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_dev_flush(struct vhost_dev *dev)
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_dev_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_dev_flush(poll->dev);
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));
307 bool vhost_vq_is_setup(struct vhost_virtqueue *vq)
309 return vq->avail && vq->desc && vq->used && vhost_vq_access_ok(vq);
311 EXPORT_SYMBOL_GPL(vhost_vq_is_setup);
313 static void vhost_vq_reset(struct vhost_dev *dev,
314 struct vhost_virtqueue *vq)
320 vq->last_avail_idx = 0;
322 vq->last_used_idx = 0;
323 vq->signalled_used = 0;
324 vq->signalled_used_valid = false;
326 vq->log_used = false;
327 vq->log_addr = -1ull;
328 vq->private_data = NULL;
329 vq->acked_features = 0;
330 vq->acked_backend_features = 0;
332 vq->error_ctx = NULL;
335 vhost_disable_cross_endian(vq);
336 vhost_reset_is_le(vq);
337 vq->busyloop_timeout = 0;
340 vhost_vring_call_reset(&vq->call_ctx);
341 __vhost_vq_meta_reset(vq);
344 static int vhost_worker(void *data)
346 struct vhost_dev *dev = data;
347 struct vhost_work *work, *work_next;
348 struct llist_node *node;
350 kthread_use_mm(dev->mm);
353 /* mb paired w/ kthread_stop */
354 set_current_state(TASK_INTERRUPTIBLE);
356 if (kthread_should_stop()) {
357 __set_current_state(TASK_RUNNING);
361 node = llist_del_all(&dev->work_list);
365 node = llist_reverse_order(node);
366 /* make sure flag is seen after deletion */
368 llist_for_each_entry_safe(work, work_next, node, node) {
369 clear_bit(VHOST_WORK_QUEUED, &work->flags);
370 __set_current_state(TASK_RUNNING);
371 kcov_remote_start_common(dev->kcov_handle);
378 kthread_unuse_mm(dev->mm);
382 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
392 /* Helper to allocate iovec buffers for all vqs. */
393 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
395 struct vhost_virtqueue *vq;
398 for (i = 0; i < dev->nvqs; ++i) {
400 vq->indirect = kmalloc_array(UIO_MAXIOV,
401 sizeof(*vq->indirect),
403 vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
405 vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
407 if (!vq->indirect || !vq->log || !vq->heads)
414 vhost_vq_free_iovecs(dev->vqs[i]);
418 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
422 for (i = 0; i < dev->nvqs; ++i)
423 vhost_vq_free_iovecs(dev->vqs[i]);
426 bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
427 int pkts, int total_len)
429 struct vhost_dev *dev = vq->dev;
431 if ((dev->byte_weight && total_len >= dev->byte_weight) ||
432 pkts >= dev->weight) {
433 vhost_poll_queue(&vq->poll);
439 EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
441 static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
444 size_t event __maybe_unused =
445 vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
447 return sizeof(*vq->avail) +
448 sizeof(*vq->avail->ring) * num + event;
451 static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
454 size_t event __maybe_unused =
455 vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
457 return sizeof(*vq->used) +
458 sizeof(*vq->used->ring) * num + event;
461 static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
464 return sizeof(*vq->desc) * num;
467 void vhost_dev_init(struct vhost_dev *dev,
468 struct vhost_virtqueue **vqs, int nvqs,
469 int iov_limit, int weight, int byte_weight,
471 int (*msg_handler)(struct vhost_dev *dev,
472 struct vhost_iotlb_msg *msg))
474 struct vhost_virtqueue *vq;
479 mutex_init(&dev->mutex);
485 dev->iov_limit = iov_limit;
486 dev->weight = weight;
487 dev->byte_weight = byte_weight;
488 dev->use_worker = use_worker;
489 dev->msg_handler = msg_handler;
490 init_llist_head(&dev->work_list);
491 init_waitqueue_head(&dev->wait);
492 INIT_LIST_HEAD(&dev->read_list);
493 INIT_LIST_HEAD(&dev->pending_list);
494 spin_lock_init(&dev->iotlb_lock);
497 for (i = 0; i < dev->nvqs; ++i) {
503 mutex_init(&vq->mutex);
504 vhost_vq_reset(dev, vq);
506 vhost_poll_init(&vq->poll, vq->handle_kick,
510 EXPORT_SYMBOL_GPL(vhost_dev_init);
512 /* Caller should have device mutex */
513 long vhost_dev_check_owner(struct vhost_dev *dev)
515 /* Are you the owner? If not, I don't think you mean to do that */
516 return dev->mm == current->mm ? 0 : -EPERM;
518 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
520 struct vhost_attach_cgroups_struct {
521 struct vhost_work work;
522 struct task_struct *owner;
526 static void vhost_attach_cgroups_work(struct vhost_work *work)
528 struct vhost_attach_cgroups_struct *s;
530 s = container_of(work, struct vhost_attach_cgroups_struct, work);
531 s->ret = cgroup_attach_task_all(s->owner, current);
534 static int vhost_attach_cgroups(struct vhost_dev *dev)
536 struct vhost_attach_cgroups_struct attach;
538 attach.owner = current;
539 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
540 vhost_work_queue(dev, &attach.work);
541 vhost_work_dev_flush(dev);
545 /* Caller should have device mutex */
546 bool vhost_dev_has_owner(struct vhost_dev *dev)
550 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
552 static void vhost_attach_mm(struct vhost_dev *dev)
554 /* No owner, become one */
555 if (dev->use_worker) {
556 dev->mm = get_task_mm(current);
558 /* vDPA device does not use worker thead, so there's
559 * no need to hold the address space for mm. This help
560 * to avoid deadlock in the case of mmap() which may
561 * held the refcnt of the file and depends on release
562 * method to remove vma.
564 dev->mm = current->mm;
569 static void vhost_detach_mm(struct vhost_dev *dev)
582 /* Caller should have device mutex */
583 long vhost_dev_set_owner(struct vhost_dev *dev)
585 struct task_struct *worker;
588 /* Is there an owner already? */
589 if (vhost_dev_has_owner(dev)) {
594 vhost_attach_mm(dev);
596 dev->kcov_handle = kcov_common_handle();
597 if (dev->use_worker) {
598 worker = kthread_create(vhost_worker, dev,
599 "vhost-%d", current->pid);
600 if (IS_ERR(worker)) {
601 err = PTR_ERR(worker);
605 dev->worker = worker;
606 wake_up_process(worker); /* avoid contributing to loadavg */
608 err = vhost_attach_cgroups(dev);
613 err = vhost_dev_alloc_iovecs(dev);
620 kthread_stop(dev->worker);
624 vhost_detach_mm(dev);
625 dev->kcov_handle = 0;
629 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
631 static struct vhost_iotlb *iotlb_alloc(void)
633 return vhost_iotlb_alloc(max_iotlb_entries,
634 VHOST_IOTLB_FLAG_RETIRE);
637 struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
639 return iotlb_alloc();
641 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
643 /* Caller should have device mutex */
644 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
648 vhost_dev_cleanup(dev);
651 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
652 * VQs aren't running.
654 for (i = 0; i < dev->nvqs; ++i)
655 dev->vqs[i]->umem = umem;
657 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
659 void vhost_dev_stop(struct vhost_dev *dev)
663 for (i = 0; i < dev->nvqs; ++i) {
664 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
665 vhost_poll_stop(&dev->vqs[i]->poll);
666 vhost_poll_flush(&dev->vqs[i]->poll);
670 EXPORT_SYMBOL_GPL(vhost_dev_stop);
672 static void vhost_clear_msg(struct vhost_dev *dev)
674 struct vhost_msg_node *node, *n;
676 spin_lock(&dev->iotlb_lock);
678 list_for_each_entry_safe(node, n, &dev->read_list, node) {
679 list_del(&node->node);
683 list_for_each_entry_safe(node, n, &dev->pending_list, node) {
684 list_del(&node->node);
688 spin_unlock(&dev->iotlb_lock);
691 void vhost_dev_cleanup(struct vhost_dev *dev)
695 for (i = 0; i < dev->nvqs; ++i) {
696 if (dev->vqs[i]->error_ctx)
697 eventfd_ctx_put(dev->vqs[i]->error_ctx);
698 if (dev->vqs[i]->kick)
699 fput(dev->vqs[i]->kick);
700 if (dev->vqs[i]->call_ctx.ctx)
701 eventfd_ctx_put(dev->vqs[i]->call_ctx.ctx);
702 vhost_vq_reset(dev, dev->vqs[i]);
704 vhost_dev_free_iovecs(dev);
706 eventfd_ctx_put(dev->log_ctx);
708 /* No one will access memory at this point */
709 vhost_iotlb_free(dev->umem);
711 vhost_iotlb_free(dev->iotlb);
713 vhost_clear_msg(dev);
714 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
715 WARN_ON(!llist_empty(&dev->work_list));
717 kthread_stop(dev->worker);
719 dev->kcov_handle = 0;
721 vhost_detach_mm(dev);
723 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
725 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
727 u64 a = addr / VHOST_PAGE_SIZE / 8;
729 /* Make sure 64 bit math will not overflow. */
730 if (a > ULONG_MAX - (unsigned long)log_base ||
731 a + (unsigned long)log_base > ULONG_MAX)
734 return access_ok(log_base + a,
735 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
738 /* Make sure 64 bit math will not overflow. */
739 static bool vhost_overflow(u64 uaddr, u64 size)
741 if (uaddr > ULONG_MAX || size > ULONG_MAX)
747 return uaddr > ULONG_MAX - size + 1;
750 /* Caller should have vq mutex and device mutex. */
751 static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
754 struct vhost_iotlb_map *map;
759 list_for_each_entry(map, &umem->list, link) {
760 unsigned long a = map->addr;
762 if (vhost_overflow(map->addr, map->size))
766 if (!access_ok((void __user *)a, map->size))
768 else if (log_all && !log_access_ok(log_base,
776 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
777 u64 addr, unsigned int size,
780 const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
785 return (void __user *)(uintptr_t)(map->addr + addr - map->start);
788 /* Can we switch to this memory table? */
789 /* Caller should have device mutex but not vq mutex */
790 static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
795 for (i = 0; i < d->nvqs; ++i) {
799 mutex_lock(&d->vqs[i]->mutex);
800 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
801 /* If ring is inactive, will check when it's enabled. */
802 if (d->vqs[i]->private_data)
803 ok = vq_memory_access_ok(d->vqs[i]->log_base,
807 mutex_unlock(&d->vqs[i]->mutex);
814 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
815 struct iovec iov[], int iov_size, int access);
817 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
818 const void *from, unsigned size)
823 return __copy_to_user(to, from, size);
825 /* This function should be called after iotlb
826 * prefetch, which means we're sure that all vq
827 * could be access through iotlb. So -EAGAIN should
828 * not happen in this case.
831 void __user *uaddr = vhost_vq_meta_fetch(vq,
832 (u64)(uintptr_t)to, size,
836 return __copy_to_user(uaddr, from, size);
838 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
839 ARRAY_SIZE(vq->iotlb_iov),
843 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
844 ret = copy_to_iter(from, size, &t);
852 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
853 void __user *from, unsigned size)
858 return __copy_from_user(to, from, size);
860 /* This function should be called after iotlb
861 * prefetch, which means we're sure that vq
862 * could be access through iotlb. So -EAGAIN should
863 * not happen in this case.
865 void __user *uaddr = vhost_vq_meta_fetch(vq,
866 (u64)(uintptr_t)from, size,
871 return __copy_from_user(to, uaddr, size);
873 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
874 ARRAY_SIZE(vq->iotlb_iov),
877 vq_err(vq, "IOTLB translation failure: uaddr "
878 "%p size 0x%llx\n", from,
879 (unsigned long long) size);
882 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
883 ret = copy_from_iter(to, size, &f);
892 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
893 void __user *addr, unsigned int size,
898 ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
899 ARRAY_SIZE(vq->iotlb_iov),
902 vq_err(vq, "IOTLB translation failure: uaddr "
903 "%p size 0x%llx\n", addr,
904 (unsigned long long) size);
908 if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
909 vq_err(vq, "Non atomic userspace memory access: uaddr "
910 "%p size 0x%llx\n", addr,
911 (unsigned long long) size);
915 return vq->iotlb_iov[0].iov_base;
918 /* This function should be called after iotlb
919 * prefetch, which means we're sure that vq
920 * could be access through iotlb. So -EAGAIN should
921 * not happen in this case.
923 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
924 void __user *addr, unsigned int size,
927 void __user *uaddr = vhost_vq_meta_fetch(vq,
928 (u64)(uintptr_t)addr, size, type);
932 return __vhost_get_user_slow(vq, addr, size, type);
935 #define vhost_put_user(vq, x, ptr) \
939 ret = __put_user(x, ptr); \
941 __typeof__(ptr) to = \
942 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
943 sizeof(*ptr), VHOST_ADDR_USED); \
945 ret = __put_user(x, to); \
952 static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
954 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
955 vhost_avail_event(vq));
958 static inline int vhost_put_used(struct vhost_virtqueue *vq,
959 struct vring_used_elem *head, int idx,
962 return vhost_copy_to_user(vq, vq->used->ring + idx, head,
963 count * sizeof(*head));
966 static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
969 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
973 static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
976 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
980 #define vhost_get_user(vq, x, ptr, type) \
984 ret = __get_user(x, ptr); \
986 __typeof__(ptr) from = \
987 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
991 ret = __get_user(x, from); \
998 #define vhost_get_avail(vq, x, ptr) \
999 vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
1001 #define vhost_get_used(vq, x, ptr) \
1002 vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
1004 static void vhost_dev_lock_vqs(struct vhost_dev *d)
1007 for (i = 0; i < d->nvqs; ++i)
1008 mutex_lock_nested(&d->vqs[i]->mutex, i);
1011 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
1014 for (i = 0; i < d->nvqs; ++i)
1015 mutex_unlock(&d->vqs[i]->mutex);
1018 static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
1021 return vhost_get_avail(vq, *idx, &vq->avail->idx);
1024 static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
1025 __virtio16 *head, int idx)
1027 return vhost_get_avail(vq, *head,
1028 &vq->avail->ring[idx & (vq->num - 1)]);
1031 static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
1034 return vhost_get_avail(vq, *flags, &vq->avail->flags);
1037 static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1040 return vhost_get_avail(vq, *event, vhost_used_event(vq));
1043 static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1046 return vhost_get_used(vq, *idx, &vq->used->idx);
1049 static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1050 struct vring_desc *desc, int idx)
1052 return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1055 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1056 struct vhost_iotlb_msg *msg)
1058 struct vhost_msg_node *node, *n;
1060 spin_lock(&d->iotlb_lock);
1062 list_for_each_entry_safe(node, n, &d->pending_list, node) {
1063 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1064 if (msg->iova <= vq_msg->iova &&
1065 msg->iova + msg->size - 1 >= vq_msg->iova &&
1066 vq_msg->type == VHOST_IOTLB_MISS) {
1067 vhost_poll_queue(&node->vq->poll);
1068 list_del(&node->node);
1073 spin_unlock(&d->iotlb_lock);
1076 static bool umem_access_ok(u64 uaddr, u64 size, int access)
1078 unsigned long a = uaddr;
1080 /* Make sure 64 bit math will not overflow. */
1081 if (vhost_overflow(uaddr, size))
1084 if ((access & VHOST_ACCESS_RO) &&
1085 !access_ok((void __user *)a, size))
1087 if ((access & VHOST_ACCESS_WO) &&
1088 !access_ok((void __user *)a, size))
1093 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
1094 struct vhost_iotlb_msg *msg)
1098 mutex_lock(&dev->mutex);
1099 vhost_dev_lock_vqs(dev);
1100 switch (msg->type) {
1101 case VHOST_IOTLB_UPDATE:
1106 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1110 vhost_vq_meta_reset(dev);
1111 if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1112 msg->iova + msg->size - 1,
1113 msg->uaddr, msg->perm)) {
1117 vhost_iotlb_notify_vq(dev, msg);
1119 case VHOST_IOTLB_INVALIDATE:
1124 vhost_vq_meta_reset(dev);
1125 vhost_iotlb_del_range(dev->iotlb, msg->iova,
1126 msg->iova + msg->size - 1);
1133 vhost_dev_unlock_vqs(dev);
1134 mutex_unlock(&dev->mutex);
1138 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1139 struct iov_iter *from)
1141 struct vhost_iotlb_msg msg;
1145 ret = copy_from_iter(&type, sizeof(type), from);
1146 if (ret != sizeof(type)) {
1152 case VHOST_IOTLB_MSG:
1153 /* There maybe a hole after type for V1 message type,
1156 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1158 case VHOST_IOTLB_MSG_V2:
1159 offset = sizeof(__u32);
1166 iov_iter_advance(from, offset);
1167 ret = copy_from_iter(&msg, sizeof(msg), from);
1168 if (ret != sizeof(msg)) {
1173 if (msg.size == 0) {
1178 if (dev->msg_handler)
1179 ret = dev->msg_handler(dev, &msg);
1181 ret = vhost_process_iotlb_msg(dev, &msg);
1187 ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1188 sizeof(struct vhost_msg_v2);
1192 EXPORT_SYMBOL(vhost_chr_write_iter);
1194 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1199 poll_wait(file, &dev->wait, wait);
1201 if (!list_empty(&dev->read_list))
1202 mask |= EPOLLIN | EPOLLRDNORM;
1206 EXPORT_SYMBOL(vhost_chr_poll);
1208 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1212 struct vhost_msg_node *node;
1214 unsigned size = sizeof(struct vhost_msg);
1216 if (iov_iter_count(to) < size)
1221 prepare_to_wait(&dev->wait, &wait,
1222 TASK_INTERRUPTIBLE);
1224 node = vhost_dequeue_msg(dev, &dev->read_list);
1231 if (signal_pending(current)) {
1244 finish_wait(&dev->wait, &wait);
1247 struct vhost_iotlb_msg *msg;
1248 void *start = &node->msg;
1250 switch (node->msg.type) {
1251 case VHOST_IOTLB_MSG:
1252 size = sizeof(node->msg);
1253 msg = &node->msg.iotlb;
1255 case VHOST_IOTLB_MSG_V2:
1256 size = sizeof(node->msg_v2);
1257 msg = &node->msg_v2.iotlb;
1264 ret = copy_to_iter(start, size, to);
1265 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1269 vhost_enqueue_msg(dev, &dev->pending_list, node);
1274 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1276 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1278 struct vhost_dev *dev = vq->dev;
1279 struct vhost_msg_node *node;
1280 struct vhost_iotlb_msg *msg;
1281 bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1283 node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1288 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1289 msg = &node->msg_v2.iotlb;
1291 msg = &node->msg.iotlb;
1294 msg->type = VHOST_IOTLB_MISS;
1298 vhost_enqueue_msg(dev, &dev->read_list, node);
1303 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1304 vring_desc_t __user *desc,
1305 vring_avail_t __user *avail,
1306 vring_used_t __user *used)
1309 /* If an IOTLB device is present, the vring addresses are
1310 * GIOVAs. Access validation occurs at prefetch time. */
1314 return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1315 access_ok(avail, vhost_get_avail_size(vq, num)) &&
1316 access_ok(used, vhost_get_used_size(vq, num));
1319 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1320 const struct vhost_iotlb_map *map,
1323 int access = (type == VHOST_ADDR_USED) ?
1324 VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1326 if (likely(map->perm & access))
1327 vq->meta_iotlb[type] = map;
1330 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1331 int access, u64 addr, u64 len, int type)
1333 const struct vhost_iotlb_map *map;
1334 struct vhost_iotlb *umem = vq->iotlb;
1335 u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1337 if (vhost_vq_meta_fetch(vq, addr, len, type))
1341 map = vhost_iotlb_itree_first(umem, addr, last);
1342 if (map == NULL || map->start > addr) {
1343 vhost_iotlb_miss(vq, addr, access);
1345 } else if (!(map->perm & access)) {
1346 /* Report the possible access violation by
1347 * request another translation from userspace.
1352 size = map->size - addr + map->start;
1354 if (orig_addr == addr && size >= len)
1355 vhost_vq_meta_update(vq, map, type);
1364 int vq_meta_prefetch(struct vhost_virtqueue *vq)
1366 unsigned int num = vq->num;
1371 return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1372 vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1373 iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1374 vhost_get_avail_size(vq, num),
1375 VHOST_ADDR_AVAIL) &&
1376 iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1377 vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1379 EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1381 /* Can we log writes? */
1382 /* Caller should have device mutex but not vq mutex */
1383 bool vhost_log_access_ok(struct vhost_dev *dev)
1385 return memory_access_ok(dev, dev->umem, 1);
1387 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1389 static bool vq_log_used_access_ok(struct vhost_virtqueue *vq,
1390 void __user *log_base,
1394 /* If an IOTLB device is present, log_addr is a GIOVA that
1395 * will never be logged by log_used(). */
1399 return !log_used || log_access_ok(log_base, log_addr,
1400 vhost_get_used_size(vq, vq->num));
1403 /* Verify access for write logging. */
1404 /* Caller should have vq mutex and device mutex */
1405 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1406 void __user *log_base)
1408 return vq_memory_access_ok(log_base, vq->umem,
1409 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1410 vq_log_used_access_ok(vq, log_base, vq->log_used, vq->log_addr);
1413 /* Can we start vq? */
1414 /* Caller should have vq mutex and device mutex */
1415 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1417 if (!vq_log_access_ok(vq, vq->log_base))
1420 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1422 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1424 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1426 struct vhost_memory mem, *newmem;
1427 struct vhost_memory_region *region;
1428 struct vhost_iotlb *newumem, *oldumem;
1429 unsigned long size = offsetof(struct vhost_memory, regions);
1432 if (copy_from_user(&mem, m, size))
1436 if (mem.nregions > max_mem_regions)
1438 newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1443 memcpy(newmem, &mem, size);
1444 if (copy_from_user(newmem->regions, m->regions,
1445 flex_array_size(newmem, regions, mem.nregions))) {
1450 newumem = iotlb_alloc();
1456 for (region = newmem->regions;
1457 region < newmem->regions + mem.nregions;
1459 if (vhost_iotlb_add_range(newumem,
1460 region->guest_phys_addr,
1461 region->guest_phys_addr +
1462 region->memory_size - 1,
1463 region->userspace_addr,
1468 if (!memory_access_ok(d, newumem, 0))
1474 /* All memory accesses are done under some VQ mutex. */
1475 for (i = 0; i < d->nvqs; ++i) {
1476 mutex_lock(&d->vqs[i]->mutex);
1477 d->vqs[i]->umem = newumem;
1478 mutex_unlock(&d->vqs[i]->mutex);
1482 vhost_iotlb_free(oldumem);
1486 vhost_iotlb_free(newumem);
1491 static long vhost_vring_set_num(struct vhost_dev *d,
1492 struct vhost_virtqueue *vq,
1495 struct vhost_vring_state s;
1497 /* Resizing ring with an active backend?
1498 * You don't want to do that. */
1499 if (vq->private_data)
1502 if (copy_from_user(&s, argp, sizeof s))
1505 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1512 static long vhost_vring_set_addr(struct vhost_dev *d,
1513 struct vhost_virtqueue *vq,
1516 struct vhost_vring_addr a;
1518 if (copy_from_user(&a, argp, sizeof a))
1520 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1523 /* For 32bit, verify that the top 32bits of the user
1524 data are set to zero. */
1525 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1526 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1527 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1530 /* Make sure it's safe to cast pointers to vring types. */
1531 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1532 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1533 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1534 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1535 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1538 /* We only verify access here if backend is configured.
1539 * If it is not, we don't as size might not have been setup.
1540 * We will verify when backend is configured. */
1541 if (vq->private_data) {
1542 if (!vq_access_ok(vq, vq->num,
1543 (void __user *)(unsigned long)a.desc_user_addr,
1544 (void __user *)(unsigned long)a.avail_user_addr,
1545 (void __user *)(unsigned long)a.used_user_addr))
1548 /* Also validate log access for used ring if enabled. */
1549 if (!vq_log_used_access_ok(vq, vq->log_base,
1550 a.flags & (0x1 << VHOST_VRING_F_LOG),
1555 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1556 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1557 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1558 vq->log_addr = a.log_guest_addr;
1559 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1564 static long vhost_vring_set_num_addr(struct vhost_dev *d,
1565 struct vhost_virtqueue *vq,
1571 mutex_lock(&vq->mutex);
1574 case VHOST_SET_VRING_NUM:
1575 r = vhost_vring_set_num(d, vq, argp);
1577 case VHOST_SET_VRING_ADDR:
1578 r = vhost_vring_set_addr(d, vq, argp);
1584 mutex_unlock(&vq->mutex);
1588 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1590 struct file *eventfp, *filep = NULL;
1591 bool pollstart = false, pollstop = false;
1592 struct eventfd_ctx *ctx = NULL;
1593 u32 __user *idxp = argp;
1594 struct vhost_virtqueue *vq;
1595 struct vhost_vring_state s;
1596 struct vhost_vring_file f;
1600 r = get_user(idx, idxp);
1606 idx = array_index_nospec(idx, d->nvqs);
1609 if (ioctl == VHOST_SET_VRING_NUM ||
1610 ioctl == VHOST_SET_VRING_ADDR) {
1611 return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1614 mutex_lock(&vq->mutex);
1617 case VHOST_SET_VRING_BASE:
1618 /* Moving base with an active backend?
1619 * You don't want to do that. */
1620 if (vq->private_data) {
1624 if (copy_from_user(&s, argp, sizeof s)) {
1628 if (s.num > 0xffff) {
1632 vq->last_avail_idx = s.num;
1633 /* Forget the cached index value. */
1634 vq->avail_idx = vq->last_avail_idx;
1636 case VHOST_GET_VRING_BASE:
1638 s.num = vq->last_avail_idx;
1639 if (copy_to_user(argp, &s, sizeof s))
1642 case VHOST_SET_VRING_KICK:
1643 if (copy_from_user(&f, argp, sizeof f)) {
1647 eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
1648 if (IS_ERR(eventfp)) {
1649 r = PTR_ERR(eventfp);
1652 if (eventfp != vq->kick) {
1653 pollstop = (filep = vq->kick) != NULL;
1654 pollstart = (vq->kick = eventfp) != NULL;
1658 case VHOST_SET_VRING_CALL:
1659 if (copy_from_user(&f, argp, sizeof f)) {
1663 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1669 swap(ctx, vq->call_ctx.ctx);
1671 case VHOST_SET_VRING_ERR:
1672 if (copy_from_user(&f, argp, sizeof f)) {
1676 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1681 swap(ctx, vq->error_ctx);
1683 case VHOST_SET_VRING_ENDIAN:
1684 r = vhost_set_vring_endian(vq, argp);
1686 case VHOST_GET_VRING_ENDIAN:
1687 r = vhost_get_vring_endian(vq, idx, argp);
1689 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1690 if (copy_from_user(&s, argp, sizeof(s))) {
1694 vq->busyloop_timeout = s.num;
1696 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1698 s.num = vq->busyloop_timeout;
1699 if (copy_to_user(argp, &s, sizeof(s)))
1706 if (pollstop && vq->handle_kick)
1707 vhost_poll_stop(&vq->poll);
1709 if (!IS_ERR_OR_NULL(ctx))
1710 eventfd_ctx_put(ctx);
1714 if (pollstart && vq->handle_kick)
1715 r = vhost_poll_start(&vq->poll, vq->kick);
1717 mutex_unlock(&vq->mutex);
1719 if (pollstop && vq->handle_kick)
1720 vhost_poll_flush(&vq->poll);
1723 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1725 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1727 struct vhost_iotlb *niotlb, *oiotlb;
1730 niotlb = iotlb_alloc();
1737 for (i = 0; i < d->nvqs; ++i) {
1738 struct vhost_virtqueue *vq = d->vqs[i];
1740 mutex_lock(&vq->mutex);
1742 __vhost_vq_meta_reset(vq);
1743 mutex_unlock(&vq->mutex);
1746 vhost_iotlb_free(oiotlb);
1750 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1752 /* Caller must have device mutex */
1753 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1755 struct eventfd_ctx *ctx;
1760 /* If you are not the owner, you can become one */
1761 if (ioctl == VHOST_SET_OWNER) {
1762 r = vhost_dev_set_owner(d);
1766 /* You must be the owner to do anything else */
1767 r = vhost_dev_check_owner(d);
1772 case VHOST_SET_MEM_TABLE:
1773 r = vhost_set_memory(d, argp);
1775 case VHOST_SET_LOG_BASE:
1776 if (copy_from_user(&p, argp, sizeof p)) {
1780 if ((u64)(unsigned long)p != p) {
1784 for (i = 0; i < d->nvqs; ++i) {
1785 struct vhost_virtqueue *vq;
1786 void __user *base = (void __user *)(unsigned long)p;
1788 mutex_lock(&vq->mutex);
1789 /* If ring is inactive, will check when it's enabled. */
1790 if (vq->private_data && !vq_log_access_ok(vq, base))
1793 vq->log_base = base;
1794 mutex_unlock(&vq->mutex);
1797 case VHOST_SET_LOG_FD:
1798 r = get_user(fd, (int __user *)argp);
1801 ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
1806 swap(ctx, d->log_ctx);
1807 for (i = 0; i < d->nvqs; ++i) {
1808 mutex_lock(&d->vqs[i]->mutex);
1809 d->vqs[i]->log_ctx = d->log_ctx;
1810 mutex_unlock(&d->vqs[i]->mutex);
1813 eventfd_ctx_put(ctx);
1822 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1824 /* TODO: This is really inefficient. We need something like get_user()
1825 * (instruction directly accesses the data, with an exception table entry
1826 * returning -EFAULT). See Documentation/x86/exception-tables.rst.
1828 static int set_bit_to_user(int nr, void __user *addr)
1830 unsigned long log = (unsigned long)addr;
1833 int bit = nr + (log % PAGE_SIZE) * 8;
1836 r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
1840 base = kmap_atomic(page);
1842 kunmap_atomic(base);
1843 unpin_user_pages_dirty_lock(&page, 1, true);
1847 static int log_write(void __user *log_base,
1848 u64 write_address, u64 write_length)
1850 u64 write_page = write_address / VHOST_PAGE_SIZE;
1855 write_length += write_address % VHOST_PAGE_SIZE;
1857 u64 base = (u64)(unsigned long)log_base;
1858 u64 log = base + write_page / 8;
1859 int bit = write_page % 8;
1860 if ((u64)(unsigned long)log != log)
1862 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1865 if (write_length <= VHOST_PAGE_SIZE)
1867 write_length -= VHOST_PAGE_SIZE;
1873 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1875 struct vhost_iotlb *umem = vq->umem;
1876 struct vhost_iotlb_map *u;
1877 u64 start, end, l, min;
1883 /* More than one GPAs can be mapped into a single HVA. So
1884 * iterate all possible umems here to be safe.
1886 list_for_each_entry(u, &umem->list, link) {
1887 if (u->addr > hva - 1 + len ||
1888 u->addr - 1 + u->size < hva)
1890 start = max(u->addr, hva);
1891 end = min(u->addr - 1 + u->size, hva - 1 + len);
1892 l = end - start + 1;
1893 r = log_write(vq->log_base,
1894 u->start + start - u->addr,
1912 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1914 struct iovec *iov = vq->log_iov;
1918 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1920 ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1921 len, iov, 64, VHOST_ACCESS_WO);
1925 for (i = 0; i < ret; i++) {
1926 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1935 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1936 unsigned int log_num, u64 len, struct iovec *iov, int count)
1940 /* Make sure data written is seen before log. */
1944 for (i = 0; i < count; i++) {
1945 r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1953 for (i = 0; i < log_num; ++i) {
1954 u64 l = min(log[i].len, len);
1955 r = log_write(vq->log_base, log[i].addr, l);
1961 eventfd_signal(vq->log_ctx, 1);
1965 /* Length written exceeds what we have stored. This is a bug. */
1969 EXPORT_SYMBOL_GPL(vhost_log_write);
1971 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1974 if (vhost_put_used_flags(vq))
1976 if (unlikely(vq->log_used)) {
1977 /* Make sure the flag is seen before log. */
1979 /* Log used flag write. */
1980 used = &vq->used->flags;
1981 log_used(vq, (used - (void __user *)vq->used),
1982 sizeof vq->used->flags);
1984 eventfd_signal(vq->log_ctx, 1);
1989 static int vhost_update_avail_event(struct vhost_virtqueue *vq)
1991 if (vhost_put_avail_event(vq))
1993 if (unlikely(vq->log_used)) {
1995 /* Make sure the event is seen before log. */
1997 /* Log avail event write */
1998 used = vhost_avail_event(vq);
1999 log_used(vq, (used - (void __user *)vq->used),
2000 sizeof *vhost_avail_event(vq));
2002 eventfd_signal(vq->log_ctx, 1);
2007 int vhost_vq_init_access(struct vhost_virtqueue *vq)
2009 __virtio16 last_used_idx;
2011 bool is_le = vq->is_le;
2013 if (!vq->private_data)
2016 vhost_init_is_le(vq);
2018 r = vhost_update_used_flags(vq);
2021 vq->signalled_used_valid = false;
2023 !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2027 r = vhost_get_used_idx(vq, &last_used_idx);
2029 vq_err(vq, "Can't access used idx at %p\n",
2033 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2040 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2042 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2043 struct iovec iov[], int iov_size, int access)
2045 const struct vhost_iotlb_map *map;
2046 struct vhost_dev *dev = vq->dev;
2047 struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2052 while ((u64)len > s) {
2054 if (unlikely(ret >= iov_size)) {
2059 map = vhost_iotlb_itree_first(umem, addr, addr + len - 1);
2060 if (map == NULL || map->start > addr) {
2061 if (umem != dev->iotlb) {
2067 } else if (!(map->perm & access)) {
2073 size = map->size - addr + map->start;
2074 _iov->iov_len = min((u64)len - s, size);
2075 _iov->iov_base = (void __user *)(unsigned long)
2076 (map->addr + addr - map->start);
2083 vhost_iotlb_miss(vq, addr, access);
2087 /* Each buffer in the virtqueues is actually a chain of descriptors. This
2088 * function returns the next descriptor in the chain,
2089 * or -1U if we're at the end. */
2090 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2094 /* If this descriptor says it doesn't chain, we're done. */
2095 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2098 /* Check they're not leading us off end of descriptors. */
2099 next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2103 static int get_indirect(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,
2107 struct vring_desc *indirect)
2109 struct vring_desc desc;
2110 unsigned int i = 0, count, found = 0;
2111 u32 len = vhost32_to_cpu(vq, indirect->len);
2112 struct iov_iter from;
2116 if (unlikely(len % sizeof desc)) {
2117 vq_err(vq, "Invalid length in indirect descriptor: "
2118 "len 0x%llx not multiple of 0x%zx\n",
2119 (unsigned long long)len,
2124 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2125 UIO_MAXIOV, VHOST_ACCESS_RO);
2126 if (unlikely(ret < 0)) {
2128 vq_err(vq, "Translation failure %d in indirect.\n", ret);
2131 iov_iter_init(&from, READ, vq->indirect, ret, len);
2132 count = len / sizeof desc;
2133 /* Buffers are chained via a 16 bit next field, so
2134 * we can have at most 2^16 of these. */
2135 if (unlikely(count > USHRT_MAX + 1)) {
2136 vq_err(vq, "Indirect buffer length too big: %d\n",
2142 unsigned iov_count = *in_num + *out_num;
2143 if (unlikely(++found > count)) {
2144 vq_err(vq, "Loop detected: last one at %u "
2145 "indirect size %u\n",
2149 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2150 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2151 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2154 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2155 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2156 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2160 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2161 access = VHOST_ACCESS_WO;
2163 access = VHOST_ACCESS_RO;
2165 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2166 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2167 iov_size - iov_count, access);
2168 if (unlikely(ret < 0)) {
2170 vq_err(vq, "Translation failure %d indirect idx %d\n",
2174 /* If this is an input descriptor, increment that count. */
2175 if (access == VHOST_ACCESS_WO) {
2177 if (unlikely(log && ret)) {
2178 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2179 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2183 /* If it's an output descriptor, they're all supposed
2184 * to come before any input descriptors. */
2185 if (unlikely(*in_num)) {
2186 vq_err(vq, "Indirect descriptor "
2187 "has out after in: idx %d\n", i);
2192 } while ((i = next_desc(vq, &desc)) != -1);
2196 /* This looks in the virtqueue and for the first available buffer, and converts
2197 * it to an iovec for convenient access. Since descriptors consist of some
2198 * number of output then some number of input descriptors, it's actually two
2199 * iovecs, but we pack them into one and note how many of each there were.
2201 * This function returns the descriptor number found, or vq->num (which is
2202 * never a valid descriptor number) if none was found. A negative code is
2203 * returned on error. */
2204 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2205 struct iovec iov[], unsigned int iov_size,
2206 unsigned int *out_num, unsigned int *in_num,
2207 struct vhost_log *log, unsigned int *log_num)
2209 struct vring_desc desc;
2210 unsigned int i, head, found = 0;
2212 __virtio16 avail_idx;
2213 __virtio16 ring_head;
2216 /* Check it isn't doing very strange things with descriptor numbers. */
2217 last_avail_idx = vq->last_avail_idx;
2219 if (vq->avail_idx == vq->last_avail_idx) {
2220 if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2221 vq_err(vq, "Failed to access avail idx at %p\n",
2225 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2227 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2228 vq_err(vq, "Guest moved used index from %u to %u",
2229 last_avail_idx, vq->avail_idx);
2233 /* If there's nothing new since last we looked, return
2236 if (vq->avail_idx == last_avail_idx)
2239 /* Only get avail ring entries after they have been
2245 /* Grab the next descriptor number they're advertising, and increment
2246 * the index we've seen. */
2247 if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2248 vq_err(vq, "Failed to read head: idx %d address %p\n",
2250 &vq->avail->ring[last_avail_idx % vq->num]);
2254 head = vhost16_to_cpu(vq, ring_head);
2256 /* If their number is silly, that's an error. */
2257 if (unlikely(head >= vq->num)) {
2258 vq_err(vq, "Guest says index %u > %u is available",
2263 /* When we start there are none of either input nor output. */
2264 *out_num = *in_num = 0;
2270 unsigned iov_count = *in_num + *out_num;
2271 if (unlikely(i >= vq->num)) {
2272 vq_err(vq, "Desc index is %u > %u, head = %u",
2276 if (unlikely(++found > vq->num)) {
2277 vq_err(vq, "Loop detected: last one at %u "
2278 "vq size %u head %u\n",
2282 ret = vhost_get_desc(vq, &desc, i);
2283 if (unlikely(ret)) {
2284 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2288 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2289 ret = get_indirect(vq, iov, iov_size,
2291 log, log_num, &desc);
2292 if (unlikely(ret < 0)) {
2294 vq_err(vq, "Failure detected "
2295 "in indirect descriptor at idx %d\n", i);
2301 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2302 access = VHOST_ACCESS_WO;
2304 access = VHOST_ACCESS_RO;
2305 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2306 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2307 iov_size - iov_count, access);
2308 if (unlikely(ret < 0)) {
2310 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2314 if (access == VHOST_ACCESS_WO) {
2315 /* If this is an input descriptor,
2316 * increment that count. */
2318 if (unlikely(log && ret)) {
2319 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2320 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2324 /* If it's an output descriptor, they're all supposed
2325 * to come before any input descriptors. */
2326 if (unlikely(*in_num)) {
2327 vq_err(vq, "Descriptor has out after in: "
2333 } while ((i = next_desc(vq, &desc)) != -1);
2335 /* On success, increment avail index. */
2336 vq->last_avail_idx++;
2338 /* Assume notifications from guest are disabled at this point,
2339 * if they aren't we would need to update avail_event index. */
2340 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2343 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2345 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2346 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2348 vq->last_avail_idx -= n;
2350 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2352 /* After we've used one of their buffers, we tell them about it. We'll then
2353 * want to notify the guest, using eventfd. */
2354 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2356 struct vring_used_elem heads = {
2357 cpu_to_vhost32(vq, head),
2358 cpu_to_vhost32(vq, len)
2361 return vhost_add_used_n(vq, &heads, 1);
2363 EXPORT_SYMBOL_GPL(vhost_add_used);
2365 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2366 struct vring_used_elem *heads,
2369 vring_used_elem_t __user *used;
2373 start = vq->last_used_idx & (vq->num - 1);
2374 used = vq->used->ring + start;
2375 if (vhost_put_used(vq, heads, start, count)) {
2376 vq_err(vq, "Failed to write used");
2379 if (unlikely(vq->log_used)) {
2380 /* Make sure data is seen before log. */
2382 /* Log used ring entry write. */
2383 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2384 count * sizeof *used);
2386 old = vq->last_used_idx;
2387 new = (vq->last_used_idx += count);
2388 /* If the driver never bothers to signal in a very long while,
2389 * used index might wrap around. If that happens, invalidate
2390 * signalled_used index we stored. TODO: make sure driver
2391 * signals at least once in 2^16 and remove this. */
2392 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2393 vq->signalled_used_valid = false;
2397 /* After we've used one of their buffers, we tell them about it. We'll then
2398 * want to notify the guest, using eventfd. */
2399 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2404 start = vq->last_used_idx & (vq->num - 1);
2405 n = vq->num - start;
2407 r = __vhost_add_used_n(vq, heads, n);
2413 r = __vhost_add_used_n(vq, heads, count);
2415 /* Make sure buffer is written before we update index. */
2417 if (vhost_put_used_idx(vq)) {
2418 vq_err(vq, "Failed to increment used idx");
2421 if (unlikely(vq->log_used)) {
2422 /* Make sure used idx is seen before log. */
2424 /* Log used index update. */
2425 log_used(vq, offsetof(struct vring_used, idx),
2426 sizeof vq->used->idx);
2428 eventfd_signal(vq->log_ctx, 1);
2432 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2434 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2439 /* Flush out used index updates. This is paired
2440 * with the barrier that the Guest executes when enabling
2444 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2445 unlikely(vq->avail_idx == vq->last_avail_idx))
2448 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2450 if (vhost_get_avail_flags(vq, &flags)) {
2451 vq_err(vq, "Failed to get flags");
2454 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2456 old = vq->signalled_used;
2457 v = vq->signalled_used_valid;
2458 new = vq->signalled_used = vq->last_used_idx;
2459 vq->signalled_used_valid = true;
2464 if (vhost_get_used_event(vq, &event)) {
2465 vq_err(vq, "Failed to get used event idx");
2468 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2471 /* This actually signals the guest, using eventfd. */
2472 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2474 /* Signal the Guest tell them we used something up. */
2475 if (vq->call_ctx.ctx && vhost_notify(dev, vq))
2476 eventfd_signal(vq->call_ctx.ctx, 1);
2478 EXPORT_SYMBOL_GPL(vhost_signal);
2480 /* And here's the combo meal deal. Supersize me! */
2481 void vhost_add_used_and_signal(struct vhost_dev *dev,
2482 struct vhost_virtqueue *vq,
2483 unsigned int head, int len)
2485 vhost_add_used(vq, head, len);
2486 vhost_signal(dev, vq);
2488 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2490 /* multi-buffer version of vhost_add_used_and_signal */
2491 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2492 struct vhost_virtqueue *vq,
2493 struct vring_used_elem *heads, unsigned count)
2495 vhost_add_used_n(vq, heads, count);
2496 vhost_signal(dev, vq);
2498 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2500 /* return true if we're sure that avaiable ring is empty */
2501 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2503 __virtio16 avail_idx;
2506 if (vq->avail_idx != vq->last_avail_idx)
2509 r = vhost_get_avail_idx(vq, &avail_idx);
2512 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2514 return vq->avail_idx == vq->last_avail_idx;
2516 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2518 /* OK, now we need to know about added descriptors. */
2519 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2521 __virtio16 avail_idx;
2524 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2526 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2527 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2528 r = vhost_update_used_flags(vq);
2530 vq_err(vq, "Failed to enable notification at %p: %d\n",
2531 &vq->used->flags, r);
2535 r = vhost_update_avail_event(vq);
2537 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2538 vhost_avail_event(vq), r);
2542 /* They could have slipped one in as we were doing that: make
2543 * sure it's written, then check again. */
2545 r = vhost_get_avail_idx(vq, &avail_idx);
2547 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2548 &vq->avail->idx, r);
2552 return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2554 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2556 /* We don't need to be notified again. */
2557 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2561 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2563 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2564 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2565 r = vhost_update_used_flags(vq);
2567 vq_err(vq, "Failed to disable notification at %p: %d\n",
2568 &vq->used->flags, r);
2571 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2573 /* Create a new message. */
2574 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2576 struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2580 /* Make sure all padding within the structure is initialized. */
2581 memset(&node->msg, 0, sizeof node->msg);
2583 node->msg.type = type;
2586 EXPORT_SYMBOL_GPL(vhost_new_msg);
2588 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2589 struct vhost_msg_node *node)
2591 spin_lock(&dev->iotlb_lock);
2592 list_add_tail(&node->node, head);
2593 spin_unlock(&dev->iotlb_lock);
2595 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2597 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2599 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2600 struct list_head *head)
2602 struct vhost_msg_node *node = NULL;
2604 spin_lock(&dev->iotlb_lock);
2605 if (!list_empty(head)) {
2606 node = list_first_entry(head, struct vhost_msg_node,
2608 list_del(&node->node);
2610 spin_unlock(&dev->iotlb_lock);
2614 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2616 void vhost_set_backend_features(struct vhost_dev *dev, u64 features)
2618 struct vhost_virtqueue *vq;
2621 mutex_lock(&dev->mutex);
2622 for (i = 0; i < dev->nvqs; ++i) {
2624 mutex_lock(&vq->mutex);
2625 vq->acked_backend_features = features;
2626 mutex_unlock(&vq->mutex);
2628 mutex_unlock(&dev->mutex);
2630 EXPORT_SYMBOL_GPL(vhost_set_backend_features);
2632 static int __init vhost_init(void)
2637 static void __exit vhost_exit(void)
2641 module_init(vhost_init);
2642 module_exit(vhost_exit);
2644 MODULE_VERSION("0.0.1");
2645 MODULE_LICENSE("GPL v2");
2646 MODULE_AUTHOR("Michael S. Tsirkin");
2647 MODULE_DESCRIPTION("Host kernel accelerator for virtio");