 bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1169 node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1174 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1175 msg = &node->msg_v2.iotlb;
1177 msg = &node->msg.iotlb;
1180 msg->type = VHOST_IOTLB_MISS;
1184 vhost_enqueue_msg(dev, &dev->read_list, node);
1189 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1190 struct vring_desc __user *desc,
1191 struct vring_avail __user *avail,
1192 struct vring_used __user *used)
1195 size_t s __maybe_unused = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1197 return access_ok(desc, num * sizeof *desc) &&
1199 sizeof *avail + num * sizeof *avail->ring + s) &&
1201 sizeof *used + num * sizeof *used->ring + s);
1204 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1205 const struct vhost_umem_node *node,
1208 int access = (type == VHOST_ADDR_USED) ?
1209 VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1211 if (likely(node->perm & access))
1212 vq->meta_iotlb[type] = node;
1215 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1216 int access, u64 addr, u64 len, int type)
1218 const struct vhost_umem_node *node;
1219 struct vhost_umem *umem = vq->iotlb;
1220 u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1222 if (vhost_vq_meta_fetch(vq, addr, len, type))
1226 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1229 if (node == NULL || node->start > addr) {
1230 vhost_iotlb_miss(vq, addr, access);
1232 } else if (!(node->perm & access)) {
1233 /* Report the possible access violation by
1234 * request another translation from userspace.
1239 size = node->size - addr + node->start;
1241 if (orig_addr == addr && size >= len)
1242 vhost_vq_meta_update(vq, node, type);
1251 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1253 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1254 unsigned int num = vq->num;
1259 return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1260 num * sizeof(*vq->desc), VHOST_ADDR_DESC) &&
1261 iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1263 num * sizeof(*vq->avail->ring) + s,
1264 VHOST_ADDR_AVAIL) &&
1265 iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1267 num * sizeof(*vq->used->ring) + s,
1270 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1272 /* Can we log writes? */
1273 /* Caller should have device mutex but not vq mutex */
1274 bool vhost_log_access_ok(struct vhost_dev *dev)
1276 return memory_access_ok(dev, dev->umem, 1);
1278 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1280 /* Verify access for write logging. */
1281 /* Caller should have vq mutex and device mutex */
1282 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1283 void __user *log_base)
1285 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1287 return vq_memory_access_ok(log_base, vq->umem,
1288 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1289 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1291 vq->num * sizeof *vq->used->ring + s));
1294 /* Can we start vq? */
1295 /* Caller should have vq mutex and device mutex */
1296 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1298 if (!vq_log_access_ok(vq, vq->log_base))
1301 /* Access validation occurs at prefetch time with IOTLB */
1305 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1307 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1309 static struct vhost_umem *vhost_umem_alloc(void)
1311 struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1316 umem->umem_tree = RB_ROOT_CACHED;
1318 INIT_LIST_HEAD(&umem->umem_list);
1323 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1325 struct vhost_memory mem, *newmem;
1326 struct vhost_memory_region *region;
1327 struct vhost_umem *newumem, *oldumem;
1328 unsigned long size = offsetof(struct vhost_memory, regions);
1331 if (copy_from_user(&mem, m, size))
1335 if (mem.nregions > max_mem_regions)
1337 newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1342 memcpy(newmem, &mem, size);
1343 if (copy_from_user(newmem->regions, m->regions,
1344 mem.nregions * sizeof *m->regions)) {
1349 newumem = vhost_umem_alloc();
1355 for (region = newmem->regions;
1356 region < newmem->regions + mem.nregions;
1358 if (vhost_new_umem_range(newumem,
1359 region->guest_phys_addr,
1360 region->memory_size,
1361 region->guest_phys_addr +
1362 region->memory_size - 1,
1363 region->userspace_addr,
1368 if (!memory_access_ok(d, newumem, 0))
1374 /* All memory accesses are done under some VQ mutex. */
1375 for (i = 0; i < d->nvqs; ++i) {
1376 mutex_lock(&d->vqs[i]->mutex);
1377 d->vqs[i]->umem = newumem;
1378 mutex_unlock(&d->vqs[i]->mutex);
1382 vhost_umem_clean(oldumem);
1386 vhost_umem_clean(newumem);
1391 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1393 struct file *eventfp, *filep = NULL;
1394 bool pollstart = false, pollstop = false;
1395 struct eventfd_ctx *ctx = NULL;
1396 u32 __user *idxp = argp;
1397 struct vhost_virtqueue *vq;
1398 struct vhost_vring_state s;
1399 struct vhost_vring_file f;
1400 struct vhost_vring_addr a;
1404 r = get_user(idx, idxp);
1410 idx = array_index_nospec(idx, d->nvqs);
1413 mutex_lock(&vq->mutex);
1416 case VHOST_SET_VRING_NUM:
1417 /* Resizing ring with an active backend?
1418 * You don't want to do that. */
1419 if (vq->private_data) {
1423 if (copy_from_user(&s, argp, sizeof s)) {
1427 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1433 case VHOST_SET_VRING_BASE:
1434 /* Moving base with an active backend?
1435 * You don't want to do that. */
1436 if (vq->private_data) {
1440 if (copy_from_user(&s, argp, sizeof s)) {
1444 if (s.num > 0xffff) {
1448 vq->last_avail_idx = s.num;
1449 /* Forget the cached index value. */
1450 vq->avail_idx = vq->last_avail_idx;
1452 case VHOST_GET_VRING_BASE:
1454 s.num = vq->last_avail_idx;
1455 if (copy_to_user(argp, &s, sizeof s))
1458 case VHOST_SET_VRING_ADDR:
1459 if (copy_from_user(&a, argp, sizeof a)) {
1463 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1467 /* For 32bit, verify that the top 32bits of the user
1468 data are set to zero. */
1469 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1470 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1471 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1476 /* Make sure it's safe to cast pointers to vring types. */
1477 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1478 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1479 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1480 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1481 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1486 /* We only verify access here if backend is configured.
1487 * If it is not, we don't as size might not have been setup.
1488 * We will verify when backend is configured. */
1489 if (vq->private_data) {
1490 if (!vq_access_ok(vq, vq->num,
1491 (void __user *)(unsigned long)a.desc_user_addr,
1492 (void __user *)(unsigned long)a.avail_user_addr,
1493 (void __user *)(unsigned long)a.used_user_addr)) {
1498 /* Also validate log access for used ring if enabled. */
1499 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1500 !log_access_ok(vq->log_base, a.log_guest_addr,
1502 vq->num * sizeof *vq->used->ring)) {
1508 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1509 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1510 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1511 vq->log_addr = a.log_guest_addr;
1512 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1514 case VHOST_SET_VRING_KICK:
1515 if (copy_from_user(&f, argp, sizeof f)) {
1519 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1520 if (IS_ERR(eventfp)) {
1521 r = PTR_ERR(eventfp);
1524 if (eventfp != vq->kick) {
1525 pollstop = (filep = vq->kick) != NULL;
1526 pollstart = (vq->kick = eventfp) != NULL;
1530 case VHOST_SET_VRING_CALL:
1531 if (copy_from_user(&f, argp, sizeof f)) {
1535 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1540 swap(ctx, vq->call_ctx);
1542 case VHOST_SET_VRING_ERR:
1543 if (copy_from_user(&f, argp, sizeof f)) {
1547 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1552 swap(ctx, vq->error_ctx);
1554 case VHOST_SET_VRING_ENDIAN:
1555 r = vhost_set_vring_endian(vq, argp);
1557 case VHOST_GET_VRING_ENDIAN:
1558 r = vhost_get_vring_endian(vq, idx, argp);
1560 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1561 if (copy_from_user(&s, argp, sizeof(s))) {
1565 vq->busyloop_timeout = s.num;
1567 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1569 s.num = vq->busyloop_timeout;
1570 if (copy_to_user(argp, &s, sizeof(s)))
1577 if (pollstop && vq->handle_kick)
1578 vhost_poll_stop(&vq->poll);
1580 if (!IS_ERR_OR_NULL(ctx))
1581 eventfd_ctx_put(ctx);
1585 if (pollstart && vq->handle_kick)
1586 r = vhost_poll_start(&vq->poll, vq->kick);
1588 mutex_unlock(&vq->mutex);
1590 if (pollstop && vq->handle_kick)
1591 vhost_poll_flush(&vq->poll);
1594 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1596 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1598 struct vhost_umem *niotlb, *oiotlb;
1601 niotlb = vhost_umem_alloc();
1608 for (i = 0; i < d->nvqs; ++i) {
1609 struct vhost_virtqueue *vq = d->vqs[i];
1611 mutex_lock(&vq->mutex);
1613 __vhost_vq_meta_reset(vq);
1614 mutex_unlock(&vq->mutex);
1617 vhost_umem_clean(oiotlb);
1621 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1623 /* Caller must have device mutex */
1624 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1626 struct eventfd_ctx *ctx;
1631 /* If you are not the owner, you can become one */
1632 if (ioctl == VHOST_SET_OWNER) {
1633 r = vhost_dev_set_owner(d);
1637 /* You must be the owner to do anything else */
1638 r = vhost_dev_check_owner(d);
1643 case VHOST_SET_MEM_TABLE:
1644 r = vhost_set_memory(d, argp);
1646 case VHOST_SET_LOG_BASE:
1647 if (copy_from_user(&p, argp, sizeof p)) {
1651 if ((u64)(unsigned long)p != p) {
1655 for (i = 0; i < d->nvqs; ++i) {
1656 struct vhost_virtqueue *vq;
1657 void __user *base = (void __user *)(unsigned long)p;
1659 mutex_lock(&vq->mutex);
1660 /* If ring is inactive, will check when it's enabled. */
1661 if (vq->private_data && !vq_log_access_ok(vq, base))
1664 vq->log_base = base;
1665 mutex_unlock(&vq->mutex);
1668 case VHOST_SET_LOG_FD:
1669 r = get_user(fd, (int __user *)argp);
1672 ctx = fd == -1 ? NULL : eventfd_ctx_fdget(fd);
1677 swap(ctx, d->log_ctx);
1678 for (i = 0; i < d->nvqs; ++i) {
1679 mutex_lock(&d->vqs[i]->mutex);
1680 d->vqs[i]->log_ctx = d->log_ctx;
1681 mutex_unlock(&d->vqs[i]->mutex);
1684 eventfd_ctx_put(ctx);
1693 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1695 /* TODO: This is really inefficient. We need something like get_user()
1696 * (instruction directly accesses the data, with an exception table entry
1697 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1699 static int set_bit_to_user(int nr, void __user *addr)
1701 unsigned long log = (unsigned long)addr;
1704 int bit = nr + (log % PAGE_SIZE) * 8;
1707 r = get_user_pages_fast(log, 1, 1, &page);
1711 base = kmap_atomic(page);
1713 kunmap_atomic(base);
1714 set_page_dirty_lock(page);
1719 static int log_write(void __user *log_base,
1720 u64 write_address, u64 write_length)
1722 u64 write_page = write_address / VHOST_PAGE_SIZE;
1727 write_length += write_address % VHOST_PAGE_SIZE;
1729 u64 base = (u64)(unsigned long)log_base;
1730 u64 log = base + write_page / 8;
1731 int bit = write_page % 8;
1732 if ((u64)(unsigned long)log != log)
1734 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1737 if (write_length <= VHOST_PAGE_SIZE)
1739 write_length -= VHOST_PAGE_SIZE;
1745 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1747 struct vhost_umem *umem = vq->umem;
1748 struct vhost_umem_node *u;
1749 u64 start, end, l, min;
1755 /* More than one GPAs can be mapped into a single HVA. So
1756 * iterate all possible umems here to be safe.
1758 list_for_each_entry(u, &umem->umem_list, link) {
1759 if (u->userspace_addr > hva - 1 + len ||
1760 u->userspace_addr - 1 + u->size < hva)
1762 start = max(u->userspace_addr, hva);
1763 end = min(u->userspace_addr - 1 + u->size,
1765 l = end - start + 1;
1766 r = log_write(vq->log_base,
1767 u->start + start - u->userspace_addr,
1785 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1787 struct iovec iov[64];
1791 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1793 ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1794 len, iov, 64, VHOST_ACCESS_WO);
1798 for (i = 0; i < ret; i++) {
1799 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1808 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1809 unsigned int log_num, u64 len, struct iovec *iov, int count)
1813 /* Make sure data written is seen before log. */
1817 for (i = 0; i < count; i++) {
1818 r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1826 for (i = 0; i < log_num; ++i) {
1827 u64 l = min(log[i].len, len);
1828 r = log_write(vq->log_base, log[i].addr, l);
1834 eventfd_signal(vq->log_ctx, 1);
1838 /* Length written exceeds what we have stored. This is a bug. */
1842 EXPORT_SYMBOL_GPL(vhost_log_write);
1844 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1847 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1848 &vq->used->flags) < 0)
1850 if (unlikely(vq->log_used)) {
1851 /* Make sure the flag is seen before log. */
1853 /* Log used flag write. */
1854 used = &vq->used->flags;
1855 log_used(vq, (used - (void __user *)vq->used),
1856 sizeof vq->used->flags);
1858 eventfd_signal(vq->log_ctx, 1);
1863 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1865 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1866 vhost_avail_event(vq)))
1868 if (unlikely(vq->log_used)) {
1870 /* Make sure the event is seen before log. */
1872 /* Log avail event write */
1873 used = vhost_avail_event(vq);
1874 log_used(vq, (used - (void __user *)vq->used),
1875 sizeof *vhost_avail_event(vq));
1877 eventfd_signal(vq->log_ctx, 1);
1882 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1884 __virtio16 last_used_idx;
1886 bool is_le = vq->is_le;
1888 if (!vq->private_data)
1891 vhost_init_is_le(vq);
1893 r = vhost_update_used_flags(vq);
1896 vq->signalled_used_valid = false;
1898 !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
1902 r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1904 vq_err(vq, "Can't access used idx at %p\n",
1908 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1915 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1917 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1918 struct iovec iov[], int iov_size, int access)
1920 const struct vhost_umem_node *node;
1921 struct vhost_dev *dev = vq->dev;
1922 struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1927 while ((u64)len > s) {
1929 if (unlikely(ret >= iov_size)) {
1934 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1935 addr, addr + len - 1);
1936 if (node == NULL || node->start > addr) {
1937 if (umem != dev->iotlb) {
1943 } else if (!(node->perm & access)) {
1949 size = node->size - addr + node->start;
1950 _iov->iov_len = min((u64)len - s, size);
1951 _iov->iov_base = (void __user *)(unsigned long)
1952 (node->userspace_addr + addr - node->start);
1959 vhost_iotlb_miss(vq, addr, access);
1963 /* Each buffer in the virtqueues is actually a chain of descriptors. This
1964 * function returns the next descriptor in the chain,
1965 * or -1U if we're at the end. */
1966 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1970 /* If this descriptor says it doesn't chain, we're done. */
1971 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1974 /* Check they're not leading us off end of descriptors. */
1975 next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
1979 static int get_indirect(struct vhost_virtqueue *vq,
1980 struct iovec iov[], unsigned int iov_size,
1981 unsigned int *out_num, unsigned int *in_num,
1982 struct vhost_log *log, unsigned int *log_num,
1983 struct vring_desc *indirect)
1985 struct vring_desc desc;
1986 unsigned int i = 0, count, found = 0;
1987 u32 len = vhost32_to_cpu(vq, indirect->len);
1988 struct iov_iter from;
1992 if (unlikely(len % sizeof desc)) {
1993 vq_err(vq, "Invalid length in indirect descriptor: "
1994 "len 0x%llx not multiple of 0x%zx\n",
1995 (unsigned long long)len,
2000 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2001 UIO_MAXIOV, VHOST_ACCESS_RO);
2002 if (unlikely(ret < 0)) {
2004 vq_err(vq, "Translation failure %d in indirect.\n", ret);
2007 iov_iter_init(&from, READ, vq->indirect, ret, len);
2009 /* We will use the result as an address to read from, so most
2010 * architectures only need a compiler barrier here. */
2011 read_barrier_depends();
2013 count = len / sizeof desc;
2014 /* Buffers are chained via a 16 bit next field, so
2015 * we can have at most 2^16 of these. */
2016 if (unlikely(count > USHRT_MAX + 1)) {
2017 vq_err(vq, "Indirect buffer length too big: %d\n",
2023 unsigned iov_count = *in_num + *out_num;
2024 if (unlikely(++found > count)) {
2025 vq_err(vq, "Loop detected: last one at %u "
2026 "indirect size %u\n",
2030 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2031 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2032 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2035 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2036 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2037 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2041 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2042 access = VHOST_ACCESS_WO;
2044 access = VHOST_ACCESS_RO;
2046 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2047 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2048 iov_size - iov_count, access);
2049 if (unlikely(ret < 0)) {
2051 vq_err(vq, "Translation failure %d indirect idx %d\n",
2055 /* If this is an input descriptor, increment that count. */
2056 if (access == VHOST_ACCESS_WO) {
2058 if (unlikely(log)) {
2059 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2060 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2064 /* If it's an output descriptor, they're all supposed
2065 * to come before any input descriptors. */
2066 if (unlikely(*in_num)) {
2067 vq_err(vq, "Indirect descriptor "
2068 "has out after in: idx %d\n", i);
2073 } while ((i = next_desc(vq, &desc)) != -1);
2077 /* This looks in the virtqueue and for the first available buffer, and converts
2078 * it to an iovec for convenient access. Since descriptors consist of some
2079 * number of output then some number of input descriptors, it's actually two
2080 * iovecs, but we pack them into one and note how many of each there were.
2082 * This function returns the descriptor number found, or vq->num (which is
2083 * never a valid descriptor number) if none was found. A negative code is
2084 * returned on error. */
2085 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2086 struct iovec iov[], unsigned int iov_size,
2087 unsigned int *out_num, unsigned int *in_num,
2088 struct vhost_log *log, unsigned int *log_num)
2090 struct vring_desc desc;
2091 unsigned int i, head, found = 0;
2093 __virtio16 avail_idx;
2094 __virtio16 ring_head;
2097 /* Check it isn't doing very strange things with descriptor numbers. */
2098 last_avail_idx = vq->last_avail_idx;
2100 if (vq->avail_idx == vq->last_avail_idx) {
2101 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
2102 vq_err(vq, "Failed to access avail idx at %p\n",
2106 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2108 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2109 vq_err(vq, "Guest moved used index from %u to %u",
2110 last_avail_idx, vq->avail_idx);
2114 /* If there's nothing new since last we looked, return
2117 if (vq->avail_idx == last_avail_idx)
2120 /* Only get avail ring entries after they have been
2126 /* Grab the next descriptor number they're advertising, and increment
2127 * the index we've seen. */
2128 if (unlikely(vhost_get_avail(vq, ring_head,
2129 &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2130 vq_err(vq, "Failed to read head: idx %d address %p\n",
2132 &vq->avail->ring[last_avail_idx % vq->num]);
2136 head = vhost16_to_cpu(vq, ring_head);
2138 /* If their number is silly, that's an error. */
2139 if (unlikely(head >= vq->num)) {
2140 vq_err(vq, "Guest says index %u > %u is available",
2145 /* When we start there are none of either input nor output. */
2146 *out_num = *in_num = 0;
2152 unsigned iov_count = *in_num + *out_num;
2153 if (unlikely(i >= vq->num)) {
2154 vq_err(vq, "Desc index is %u > %u, head = %u",
2158 if (unlikely(++found > vq->num)) {
2159 vq_err(vq, "Loop detected: last one at %u "
2160 "vq size %u head %u\n",
2164 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2166 if (unlikely(ret)) {
2167 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2171 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2172 ret = get_indirect(vq, iov, iov_size,
2174 log, log_num, &desc);
2175 if (unlikely(ret < 0)) {
2177 vq_err(vq, "Failure detected "
2178 "in indirect descriptor at idx %d\n", i);
2184 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2185 access = VHOST_ACCESS_WO;
2187 access = VHOST_ACCESS_RO;
2188 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2189 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2190 iov_size - iov_count, access);
2191 if (unlikely(ret < 0)) {
2193 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2197 if (access == VHOST_ACCESS_WO) {
2198 /* If this is an input descriptor,
2199 * increment that count. */
2201 if (unlikely(log)) {
2202 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2203 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2207 /* If it's an output descriptor, they're all supposed
2208 * to come before any input descriptors. */
2209 if (unlikely(*in_num)) {
2210 vq_err(vq, "Descriptor has out after in: "
2216 } while ((i = next_desc(vq, &desc)) != -1);
2218 /* On success, increment avail index. */
2219 vq->last_avail_idx++;
2221 /* Assume notifications from guest are disabled at this point,
2222 * if they aren't we would need to update avail_event index. */
2223 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2226 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2228 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2229 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2231 vq->last_avail_idx -= n;
2233 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2235 /* After we've used one of their buffers, we tell them about it. We'll then
2236 * want to notify the guest, using eventfd. */
2237 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2239 struct vring_used_elem heads = {
2240 cpu_to_vhost32(vq, head),
2241 cpu_to_vhost32(vq, len)
2244 return vhost_add_used_n(vq, &heads, 1);
2246 EXPORT_SYMBOL_GPL(vhost_add_used);
2248 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2249 struct vring_used_elem *heads,
2252 struct vring_used_elem __user *used;
2256 start = vq->last_used_idx & (vq->num - 1);
2257 used = vq->used->ring + start;
2259 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2260 vq_err(vq, "Failed to write used id");
2263 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2264 vq_err(vq, "Failed to write used len");
2267 } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2268 vq_err(vq, "Failed to write used");
2271 if (unlikely(vq->log_used)) {
2272 /* Make sure data is seen before log. */
2274 /* Log used ring entry write. */
2275 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2276 count * sizeof *used);
2278 old = vq->last_used_idx;
2279 new = (vq->last_used_idx += count);
2280 /* If the driver never bothers to signal in a very long while,
2281 * used index might wrap around. If that happens, invalidate
2282 * signalled_used index we stored. TODO: make sure driver
2283 * signals at least once in 2^16 and remove this. */
2284 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2285 vq->signalled_used_valid = false;
2289 /* After we've used one of their buffers, we tell them about it. We'll then
2290 * want to notify the guest, using eventfd. */
2291 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2296 start = vq->last_used_idx & (vq->num - 1);
2297 n = vq->num - start;
2299 r = __vhost_add_used_n(vq, heads, n);
2305 r = __vhost_add_used_n(vq, heads, count);
2307 /* Make sure buffer is written before we update index. */
2309 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2311 vq_err(vq, "Failed to increment used idx");
2314 if (unlikely(vq->log_used)) {
2315 /* Make sure used idx is seen before log. */
2317 /* Log used index update. */
2318 log_used(vq, offsetof(struct vring_used, idx),
2319 sizeof vq->used->idx);
2321 eventfd_signal(vq->log_ctx, 1);
2325 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2327 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2332 /* Flush out used index updates. This is paired
2333 * with the barrier that the Guest executes when enabling
2337 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2338 unlikely(vq->avail_idx == vq->last_avail_idx))
2341 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2343 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2344 vq_err(vq, "Failed to get flags");
2347 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2349 old = vq->signalled_used;
2350 v = vq->signalled_used_valid;
2351 new = vq->signalled_used = vq->last_used_idx;
2352 vq->signalled_used_valid = true;
2357 if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2358 vq_err(vq, "Failed to get used event idx");
2361 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2364 /* This actually signals the guest, using eventfd. */
2365 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2367 /* Signal the Guest tell them we used something up. */
2368 if (vq->call_ctx && vhost_notify(dev, vq))
2369 eventfd_signal(vq->call_ctx, 1);
2371 EXPORT_SYMBOL_GPL(vhost_signal);
2373 /* And here's the combo meal deal. Supersize me! */
2374 void vhost_add_used_and_signal(struct vhost_dev *dev,
2375 struct vhost_virtqueue *vq,
2376 unsigned int head, int len)
2378 vhost_add_used(vq, head, len);
2379 vhost_signal(dev, vq);
2381 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2383 /* multi-buffer version of vhost_add_used_and_signal */
2384 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2385 struct vhost_virtqueue *vq,
2386 struct vring_used_elem *heads, unsigned count)
2388 vhost_add_used_n(vq, heads, count);
2389 vhost_signal(dev, vq);
2391 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2393 /* return true if we're sure that avaiable ring is empty */
2394 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2396 __virtio16 avail_idx;
2399 if (vq->avail_idx != vq->last_avail_idx)
2402 r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2405 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2407 return vq->avail_idx == vq->last_avail_idx;
2409 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2411 /* OK, now we need to know about added descriptors. */
2412 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2414 __virtio16 avail_idx;
2417 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2419 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2420 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2421 r = vhost_update_used_flags(vq);
2423 vq_err(vq, "Failed to enable notification at %p: %d\n",
2424 &vq->used->flags, r);
2428 r = vhost_update_avail_event(vq, vq->avail_idx);
2430 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2431 vhost_avail_event(vq), r);
2435 /* They could have slipped one in as we were doing that: make
2436 * sure it's written, then check again. */
2438 r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2440 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2441 &vq->avail->idx, r);
2445 return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2447 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2449 /* We don't need to be notified again. */
2450 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2454 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2456 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2457 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2458 r = vhost_update_used_flags(vq);
2460 vq_err(vq, "Failed to enable notification at %p: %d\n",
2461 &vq->used->flags, r);
2464 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2466 /* Create a new message. */
2467 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2469 struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2473 /* Make sure all padding within the structure is initialized. */
2474 memset(&node->msg, 0, sizeof node->msg);
2476 node->msg.type = type;
2479 EXPORT_SYMBOL_GPL(vhost_new_msg);
2481 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2482 struct vhost_msg_node *node)
2484 spin_lock(&dev->iotlb_lock);
2485 list_add_tail(&node->node, head);
2486 spin_unlock(&dev->iotlb_lock);
2488 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2490 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2492 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2493 struct list_head *head)
2495 struct vhost_msg_node *node = NULL;
2497 spin_lock(&dev->iotlb_lock);
2498 if (!list_empty(head)) {
2499 node = list_first_entry(head, struct vhost_msg_node,
2501 list_del(&node->node);
2503 spin_unlock(&dev->iotlb_lock);
2507 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2510 static int __init vhost_init(void)
2515 static void __exit vhost_exit(void)
2519 module_init(vhost_init);
2520 module_exit(vhost_exit);
2522 MODULE_VERSION("0.0.1");
2523 MODULE_LICENSE("GPL v2");
2524 MODULE_AUTHOR("Michael S. Tsirkin");
2525 MODULE_DESCRIPTION("Host kernel accelerator for virtio");