1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (C) 2009 Red Hat, Inc.
3 * Copyright (C) 2006 Rusty Russell IBM Corporation
5 * Author: Michael S. Tsirkin <mst@redhat.com>
7 * Inspiration, some code, and most witty comments come from
8 * Documentation/virtual/lguest/lguest.c, by Rusty Russell
10 * Generic code for virtio server in host kernel.
13 #include <linux/eventfd.h>
14 #include <linux/vhost.h>
15 #include <linux/uio.h>
17 #include <linux/miscdevice.h>
18 #include <linux/mutex.h>
19 #include <linux/poll.h>
20 #include <linux/file.h>
21 #include <linux/highmem.h>
22 #include <linux/slab.h>
23 #include <linux/vmalloc.h>
24 #include <linux/kthread.h>
25 #include <linux/cgroup.h>
26 #include <linux/module.h>
27 #include <linux/sort.h>
28 #include <linux/sched/mm.h>
29 #include <linux/sched/signal.h>
30 #include <linux/interval_tree_generic.h>
31 #include <linux/nospec.h>
32 #include <linux/kcov.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 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
53 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
55 vq->user_be = !virtio_legacy_is_little_endian();
58 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
63 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
68 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
70 struct vhost_vring_state s;
75 if (copy_from_user(&s, argp, sizeof(s)))
78 if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
79 s.num != VHOST_VRING_BIG_ENDIAN)
82 if (s.num == VHOST_VRING_BIG_ENDIAN)
83 vhost_enable_cross_endian_big(vq);
85 vhost_enable_cross_endian_little(vq);
90 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
93 struct vhost_vring_state s = {
98 if (copy_to_user(argp, &s, sizeof(s)))
104 static void vhost_init_is_le(struct vhost_virtqueue *vq)
106 /* Note for legacy virtio: user_be is initialized at reset time
107 * according to the host endianness. If userspace does not set an
108 * explicit endianness, the default behavior is native endian, as
109 * expected by legacy virtio.
111 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
114 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
118 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
123 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
129 static void vhost_init_is_le(struct vhost_virtqueue *vq)
131 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
132 || virtio_legacy_is_little_endian();
134 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
136 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
138 vhost_init_is_le(vq);
141 struct vhost_flush_struct {
142 struct vhost_work work;
143 struct completion wait_event;
146 static void vhost_flush_work(struct vhost_work *work)
148 struct vhost_flush_struct *s;
150 s = container_of(work, struct vhost_flush_struct, work);
151 complete(&s->wait_event);
154 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
157 struct vhost_poll *poll;
159 poll = container_of(pt, struct vhost_poll, table);
161 add_wait_queue(wqh, &poll->wait);
164 static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
167 struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
168 struct vhost_work *work = &poll->work;
170 if (!(key_to_poll(key) & poll->mask))
173 if (!poll->dev->use_worker)
176 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)
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_vring_call_reset(struct vhost_vring_call *call_ctx)
303 call_ctx->ctx = NULL;
304 memset(&call_ctx->producer, 0x0, sizeof(struct irq_bypass_producer));
305 spin_lock_init(&call_ctx->ctx_lock);
308 static void vhost_vq_reset(struct vhost_dev *dev,
309 struct vhost_virtqueue *vq)
315 vq->last_avail_idx = 0;
317 vq->last_used_idx = 0;
318 vq->signalled_used = 0;
319 vq->signalled_used_valid = false;
321 vq->log_used = false;
322 vq->log_addr = -1ull;
323 vq->private_data = NULL;
324 vq->acked_features = 0;
325 vq->acked_backend_features = 0;
327 vq->error_ctx = NULL;
330 vhost_reset_is_le(vq);
331 vhost_disable_cross_endian(vq);
332 vq->busyloop_timeout = 0;
335 vhost_vring_call_reset(&vq->call_ctx);
336 __vhost_vq_meta_reset(vq);
339 static int vhost_worker(void *data)
341 struct vhost_dev *dev = data;
342 struct vhost_work *work, *work_next;
343 struct llist_node *node;
345 kthread_use_mm(dev->mm);
348 /* mb paired w/ kthread_stop */
349 set_current_state(TASK_INTERRUPTIBLE);
351 if (kthread_should_stop()) {
352 __set_current_state(TASK_RUNNING);
356 node = llist_del_all(&dev->work_list);
360 node = llist_reverse_order(node);
361 /* make sure flag is seen after deletion */
363 llist_for_each_entry_safe(work, work_next, node, node) {
364 clear_bit(VHOST_WORK_QUEUED, &work->flags);
365 __set_current_state(TASK_RUNNING);
366 kcov_remote_start_common(dev->kcov_handle);
373 kthread_unuse_mm(dev->mm);
377 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
387 /* Helper to allocate iovec buffers for all vqs. */
388 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
390 struct vhost_virtqueue *vq;
393 for (i = 0; i < dev->nvqs; ++i) {
395 vq->indirect = kmalloc_array(UIO_MAXIOV,
396 sizeof(*vq->indirect),
398 vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
400 vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
402 if (!vq->indirect || !vq->log || !vq->heads)
409 vhost_vq_free_iovecs(dev->vqs[i]);
413 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
417 for (i = 0; i < dev->nvqs; ++i)
418 vhost_vq_free_iovecs(dev->vqs[i]);
421 bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
422 int pkts, int total_len)
424 struct vhost_dev *dev = vq->dev;
426 if ((dev->byte_weight && total_len >= dev->byte_weight) ||
427 pkts >= dev->weight) {
428 vhost_poll_queue(&vq->poll);
434 EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
436 static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
439 size_t event __maybe_unused =
440 vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
442 return sizeof(*vq->avail) +
443 sizeof(*vq->avail->ring) * num + event;
446 static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
449 size_t event __maybe_unused =
450 vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
452 return sizeof(*vq->used) +
453 sizeof(*vq->used->ring) * num + event;
456 static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
459 return sizeof(*vq->desc) * num;
462 void vhost_dev_init(struct vhost_dev *dev,
463 struct vhost_virtqueue **vqs, int nvqs,
464 int iov_limit, int weight, int byte_weight,
466 int (*msg_handler)(struct vhost_dev *dev,
467 struct vhost_iotlb_msg *msg))
469 struct vhost_virtqueue *vq;
474 mutex_init(&dev->mutex);
480 dev->iov_limit = iov_limit;
481 dev->weight = weight;
482 dev->byte_weight = byte_weight;
483 dev->use_worker = use_worker;
484 dev->msg_handler = msg_handler;
485 init_llist_head(&dev->work_list);
486 init_waitqueue_head(&dev->wait);
487 INIT_LIST_HEAD(&dev->read_list);
488 INIT_LIST_HEAD(&dev->pending_list);
489 spin_lock_init(&dev->iotlb_lock);
492 for (i = 0; i < dev->nvqs; ++i) {
498 mutex_init(&vq->mutex);
499 vhost_vq_reset(dev, vq);
501 vhost_poll_init(&vq->poll, vq->handle_kick,
505 EXPORT_SYMBOL_GPL(vhost_dev_init);
507 /* Caller should have device mutex */
508 long vhost_dev_check_owner(struct vhost_dev *dev)
510 /* Are you the owner? If not, I don't think you mean to do that */
511 return dev->mm == current->mm ? 0 : -EPERM;
513 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
515 struct vhost_attach_cgroups_struct {
516 struct vhost_work work;
517 struct task_struct *owner;
521 static void vhost_attach_cgroups_work(struct vhost_work *work)
523 struct vhost_attach_cgroups_struct *s;
525 s = container_of(work, struct vhost_attach_cgroups_struct, work);
526 s->ret = cgroup_attach_task_all(s->owner, current);
529 static int vhost_attach_cgroups(struct vhost_dev *dev)
531 struct vhost_attach_cgroups_struct attach;
533 attach.owner = current;
534 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
535 vhost_work_queue(dev, &attach.work);
536 vhost_work_flush(dev, &attach.work);
540 /* Caller should have device mutex */
541 bool vhost_dev_has_owner(struct vhost_dev *dev)
545 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
547 static void vhost_attach_mm(struct vhost_dev *dev)
549 /* No owner, become one */
550 if (dev->use_worker) {
551 dev->mm = get_task_mm(current);
553 /* vDPA device does not use worker thead, so there's
554 * no need to hold the address space for mm. This help
555 * to avoid deadlock in the case of mmap() which may
556 * held the refcnt of the file and depends on release
557 * method to remove vma.
559 dev->mm = current->mm;
564 static void vhost_detach_mm(struct vhost_dev *dev)
577 /* Caller should have device mutex */
578 long vhost_dev_set_owner(struct vhost_dev *dev)
580 struct task_struct *worker;
583 /* Is there an owner already? */
584 if (vhost_dev_has_owner(dev)) {
589 vhost_attach_mm(dev);
591 dev->kcov_handle = kcov_common_handle();
592 if (dev->use_worker) {
593 worker = kthread_create(vhost_worker, dev,
594 "vhost-%d", current->pid);
595 if (IS_ERR(worker)) {
596 err = PTR_ERR(worker);
600 dev->worker = worker;
601 wake_up_process(worker); /* avoid contributing to loadavg */
603 err = vhost_attach_cgroups(dev);
608 err = vhost_dev_alloc_iovecs(dev);
615 kthread_stop(dev->worker);
619 vhost_detach_mm(dev);
620 dev->kcov_handle = 0;
624 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
626 static struct vhost_iotlb *iotlb_alloc(void)
628 return vhost_iotlb_alloc(max_iotlb_entries,
629 VHOST_IOTLB_FLAG_RETIRE);
632 struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
634 return iotlb_alloc();
636 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
638 /* Caller should have device mutex */
639 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
643 vhost_dev_cleanup(dev);
646 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
647 * VQs aren't running.
649 for (i = 0; i < dev->nvqs; ++i)
650 dev->vqs[i]->umem = umem;
652 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
654 void vhost_dev_stop(struct vhost_dev *dev)
658 for (i = 0; i < dev->nvqs; ++i) {
659 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
660 vhost_poll_stop(&dev->vqs[i]->poll);
661 vhost_poll_flush(&dev->vqs[i]->poll);
665 EXPORT_SYMBOL_GPL(vhost_dev_stop);
667 static void vhost_clear_msg(struct vhost_dev *dev)
669 struct vhost_msg_node *node, *n;
671 spin_lock(&dev->iotlb_lock);
673 list_for_each_entry_safe(node, n, &dev->read_list, node) {
674 list_del(&node->node);
678 list_for_each_entry_safe(node, n, &dev->pending_list, node) {
679 list_del(&node->node);
683 spin_unlock(&dev->iotlb_lock);
686 void vhost_dev_cleanup(struct vhost_dev *dev)
690 for (i = 0; i < dev->nvqs; ++i) {
691 if (dev->vqs[i]->error_ctx)
692 eventfd_ctx_put(dev->vqs[i]->error_ctx);
693 if (dev->vqs[i]->kick)
694 fput(dev->vqs[i]->kick);
695 if (dev->vqs[i]->call_ctx.ctx)
696 eventfd_ctx_put(dev->vqs[i]->call_ctx.ctx);
697 vhost_vq_reset(dev, dev->vqs[i]);
699 vhost_dev_free_iovecs(dev);
701 eventfd_ctx_put(dev->log_ctx);
703 /* No one will access memory at this point */
704 vhost_iotlb_free(dev->umem);
706 vhost_iotlb_free(dev->iotlb);
708 vhost_clear_msg(dev);
709 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
710 WARN_ON(!llist_empty(&dev->work_list));
712 kthread_stop(dev->worker);
714 dev->kcov_handle = 0;
716 vhost_detach_mm(dev);
718 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
720 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
722 u64 a = addr / VHOST_PAGE_SIZE / 8;
724 /* Make sure 64 bit math will not overflow. */
725 if (a > ULONG_MAX - (unsigned long)log_base ||
726 a + (unsigned long)log_base > ULONG_MAX)
729 return access_ok(log_base + a,
730 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
733 static bool vhost_overflow(u64 uaddr, u64 size)
735 /* Make sure 64 bit math will not overflow. */
736 return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
739 /* Caller should have vq mutex and device mutex. */
740 static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
743 struct vhost_iotlb_map *map;
748 list_for_each_entry(map, &umem->list, link) {
749 unsigned long a = map->addr;
751 if (vhost_overflow(map->addr, map->size))
755 if (!access_ok((void __user *)a, map->size))
757 else if (log_all && !log_access_ok(log_base,
765 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
766 u64 addr, unsigned int size,
769 const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
774 return (void __user *)(uintptr_t)(map->addr + addr - map->start);
777 /* Can we switch to this memory table? */
778 /* Caller should have device mutex but not vq mutex */
779 static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
784 for (i = 0; i < d->nvqs; ++i) {
788 mutex_lock(&d->vqs[i]->mutex);
789 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
790 /* If ring is inactive, will check when it's enabled. */
791 if (d->vqs[i]->private_data)
792 ok = vq_memory_access_ok(d->vqs[i]->log_base,
796 mutex_unlock(&d->vqs[i]->mutex);
803 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
804 struct iovec iov[], int iov_size, int access);
806 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
807 const void *from, unsigned size)
812 return __copy_to_user(to, from, size);
814 /* This function should be called after iotlb
815 * prefetch, which means we're sure that all vq
816 * could be access through iotlb. So -EAGAIN should
817 * not happen in this case.
820 void __user *uaddr = vhost_vq_meta_fetch(vq,
821 (u64)(uintptr_t)to, size,
825 return __copy_to_user(uaddr, from, size);
827 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
828 ARRAY_SIZE(vq->iotlb_iov),
832 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
833 ret = copy_to_iter(from, size, &t);
841 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
842 void __user *from, unsigned size)
847 return __copy_from_user(to, from, size);
849 /* This function should be called after iotlb
850 * prefetch, which means we're sure that vq
851 * could be access through iotlb. So -EAGAIN should
852 * not happen in this case.
854 void __user *uaddr = vhost_vq_meta_fetch(vq,
855 (u64)(uintptr_t)from, size,
860 return __copy_from_user(to, uaddr, size);
862 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
863 ARRAY_SIZE(vq->iotlb_iov),
866 vq_err(vq, "IOTLB translation failure: uaddr "
867 "%p size 0x%llx\n", from,
868 (unsigned long long) size);
871 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
872 ret = copy_from_iter(to, size, &f);
881 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
882 void __user *addr, unsigned int size,
887 ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
888 ARRAY_SIZE(vq->iotlb_iov),
891 vq_err(vq, "IOTLB translation failure: uaddr "
892 "%p size 0x%llx\n", addr,
893 (unsigned long long) size);
897 if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
898 vq_err(vq, "Non atomic userspace memory access: uaddr "
899 "%p size 0x%llx\n", addr,
900 (unsigned long long) size);
904 return vq->iotlb_iov[0].iov_base;
907 /* This function should be called after iotlb
908 * prefetch, which means we're sure that vq
909 * could be access through iotlb. So -EAGAIN should
910 * not happen in this case.
912 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
913 void __user *addr, unsigned int size,
916 void __user *uaddr = vhost_vq_meta_fetch(vq,
917 (u64)(uintptr_t)addr, size, type);
921 return __vhost_get_user_slow(vq, addr, size, type);
924 #define vhost_put_user(vq, x, ptr) \
928 ret = __put_user(x, ptr); \
930 __typeof__(ptr) to = \
931 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
932 sizeof(*ptr), VHOST_ADDR_USED); \
934 ret = __put_user(x, to); \
941 static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
943 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
944 vhost_avail_event(vq));
947 static inline int vhost_put_used(struct vhost_virtqueue *vq,
948 struct vring_used_elem *head, int idx,
951 return vhost_copy_to_user(vq, vq->used->ring + idx, head,
952 count * sizeof(*head));
955 static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
958 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
962 static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
965 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
969 #define vhost_get_user(vq, x, ptr, type) \
973 ret = __get_user(x, ptr); \
975 __typeof__(ptr) from = \
976 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
980 ret = __get_user(x, from); \
987 #define vhost_get_avail(vq, x, ptr) \
988 vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
990 #define vhost_get_used(vq, x, ptr) \
991 vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
993 static void vhost_dev_lock_vqs(struct vhost_dev *d)
996 for (i = 0; i < d->nvqs; ++i)
997 mutex_lock_nested(&d->vqs[i]->mutex, i);
1000 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
1003 for (i = 0; i < d->nvqs; ++i)
1004 mutex_unlock(&d->vqs[i]->mutex);
1007 static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
1010 return vhost_get_avail(vq, *idx, &vq->avail->idx);
1013 static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
1014 __virtio16 *head, int idx)
1016 return vhost_get_avail(vq, *head,
1017 &vq->avail->ring[idx & (vq->num - 1)]);
1020 static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
1023 return vhost_get_avail(vq, *flags, &vq->avail->flags);
1026 static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1029 return vhost_get_avail(vq, *event, vhost_used_event(vq));
1032 static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1035 return vhost_get_used(vq, *idx, &vq->used->idx);
1038 static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1039 struct vring_desc *desc, int idx)
1041 return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1044 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1045 struct vhost_iotlb_msg *msg)
1047 struct vhost_msg_node *node, *n;
1049 spin_lock(&d->iotlb_lock);
1051 list_for_each_entry_safe(node, n, &d->pending_list, node) {
1052 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1053 if (msg->iova <= vq_msg->iova &&
1054 msg->iova + msg->size - 1 >= vq_msg->iova &&
1055 vq_msg->type == VHOST_IOTLB_MISS) {
1056 vhost_poll_queue(&node->vq->poll);
1057 list_del(&node->node);
1062 spin_unlock(&d->iotlb_lock);
1065 static bool umem_access_ok(u64 uaddr, u64 size, int access)
1067 unsigned long a = uaddr;
1069 /* Make sure 64 bit math will not overflow. */
1070 if (vhost_overflow(uaddr, size))
1073 if ((access & VHOST_ACCESS_RO) &&
1074 !access_ok((void __user *)a, size))
1076 if ((access & VHOST_ACCESS_WO) &&
1077 !access_ok((void __user *)a, size))
1082 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
1083 struct vhost_iotlb_msg *msg)
1087 mutex_lock(&dev->mutex);
1088 vhost_dev_lock_vqs(dev);
1089 switch (msg->type) {
1090 case VHOST_IOTLB_UPDATE:
1095 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1099 vhost_vq_meta_reset(dev);
1100 if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1101 msg->iova + msg->size - 1,
1102 msg->uaddr, msg->perm)) {
1106 vhost_iotlb_notify_vq(dev, msg);
1108 case VHOST_IOTLB_INVALIDATE:
1113 vhost_vq_meta_reset(dev);
1114 vhost_iotlb_del_range(dev->iotlb, msg->iova,
1115 msg->iova + msg->size - 1);
1122 vhost_dev_unlock_vqs(dev);
1123 mutex_unlock(&dev->mutex);
1127 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1128 struct iov_iter *from)
1130 struct vhost_iotlb_msg msg;
1134 ret = copy_from_iter(&type, sizeof(type), from);
1135 if (ret != sizeof(type)) {
1141 case VHOST_IOTLB_MSG:
1142 /* There maybe a hole after type for V1 message type,
1145 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1147 case VHOST_IOTLB_MSG_V2:
1148 offset = sizeof(__u32);
1155 iov_iter_advance(from, offset);
1156 ret = copy_from_iter(&msg, sizeof(msg), from);
1157 if (ret != sizeof(msg)) {
1162 if (dev->msg_handler)
1163 ret = dev->msg_handler(dev, &msg);
1165 ret = vhost_process_iotlb_msg(dev, &msg);
1171 ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1172 sizeof(struct vhost_msg_v2);
1176 EXPORT_SYMBOL(vhost_chr_write_iter);
1178 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1183 poll_wait(file, &dev->wait, wait);
1185 if (!list_empty(&dev->read_list))
1186 mask |= EPOLLIN | EPOLLRDNORM;
1190 EXPORT_SYMBOL(vhost_chr_poll);
1192 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1196 struct vhost_msg_node *node;
1198 unsigned size = sizeof(struct vhost_msg);
1200 if (iov_iter_count(to) < size)
1205 prepare_to_wait(&dev->wait, &wait,
1206 TASK_INTERRUPTIBLE);
1208 node = vhost_dequeue_msg(dev, &dev->read_list);
1215 if (signal_pending(current)) {
1228 finish_wait(&dev->wait, &wait);
1231 struct vhost_iotlb_msg *msg;
1232 void *start = &node->msg;
1234 switch (node->msg.type) {
1235 case VHOST_IOTLB_MSG:
1236 size = sizeof(node->msg);
1237 msg = &node->msg.iotlb;
1239 case VHOST_IOTLB_MSG_V2:
1240 size = sizeof(node->msg_v2);
1241 msg = &node->msg_v2.iotlb;
1248 ret = copy_to_iter(start, size, to);
1249 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1253 vhost_enqueue_msg(dev, &dev->pending_list, node);
1258 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1260 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1262 struct vhost_dev *dev = vq->dev;
1263 struct vhost_msg_node *node;
1264 struct vhost_iotlb_msg *msg;
1265 bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1267 node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1272 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1273 msg = &node->msg_v2.iotlb;
1275 msg = &node->msg.iotlb;
1278 msg->type = VHOST_IOTLB_MISS;
1282 vhost_enqueue_msg(dev, &dev->read_list, node);
1287 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1288 vring_desc_t __user *desc,
1289 vring_avail_t __user *avail,
1290 vring_used_t __user *used)
1293 return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1294 access_ok(avail, vhost_get_avail_size(vq, num)) &&
1295 access_ok(used, vhost_get_used_size(vq, num));
1298 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1299 const struct vhost_iotlb_map *map,
1302 int access = (type == VHOST_ADDR_USED) ?
1303 VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1305 if (likely(map->perm & access))
1306 vq->meta_iotlb[type] = map;
1309 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1310 int access, u64 addr, u64 len, int type)
1312 const struct vhost_iotlb_map *map;
1313 struct vhost_iotlb *umem = vq->iotlb;
1314 u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1316 if (vhost_vq_meta_fetch(vq, addr, len, type))
1320 map = vhost_iotlb_itree_first(umem, addr, last);
1321 if (map == NULL || map->start > addr) {
1322 vhost_iotlb_miss(vq, addr, access);
1324 } else if (!(map->perm & access)) {
1325 /* Report the possible access violation by
1326 * request another translation from userspace.
1331 size = map->size - addr + map->start;
1333 if (orig_addr == addr && size >= len)
1334 vhost_vq_meta_update(vq, map, type);
1343 int vq_meta_prefetch(struct vhost_virtqueue *vq)
1345 unsigned int num = vq->num;
1350 return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1351 vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1352 iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1353 vhost_get_avail_size(vq, num),
1354 VHOST_ADDR_AVAIL) &&
1355 iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1356 vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1358 EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1360 /* Can we log writes? */
1361 /* Caller should have device mutex but not vq mutex */
1362 bool vhost_log_access_ok(struct vhost_dev *dev)
1364 return memory_access_ok(dev, dev->umem, 1);
1366 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1368 /* Verify access for write logging. */
1369 /* Caller should have vq mutex and device mutex */
1370 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1371 void __user *log_base)
1373 return vq_memory_access_ok(log_base, vq->umem,
1374 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1375 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1376 vhost_get_used_size(vq, vq->num)));
1379 /* Can we start vq? */
1380 /* Caller should have vq mutex and device mutex */
1381 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1383 if (!vq_log_access_ok(vq, vq->log_base))
1386 /* Access validation occurs at prefetch time with IOTLB */
1390 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1392 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1394 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1396 struct vhost_memory mem, *newmem;
1397 struct vhost_memory_region *region;
1398 struct vhost_iotlb *newumem, *oldumem;
1399 unsigned long size = offsetof(struct vhost_memory, regions);
1402 if (copy_from_user(&mem, m, size))
1406 if (mem.nregions > max_mem_regions)
1408 newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1413 memcpy(newmem, &mem, size);
1414 if (copy_from_user(newmem->regions, m->regions,
1415 flex_array_size(newmem, regions, mem.nregions))) {
1420 newumem = iotlb_alloc();
1426 for (region = newmem->regions;
1427 region < newmem->regions + mem.nregions;
1429 if (vhost_iotlb_add_range(newumem,
1430 region->guest_phys_addr,
1431 region->guest_phys_addr +
1432 region->memory_size - 1,
1433 region->userspace_addr,
1438 if (!memory_access_ok(d, newumem, 0))
1444 /* All memory accesses are done under some VQ mutex. */
1445 for (i = 0; i < d->nvqs; ++i) {
1446 mutex_lock(&d->vqs[i]->mutex);
1447 d->vqs[i]->umem = newumem;
1448 mutex_unlock(&d->vqs[i]->mutex);
1452 vhost_iotlb_free(oldumem);
1456 vhost_iotlb_free(newumem);
1461 static long vhost_vring_set_num(struct vhost_dev *d,
1462 struct vhost_virtqueue *vq,
1465 struct vhost_vring_state s;
1467 /* Resizing ring with an active backend?
1468 * You don't want to do that. */
1469 if (vq->private_data)
1472 if (copy_from_user(&s, argp, sizeof s))
1475 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1482 static long vhost_vring_set_addr(struct vhost_dev *d,
1483 struct vhost_virtqueue *vq,
1486 struct vhost_vring_addr a;
1488 if (copy_from_user(&a, argp, sizeof a))
1490 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1493 /* For 32bit, verify that the top 32bits of the user
1494 data are set to zero. */
1495 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1496 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1497 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1500 /* Make sure it's safe to cast pointers to vring types. */
1501 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1502 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1503 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1504 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1505 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1508 /* We only verify access here if backend is configured.
1509 * If it is not, we don't as size might not have been setup.
1510 * We will verify when backend is configured. */
1511 if (vq->private_data) {
1512 if (!vq_access_ok(vq, vq->num,
1513 (void __user *)(unsigned long)a.desc_user_addr,
1514 (void __user *)(unsigned long)a.avail_user_addr,
1515 (void __user *)(unsigned long)a.used_user_addr))
1518 /* Also validate log access for used ring if enabled. */
1519 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1520 !log_access_ok(vq->log_base, a.log_guest_addr,
1522 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;
1535 static long vhost_vring_set_num_addr(struct vhost_dev *d,
1536 struct vhost_virtqueue *vq,
1542 mutex_lock(&vq->mutex);
1545 case VHOST_SET_VRING_NUM:
1546 r = vhost_vring_set_num(d, vq, argp);
1548 case VHOST_SET_VRING_ADDR:
1549 r = vhost_vring_set_addr(d, vq, argp);
1555 mutex_unlock(&vq->mutex);
1559 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1561 struct file *eventfp, *filep = NULL;
1562 bool pollstart = false, pollstop = false;
1563 struct eventfd_ctx *ctx = NULL;
1564 u32 __user *idxp = argp;
1565 struct vhost_virtqueue *vq;
1566 struct vhost_vring_state s;
1567 struct vhost_vring_file f;
1571 r = get_user(idx, idxp);
1577 idx = array_index_nospec(idx, d->nvqs);
1580 if (ioctl == VHOST_SET_VRING_NUM ||
1581 ioctl == VHOST_SET_VRING_ADDR) {
1582 return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1585 mutex_lock(&vq->mutex);
1588 case VHOST_SET_VRING_BASE:
1589 /* Moving base with an active backend?
1590 * You don't want to do that. */
1591 if (vq->private_data) {
1595 if (copy_from_user(&s, argp, sizeof s)) {
1599 if (s.num > 0xffff) {
1603 vq->last_avail_idx = s.num;
1604 /* Forget the cached index value. */
1605 vq->avail_idx = vq->last_avail_idx;
1607 case VHOST_GET_VRING_BASE:
1609 s.num = vq->last_avail_idx;
1610 if (copy_to_user(argp, &s, sizeof s))
1613 case VHOST_SET_VRING_KICK:
1614 if (copy_from_user(&f, argp, sizeof f)) {
1618 eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
1619 if (IS_ERR(eventfp)) {
1620 r = PTR_ERR(eventfp);
1623 if (eventfp != vq->kick) {
1624 pollstop = (filep = vq->kick) != NULL;
1625 pollstart = (vq->kick = eventfp) != NULL;
1629 case VHOST_SET_VRING_CALL:
1630 if (copy_from_user(&f, argp, sizeof f)) {
1634 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1640 spin_lock(&vq->call_ctx.ctx_lock);
1641 swap(ctx, vq->call_ctx.ctx);
1642 spin_unlock(&vq->call_ctx.ctx_lock);
1644 case VHOST_SET_VRING_ERR:
1645 if (copy_from_user(&f, argp, sizeof f)) {
1649 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1654 swap(ctx, vq->error_ctx);
1656 case VHOST_SET_VRING_ENDIAN:
1657 r = vhost_set_vring_endian(vq, argp);
1659 case VHOST_GET_VRING_ENDIAN:
1660 r = vhost_get_vring_endian(vq, idx, argp);
1662 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1663 if (copy_from_user(&s, argp, sizeof(s))) {
1667 vq->busyloop_timeout = s.num;
1669 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1671 s.num = vq->busyloop_timeout;
1672 if (copy_to_user(argp, &s, sizeof(s)))
1679 if (pollstop && vq->handle_kick)
1680 vhost_poll_stop(&vq->poll);
1682 if (!IS_ERR_OR_NULL(ctx))
1683 eventfd_ctx_put(ctx);
1687 if (pollstart && vq->handle_kick)
1688 r = vhost_poll_start(&vq->poll, vq->kick);
1690 mutex_unlock(&vq->mutex);
1692 if (pollstop && vq->handle_kick)
1693 vhost_poll_flush(&vq->poll);
1696 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1698 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1700 struct vhost_iotlb *niotlb, *oiotlb;
1703 niotlb = iotlb_alloc();
1710 for (i = 0; i < d->nvqs; ++i) {
1711 struct vhost_virtqueue *vq = d->vqs[i];
1713 mutex_lock(&vq->mutex);
1715 __vhost_vq_meta_reset(vq);
1716 mutex_unlock(&vq->mutex);
1719 vhost_iotlb_free(oiotlb);
1723 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1725 /* Caller must have device mutex */
1726 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1728 struct eventfd_ctx *ctx;
1733 /* If you are not the owner, you can become one */
1734 if (ioctl == VHOST_SET_OWNER) {
1735 r = vhost_dev_set_owner(d);
1739 /* You must be the owner to do anything else */
1740 r = vhost_dev_check_owner(d);
1745 case VHOST_SET_MEM_TABLE:
1746 r = vhost_set_memory(d, argp);
1748 case VHOST_SET_LOG_BASE:
1749 if (copy_from_user(&p, argp, sizeof p)) {
1753 if ((u64)(unsigned long)p != p) {
1757 for (i = 0; i < d->nvqs; ++i) {
1758 struct vhost_virtqueue *vq;
1759 void __user *base = (void __user *)(unsigned long)p;
1761 mutex_lock(&vq->mutex);
1762 /* If ring is inactive, will check when it's enabled. */
1763 if (vq->private_data && !vq_log_access_ok(vq, base))
1766 vq->log_base = base;
1767 mutex_unlock(&vq->mutex);
1770 case VHOST_SET_LOG_FD:
1771 r = get_user(fd, (int __user *)argp);
1774 ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
1779 swap(ctx, d->log_ctx);
1780 for (i = 0; i < d->nvqs; ++i) {
1781 mutex_lock(&d->vqs[i]->mutex);
1782 d->vqs[i]->log_ctx = d->log_ctx;
1783 mutex_unlock(&d->vqs[i]->mutex);
1786 eventfd_ctx_put(ctx);
1795 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1797 /* TODO: This is really inefficient. We need something like get_user()
1798 * (instruction directly accesses the data, with an exception table entry
1799 * returning -EFAULT). See Documentation/x86/exception-tables.rst.
1801 static int set_bit_to_user(int nr, void __user *addr)
1803 unsigned long log = (unsigned long)addr;
1806 int bit = nr + (log % PAGE_SIZE) * 8;
1809 r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
1813 base = kmap_atomic(page);
1815 kunmap_atomic(base);
1816 unpin_user_pages_dirty_lock(&page, 1, true);
1820 static int log_write(void __user *log_base,
1821 u64 write_address, u64 write_length)
1823 u64 write_page = write_address / VHOST_PAGE_SIZE;
1828 write_length += write_address % VHOST_PAGE_SIZE;
1830 u64 base = (u64)(unsigned long)log_base;
1831 u64 log = base + write_page / 8;
1832 int bit = write_page % 8;
1833 if ((u64)(unsigned long)log != log)
1835 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1838 if (write_length <= VHOST_PAGE_SIZE)
1840 write_length -= VHOST_PAGE_SIZE;
1846 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1848 struct vhost_iotlb *umem = vq->umem;
1849 struct vhost_iotlb_map *u;
1850 u64 start, end, l, min;
1856 /* More than one GPAs can be mapped into a single HVA. So
1857 * iterate all possible umems here to be safe.
1859 list_for_each_entry(u, &umem->list, link) {
1860 if (u->addr > hva - 1 + len ||
1861 u->addr - 1 + u->size < hva)
1863 start = max(u->addr, hva);
1864 end = min(u->addr - 1 + u->size, hva - 1 + len);
1865 l = end - start + 1;
1866 r = log_write(vq->log_base,
1867 u->start + start - u->addr,
1885 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1887 struct iovec iov[64];
1891 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1893 ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1894 len, iov, 64, VHOST_ACCESS_WO);
1898 for (i = 0; i < ret; i++) {
1899 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1908 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1909 unsigned int log_num, u64 len, struct iovec *iov, int count)
1913 /* Make sure data written is seen before log. */
1917 for (i = 0; i < count; i++) {
1918 r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1926 for (i = 0; i < log_num; ++i) {
1927 u64 l = min(log[i].len, len);
1928 r = log_write(vq->log_base, log[i].addr, l);
1934 eventfd_signal(vq->log_ctx, 1);
1938 /* Length written exceeds what we have stored. This is a bug. */
1942 EXPORT_SYMBOL_GPL(vhost_log_write);
1944 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1947 if (vhost_put_used_flags(vq))
1949 if (unlikely(vq->log_used)) {
1950 /* Make sure the flag is seen before log. */
1952 /* Log used flag write. */
1953 used = &vq->used->flags;
1954 log_used(vq, (used - (void __user *)vq->used),
1955 sizeof vq->used->flags);
1957 eventfd_signal(vq->log_ctx, 1);
1962 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1964 if (vhost_put_avail_event(vq))
1966 if (unlikely(vq->log_used)) {
1968 /* Make sure the event is seen before log. */
1970 /* Log avail event write */
1971 used = vhost_avail_event(vq);
1972 log_used(vq, (used - (void __user *)vq->used),
1973 sizeof *vhost_avail_event(vq));
1975 eventfd_signal(vq->log_ctx, 1);
1980 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1982 __virtio16 last_used_idx;
1984 bool is_le = vq->is_le;
1986 if (!vq->private_data)
1989 vhost_init_is_le(vq);
1991 r = vhost_update_used_flags(vq);
1994 vq->signalled_used_valid = false;
1996 !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2000 r = vhost_get_used_idx(vq, &last_used_idx);
2002 vq_err(vq, "Can't access used idx at %p\n",
2006 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2013 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2015 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2016 struct iovec iov[], int iov_size, int access)
2018 const struct vhost_iotlb_map *map;
2019 struct vhost_dev *dev = vq->dev;
2020 struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2025 while ((u64)len > s) {
2027 if (unlikely(ret >= iov_size)) {
2032 map = vhost_iotlb_itree_first(umem, addr, addr + len - 1);
2033 if (map == NULL || map->start > addr) {
2034 if (umem != dev->iotlb) {
2040 } else if (!(map->perm & access)) {
2046 size = map->size - addr + map->start;
2047 _iov->iov_len = min((u64)len - s, size);
2048 _iov->iov_base = (void __user *)(unsigned long)
2049 (map->addr + addr - map->start);
2056 vhost_iotlb_miss(vq, addr, access);
2060 /* Each buffer in the virtqueues is actually a chain of descriptors. This
2061 * function returns the next descriptor in the chain,
2062 * or -1U if we're at the end. */
2063 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2067 /* If this descriptor says it doesn't chain, we're done. */
2068 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2071 /* Check they're not leading us off end of descriptors. */
2072 next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2076 static int get_indirect(struct vhost_virtqueue *vq,
2077 struct iovec iov[], unsigned int iov_size,
2078 unsigned int *out_num, unsigned int *in_num,
2079 struct vhost_log *log, unsigned int *log_num,
2080 struct vring_desc *indirect)
2082 struct vring_desc desc;
2083 unsigned int i = 0, count, found = 0;
2084 u32 len = vhost32_to_cpu(vq, indirect->len);
2085 struct iov_iter from;
2089 if (unlikely(len % sizeof desc)) {
2090 vq_err(vq, "Invalid length in indirect descriptor: "
2091 "len 0x%llx not multiple of 0x%zx\n",
2092 (unsigned long long)len,
2097 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2098 UIO_MAXIOV, VHOST_ACCESS_RO);
2099 if (unlikely(ret < 0)) {
2101 vq_err(vq, "Translation failure %d in indirect.\n", ret);
2104 iov_iter_init(&from, READ, vq->indirect, ret, len);
2105 count = len / sizeof desc;
2106 /* Buffers are chained via a 16 bit next field, so
2107 * we can have at most 2^16 of these. */
2108 if (unlikely(count > USHRT_MAX + 1)) {
2109 vq_err(vq, "Indirect buffer length too big: %d\n",
2115 unsigned iov_count = *in_num + *out_num;
2116 if (unlikely(++found > count)) {
2117 vq_err(vq, "Loop detected: last one at %u "
2118 "indirect size %u\n",
2122 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2123 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2124 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2127 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2128 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2129 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2133 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2134 access = VHOST_ACCESS_WO;
2136 access = VHOST_ACCESS_RO;
2138 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2139 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2140 iov_size - iov_count, access);
2141 if (unlikely(ret < 0)) {
2143 vq_err(vq, "Translation failure %d indirect idx %d\n",
2147 /* If this is an input descriptor, increment that count. */
2148 if (access == VHOST_ACCESS_WO) {
2150 if (unlikely(log && ret)) {
2151 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2152 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2156 /* If it's an output descriptor, they're all supposed
2157 * to come before any input descriptors. */
2158 if (unlikely(*in_num)) {
2159 vq_err(vq, "Indirect descriptor "
2160 "has out after in: idx %d\n", i);
2165 } while ((i = next_desc(vq, &desc)) != -1);
2169 /* This looks in the virtqueue and for the first available buffer, and converts
2170 * it to an iovec for convenient access. Since descriptors consist of some
2171 * number of output then some number of input descriptors, it's actually two
2172 * iovecs, but we pack them into one and note how many of each there were.
2174 * This function returns the descriptor number found, or vq->num (which is
2175 * never a valid descriptor number) if none was found. A negative code is
2176 * returned on error. */
2177 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2178 struct iovec iov[], unsigned int iov_size,
2179 unsigned int *out_num, unsigned int *in_num,
2180 struct vhost_log *log, unsigned int *log_num)
2182 struct vring_desc desc;
2183 unsigned int i, head, found = 0;
2185 __virtio16 avail_idx;
2186 __virtio16 ring_head;
2189 /* Check it isn't doing very strange things with descriptor numbers. */
2190 last_avail_idx = vq->last_avail_idx;
2192 if (vq->avail_idx == vq->last_avail_idx) {
2193 if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2194 vq_err(vq, "Failed to access avail idx at %p\n",
2198 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2200 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2201 vq_err(vq, "Guest moved used index from %u to %u",
2202 last_avail_idx, vq->avail_idx);
2206 /* If there's nothing new since last we looked, return
2209 if (vq->avail_idx == last_avail_idx)
2212 /* Only get avail ring entries after they have been
2218 /* Grab the next descriptor number they're advertising, and increment
2219 * the index we've seen. */
2220 if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2221 vq_err(vq, "Failed to read head: idx %d address %p\n",
2223 &vq->avail->ring[last_avail_idx % vq->num]);
2227 head = vhost16_to_cpu(vq, ring_head);
2229 /* If their number is silly, that's an error. */
2230 if (unlikely(head >= vq->num)) {
2231 vq_err(vq, "Guest says index %u > %u is available",
2236 /* When we start there are none of either input nor output. */
2237 *out_num = *in_num = 0;
2243 unsigned iov_count = *in_num + *out_num;
2244 if (unlikely(i >= vq->num)) {
2245 vq_err(vq, "Desc index is %u > %u, head = %u",
2249 if (unlikely(++found > vq->num)) {
2250 vq_err(vq, "Loop detected: last one at %u "
2251 "vq size %u head %u\n",
2255 ret = vhost_get_desc(vq, &desc, i);
2256 if (unlikely(ret)) {
2257 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2261 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2262 ret = get_indirect(vq, iov, iov_size,
2264 log, log_num, &desc);
2265 if (unlikely(ret < 0)) {
2267 vq_err(vq, "Failure detected "
2268 "in indirect descriptor at idx %d\n", i);
2274 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2275 access = VHOST_ACCESS_WO;
2277 access = VHOST_ACCESS_RO;
2278 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2279 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2280 iov_size - iov_count, access);
2281 if (unlikely(ret < 0)) {
2283 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2287 if (access == VHOST_ACCESS_WO) {
2288 /* If this is an input descriptor,
2289 * increment that count. */
2291 if (unlikely(log && ret)) {
2292 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2293 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2297 /* If it's an output descriptor, they're all supposed
2298 * to come before any input descriptors. */
2299 if (unlikely(*in_num)) {
2300 vq_err(vq, "Descriptor has out after in: "
2306 } while ((i = next_desc(vq, &desc)) != -1);
2308 /* On success, increment avail index. */
2309 vq->last_avail_idx++;
2311 /* Assume notifications from guest are disabled at this point,
2312 * if they aren't we would need to update avail_event index. */
2313 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2316 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2318 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2319 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2321 vq->last_avail_idx -= n;
2323 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2325 /* After we've used one of their buffers, we tell them about it. We'll then
2326 * want to notify the guest, using eventfd. */
2327 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2329 struct vring_used_elem heads = {
2330 cpu_to_vhost32(vq, head),
2331 cpu_to_vhost32(vq, len)
2334 return vhost_add_used_n(vq, &heads, 1);
2336 EXPORT_SYMBOL_GPL(vhost_add_used);
2338 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2339 struct vring_used_elem *heads,
2342 vring_used_elem_t __user *used;
2346 start = vq->last_used_idx & (vq->num - 1);
2347 used = vq->used->ring + start;
2348 if (vhost_put_used(vq, heads, start, count)) {
2349 vq_err(vq, "Failed to write used");
2352 if (unlikely(vq->log_used)) {
2353 /* Make sure data is seen before log. */
2355 /* Log used ring entry write. */
2356 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2357 count * sizeof *used);
2359 old = vq->last_used_idx;
2360 new = (vq->last_used_idx += count);
2361 /* If the driver never bothers to signal in a very long while,
2362 * used index might wrap around. If that happens, invalidate
2363 * signalled_used index we stored. TODO: make sure driver
2364 * signals at least once in 2^16 and remove this. */
2365 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2366 vq->signalled_used_valid = false;
2370 /* After we've used one of their buffers, we tell them about it. We'll then
2371 * want to notify the guest, using eventfd. */
2372 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2377 start = vq->last_used_idx & (vq->num - 1);
2378 n = vq->num - start;
2380 r = __vhost_add_used_n(vq, heads, n);
2386 r = __vhost_add_used_n(vq, heads, count);
2388 /* Make sure buffer is written before we update index. */
2390 if (vhost_put_used_idx(vq)) {
2391 vq_err(vq, "Failed to increment used idx");
2394 if (unlikely(vq->log_used)) {
2395 /* Make sure used idx is seen before log. */
2397 /* Log used index update. */
2398 log_used(vq, offsetof(struct vring_used, idx),
2399 sizeof vq->used->idx);
2401 eventfd_signal(vq->log_ctx, 1);
2405 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2407 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2412 /* Flush out used index updates. This is paired
2413 * with the barrier that the Guest executes when enabling
2417 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2418 unlikely(vq->avail_idx == vq->last_avail_idx))
2421 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2423 if (vhost_get_avail_flags(vq, &flags)) {
2424 vq_err(vq, "Failed to get flags");
2427 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2429 old = vq->signalled_used;
2430 v = vq->signalled_used_valid;
2431 new = vq->signalled_used = vq->last_used_idx;
2432 vq->signalled_used_valid = true;
2437 if (vhost_get_used_event(vq, &event)) {
2438 vq_err(vq, "Failed to get used event idx");
2441 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2444 /* This actually signals the guest, using eventfd. */
2445 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2447 /* Signal the Guest tell them we used something up. */
2448 if (vq->call_ctx.ctx && vhost_notify(dev, vq))
2449 eventfd_signal(vq->call_ctx.ctx, 1);
2451 EXPORT_SYMBOL_GPL(vhost_signal);
2453 /* And here's the combo meal deal. Supersize me! */
2454 void vhost_add_used_and_signal(struct vhost_dev *dev,
2455 struct vhost_virtqueue *vq,
2456 unsigned int head, int len)
2458 vhost_add_used(vq, head, len);
2459 vhost_signal(dev, vq);
2461 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2463 /* multi-buffer version of vhost_add_used_and_signal */
2464 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2465 struct vhost_virtqueue *vq,
2466 struct vring_used_elem *heads, unsigned count)
2468 vhost_add_used_n(vq, heads, count);
2469 vhost_signal(dev, vq);
2471 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2473 /* return true if we're sure that avaiable ring is empty */
2474 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2476 __virtio16 avail_idx;
2479 if (vq->avail_idx != vq->last_avail_idx)
2482 r = vhost_get_avail_idx(vq, &avail_idx);
2485 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2487 return vq->avail_idx == vq->last_avail_idx;
2489 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2491 /* OK, now we need to know about added descriptors. */
2492 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2494 __virtio16 avail_idx;
2497 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2499 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2500 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2501 r = vhost_update_used_flags(vq);
2503 vq_err(vq, "Failed to enable notification at %p: %d\n",
2504 &vq->used->flags, r);
2508 r = vhost_update_avail_event(vq, vq->avail_idx);
2510 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2511 vhost_avail_event(vq), r);
2515 /* They could have slipped one in as we were doing that: make
2516 * sure it's written, then check again. */
2518 r = vhost_get_avail_idx(vq, &avail_idx);
2520 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2521 &vq->avail->idx, r);
2525 return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2527 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2529 /* We don't need to be notified again. */
2530 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2534 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2536 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2537 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2538 r = vhost_update_used_flags(vq);
2540 vq_err(vq, "Failed to enable notification at %p: %d\n",
2541 &vq->used->flags, r);
2544 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2546 /* Create a new message. */
2547 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2549 struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2553 /* Make sure all padding within the structure is initialized. */
2554 memset(&node->msg, 0, sizeof node->msg);
2556 node->msg.type = type;
2559 EXPORT_SYMBOL_GPL(vhost_new_msg);
2561 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2562 struct vhost_msg_node *node)
2564 spin_lock(&dev->iotlb_lock);
2565 list_add_tail(&node->node, head);
2566 spin_unlock(&dev->iotlb_lock);
2568 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2570 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2572 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2573 struct list_head *head)
2575 struct vhost_msg_node *node = NULL;
2577 spin_lock(&dev->iotlb_lock);
2578 if (!list_empty(head)) {
2579 node = list_first_entry(head, struct vhost_msg_node,
2581 list_del(&node->node);
2583 spin_unlock(&dev->iotlb_lock);
2587 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2589 void vhost_set_backend_features(struct vhost_dev *dev, u64 features)
2591 struct vhost_virtqueue *vq;
2594 mutex_lock(&dev->mutex);
2595 for (i = 0; i < dev->nvqs; ++i) {
2597 mutex_lock(&vq->mutex);
2598 vq->acked_backend_features = features;
2599 mutex_unlock(&vq->mutex);
2601 mutex_unlock(&dev->mutex);
2603 EXPORT_SYMBOL_GPL(vhost_set_backend_features);
2605 static int __init vhost_init(void)
2610 static void __exit vhost_exit(void)
2614 module_init(vhost_init);
2615 module_exit(vhost_exit);
2617 MODULE_VERSION("0.0.1");
2618 MODULE_LICENSE("GPL v2");
2619 MODULE_AUTHOR("Michael S. Tsirkin");
2620 MODULE_DESCRIPTION("Host kernel accelerator for virtio");