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/module.h>
26 #include <linux/sort.h>
27 #include <linux/sched/mm.h>
28 #include <linux/sched/signal.h>
29 #include <linux/sched/vhost_task.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_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_dev_flush);
248 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
253 if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
254 /* We can only add the work to the list after we're
255 * sure it was not in the list.
256 * test_and_set_bit() implies a memory barrier.
258 llist_add(&work->node, &dev->worker->work_list);
259 vhost_task_wake(dev->worker->vtsk);
262 EXPORT_SYMBOL_GPL(vhost_work_queue);
264 /* A lockless hint for busy polling code to exit the loop */
265 bool vhost_has_work(struct vhost_dev *dev)
267 return dev->worker && !llist_empty(&dev->worker->work_list);
269 EXPORT_SYMBOL_GPL(vhost_has_work);
271 void vhost_poll_queue(struct vhost_poll *poll)
273 vhost_work_queue(poll->dev, &poll->work);
275 EXPORT_SYMBOL_GPL(vhost_poll_queue);
277 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
281 for (j = 0; j < VHOST_NUM_ADDRS; j++)
282 vq->meta_iotlb[j] = NULL;
285 static void vhost_vq_meta_reset(struct vhost_dev *d)
289 for (i = 0; i < d->nvqs; ++i)
290 __vhost_vq_meta_reset(d->vqs[i]);
293 static void vhost_vring_call_reset(struct vhost_vring_call *call_ctx)
295 call_ctx->ctx = NULL;
296 memset(&call_ctx->producer, 0x0, sizeof(struct irq_bypass_producer));
299 bool vhost_vq_is_setup(struct vhost_virtqueue *vq)
301 return vq->avail && vq->desc && vq->used && vhost_vq_access_ok(vq);
303 EXPORT_SYMBOL_GPL(vhost_vq_is_setup);
305 static void vhost_vq_reset(struct vhost_dev *dev,
306 struct vhost_virtqueue *vq)
312 vq->last_avail_idx = 0;
314 vq->last_used_idx = 0;
315 vq->signalled_used = 0;
316 vq->signalled_used_valid = false;
318 vq->log_used = false;
319 vq->log_addr = -1ull;
320 vq->private_data = NULL;
321 vq->acked_features = 0;
322 vq->acked_backend_features = 0;
324 vq->error_ctx = NULL;
327 vhost_disable_cross_endian(vq);
328 vhost_reset_is_le(vq);
329 vq->busyloop_timeout = 0;
332 vhost_vring_call_reset(&vq->call_ctx);
333 __vhost_vq_meta_reset(vq);
336 static bool vhost_worker(void *data)
338 struct vhost_worker *worker = data;
339 struct vhost_work *work, *work_next;
340 struct llist_node *node;
342 node = llist_del_all(&worker->work_list);
344 node = llist_reverse_order(node);
345 /* make sure flag is seen after deletion */
347 llist_for_each_entry_safe(work, work_next, node, node) {
348 clear_bit(VHOST_WORK_QUEUED, &work->flags);
349 kcov_remote_start_common(worker->kcov_handle);
359 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
369 /* Helper to allocate iovec buffers for all vqs. */
370 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
372 struct vhost_virtqueue *vq;
375 for (i = 0; i < dev->nvqs; ++i) {
377 vq->indirect = kmalloc_array(UIO_MAXIOV,
378 sizeof(*vq->indirect),
380 vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
382 vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
384 if (!vq->indirect || !vq->log || !vq->heads)
391 vhost_vq_free_iovecs(dev->vqs[i]);
395 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
399 for (i = 0; i < dev->nvqs; ++i)
400 vhost_vq_free_iovecs(dev->vqs[i]);
403 bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
404 int pkts, int total_len)
406 struct vhost_dev *dev = vq->dev;
408 if ((dev->byte_weight && total_len >= dev->byte_weight) ||
409 pkts >= dev->weight) {
410 vhost_poll_queue(&vq->poll);
416 EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
418 static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
421 size_t event __maybe_unused =
422 vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
424 return size_add(struct_size(vq->avail, ring, num), event);
427 static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
430 size_t event __maybe_unused =
431 vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
433 return size_add(struct_size(vq->used, ring, num), event);
436 static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
439 return sizeof(*vq->desc) * num;
442 void vhost_dev_init(struct vhost_dev *dev,
443 struct vhost_virtqueue **vqs, int nvqs,
444 int iov_limit, int weight, int byte_weight,
446 int (*msg_handler)(struct vhost_dev *dev, u32 asid,
447 struct vhost_iotlb_msg *msg))
449 struct vhost_virtqueue *vq;
454 mutex_init(&dev->mutex);
460 dev->iov_limit = iov_limit;
461 dev->weight = weight;
462 dev->byte_weight = byte_weight;
463 dev->use_worker = use_worker;
464 dev->msg_handler = msg_handler;
465 init_waitqueue_head(&dev->wait);
466 INIT_LIST_HEAD(&dev->read_list);
467 INIT_LIST_HEAD(&dev->pending_list);
468 spin_lock_init(&dev->iotlb_lock);
471 for (i = 0; i < dev->nvqs; ++i) {
477 mutex_init(&vq->mutex);
478 vhost_vq_reset(dev, vq);
480 vhost_poll_init(&vq->poll, vq->handle_kick,
484 EXPORT_SYMBOL_GPL(vhost_dev_init);
486 /* Caller should have device mutex */
487 long vhost_dev_check_owner(struct vhost_dev *dev)
489 /* Are you the owner? If not, I don't think you mean to do that */
490 return dev->mm == current->mm ? 0 : -EPERM;
492 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
494 /* Caller should have device mutex */
495 bool vhost_dev_has_owner(struct vhost_dev *dev)
499 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
501 static void vhost_attach_mm(struct vhost_dev *dev)
503 /* No owner, become one */
504 if (dev->use_worker) {
505 dev->mm = get_task_mm(current);
507 /* vDPA device does not use worker thead, so there's
508 * no need to hold the address space for mm. This help
509 * to avoid deadlock in the case of mmap() which may
510 * held the refcnt of the file and depends on release
511 * method to remove vma.
513 dev->mm = current->mm;
518 static void vhost_detach_mm(struct vhost_dev *dev)
531 static void vhost_worker_free(struct vhost_dev *dev)
533 struct vhost_worker *worker = dev->worker;
539 WARN_ON(!llist_empty(&worker->work_list));
540 vhost_task_stop(worker->vtsk);
544 static int vhost_worker_create(struct vhost_dev *dev)
546 struct vhost_worker *worker;
547 struct vhost_task *vtsk;
548 char name[TASK_COMM_LEN];
551 worker = kzalloc(sizeof(*worker), GFP_KERNEL_ACCOUNT);
555 dev->worker = worker;
556 worker->kcov_handle = kcov_common_handle();
557 init_llist_head(&worker->work_list);
558 snprintf(name, sizeof(name), "vhost-%d", current->pid);
560 vtsk = vhost_task_create(vhost_worker, worker, name);
567 vhost_task_start(vtsk);
576 /* Caller should have device mutex */
577 long vhost_dev_set_owner(struct vhost_dev *dev)
581 /* Is there an owner already? */
582 if (vhost_dev_has_owner(dev)) {
587 vhost_attach_mm(dev);
589 if (dev->use_worker) {
590 err = vhost_worker_create(dev);
595 err = vhost_dev_alloc_iovecs(dev);
601 vhost_worker_free(dev);
603 vhost_detach_mm(dev);
607 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
609 static struct vhost_iotlb *iotlb_alloc(void)
611 return vhost_iotlb_alloc(max_iotlb_entries,
612 VHOST_IOTLB_FLAG_RETIRE);
615 struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
617 return iotlb_alloc();
619 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
621 /* Caller should have device mutex */
622 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
626 vhost_dev_cleanup(dev);
629 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
630 * VQs aren't running.
632 for (i = 0; i < dev->nvqs; ++i)
633 dev->vqs[i]->umem = umem;
635 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
637 void vhost_dev_stop(struct vhost_dev *dev)
641 for (i = 0; i < dev->nvqs; ++i) {
642 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick)
643 vhost_poll_stop(&dev->vqs[i]->poll);
646 vhost_dev_flush(dev);
648 EXPORT_SYMBOL_GPL(vhost_dev_stop);
650 void vhost_clear_msg(struct vhost_dev *dev)
652 struct vhost_msg_node *node, *n;
654 spin_lock(&dev->iotlb_lock);
656 list_for_each_entry_safe(node, n, &dev->read_list, node) {
657 list_del(&node->node);
661 list_for_each_entry_safe(node, n, &dev->pending_list, node) {
662 list_del(&node->node);
666 spin_unlock(&dev->iotlb_lock);
668 EXPORT_SYMBOL_GPL(vhost_clear_msg);
670 void vhost_dev_cleanup(struct vhost_dev *dev)
674 for (i = 0; i < dev->nvqs; ++i) {
675 if (dev->vqs[i]->error_ctx)
676 eventfd_ctx_put(dev->vqs[i]->error_ctx);
677 if (dev->vqs[i]->kick)
678 fput(dev->vqs[i]->kick);
679 if (dev->vqs[i]->call_ctx.ctx)
680 eventfd_ctx_put(dev->vqs[i]->call_ctx.ctx);
681 vhost_vq_reset(dev, dev->vqs[i]);
683 vhost_dev_free_iovecs(dev);
685 eventfd_ctx_put(dev->log_ctx);
687 /* No one will access memory at this point */
688 vhost_iotlb_free(dev->umem);
690 vhost_iotlb_free(dev->iotlb);
692 vhost_clear_msg(dev);
693 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
694 vhost_worker_free(dev);
695 vhost_detach_mm(dev);
697 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
699 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
701 u64 a = addr / VHOST_PAGE_SIZE / 8;
703 /* Make sure 64 bit math will not overflow. */
704 if (a > ULONG_MAX - (unsigned long)log_base ||
705 a + (unsigned long)log_base > ULONG_MAX)
708 return access_ok(log_base + a,
709 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
712 /* Make sure 64 bit math will not overflow. */
713 static bool vhost_overflow(u64 uaddr, u64 size)
715 if (uaddr > ULONG_MAX || size > ULONG_MAX)
721 return uaddr > ULONG_MAX - size + 1;
724 /* Caller should have vq mutex and device mutex. */
725 static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
728 struct vhost_iotlb_map *map;
733 list_for_each_entry(map, &umem->list, link) {
734 unsigned long a = map->addr;
736 if (vhost_overflow(map->addr, map->size))
740 if (!access_ok((void __user *)a, map->size))
742 else if (log_all && !log_access_ok(log_base,
750 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
751 u64 addr, unsigned int size,
754 const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
759 return (void __user *)(uintptr_t)(map->addr + addr - map->start);
762 /* Can we switch to this memory table? */
763 /* Caller should have device mutex but not vq mutex */
764 static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
769 for (i = 0; i < d->nvqs; ++i) {
773 mutex_lock(&d->vqs[i]->mutex);
774 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
775 /* If ring is inactive, will check when it's enabled. */
776 if (d->vqs[i]->private_data)
777 ok = vq_memory_access_ok(d->vqs[i]->log_base,
781 mutex_unlock(&d->vqs[i]->mutex);
788 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
789 struct iovec iov[], int iov_size, int access);
791 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
792 const void *from, unsigned size)
797 return __copy_to_user(to, from, size);
799 /* This function should be called after iotlb
800 * prefetch, which means we're sure that all vq
801 * could be access through iotlb. So -EAGAIN should
802 * not happen in this case.
805 void __user *uaddr = vhost_vq_meta_fetch(vq,
806 (u64)(uintptr_t)to, size,
810 return __copy_to_user(uaddr, from, size);
812 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
813 ARRAY_SIZE(vq->iotlb_iov),
817 iov_iter_init(&t, ITER_DEST, vq->iotlb_iov, ret, size);
818 ret = copy_to_iter(from, size, &t);
826 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
827 void __user *from, unsigned size)
832 return __copy_from_user(to, from, size);
834 /* This function should be called after iotlb
835 * prefetch, which means we're sure that vq
836 * could be access through iotlb. So -EAGAIN should
837 * not happen in this case.
839 void __user *uaddr = vhost_vq_meta_fetch(vq,
840 (u64)(uintptr_t)from, size,
845 return __copy_from_user(to, uaddr, size);
847 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
848 ARRAY_SIZE(vq->iotlb_iov),
851 vq_err(vq, "IOTLB translation failure: uaddr "
852 "%p size 0x%llx\n", from,
853 (unsigned long long) size);
856 iov_iter_init(&f, ITER_SOURCE, vq->iotlb_iov, ret, size);
857 ret = copy_from_iter(to, size, &f);
866 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
867 void __user *addr, unsigned int size,
872 ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
873 ARRAY_SIZE(vq->iotlb_iov),
876 vq_err(vq, "IOTLB translation failure: uaddr "
877 "%p size 0x%llx\n", addr,
878 (unsigned long long) size);
882 if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
883 vq_err(vq, "Non atomic userspace memory access: uaddr "
884 "%p size 0x%llx\n", addr,
885 (unsigned long long) size);
889 return vq->iotlb_iov[0].iov_base;
892 /* This function should be called after iotlb
893 * prefetch, which means we're sure that vq
894 * could be access through iotlb. So -EAGAIN should
895 * not happen in this case.
897 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
898 void __user *addr, unsigned int size,
901 void __user *uaddr = vhost_vq_meta_fetch(vq,
902 (u64)(uintptr_t)addr, size, type);
906 return __vhost_get_user_slow(vq, addr, size, type);
909 #define vhost_put_user(vq, x, ptr) \
913 ret = __put_user(x, ptr); \
915 __typeof__(ptr) to = \
916 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
917 sizeof(*ptr), VHOST_ADDR_USED); \
919 ret = __put_user(x, to); \
926 static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
928 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
929 vhost_avail_event(vq));
932 static inline int vhost_put_used(struct vhost_virtqueue *vq,
933 struct vring_used_elem *head, int idx,
936 return vhost_copy_to_user(vq, vq->used->ring + idx, head,
937 count * sizeof(*head));
940 static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
943 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
947 static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
950 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
954 #define vhost_get_user(vq, x, ptr, type) \
958 ret = __get_user(x, ptr); \
960 __typeof__(ptr) from = \
961 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
965 ret = __get_user(x, from); \
972 #define vhost_get_avail(vq, x, ptr) \
973 vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
975 #define vhost_get_used(vq, x, ptr) \
976 vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
978 static void vhost_dev_lock_vqs(struct vhost_dev *d)
981 for (i = 0; i < d->nvqs; ++i)
982 mutex_lock_nested(&d->vqs[i]->mutex, i);
985 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
988 for (i = 0; i < d->nvqs; ++i)
989 mutex_unlock(&d->vqs[i]->mutex);
992 static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
995 return vhost_get_avail(vq, *idx, &vq->avail->idx);
998 static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
999 __virtio16 *head, int idx)
1001 return vhost_get_avail(vq, *head,
1002 &vq->avail->ring[idx & (vq->num - 1)]);
1005 static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
1008 return vhost_get_avail(vq, *flags, &vq->avail->flags);
1011 static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1014 return vhost_get_avail(vq, *event, vhost_used_event(vq));
1017 static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1020 return vhost_get_used(vq, *idx, &vq->used->idx);
1023 static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1024 struct vring_desc *desc, int idx)
1026 return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1029 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1030 struct vhost_iotlb_msg *msg)
1032 struct vhost_msg_node *node, *n;
1034 spin_lock(&d->iotlb_lock);
1036 list_for_each_entry_safe(node, n, &d->pending_list, node) {
1037 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1038 if (msg->iova <= vq_msg->iova &&
1039 msg->iova + msg->size - 1 >= vq_msg->iova &&
1040 vq_msg->type == VHOST_IOTLB_MISS) {
1041 vhost_poll_queue(&node->vq->poll);
1042 list_del(&node->node);
1047 spin_unlock(&d->iotlb_lock);
1050 static bool umem_access_ok(u64 uaddr, u64 size, int access)
1052 unsigned long a = uaddr;
1054 /* Make sure 64 bit math will not overflow. */
1055 if (vhost_overflow(uaddr, size))
1058 if ((access & VHOST_ACCESS_RO) &&
1059 !access_ok((void __user *)a, size))
1061 if ((access & VHOST_ACCESS_WO) &&
1062 !access_ok((void __user *)a, size))
1067 static int vhost_process_iotlb_msg(struct vhost_dev *dev, u32 asid,
1068 struct vhost_iotlb_msg *msg)
1075 mutex_lock(&dev->mutex);
1076 vhost_dev_lock_vqs(dev);
1077 switch (msg->type) {
1078 case VHOST_IOTLB_UPDATE:
1083 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1087 vhost_vq_meta_reset(dev);
1088 if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1089 msg->iova + msg->size - 1,
1090 msg->uaddr, msg->perm)) {
1094 vhost_iotlb_notify_vq(dev, msg);
1096 case VHOST_IOTLB_INVALIDATE:
1101 vhost_vq_meta_reset(dev);
1102 vhost_iotlb_del_range(dev->iotlb, msg->iova,
1103 msg->iova + msg->size - 1);
1110 vhost_dev_unlock_vqs(dev);
1111 mutex_unlock(&dev->mutex);
1115 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1116 struct iov_iter *from)
1118 struct vhost_iotlb_msg msg;
1123 ret = copy_from_iter(&type, sizeof(type), from);
1124 if (ret != sizeof(type)) {
1130 case VHOST_IOTLB_MSG:
1131 /* There maybe a hole after type for V1 message type,
1134 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1136 case VHOST_IOTLB_MSG_V2:
1137 if (vhost_backend_has_feature(dev->vqs[0],
1138 VHOST_BACKEND_F_IOTLB_ASID)) {
1139 ret = copy_from_iter(&asid, sizeof(asid), from);
1140 if (ret != sizeof(asid)) {
1146 offset = sizeof(__u32);
1153 iov_iter_advance(from, offset);
1154 ret = copy_from_iter(&msg, sizeof(msg), from);
1155 if (ret != sizeof(msg)) {
1160 if ((msg.type == VHOST_IOTLB_UPDATE ||
1161 msg.type == VHOST_IOTLB_INVALIDATE) &&
1167 if (dev->msg_handler)
1168 ret = dev->msg_handler(dev, asid, &msg);
1170 ret = vhost_process_iotlb_msg(dev, asid, &msg);
1176 ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1177 sizeof(struct vhost_msg_v2);
1181 EXPORT_SYMBOL(vhost_chr_write_iter);
1183 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1188 poll_wait(file, &dev->wait, wait);
1190 if (!list_empty(&dev->read_list))
1191 mask |= EPOLLIN | EPOLLRDNORM;
1195 EXPORT_SYMBOL(vhost_chr_poll);
1197 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1201 struct vhost_msg_node *node;
1203 unsigned size = sizeof(struct vhost_msg);
1205 if (iov_iter_count(to) < size)
1210 prepare_to_wait(&dev->wait, &wait,
1211 TASK_INTERRUPTIBLE);
1213 node = vhost_dequeue_msg(dev, &dev->read_list);
1220 if (signal_pending(current)) {
1233 finish_wait(&dev->wait, &wait);
1236 struct vhost_iotlb_msg *msg;
1237 void *start = &node->msg;
1239 switch (node->msg.type) {
1240 case VHOST_IOTLB_MSG:
1241 size = sizeof(node->msg);
1242 msg = &node->msg.iotlb;
1244 case VHOST_IOTLB_MSG_V2:
1245 size = sizeof(node->msg_v2);
1246 msg = &node->msg_v2.iotlb;
1253 ret = copy_to_iter(start, size, to);
1254 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1258 vhost_enqueue_msg(dev, &dev->pending_list, node);
1263 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1265 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1267 struct vhost_dev *dev = vq->dev;
1268 struct vhost_msg_node *node;
1269 struct vhost_iotlb_msg *msg;
1270 bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1272 node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1277 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1278 msg = &node->msg_v2.iotlb;
1280 msg = &node->msg.iotlb;
1283 msg->type = VHOST_IOTLB_MISS;
1287 vhost_enqueue_msg(dev, &dev->read_list, node);
1292 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1293 vring_desc_t __user *desc,
1294 vring_avail_t __user *avail,
1295 vring_used_t __user *used)
1298 /* If an IOTLB device is present, the vring addresses are
1299 * GIOVAs. Access validation occurs at prefetch time. */
1303 return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1304 access_ok(avail, vhost_get_avail_size(vq, num)) &&
1305 access_ok(used, vhost_get_used_size(vq, num));
1308 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1309 const struct vhost_iotlb_map *map,
1312 int access = (type == VHOST_ADDR_USED) ?
1313 VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1315 if (likely(map->perm & access))
1316 vq->meta_iotlb[type] = map;
1319 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1320 int access, u64 addr, u64 len, int type)
1322 const struct vhost_iotlb_map *map;
1323 struct vhost_iotlb *umem = vq->iotlb;
1324 u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1326 if (vhost_vq_meta_fetch(vq, addr, len, type))
1330 map = vhost_iotlb_itree_first(umem, addr, last);
1331 if (map == NULL || map->start > addr) {
1332 vhost_iotlb_miss(vq, addr, access);
1334 } else if (!(map->perm & access)) {
1335 /* Report the possible access violation by
1336 * request another translation from userspace.
1341 size = map->size - addr + map->start;
1343 if (orig_addr == addr && size >= len)
1344 vhost_vq_meta_update(vq, map, type);
1353 int vq_meta_prefetch(struct vhost_virtqueue *vq)
1355 unsigned int num = vq->num;
1360 return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1361 vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1362 iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1363 vhost_get_avail_size(vq, num),
1364 VHOST_ADDR_AVAIL) &&
1365 iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1366 vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1368 EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1370 /* Can we log writes? */
1371 /* Caller should have device mutex but not vq mutex */
1372 bool vhost_log_access_ok(struct vhost_dev *dev)
1374 return memory_access_ok(dev, dev->umem, 1);
1376 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1378 static bool vq_log_used_access_ok(struct vhost_virtqueue *vq,
1379 void __user *log_base,
1383 /* If an IOTLB device is present, log_addr is a GIOVA that
1384 * will never be logged by log_used(). */
1388 return !log_used || log_access_ok(log_base, log_addr,
1389 vhost_get_used_size(vq, vq->num));
1392 /* Verify access for write logging. */
1393 /* Caller should have vq mutex and device mutex */
1394 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1395 void __user *log_base)
1397 return vq_memory_access_ok(log_base, vq->umem,
1398 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1399 vq_log_used_access_ok(vq, log_base, vq->log_used, vq->log_addr);
1402 /* Can we start vq? */
1403 /* Caller should have vq mutex and device mutex */
1404 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1406 if (!vq_log_access_ok(vq, vq->log_base))
1409 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1411 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1413 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1415 struct vhost_memory mem, *newmem;
1416 struct vhost_memory_region *region;
1417 struct vhost_iotlb *newumem, *oldumem;
1418 unsigned long size = offsetof(struct vhost_memory, regions);
1421 if (copy_from_user(&mem, m, size))
1425 if (mem.nregions > max_mem_regions)
1427 newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1432 memcpy(newmem, &mem, size);
1433 if (copy_from_user(newmem->regions, m->regions,
1434 flex_array_size(newmem, regions, mem.nregions))) {
1439 newumem = iotlb_alloc();
1445 for (region = newmem->regions;
1446 region < newmem->regions + mem.nregions;
1448 if (vhost_iotlb_add_range(newumem,
1449 region->guest_phys_addr,
1450 region->guest_phys_addr +
1451 region->memory_size - 1,
1452 region->userspace_addr,
1457 if (!memory_access_ok(d, newumem, 0))
1463 /* All memory accesses are done under some VQ mutex. */
1464 for (i = 0; i < d->nvqs; ++i) {
1465 mutex_lock(&d->vqs[i]->mutex);
1466 d->vqs[i]->umem = newumem;
1467 mutex_unlock(&d->vqs[i]->mutex);
1471 vhost_iotlb_free(oldumem);
1475 vhost_iotlb_free(newumem);
1480 static long vhost_vring_set_num(struct vhost_dev *d,
1481 struct vhost_virtqueue *vq,
1484 struct vhost_vring_state s;
1486 /* Resizing ring with an active backend?
1487 * You don't want to do that. */
1488 if (vq->private_data)
1491 if (copy_from_user(&s, argp, sizeof s))
1494 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1501 static long vhost_vring_set_addr(struct vhost_dev *d,
1502 struct vhost_virtqueue *vq,
1505 struct vhost_vring_addr a;
1507 if (copy_from_user(&a, argp, sizeof a))
1509 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1512 /* For 32bit, verify that the top 32bits of the user
1513 data are set to zero. */
1514 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1515 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1516 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1519 /* Make sure it's safe to cast pointers to vring types. */
1520 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1521 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1522 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1523 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1524 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1527 /* We only verify access here if backend is configured.
1528 * If it is not, we don't as size might not have been setup.
1529 * We will verify when backend is configured. */
1530 if (vq->private_data) {
1531 if (!vq_access_ok(vq, vq->num,
1532 (void __user *)(unsigned long)a.desc_user_addr,
1533 (void __user *)(unsigned long)a.avail_user_addr,
1534 (void __user *)(unsigned long)a.used_user_addr))
1537 /* Also validate log access for used ring if enabled. */
1538 if (!vq_log_used_access_ok(vq, vq->log_base,
1539 a.flags & (0x1 << VHOST_VRING_F_LOG),
1544 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1545 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1546 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1547 vq->log_addr = a.log_guest_addr;
1548 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1553 static long vhost_vring_set_num_addr(struct vhost_dev *d,
1554 struct vhost_virtqueue *vq,
1560 mutex_lock(&vq->mutex);
1563 case VHOST_SET_VRING_NUM:
1564 r = vhost_vring_set_num(d, vq, argp);
1566 case VHOST_SET_VRING_ADDR:
1567 r = vhost_vring_set_addr(d, vq, argp);
1573 mutex_unlock(&vq->mutex);
1577 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1579 struct file *eventfp, *filep = NULL;
1580 bool pollstart = false, pollstop = false;
1581 struct eventfd_ctx *ctx = NULL;
1582 u32 __user *idxp = argp;
1583 struct vhost_virtqueue *vq;
1584 struct vhost_vring_state s;
1585 struct vhost_vring_file f;
1589 r = get_user(idx, idxp);
1595 idx = array_index_nospec(idx, d->nvqs);
1598 if (ioctl == VHOST_SET_VRING_NUM ||
1599 ioctl == VHOST_SET_VRING_ADDR) {
1600 return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1603 mutex_lock(&vq->mutex);
1606 case VHOST_SET_VRING_BASE:
1607 /* Moving base with an active backend?
1608 * You don't want to do that. */
1609 if (vq->private_data) {
1613 if (copy_from_user(&s, argp, sizeof s)) {
1617 if (s.num > 0xffff) {
1621 vq->last_avail_idx = s.num;
1622 /* Forget the cached index value. */
1623 vq->avail_idx = vq->last_avail_idx;
1625 case VHOST_GET_VRING_BASE:
1627 s.num = vq->last_avail_idx;
1628 if (copy_to_user(argp, &s, sizeof s))
1631 case VHOST_SET_VRING_KICK:
1632 if (copy_from_user(&f, argp, sizeof f)) {
1636 eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
1637 if (IS_ERR(eventfp)) {
1638 r = PTR_ERR(eventfp);
1641 if (eventfp != vq->kick) {
1642 pollstop = (filep = vq->kick) != NULL;
1643 pollstart = (vq->kick = eventfp) != NULL;
1647 case VHOST_SET_VRING_CALL:
1648 if (copy_from_user(&f, argp, sizeof f)) {
1652 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1658 swap(ctx, vq->call_ctx.ctx);
1660 case VHOST_SET_VRING_ERR:
1661 if (copy_from_user(&f, argp, sizeof f)) {
1665 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1670 swap(ctx, vq->error_ctx);
1672 case VHOST_SET_VRING_ENDIAN:
1673 r = vhost_set_vring_endian(vq, argp);
1675 case VHOST_GET_VRING_ENDIAN:
1676 r = vhost_get_vring_endian(vq, idx, argp);
1678 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1679 if (copy_from_user(&s, argp, sizeof(s))) {
1683 vq->busyloop_timeout = s.num;
1685 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1687 s.num = vq->busyloop_timeout;
1688 if (copy_to_user(argp, &s, sizeof(s)))
1695 if (pollstop && vq->handle_kick)
1696 vhost_poll_stop(&vq->poll);
1698 if (!IS_ERR_OR_NULL(ctx))
1699 eventfd_ctx_put(ctx);
1703 if (pollstart && vq->handle_kick)
1704 r = vhost_poll_start(&vq->poll, vq->kick);
1706 mutex_unlock(&vq->mutex);
1708 if (pollstop && vq->handle_kick)
1709 vhost_dev_flush(vq->poll.dev);
1712 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1714 int vhost_init_device_iotlb(struct vhost_dev *d)
1716 struct vhost_iotlb *niotlb, *oiotlb;
1719 niotlb = iotlb_alloc();
1726 for (i = 0; i < d->nvqs; ++i) {
1727 struct vhost_virtqueue *vq = d->vqs[i];
1729 mutex_lock(&vq->mutex);
1731 __vhost_vq_meta_reset(vq);
1732 mutex_unlock(&vq->mutex);
1735 vhost_iotlb_free(oiotlb);
1739 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1741 /* Caller must have device mutex */
1742 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1744 struct eventfd_ctx *ctx;
1749 /* If you are not the owner, you can become one */
1750 if (ioctl == VHOST_SET_OWNER) {
1751 r = vhost_dev_set_owner(d);
1755 /* You must be the owner to do anything else */
1756 r = vhost_dev_check_owner(d);
1761 case VHOST_SET_MEM_TABLE:
1762 r = vhost_set_memory(d, argp);
1764 case VHOST_SET_LOG_BASE:
1765 if (copy_from_user(&p, argp, sizeof p)) {
1769 if ((u64)(unsigned long)p != p) {
1773 for (i = 0; i < d->nvqs; ++i) {
1774 struct vhost_virtqueue *vq;
1775 void __user *base = (void __user *)(unsigned long)p;
1777 mutex_lock(&vq->mutex);
1778 /* If ring is inactive, will check when it's enabled. */
1779 if (vq->private_data && !vq_log_access_ok(vq, base))
1782 vq->log_base = base;
1783 mutex_unlock(&vq->mutex);
1786 case VHOST_SET_LOG_FD:
1787 r = get_user(fd, (int __user *)argp);
1790 ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
1795 swap(ctx, d->log_ctx);
1796 for (i = 0; i < d->nvqs; ++i) {
1797 mutex_lock(&d->vqs[i]->mutex);
1798 d->vqs[i]->log_ctx = d->log_ctx;
1799 mutex_unlock(&d->vqs[i]->mutex);
1802 eventfd_ctx_put(ctx);
1811 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1813 /* TODO: This is really inefficient. We need something like get_user()
1814 * (instruction directly accesses the data, with an exception table entry
1815 * returning -EFAULT). See Documentation/arch/x86/exception-tables.rst.
1817 static int set_bit_to_user(int nr, void __user *addr)
1819 unsigned long log = (unsigned long)addr;
1822 int bit = nr + (log % PAGE_SIZE) * 8;
1825 r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
1829 base = kmap_atomic(page);
1831 kunmap_atomic(base);
1832 unpin_user_pages_dirty_lock(&page, 1, true);
1836 static int log_write(void __user *log_base,
1837 u64 write_address, u64 write_length)
1839 u64 write_page = write_address / VHOST_PAGE_SIZE;
1844 write_length += write_address % VHOST_PAGE_SIZE;
1846 u64 base = (u64)(unsigned long)log_base;
1847 u64 log = base + write_page / 8;
1848 int bit = write_page % 8;
1849 if ((u64)(unsigned long)log != log)
1851 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1854 if (write_length <= VHOST_PAGE_SIZE)
1856 write_length -= VHOST_PAGE_SIZE;
1862 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1864 struct vhost_iotlb *umem = vq->umem;
1865 struct vhost_iotlb_map *u;
1866 u64 start, end, l, min;
1872 /* More than one GPAs can be mapped into a single HVA. So
1873 * iterate all possible umems here to be safe.
1875 list_for_each_entry(u, &umem->list, link) {
1876 if (u->addr > hva - 1 + len ||
1877 u->addr - 1 + u->size < hva)
1879 start = max(u->addr, hva);
1880 end = min(u->addr - 1 + u->size, hva - 1 + len);
1881 l = end - start + 1;
1882 r = log_write(vq->log_base,
1883 u->start + start - u->addr,
1901 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1903 struct iovec *iov = vq->log_iov;
1907 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1909 ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1910 len, iov, 64, VHOST_ACCESS_WO);
1914 for (i = 0; i < ret; i++) {
1915 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1924 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1925 unsigned int log_num, u64 len, struct iovec *iov, int count)
1929 /* Make sure data written is seen before log. */
1933 for (i = 0; i < count; i++) {
1934 r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1942 for (i = 0; i < log_num; ++i) {
1943 u64 l = min(log[i].len, len);
1944 r = log_write(vq->log_base, log[i].addr, l);
1950 eventfd_signal(vq->log_ctx, 1);
1954 /* Length written exceeds what we have stored. This is a bug. */
1958 EXPORT_SYMBOL_GPL(vhost_log_write);
1960 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1963 if (vhost_put_used_flags(vq))
1965 if (unlikely(vq->log_used)) {
1966 /* Make sure the flag is seen before log. */
1968 /* Log used flag write. */
1969 used = &vq->used->flags;
1970 log_used(vq, (used - (void __user *)vq->used),
1971 sizeof vq->used->flags);
1973 eventfd_signal(vq->log_ctx, 1);
1978 static int vhost_update_avail_event(struct vhost_virtqueue *vq)
1980 if (vhost_put_avail_event(vq))
1982 if (unlikely(vq->log_used)) {
1984 /* Make sure the event is seen before log. */
1986 /* Log avail event write */
1987 used = vhost_avail_event(vq);
1988 log_used(vq, (used - (void __user *)vq->used),
1989 sizeof *vhost_avail_event(vq));
1991 eventfd_signal(vq->log_ctx, 1);
1996 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1998 __virtio16 last_used_idx;
2000 bool is_le = vq->is_le;
2002 if (!vq->private_data)
2005 vhost_init_is_le(vq);
2007 r = vhost_update_used_flags(vq);
2010 vq->signalled_used_valid = false;
2012 !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2016 r = vhost_get_used_idx(vq, &last_used_idx);
2018 vq_err(vq, "Can't access used idx at %p\n",
2022 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2029 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2031 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2032 struct iovec iov[], int iov_size, int access)
2034 const struct vhost_iotlb_map *map;
2035 struct vhost_dev *dev = vq->dev;
2036 struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2038 u64 s = 0, last = addr + len - 1;
2041 while ((u64)len > s) {
2043 if (unlikely(ret >= iov_size)) {
2048 map = vhost_iotlb_itree_first(umem, addr, last);
2049 if (map == NULL || map->start > addr) {
2050 if (umem != dev->iotlb) {
2056 } else if (!(map->perm & access)) {
2062 size = map->size - addr + map->start;
2063 _iov->iov_len = min((u64)len - s, size);
2064 _iov->iov_base = (void __user *)(unsigned long)
2065 (map->addr + addr - map->start);
2072 vhost_iotlb_miss(vq, addr, access);
2076 /* Each buffer in the virtqueues is actually a chain of descriptors. This
2077 * function returns the next descriptor in the chain,
2078 * or -1U if we're at the end. */
2079 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2083 /* If this descriptor says it doesn't chain, we're done. */
2084 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2087 /* Check they're not leading us off end of descriptors. */
2088 next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2092 static int get_indirect(struct vhost_virtqueue *vq,
2093 struct iovec iov[], unsigned int iov_size,
2094 unsigned int *out_num, unsigned int *in_num,
2095 struct vhost_log *log, unsigned int *log_num,
2096 struct vring_desc *indirect)
2098 struct vring_desc desc;
2099 unsigned int i = 0, count, found = 0;
2100 u32 len = vhost32_to_cpu(vq, indirect->len);
2101 struct iov_iter from;
2105 if (unlikely(len % sizeof desc)) {
2106 vq_err(vq, "Invalid length in indirect descriptor: "
2107 "len 0x%llx not multiple of 0x%zx\n",
2108 (unsigned long long)len,
2113 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2114 UIO_MAXIOV, VHOST_ACCESS_RO);
2115 if (unlikely(ret < 0)) {
2117 vq_err(vq, "Translation failure %d in indirect.\n", ret);
2120 iov_iter_init(&from, ITER_SOURCE, vq->indirect, ret, len);
2121 count = len / sizeof desc;
2122 /* Buffers are chained via a 16 bit next field, so
2123 * we can have at most 2^16 of these. */
2124 if (unlikely(count > USHRT_MAX + 1)) {
2125 vq_err(vq, "Indirect buffer length too big: %d\n",
2131 unsigned iov_count = *in_num + *out_num;
2132 if (unlikely(++found > count)) {
2133 vq_err(vq, "Loop detected: last one at %u "
2134 "indirect size %u\n",
2138 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2139 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2140 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2143 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2144 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2145 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2149 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2150 access = VHOST_ACCESS_WO;
2152 access = VHOST_ACCESS_RO;
2154 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2155 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2156 iov_size - iov_count, access);
2157 if (unlikely(ret < 0)) {
2159 vq_err(vq, "Translation failure %d indirect idx %d\n",
2163 /* If this is an input descriptor, increment that count. */
2164 if (access == VHOST_ACCESS_WO) {
2166 if (unlikely(log && ret)) {
2167 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2168 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2172 /* If it's an output descriptor, they're all supposed
2173 * to come before any input descriptors. */
2174 if (unlikely(*in_num)) {
2175 vq_err(vq, "Indirect descriptor "
2176 "has out after in: idx %d\n", i);
2181 } while ((i = next_desc(vq, &desc)) != -1);
2185 /* This looks in the virtqueue and for the first available buffer, and converts
2186 * it to an iovec for convenient access. Since descriptors consist of some
2187 * number of output then some number of input descriptors, it's actually two
2188 * iovecs, but we pack them into one and note how many of each there were.
2190 * This function returns the descriptor number found, or vq->num (which is
2191 * never a valid descriptor number) if none was found. A negative code is
2192 * returned on error. */
2193 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2194 struct iovec iov[], unsigned int iov_size,
2195 unsigned int *out_num, unsigned int *in_num,
2196 struct vhost_log *log, unsigned int *log_num)
2198 struct vring_desc desc;
2199 unsigned int i, head, found = 0;
2201 __virtio16 avail_idx;
2202 __virtio16 ring_head;
2205 /* Check it isn't doing very strange things with descriptor numbers. */
2206 last_avail_idx = vq->last_avail_idx;
2208 if (vq->avail_idx == vq->last_avail_idx) {
2209 if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2210 vq_err(vq, "Failed to access avail idx at %p\n",
2214 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2216 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2217 vq_err(vq, "Guest moved used index from %u to %u",
2218 last_avail_idx, vq->avail_idx);
2222 /* If there's nothing new since last we looked, return
2225 if (vq->avail_idx == last_avail_idx)
2228 /* Only get avail ring entries after they have been
2234 /* Grab the next descriptor number they're advertising, and increment
2235 * the index we've seen. */
2236 if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2237 vq_err(vq, "Failed to read head: idx %d address %p\n",
2239 &vq->avail->ring[last_avail_idx % vq->num]);
2243 head = vhost16_to_cpu(vq, ring_head);
2245 /* If their number is silly, that's an error. */
2246 if (unlikely(head >= vq->num)) {
2247 vq_err(vq, "Guest says index %u > %u is available",
2252 /* When we start there are none of either input nor output. */
2253 *out_num = *in_num = 0;
2259 unsigned iov_count = *in_num + *out_num;
2260 if (unlikely(i >= vq->num)) {
2261 vq_err(vq, "Desc index is %u > %u, head = %u",
2265 if (unlikely(++found > vq->num)) {
2266 vq_err(vq, "Loop detected: last one at %u "
2267 "vq size %u head %u\n",
2271 ret = vhost_get_desc(vq, &desc, i);
2272 if (unlikely(ret)) {
2273 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2277 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2278 ret = get_indirect(vq, iov, iov_size,
2280 log, log_num, &desc);
2281 if (unlikely(ret < 0)) {
2283 vq_err(vq, "Failure detected "
2284 "in indirect descriptor at idx %d\n", i);
2290 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2291 access = VHOST_ACCESS_WO;
2293 access = VHOST_ACCESS_RO;
2294 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2295 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2296 iov_size - iov_count, access);
2297 if (unlikely(ret < 0)) {
2299 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2303 if (access == VHOST_ACCESS_WO) {
2304 /* If this is an input descriptor,
2305 * increment that count. */
2307 if (unlikely(log && ret)) {
2308 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2309 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2313 /* If it's an output descriptor, they're all supposed
2314 * to come before any input descriptors. */
2315 if (unlikely(*in_num)) {
2316 vq_err(vq, "Descriptor has out after in: "
2322 } while ((i = next_desc(vq, &desc)) != -1);
2324 /* On success, increment avail index. */
2325 vq->last_avail_idx++;
2327 /* Assume notifications from guest are disabled at this point,
2328 * if they aren't we would need to update avail_event index. */
2329 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2332 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2334 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2335 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2337 vq->last_avail_idx -= n;
2339 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2341 /* After we've used one of their buffers, we tell them about it. We'll then
2342 * want to notify the guest, using eventfd. */
2343 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2345 struct vring_used_elem heads = {
2346 cpu_to_vhost32(vq, head),
2347 cpu_to_vhost32(vq, len)
2350 return vhost_add_used_n(vq, &heads, 1);
2352 EXPORT_SYMBOL_GPL(vhost_add_used);
2354 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2355 struct vring_used_elem *heads,
2358 vring_used_elem_t __user *used;
2362 start = vq->last_used_idx & (vq->num - 1);
2363 used = vq->used->ring + start;
2364 if (vhost_put_used(vq, heads, start, count)) {
2365 vq_err(vq, "Failed to write used");
2368 if (unlikely(vq->log_used)) {
2369 /* Make sure data is seen before log. */
2371 /* Log used ring entry write. */
2372 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2373 count * sizeof *used);
2375 old = vq->last_used_idx;
2376 new = (vq->last_used_idx += count);
2377 /* If the driver never bothers to signal in a very long while,
2378 * used index might wrap around. If that happens, invalidate
2379 * signalled_used index we stored. TODO: make sure driver
2380 * signals at least once in 2^16 and remove this. */
2381 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2382 vq->signalled_used_valid = false;
2386 /* After we've used one of their buffers, we tell them about it. We'll then
2387 * want to notify the guest, using eventfd. */
2388 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2393 start = vq->last_used_idx & (vq->num - 1);
2394 n = vq->num - start;
2396 r = __vhost_add_used_n(vq, heads, n);
2402 r = __vhost_add_used_n(vq, heads, count);
2404 /* Make sure buffer is written before we update index. */
2406 if (vhost_put_used_idx(vq)) {
2407 vq_err(vq, "Failed to increment used idx");
2410 if (unlikely(vq->log_used)) {
2411 /* Make sure used idx is seen before log. */
2413 /* Log used index update. */
2414 log_used(vq, offsetof(struct vring_used, idx),
2415 sizeof vq->used->idx);
2417 eventfd_signal(vq->log_ctx, 1);
2421 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2423 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2428 /* Flush out used index updates. This is paired
2429 * with the barrier that the Guest executes when enabling
2433 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2434 unlikely(vq->avail_idx == vq->last_avail_idx))
2437 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2439 if (vhost_get_avail_flags(vq, &flags)) {
2440 vq_err(vq, "Failed to get flags");
2443 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2445 old = vq->signalled_used;
2446 v = vq->signalled_used_valid;
2447 new = vq->signalled_used = vq->last_used_idx;
2448 vq->signalled_used_valid = true;
2453 if (vhost_get_used_event(vq, &event)) {
2454 vq_err(vq, "Failed to get used event idx");
2457 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2460 /* This actually signals the guest, using eventfd. */
2461 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2463 /* Signal the Guest tell them we used something up. */
2464 if (vq->call_ctx.ctx && vhost_notify(dev, vq))
2465 eventfd_signal(vq->call_ctx.ctx, 1);
2467 EXPORT_SYMBOL_GPL(vhost_signal);
2469 /* And here's the combo meal deal. Supersize me! */
2470 void vhost_add_used_and_signal(struct vhost_dev *dev,
2471 struct vhost_virtqueue *vq,
2472 unsigned int head, int len)
2474 vhost_add_used(vq, head, len);
2475 vhost_signal(dev, vq);
2477 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2479 /* multi-buffer version of vhost_add_used_and_signal */
2480 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2481 struct vhost_virtqueue *vq,
2482 struct vring_used_elem *heads, unsigned count)
2484 vhost_add_used_n(vq, heads, count);
2485 vhost_signal(dev, vq);
2487 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2489 /* return true if we're sure that avaiable ring is empty */
2490 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2492 __virtio16 avail_idx;
2495 if (vq->avail_idx != vq->last_avail_idx)
2498 r = vhost_get_avail_idx(vq, &avail_idx);
2501 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2503 return vq->avail_idx == vq->last_avail_idx;
2505 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2507 /* OK, now we need to know about added descriptors. */
2508 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2510 __virtio16 avail_idx;
2513 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2515 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2516 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2517 r = vhost_update_used_flags(vq);
2519 vq_err(vq, "Failed to enable notification at %p: %d\n",
2520 &vq->used->flags, r);
2524 r = vhost_update_avail_event(vq);
2526 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2527 vhost_avail_event(vq), r);
2531 /* They could have slipped one in as we were doing that: make
2532 * sure it's written, then check again. */
2534 r = vhost_get_avail_idx(vq, &avail_idx);
2536 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2537 &vq->avail->idx, r);
2540 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2542 return vq->avail_idx != vq->last_avail_idx;
2544 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2546 /* We don't need to be notified again. */
2547 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2551 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2553 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2554 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2555 r = vhost_update_used_flags(vq);
2557 vq_err(vq, "Failed to disable notification at %p: %d\n",
2558 &vq->used->flags, r);
2561 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2563 /* Create a new message. */
2564 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2566 struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2570 /* Make sure all padding within the structure is initialized. */
2571 memset(&node->msg, 0, sizeof node->msg);
2573 node->msg.type = type;
2576 EXPORT_SYMBOL_GPL(vhost_new_msg);
2578 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2579 struct vhost_msg_node *node)
2581 spin_lock(&dev->iotlb_lock);
2582 list_add_tail(&node->node, head);
2583 spin_unlock(&dev->iotlb_lock);
2585 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2587 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2589 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2590 struct list_head *head)
2592 struct vhost_msg_node *node = NULL;
2594 spin_lock(&dev->iotlb_lock);
2595 if (!list_empty(head)) {
2596 node = list_first_entry(head, struct vhost_msg_node,
2598 list_del(&node->node);
2600 spin_unlock(&dev->iotlb_lock);
2604 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2606 void vhost_set_backend_features(struct vhost_dev *dev, u64 features)
2608 struct vhost_virtqueue *vq;
2611 mutex_lock(&dev->mutex);
2612 for (i = 0; i < dev->nvqs; ++i) {
2614 mutex_lock(&vq->mutex);
2615 vq->acked_backend_features = features;
2616 mutex_unlock(&vq->mutex);
2618 mutex_unlock(&dev->mutex);
2620 EXPORT_SYMBOL_GPL(vhost_set_backend_features);
2622 static int __init vhost_init(void)
2627 static void __exit vhost_exit(void)
2631 module_init(vhost_init);
2632 module_exit(vhost_exit);
2634 MODULE_VERSION("0.0.1");
2635 MODULE_LICENSE("GPL v2");
2636 MODULE_AUTHOR("Michael S. Tsirkin");
2637 MODULE_DESCRIPTION("Host kernel accelerator for virtio");