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 __vhost_vq_meta_reset(d->vqs[i]);
301 static void vhost_vq_reset(struct vhost_dev *dev,
302 struct vhost_virtqueue *vq)
308 vq->last_avail_idx = 0;
310 vq->last_used_idx = 0;
311 vq->signalled_used = 0;
312 vq->signalled_used_valid = false;
314 vq->log_used = false;
315 vq->log_addr = -1ull;
316 vq->private_data = NULL;
317 vq->acked_features = 0;
318 vq->acked_backend_features = 0;
320 vq->error_ctx = NULL;
324 vhost_reset_is_le(vq);
325 vhost_disable_cross_endian(vq);
326 vq->busyloop_timeout = 0;
329 __vhost_vq_meta_reset(vq);
332 static int vhost_worker(void *data)
334 struct vhost_dev *dev = data;
335 struct vhost_work *work, *work_next;
336 struct llist_node *node;
337 mm_segment_t oldfs = get_fs();
343 /* mb paired w/ kthread_stop */
344 set_current_state(TASK_INTERRUPTIBLE);
346 if (kthread_should_stop()) {
347 __set_current_state(TASK_RUNNING);
351 node = llist_del_all(&dev->work_list);
355 node = llist_reverse_order(node);
356 /* make sure flag is seen after deletion */
358 llist_for_each_entry_safe(work, work_next, node, node) {
359 clear_bit(VHOST_WORK_QUEUED, &work->flags);
360 __set_current_state(TASK_RUNNING);
371 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
381 /* Helper to allocate iovec buffers for all vqs. */
382 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
384 struct vhost_virtqueue *vq;
387 for (i = 0; i < dev->nvqs; ++i) {
389 vq->indirect = kmalloc_array(UIO_MAXIOV,
390 sizeof(*vq->indirect),
392 vq->log = kmalloc_array(UIO_MAXIOV, sizeof(*vq->log),
394 vq->heads = kmalloc_array(UIO_MAXIOV, sizeof(*vq->heads),
396 if (!vq->indirect || !vq->log || !vq->heads)
403 vhost_vq_free_iovecs(dev->vqs[i]);
407 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
411 for (i = 0; i < dev->nvqs; ++i)
412 vhost_vq_free_iovecs(dev->vqs[i]);
415 void vhost_dev_init(struct vhost_dev *dev,
416 struct vhost_virtqueue **vqs, int nvqs)
418 struct vhost_virtqueue *vq;
423 mutex_init(&dev->mutex);
429 init_llist_head(&dev->work_list);
430 init_waitqueue_head(&dev->wait);
431 INIT_LIST_HEAD(&dev->read_list);
432 INIT_LIST_HEAD(&dev->pending_list);
433 spin_lock_init(&dev->iotlb_lock);
436 for (i = 0; i < dev->nvqs; ++i) {
442 mutex_init(&vq->mutex);
443 vhost_vq_reset(dev, vq);
445 vhost_poll_init(&vq->poll, vq->handle_kick,
449 EXPORT_SYMBOL_GPL(vhost_dev_init);
451 /* Caller should have device mutex */
452 long vhost_dev_check_owner(struct vhost_dev *dev)
454 /* Are you the owner? If not, I don't think you mean to do that */
455 return dev->mm == current->mm ? 0 : -EPERM;
457 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
459 struct vhost_attach_cgroups_struct {
460 struct vhost_work work;
461 struct task_struct *owner;
465 static void vhost_attach_cgroups_work(struct vhost_work *work)
467 struct vhost_attach_cgroups_struct *s;
469 s = container_of(work, struct vhost_attach_cgroups_struct, work);
470 s->ret = cgroup_attach_task_all(s->owner, current);
473 static int vhost_attach_cgroups(struct vhost_dev *dev)
475 struct vhost_attach_cgroups_struct attach;
477 attach.owner = current;
478 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
479 vhost_work_queue(dev, &attach.work);
480 vhost_work_flush(dev, &attach.work);
484 /* Caller should have device mutex */
485 bool vhost_dev_has_owner(struct vhost_dev *dev)
489 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
491 /* Caller should have device mutex */
492 long vhost_dev_set_owner(struct vhost_dev *dev)
494 struct task_struct *worker;
497 /* Is there an owner already? */
498 if (vhost_dev_has_owner(dev)) {
503 /* No owner, become one */
504 dev->mm = get_task_mm(current);
505 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
506 if (IS_ERR(worker)) {
507 err = PTR_ERR(worker);
511 dev->worker = worker;
512 wake_up_process(worker); /* avoid contributing to loadavg */
514 err = vhost_attach_cgroups(dev);
518 err = vhost_dev_alloc_iovecs(dev);
524 kthread_stop(worker);
533 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
535 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
537 return kvzalloc(sizeof(struct vhost_umem), GFP_KERNEL);
539 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
541 /* Caller should have device mutex */
542 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
546 vhost_dev_cleanup(dev);
548 /* Restore memory to default empty mapping. */
549 INIT_LIST_HEAD(&umem->umem_list);
551 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
552 * VQs aren't running.
554 for (i = 0; i < dev->nvqs; ++i)
555 dev->vqs[i]->umem = umem;
557 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
559 void vhost_dev_stop(struct vhost_dev *dev)
563 for (i = 0; i < dev->nvqs; ++i) {
564 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
565 vhost_poll_stop(&dev->vqs[i]->poll);
566 vhost_poll_flush(&dev->vqs[i]->poll);
570 EXPORT_SYMBOL_GPL(vhost_dev_stop);
572 static void vhost_umem_free(struct vhost_umem *umem,
573 struct vhost_umem_node *node)
575 vhost_umem_interval_tree_remove(node, &umem->umem_tree);
576 list_del(&node->link);
581 static void vhost_umem_clean(struct vhost_umem *umem)
583 struct vhost_umem_node *node, *tmp;
588 list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
589 vhost_umem_free(umem, node);
594 static void vhost_clear_msg(struct vhost_dev *dev)
596 struct vhost_msg_node *node, *n;
598 spin_lock(&dev->iotlb_lock);
600 list_for_each_entry_safe(node, n, &dev->read_list, node) {
601 list_del(&node->node);
605 list_for_each_entry_safe(node, n, &dev->pending_list, node) {
606 list_del(&node->node);
610 spin_unlock(&dev->iotlb_lock);
613 void vhost_dev_cleanup(struct vhost_dev *dev)
617 for (i = 0; i < dev->nvqs; ++i) {
618 if (dev->vqs[i]->error_ctx)
619 eventfd_ctx_put(dev->vqs[i]->error_ctx);
620 if (dev->vqs[i]->kick)
621 fput(dev->vqs[i]->kick);
622 if (dev->vqs[i]->call_ctx)
623 eventfd_ctx_put(dev->vqs[i]->call_ctx);
624 vhost_vq_reset(dev, dev->vqs[i]);
626 vhost_dev_free_iovecs(dev);
628 eventfd_ctx_put(dev->log_ctx);
630 /* No one will access memory at this point */
631 vhost_umem_clean(dev->umem);
633 vhost_umem_clean(dev->iotlb);
635 vhost_clear_msg(dev);
636 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
637 WARN_ON(!llist_empty(&dev->work_list));
639 kthread_stop(dev->worker);
646 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
648 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
650 u64 a = addr / VHOST_PAGE_SIZE / 8;
652 /* Make sure 64 bit math will not overflow. */
653 if (a > ULONG_MAX - (unsigned long)log_base ||
654 a + (unsigned long)log_base > ULONG_MAX)
657 return access_ok(VERIFY_WRITE, log_base + a,
658 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
661 static bool vhost_overflow(u64 uaddr, u64 size)
663 /* Make sure 64 bit math will not overflow. */
664 return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
667 /* Caller should have vq mutex and device mutex. */
668 static bool vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
671 struct vhost_umem_node *node;
676 list_for_each_entry(node, &umem->umem_list, link) {
677 unsigned long a = node->userspace_addr;
679 if (vhost_overflow(node->userspace_addr, node->size))
683 if (!access_ok(VERIFY_WRITE, (void __user *)a,
686 else if (log_all && !log_access_ok(log_base,
694 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
695 u64 addr, unsigned int size,
698 const struct vhost_umem_node *node = vq->meta_iotlb[type];
703 return (void *)(uintptr_t)(node->userspace_addr + addr - node->start);
706 /* Can we switch to this memory table? */
707 /* Caller should have device mutex but not vq mutex */
708 static bool memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
713 for (i = 0; i < d->nvqs; ++i) {
717 mutex_lock(&d->vqs[i]->mutex);
718 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
719 /* If ring is inactive, will check when it's enabled. */
720 if (d->vqs[i]->private_data)
721 ok = vq_memory_access_ok(d->vqs[i]->log_base,
725 mutex_unlock(&d->vqs[i]->mutex);
732 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
733 struct iovec iov[], int iov_size, int access);
735 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
736 const void *from, unsigned size)
741 return __copy_to_user(to, from, size);
743 /* This function should be called after iotlb
744 * prefetch, which means we're sure that all vq
745 * could be access through iotlb. So -EAGAIN should
746 * not happen in this case.
749 void __user *uaddr = vhost_vq_meta_fetch(vq,
750 (u64)(uintptr_t)to, size,
754 return __copy_to_user(uaddr, from, size);
756 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
757 ARRAY_SIZE(vq->iotlb_iov),
761 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
762 ret = copy_to_iter(from, size, &t);
770 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
771 void __user *from, unsigned size)
776 return __copy_from_user(to, from, size);
778 /* This function should be called after iotlb
779 * prefetch, which means we're sure that vq
780 * could be access through iotlb. So -EAGAIN should
781 * not happen in this case.
783 void __user *uaddr = vhost_vq_meta_fetch(vq,
784 (u64)(uintptr_t)from, size,
789 return __copy_from_user(to, uaddr, size);
791 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
792 ARRAY_SIZE(vq->iotlb_iov),
795 vq_err(vq, "IOTLB translation failure: uaddr "
796 "%p size 0x%llx\n", from,
797 (unsigned long long) size);
800 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
801 ret = copy_from_iter(to, size, &f);
810 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
811 void __user *addr, unsigned int size,
816 ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
817 ARRAY_SIZE(vq->iotlb_iov),
820 vq_err(vq, "IOTLB translation failure: uaddr "
821 "%p size 0x%llx\n", addr,
822 (unsigned long long) size);
826 if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
827 vq_err(vq, "Non atomic userspace memory access: uaddr "
828 "%p size 0x%llx\n", addr,
829 (unsigned long long) size);
833 return vq->iotlb_iov[0].iov_base;
836 /* This function should be called after iotlb
837 * prefetch, which means we're sure that vq
838 * could be access through iotlb. So -EAGAIN should
839 * not happen in this case.
841 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
842 void *addr, unsigned int size,
845 void __user *uaddr = vhost_vq_meta_fetch(vq,
846 (u64)(uintptr_t)addr, size, type);
850 return __vhost_get_user_slow(vq, addr, size, type);
853 #define vhost_put_user(vq, x, ptr) \
857 ret = __put_user(x, ptr); \
859 __typeof__(ptr) to = \
860 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
861 sizeof(*ptr), VHOST_ADDR_USED); \
863 ret = __put_user(x, to); \
870 #define vhost_get_user(vq, x, ptr, type) \
874 ret = __get_user(x, ptr); \
876 __typeof__(ptr) from = \
877 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
881 ret = __get_user(x, from); \
888 #define vhost_get_avail(vq, x, ptr) \
889 vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
891 #define vhost_get_used(vq, x, ptr) \
892 vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
894 static void vhost_dev_lock_vqs(struct vhost_dev *d)
897 for (i = 0; i < d->nvqs; ++i)
898 mutex_lock_nested(&d->vqs[i]->mutex, i);
901 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
904 for (i = 0; i < d->nvqs; ++i)
905 mutex_unlock(&d->vqs[i]->mutex);
908 static int vhost_new_umem_range(struct vhost_umem *umem,
909 u64 start, u64 size, u64 end,
910 u64 userspace_addr, int perm)
912 struct vhost_umem_node *tmp, *node = kmalloc(sizeof(*node), GFP_ATOMIC);
917 if (umem->numem == max_iotlb_entries) {
918 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
919 vhost_umem_free(umem, tmp);
925 node->userspace_addr = userspace_addr;
927 INIT_LIST_HEAD(&node->link);
928 list_add_tail(&node->link, &umem->umem_list);
929 vhost_umem_interval_tree_insert(node, &umem->umem_tree);
935 static void vhost_del_umem_range(struct vhost_umem *umem,
938 struct vhost_umem_node *node;
940 while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
942 vhost_umem_free(umem, node);
945 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
946 struct vhost_iotlb_msg *msg)
948 struct vhost_msg_node *node, *n;
950 spin_lock(&d->iotlb_lock);
952 list_for_each_entry_safe(node, n, &d->pending_list, node) {
953 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
954 if (msg->iova <= vq_msg->iova &&
955 msg->iova + msg->size - 1 > vq_msg->iova &&
956 vq_msg->type == VHOST_IOTLB_MISS) {
957 vhost_poll_queue(&node->vq->poll);
958 list_del(&node->node);
963 spin_unlock(&d->iotlb_lock);
966 static bool umem_access_ok(u64 uaddr, u64 size, int access)
968 unsigned long a = uaddr;
970 /* Make sure 64 bit math will not overflow. */
971 if (vhost_overflow(uaddr, size))
974 if ((access & VHOST_ACCESS_RO) &&
975 !access_ok(VERIFY_READ, (void __user *)a, size))
977 if ((access & VHOST_ACCESS_WO) &&
978 !access_ok(VERIFY_WRITE, (void __user *)a, size))
983 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
984 struct vhost_iotlb_msg *msg)
988 mutex_lock(&dev->mutex);
989 vhost_dev_lock_vqs(dev);
991 case VHOST_IOTLB_UPDATE:
996 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1000 vhost_vq_meta_reset(dev);
1001 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
1002 msg->iova + msg->size - 1,
1003 msg->uaddr, msg->perm)) {
1007 vhost_iotlb_notify_vq(dev, msg);
1009 case VHOST_IOTLB_INVALIDATE:
1014 vhost_vq_meta_reset(dev);
1015 vhost_del_umem_range(dev->iotlb, msg->iova,
1016 msg->iova + msg->size - 1);
1023 vhost_dev_unlock_vqs(dev);
1024 mutex_unlock(&dev->mutex);
1028 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1029 struct iov_iter *from)
1031 struct vhost_iotlb_msg msg;
1035 ret = copy_from_iter(&type, sizeof(type), from);
1036 if (ret != sizeof(type))
1040 case VHOST_IOTLB_MSG:
1041 /* There maybe a hole after type for V1 message type,
1044 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1046 case VHOST_IOTLB_MSG_V2:
1047 offset = sizeof(__u32);
1054 iov_iter_advance(from, offset);
1055 ret = copy_from_iter(&msg, sizeof(msg), from);
1056 if (ret != sizeof(msg))
1058 if (vhost_process_iotlb_msg(dev, &msg)) {
1063 ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1064 sizeof(struct vhost_msg_v2);
1068 EXPORT_SYMBOL(vhost_chr_write_iter);
1070 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1075 poll_wait(file, &dev->wait, wait);
1077 if (!list_empty(&dev->read_list))
1078 mask |= EPOLLIN | EPOLLRDNORM;
1082 EXPORT_SYMBOL(vhost_chr_poll);
1084 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1088 struct vhost_msg_node *node;
1090 unsigned size = sizeof(struct vhost_msg);
1092 if (iov_iter_count(to) < size)
1097 prepare_to_wait(&dev->wait, &wait,
1098 TASK_INTERRUPTIBLE);
1100 node = vhost_dequeue_msg(dev, &dev->read_list);
1107 if (signal_pending(current)) {
1120 finish_wait(&dev->wait, &wait);
1123 struct vhost_iotlb_msg *msg;
1124 void *start = &node->msg;
1126 switch (node->msg.type) {
1127 case VHOST_IOTLB_MSG:
1128 size = sizeof(node->msg);
1129 msg = &node->msg.iotlb;
1131 case VHOST_IOTLB_MSG_V2:
1132 size = sizeof(node->msg_v2);
1133 msg = &node->msg_v2.iotlb;
1140 ret = copy_to_iter(start, size, to);
1141 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1145 vhost_enqueue_msg(dev, &dev->pending_list, node);
1150 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1152 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1154 struct vhost_dev *dev = vq->dev;
1155 struct vhost_msg_node *node;
1156 struct vhost_iotlb_msg *msg;
1157 bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1159 node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1164 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1165 msg = &node->msg_v2.iotlb;
1167 msg = &node->msg.iotlb;
1170 msg->type = VHOST_IOTLB_MISS;
1174 vhost_enqueue_msg(dev, &dev->read_list, node);
1179 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1180 struct vring_desc __user *desc,
1181 struct vring_avail __user *avail,
1182 struct vring_used __user *used)
1185 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1187 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
1188 access_ok(VERIFY_READ, avail,
1189 sizeof *avail + num * sizeof *avail->ring + s) &&
1190 access_ok(VERIFY_WRITE, used,
1191 sizeof *used + num * sizeof *used->ring + s);
1194 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1195 const struct vhost_umem_node *node,
1198 int access = (type == VHOST_ADDR_USED) ?
1199 VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1201 if (likely(node->perm & access))
1202 vq->meta_iotlb[type] = node;
1205 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1206 int access, u64 addr, u64 len, int type)
1208 const struct vhost_umem_node *node;
1209 struct vhost_umem *umem = vq->iotlb;
1210 u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1212 if (vhost_vq_meta_fetch(vq, addr, len, type))
1216 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1219 if (node == NULL || node->start > addr) {
1220 vhost_iotlb_miss(vq, addr, access);
1222 } else if (!(node->perm & access)) {
1223 /* Report the possible access violation by
1224 * request another translation from userspace.
1229 size = node->size - addr + node->start;
1231 if (orig_addr == addr && size >= len)
1232 vhost_vq_meta_update(vq, node, type);
1241 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1243 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1244 unsigned int num = vq->num;
1249 return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1250 num * sizeof(*vq->desc), VHOST_ADDR_DESC) &&
1251 iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1253 num * sizeof(*vq->avail->ring) + s,
1254 VHOST_ADDR_AVAIL) &&
1255 iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1257 num * sizeof(*vq->used->ring) + s,
1260 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1262 /* Can we log writes? */
1263 /* Caller should have device mutex but not vq mutex */
1264 bool vhost_log_access_ok(struct vhost_dev *dev)
1266 return memory_access_ok(dev, dev->umem, 1);
1268 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1270 /* Verify access for write logging. */
1271 /* Caller should have vq mutex and device mutex */
1272 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1273 void __user *log_base)
1275 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1277 return vq_memory_access_ok(log_base, vq->umem,
1278 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1279 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1281 vq->num * sizeof *vq->used->ring + s));
1284 /* Can we start vq? */
1285 /* Caller should have vq mutex and device mutex */
1286 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1288 if (!vq_log_access_ok(vq, vq->log_base))
1291 /* Access validation occurs at prefetch time with IOTLB */
1295 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1297 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1299 static struct vhost_umem *vhost_umem_alloc(void)
1301 struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1306 umem->umem_tree = RB_ROOT_CACHED;
1308 INIT_LIST_HEAD(&umem->umem_list);
1313 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1315 struct vhost_memory mem, *newmem;
1316 struct vhost_memory_region *region;
1317 struct vhost_umem *newumem, *oldumem;
1318 unsigned long size = offsetof(struct vhost_memory, regions);
1321 if (copy_from_user(&mem, m, size))
1325 if (mem.nregions > max_mem_regions)
1327 newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1332 memcpy(newmem, &mem, size);
1333 if (copy_from_user(newmem->regions, m->regions,
1334 mem.nregions * sizeof *m->regions)) {
1339 newumem = vhost_umem_alloc();
1345 for (region = newmem->regions;
1346 region < newmem->regions + mem.nregions;
1348 if (vhost_new_umem_range(newumem,
1349 region->guest_phys_addr,
1350 region->memory_size,
1351 region->guest_phys_addr +
1352 region->memory_size - 1,
1353 region->userspace_addr,
1358 if (!memory_access_ok(d, newumem, 0))
1364 /* All memory accesses are done under some VQ mutex. */
1365 for (i = 0; i < d->nvqs; ++i) {
1366 mutex_lock(&d->vqs[i]->mutex);
1367 d->vqs[i]->umem = newumem;
1368 mutex_unlock(&d->vqs[i]->mutex);
1372 vhost_umem_clean(oldumem);
1376 vhost_umem_clean(newumem);
1381 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1383 struct file *eventfp, *filep = NULL;
1384 bool pollstart = false, pollstop = false;
1385 struct eventfd_ctx *ctx = NULL;
1386 u32 __user *idxp = argp;
1387 struct vhost_virtqueue *vq;
1388 struct vhost_vring_state s;
1389 struct vhost_vring_file f;
1390 struct vhost_vring_addr a;
1394 r = get_user(idx, idxp);
1402 mutex_lock(&vq->mutex);
1405 case VHOST_SET_VRING_NUM:
1406 /* Resizing ring with an active backend?
1407 * You don't want to do that. */
1408 if (vq->private_data) {
1412 if (copy_from_user(&s, argp, sizeof s)) {
1416 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1422 case VHOST_SET_VRING_BASE:
1423 /* Moving base with an active backend?
1424 * You don't want to do that. */
1425 if (vq->private_data) {
1429 if (copy_from_user(&s, argp, sizeof s)) {
1433 if (s.num > 0xffff) {
1437 vq->last_avail_idx = s.num;
1438 /* Forget the cached index value. */
1439 vq->avail_idx = vq->last_avail_idx;
1441 case VHOST_GET_VRING_BASE:
1443 s.num = vq->last_avail_idx;
1444 if (copy_to_user(argp, &s, sizeof s))
1447 case VHOST_SET_VRING_ADDR:
1448 if (copy_from_user(&a, argp, sizeof a)) {
1452 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1456 /* For 32bit, verify that the top 32bits of the user
1457 data are set to zero. */
1458 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1459 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1460 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1465 /* Make sure it's safe to cast pointers to vring types. */
1466 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1467 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1468 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1469 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1470 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1475 /* We only verify access here if backend is configured.
1476 * If it is not, we don't as size might not have been setup.
1477 * We will verify when backend is configured. */
1478 if (vq->private_data) {
1479 if (!vq_access_ok(vq, vq->num,
1480 (void __user *)(unsigned long)a.desc_user_addr,
1481 (void __user *)(unsigned long)a.avail_user_addr,
1482 (void __user *)(unsigned long)a.used_user_addr)) {
1487 /* Also validate log access for used ring if enabled. */
1488 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1489 !log_access_ok(vq->log_base, a.log_guest_addr,
1491 vq->num * sizeof *vq->used->ring)) {
1497 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1498 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1499 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1500 vq->log_addr = a.log_guest_addr;
1501 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1503 case VHOST_SET_VRING_KICK:
1504 if (copy_from_user(&f, argp, sizeof f)) {
1508 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1509 if (IS_ERR(eventfp)) {
1510 r = PTR_ERR(eventfp);
1513 if (eventfp != vq->kick) {
1514 pollstop = (filep = vq->kick) != NULL;
1515 pollstart = (vq->kick = eventfp) != NULL;
1519 case VHOST_SET_VRING_CALL:
1520 if (copy_from_user(&f, argp, sizeof f)) {
1524 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1529 swap(ctx, vq->call_ctx);
1531 case VHOST_SET_VRING_ERR:
1532 if (copy_from_user(&f, argp, sizeof f)) {
1536 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1541 swap(ctx, vq->error_ctx);
1543 case VHOST_SET_VRING_ENDIAN:
1544 r = vhost_set_vring_endian(vq, argp);
1546 case VHOST_GET_VRING_ENDIAN:
1547 r = vhost_get_vring_endian(vq, idx, argp);
1549 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1550 if (copy_from_user(&s, argp, sizeof(s))) {
1554 vq->busyloop_timeout = s.num;
1556 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1558 s.num = vq->busyloop_timeout;
1559 if (copy_to_user(argp, &s, sizeof(s)))
1566 if (pollstop && vq->handle_kick)
1567 vhost_poll_stop(&vq->poll);
1569 if (!IS_ERR_OR_NULL(ctx))
1570 eventfd_ctx_put(ctx);
1574 if (pollstart && vq->handle_kick)
1575 r = vhost_poll_start(&vq->poll, vq->kick);
1577 mutex_unlock(&vq->mutex);
1579 if (pollstop && vq->handle_kick)
1580 vhost_poll_flush(&vq->poll);
1583 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1585 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1587 struct vhost_umem *niotlb, *oiotlb;
1590 niotlb = vhost_umem_alloc();
1597 for (i = 0; i < d->nvqs; ++i) {
1598 struct vhost_virtqueue *vq = d->vqs[i];
1600 mutex_lock(&vq->mutex);
1602 __vhost_vq_meta_reset(vq);
1603 mutex_unlock(&vq->mutex);
1606 vhost_umem_clean(oiotlb);
1610 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1612 /* Caller must have device mutex */
1613 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1615 struct eventfd_ctx *ctx;
1620 /* If you are not the owner, you can become one */
1621 if (ioctl == VHOST_SET_OWNER) {
1622 r = vhost_dev_set_owner(d);
1626 /* You must be the owner to do anything else */
1627 r = vhost_dev_check_owner(d);
1632 case VHOST_SET_MEM_TABLE:
1633 r = vhost_set_memory(d, argp);
1635 case VHOST_SET_LOG_BASE:
1636 if (copy_from_user(&p, argp, sizeof p)) {
1640 if ((u64)(unsigned long)p != p) {
1644 for (i = 0; i < d->nvqs; ++i) {
1645 struct vhost_virtqueue *vq;
1646 void __user *base = (void __user *)(unsigned long)p;
1648 mutex_lock(&vq->mutex);
1649 /* If ring is inactive, will check when it's enabled. */
1650 if (vq->private_data && !vq_log_access_ok(vq, base))
1653 vq->log_base = base;
1654 mutex_unlock(&vq->mutex);
1657 case VHOST_SET_LOG_FD:
1658 r = get_user(fd, (int __user *)argp);
1661 ctx = fd == -1 ? NULL : eventfd_ctx_fdget(fd);
1666 swap(ctx, d->log_ctx);
1667 for (i = 0; i < d->nvqs; ++i) {
1668 mutex_lock(&d->vqs[i]->mutex);
1669 d->vqs[i]->log_ctx = d->log_ctx;
1670 mutex_unlock(&d->vqs[i]->mutex);
1673 eventfd_ctx_put(ctx);
1682 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1684 /* TODO: This is really inefficient. We need something like get_user()
1685 * (instruction directly accesses the data, with an exception table entry
1686 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1688 static int set_bit_to_user(int nr, void __user *addr)
1690 unsigned long log = (unsigned long)addr;
1693 int bit = nr + (log % PAGE_SIZE) * 8;
1696 r = get_user_pages_fast(log, 1, 1, &page);
1700 base = kmap_atomic(page);
1702 kunmap_atomic(base);
1703 set_page_dirty_lock(page);
1708 static int log_write(void __user *log_base,
1709 u64 write_address, u64 write_length)
1711 u64 write_page = write_address / VHOST_PAGE_SIZE;
1716 write_length += write_address % VHOST_PAGE_SIZE;
1718 u64 base = (u64)(unsigned long)log_base;
1719 u64 log = base + write_page / 8;
1720 int bit = write_page % 8;
1721 if ((u64)(unsigned long)log != log)
1723 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1726 if (write_length <= VHOST_PAGE_SIZE)
1728 write_length -= VHOST_PAGE_SIZE;
1734 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1735 unsigned int log_num, u64 len)
1739 /* Make sure data written is seen before log. */
1741 for (i = 0; i < log_num; ++i) {
1742 u64 l = min(log[i].len, len);
1743 r = log_write(vq->log_base, log[i].addr, l);
1749 eventfd_signal(vq->log_ctx, 1);
1753 /* Length written exceeds what we have stored. This is a bug. */
1757 EXPORT_SYMBOL_GPL(vhost_log_write);
1759 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1762 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1763 &vq->used->flags) < 0)
1765 if (unlikely(vq->log_used)) {
1766 /* Make sure the flag is seen before log. */
1768 /* Log used flag write. */
1769 used = &vq->used->flags;
1770 log_write(vq->log_base, vq->log_addr +
1771 (used - (void __user *)vq->used),
1772 sizeof vq->used->flags);
1774 eventfd_signal(vq->log_ctx, 1);
1779 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1781 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1782 vhost_avail_event(vq)))
1784 if (unlikely(vq->log_used)) {
1786 /* Make sure the event is seen before log. */
1788 /* Log avail event write */
1789 used = vhost_avail_event(vq);
1790 log_write(vq->log_base, vq->log_addr +
1791 (used - (void __user *)vq->used),
1792 sizeof *vhost_avail_event(vq));
1794 eventfd_signal(vq->log_ctx, 1);
1799 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1801 __virtio16 last_used_idx;
1803 bool is_le = vq->is_le;
1805 if (!vq->private_data)
1808 vhost_init_is_le(vq);
1810 r = vhost_update_used_flags(vq);
1813 vq->signalled_used_valid = false;
1815 !access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1819 r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1821 vq_err(vq, "Can't access used idx at %p\n",
1825 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1832 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1834 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1835 struct iovec iov[], int iov_size, int access)
1837 const struct vhost_umem_node *node;
1838 struct vhost_dev *dev = vq->dev;
1839 struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1844 while ((u64)len > s) {
1846 if (unlikely(ret >= iov_size)) {
1851 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1852 addr, addr + len - 1);
1853 if (node == NULL || node->start > addr) {
1854 if (umem != dev->iotlb) {
1860 } else if (!(node->perm & access)) {
1866 size = node->size - addr + node->start;
1867 _iov->iov_len = min((u64)len - s, size);
1868 _iov->iov_base = (void __user *)(unsigned long)
1869 (node->userspace_addr + addr - node->start);
1876 vhost_iotlb_miss(vq, addr, access);
1880 /* Each buffer in the virtqueues is actually a chain of descriptors. This
1881 * function returns the next descriptor in the chain,
1882 * or -1U if we're at the end. */
1883 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1887 /* If this descriptor says it doesn't chain, we're done. */
1888 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1891 /* Check they're not leading us off end of descriptors. */
1892 next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
1896 static int get_indirect(struct vhost_virtqueue *vq,
1897 struct iovec iov[], unsigned int iov_size,
1898 unsigned int *out_num, unsigned int *in_num,
1899 struct vhost_log *log, unsigned int *log_num,
1900 struct vring_desc *indirect)
1902 struct vring_desc desc;
1903 unsigned int i = 0, count, found = 0;
1904 u32 len = vhost32_to_cpu(vq, indirect->len);
1905 struct iov_iter from;
1909 if (unlikely(len % sizeof desc)) {
1910 vq_err(vq, "Invalid length in indirect descriptor: "
1911 "len 0x%llx not multiple of 0x%zx\n",
1912 (unsigned long long)len,
1917 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1918 UIO_MAXIOV, VHOST_ACCESS_RO);
1919 if (unlikely(ret < 0)) {
1921 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1924 iov_iter_init(&from, READ, vq->indirect, ret, len);
1926 /* We will use the result as an address to read from, so most
1927 * architectures only need a compiler barrier here. */
1928 read_barrier_depends();
1930 count = len / sizeof desc;
1931 /* Buffers are chained via a 16 bit next field, so
1932 * we can have at most 2^16 of these. */
1933 if (unlikely(count > USHRT_MAX + 1)) {
1934 vq_err(vq, "Indirect buffer length too big: %d\n",
1940 unsigned iov_count = *in_num + *out_num;
1941 if (unlikely(++found > count)) {
1942 vq_err(vq, "Loop detected: last one at %u "
1943 "indirect size %u\n",
1947 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
1948 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1949 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1952 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1953 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1954 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1958 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
1959 access = VHOST_ACCESS_WO;
1961 access = VHOST_ACCESS_RO;
1963 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1964 vhost32_to_cpu(vq, desc.len), iov + iov_count,
1965 iov_size - iov_count, access);
1966 if (unlikely(ret < 0)) {
1968 vq_err(vq, "Translation failure %d indirect idx %d\n",
1972 /* If this is an input descriptor, increment that count. */
1973 if (access == VHOST_ACCESS_WO) {
1975 if (unlikely(log)) {
1976 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1977 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1981 /* If it's an output descriptor, they're all supposed
1982 * to come before any input descriptors. */
1983 if (unlikely(*in_num)) {
1984 vq_err(vq, "Indirect descriptor "
1985 "has out after in: idx %d\n", i);
1990 } while ((i = next_desc(vq, &desc)) != -1);
1994 /* This looks in the virtqueue and for the first available buffer, and converts
1995 * it to an iovec for convenient access. Since descriptors consist of some
1996 * number of output then some number of input descriptors, it's actually two
1997 * iovecs, but we pack them into one and note how many of each there were.
1999 * This function returns the descriptor number found, or vq->num (which is
2000 * never a valid descriptor number) if none was found. A negative code is
2001 * returned on error. */
2002 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2003 struct iovec iov[], unsigned int iov_size,
2004 unsigned int *out_num, unsigned int *in_num,
2005 struct vhost_log *log, unsigned int *log_num)
2007 struct vring_desc desc;
2008 unsigned int i, head, found = 0;
2010 __virtio16 avail_idx;
2011 __virtio16 ring_head;
2014 /* Check it isn't doing very strange things with descriptor numbers. */
2015 last_avail_idx = vq->last_avail_idx;
2017 if (vq->avail_idx == vq->last_avail_idx) {
2018 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
2019 vq_err(vq, "Failed to access avail idx at %p\n",
2023 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2025 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2026 vq_err(vq, "Guest moved used index from %u to %u",
2027 last_avail_idx, vq->avail_idx);
2031 /* If there's nothing new since last we looked, return
2034 if (vq->avail_idx == last_avail_idx)
2037 /* Only get avail ring entries after they have been
2043 /* Grab the next descriptor number they're advertising, and increment
2044 * the index we've seen. */
2045 if (unlikely(vhost_get_avail(vq, ring_head,
2046 &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2047 vq_err(vq, "Failed to read head: idx %d address %p\n",
2049 &vq->avail->ring[last_avail_idx % vq->num]);
2053 head = vhost16_to_cpu(vq, ring_head);
2055 /* If their number is silly, that's an error. */
2056 if (unlikely(head >= vq->num)) {
2057 vq_err(vq, "Guest says index %u > %u is available",
2062 /* When we start there are none of either input nor output. */
2063 *out_num = *in_num = 0;
2069 unsigned iov_count = *in_num + *out_num;
2070 if (unlikely(i >= vq->num)) {
2071 vq_err(vq, "Desc index is %u > %u, head = %u",
2075 if (unlikely(++found > vq->num)) {
2076 vq_err(vq, "Loop detected: last one at %u "
2077 "vq size %u head %u\n",
2081 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2083 if (unlikely(ret)) {
2084 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2088 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2089 ret = get_indirect(vq, iov, iov_size,
2091 log, log_num, &desc);
2092 if (unlikely(ret < 0)) {
2094 vq_err(vq, "Failure detected "
2095 "in indirect descriptor at idx %d\n", i);
2101 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2102 access = VHOST_ACCESS_WO;
2104 access = VHOST_ACCESS_RO;
2105 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2106 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2107 iov_size - iov_count, access);
2108 if (unlikely(ret < 0)) {
2110 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2114 if (access == VHOST_ACCESS_WO) {
2115 /* If this is an input descriptor,
2116 * increment that count. */
2118 if (unlikely(log)) {
2119 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2120 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2124 /* If it's an output descriptor, they're all supposed
2125 * to come before any input descriptors. */
2126 if (unlikely(*in_num)) {
2127 vq_err(vq, "Descriptor has out after in: "
2133 } while ((i = next_desc(vq, &desc)) != -1);
2135 /* On success, increment avail index. */
2136 vq->last_avail_idx++;
2138 /* Assume notifications from guest are disabled at this point,
2139 * if they aren't we would need to update avail_event index. */
2140 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2143 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2145 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2146 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2148 vq->last_avail_idx -= n;
2150 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2152 /* After we've used one of their buffers, we tell them about it. We'll then
2153 * want to notify the guest, using eventfd. */
2154 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2156 struct vring_used_elem heads = {
2157 cpu_to_vhost32(vq, head),
2158 cpu_to_vhost32(vq, len)
2161 return vhost_add_used_n(vq, &heads, 1);
2163 EXPORT_SYMBOL_GPL(vhost_add_used);
2165 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2166 struct vring_used_elem *heads,
2169 struct vring_used_elem __user *used;
2173 start = vq->last_used_idx & (vq->num - 1);
2174 used = vq->used->ring + start;
2176 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2177 vq_err(vq, "Failed to write used id");
2180 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2181 vq_err(vq, "Failed to write used len");
2184 } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2185 vq_err(vq, "Failed to write used");
2188 if (unlikely(vq->log_used)) {
2189 /* Make sure data is seen before log. */
2191 /* Log used ring entry write. */
2192 log_write(vq->log_base,
2194 ((void __user *)used - (void __user *)vq->used),
2195 count * sizeof *used);
2197 old = vq->last_used_idx;
2198 new = (vq->last_used_idx += count);
2199 /* If the driver never bothers to signal in a very long while,
2200 * used index might wrap around. If that happens, invalidate
2201 * signalled_used index we stored. TODO: make sure driver
2202 * signals at least once in 2^16 and remove this. */
2203 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2204 vq->signalled_used_valid = false;
2208 /* After we've used one of their buffers, we tell them about it. We'll then
2209 * want to notify the guest, using eventfd. */
2210 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2215 start = vq->last_used_idx & (vq->num - 1);
2216 n = vq->num - start;
2218 r = __vhost_add_used_n(vq, heads, n);
2224 r = __vhost_add_used_n(vq, heads, count);
2226 /* Make sure buffer is written before we update index. */
2228 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2230 vq_err(vq, "Failed to increment used idx");
2233 if (unlikely(vq->log_used)) {
2234 /* Log used index update. */
2235 log_write(vq->log_base,
2236 vq->log_addr + offsetof(struct vring_used, idx),
2237 sizeof vq->used->idx);
2239 eventfd_signal(vq->log_ctx, 1);
2243 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2245 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2250 /* Flush out used index updates. This is paired
2251 * with the barrier that the Guest executes when enabling
2255 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2256 unlikely(vq->avail_idx == vq->last_avail_idx))
2259 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2261 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2262 vq_err(vq, "Failed to get flags");
2265 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2267 old = vq->signalled_used;
2268 v = vq->signalled_used_valid;
2269 new = vq->signalled_used = vq->last_used_idx;
2270 vq->signalled_used_valid = true;
2275 if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2276 vq_err(vq, "Failed to get used event idx");
2279 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2282 /* This actually signals the guest, using eventfd. */
2283 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2285 /* Signal the Guest tell them we used something up. */
2286 if (vq->call_ctx && vhost_notify(dev, vq))
2287 eventfd_signal(vq->call_ctx, 1);
2289 EXPORT_SYMBOL_GPL(vhost_signal);
2291 /* And here's the combo meal deal. Supersize me! */
2292 void vhost_add_used_and_signal(struct vhost_dev *dev,
2293 struct vhost_virtqueue *vq,
2294 unsigned int head, int len)
2296 vhost_add_used(vq, head, len);
2297 vhost_signal(dev, vq);
2299 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2301 /* multi-buffer version of vhost_add_used_and_signal */
2302 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2303 struct vhost_virtqueue *vq,
2304 struct vring_used_elem *heads, unsigned count)
2306 vhost_add_used_n(vq, heads, count);
2307 vhost_signal(dev, vq);
2309 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2311 /* return true if we're sure that avaiable ring is empty */
2312 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2314 __virtio16 avail_idx;
2317 if (vq->avail_idx != vq->last_avail_idx)
2320 r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2323 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2325 return vq->avail_idx == vq->last_avail_idx;
2327 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2329 /* OK, now we need to know about added descriptors. */
2330 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2332 __virtio16 avail_idx;
2335 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2337 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2338 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2339 r = vhost_update_used_flags(vq);
2341 vq_err(vq, "Failed to enable notification at %p: %d\n",
2342 &vq->used->flags, r);
2346 r = vhost_update_avail_event(vq, vq->avail_idx);
2348 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2349 vhost_avail_event(vq), r);
2353 /* They could have slipped one in as we were doing that: make
2354 * sure it's written, then check again. */
2356 r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2358 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2359 &vq->avail->idx, r);
2363 return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2365 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2367 /* We don't need to be notified again. */
2368 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2372 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2374 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2375 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2376 r = vhost_update_used_flags(vq);
2378 vq_err(vq, "Failed to enable notification at %p: %d\n",
2379 &vq->used->flags, r);
2382 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2384 /* Create a new message. */
2385 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2387 struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2391 /* Make sure all padding within the structure is initialized. */
2392 memset(&node->msg, 0, sizeof node->msg);
2394 node->msg.type = type;
2397 EXPORT_SYMBOL_GPL(vhost_new_msg);
2399 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2400 struct vhost_msg_node *node)
2402 spin_lock(&dev->iotlb_lock);
2403 list_add_tail(&node->node, head);
2404 spin_unlock(&dev->iotlb_lock);
2406 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2408 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2410 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2411 struct list_head *head)
2413 struct vhost_msg_node *node = NULL;
2415 spin_lock(&dev->iotlb_lock);
2416 if (!list_empty(head)) {
2417 node = list_first_entry(head, struct vhost_msg_node,
2419 list_del(&node->node);
2421 spin_unlock(&dev->iotlb_lock);
2425 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2428 static int __init vhost_init(void)
2433 static void __exit vhost_exit(void)
2437 module_init(vhost_init);
2438 module_exit(vhost_exit);
2440 MODULE_VERSION("0.0.1");
2441 MODULE_LICENSE("GPL v2");
2442 MODULE_AUTHOR("Michael S. Tsirkin");
2443 MODULE_DESCRIPTION("Host kernel accelerator for virtio");