1 /* Copyright (C) 2009 Red Hat, Inc.
2 * Copyright (C) 2006 Rusty Russell IBM Corporation
4 * Author: Michael S. Tsirkin <mst@redhat.com>
6 * Inspiration, some code, and most witty comments come from
7 * Documentation/virtual/lguest/lguest.c, by Rusty Russell
9 * This work is licensed under the terms of the GNU GPL, version 2.
11 * Generic code for virtio server in host kernel.
14 #include <linux/eventfd.h>
15 #include <linux/vhost.h>
16 #include <linux/uio.h>
18 #include <linux/mmu_context.h>
19 #include <linux/miscdevice.h>
20 #include <linux/mutex.h>
21 #include <linux/poll.h>
22 #include <linux/file.h>
23 #include <linux/highmem.h>
24 #include <linux/slab.h>
25 #include <linux/vmalloc.h>
26 #include <linux/kthread.h>
27 #include <linux/cgroup.h>
28 #include <linux/module.h>
29 #include <linux/sort.h>
30 #include <linux/sched/mm.h>
31 #include <linux/sched/signal.h>
32 #include <linux/interval_tree_generic.h>
33 #include <linux/nospec.h>
37 static ushort max_mem_regions = 64;
38 module_param(max_mem_regions, ushort, 0444);
39 MODULE_PARM_DESC(max_mem_regions,
40 "Maximum number of memory regions in memory map. (default: 64)");
41 static int max_iotlb_entries = 2048;
42 module_param(max_iotlb_entries, int, 0444);
43 MODULE_PARM_DESC(max_iotlb_entries,
44 "Maximum number of iotlb entries. (default: 2048)");
47 VHOST_MEMORY_F_LOG = 0x1,
50 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
51 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
53 INTERVAL_TREE_DEFINE(struct vhost_umem_node,
54 rb, __u64, __subtree_last,
55 START, LAST, static inline, vhost_umem_interval_tree);
57 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
58 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
60 vq->user_be = !virtio_legacy_is_little_endian();
63 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
68 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
73 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
75 struct vhost_vring_state s;
80 if (copy_from_user(&s, argp, sizeof(s)))
83 if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
84 s.num != VHOST_VRING_BIG_ENDIAN)
87 if (s.num == VHOST_VRING_BIG_ENDIAN)
88 vhost_enable_cross_endian_big(vq);
90 vhost_enable_cross_endian_little(vq);
95 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
98 struct vhost_vring_state s = {
103 if (copy_to_user(argp, &s, sizeof(s)))
109 static void vhost_init_is_le(struct vhost_virtqueue *vq)
111 /* Note for legacy virtio: user_be is initialized at reset time
112 * according to the host endianness. If userspace does not set an
113 * explicit endianness, the default behavior is native endian, as
114 * expected by legacy virtio.
116 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
119 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
123 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
128 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
134 static void vhost_init_is_le(struct vhost_virtqueue *vq)
136 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
137 || virtio_legacy_is_little_endian();
139 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
141 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
143 vhost_init_is_le(vq);
146 struct vhost_flush_struct {
147 struct vhost_work work;
148 struct completion wait_event;
151 static void vhost_flush_work(struct vhost_work *work)
153 struct vhost_flush_struct *s;
155 s = container_of(work, struct vhost_flush_struct, work);
156 complete(&s->wait_event);
159 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
162 struct vhost_poll *poll;
164 poll = container_of(pt, struct vhost_poll, table);
166 add_wait_queue(wqh, &poll->wait);
169 static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
172 struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
174 if (!(key_to_poll(key) & poll->mask))
177 vhost_poll_queue(poll);
181 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
183 clear_bit(VHOST_WORK_QUEUED, &work->flags);
186 EXPORT_SYMBOL_GPL(vhost_work_init);
188 /* Init poll structure */
189 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
190 __poll_t mask, struct vhost_dev *dev)
192 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
193 init_poll_funcptr(&poll->table, vhost_poll_func);
198 vhost_work_init(&poll->work, fn);
200 EXPORT_SYMBOL_GPL(vhost_poll_init);
202 /* Start polling a file. We add ourselves to file's wait queue. The caller must
203 * keep a reference to a file until after vhost_poll_stop is called. */
204 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
212 mask = vfs_poll(file, &poll->table);
214 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
215 if (mask & EPOLLERR) {
216 vhost_poll_stop(poll);
222 EXPORT_SYMBOL_GPL(vhost_poll_start);
224 /* Stop polling a file. After this function returns, it becomes safe to drop the
225 * file reference. You must also flush afterwards. */
226 void vhost_poll_stop(struct vhost_poll *poll)
229 remove_wait_queue(poll->wqh, &poll->wait);
233 EXPORT_SYMBOL_GPL(vhost_poll_stop);
235 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
237 struct vhost_flush_struct flush;
240 init_completion(&flush.wait_event);
241 vhost_work_init(&flush.work, vhost_flush_work);
243 vhost_work_queue(dev, &flush.work);
244 wait_for_completion(&flush.wait_event);
247 EXPORT_SYMBOL_GPL(vhost_work_flush);
249 /* Flush any work that has been scheduled. When calling this, don't hold any
250 * locks that are also used by the callback. */
251 void vhost_poll_flush(struct vhost_poll *poll)
253 vhost_work_flush(poll->dev, &poll->work);
255 EXPORT_SYMBOL_GPL(vhost_poll_flush);
257 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
262 if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
263 /* We can only add the work to the list after we're
264 * sure it was not in the list.
265 * test_and_set_bit() implies a memory barrier.
267 llist_add(&work->node, &dev->work_list);
268 wake_up_process(dev->worker);
271 EXPORT_SYMBOL_GPL(vhost_work_queue);
273 /* A lockless hint for busy polling code to exit the loop */
274 bool vhost_has_work(struct vhost_dev *dev)
276 return !llist_empty(&dev->work_list);
278 EXPORT_SYMBOL_GPL(vhost_has_work);
280 void vhost_poll_queue(struct vhost_poll *poll)
282 vhost_work_queue(poll->dev, &poll->work);
284 EXPORT_SYMBOL_GPL(vhost_poll_queue);
286 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
290 for (j = 0; j < VHOST_NUM_ADDRS; j++)
291 vq->meta_iotlb[j] = NULL;
294 static void vhost_vq_meta_reset(struct vhost_dev *d)
298 for (i = 0; i < d->nvqs; ++i)
299 __vhost_vq_meta_reset(d->vqs[i]);
302 static void vhost_vq_reset(struct vhost_dev *dev,
303 struct vhost_virtqueue *vq)
309 vq->last_avail_idx = 0;
311 vq->last_used_idx = 0;
312 vq->signalled_used = 0;
313 vq->signalled_used_valid = false;
315 vq->log_used = false;
316 vq->log_addr = -1ull;
317 vq->private_data = NULL;
318 vq->acked_features = 0;
319 vq->acked_backend_features = 0;
321 vq->error_ctx = NULL;
325 vhost_reset_is_le(vq);
326 vhost_disable_cross_endian(vq);
327 vq->busyloop_timeout = 0;
330 __vhost_vq_meta_reset(vq);
333 static int vhost_worker(void *data)
335 struct vhost_dev *dev = data;
336 struct vhost_work *work, *work_next;
337 struct llist_node *node;
338 mm_segment_t oldfs = get_fs();
344 /* mb paired w/ kthread_stop */
345 set_current_state(TASK_INTERRUPTIBLE);
347 if (kthread_should_stop()) {
348 __set_current_state(TASK_RUNNING);
352 node = llist_del_all(&dev->work_list);
356 node = llist_reverse_order(node);
357 /* make sure flag is seen after deletion */
359 llist_for_each_entry_safe(work, work_next, node, node) {
360 clear_bit(VHOST_WORK_QUEUED, &work->flags);
361 __set_current_state(TASK_RUNNING);
372 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
382 /* Helper to allocate iovec buffers for all vqs. */
383 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
385 struct vhost_virtqueue *vq;
388 for (i = 0; i < dev->nvqs; ++i) {
390 vq->indirect = kmalloc_array(UIO_MAXIOV,
391 sizeof(*vq->indirect),
393 vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
395 vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
397 if (!vq->indirect || !vq->log || !vq->heads)
404 vhost_vq_free_iovecs(dev->vqs[i]);
408 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
412 for (i = 0; i < dev->nvqs; ++i)
413 vhost_vq_free_iovecs(dev->vqs[i]);
416 bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
417 int pkts, int total_len)
419 struct vhost_dev *dev = vq->dev;
421 if ((dev->byte_weight && total_len >= dev->byte_weight) ||
422 pkts >= dev->weight) {
423 vhost_poll_queue(&vq->poll);
429 EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
431 void vhost_dev_init(struct vhost_dev *dev,
432 struct vhost_virtqueue **vqs, int nvqs,
433 int iov_limit, int weight, int byte_weight)
435 struct vhost_virtqueue *vq;
440 mutex_init(&dev->mutex);
446 dev->iov_limit = iov_limit;
447 dev->weight = weight;
448 dev->byte_weight = byte_weight;
449 init_llist_head(&dev->work_list);
450 init_waitqueue_head(&dev->wait);
451 INIT_LIST_HEAD(&dev->read_list);
452 INIT_LIST_HEAD(&dev->pending_list);
453 spin_lock_init(&dev->iotlb_lock);
456 for (i = 0; i < dev->nvqs; ++i) {
462 mutex_init(&vq->mutex);
463 vhost_vq_reset(dev, vq);
465 vhost_poll_init(&vq->poll, vq->handle_kick,
469 EXPORT_SYMBOL_GPL(vhost_dev_init);
471 /* Caller should have device mutex */
472 long vhost_dev_check_owner(struct vhost_dev *dev)
474 /* Are you the owner? If not, I don't think you mean to do that */
475 return dev->mm == current->mm ? 0 : -EPERM;
477 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
479 struct vhost_attach_cgroups_struct {
480 struct vhost_work work;
481 struct task_struct *owner;
485 static void vhost_attach_cgroups_work(struct vhost_work *work)
487 struct vhost_attach_cgroups_struct *s;
489 s = container_of(work, struct vhost_attach_cgroups_struct, work);
490 s->ret = cgroup_attach_task_all(s->owner, current);
493 static int vhost_attach_cgroups(struct vhost_dev *dev)
495 struct vhost_attach_cgroups_struct attach;
497 attach.owner = current;
498 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
499 vhost_work_queue(dev, &attach.work);
500 vhost_work_flush(dev, &attach.work);
504 /* Caller should have device mutex */
505 bool vhost_dev_has_owner(struct vhost_dev *dev)
509 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
511 /* Caller should have device mutex */
512 long vhost_dev_set_owner(struct vhost_dev *dev)
514 struct task_struct *worker;
517 /* Is there an owner already? */
518 if (vhost_dev_has_owner(dev)) {
523 /* No owner, become one */
524 dev->mm = get_task_mm(current);
525 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
526 if (IS_ERR(worker)) {
527 err = PTR_ERR(worker);
531 dev->worker = worker;
532 wake_up_process(worker); /* avoid contributing to loadavg */
534 err = vhost_attach_cgroups(dev);
538 err = vhost_dev_alloc_iovecs(dev);
544 kthread_stop(worker);
553 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
555 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
557 return kvzalloc(sizeof(struct vhost_umem), GFP_KERNEL);
559 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
561 /* Caller should have device mutex */
562 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
566 vhost_dev_cleanup(dev);
568 /* Restore memory to default empty mapping. */
569 INIT_LIST_HEAD(&umem->umem_list);
571 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
572 * VQs aren't running.
574 for (i = 0; i < dev->nvqs; ++i)
575 dev->vqs[i]->umem = umem;
577 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
579 void vhost_dev_stop(struct vhost_dev *dev)
583 for (i = 0; i < dev->nvqs; ++i) {
584 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
585 vhost_poll_stop(&dev->vqs[i]->poll);
586 vhost_poll_flush(&dev->vqs[i]->poll);
590 EXPORT_SYMBOL_GPL(vhost_dev_stop);
592 static void vhost_umem_free(struct vhost_umem *umem,
593 struct vhost_umem_node *node)
595 vhost_umem_interval_tree_remove(node, &umem->umem_tree);
596 list_del(&node->link);
601 static void vhost_umem_clean(struct vhost_umem *umem)
603 struct vhost_umem_node *node, *tmp;
608 list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
609 vhost_umem_free(umem, node);
614 static void vhost_clear_msg(struct vhost_dev *dev)
616 struct vhost_msg_node *node, *n;
618 spin_lock(&dev->iotlb_lock);
620 list_for_each_entry_safe(node, n, &dev->read_list, node) {
621 list_del(&node->node);
625 list_for_each_entry_safe(node, n, &dev->pending_list, node) {
626 list_del(&node->node);
630 spin_unlock(&dev->iotlb_lock);
633 void vhost_dev_cleanup(struct vhost_dev *dev)
637 for (i = 0; i < dev->nvqs; ++i) {
638 if (dev->vqs[i]->error_ctx)
639 eventfd_ctx_put(dev->vqs[i]->error_ctx);
640 if (dev->vqs[i]->kick)
641 fput(dev->vqs[i]->kick);
642 if (dev->vqs[i]->call_ctx)
643 eventfd_ctx_put(dev->vqs[i]->call_ctx);
644 vhost_vq_reset(dev, dev->vqs[i]);
646 vhost_dev_free_iovecs(dev);
648 eventfd_ctx_put(dev->log_ctx);
650 /* No one will access memory at this point */
651 vhost_umem_clean(dev->umem);
653 vhost_umem_clean(dev->iotlb);
655 vhost_clear_msg(dev);
656 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
657 WARN_ON(!llist_empty(&dev->work_list));
659 kthread_stop(dev->worker);
666 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
668 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
670 u64 a = addr / VHOST_PAGE_SIZE / 8;
672 /* Make sure 64 bit math will not overflow. */
673 if (a > ULONG_MAX - (unsigned long)log_base ||
674 a + (unsigned long)log_base > ULONG_MAX)
677 return access_ok(log_base + a,
678 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
681 static bool vhost_overflow(u64 uaddr, u64 size)
683 /* Make sure 64 bit math will not overflow. */
684 return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
687 /* Caller should have vq mutex and device mutex. */
688 static bool vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
691 struct vhost_umem_node *node;
696 list_for_each_entry(node, &umem->umem_list, link) {
697 unsigned long a = node->userspace_addr;
699 if (vhost_overflow(node->userspace_addr, node->size))
703 if (!access_ok((void __user *)a,
706 else if (log_all && !log_access_ok(log_base,
714 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
715 u64 addr, unsigned int size,
718 const struct vhost_umem_node *node = vq->meta_iotlb[type];
723 return (void *)(uintptr_t)(node->userspace_addr + addr - node->start);
726 /* Can we switch to this memory table? */
727 /* Caller should have device mutex but not vq mutex */
728 static bool memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
733 for (i = 0; i < d->nvqs; ++i) {
737 mutex_lock(&d->vqs[i]->mutex);
738 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
739 /* If ring is inactive, will check when it's enabled. */
740 if (d->vqs[i]->private_data)
741 ok = vq_memory_access_ok(d->vqs[i]->log_base,
745 mutex_unlock(&d->vqs[i]->mutex);
752 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
753 struct iovec iov[], int iov_size, int access);
755 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
756 const void *from, unsigned size)
761 return __copy_to_user(to, from, size);
763 /* This function should be called after iotlb
764 * prefetch, which means we're sure that all vq
765 * could be access through iotlb. So -EAGAIN should
766 * not happen in this case.
769 void __user *uaddr = vhost_vq_meta_fetch(vq,
770 (u64)(uintptr_t)to, size,
774 return __copy_to_user(uaddr, from, size);
776 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
777 ARRAY_SIZE(vq->iotlb_iov),
781 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
782 ret = copy_to_iter(from, size, &t);
790 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
791 void __user *from, unsigned size)
796 return __copy_from_user(to, from, size);
798 /* This function should be called after iotlb
799 * prefetch, which means we're sure that vq
800 * could be access through iotlb. So -EAGAIN should
801 * not happen in this case.
803 void __user *uaddr = vhost_vq_meta_fetch(vq,
804 (u64)(uintptr_t)from, size,
809 return __copy_from_user(to, uaddr, size);
811 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
812 ARRAY_SIZE(vq->iotlb_iov),
815 vq_err(vq, "IOTLB translation failure: uaddr "
816 "%p size 0x%llx\n", from,
817 (unsigned long long) size);
820 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
821 ret = copy_from_iter(to, size, &f);
830 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
831 void __user *addr, unsigned int size,
836 ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
837 ARRAY_SIZE(vq->iotlb_iov),
840 vq_err(vq, "IOTLB translation failure: uaddr "
841 "%p size 0x%llx\n", addr,
842 (unsigned long long) size);
846 if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
847 vq_err(vq, "Non atomic userspace memory access: uaddr "
848 "%p size 0x%llx\n", addr,
849 (unsigned long long) size);
853 return vq->iotlb_iov[0].iov_base;
856 /* This function should be called after iotlb
857 * prefetch, which means we're sure that vq
858 * could be access through iotlb. So -EAGAIN should
859 * not happen in this case.
861 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
862 void *addr, unsigned int size,
865 void __user *uaddr = vhost_vq_meta_fetch(vq,
866 (u64)(uintptr_t)addr, size, type);
870 return __vhost_get_user_slow(vq, addr, size, type);
873 #define vhost_put_user(vq, x, ptr) \
877 ret = __put_user(x, ptr); \
879 __typeof__(ptr) to = \
880 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
881 sizeof(*ptr), VHOST_ADDR_USED); \
883 ret = __put_user(x, to); \
890 #define vhost_get_user(vq, x, ptr, type) \
894 ret = __get_user(x, ptr); \
896 __typeof__(ptr) from = \
897 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
901 ret = __get_user(x, from); \
908 #define vhost_get_avail(vq, x, ptr) \
909 vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
911 #define vhost_get_used(vq, x, ptr) \
912 vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
914 static void vhost_dev_lock_vqs(struct vhost_dev *d)
917 for (i = 0; i < d->nvqs; ++i)
918 mutex_lock_nested(&d->vqs[i]->mutex, i);
921 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
924 for (i = 0; i < d->nvqs; ++i)
925 mutex_unlock(&d->vqs[i]->mutex);
928 static int vhost_new_umem_range(struct vhost_umem *umem,
929 u64 start, u64 size, u64 end,
930 u64 userspace_addr, int perm)
932 struct vhost_umem_node *tmp, *node;
937 node = kmalloc(sizeof(*node), GFP_ATOMIC);
941 if (umem->numem == max_iotlb_entries) {
942 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
943 vhost_umem_free(umem, tmp);
949 node->userspace_addr = userspace_addr;
951 INIT_LIST_HEAD(&node->link);
952 list_add_tail(&node->link, &umem->umem_list);
953 vhost_umem_interval_tree_insert(node, &umem->umem_tree);
959 static void vhost_del_umem_range(struct vhost_umem *umem,
962 struct vhost_umem_node *node;
964 while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
966 vhost_umem_free(umem, node);
969 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
970 struct vhost_iotlb_msg *msg)
972 struct vhost_msg_node *node, *n;
974 spin_lock(&d->iotlb_lock);
976 list_for_each_entry_safe(node, n, &d->pending_list, node) {
977 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
978 if (msg->iova <= vq_msg->iova &&
979 msg->iova + msg->size - 1 >= vq_msg->iova &&
980 vq_msg->type == VHOST_IOTLB_MISS) {
981 vhost_poll_queue(&node->vq->poll);
982 list_del(&node->node);
987 spin_unlock(&d->iotlb_lock);
990 static bool umem_access_ok(u64 uaddr, u64 size, int access)
992 unsigned long a = uaddr;
994 /* Make sure 64 bit math will not overflow. */
995 if (vhost_overflow(uaddr, size))
998 if ((access & VHOST_ACCESS_RO) &&
999 !access_ok((void __user *)a, size))
1001 if ((access & VHOST_ACCESS_WO) &&
1002 !access_ok((void __user *)a, size))
1007 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
1008 struct vhost_iotlb_msg *msg)
1012 mutex_lock(&dev->mutex);
1013 vhost_dev_lock_vqs(dev);
1014 switch (msg->type) {
1015 case VHOST_IOTLB_UPDATE:
1020 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1024 vhost_vq_meta_reset(dev);
1025 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
1026 msg->iova + msg->size - 1,
1027 msg->uaddr, msg->perm)) {
1031 vhost_iotlb_notify_vq(dev, msg);
1033 case VHOST_IOTLB_INVALIDATE:
1038 vhost_vq_meta_reset(dev);
1039 vhost_del_umem_range(dev->iotlb, msg->iova,
1040 msg->iova + msg->size - 1);
1047 vhost_dev_unlock_vqs(dev);
1048 mutex_unlock(&dev->mutex);
1052 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1053 struct iov_iter *from)
1055 struct vhost_iotlb_msg msg;
1059 ret = copy_from_iter(&type, sizeof(type), from);
1060 if (ret != sizeof(type)) {
1066 case VHOST_IOTLB_MSG:
1067 /* There maybe a hole after type for V1 message type,
1070 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1072 case VHOST_IOTLB_MSG_V2:
1073 offset = sizeof(__u32);
1080 iov_iter_advance(from, offset);
1081 ret = copy_from_iter(&msg, sizeof(msg), from);
1082 if (ret != sizeof(msg)) {
1086 if (vhost_process_iotlb_msg(dev, &msg)) {
1091 ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1092 sizeof(struct vhost_msg_v2);
1096 EXPORT_SYMBOL(vhost_chr_write_iter);
1098 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1103 poll_wait(file, &dev->wait, wait);
1105 if (!list_empty(&dev->read_list))
1106 mask |= EPOLLIN | EPOLLRDNORM;
1110 EXPORT_SYMBOL(vhost_chr_poll);
1112 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1116 struct vhost_msg_node *node;
1118 unsigned size = sizeof(struct vhost_msg);
1120 if (iov_iter_count(to) < size)
1125 prepare_to_wait(&dev->wait, &wait,
1126 TASK_INTERRUPTIBLE);
1128 node = vhost_dequeue_msg(dev, &dev->read_list);
1135 if (signal_pending(current)) {
1148 finish_wait(&dev->wait, &wait);
1151 struct vhost_iotlb_msg *msg;
1152 void *start = &node->msg;
1154 switch (node->msg.type) {
1155 case VHOST_IOTLB_MSG:
1156 size = sizeof(node->msg);
1157 msg = &node->msg.iotlb;
1159 case VHOST_IOTLB_MSG_V2:
1160 size = sizeof(node->msg_v2);
1161 msg = &node->msg_v2.iotlb;
1168 ret = copy_to_iter(start, size, to);
1169 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1173 vhost_enqueue_msg(dev, &dev->pending_list, node);
1178 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1180 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1182 struct vhost_dev *dev = vq->dev;
1183 struct vhost_msg_node *node;
1184 struct vhost_iotlb_msg *msg;
1185 bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1187 node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1192 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1193 msg = &node->msg_v2.iotlb;
1195 msg = &node->msg.iotlb;
1198 msg->type = VHOST_IOTLB_MISS;
1202 vhost_enqueue_msg(dev, &dev->read_list, node);
1207 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1208 struct vring_desc __user *desc,
1209 struct vring_avail __user *avail,
1210 struct vring_used __user *used)
1213 size_t s __maybe_unused = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1215 return access_ok(desc, num * sizeof *desc) &&
1217 sizeof *avail + num * sizeof *avail->ring + s) &&
1219 sizeof *used + num * sizeof *used->ring + s);
1222 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1223 const struct vhost_umem_node *node,
1226 int access = (type == VHOST_ADDR_USED) ?
1227 VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1229 if (likely(node->perm & access))
1230 vq->meta_iotlb[type] = node;
1233 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1234 int access, u64 addr, u64 len, int type)
1236 const struct vhost_umem_node *node;
1237 struct vhost_umem *umem = vq->iotlb;
1238 u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1240 if (vhost_vq_meta_fetch(vq, addr, len, type))
1244 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1247 if (node == NULL || node->start > addr) {
1248 vhost_iotlb_miss(vq, addr, access);
1250 } else if (!(node->perm & access)) {
1251 /* Report the possible access violation by
1252 * request another translation from userspace.
1257 size = node->size - addr + node->start;
1259 if (orig_addr == addr && size >= len)
1260 vhost_vq_meta_update(vq, node, type);
1269 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1271 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1272 unsigned int num = vq->num;
1277 return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1278 num * sizeof(*vq->desc), VHOST_ADDR_DESC) &&
1279 iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1281 num * sizeof(*vq->avail->ring) + s,
1282 VHOST_ADDR_AVAIL) &&
1283 iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1285 num * sizeof(*vq->used->ring) + s,
1288 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1290 /* Can we log writes? */
1291 /* Caller should have device mutex but not vq mutex */
1292 bool vhost_log_access_ok(struct vhost_dev *dev)
1294 return memory_access_ok(dev, dev->umem, 1);
1296 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1298 /* Verify access for write logging. */
1299 /* Caller should have vq mutex and device mutex */
1300 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1301 void __user *log_base)
1303 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1305 return vq_memory_access_ok(log_base, vq->umem,
1306 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1307 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1309 vq->num * sizeof *vq->used->ring + s));
1312 /* Can we start vq? */
1313 /* Caller should have vq mutex and device mutex */
1314 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1316 if (!vq_log_access_ok(vq, vq->log_base))
1319 /* Access validation occurs at prefetch time with IOTLB */
1323 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1325 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1327 static struct vhost_umem *vhost_umem_alloc(void)
1329 struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1334 umem->umem_tree = RB_ROOT_CACHED;
1336 INIT_LIST_HEAD(&umem->umem_list);
1341 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1343 struct vhost_memory mem, *newmem;
1344 struct vhost_memory_region *region;
1345 struct vhost_umem *newumem, *oldumem;
1346 unsigned long size = offsetof(struct vhost_memory, regions);
1349 if (copy_from_user(&mem, m, size))
1353 if (mem.nregions > max_mem_regions)
1355 newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1360 memcpy(newmem, &mem, size);
1361 if (copy_from_user(newmem->regions, m->regions,
1362 mem.nregions * sizeof *m->regions)) {
1367 newumem = vhost_umem_alloc();
1373 for (region = newmem->regions;
1374 region < newmem->regions + mem.nregions;
1376 if (vhost_new_umem_range(newumem,
1377 region->guest_phys_addr,
1378 region->memory_size,
1379 region->guest_phys_addr +
1380 region->memory_size - 1,
1381 region->userspace_addr,
1386 if (!memory_access_ok(d, newumem, 0))
1392 /* All memory accesses are done under some VQ mutex. */
1393 for (i = 0; i < d->nvqs; ++i) {
1394 mutex_lock(&d->vqs[i]->mutex);
1395 d->vqs[i]->umem = newumem;
1396 mutex_unlock(&d->vqs[i]->mutex);
1400 vhost_umem_clean(oldumem);
1404 vhost_umem_clean(newumem);
1409 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1411 struct file *eventfp, *filep = NULL;
1412 bool pollstart = false, pollstop = false;
1413 struct eventfd_ctx *ctx = NULL;
1414 u32 __user *idxp = argp;
1415 struct vhost_virtqueue *vq;
1416 struct vhost_vring_state s;
1417 struct vhost_vring_file f;
1418 struct vhost_vring_addr a;
1422 r = get_user(idx, idxp);
1428 idx = array_index_nospec(idx, d->nvqs);
1431 mutex_lock(&vq->mutex);
1434 case VHOST_SET_VRING_NUM:
1435 /* Resizing ring with an active backend?
1436 * You don't want to do that. */
1437 if (vq->private_data) {
1441 if (copy_from_user(&s, argp, sizeof s)) {
1445 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1451 case VHOST_SET_VRING_BASE:
1452 /* Moving base with an active backend?
1453 * You don't want to do that. */
1454 if (vq->private_data) {
1458 if (copy_from_user(&s, argp, sizeof s)) {
1462 if (s.num > 0xffff) {
1466 vq->last_avail_idx = s.num;
1467 /* Forget the cached index value. */
1468 vq->avail_idx = vq->last_avail_idx;
1470 case VHOST_GET_VRING_BASE:
1472 s.num = vq->last_avail_idx;
1473 if (copy_to_user(argp, &s, sizeof s))
1476 case VHOST_SET_VRING_ADDR:
1477 if (copy_from_user(&a, argp, sizeof a)) {
1481 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1485 /* For 32bit, verify that the top 32bits of the user
1486 data are set to zero. */
1487 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1488 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1489 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1494 /* Make sure it's safe to cast pointers to vring types. */
1495 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1496 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1497 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1498 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1499 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1504 /* We only verify access here if backend is configured.
1505 * If it is not, we don't as size might not have been setup.
1506 * We will verify when backend is configured. */
1507 if (vq->private_data) {
1508 if (!vq_access_ok(vq, vq->num,
1509 (void __user *)(unsigned long)a.desc_user_addr,
1510 (void __user *)(unsigned long)a.avail_user_addr,
1511 (void __user *)(unsigned long)a.used_user_addr)) {
1516 /* Also validate log access for used ring if enabled. */
1517 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1518 !log_access_ok(vq->log_base, a.log_guest_addr,
1520 vq->num * sizeof *vq->used->ring)) {
1526 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1527 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1528 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1529 vq->log_addr = a.log_guest_addr;
1530 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1532 case VHOST_SET_VRING_KICK:
1533 if (copy_from_user(&f, argp, sizeof f)) {
1537 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1538 if (IS_ERR(eventfp)) {
1539 r = PTR_ERR(eventfp);
1542 if (eventfp != vq->kick) {
1543 pollstop = (filep = vq->kick) != NULL;
1544 pollstart = (vq->kick = eventfp) != NULL;
1548 case VHOST_SET_VRING_CALL:
1549 if (copy_from_user(&f, argp, sizeof f)) {
1553 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1558 swap(ctx, vq->call_ctx);
1560 case VHOST_SET_VRING_ERR:
1561 if (copy_from_user(&f, argp, sizeof f)) {
1565 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1570 swap(ctx, vq->error_ctx);
1572 case VHOST_SET_VRING_ENDIAN:
1573 r = vhost_set_vring_endian(vq, argp);
1575 case VHOST_GET_VRING_ENDIAN:
1576 r = vhost_get_vring_endian(vq, idx, argp);
1578 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1579 if (copy_from_user(&s, argp, sizeof(s))) {
1583 vq->busyloop_timeout = s.num;
1585 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1587 s.num = vq->busyloop_timeout;
1588 if (copy_to_user(argp, &s, sizeof(s)))
1595 if (pollstop && vq->handle_kick)
1596 vhost_poll_stop(&vq->poll);
1598 if (!IS_ERR_OR_NULL(ctx))
1599 eventfd_ctx_put(ctx);
1603 if (pollstart && vq->handle_kick)
1604 r = vhost_poll_start(&vq->poll, vq->kick);
1606 mutex_unlock(&vq->mutex);
1608 if (pollstop && vq->handle_kick)
1609 vhost_poll_flush(&vq->poll);
1612 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1614 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1616 struct vhost_umem *niotlb, *oiotlb;
1619 niotlb = vhost_umem_alloc();
1626 for (i = 0; i < d->nvqs; ++i) {
1627 struct vhost_virtqueue *vq = d->vqs[i];
1629 mutex_lock(&vq->mutex);
1631 __vhost_vq_meta_reset(vq);
1632 mutex_unlock(&vq->mutex);
1635 vhost_umem_clean(oiotlb);
1639 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1641 /* Caller must have device mutex */
1642 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1644 struct eventfd_ctx *ctx;
1649 /* If you are not the owner, you can become one */
1650 if (ioctl == VHOST_SET_OWNER) {
1651 r = vhost_dev_set_owner(d);
1655 /* You must be the owner to do anything else */
1656 r = vhost_dev_check_owner(d);
1661 case VHOST_SET_MEM_TABLE:
1662 r = vhost_set_memory(d, argp);
1664 case VHOST_SET_LOG_BASE:
1665 if (copy_from_user(&p, argp, sizeof p)) {
1669 if ((u64)(unsigned long)p != p) {
1673 for (i = 0; i < d->nvqs; ++i) {
1674 struct vhost_virtqueue *vq;
1675 void __user *base = (void __user *)(unsigned long)p;
1677 mutex_lock(&vq->mutex);
1678 /* If ring is inactive, will check when it's enabled. */
1679 if (vq->private_data && !vq_log_access_ok(vq, base))
1682 vq->log_base = base;
1683 mutex_unlock(&vq->mutex);
1686 case VHOST_SET_LOG_FD:
1687 r = get_user(fd, (int __user *)argp);
1690 ctx = fd == -1 ? NULL : eventfd_ctx_fdget(fd);
1695 swap(ctx, d->log_ctx);
1696 for (i = 0; i < d->nvqs; ++i) {
1697 mutex_lock(&d->vqs[i]->mutex);
1698 d->vqs[i]->log_ctx = d->log_ctx;
1699 mutex_unlock(&d->vqs[i]->mutex);
1702 eventfd_ctx_put(ctx);
1711 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1713 /* TODO: This is really inefficient. We need something like get_user()
1714 * (instruction directly accesses the data, with an exception table entry
1715 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1717 static int set_bit_to_user(int nr, void __user *addr)
1719 unsigned long log = (unsigned long)addr;
1722 int bit = nr + (log % PAGE_SIZE) * 8;
1725 r = get_user_pages_fast(log, 1, FOLL_WRITE, &page);
1729 base = kmap_atomic(page);
1731 kunmap_atomic(base);
1732 set_page_dirty_lock(page);
1737 static int log_write(void __user *log_base,
1738 u64 write_address, u64 write_length)
1740 u64 write_page = write_address / VHOST_PAGE_SIZE;
1745 write_length += write_address % VHOST_PAGE_SIZE;
1747 u64 base = (u64)(unsigned long)log_base;
1748 u64 log = base + write_page / 8;
1749 int bit = write_page % 8;
1750 if ((u64)(unsigned long)log != log)
1752 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1755 if (write_length <= VHOST_PAGE_SIZE)
1757 write_length -= VHOST_PAGE_SIZE;
1763 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1765 struct vhost_umem *umem = vq->umem;
1766 struct vhost_umem_node *u;
1767 u64 start, end, l, min;
1773 /* More than one GPAs can be mapped into a single HVA. So
1774 * iterate all possible umems here to be safe.
1776 list_for_each_entry(u, &umem->umem_list, link) {
1777 if (u->userspace_addr > hva - 1 + len ||
1778 u->userspace_addr - 1 + u->size < hva)
1780 start = max(u->userspace_addr, hva);
1781 end = min(u->userspace_addr - 1 + u->size,
1783 l = end - start + 1;
1784 r = log_write(vq->log_base,
1785 u->start + start - u->userspace_addr,
1803 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1805 struct iovec iov[64];
1809 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1811 ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1812 len, iov, 64, VHOST_ACCESS_WO);
1816 for (i = 0; i < ret; i++) {
1817 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1826 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1827 unsigned int log_num, u64 len, struct iovec *iov, int count)
1831 /* Make sure data written is seen before log. */
1835 for (i = 0; i < count; i++) {
1836 r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1844 for (i = 0; i < log_num; ++i) {
1845 u64 l = min(log[i].len, len);
1846 r = log_write(vq->log_base, log[i].addr, l);
1852 eventfd_signal(vq->log_ctx, 1);
1856 /* Length written exceeds what we have stored. This is a bug. */
1860 EXPORT_SYMBOL_GPL(vhost_log_write);
1862 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1865 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1866 &vq->used->flags) < 0)
1868 if (unlikely(vq->log_used)) {
1869 /* Make sure the flag is seen before log. */
1871 /* Log used flag write. */
1872 used = &vq->used->flags;
1873 log_used(vq, (used - (void __user *)vq->used),
1874 sizeof vq->used->flags);
1876 eventfd_signal(vq->log_ctx, 1);
1881 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1883 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1884 vhost_avail_event(vq)))
1886 if (unlikely(vq->log_used)) {
1888 /* Make sure the event is seen before log. */
1890 /* Log avail event write */
1891 used = vhost_avail_event(vq);
1892 log_used(vq, (used - (void __user *)vq->used),
1893 sizeof *vhost_avail_event(vq));
1895 eventfd_signal(vq->log_ctx, 1);
1900 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1902 __virtio16 last_used_idx;
1904 bool is_le = vq->is_le;
1906 if (!vq->private_data)
1909 vhost_init_is_le(vq);
1911 r = vhost_update_used_flags(vq);
1914 vq->signalled_used_valid = false;
1916 !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
1920 r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1922 vq_err(vq, "Can't access used idx at %p\n",
1926 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1933 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1935 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1936 struct iovec iov[], int iov_size, int access)
1938 const struct vhost_umem_node *node;
1939 struct vhost_dev *dev = vq->dev;
1940 struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1945 while ((u64)len > s) {
1947 if (unlikely(ret >= iov_size)) {
1952 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1953 addr, addr + len - 1);
1954 if (node == NULL || node->start > addr) {
1955 if (umem != dev->iotlb) {
1961 } else if (!(node->perm & access)) {
1967 size = node->size - addr + node->start;
1968 _iov->iov_len = min((u64)len - s, size);
1969 _iov->iov_base = (void __user *)(unsigned long)
1970 (node->userspace_addr + addr - node->start);
1977 vhost_iotlb_miss(vq, addr, access);
1981 /* Each buffer in the virtqueues is actually a chain of descriptors. This
1982 * function returns the next descriptor in the chain,
1983 * or -1U if we're at the end. */
1984 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1988 /* If this descriptor says it doesn't chain, we're done. */
1989 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1992 /* Check they're not leading us off end of descriptors. */
1993 next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
1997 static int get_indirect(struct vhost_virtqueue *vq,
1998 struct iovec iov[], unsigned int iov_size,
1999 unsigned int *out_num, unsigned int *in_num,
2000 struct vhost_log *log, unsigned int *log_num,
2001 struct vring_desc *indirect)
2003 struct vring_desc desc;
2004 unsigned int i = 0, count, found = 0;
2005 u32 len = vhost32_to_cpu(vq, indirect->len);
2006 struct iov_iter from;
2010 if (unlikely(len % sizeof desc)) {
2011 vq_err(vq, "Invalid length in indirect descriptor: "
2012 "len 0x%llx not multiple of 0x%zx\n",
2013 (unsigned long long)len,
2018 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2019 UIO_MAXIOV, VHOST_ACCESS_RO);
2020 if (unlikely(ret < 0)) {
2022 vq_err(vq, "Translation failure %d in indirect.\n", ret);
2025 iov_iter_init(&from, READ, vq->indirect, ret, len);
2027 /* We will use the result as an address to read from, so most
2028 * architectures only need a compiler barrier here. */
2029 read_barrier_depends();
2031 count = len / sizeof desc;
2032 /* Buffers are chained via a 16 bit next field, so
2033 * we can have at most 2^16 of these. */
2034 if (unlikely(count > USHRT_MAX + 1)) {
2035 vq_err(vq, "Indirect buffer length too big: %d\n",
2041 unsigned iov_count = *in_num + *out_num;
2042 if (unlikely(++found > count)) {
2043 vq_err(vq, "Loop detected: last one at %u "
2044 "indirect size %u\n",
2048 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2049 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2050 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2053 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2054 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2055 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2059 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2060 access = VHOST_ACCESS_WO;
2062 access = VHOST_ACCESS_RO;
2064 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2065 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2066 iov_size - iov_count, access);
2067 if (unlikely(ret < 0)) {
2069 vq_err(vq, "Translation failure %d indirect idx %d\n",
2073 /* If this is an input descriptor, increment that count. */
2074 if (access == VHOST_ACCESS_WO) {
2076 if (unlikely(log)) {
2077 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2078 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2082 /* If it's an output descriptor, they're all supposed
2083 * to come before any input descriptors. */
2084 if (unlikely(*in_num)) {
2085 vq_err(vq, "Indirect descriptor "
2086 "has out after in: idx %d\n", i);
2091 } while ((i = next_desc(vq, &desc)) != -1);
2095 /* This looks in the virtqueue and for the first available buffer, and converts
2096 * it to an iovec for convenient access. Since descriptors consist of some
2097 * number of output then some number of input descriptors, it's actually two
2098 * iovecs, but we pack them into one and note how many of each there were.
2100 * This function returns the descriptor number found, or vq->num (which is
2101 * never a valid descriptor number) if none was found. A negative code is
2102 * returned on error. */
2103 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2104 struct iovec iov[], unsigned int iov_size,
2105 unsigned int *out_num, unsigned int *in_num,
2106 struct vhost_log *log, unsigned int *log_num)
2108 struct vring_desc desc;
2109 unsigned int i, head, found = 0;
2111 __virtio16 avail_idx;
2112 __virtio16 ring_head;
2115 /* Check it isn't doing very strange things with descriptor numbers. */
2116 last_avail_idx = vq->last_avail_idx;
2118 if (vq->avail_idx == vq->last_avail_idx) {
2119 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
2120 vq_err(vq, "Failed to access avail idx at %p\n",
2124 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2126 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2127 vq_err(vq, "Guest moved used index from %u to %u",
2128 last_avail_idx, vq->avail_idx);
2132 /* If there's nothing new since last we looked, return
2135 if (vq->avail_idx == last_avail_idx)
2138 /* Only get avail ring entries after they have been
2144 /* Grab the next descriptor number they're advertising, and increment
2145 * the index we've seen. */
2146 if (unlikely(vhost_get_avail(vq, ring_head,
2147 &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2148 vq_err(vq, "Failed to read head: idx %d address %p\n",
2150 &vq->avail->ring[last_avail_idx % vq->num]);
2154 head = vhost16_to_cpu(vq, ring_head);
2156 /* If their number is silly, that's an error. */
2157 if (unlikely(head >= vq->num)) {
2158 vq_err(vq, "Guest says index %u > %u is available",
2163 /* When we start there are none of either input nor output. */
2164 *out_num = *in_num = 0;
2170 unsigned iov_count = *in_num + *out_num;
2171 if (unlikely(i >= vq->num)) {
2172 vq_err(vq, "Desc index is %u > %u, head = %u",
2176 if (unlikely(++found > vq->num)) {
2177 vq_err(vq, "Loop detected: last one at %u "
2178 "vq size %u head %u\n",
2182 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2184 if (unlikely(ret)) {
2185 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2189 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2190 ret = get_indirect(vq, iov, iov_size,
2192 log, log_num, &desc);
2193 if (unlikely(ret < 0)) {
2195 vq_err(vq, "Failure detected "
2196 "in indirect descriptor at idx %d\n", i);
2202 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2203 access = VHOST_ACCESS_WO;
2205 access = VHOST_ACCESS_RO;
2206 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2207 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2208 iov_size - iov_count, access);
2209 if (unlikely(ret < 0)) {
2211 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2215 if (access == VHOST_ACCESS_WO) {
2216 /* If this is an input descriptor,
2217 * increment that count. */
2219 if (unlikely(log)) {
2220 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2221 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2225 /* If it's an output descriptor, they're all supposed
2226 * to come before any input descriptors. */
2227 if (unlikely(*in_num)) {
2228 vq_err(vq, "Descriptor has out after in: "
2234 } while ((i = next_desc(vq, &desc)) != -1);
2236 /* On success, increment avail index. */
2237 vq->last_avail_idx++;
2239 /* Assume notifications from guest are disabled at this point,
2240 * if they aren't we would need to update avail_event index. */
2241 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2244 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2246 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2247 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2249 vq->last_avail_idx -= n;
2251 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2253 /* After we've used one of their buffers, we tell them about it. We'll then
2254 * want to notify the guest, using eventfd. */
2255 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2257 struct vring_used_elem heads = {
2258 cpu_to_vhost32(vq, head),
2259 cpu_to_vhost32(vq, len)
2262 return vhost_add_used_n(vq, &heads, 1);
2264 EXPORT_SYMBOL_GPL(vhost_add_used);
2266 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2267 struct vring_used_elem *heads,
2270 struct vring_used_elem __user *used;
2274 start = vq->last_used_idx & (vq->num - 1);
2275 used = vq->used->ring + start;
2277 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2278 vq_err(vq, "Failed to write used id");
2281 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2282 vq_err(vq, "Failed to write used len");
2285 } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2286 vq_err(vq, "Failed to write used");
2289 if (unlikely(vq->log_used)) {
2290 /* Make sure data is seen before log. */
2292 /* Log used ring entry write. */
2293 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2294 count * sizeof *used);
2296 old = vq->last_used_idx;
2297 new = (vq->last_used_idx += count);
2298 /* If the driver never bothers to signal in a very long while,
2299 * used index might wrap around. If that happens, invalidate
2300 * signalled_used index we stored. TODO: make sure driver
2301 * signals at least once in 2^16 and remove this. */
2302 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2303 vq->signalled_used_valid = false;
2307 /* After we've used one of their buffers, we tell them about it. We'll then
2308 * want to notify the guest, using eventfd. */
2309 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2314 start = vq->last_used_idx & (vq->num - 1);
2315 n = vq->num - start;
2317 r = __vhost_add_used_n(vq, heads, n);
2323 r = __vhost_add_used_n(vq, heads, count);
2325 /* Make sure buffer is written before we update index. */
2327 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2329 vq_err(vq, "Failed to increment used idx");
2332 if (unlikely(vq->log_used)) {
2333 /* Make sure used idx is seen before log. */
2335 /* Log used index update. */
2336 log_used(vq, offsetof(struct vring_used, idx),
2337 sizeof vq->used->idx);
2339 eventfd_signal(vq->log_ctx, 1);
2343 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2345 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2350 /* Flush out used index updates. This is paired
2351 * with the barrier that the Guest executes when enabling
2355 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2356 unlikely(vq->avail_idx == vq->last_avail_idx))
2359 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2361 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2362 vq_err(vq, "Failed to get flags");
2365 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2367 old = vq->signalled_used;
2368 v = vq->signalled_used_valid;
2369 new = vq->signalled_used = vq->last_used_idx;
2370 vq->signalled_used_valid = true;
2375 if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2376 vq_err(vq, "Failed to get used event idx");
2379 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2382 /* This actually signals the guest, using eventfd. */
2383 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2385 /* Signal the Guest tell them we used something up. */
2386 if (vq->call_ctx && vhost_notify(dev, vq))
2387 eventfd_signal(vq->call_ctx, 1);
2389 EXPORT_SYMBOL_GPL(vhost_signal);
2391 /* And here's the combo meal deal. Supersize me! */
2392 void vhost_add_used_and_signal(struct vhost_dev *dev,
2393 struct vhost_virtqueue *vq,
2394 unsigned int head, int len)
2396 vhost_add_used(vq, head, len);
2397 vhost_signal(dev, vq);
2399 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2401 /* multi-buffer version of vhost_add_used_and_signal */
2402 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2403 struct vhost_virtqueue *vq,
2404 struct vring_used_elem *heads, unsigned count)
2406 vhost_add_used_n(vq, heads, count);
2407 vhost_signal(dev, vq);
2409 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2411 /* return true if we're sure that avaiable ring is empty */
2412 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2414 __virtio16 avail_idx;
2417 if (vq->avail_idx != vq->last_avail_idx)
2420 r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2423 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2425 return vq->avail_idx == vq->last_avail_idx;
2427 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2429 /* OK, now we need to know about added descriptors. */
2430 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2432 __virtio16 avail_idx;
2435 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2437 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2438 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2439 r = vhost_update_used_flags(vq);
2441 vq_err(vq, "Failed to enable notification at %p: %d\n",
2442 &vq->used->flags, r);
2446 r = vhost_update_avail_event(vq, vq->avail_idx);
2448 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2449 vhost_avail_event(vq), r);
2453 /* They could have slipped one in as we were doing that: make
2454 * sure it's written, then check again. */
2456 r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2458 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2459 &vq->avail->idx, r);
2463 return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2465 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2467 /* We don't need to be notified again. */
2468 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2472 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2474 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2475 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2476 r = vhost_update_used_flags(vq);
2478 vq_err(vq, "Failed to enable notification at %p: %d\n",
2479 &vq->used->flags, r);
2482 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2484 /* Create a new message. */
2485 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2487 struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2491 /* Make sure all padding within the structure is initialized. */
2492 memset(&node->msg, 0, sizeof node->msg);
2494 node->msg.type = type;
2497 EXPORT_SYMBOL_GPL(vhost_new_msg);
2499 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2500 struct vhost_msg_node *node)
2502 spin_lock(&dev->iotlb_lock);
2503 list_add_tail(&node->node, head);
2504 spin_unlock(&dev->iotlb_lock);
2506 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2508 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2510 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2511 struct list_head *head)
2513 struct vhost_msg_node *node = NULL;
2515 spin_lock(&dev->iotlb_lock);
2516 if (!list_empty(head)) {
2517 node = list_first_entry(head, struct vhost_msg_node,
2519 list_del(&node->node);
2521 spin_unlock(&dev->iotlb_lock);
2525 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2528 static int __init vhost_init(void)
2533 static void __exit vhost_exit(void)
2537 module_init(vhost_init);
2538 module_exit(vhost_exit);
2540 MODULE_VERSION("0.0.1");
2541 MODULE_LICENSE("GPL v2");
2542 MODULE_AUTHOR("Michael S. Tsirkin");
2543 MODULE_DESCRIPTION("Host kernel accelerator for virtio");