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>
36 static ushort max_mem_regions = 64;
37 module_param(max_mem_regions, ushort, 0444);
38 MODULE_PARM_DESC(max_mem_regions,
39 "Maximum number of memory regions in memory map. (default: 64)");
40 static int max_iotlb_entries = 2048;
41 module_param(max_iotlb_entries, int, 0444);
42 MODULE_PARM_DESC(max_iotlb_entries,
43 "Maximum number of iotlb entries. (default: 2048)");
46 VHOST_MEMORY_F_LOG = 0x1,
49 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
50 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
52 INTERVAL_TREE_DEFINE(struct vhost_umem_node,
53 rb, __u64, __subtree_last,
54 START, LAST, static inline, vhost_umem_interval_tree);
56 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
57 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
59 vq->user_be = !virtio_legacy_is_little_endian();
62 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
67 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
72 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
74 struct vhost_vring_state s;
79 if (copy_from_user(&s, argp, sizeof(s)))
82 if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
83 s.num != VHOST_VRING_BIG_ENDIAN)
86 if (s.num == VHOST_VRING_BIG_ENDIAN)
87 vhost_enable_cross_endian_big(vq);
89 vhost_enable_cross_endian_little(vq);
94 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
97 struct vhost_vring_state s = {
102 if (copy_to_user(argp, &s, sizeof(s)))
108 static void vhost_init_is_le(struct vhost_virtqueue *vq)
110 /* Note for legacy virtio: user_be is initialized at reset time
111 * according to the host endianness. If userspace does not set an
112 * explicit endianness, the default behavior is native endian, as
113 * expected by legacy virtio.
115 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
118 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
122 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
127 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
133 static void vhost_init_is_le(struct vhost_virtqueue *vq)
135 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
136 || virtio_legacy_is_little_endian();
138 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
140 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
142 vhost_init_is_le(vq);
145 struct vhost_flush_struct {
146 struct vhost_work work;
147 struct completion wait_event;
150 static void vhost_flush_work(struct vhost_work *work)
152 struct vhost_flush_struct *s;
154 s = container_of(work, struct vhost_flush_struct, work);
155 complete(&s->wait_event);
158 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
161 struct vhost_poll *poll;
163 poll = container_of(pt, struct vhost_poll, table);
165 add_wait_queue(wqh, &poll->wait);
168 static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
171 struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
173 if (!(key_to_poll(key) & poll->mask))
176 vhost_poll_queue(poll);
180 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
182 clear_bit(VHOST_WORK_QUEUED, &work->flags);
185 EXPORT_SYMBOL_GPL(vhost_work_init);
187 /* Init poll structure */
188 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
189 __poll_t mask, struct vhost_dev *dev)
191 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
192 init_poll_funcptr(&poll->table, vhost_poll_func);
197 vhost_work_init(&poll->work, fn);
199 EXPORT_SYMBOL_GPL(vhost_poll_init);
201 /* Start polling a file. We add ourselves to file's wait queue. The caller must
202 * keep a reference to a file until after vhost_poll_stop is called. */
203 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
211 mask = vfs_poll(file, &poll->table);
213 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
214 if (mask & EPOLLERR) {
215 vhost_poll_stop(poll);
221 EXPORT_SYMBOL_GPL(vhost_poll_start);
223 /* Stop polling a file. After this function returns, it becomes safe to drop the
224 * file reference. You must also flush afterwards. */
225 void vhost_poll_stop(struct vhost_poll *poll)
228 remove_wait_queue(poll->wqh, &poll->wait);
232 EXPORT_SYMBOL_GPL(vhost_poll_stop);
234 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
236 struct vhost_flush_struct flush;
239 init_completion(&flush.wait_event);
240 vhost_work_init(&flush.work, vhost_flush_work);
242 vhost_work_queue(dev, &flush.work);
243 wait_for_completion(&flush.wait_event);
246 EXPORT_SYMBOL_GPL(vhost_work_flush);
248 /* Flush any work that has been scheduled. When calling this, don't hold any
249 * locks that are also used by the callback. */
250 void vhost_poll_flush(struct vhost_poll *poll)
252 vhost_work_flush(poll->dev, &poll->work);
254 EXPORT_SYMBOL_GPL(vhost_poll_flush);
256 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
261 if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
262 /* We can only add the work to the list after we're
263 * sure it was not in the list.
264 * test_and_set_bit() implies a memory barrier.
266 llist_add(&work->node, &dev->work_list);
267 wake_up_process(dev->worker);
270 EXPORT_SYMBOL_GPL(vhost_work_queue);
272 /* A lockless hint for busy polling code to exit the loop */
273 bool vhost_has_work(struct vhost_dev *dev)
275 return !llist_empty(&dev->work_list);
277 EXPORT_SYMBOL_GPL(vhost_has_work);
279 void vhost_poll_queue(struct vhost_poll *poll)
281 vhost_work_queue(poll->dev, &poll->work);
283 EXPORT_SYMBOL_GPL(vhost_poll_queue);
285 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
289 for (j = 0; j < VHOST_NUM_ADDRS; j++)
290 vq->meta_iotlb[j] = NULL;
293 static void vhost_vq_meta_reset(struct vhost_dev *d)
297 for (i = 0; i < d->nvqs; ++i) {
298 mutex_lock(&d->vqs[i]->mutex);
299 __vhost_vq_meta_reset(d->vqs[i]);
300 mutex_unlock(&d->vqs[i]->mutex);
304 static void vhost_vq_reset(struct vhost_dev *dev,
305 struct vhost_virtqueue *vq)
311 vq->last_avail_idx = 0;
313 vq->last_used_idx = 0;
314 vq->signalled_used = 0;
315 vq->signalled_used_valid = false;
317 vq->log_used = false;
318 vq->log_addr = -1ull;
319 vq->private_data = NULL;
320 vq->acked_features = 0;
321 vq->acked_backend_features = 0;
323 vq->error_ctx = NULL;
327 vhost_reset_is_le(vq);
328 vhost_disable_cross_endian(vq);
329 vq->busyloop_timeout = 0;
332 __vhost_vq_meta_reset(vq);
335 static int vhost_worker(void *data)
337 struct vhost_dev *dev = data;
338 struct vhost_work *work, *work_next;
339 struct llist_node *node;
340 mm_segment_t oldfs = get_fs();
346 /* mb paired w/ kthread_stop */
347 set_current_state(TASK_INTERRUPTIBLE);
349 if (kthread_should_stop()) {
350 __set_current_state(TASK_RUNNING);
354 node = llist_del_all(&dev->work_list);
358 node = llist_reverse_order(node);
359 /* make sure flag is seen after deletion */
361 llist_for_each_entry_safe(work, work_next, node, node) {
362 clear_bit(VHOST_WORK_QUEUED, &work->flags);
363 __set_current_state(TASK_RUNNING);
374 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
384 /* Helper to allocate iovec buffers for all vqs. */
385 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
387 struct vhost_virtqueue *vq;
390 for (i = 0; i < dev->nvqs; ++i) {
392 vq->indirect = kmalloc_array(UIO_MAXIOV,
393 sizeof(*vq->indirect),
395 vq->log = kmalloc_array(UIO_MAXIOV, sizeof(*vq->log),
397 vq->heads = kmalloc_array(UIO_MAXIOV, sizeof(*vq->heads),
399 if (!vq->indirect || !vq->log || !vq->heads)
406 vhost_vq_free_iovecs(dev->vqs[i]);
410 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
414 for (i = 0; i < dev->nvqs; ++i)
415 vhost_vq_free_iovecs(dev->vqs[i]);
418 void vhost_dev_init(struct vhost_dev *dev,
419 struct vhost_virtqueue **vqs, int nvqs)
421 struct vhost_virtqueue *vq;
426 mutex_init(&dev->mutex);
432 init_llist_head(&dev->work_list);
433 init_waitqueue_head(&dev->wait);
434 INIT_LIST_HEAD(&dev->read_list);
435 INIT_LIST_HEAD(&dev->pending_list);
436 spin_lock_init(&dev->iotlb_lock);
439 for (i = 0; i < dev->nvqs; ++i) {
445 mutex_init(&vq->mutex);
446 vhost_vq_reset(dev, vq);
448 vhost_poll_init(&vq->poll, vq->handle_kick,
452 EXPORT_SYMBOL_GPL(vhost_dev_init);
454 /* Caller should have device mutex */
455 long vhost_dev_check_owner(struct vhost_dev *dev)
457 /* Are you the owner? If not, I don't think you mean to do that */
458 return dev->mm == current->mm ? 0 : -EPERM;
460 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
462 struct vhost_attach_cgroups_struct {
463 struct vhost_work work;
464 struct task_struct *owner;
468 static void vhost_attach_cgroups_work(struct vhost_work *work)
470 struct vhost_attach_cgroups_struct *s;
472 s = container_of(work, struct vhost_attach_cgroups_struct, work);
473 s->ret = cgroup_attach_task_all(s->owner, current);
476 static int vhost_attach_cgroups(struct vhost_dev *dev)
478 struct vhost_attach_cgroups_struct attach;
480 attach.owner = current;
481 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
482 vhost_work_queue(dev, &attach.work);
483 vhost_work_flush(dev, &attach.work);
487 /* Caller should have device mutex */
488 bool vhost_dev_has_owner(struct vhost_dev *dev)
492 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
494 /* Caller should have device mutex */
495 long vhost_dev_set_owner(struct vhost_dev *dev)
497 struct task_struct *worker;
500 /* Is there an owner already? */
501 if (vhost_dev_has_owner(dev)) {
506 /* No owner, become one */
507 dev->mm = get_task_mm(current);
508 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
509 if (IS_ERR(worker)) {
510 err = PTR_ERR(worker);
514 dev->worker = worker;
515 wake_up_process(worker); /* avoid contributing to loadavg */
517 err = vhost_attach_cgroups(dev);
521 err = vhost_dev_alloc_iovecs(dev);
527 kthread_stop(worker);
536 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
538 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
540 return kvzalloc(sizeof(struct vhost_umem), GFP_KERNEL);
542 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
544 /* Caller should have device mutex */
545 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
549 vhost_dev_cleanup(dev);
551 /* Restore memory to default empty mapping. */
552 INIT_LIST_HEAD(&umem->umem_list);
554 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
555 * VQs aren't running.
557 for (i = 0; i < dev->nvqs; ++i)
558 dev->vqs[i]->umem = umem;
560 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
562 void vhost_dev_stop(struct vhost_dev *dev)
566 for (i = 0; i < dev->nvqs; ++i) {
567 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
568 vhost_poll_stop(&dev->vqs[i]->poll);
569 vhost_poll_flush(&dev->vqs[i]->poll);
573 EXPORT_SYMBOL_GPL(vhost_dev_stop);
575 static void vhost_umem_free(struct vhost_umem *umem,
576 struct vhost_umem_node *node)
578 vhost_umem_interval_tree_remove(node, &umem->umem_tree);
579 list_del(&node->link);
584 static void vhost_umem_clean(struct vhost_umem *umem)
586 struct vhost_umem_node *node, *tmp;
591 list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
592 vhost_umem_free(umem, node);
597 static void vhost_clear_msg(struct vhost_dev *dev)
599 struct vhost_msg_node *node, *n;
601 spin_lock(&dev->iotlb_lock);
603 list_for_each_entry_safe(node, n, &dev->read_list, node) {
604 list_del(&node->node);
608 list_for_each_entry_safe(node, n, &dev->pending_list, node) {
609 list_del(&node->node);
613 spin_unlock(&dev->iotlb_lock);
616 void vhost_dev_cleanup(struct vhost_dev *dev)
620 for (i = 0; i < dev->nvqs; ++i) {
621 if (dev->vqs[i]->error_ctx)
622 eventfd_ctx_put(dev->vqs[i]->error_ctx);
623 if (dev->vqs[i]->kick)
624 fput(dev->vqs[i]->kick);
625 if (dev->vqs[i]->call_ctx)
626 eventfd_ctx_put(dev->vqs[i]->call_ctx);
627 vhost_vq_reset(dev, dev->vqs[i]);
629 vhost_dev_free_iovecs(dev);
631 eventfd_ctx_put(dev->log_ctx);
633 /* No one will access memory at this point */
634 vhost_umem_clean(dev->umem);
636 vhost_umem_clean(dev->iotlb);
638 vhost_clear_msg(dev);
639 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
640 WARN_ON(!llist_empty(&dev->work_list));
642 kthread_stop(dev->worker);
649 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
651 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
653 u64 a = addr / VHOST_PAGE_SIZE / 8;
655 /* Make sure 64 bit math will not overflow. */
656 if (a > ULONG_MAX - (unsigned long)log_base ||
657 a + (unsigned long)log_base > ULONG_MAX)
660 return access_ok(VERIFY_WRITE, log_base + a,
661 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
664 static bool vhost_overflow(u64 uaddr, u64 size)
666 /* Make sure 64 bit math will not overflow. */
667 return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
670 /* Caller should have vq mutex and device mutex. */
671 static bool vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
674 struct vhost_umem_node *node;
679 list_for_each_entry(node, &umem->umem_list, link) {
680 unsigned long a = node->userspace_addr;
682 if (vhost_overflow(node->userspace_addr, node->size))
686 if (!access_ok(VERIFY_WRITE, (void __user *)a,
689 else if (log_all && !log_access_ok(log_base,
697 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
698 u64 addr, unsigned int size,
701 const struct vhost_umem_node *node = vq->meta_iotlb[type];
706 return (void *)(uintptr_t)(node->userspace_addr + addr - node->start);
709 /* Can we switch to this memory table? */
710 /* Caller should have device mutex but not vq mutex */
711 static bool memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
716 for (i = 0; i < d->nvqs; ++i) {
720 mutex_lock(&d->vqs[i]->mutex);
721 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
722 /* If ring is inactive, will check when it's enabled. */
723 if (d->vqs[i]->private_data)
724 ok = vq_memory_access_ok(d->vqs[i]->log_base,
728 mutex_unlock(&d->vqs[i]->mutex);
735 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
736 struct iovec iov[], int iov_size, int access);
738 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
739 const void *from, unsigned size)
744 return __copy_to_user(to, from, size);
746 /* This function should be called after iotlb
747 * prefetch, which means we're sure that all vq
748 * could be access through iotlb. So -EAGAIN should
749 * not happen in this case.
752 void __user *uaddr = vhost_vq_meta_fetch(vq,
753 (u64)(uintptr_t)to, size,
757 return __copy_to_user(uaddr, from, size);
759 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
760 ARRAY_SIZE(vq->iotlb_iov),
764 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
765 ret = copy_to_iter(from, size, &t);
773 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
774 void __user *from, unsigned size)
779 return __copy_from_user(to, from, size);
781 /* This function should be called after iotlb
782 * prefetch, which means we're sure that vq
783 * could be access through iotlb. So -EAGAIN should
784 * not happen in this case.
786 void __user *uaddr = vhost_vq_meta_fetch(vq,
787 (u64)(uintptr_t)from, size,
792 return __copy_from_user(to, uaddr, size);
794 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
795 ARRAY_SIZE(vq->iotlb_iov),
798 vq_err(vq, "IOTLB translation failure: uaddr "
799 "%p size 0x%llx\n", from,
800 (unsigned long long) size);
803 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
804 ret = copy_from_iter(to, size, &f);
813 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
814 void __user *addr, unsigned int size,
819 ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
820 ARRAY_SIZE(vq->iotlb_iov),
823 vq_err(vq, "IOTLB translation failure: uaddr "
824 "%p size 0x%llx\n", addr,
825 (unsigned long long) size);
829 if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
830 vq_err(vq, "Non atomic userspace memory access: uaddr "
831 "%p size 0x%llx\n", addr,
832 (unsigned long long) size);
836 return vq->iotlb_iov[0].iov_base;
839 /* This function should be called after iotlb
840 * prefetch, which means we're sure that vq
841 * could be access through iotlb. So -EAGAIN should
842 * not happen in this case.
844 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
845 void *addr, unsigned int size,
848 void __user *uaddr = vhost_vq_meta_fetch(vq,
849 (u64)(uintptr_t)addr, size, type);
853 return __vhost_get_user_slow(vq, addr, size, type);
856 #define vhost_put_user(vq, x, ptr) \
860 ret = __put_user(x, ptr); \
862 __typeof__(ptr) to = \
863 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
864 sizeof(*ptr), VHOST_ADDR_USED); \
866 ret = __put_user(x, to); \
873 #define vhost_get_user(vq, x, ptr, type) \
877 ret = __get_user(x, ptr); \
879 __typeof__(ptr) from = \
880 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
884 ret = __get_user(x, from); \
891 #define vhost_get_avail(vq, x, ptr) \
892 vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
894 #define vhost_get_used(vq, x, ptr) \
895 vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
897 static int vhost_new_umem_range(struct vhost_umem *umem,
898 u64 start, u64 size, u64 end,
899 u64 userspace_addr, int perm)
901 struct vhost_umem_node *tmp, *node = kmalloc(sizeof(*node), GFP_ATOMIC);
906 if (umem->numem == max_iotlb_entries) {
907 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
908 vhost_umem_free(umem, tmp);
914 node->userspace_addr = userspace_addr;
916 INIT_LIST_HEAD(&node->link);
917 list_add_tail(&node->link, &umem->umem_list);
918 vhost_umem_interval_tree_insert(node, &umem->umem_tree);
924 static void vhost_del_umem_range(struct vhost_umem *umem,
927 struct vhost_umem_node *node;
929 while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
931 vhost_umem_free(umem, node);
934 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
935 struct vhost_iotlb_msg *msg)
937 struct vhost_msg_node *node, *n;
939 spin_lock(&d->iotlb_lock);
941 list_for_each_entry_safe(node, n, &d->pending_list, node) {
942 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
943 if (msg->iova <= vq_msg->iova &&
944 msg->iova + msg->size - 1 >= vq_msg->iova &&
945 vq_msg->type == VHOST_IOTLB_MISS) {
946 mutex_lock(&node->vq->mutex);
947 vhost_poll_queue(&node->vq->poll);
948 mutex_unlock(&node->vq->mutex);
950 list_del(&node->node);
955 spin_unlock(&d->iotlb_lock);
958 static bool umem_access_ok(u64 uaddr, u64 size, int access)
960 unsigned long a = uaddr;
962 /* Make sure 64 bit math will not overflow. */
963 if (vhost_overflow(uaddr, size))
966 if ((access & VHOST_ACCESS_RO) &&
967 !access_ok(VERIFY_READ, (void __user *)a, size))
969 if ((access & VHOST_ACCESS_WO) &&
970 !access_ok(VERIFY_WRITE, (void __user *)a, size))
975 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
976 struct vhost_iotlb_msg *msg)
980 mutex_lock(&dev->mutex);
982 case VHOST_IOTLB_UPDATE:
987 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
991 vhost_vq_meta_reset(dev);
992 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
993 msg->iova + msg->size - 1,
994 msg->uaddr, msg->perm)) {
998 vhost_iotlb_notify_vq(dev, msg);
1000 case VHOST_IOTLB_INVALIDATE:
1005 vhost_vq_meta_reset(dev);
1006 vhost_del_umem_range(dev->iotlb, msg->iova,
1007 msg->iova + msg->size - 1);
1014 mutex_unlock(&dev->mutex);
1018 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1019 struct iov_iter *from)
1021 struct vhost_iotlb_msg msg;
1025 ret = copy_from_iter(&type, sizeof(type), from);
1026 if (ret != sizeof(type))
1030 case VHOST_IOTLB_MSG:
1031 /* There maybe a hole after type for V1 message type,
1034 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1036 case VHOST_IOTLB_MSG_V2:
1037 offset = sizeof(__u32);
1044 iov_iter_advance(from, offset);
1045 ret = copy_from_iter(&msg, sizeof(msg), from);
1046 if (ret != sizeof(msg))
1048 if (vhost_process_iotlb_msg(dev, &msg)) {
1053 ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1054 sizeof(struct vhost_msg_v2);
1058 EXPORT_SYMBOL(vhost_chr_write_iter);
1060 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1065 poll_wait(file, &dev->wait, wait);
1067 if (!list_empty(&dev->read_list))
1068 mask |= EPOLLIN | EPOLLRDNORM;
1072 EXPORT_SYMBOL(vhost_chr_poll);
1074 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1078 struct vhost_msg_node *node;
1080 unsigned size = sizeof(struct vhost_msg);
1082 if (iov_iter_count(to) < size)
1087 prepare_to_wait(&dev->wait, &wait,
1088 TASK_INTERRUPTIBLE);
1090 node = vhost_dequeue_msg(dev, &dev->read_list);
1097 if (signal_pending(current)) {
1110 finish_wait(&dev->wait, &wait);
1113 struct vhost_iotlb_msg *msg;
1114 void *start = &node->msg;
1116 switch (node->msg.type) {
1117 case VHOST_IOTLB_MSG:
1118 size = sizeof(node->msg);
1119 msg = &node->msg.iotlb;
1121 case VHOST_IOTLB_MSG_V2:
1122 size = sizeof(node->msg_v2);
1123 msg = &node->msg_v2.iotlb;
1130 ret = copy_to_iter(start, size, to);
1131 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1135 vhost_enqueue_msg(dev, &dev->pending_list, node);
1140 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1142 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1144 struct vhost_dev *dev = vq->dev;
1145 struct vhost_msg_node *node;
1146 struct vhost_iotlb_msg *msg;
1147 bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1149 node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1154 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1155 msg = &node->msg_v2.iotlb;
1157 msg = &node->msg.iotlb;
1160 msg->type = VHOST_IOTLB_MISS;
1164 vhost_enqueue_msg(dev, &dev->read_list, node);
1169 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1170 struct vring_desc __user *desc,
1171 struct vring_avail __user *avail,
1172 struct vring_used __user *used)
1175 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1177 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
1178 access_ok(VERIFY_READ, avail,
1179 sizeof *avail + num * sizeof *avail->ring + s) &&
1180 access_ok(VERIFY_WRITE, used,
1181 sizeof *used + num * sizeof *used->ring + s);
1184 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1185 const struct vhost_umem_node *node,
1188 int access = (type == VHOST_ADDR_USED) ?
1189 VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1191 if (likely(node->perm & access))
1192 vq->meta_iotlb[type] = node;
1195 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1196 int access, u64 addr, u64 len, int type)
1198 const struct vhost_umem_node *node;
1199 struct vhost_umem *umem = vq->iotlb;
1200 u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1202 if (vhost_vq_meta_fetch(vq, addr, len, type))
1206 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1209 if (node == NULL || node->start > addr) {
1210 vhost_iotlb_miss(vq, addr, access);
1212 } else if (!(node->perm & access)) {
1213 /* Report the possible access violation by
1214 * request another translation from userspace.
1219 size = node->size - addr + node->start;
1221 if (orig_addr == addr && size >= len)
1222 vhost_vq_meta_update(vq, node, type);
1231 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1233 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1234 unsigned int num = vq->num;
1239 return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1240 num * sizeof(*vq->desc), VHOST_ADDR_DESC) &&
1241 iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1243 num * sizeof(*vq->avail->ring) + s,
1244 VHOST_ADDR_AVAIL) &&
1245 iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1247 num * sizeof(*vq->used->ring) + s,
1250 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1252 /* Can we log writes? */
1253 /* Caller should have device mutex but not vq mutex */
1254 bool vhost_log_access_ok(struct vhost_dev *dev)
1256 return memory_access_ok(dev, dev->umem, 1);
1258 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1260 /* Verify access for write logging. */
1261 /* Caller should have vq mutex and device mutex */
1262 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1263 void __user *log_base)
1265 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1267 return vq_memory_access_ok(log_base, vq->umem,
1268 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1269 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1271 vq->num * sizeof *vq->used->ring + s));
1274 /* Can we start vq? */
1275 /* Caller should have vq mutex and device mutex */
1276 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1278 if (!vq_log_access_ok(vq, vq->log_base))
1281 /* Access validation occurs at prefetch time with IOTLB */
1285 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1287 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1289 static struct vhost_umem *vhost_umem_alloc(void)
1291 struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1296 umem->umem_tree = RB_ROOT_CACHED;
1298 INIT_LIST_HEAD(&umem->umem_list);
1303 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1305 struct vhost_memory mem, *newmem;
1306 struct vhost_memory_region *region;
1307 struct vhost_umem *newumem, *oldumem;
1308 unsigned long size = offsetof(struct vhost_memory, regions);
1311 if (copy_from_user(&mem, m, size))
1315 if (mem.nregions > max_mem_regions)
1317 newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1322 memcpy(newmem, &mem, size);
1323 if (copy_from_user(newmem->regions, m->regions,
1324 mem.nregions * sizeof *m->regions)) {
1329 newumem = vhost_umem_alloc();
1335 for (region = newmem->regions;
1336 region < newmem->regions + mem.nregions;
1338 if (vhost_new_umem_range(newumem,
1339 region->guest_phys_addr,
1340 region->memory_size,
1341 region->guest_phys_addr +
1342 region->memory_size - 1,
1343 region->userspace_addr,
1348 if (!memory_access_ok(d, newumem, 0))
1354 /* All memory accesses are done under some VQ mutex. */
1355 for (i = 0; i < d->nvqs; ++i) {
1356 mutex_lock(&d->vqs[i]->mutex);
1357 d->vqs[i]->umem = newumem;
1358 mutex_unlock(&d->vqs[i]->mutex);
1362 vhost_umem_clean(oldumem);
1366 vhost_umem_clean(newumem);
1371 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1373 struct file *eventfp, *filep = NULL;
1374 bool pollstart = false, pollstop = false;
1375 struct eventfd_ctx *ctx = NULL;
1376 u32 __user *idxp = argp;
1377 struct vhost_virtqueue *vq;
1378 struct vhost_vring_state s;
1379 struct vhost_vring_file f;
1380 struct vhost_vring_addr a;
1384 r = get_user(idx, idxp);
1392 mutex_lock(&vq->mutex);
1395 case VHOST_SET_VRING_NUM:
1396 /* Resizing ring with an active backend?
1397 * You don't want to do that. */
1398 if (vq->private_data) {
1402 if (copy_from_user(&s, argp, sizeof s)) {
1406 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1412 case VHOST_SET_VRING_BASE:
1413 /* Moving base with an active backend?
1414 * You don't want to do that. */
1415 if (vq->private_data) {
1419 if (copy_from_user(&s, argp, sizeof s)) {
1423 if (s.num > 0xffff) {
1427 vq->last_avail_idx = s.num;
1428 /* Forget the cached index value. */
1429 vq->avail_idx = vq->last_avail_idx;
1431 case VHOST_GET_VRING_BASE:
1433 s.num = vq->last_avail_idx;
1434 if (copy_to_user(argp, &s, sizeof s))
1437 case VHOST_SET_VRING_ADDR:
1438 if (copy_from_user(&a, argp, sizeof a)) {
1442 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1446 /* For 32bit, verify that the top 32bits of the user
1447 data are set to zero. */
1448 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1449 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1450 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1455 /* Make sure it's safe to cast pointers to vring types. */
1456 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1457 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1458 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1459 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1460 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1465 /* We only verify access here if backend is configured.
1466 * If it is not, we don't as size might not have been setup.
1467 * We will verify when backend is configured. */
1468 if (vq->private_data) {
1469 if (!vq_access_ok(vq, vq->num,
1470 (void __user *)(unsigned long)a.desc_user_addr,
1471 (void __user *)(unsigned long)a.avail_user_addr,
1472 (void __user *)(unsigned long)a.used_user_addr)) {
1477 /* Also validate log access for used ring if enabled. */
1478 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1479 !log_access_ok(vq->log_base, a.log_guest_addr,
1481 vq->num * sizeof *vq->used->ring)) {
1487 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1488 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1489 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1490 vq->log_addr = a.log_guest_addr;
1491 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1493 case VHOST_SET_VRING_KICK:
1494 if (copy_from_user(&f, argp, sizeof f)) {
1498 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1499 if (IS_ERR(eventfp)) {
1500 r = PTR_ERR(eventfp);
1503 if (eventfp != vq->kick) {
1504 pollstop = (filep = vq->kick) != NULL;
1505 pollstart = (vq->kick = eventfp) != NULL;
1509 case VHOST_SET_VRING_CALL:
1510 if (copy_from_user(&f, argp, sizeof f)) {
1514 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1519 swap(ctx, vq->call_ctx);
1521 case VHOST_SET_VRING_ERR:
1522 if (copy_from_user(&f, argp, sizeof f)) {
1526 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1531 swap(ctx, vq->error_ctx);
1533 case VHOST_SET_VRING_ENDIAN:
1534 r = vhost_set_vring_endian(vq, argp);
1536 case VHOST_GET_VRING_ENDIAN:
1537 r = vhost_get_vring_endian(vq, idx, argp);
1539 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1540 if (copy_from_user(&s, argp, sizeof(s))) {
1544 vq->busyloop_timeout = s.num;
1546 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1548 s.num = vq->busyloop_timeout;
1549 if (copy_to_user(argp, &s, sizeof(s)))
1556 if (pollstop && vq->handle_kick)
1557 vhost_poll_stop(&vq->poll);
1559 if (!IS_ERR_OR_NULL(ctx))
1560 eventfd_ctx_put(ctx);
1564 if (pollstart && vq->handle_kick)
1565 r = vhost_poll_start(&vq->poll, vq->kick);
1567 mutex_unlock(&vq->mutex);
1569 if (pollstop && vq->handle_kick)
1570 vhost_poll_flush(&vq->poll);
1573 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1575 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1577 struct vhost_umem *niotlb, *oiotlb;
1580 niotlb = vhost_umem_alloc();
1587 for (i = 0; i < d->nvqs; ++i) {
1588 struct vhost_virtqueue *vq = d->vqs[i];
1590 mutex_lock(&vq->mutex);
1592 __vhost_vq_meta_reset(vq);
1593 mutex_unlock(&vq->mutex);
1596 vhost_umem_clean(oiotlb);
1600 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1602 /* Caller must have device mutex */
1603 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1605 struct eventfd_ctx *ctx;
1610 /* If you are not the owner, you can become one */
1611 if (ioctl == VHOST_SET_OWNER) {
1612 r = vhost_dev_set_owner(d);
1616 /* You must be the owner to do anything else */
1617 r = vhost_dev_check_owner(d);
1622 case VHOST_SET_MEM_TABLE:
1623 r = vhost_set_memory(d, argp);
1625 case VHOST_SET_LOG_BASE:
1626 if (copy_from_user(&p, argp, sizeof p)) {
1630 if ((u64)(unsigned long)p != p) {
1634 for (i = 0; i < d->nvqs; ++i) {
1635 struct vhost_virtqueue *vq;
1636 void __user *base = (void __user *)(unsigned long)p;
1638 mutex_lock(&vq->mutex);
1639 /* If ring is inactive, will check when it's enabled. */
1640 if (vq->private_data && !vq_log_access_ok(vq, base))
1643 vq->log_base = base;
1644 mutex_unlock(&vq->mutex);
1647 case VHOST_SET_LOG_FD:
1648 r = get_user(fd, (int __user *)argp);
1651 ctx = fd == -1 ? NULL : eventfd_ctx_fdget(fd);
1656 swap(ctx, d->log_ctx);
1657 for (i = 0; i < d->nvqs; ++i) {
1658 mutex_lock(&d->vqs[i]->mutex);
1659 d->vqs[i]->log_ctx = d->log_ctx;
1660 mutex_unlock(&d->vqs[i]->mutex);
1663 eventfd_ctx_put(ctx);
1672 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1674 /* TODO: This is really inefficient. We need something like get_user()
1675 * (instruction directly accesses the data, with an exception table entry
1676 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1678 static int set_bit_to_user(int nr, void __user *addr)
1680 unsigned long log = (unsigned long)addr;
1683 int bit = nr + (log % PAGE_SIZE) * 8;
1686 r = get_user_pages_fast(log, 1, 1, &page);
1690 base = kmap_atomic(page);
1692 kunmap_atomic(base);
1693 set_page_dirty_lock(page);
1698 static int log_write(void __user *log_base,
1699 u64 write_address, u64 write_length)
1701 u64 write_page = write_address / VHOST_PAGE_SIZE;
1706 write_length += write_address % VHOST_PAGE_SIZE;
1708 u64 base = (u64)(unsigned long)log_base;
1709 u64 log = base + write_page / 8;
1710 int bit = write_page % 8;
1711 if ((u64)(unsigned long)log != log)
1713 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1716 if (write_length <= VHOST_PAGE_SIZE)
1718 write_length -= VHOST_PAGE_SIZE;
1724 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1725 unsigned int log_num, u64 len)
1729 /* Make sure data written is seen before log. */
1731 for (i = 0; i < log_num; ++i) {
1732 u64 l = min(log[i].len, len);
1733 r = log_write(vq->log_base, log[i].addr, l);
1739 eventfd_signal(vq->log_ctx, 1);
1743 /* Length written exceeds what we have stored. This is a bug. */
1747 EXPORT_SYMBOL_GPL(vhost_log_write);
1749 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1752 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1753 &vq->used->flags) < 0)
1755 if (unlikely(vq->log_used)) {
1756 /* Make sure the flag is seen before log. */
1758 /* Log used flag write. */
1759 used = &vq->used->flags;
1760 log_write(vq->log_base, vq->log_addr +
1761 (used - (void __user *)vq->used),
1762 sizeof vq->used->flags);
1764 eventfd_signal(vq->log_ctx, 1);
1769 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1771 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1772 vhost_avail_event(vq)))
1774 if (unlikely(vq->log_used)) {
1776 /* Make sure the event is seen before log. */
1778 /* Log avail event write */
1779 used = vhost_avail_event(vq);
1780 log_write(vq->log_base, vq->log_addr +
1781 (used - (void __user *)vq->used),
1782 sizeof *vhost_avail_event(vq));
1784 eventfd_signal(vq->log_ctx, 1);
1789 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1791 __virtio16 last_used_idx;
1793 bool is_le = vq->is_le;
1795 if (!vq->private_data)
1798 vhost_init_is_le(vq);
1800 r = vhost_update_used_flags(vq);
1803 vq->signalled_used_valid = false;
1805 !access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1809 r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1811 vq_err(vq, "Can't access used idx at %p\n",
1815 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1822 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1824 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1825 struct iovec iov[], int iov_size, int access)
1827 const struct vhost_umem_node *node;
1828 struct vhost_dev *dev = vq->dev;
1829 struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1834 while ((u64)len > s) {
1836 if (unlikely(ret >= iov_size)) {
1841 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1842 addr, addr + len - 1);
1843 if (node == NULL || node->start > addr) {
1844 if (umem != dev->iotlb) {
1850 } else if (!(node->perm & access)) {
1856 size = node->size - addr + node->start;
1857 _iov->iov_len = min((u64)len - s, size);
1858 _iov->iov_base = (void __user *)(unsigned long)
1859 (node->userspace_addr + addr - node->start);
1866 vhost_iotlb_miss(vq, addr, access);
1870 /* Each buffer in the virtqueues is actually a chain of descriptors. This
1871 * function returns the next descriptor in the chain,
1872 * or -1U if we're at the end. */
1873 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1877 /* If this descriptor says it doesn't chain, we're done. */
1878 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1881 /* Check they're not leading us off end of descriptors. */
1882 next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
1886 static int get_indirect(struct vhost_virtqueue *vq,
1887 struct iovec iov[], unsigned int iov_size,
1888 unsigned int *out_num, unsigned int *in_num,
1889 struct vhost_log *log, unsigned int *log_num,
1890 struct vring_desc *indirect)
1892 struct vring_desc desc;
1893 unsigned int i = 0, count, found = 0;
1894 u32 len = vhost32_to_cpu(vq, indirect->len);
1895 struct iov_iter from;
1899 if (unlikely(len % sizeof desc)) {
1900 vq_err(vq, "Invalid length in indirect descriptor: "
1901 "len 0x%llx not multiple of 0x%zx\n",
1902 (unsigned long long)len,
1907 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1908 UIO_MAXIOV, VHOST_ACCESS_RO);
1909 if (unlikely(ret < 0)) {
1911 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1914 iov_iter_init(&from, READ, vq->indirect, ret, len);
1916 /* We will use the result as an address to read from, so most
1917 * architectures only need a compiler barrier here. */
1918 read_barrier_depends();
1920 count = len / sizeof desc;
1921 /* Buffers are chained via a 16 bit next field, so
1922 * we can have at most 2^16 of these. */
1923 if (unlikely(count > USHRT_MAX + 1)) {
1924 vq_err(vq, "Indirect buffer length too big: %d\n",
1930 unsigned iov_count = *in_num + *out_num;
1931 if (unlikely(++found > count)) {
1932 vq_err(vq, "Loop detected: last one at %u "
1933 "indirect size %u\n",
1937 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
1938 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1939 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1942 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1943 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1944 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1948 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
1949 access = VHOST_ACCESS_WO;
1951 access = VHOST_ACCESS_RO;
1953 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1954 vhost32_to_cpu(vq, desc.len), iov + iov_count,
1955 iov_size - iov_count, access);
1956 if (unlikely(ret < 0)) {
1958 vq_err(vq, "Translation failure %d indirect idx %d\n",
1962 /* If this is an input descriptor, increment that count. */
1963 if (access == VHOST_ACCESS_WO) {
1965 if (unlikely(log)) {
1966 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1967 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1971 /* If it's an output descriptor, they're all supposed
1972 * to come before any input descriptors. */
1973 if (unlikely(*in_num)) {
1974 vq_err(vq, "Indirect descriptor "
1975 "has out after in: idx %d\n", i);
1980 } while ((i = next_desc(vq, &desc)) != -1);
1984 /* This looks in the virtqueue and for the first available buffer, and converts
1985 * it to an iovec for convenient access. Since descriptors consist of some
1986 * number of output then some number of input descriptors, it's actually two
1987 * iovecs, but we pack them into one and note how many of each there were.
1989 * This function returns the descriptor number found, or vq->num (which is
1990 * never a valid descriptor number) if none was found. A negative code is
1991 * returned on error. */
1992 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1993 struct iovec iov[], unsigned int iov_size,
1994 unsigned int *out_num, unsigned int *in_num,
1995 struct vhost_log *log, unsigned int *log_num)
1997 struct vring_desc desc;
1998 unsigned int i, head, found = 0;
2000 __virtio16 avail_idx;
2001 __virtio16 ring_head;
2004 /* Check it isn't doing very strange things with descriptor numbers. */
2005 last_avail_idx = vq->last_avail_idx;
2007 if (vq->avail_idx == vq->last_avail_idx) {
2008 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
2009 vq_err(vq, "Failed to access avail idx at %p\n",
2013 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2015 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2016 vq_err(vq, "Guest moved used index from %u to %u",
2017 last_avail_idx, vq->avail_idx);
2021 /* If there's nothing new since last we looked, return
2024 if (vq->avail_idx == last_avail_idx)
2027 /* Only get avail ring entries after they have been
2033 /* Grab the next descriptor number they're advertising, and increment
2034 * the index we've seen. */
2035 if (unlikely(vhost_get_avail(vq, ring_head,
2036 &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2037 vq_err(vq, "Failed to read head: idx %d address %p\n",
2039 &vq->avail->ring[last_avail_idx % vq->num]);
2043 head = vhost16_to_cpu(vq, ring_head);
2045 /* If their number is silly, that's an error. */
2046 if (unlikely(head >= vq->num)) {
2047 vq_err(vq, "Guest says index %u > %u is available",
2052 /* When we start there are none of either input nor output. */
2053 *out_num = *in_num = 0;
2059 unsigned iov_count = *in_num + *out_num;
2060 if (unlikely(i >= vq->num)) {
2061 vq_err(vq, "Desc index is %u > %u, head = %u",
2065 if (unlikely(++found > vq->num)) {
2066 vq_err(vq, "Loop detected: last one at %u "
2067 "vq size %u head %u\n",
2071 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2073 if (unlikely(ret)) {
2074 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2078 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2079 ret = get_indirect(vq, iov, iov_size,
2081 log, log_num, &desc);
2082 if (unlikely(ret < 0)) {
2084 vq_err(vq, "Failure detected "
2085 "in indirect descriptor at idx %d\n", i);
2091 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2092 access = VHOST_ACCESS_WO;
2094 access = VHOST_ACCESS_RO;
2095 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2096 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2097 iov_size - iov_count, access);
2098 if (unlikely(ret < 0)) {
2100 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2104 if (access == VHOST_ACCESS_WO) {
2105 /* If this is an input descriptor,
2106 * increment that count. */
2108 if (unlikely(log)) {
2109 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2110 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2114 /* If it's an output descriptor, they're all supposed
2115 * to come before any input descriptors. */
2116 if (unlikely(*in_num)) {
2117 vq_err(vq, "Descriptor has out after in: "
2123 } while ((i = next_desc(vq, &desc)) != -1);
2125 /* On success, increment avail index. */
2126 vq->last_avail_idx++;
2128 /* Assume notifications from guest are disabled at this point,
2129 * if they aren't we would need to update avail_event index. */
2130 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2133 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2135 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2136 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2138 vq->last_avail_idx -= n;
2140 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2142 /* After we've used one of their buffers, we tell them about it. We'll then
2143 * want to notify the guest, using eventfd. */
2144 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2146 struct vring_used_elem heads = {
2147 cpu_to_vhost32(vq, head),
2148 cpu_to_vhost32(vq, len)
2151 return vhost_add_used_n(vq, &heads, 1);
2153 EXPORT_SYMBOL_GPL(vhost_add_used);
2155 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2156 struct vring_used_elem *heads,
2159 struct vring_used_elem __user *used;
2163 start = vq->last_used_idx & (vq->num - 1);
2164 used = vq->used->ring + start;
2166 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2167 vq_err(vq, "Failed to write used id");
2170 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2171 vq_err(vq, "Failed to write used len");
2174 } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2175 vq_err(vq, "Failed to write used");
2178 if (unlikely(vq->log_used)) {
2179 /* Make sure data is seen before log. */
2181 /* Log used ring entry write. */
2182 log_write(vq->log_base,
2184 ((void __user *)used - (void __user *)vq->used),
2185 count * sizeof *used);
2187 old = vq->last_used_idx;
2188 new = (vq->last_used_idx += count);
2189 /* If the driver never bothers to signal in a very long while,
2190 * used index might wrap around. If that happens, invalidate
2191 * signalled_used index we stored. TODO: make sure driver
2192 * signals at least once in 2^16 and remove this. */
2193 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2194 vq->signalled_used_valid = false;
2198 /* After we've used one of their buffers, we tell them about it. We'll then
2199 * want to notify the guest, using eventfd. */
2200 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2205 start = vq->last_used_idx & (vq->num - 1);
2206 n = vq->num - start;
2208 r = __vhost_add_used_n(vq, heads, n);
2214 r = __vhost_add_used_n(vq, heads, count);
2216 /* Make sure buffer is written before we update index. */
2218 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2220 vq_err(vq, "Failed to increment used idx");
2223 if (unlikely(vq->log_used)) {
2224 /* Log used index update. */
2225 log_write(vq->log_base,
2226 vq->log_addr + offsetof(struct vring_used, idx),
2227 sizeof vq->used->idx);
2229 eventfd_signal(vq->log_ctx, 1);
2233 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2235 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2240 /* Flush out used index updates. This is paired
2241 * with the barrier that the Guest executes when enabling
2245 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2246 unlikely(vq->avail_idx == vq->last_avail_idx))
2249 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2251 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2252 vq_err(vq, "Failed to get flags");
2255 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2257 old = vq->signalled_used;
2258 v = vq->signalled_used_valid;
2259 new = vq->signalled_used = vq->last_used_idx;
2260 vq->signalled_used_valid = true;
2265 if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2266 vq_err(vq, "Failed to get used event idx");
2269 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2272 /* This actually signals the guest, using eventfd. */
2273 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2275 /* Signal the Guest tell them we used something up. */
2276 if (vq->call_ctx && vhost_notify(dev, vq))
2277 eventfd_signal(vq->call_ctx, 1);
2279 EXPORT_SYMBOL_GPL(vhost_signal);
2281 /* And here's the combo meal deal. Supersize me! */
2282 void vhost_add_used_and_signal(struct vhost_dev *dev,
2283 struct vhost_virtqueue *vq,
2284 unsigned int head, int len)
2286 vhost_add_used(vq, head, len);
2287 vhost_signal(dev, vq);
2289 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2291 /* multi-buffer version of vhost_add_used_and_signal */
2292 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2293 struct vhost_virtqueue *vq,
2294 struct vring_used_elem *heads, unsigned count)
2296 vhost_add_used_n(vq, heads, count);
2297 vhost_signal(dev, vq);
2299 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2301 /* return true if we're sure that avaiable ring is empty */
2302 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2304 __virtio16 avail_idx;
2307 if (vq->avail_idx != vq->last_avail_idx)
2310 r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2313 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2315 return vq->avail_idx == vq->last_avail_idx;
2317 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2319 /* OK, now we need to know about added descriptors. */
2320 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2322 __virtio16 avail_idx;
2325 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2327 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2328 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2329 r = vhost_update_used_flags(vq);
2331 vq_err(vq, "Failed to enable notification at %p: %d\n",
2332 &vq->used->flags, r);
2336 r = vhost_update_avail_event(vq, vq->avail_idx);
2338 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2339 vhost_avail_event(vq), r);
2343 /* They could have slipped one in as we were doing that: make
2344 * sure it's written, then check again. */
2346 r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2348 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2349 &vq->avail->idx, r);
2353 return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2355 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2357 /* We don't need to be notified again. */
2358 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2362 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2364 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2365 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2366 r = vhost_update_used_flags(vq);
2368 vq_err(vq, "Failed to enable notification at %p: %d\n",
2369 &vq->used->flags, r);
2372 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2374 /* Create a new message. */
2375 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2377 struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2381 /* Make sure all padding within the structure is initialized. */
2382 memset(&node->msg, 0, sizeof node->msg);
2384 node->msg.type = type;
2387 EXPORT_SYMBOL_GPL(vhost_new_msg);
2389 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2390 struct vhost_msg_node *node)
2392 spin_lock(&dev->iotlb_lock);
2393 list_add_tail(&node->node, head);
2394 spin_unlock(&dev->iotlb_lock);
2396 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2398 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2400 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2401 struct list_head *head)
2403 struct vhost_msg_node *node = NULL;
2405 spin_lock(&dev->iotlb_lock);
2406 if (!list_empty(head)) {
2407 node = list_first_entry(head, struct vhost_msg_node,
2409 list_del(&node->node);
2411 spin_unlock(&dev->iotlb_lock);
2415 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2418 static int __init vhost_init(void)
2423 static void __exit vhost_exit(void)
2427 module_init(vhost_init);
2428 module_exit(vhost_exit);
2430 MODULE_VERSION("0.0.1");
2431 MODULE_LICENSE("GPL v2");
2432 MODULE_AUTHOR("Michael S. Tsirkin");
2433 MODULE_DESCRIPTION("Host kernel accelerator for virtio");