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 wake_up_process(dev->worker->vtsk->task);
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 int vhost_worker(void *data)
338 struct vhost_worker *worker = data;
339 struct vhost_work *work, *work_next;
340 struct llist_node *node;
343 /* mb paired w/ kthread_stop */
344 set_current_state(TASK_INTERRUPTIBLE);
346 if (vhost_task_should_stop(worker->vtsk)) {
347 __set_current_state(TASK_RUNNING);
351 node = llist_del_all(&worker->work_list);
355 node = llist_reverse_order(node);
356 /* make sure flag is seen after deletion */
358 llist_for_each_entry_safe(work, work_next, node, node) {
359 clear_bit(VHOST_WORK_QUEUED, &work->flags);
360 __set_current_state(TASK_RUNNING);
361 kcov_remote_start_common(worker->kcov_handle);
371 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
381 /* Helper to allocate iovec buffers for all vqs. */
382 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
384 struct vhost_virtqueue *vq;
387 for (i = 0; i < dev->nvqs; ++i) {
389 vq->indirect = kmalloc_array(UIO_MAXIOV,
390 sizeof(*vq->indirect),
392 vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
394 vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
396 if (!vq->indirect || !vq->log || !vq->heads)
403 vhost_vq_free_iovecs(dev->vqs[i]);
407 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
411 for (i = 0; i < dev->nvqs; ++i)
412 vhost_vq_free_iovecs(dev->vqs[i]);
415 bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
416 int pkts, int total_len)
418 struct vhost_dev *dev = vq->dev;
420 if ((dev->byte_weight && total_len >= dev->byte_weight) ||
421 pkts >= dev->weight) {
422 vhost_poll_queue(&vq->poll);
428 EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
430 static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
433 size_t event __maybe_unused =
434 vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
436 return size_add(struct_size(vq->avail, ring, num), event);
439 static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
442 size_t event __maybe_unused =
443 vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
445 return size_add(struct_size(vq->used, ring, num), event);
448 static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
451 return sizeof(*vq->desc) * num;
454 void vhost_dev_init(struct vhost_dev *dev,
455 struct vhost_virtqueue **vqs, int nvqs,
456 int iov_limit, int weight, int byte_weight,
458 int (*msg_handler)(struct vhost_dev *dev, u32 asid,
459 struct vhost_iotlb_msg *msg))
461 struct vhost_virtqueue *vq;
466 mutex_init(&dev->mutex);
472 dev->iov_limit = iov_limit;
473 dev->weight = weight;
474 dev->byte_weight = byte_weight;
475 dev->use_worker = use_worker;
476 dev->msg_handler = msg_handler;
477 init_waitqueue_head(&dev->wait);
478 INIT_LIST_HEAD(&dev->read_list);
479 INIT_LIST_HEAD(&dev->pending_list);
480 spin_lock_init(&dev->iotlb_lock);
483 for (i = 0; i < dev->nvqs; ++i) {
489 mutex_init(&vq->mutex);
490 vhost_vq_reset(dev, vq);
492 vhost_poll_init(&vq->poll, vq->handle_kick,
496 EXPORT_SYMBOL_GPL(vhost_dev_init);
498 /* Caller should have device mutex */
499 long vhost_dev_check_owner(struct vhost_dev *dev)
501 /* Are you the owner? If not, I don't think you mean to do that */
502 return dev->mm == current->mm ? 0 : -EPERM;
504 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
506 /* Caller should have device mutex */
507 bool vhost_dev_has_owner(struct vhost_dev *dev)
511 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
513 static void vhost_attach_mm(struct vhost_dev *dev)
515 /* No owner, become one */
516 if (dev->use_worker) {
517 dev->mm = get_task_mm(current);
519 /* vDPA device does not use worker thead, so there's
520 * no need to hold the address space for mm. This help
521 * to avoid deadlock in the case of mmap() which may
522 * held the refcnt of the file and depends on release
523 * method to remove vma.
525 dev->mm = current->mm;
530 static void vhost_detach_mm(struct vhost_dev *dev)
543 static void vhost_worker_free(struct vhost_dev *dev)
545 struct vhost_worker *worker = dev->worker;
551 WARN_ON(!llist_empty(&worker->work_list));
552 vhost_task_stop(worker->vtsk);
556 static int vhost_worker_create(struct vhost_dev *dev)
558 struct vhost_worker *worker;
559 struct vhost_task *vtsk;
560 char name[TASK_COMM_LEN];
563 worker = kzalloc(sizeof(*worker), GFP_KERNEL_ACCOUNT);
567 dev->worker = worker;
568 worker->kcov_handle = kcov_common_handle();
569 init_llist_head(&worker->work_list);
570 snprintf(name, sizeof(name), "vhost-%d", current->pid);
572 vtsk = vhost_task_create(vhost_worker, worker, name);
579 vhost_task_start(vtsk);
588 /* Caller should have device mutex */
589 long vhost_dev_set_owner(struct vhost_dev *dev)
593 /* Is there an owner already? */
594 if (vhost_dev_has_owner(dev)) {
599 vhost_attach_mm(dev);
601 if (dev->use_worker) {
602 err = vhost_worker_create(dev);
607 err = vhost_dev_alloc_iovecs(dev);
613 vhost_worker_free(dev);
615 vhost_detach_mm(dev);
619 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
621 static struct vhost_iotlb *iotlb_alloc(void)
623 return vhost_iotlb_alloc(max_iotlb_entries,
624 VHOST_IOTLB_FLAG_RETIRE);
627 struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
629 return iotlb_alloc();
631 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
633 /* Caller should have device mutex */
634 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
638 vhost_dev_cleanup(dev);
641 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
642 * VQs aren't running.
644 for (i = 0; i < dev->nvqs; ++i)
645 dev->vqs[i]->umem = umem;
647 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
649 void vhost_dev_stop(struct vhost_dev *dev)
653 for (i = 0; i < dev->nvqs; ++i) {
654 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick)
655 vhost_poll_stop(&dev->vqs[i]->poll);
658 vhost_dev_flush(dev);
660 EXPORT_SYMBOL_GPL(vhost_dev_stop);
662 void vhost_clear_msg(struct vhost_dev *dev)
664 struct vhost_msg_node *node, *n;
666 spin_lock(&dev->iotlb_lock);
668 list_for_each_entry_safe(node, n, &dev->read_list, node) {
669 list_del(&node->node);
673 list_for_each_entry_safe(node, n, &dev->pending_list, node) {
674 list_del(&node->node);
678 spin_unlock(&dev->iotlb_lock);
680 EXPORT_SYMBOL_GPL(vhost_clear_msg);
682 void vhost_dev_cleanup(struct vhost_dev *dev)
686 for (i = 0; i < dev->nvqs; ++i) {
687 if (dev->vqs[i]->error_ctx)
688 eventfd_ctx_put(dev->vqs[i]->error_ctx);
689 if (dev->vqs[i]->kick)
690 fput(dev->vqs[i]->kick);
691 if (dev->vqs[i]->call_ctx.ctx)
692 eventfd_ctx_put(dev->vqs[i]->call_ctx.ctx);
693 vhost_vq_reset(dev, dev->vqs[i]);
695 vhost_dev_free_iovecs(dev);
697 eventfd_ctx_put(dev->log_ctx);
699 /* No one will access memory at this point */
700 vhost_iotlb_free(dev->umem);
702 vhost_iotlb_free(dev->iotlb);
704 vhost_clear_msg(dev);
705 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
706 vhost_worker_free(dev);
707 vhost_detach_mm(dev);
709 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
711 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
713 u64 a = addr / VHOST_PAGE_SIZE / 8;
715 /* Make sure 64 bit math will not overflow. */
716 if (a > ULONG_MAX - (unsigned long)log_base ||
717 a + (unsigned long)log_base > ULONG_MAX)
720 return access_ok(log_base + a,
721 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
724 /* Make sure 64 bit math will not overflow. */
725 static bool vhost_overflow(u64 uaddr, u64 size)
727 if (uaddr > ULONG_MAX || size > ULONG_MAX)
733 return uaddr > ULONG_MAX - size + 1;
736 /* Caller should have vq mutex and device mutex. */
737 static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
740 struct vhost_iotlb_map *map;
745 list_for_each_entry(map, &umem->list, link) {
746 unsigned long a = map->addr;
748 if (vhost_overflow(map->addr, map->size))
752 if (!access_ok((void __user *)a, map->size))
754 else if (log_all && !log_access_ok(log_base,
762 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
763 u64 addr, unsigned int size,
766 const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
771 return (void __user *)(uintptr_t)(map->addr + addr - map->start);
774 /* Can we switch to this memory table? */
775 /* Caller should have device mutex but not vq mutex */
776 static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
781 for (i = 0; i < d->nvqs; ++i) {
785 mutex_lock(&d->vqs[i]->mutex);
786 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
787 /* If ring is inactive, will check when it's enabled. */
788 if (d->vqs[i]->private_data)
789 ok = vq_memory_access_ok(d->vqs[i]->log_base,
793 mutex_unlock(&d->vqs[i]->mutex);
800 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
801 struct iovec iov[], int iov_size, int access);
803 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
804 const void *from, unsigned size)
809 return __copy_to_user(to, from, size);
811 /* This function should be called after iotlb
812 * prefetch, which means we're sure that all vq
813 * could be access through iotlb. So -EAGAIN should
814 * not happen in this case.
817 void __user *uaddr = vhost_vq_meta_fetch(vq,
818 (u64)(uintptr_t)to, size,
822 return __copy_to_user(uaddr, from, size);
824 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
825 ARRAY_SIZE(vq->iotlb_iov),
829 iov_iter_init(&t, ITER_DEST, vq->iotlb_iov, ret, size);
830 ret = copy_to_iter(from, size, &t);
838 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
839 void __user *from, unsigned size)
844 return __copy_from_user(to, from, size);
846 /* This function should be called after iotlb
847 * prefetch, which means we're sure that vq
848 * could be access through iotlb. So -EAGAIN should
849 * not happen in this case.
851 void __user *uaddr = vhost_vq_meta_fetch(vq,
852 (u64)(uintptr_t)from, size,
857 return __copy_from_user(to, uaddr, size);
859 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
860 ARRAY_SIZE(vq->iotlb_iov),
863 vq_err(vq, "IOTLB translation failure: uaddr "
864 "%p size 0x%llx\n", from,
865 (unsigned long long) size);
868 iov_iter_init(&f, ITER_SOURCE, vq->iotlb_iov, ret, size);
869 ret = copy_from_iter(to, size, &f);
878 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
879 void __user *addr, unsigned int size,
884 ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
885 ARRAY_SIZE(vq->iotlb_iov),
888 vq_err(vq, "IOTLB translation failure: uaddr "
889 "%p size 0x%llx\n", addr,
890 (unsigned long long) size);
894 if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
895 vq_err(vq, "Non atomic userspace memory access: uaddr "
896 "%p size 0x%llx\n", addr,
897 (unsigned long long) size);
901 return vq->iotlb_iov[0].iov_base;
904 /* This function should be called after iotlb
905 * prefetch, which means we're sure that vq
906 * could be access through iotlb. So -EAGAIN should
907 * not happen in this case.
909 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
910 void __user *addr, unsigned int size,
913 void __user *uaddr = vhost_vq_meta_fetch(vq,
914 (u64)(uintptr_t)addr, size, type);
918 return __vhost_get_user_slow(vq, addr, size, type);
921 #define vhost_put_user(vq, x, ptr) \
925 ret = __put_user(x, ptr); \
927 __typeof__(ptr) to = \
928 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
929 sizeof(*ptr), VHOST_ADDR_USED); \
931 ret = __put_user(x, to); \
938 static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
940 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
941 vhost_avail_event(vq));
944 static inline int vhost_put_used(struct vhost_virtqueue *vq,
945 struct vring_used_elem *head, int idx,
948 return vhost_copy_to_user(vq, vq->used->ring + idx, head,
949 count * sizeof(*head));
952 static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
955 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
959 static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
962 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
966 #define vhost_get_user(vq, x, ptr, type) \
970 ret = __get_user(x, ptr); \
972 __typeof__(ptr) from = \
973 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
977 ret = __get_user(x, from); \
984 #define vhost_get_avail(vq, x, ptr) \
985 vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
987 #define vhost_get_used(vq, x, ptr) \
988 vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
990 static void vhost_dev_lock_vqs(struct vhost_dev *d)
993 for (i = 0; i < d->nvqs; ++i)
994 mutex_lock_nested(&d->vqs[i]->mutex, i);
997 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
1000 for (i = 0; i < d->nvqs; ++i)
1001 mutex_unlock(&d->vqs[i]->mutex);
1004 static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
1007 return vhost_get_avail(vq, *idx, &vq->avail->idx);
1010 static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
1011 __virtio16 *head, int idx)
1013 return vhost_get_avail(vq, *head,
1014 &vq->avail->ring[idx & (vq->num - 1)]);
1017 static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
1020 return vhost_get_avail(vq, *flags, &vq->avail->flags);
1023 static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1026 return vhost_get_avail(vq, *event, vhost_used_event(vq));
1029 static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1032 return vhost_get_used(vq, *idx, &vq->used->idx);
1035 static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1036 struct vring_desc *desc, int idx)
1038 return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1041 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1042 struct vhost_iotlb_msg *msg)
1044 struct vhost_msg_node *node, *n;
1046 spin_lock(&d->iotlb_lock);
1048 list_for_each_entry_safe(node, n, &d->pending_list, node) {
1049 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1050 if (msg->iova <= vq_msg->iova &&
1051 msg->iova + msg->size - 1 >= vq_msg->iova &&
1052 vq_msg->type == VHOST_IOTLB_MISS) {
1053 vhost_poll_queue(&node->vq->poll);
1054 list_del(&node->node);
1059 spin_unlock(&d->iotlb_lock);
1062 static bool umem_access_ok(u64 uaddr, u64 size, int access)
1064 unsigned long a = uaddr;
1066 /* Make sure 64 bit math will not overflow. */
1067 if (vhost_overflow(uaddr, size))
1070 if ((access & VHOST_ACCESS_RO) &&
1071 !access_ok((void __user *)a, size))
1073 if ((access & VHOST_ACCESS_WO) &&
1074 !access_ok((void __user *)a, size))
1079 static int vhost_process_iotlb_msg(struct vhost_dev *dev, u32 asid,
1080 struct vhost_iotlb_msg *msg)
1087 mutex_lock(&dev->mutex);
1088 vhost_dev_lock_vqs(dev);
1089 switch (msg->type) {
1090 case VHOST_IOTLB_UPDATE:
1095 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1099 vhost_vq_meta_reset(dev);
1100 if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1101 msg->iova + msg->size - 1,
1102 msg->uaddr, msg->perm)) {
1106 vhost_iotlb_notify_vq(dev, msg);
1108 case VHOST_IOTLB_INVALIDATE:
1113 vhost_vq_meta_reset(dev);
1114 vhost_iotlb_del_range(dev->iotlb, msg->iova,
1115 msg->iova + msg->size - 1);
1122 vhost_dev_unlock_vqs(dev);
1123 mutex_unlock(&dev->mutex);
1127 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1128 struct iov_iter *from)
1130 struct vhost_iotlb_msg msg;
1135 ret = copy_from_iter(&type, sizeof(type), from);
1136 if (ret != sizeof(type)) {
1142 case VHOST_IOTLB_MSG:
1143 /* There maybe a hole after type for V1 message type,
1146 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1148 case VHOST_IOTLB_MSG_V2:
1149 if (vhost_backend_has_feature(dev->vqs[0],
1150 VHOST_BACKEND_F_IOTLB_ASID)) {
1151 ret = copy_from_iter(&asid, sizeof(asid), from);
1152 if (ret != sizeof(asid)) {
1158 offset = sizeof(__u32);
1165 iov_iter_advance(from, offset);
1166 ret = copy_from_iter(&msg, sizeof(msg), from);
1167 if (ret != sizeof(msg)) {
1172 if ((msg.type == VHOST_IOTLB_UPDATE ||
1173 msg.type == VHOST_IOTLB_INVALIDATE) &&
1179 if (dev->msg_handler)
1180 ret = dev->msg_handler(dev, asid, &msg);
1182 ret = vhost_process_iotlb_msg(dev, asid, &msg);
1188 ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1189 sizeof(struct vhost_msg_v2);
1193 EXPORT_SYMBOL(vhost_chr_write_iter);
1195 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1200 poll_wait(file, &dev->wait, wait);
1202 if (!list_empty(&dev->read_list))
1203 mask |= EPOLLIN | EPOLLRDNORM;
1207 EXPORT_SYMBOL(vhost_chr_poll);
1209 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1213 struct vhost_msg_node *node;
1215 unsigned size = sizeof(struct vhost_msg);
1217 if (iov_iter_count(to) < size)
1222 prepare_to_wait(&dev->wait, &wait,
1223 TASK_INTERRUPTIBLE);
1225 node = vhost_dequeue_msg(dev, &dev->read_list);
1232 if (signal_pending(current)) {
1245 finish_wait(&dev->wait, &wait);
1248 struct vhost_iotlb_msg *msg;
1249 void *start = &node->msg;
1251 switch (node->msg.type) {
1252 case VHOST_IOTLB_MSG:
1253 size = sizeof(node->msg);
1254 msg = &node->msg.iotlb;
1256 case VHOST_IOTLB_MSG_V2:
1257 size = sizeof(node->msg_v2);
1258 msg = &node->msg_v2.iotlb;
1265 ret = copy_to_iter(start, size, to);
1266 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1270 vhost_enqueue_msg(dev, &dev->pending_list, node);
1275 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1277 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1279 struct vhost_dev *dev = vq->dev;
1280 struct vhost_msg_node *node;
1281 struct vhost_iotlb_msg *msg;
1282 bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1284 node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1289 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1290 msg = &node->msg_v2.iotlb;
1292 msg = &node->msg.iotlb;
1295 msg->type = VHOST_IOTLB_MISS;
1299 vhost_enqueue_msg(dev, &dev->read_list, node);
1304 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1305 vring_desc_t __user *desc,
1306 vring_avail_t __user *avail,
1307 vring_used_t __user *used)
1310 /* If an IOTLB device is present, the vring addresses are
1311 * GIOVAs. Access validation occurs at prefetch time. */
1315 return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1316 access_ok(avail, vhost_get_avail_size(vq, num)) &&
1317 access_ok(used, vhost_get_used_size(vq, num));
1320 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1321 const struct vhost_iotlb_map *map,
1324 int access = (type == VHOST_ADDR_USED) ?
1325 VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1327 if (likely(map->perm & access))
1328 vq->meta_iotlb[type] = map;
1331 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1332 int access, u64 addr, u64 len, int type)
1334 const struct vhost_iotlb_map *map;
1335 struct vhost_iotlb *umem = vq->iotlb;
1336 u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1338 if (vhost_vq_meta_fetch(vq, addr, len, type))
1342 map = vhost_iotlb_itree_first(umem, addr, last);
1343 if (map == NULL || map->start > addr) {
1344 vhost_iotlb_miss(vq, addr, access);
1346 } else if (!(map->perm & access)) {
1347 /* Report the possible access violation by
1348 * request another translation from userspace.
1353 size = map->size - addr + map->start;
1355 if (orig_addr == addr && size >= len)
1356 vhost_vq_meta_update(vq, map, type);
1365 int vq_meta_prefetch(struct vhost_virtqueue *vq)
1367 unsigned int num = vq->num;
1372 return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1373 vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1374 iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1375 vhost_get_avail_size(vq, num),
1376 VHOST_ADDR_AVAIL) &&
1377 iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1378 vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1380 EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1382 /* Can we log writes? */
1383 /* Caller should have device mutex but not vq mutex */
1384 bool vhost_log_access_ok(struct vhost_dev *dev)
1386 return memory_access_ok(dev, dev->umem, 1);
1388 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1390 static bool vq_log_used_access_ok(struct vhost_virtqueue *vq,
1391 void __user *log_base,
1395 /* If an IOTLB device is present, log_addr is a GIOVA that
1396 * will never be logged by log_used(). */
1400 return !log_used || log_access_ok(log_base, log_addr,
1401 vhost_get_used_size(vq, vq->num));
1404 /* Verify access for write logging. */
1405 /* Caller should have vq mutex and device mutex */
1406 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1407 void __user *log_base)
1409 return vq_memory_access_ok(log_base, vq->umem,
1410 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1411 vq_log_used_access_ok(vq, log_base, vq->log_used, vq->log_addr);
1414 /* Can we start vq? */
1415 /* Caller should have vq mutex and device mutex */
1416 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1418 if (!vq_log_access_ok(vq, vq->log_base))
1421 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1423 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1425 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1427 struct vhost_memory mem, *newmem;
1428 struct vhost_memory_region *region;
1429 struct vhost_iotlb *newumem, *oldumem;
1430 unsigned long size = offsetof(struct vhost_memory, regions);
1433 if (copy_from_user(&mem, m, size))
1437 if (mem.nregions > max_mem_regions)
1439 newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1444 memcpy(newmem, &mem, size);
1445 if (copy_from_user(newmem->regions, m->regions,
1446 flex_array_size(newmem, regions, mem.nregions))) {
1451 newumem = iotlb_alloc();
1457 for (region = newmem->regions;
1458 region < newmem->regions + mem.nregions;
1460 if (vhost_iotlb_add_range(newumem,
1461 region->guest_phys_addr,
1462 region->guest_phys_addr +
1463 region->memory_size - 1,
1464 region->userspace_addr,
1469 if (!memory_access_ok(d, newumem, 0))
1475 /* All memory accesses are done under some VQ mutex. */
1476 for (i = 0; i < d->nvqs; ++i) {
1477 mutex_lock(&d->vqs[i]->mutex);
1478 d->vqs[i]->umem = newumem;
1479 mutex_unlock(&d->vqs[i]->mutex);
1483 vhost_iotlb_free(oldumem);
1487 vhost_iotlb_free(newumem);
1492 static long vhost_vring_set_num(struct vhost_dev *d,
1493 struct vhost_virtqueue *vq,
1496 struct vhost_vring_state s;
1498 /* Resizing ring with an active backend?
1499 * You don't want to do that. */
1500 if (vq->private_data)
1503 if (copy_from_user(&s, argp, sizeof s))
1506 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1513 static long vhost_vring_set_addr(struct vhost_dev *d,
1514 struct vhost_virtqueue *vq,
1517 struct vhost_vring_addr a;
1519 if (copy_from_user(&a, argp, sizeof a))
1521 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1524 /* For 32bit, verify that the top 32bits of the user
1525 data are set to zero. */
1526 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1527 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1528 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1531 /* Make sure it's safe to cast pointers to vring types. */
1532 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1533 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1534 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1535 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1536 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1539 /* We only verify access here if backend is configured.
1540 * If it is not, we don't as size might not have been setup.
1541 * We will verify when backend is configured. */
1542 if (vq->private_data) {
1543 if (!vq_access_ok(vq, vq->num,
1544 (void __user *)(unsigned long)a.desc_user_addr,
1545 (void __user *)(unsigned long)a.avail_user_addr,
1546 (void __user *)(unsigned long)a.used_user_addr))
1549 /* Also validate log access for used ring if enabled. */
1550 if (!vq_log_used_access_ok(vq, vq->log_base,
1551 a.flags & (0x1 << VHOST_VRING_F_LOG),
1556 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1557 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1558 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1559 vq->log_addr = a.log_guest_addr;
1560 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1565 static long vhost_vring_set_num_addr(struct vhost_dev *d,
1566 struct vhost_virtqueue *vq,
1572 mutex_lock(&vq->mutex);
1575 case VHOST_SET_VRING_NUM:
1576 r = vhost_vring_set_num(d, vq, argp);
1578 case VHOST_SET_VRING_ADDR:
1579 r = vhost_vring_set_addr(d, vq, argp);
1585 mutex_unlock(&vq->mutex);
1589 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1591 struct file *eventfp, *filep = NULL;
1592 bool pollstart = false, pollstop = false;
1593 struct eventfd_ctx *ctx = NULL;
1594 u32 __user *idxp = argp;
1595 struct vhost_virtqueue *vq;
1596 struct vhost_vring_state s;
1597 struct vhost_vring_file f;
1601 r = get_user(idx, idxp);
1607 idx = array_index_nospec(idx, d->nvqs);
1610 if (ioctl == VHOST_SET_VRING_NUM ||
1611 ioctl == VHOST_SET_VRING_ADDR) {
1612 return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1615 mutex_lock(&vq->mutex);
1618 case VHOST_SET_VRING_BASE:
1619 /* Moving base with an active backend?
1620 * You don't want to do that. */
1621 if (vq->private_data) {
1625 if (copy_from_user(&s, argp, sizeof s)) {
1629 if (s.num > 0xffff) {
1633 vq->last_avail_idx = s.num;
1634 /* Forget the cached index value. */
1635 vq->avail_idx = vq->last_avail_idx;
1637 case VHOST_GET_VRING_BASE:
1639 s.num = vq->last_avail_idx;
1640 if (copy_to_user(argp, &s, sizeof s))
1643 case VHOST_SET_VRING_KICK:
1644 if (copy_from_user(&f, argp, sizeof f)) {
1648 eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
1649 if (IS_ERR(eventfp)) {
1650 r = PTR_ERR(eventfp);
1653 if (eventfp != vq->kick) {
1654 pollstop = (filep = vq->kick) != NULL;
1655 pollstart = (vq->kick = eventfp) != NULL;
1659 case VHOST_SET_VRING_CALL:
1660 if (copy_from_user(&f, argp, sizeof f)) {
1664 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1670 swap(ctx, vq->call_ctx.ctx);
1672 case VHOST_SET_VRING_ERR:
1673 if (copy_from_user(&f, argp, sizeof f)) {
1677 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1682 swap(ctx, vq->error_ctx);
1684 case VHOST_SET_VRING_ENDIAN:
1685 r = vhost_set_vring_endian(vq, argp);
1687 case VHOST_GET_VRING_ENDIAN:
1688 r = vhost_get_vring_endian(vq, idx, argp);
1690 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1691 if (copy_from_user(&s, argp, sizeof(s))) {
1695 vq->busyloop_timeout = s.num;
1697 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1699 s.num = vq->busyloop_timeout;
1700 if (copy_to_user(argp, &s, sizeof(s)))
1707 if (pollstop && vq->handle_kick)
1708 vhost_poll_stop(&vq->poll);
1710 if (!IS_ERR_OR_NULL(ctx))
1711 eventfd_ctx_put(ctx);
1715 if (pollstart && vq->handle_kick)
1716 r = vhost_poll_start(&vq->poll, vq->kick);
1718 mutex_unlock(&vq->mutex);
1720 if (pollstop && vq->handle_kick)
1721 vhost_dev_flush(vq->poll.dev);
1724 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1726 int vhost_init_device_iotlb(struct vhost_dev *d)
1728 struct vhost_iotlb *niotlb, *oiotlb;
1731 niotlb = iotlb_alloc();
1738 for (i = 0; i < d->nvqs; ++i) {
1739 struct vhost_virtqueue *vq = d->vqs[i];
1741 mutex_lock(&vq->mutex);
1743 __vhost_vq_meta_reset(vq);
1744 mutex_unlock(&vq->mutex);
1747 vhost_iotlb_free(oiotlb);
1751 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1753 /* Caller must have device mutex */
1754 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1756 struct eventfd_ctx *ctx;
1761 /* If you are not the owner, you can become one */
1762 if (ioctl == VHOST_SET_OWNER) {
1763 r = vhost_dev_set_owner(d);
1767 /* You must be the owner to do anything else */
1768 r = vhost_dev_check_owner(d);
1773 case VHOST_SET_MEM_TABLE:
1774 r = vhost_set_memory(d, argp);
1776 case VHOST_SET_LOG_BASE:
1777 if (copy_from_user(&p, argp, sizeof p)) {
1781 if ((u64)(unsigned long)p != p) {
1785 for (i = 0; i < d->nvqs; ++i) {
1786 struct vhost_virtqueue *vq;
1787 void __user *base = (void __user *)(unsigned long)p;
1789 mutex_lock(&vq->mutex);
1790 /* If ring is inactive, will check when it's enabled. */
1791 if (vq->private_data && !vq_log_access_ok(vq, base))
1794 vq->log_base = base;
1795 mutex_unlock(&vq->mutex);
1798 case VHOST_SET_LOG_FD:
1799 r = get_user(fd, (int __user *)argp);
1802 ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
1807 swap(ctx, d->log_ctx);
1808 for (i = 0; i < d->nvqs; ++i) {
1809 mutex_lock(&d->vqs[i]->mutex);
1810 d->vqs[i]->log_ctx = d->log_ctx;
1811 mutex_unlock(&d->vqs[i]->mutex);
1814 eventfd_ctx_put(ctx);
1823 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1825 /* TODO: This is really inefficient. We need something like get_user()
1826 * (instruction directly accesses the data, with an exception table entry
1827 * returning -EFAULT). See Documentation/arch/x86/exception-tables.rst.
1829 static int set_bit_to_user(int nr, void __user *addr)
1831 unsigned long log = (unsigned long)addr;
1834 int bit = nr + (log % PAGE_SIZE) * 8;
1837 r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
1841 base = kmap_atomic(page);
1843 kunmap_atomic(base);
1844 unpin_user_pages_dirty_lock(&page, 1, true);
1848 static int log_write(void __user *log_base,
1849 u64 write_address, u64 write_length)
1851 u64 write_page = write_address / VHOST_PAGE_SIZE;
1856 write_length += write_address % VHOST_PAGE_SIZE;
1858 u64 base = (u64)(unsigned long)log_base;
1859 u64 log = base + write_page / 8;
1860 int bit = write_page % 8;
1861 if ((u64)(unsigned long)log != log)
1863 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1866 if (write_length <= VHOST_PAGE_SIZE)
1868 write_length -= VHOST_PAGE_SIZE;
1874 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1876 struct vhost_iotlb *umem = vq->umem;
1877 struct vhost_iotlb_map *u;
1878 u64 start, end, l, min;
1884 /* More than one GPAs can be mapped into a single HVA. So
1885 * iterate all possible umems here to be safe.
1887 list_for_each_entry(u, &umem->list, link) {
1888 if (u->addr > hva - 1 + len ||
1889 u->addr - 1 + u->size < hva)
1891 start = max(u->addr, hva);
1892 end = min(u->addr - 1 + u->size, hva - 1 + len);
1893 l = end - start + 1;
1894 r = log_write(vq->log_base,
1895 u->start + start - u->addr,
1913 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1915 struct iovec *iov = vq->log_iov;
1919 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1921 ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1922 len, iov, 64, VHOST_ACCESS_WO);
1926 for (i = 0; i < ret; i++) {
1927 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1936 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1937 unsigned int log_num, u64 len, struct iovec *iov, int count)
1941 /* Make sure data written is seen before log. */
1945 for (i = 0; i < count; i++) {
1946 r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1954 for (i = 0; i < log_num; ++i) {
1955 u64 l = min(log[i].len, len);
1956 r = log_write(vq->log_base, log[i].addr, l);
1962 eventfd_signal(vq->log_ctx, 1);
1966 /* Length written exceeds what we have stored. This is a bug. */
1970 EXPORT_SYMBOL_GPL(vhost_log_write);
1972 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1975 if (vhost_put_used_flags(vq))
1977 if (unlikely(vq->log_used)) {
1978 /* Make sure the flag is seen before log. */
1980 /* Log used flag write. */
1981 used = &vq->used->flags;
1982 log_used(vq, (used - (void __user *)vq->used),
1983 sizeof vq->used->flags);
1985 eventfd_signal(vq->log_ctx, 1);
1990 static int vhost_update_avail_event(struct vhost_virtqueue *vq)
1992 if (vhost_put_avail_event(vq))
1994 if (unlikely(vq->log_used)) {
1996 /* Make sure the event is seen before log. */
1998 /* Log avail event write */
1999 used = vhost_avail_event(vq);
2000 log_used(vq, (used - (void __user *)vq->used),
2001 sizeof *vhost_avail_event(vq));
2003 eventfd_signal(vq->log_ctx, 1);
2008 int vhost_vq_init_access(struct vhost_virtqueue *vq)
2010 __virtio16 last_used_idx;
2012 bool is_le = vq->is_le;
2014 if (!vq->private_data)
2017 vhost_init_is_le(vq);
2019 r = vhost_update_used_flags(vq);
2022 vq->signalled_used_valid = false;
2024 !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2028 r = vhost_get_used_idx(vq, &last_used_idx);
2030 vq_err(vq, "Can't access used idx at %p\n",
2034 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2041 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2043 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2044 struct iovec iov[], int iov_size, int access)
2046 const struct vhost_iotlb_map *map;
2047 struct vhost_dev *dev = vq->dev;
2048 struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2050 u64 s = 0, last = addr + len - 1;
2053 while ((u64)len > s) {
2055 if (unlikely(ret >= iov_size)) {
2060 map = vhost_iotlb_itree_first(umem, addr, last);
2061 if (map == NULL || map->start > addr) {
2062 if (umem != dev->iotlb) {
2068 } else if (!(map->perm & access)) {
2074 size = map->size - addr + map->start;
2075 _iov->iov_len = min((u64)len - s, size);
2076 _iov->iov_base = (void __user *)(unsigned long)
2077 (map->addr + addr - map->start);
2084 vhost_iotlb_miss(vq, addr, access);
2088 /* Each buffer in the virtqueues is actually a chain of descriptors. This
2089 * function returns the next descriptor in the chain,
2090 * or -1U if we're at the end. */
2091 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2095 /* If this descriptor says it doesn't chain, we're done. */
2096 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2099 /* Check they're not leading us off end of descriptors. */
2100 next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2104 static int get_indirect(struct vhost_virtqueue *vq,
2105 struct iovec iov[], unsigned int iov_size,
2106 unsigned int *out_num, unsigned int *in_num,
2107 struct vhost_log *log, unsigned int *log_num,
2108 struct vring_desc *indirect)
2110 struct vring_desc desc;
2111 unsigned int i = 0, count, found = 0;
2112 u32 len = vhost32_to_cpu(vq, indirect->len);
2113 struct iov_iter from;
2117 if (unlikely(len % sizeof desc)) {
2118 vq_err(vq, "Invalid length in indirect descriptor: "
2119 "len 0x%llx not multiple of 0x%zx\n",
2120 (unsigned long long)len,
2125 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2126 UIO_MAXIOV, VHOST_ACCESS_RO);
2127 if (unlikely(ret < 0)) {
2129 vq_err(vq, "Translation failure %d in indirect.\n", ret);
2132 iov_iter_init(&from, ITER_SOURCE, vq->indirect, ret, len);
2133 count = len / sizeof desc;
2134 /* Buffers are chained via a 16 bit next field, so
2135 * we can have at most 2^16 of these. */
2136 if (unlikely(count > USHRT_MAX + 1)) {
2137 vq_err(vq, "Indirect buffer length too big: %d\n",
2143 unsigned iov_count = *in_num + *out_num;
2144 if (unlikely(++found > count)) {
2145 vq_err(vq, "Loop detected: last one at %u "
2146 "indirect size %u\n",
2150 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2151 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2152 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2155 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2156 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2157 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2161 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2162 access = VHOST_ACCESS_WO;
2164 access = VHOST_ACCESS_RO;
2166 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2167 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2168 iov_size - iov_count, access);
2169 if (unlikely(ret < 0)) {
2171 vq_err(vq, "Translation failure %d indirect idx %d\n",
2175 /* If this is an input descriptor, increment that count. */
2176 if (access == VHOST_ACCESS_WO) {
2178 if (unlikely(log && ret)) {
2179 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2180 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2184 /* If it's an output descriptor, they're all supposed
2185 * to come before any input descriptors. */
2186 if (unlikely(*in_num)) {
2187 vq_err(vq, "Indirect descriptor "
2188 "has out after in: idx %d\n", i);
2193 } while ((i = next_desc(vq, &desc)) != -1);
2197 /* This looks in the virtqueue and for the first available buffer, and converts
2198 * it to an iovec for convenient access. Since descriptors consist of some
2199 * number of output then some number of input descriptors, it's actually two
2200 * iovecs, but we pack them into one and note how many of each there were.
2202 * This function returns the descriptor number found, or vq->num (which is
2203 * never a valid descriptor number) if none was found. A negative code is
2204 * returned on error. */
2205 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2206 struct iovec iov[], unsigned int iov_size,
2207 unsigned int *out_num, unsigned int *in_num,
2208 struct vhost_log *log, unsigned int *log_num)
2210 struct vring_desc desc;
2211 unsigned int i, head, found = 0;
2213 __virtio16 avail_idx;
2214 __virtio16 ring_head;
2217 /* Check it isn't doing very strange things with descriptor numbers. */
2218 last_avail_idx = vq->last_avail_idx;
2220 if (vq->avail_idx == vq->last_avail_idx) {
2221 if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2222 vq_err(vq, "Failed to access avail idx at %p\n",
2226 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2228 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2229 vq_err(vq, "Guest moved used index from %u to %u",
2230 last_avail_idx, vq->avail_idx);
2234 /* If there's nothing new since last we looked, return
2237 if (vq->avail_idx == last_avail_idx)
2240 /* Only get avail ring entries after they have been
2246 /* Grab the next descriptor number they're advertising, and increment
2247 * the index we've seen. */
2248 if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2249 vq_err(vq, "Failed to read head: idx %d address %p\n",
2251 &vq->avail->ring[last_avail_idx % vq->num]);
2255 head = vhost16_to_cpu(vq, ring_head);
2257 /* If their number is silly, that's an error. */
2258 if (unlikely(head >= vq->num)) {
2259 vq_err(vq, "Guest says index %u > %u is available",
2264 /* When we start there are none of either input nor output. */
2265 *out_num = *in_num = 0;
2271 unsigned iov_count = *in_num + *out_num;
2272 if (unlikely(i >= vq->num)) {
2273 vq_err(vq, "Desc index is %u > %u, head = %u",
2277 if (unlikely(++found > vq->num)) {
2278 vq_err(vq, "Loop detected: last one at %u "
2279 "vq size %u head %u\n",
2283 ret = vhost_get_desc(vq, &desc, i);
2284 if (unlikely(ret)) {
2285 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2289 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2290 ret = get_indirect(vq, iov, iov_size,
2292 log, log_num, &desc);
2293 if (unlikely(ret < 0)) {
2295 vq_err(vq, "Failure detected "
2296 "in indirect descriptor at idx %d\n", i);
2302 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2303 access = VHOST_ACCESS_WO;
2305 access = VHOST_ACCESS_RO;
2306 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2307 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2308 iov_size - iov_count, access);
2309 if (unlikely(ret < 0)) {
2311 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2315 if (access == VHOST_ACCESS_WO) {
2316 /* If this is an input descriptor,
2317 * increment that count. */
2319 if (unlikely(log && ret)) {
2320 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2321 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2325 /* If it's an output descriptor, they're all supposed
2326 * to come before any input descriptors. */
2327 if (unlikely(*in_num)) {
2328 vq_err(vq, "Descriptor has out after in: "
2334 } while ((i = next_desc(vq, &desc)) != -1);
2336 /* On success, increment avail index. */
2337 vq->last_avail_idx++;
2339 /* Assume notifications from guest are disabled at this point,
2340 * if they aren't we would need to update avail_event index. */
2341 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2344 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2346 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2347 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2349 vq->last_avail_idx -= n;
2351 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2353 /* After we've used one of their buffers, we tell them about it. We'll then
2354 * want to notify the guest, using eventfd. */
2355 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2357 struct vring_used_elem heads = {
2358 cpu_to_vhost32(vq, head),
2359 cpu_to_vhost32(vq, len)
2362 return vhost_add_used_n(vq, &heads, 1);
2364 EXPORT_SYMBOL_GPL(vhost_add_used);
2366 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2367 struct vring_used_elem *heads,
2370 vring_used_elem_t __user *used;
2374 start = vq->last_used_idx & (vq->num - 1);
2375 used = vq->used->ring + start;
2376 if (vhost_put_used(vq, heads, start, count)) {
2377 vq_err(vq, "Failed to write used");
2380 if (unlikely(vq->log_used)) {
2381 /* Make sure data is seen before log. */
2383 /* Log used ring entry write. */
2384 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2385 count * sizeof *used);
2387 old = vq->last_used_idx;
2388 new = (vq->last_used_idx += count);
2389 /* If the driver never bothers to signal in a very long while,
2390 * used index might wrap around. If that happens, invalidate
2391 * signalled_used index we stored. TODO: make sure driver
2392 * signals at least once in 2^16 and remove this. */
2393 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2394 vq->signalled_used_valid = false;
2398 /* After we've used one of their buffers, we tell them about it. We'll then
2399 * want to notify the guest, using eventfd. */
2400 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2405 start = vq->last_used_idx & (vq->num - 1);
2406 n = vq->num - start;
2408 r = __vhost_add_used_n(vq, heads, n);
2414 r = __vhost_add_used_n(vq, heads, count);
2416 /* Make sure buffer is written before we update index. */
2418 if (vhost_put_used_idx(vq)) {
2419 vq_err(vq, "Failed to increment used idx");
2422 if (unlikely(vq->log_used)) {
2423 /* Make sure used idx is seen before log. */
2425 /* Log used index update. */
2426 log_used(vq, offsetof(struct vring_used, idx),
2427 sizeof vq->used->idx);
2429 eventfd_signal(vq->log_ctx, 1);
2433 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2435 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2440 /* Flush out used index updates. This is paired
2441 * with the barrier that the Guest executes when enabling
2445 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2446 unlikely(vq->avail_idx == vq->last_avail_idx))
2449 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2451 if (vhost_get_avail_flags(vq, &flags)) {
2452 vq_err(vq, "Failed to get flags");
2455 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2457 old = vq->signalled_used;
2458 v = vq->signalled_used_valid;
2459 new = vq->signalled_used = vq->last_used_idx;
2460 vq->signalled_used_valid = true;
2465 if (vhost_get_used_event(vq, &event)) {
2466 vq_err(vq, "Failed to get used event idx");
2469 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2472 /* This actually signals the guest, using eventfd. */
2473 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2475 /* Signal the Guest tell them we used something up. */
2476 if (vq->call_ctx.ctx && vhost_notify(dev, vq))
2477 eventfd_signal(vq->call_ctx.ctx, 1);
2479 EXPORT_SYMBOL_GPL(vhost_signal);
2481 /* And here's the combo meal deal. Supersize me! */
2482 void vhost_add_used_and_signal(struct vhost_dev *dev,
2483 struct vhost_virtqueue *vq,
2484 unsigned int head, int len)
2486 vhost_add_used(vq, head, len);
2487 vhost_signal(dev, vq);
2489 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2491 /* multi-buffer version of vhost_add_used_and_signal */
2492 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2493 struct vhost_virtqueue *vq,
2494 struct vring_used_elem *heads, unsigned count)
2496 vhost_add_used_n(vq, heads, count);
2497 vhost_signal(dev, vq);
2499 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2501 /* return true if we're sure that avaiable ring is empty */
2502 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2504 __virtio16 avail_idx;
2507 if (vq->avail_idx != vq->last_avail_idx)
2510 r = vhost_get_avail_idx(vq, &avail_idx);
2513 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2515 return vq->avail_idx == vq->last_avail_idx;
2517 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2519 /* OK, now we need to know about added descriptors. */
2520 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2522 __virtio16 avail_idx;
2525 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2527 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2528 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2529 r = vhost_update_used_flags(vq);
2531 vq_err(vq, "Failed to enable notification at %p: %d\n",
2532 &vq->used->flags, r);
2536 r = vhost_update_avail_event(vq);
2538 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2539 vhost_avail_event(vq), r);
2543 /* They could have slipped one in as we were doing that: make
2544 * sure it's written, then check again. */
2546 r = vhost_get_avail_idx(vq, &avail_idx);
2548 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2549 &vq->avail->idx, r);
2552 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2554 return vq->avail_idx != vq->last_avail_idx;
2556 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2558 /* We don't need to be notified again. */
2559 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2563 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2565 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2566 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2567 r = vhost_update_used_flags(vq);
2569 vq_err(vq, "Failed to disable notification at %p: %d\n",
2570 &vq->used->flags, r);
2573 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2575 /* Create a new message. */
2576 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2578 struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2582 /* Make sure all padding within the structure is initialized. */
2583 memset(&node->msg, 0, sizeof node->msg);
2585 node->msg.type = type;
2588 EXPORT_SYMBOL_GPL(vhost_new_msg);
2590 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2591 struct vhost_msg_node *node)
2593 spin_lock(&dev->iotlb_lock);
2594 list_add_tail(&node->node, head);
2595 spin_unlock(&dev->iotlb_lock);
2597 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2599 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2601 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2602 struct list_head *head)
2604 struct vhost_msg_node *node = NULL;
2606 spin_lock(&dev->iotlb_lock);
2607 if (!list_empty(head)) {
2608 node = list_first_entry(head, struct vhost_msg_node,
2610 list_del(&node->node);
2612 spin_unlock(&dev->iotlb_lock);
2616 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2618 void vhost_set_backend_features(struct vhost_dev *dev, u64 features)
2620 struct vhost_virtqueue *vq;
2623 mutex_lock(&dev->mutex);
2624 for (i = 0; i < dev->nvqs; ++i) {
2626 mutex_lock(&vq->mutex);
2627 vq->acked_backend_features = features;
2628 mutex_unlock(&vq->mutex);
2630 mutex_unlock(&dev->mutex);
2632 EXPORT_SYMBOL_GPL(vhost_set_backend_features);
2634 static int __init vhost_init(void)
2639 static void __exit vhost_exit(void)
2643 module_init(vhost_init);
2644 module_exit(vhost_exit);
2646 MODULE_VERSION("0.0.1");
2647 MODULE_LICENSE("GPL v2");
2648 MODULE_AUTHOR("Michael S. Tsirkin");
2649 MODULE_DESCRIPTION("Host kernel accelerator for virtio");