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);
372 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
382 /* Helper to allocate iovec buffers for all vqs. */
383 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
385 struct vhost_virtqueue *vq;
388 for (i = 0; i < dev->nvqs; ++i) {
390 vq->indirect = kmalloc_array(UIO_MAXIOV,
391 sizeof(*vq->indirect),
393 vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
395 vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
397 if (!vq->indirect || !vq->log || !vq->heads)
404 vhost_vq_free_iovecs(dev->vqs[i]);
408 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
412 for (i = 0; i < dev->nvqs; ++i)
413 vhost_vq_free_iovecs(dev->vqs[i]);
416 bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
417 int pkts, int total_len)
419 struct vhost_dev *dev = vq->dev;
421 if ((dev->byte_weight && total_len >= dev->byte_weight) ||
422 pkts >= dev->weight) {
423 vhost_poll_queue(&vq->poll);
429 EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
431 static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
434 size_t event __maybe_unused =
435 vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
437 return size_add(struct_size(vq->avail, ring, num), event);
440 static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
443 size_t event __maybe_unused =
444 vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
446 return size_add(struct_size(vq->used, ring, num), event);
449 static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
452 return sizeof(*vq->desc) * num;
455 void vhost_dev_init(struct vhost_dev *dev,
456 struct vhost_virtqueue **vqs, int nvqs,
457 int iov_limit, int weight, int byte_weight,
459 int (*msg_handler)(struct vhost_dev *dev, u32 asid,
460 struct vhost_iotlb_msg *msg))
462 struct vhost_virtqueue *vq;
467 mutex_init(&dev->mutex);
473 dev->iov_limit = iov_limit;
474 dev->weight = weight;
475 dev->byte_weight = byte_weight;
476 dev->use_worker = use_worker;
477 dev->msg_handler = msg_handler;
478 init_waitqueue_head(&dev->wait);
479 INIT_LIST_HEAD(&dev->read_list);
480 INIT_LIST_HEAD(&dev->pending_list);
481 spin_lock_init(&dev->iotlb_lock);
484 for (i = 0; i < dev->nvqs; ++i) {
490 mutex_init(&vq->mutex);
491 vhost_vq_reset(dev, vq);
493 vhost_poll_init(&vq->poll, vq->handle_kick,
497 EXPORT_SYMBOL_GPL(vhost_dev_init);
499 /* Caller should have device mutex */
500 long vhost_dev_check_owner(struct vhost_dev *dev)
502 /* Are you the owner? If not, I don't think you mean to do that */
503 return dev->mm == current->mm ? 0 : -EPERM;
505 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
507 /* Caller should have device mutex */
508 bool vhost_dev_has_owner(struct vhost_dev *dev)
512 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
514 static void vhost_attach_mm(struct vhost_dev *dev)
516 /* No owner, become one */
517 if (dev->use_worker) {
518 dev->mm = get_task_mm(current);
520 /* vDPA device does not use worker thead, so there's
521 * no need to hold the address space for mm. This help
522 * to avoid deadlock in the case of mmap() which may
523 * held the refcnt of the file and depends on release
524 * method to remove vma.
526 dev->mm = current->mm;
531 static void vhost_detach_mm(struct vhost_dev *dev)
544 static void vhost_worker_free(struct vhost_dev *dev)
546 struct vhost_worker *worker = dev->worker;
552 WARN_ON(!llist_empty(&worker->work_list));
553 vhost_task_stop(worker->vtsk);
557 static int vhost_worker_create(struct vhost_dev *dev)
559 struct vhost_worker *worker;
560 struct vhost_task *vtsk;
561 char name[TASK_COMM_LEN];
564 worker = kzalloc(sizeof(*worker), GFP_KERNEL_ACCOUNT);
568 dev->worker = worker;
569 worker->kcov_handle = kcov_common_handle();
570 init_llist_head(&worker->work_list);
571 snprintf(name, sizeof(name), "vhost-%d", current->pid);
573 vtsk = vhost_task_create(vhost_worker, worker, name);
580 vhost_task_start(vtsk);
589 /* Caller should have device mutex */
590 long vhost_dev_set_owner(struct vhost_dev *dev)
594 /* Is there an owner already? */
595 if (vhost_dev_has_owner(dev)) {
600 vhost_attach_mm(dev);
602 if (dev->use_worker) {
603 err = vhost_worker_create(dev);
608 err = vhost_dev_alloc_iovecs(dev);
614 vhost_worker_free(dev);
616 vhost_detach_mm(dev);
620 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
622 static struct vhost_iotlb *iotlb_alloc(void)
624 return vhost_iotlb_alloc(max_iotlb_entries,
625 VHOST_IOTLB_FLAG_RETIRE);
628 struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
630 return iotlb_alloc();
632 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
634 /* Caller should have device mutex */
635 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
639 vhost_dev_cleanup(dev);
642 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
643 * VQs aren't running.
645 for (i = 0; i < dev->nvqs; ++i)
646 dev->vqs[i]->umem = umem;
648 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
650 void vhost_dev_stop(struct vhost_dev *dev)
654 for (i = 0; i < dev->nvqs; ++i) {
655 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick)
656 vhost_poll_stop(&dev->vqs[i]->poll);
659 vhost_dev_flush(dev);
661 EXPORT_SYMBOL_GPL(vhost_dev_stop);
663 void vhost_clear_msg(struct vhost_dev *dev)
665 struct vhost_msg_node *node, *n;
667 spin_lock(&dev->iotlb_lock);
669 list_for_each_entry_safe(node, n, &dev->read_list, node) {
670 list_del(&node->node);
674 list_for_each_entry_safe(node, n, &dev->pending_list, node) {
675 list_del(&node->node);
679 spin_unlock(&dev->iotlb_lock);
681 EXPORT_SYMBOL_GPL(vhost_clear_msg);
683 void vhost_dev_cleanup(struct vhost_dev *dev)
687 for (i = 0; i < dev->nvqs; ++i) {
688 if (dev->vqs[i]->error_ctx)
689 eventfd_ctx_put(dev->vqs[i]->error_ctx);
690 if (dev->vqs[i]->kick)
691 fput(dev->vqs[i]->kick);
692 if (dev->vqs[i]->call_ctx.ctx)
693 eventfd_ctx_put(dev->vqs[i]->call_ctx.ctx);
694 vhost_vq_reset(dev, dev->vqs[i]);
696 vhost_dev_free_iovecs(dev);
698 eventfd_ctx_put(dev->log_ctx);
700 /* No one will access memory at this point */
701 vhost_iotlb_free(dev->umem);
703 vhost_iotlb_free(dev->iotlb);
705 vhost_clear_msg(dev);
706 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
707 vhost_worker_free(dev);
708 vhost_detach_mm(dev);
710 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
712 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
714 u64 a = addr / VHOST_PAGE_SIZE / 8;
716 /* Make sure 64 bit math will not overflow. */
717 if (a > ULONG_MAX - (unsigned long)log_base ||
718 a + (unsigned long)log_base > ULONG_MAX)
721 return access_ok(log_base + a,
722 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
725 /* Make sure 64 bit math will not overflow. */
726 static bool vhost_overflow(u64 uaddr, u64 size)
728 if (uaddr > ULONG_MAX || size > ULONG_MAX)
734 return uaddr > ULONG_MAX - size + 1;
737 /* Caller should have vq mutex and device mutex. */
738 static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
741 struct vhost_iotlb_map *map;
746 list_for_each_entry(map, &umem->list, link) {
747 unsigned long a = map->addr;
749 if (vhost_overflow(map->addr, map->size))
753 if (!access_ok((void __user *)a, map->size))
755 else if (log_all && !log_access_ok(log_base,
763 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
764 u64 addr, unsigned int size,
767 const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
772 return (void __user *)(uintptr_t)(map->addr + addr - map->start);
775 /* Can we switch to this memory table? */
776 /* Caller should have device mutex but not vq mutex */
777 static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
782 for (i = 0; i < d->nvqs; ++i) {
786 mutex_lock(&d->vqs[i]->mutex);
787 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
788 /* If ring is inactive, will check when it's enabled. */
789 if (d->vqs[i]->private_data)
790 ok = vq_memory_access_ok(d->vqs[i]->log_base,
794 mutex_unlock(&d->vqs[i]->mutex);
801 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
802 struct iovec iov[], int iov_size, int access);
804 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
805 const void *from, unsigned size)
810 return __copy_to_user(to, from, size);
812 /* This function should be called after iotlb
813 * prefetch, which means we're sure that all vq
814 * could be access through iotlb. So -EAGAIN should
815 * not happen in this case.
818 void __user *uaddr = vhost_vq_meta_fetch(vq,
819 (u64)(uintptr_t)to, size,
823 return __copy_to_user(uaddr, from, size);
825 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
826 ARRAY_SIZE(vq->iotlb_iov),
830 iov_iter_init(&t, ITER_DEST, vq->iotlb_iov, ret, size);
831 ret = copy_to_iter(from, size, &t);
839 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
840 void __user *from, unsigned size)
845 return __copy_from_user(to, from, size);
847 /* This function should be called after iotlb
848 * prefetch, which means we're sure that vq
849 * could be access through iotlb. So -EAGAIN should
850 * not happen in this case.
852 void __user *uaddr = vhost_vq_meta_fetch(vq,
853 (u64)(uintptr_t)from, size,
858 return __copy_from_user(to, uaddr, size);
860 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
861 ARRAY_SIZE(vq->iotlb_iov),
864 vq_err(vq, "IOTLB translation failure: uaddr "
865 "%p size 0x%llx\n", from,
866 (unsigned long long) size);
869 iov_iter_init(&f, ITER_SOURCE, vq->iotlb_iov, ret, size);
870 ret = copy_from_iter(to, size, &f);
879 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
880 void __user *addr, unsigned int size,
885 ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
886 ARRAY_SIZE(vq->iotlb_iov),
889 vq_err(vq, "IOTLB translation failure: uaddr "
890 "%p size 0x%llx\n", addr,
891 (unsigned long long) size);
895 if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
896 vq_err(vq, "Non atomic userspace memory access: uaddr "
897 "%p size 0x%llx\n", addr,
898 (unsigned long long) size);
902 return vq->iotlb_iov[0].iov_base;
905 /* This function should be called after iotlb
906 * prefetch, which means we're sure that vq
907 * could be access through iotlb. So -EAGAIN should
908 * not happen in this case.
910 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
911 void __user *addr, unsigned int size,
914 void __user *uaddr = vhost_vq_meta_fetch(vq,
915 (u64)(uintptr_t)addr, size, type);
919 return __vhost_get_user_slow(vq, addr, size, type);
922 #define vhost_put_user(vq, x, ptr) \
926 ret = __put_user(x, ptr); \
928 __typeof__(ptr) to = \
929 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
930 sizeof(*ptr), VHOST_ADDR_USED); \
932 ret = __put_user(x, to); \
939 static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
941 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
942 vhost_avail_event(vq));
945 static inline int vhost_put_used(struct vhost_virtqueue *vq,
946 struct vring_used_elem *head, int idx,
949 return vhost_copy_to_user(vq, vq->used->ring + idx, head,
950 count * sizeof(*head));
953 static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
956 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
960 static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
963 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
967 #define vhost_get_user(vq, x, ptr, type) \
971 ret = __get_user(x, ptr); \
973 __typeof__(ptr) from = \
974 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
978 ret = __get_user(x, from); \
985 #define vhost_get_avail(vq, x, ptr) \
986 vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
988 #define vhost_get_used(vq, x, ptr) \
989 vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
991 static void vhost_dev_lock_vqs(struct vhost_dev *d)
994 for (i = 0; i < d->nvqs; ++i)
995 mutex_lock_nested(&d->vqs[i]->mutex, i);
998 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
1001 for (i = 0; i < d->nvqs; ++i)
1002 mutex_unlock(&d->vqs[i]->mutex);
1005 static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
1008 return vhost_get_avail(vq, *idx, &vq->avail->idx);
1011 static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
1012 __virtio16 *head, int idx)
1014 return vhost_get_avail(vq, *head,
1015 &vq->avail->ring[idx & (vq->num - 1)]);
1018 static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
1021 return vhost_get_avail(vq, *flags, &vq->avail->flags);
1024 static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1027 return vhost_get_avail(vq, *event, vhost_used_event(vq));
1030 static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1033 return vhost_get_used(vq, *idx, &vq->used->idx);
1036 static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1037 struct vring_desc *desc, int idx)
1039 return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1042 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1043 struct vhost_iotlb_msg *msg)
1045 struct vhost_msg_node *node, *n;
1047 spin_lock(&d->iotlb_lock);
1049 list_for_each_entry_safe(node, n, &d->pending_list, node) {
1050 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1051 if (msg->iova <= vq_msg->iova &&
1052 msg->iova + msg->size - 1 >= vq_msg->iova &&
1053 vq_msg->type == VHOST_IOTLB_MISS) {
1054 vhost_poll_queue(&node->vq->poll);
1055 list_del(&node->node);
1060 spin_unlock(&d->iotlb_lock);
1063 static bool umem_access_ok(u64 uaddr, u64 size, int access)
1065 unsigned long a = uaddr;
1067 /* Make sure 64 bit math will not overflow. */
1068 if (vhost_overflow(uaddr, size))
1071 if ((access & VHOST_ACCESS_RO) &&
1072 !access_ok((void __user *)a, size))
1074 if ((access & VHOST_ACCESS_WO) &&
1075 !access_ok((void __user *)a, size))
1080 static int vhost_process_iotlb_msg(struct vhost_dev *dev, u32 asid,
1081 struct vhost_iotlb_msg *msg)
1088 mutex_lock(&dev->mutex);
1089 vhost_dev_lock_vqs(dev);
1090 switch (msg->type) {
1091 case VHOST_IOTLB_UPDATE:
1096 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1100 vhost_vq_meta_reset(dev);
1101 if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1102 msg->iova + msg->size - 1,
1103 msg->uaddr, msg->perm)) {
1107 vhost_iotlb_notify_vq(dev, msg);
1109 case VHOST_IOTLB_INVALIDATE:
1114 vhost_vq_meta_reset(dev);
1115 vhost_iotlb_del_range(dev->iotlb, msg->iova,
1116 msg->iova + msg->size - 1);
1123 vhost_dev_unlock_vqs(dev);
1124 mutex_unlock(&dev->mutex);
1128 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1129 struct iov_iter *from)
1131 struct vhost_iotlb_msg msg;
1136 ret = copy_from_iter(&type, sizeof(type), from);
1137 if (ret != sizeof(type)) {
1143 case VHOST_IOTLB_MSG:
1144 /* There maybe a hole after type for V1 message type,
1147 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1149 case VHOST_IOTLB_MSG_V2:
1150 if (vhost_backend_has_feature(dev->vqs[0],
1151 VHOST_BACKEND_F_IOTLB_ASID)) {
1152 ret = copy_from_iter(&asid, sizeof(asid), from);
1153 if (ret != sizeof(asid)) {
1159 offset = sizeof(__u32);
1166 iov_iter_advance(from, offset);
1167 ret = copy_from_iter(&msg, sizeof(msg), from);
1168 if (ret != sizeof(msg)) {
1173 if ((msg.type == VHOST_IOTLB_UPDATE ||
1174 msg.type == VHOST_IOTLB_INVALIDATE) &&
1180 if (dev->msg_handler)
1181 ret = dev->msg_handler(dev, asid, &msg);
1183 ret = vhost_process_iotlb_msg(dev, asid, &msg);
1189 ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1190 sizeof(struct vhost_msg_v2);
1194 EXPORT_SYMBOL(vhost_chr_write_iter);
1196 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1201 poll_wait(file, &dev->wait, wait);
1203 if (!list_empty(&dev->read_list))
1204 mask |= EPOLLIN | EPOLLRDNORM;
1208 EXPORT_SYMBOL(vhost_chr_poll);
1210 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1214 struct vhost_msg_node *node;
1216 unsigned size = sizeof(struct vhost_msg);
1218 if (iov_iter_count(to) < size)
1223 prepare_to_wait(&dev->wait, &wait,
1224 TASK_INTERRUPTIBLE);
1226 node = vhost_dequeue_msg(dev, &dev->read_list);
1233 if (signal_pending(current)) {
1246 finish_wait(&dev->wait, &wait);
1249 struct vhost_iotlb_msg *msg;
1250 void *start = &node->msg;
1252 switch (node->msg.type) {
1253 case VHOST_IOTLB_MSG:
1254 size = sizeof(node->msg);
1255 msg = &node->msg.iotlb;
1257 case VHOST_IOTLB_MSG_V2:
1258 size = sizeof(node->msg_v2);
1259 msg = &node->msg_v2.iotlb;
1266 ret = copy_to_iter(start, size, to);
1267 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1271 vhost_enqueue_msg(dev, &dev->pending_list, node);
1276 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1278 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1280 struct vhost_dev *dev = vq->dev;
1281 struct vhost_msg_node *node;
1282 struct vhost_iotlb_msg *msg;
1283 bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1285 node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1290 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1291 msg = &node->msg_v2.iotlb;
1293 msg = &node->msg.iotlb;
1296 msg->type = VHOST_IOTLB_MISS;
1300 vhost_enqueue_msg(dev, &dev->read_list, node);
1305 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1306 vring_desc_t __user *desc,
1307 vring_avail_t __user *avail,
1308 vring_used_t __user *used)
1311 /* If an IOTLB device is present, the vring addresses are
1312 * GIOVAs. Access validation occurs at prefetch time. */
1316 return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1317 access_ok(avail, vhost_get_avail_size(vq, num)) &&
1318 access_ok(used, vhost_get_used_size(vq, num));
1321 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1322 const struct vhost_iotlb_map *map,
1325 int access = (type == VHOST_ADDR_USED) ?
1326 VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1328 if (likely(map->perm & access))
1329 vq->meta_iotlb[type] = map;
1332 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1333 int access, u64 addr, u64 len, int type)
1335 const struct vhost_iotlb_map *map;
1336 struct vhost_iotlb *umem = vq->iotlb;
1337 u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1339 if (vhost_vq_meta_fetch(vq, addr, len, type))
1343 map = vhost_iotlb_itree_first(umem, addr, last);
1344 if (map == NULL || map->start > addr) {
1345 vhost_iotlb_miss(vq, addr, access);
1347 } else if (!(map->perm & access)) {
1348 /* Report the possible access violation by
1349 * request another translation from userspace.
1354 size = map->size - addr + map->start;
1356 if (orig_addr == addr && size >= len)
1357 vhost_vq_meta_update(vq, map, type);
1366 int vq_meta_prefetch(struct vhost_virtqueue *vq)
1368 unsigned int num = vq->num;
1373 return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1374 vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1375 iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1376 vhost_get_avail_size(vq, num),
1377 VHOST_ADDR_AVAIL) &&
1378 iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1379 vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1381 EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1383 /* Can we log writes? */
1384 /* Caller should have device mutex but not vq mutex */
1385 bool vhost_log_access_ok(struct vhost_dev *dev)
1387 return memory_access_ok(dev, dev->umem, 1);
1389 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1391 static bool vq_log_used_access_ok(struct vhost_virtqueue *vq,
1392 void __user *log_base,
1396 /* If an IOTLB device is present, log_addr is a GIOVA that
1397 * will never be logged by log_used(). */
1401 return !log_used || log_access_ok(log_base, log_addr,
1402 vhost_get_used_size(vq, vq->num));
1405 /* Verify access for write logging. */
1406 /* Caller should have vq mutex and device mutex */
1407 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1408 void __user *log_base)
1410 return vq_memory_access_ok(log_base, vq->umem,
1411 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1412 vq_log_used_access_ok(vq, log_base, vq->log_used, vq->log_addr);
1415 /* Can we start vq? */
1416 /* Caller should have vq mutex and device mutex */
1417 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1419 if (!vq_log_access_ok(vq, vq->log_base))
1422 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1424 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1426 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1428 struct vhost_memory mem, *newmem;
1429 struct vhost_memory_region *region;
1430 struct vhost_iotlb *newumem, *oldumem;
1431 unsigned long size = offsetof(struct vhost_memory, regions);
1434 if (copy_from_user(&mem, m, size))
1438 if (mem.nregions > max_mem_regions)
1440 newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1445 memcpy(newmem, &mem, size);
1446 if (copy_from_user(newmem->regions, m->regions,
1447 flex_array_size(newmem, regions, mem.nregions))) {
1452 newumem = iotlb_alloc();
1458 for (region = newmem->regions;
1459 region < newmem->regions + mem.nregions;
1461 if (vhost_iotlb_add_range(newumem,
1462 region->guest_phys_addr,
1463 region->guest_phys_addr +
1464 region->memory_size - 1,
1465 region->userspace_addr,
1470 if (!memory_access_ok(d, newumem, 0))
1476 /* All memory accesses are done under some VQ mutex. */
1477 for (i = 0; i < d->nvqs; ++i) {
1478 mutex_lock(&d->vqs[i]->mutex);
1479 d->vqs[i]->umem = newumem;
1480 mutex_unlock(&d->vqs[i]->mutex);
1484 vhost_iotlb_free(oldumem);
1488 vhost_iotlb_free(newumem);
1493 static long vhost_vring_set_num(struct vhost_dev *d,
1494 struct vhost_virtqueue *vq,
1497 struct vhost_vring_state s;
1499 /* Resizing ring with an active backend?
1500 * You don't want to do that. */
1501 if (vq->private_data)
1504 if (copy_from_user(&s, argp, sizeof s))
1507 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1514 static long vhost_vring_set_addr(struct vhost_dev *d,
1515 struct vhost_virtqueue *vq,
1518 struct vhost_vring_addr a;
1520 if (copy_from_user(&a, argp, sizeof a))
1522 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1525 /* For 32bit, verify that the top 32bits of the user
1526 data are set to zero. */
1527 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1528 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1529 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1532 /* Make sure it's safe to cast pointers to vring types. */
1533 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1534 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1535 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1536 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1537 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1540 /* We only verify access here if backend is configured.
1541 * If it is not, we don't as size might not have been setup.
1542 * We will verify when backend is configured. */
1543 if (vq->private_data) {
1544 if (!vq_access_ok(vq, vq->num,
1545 (void __user *)(unsigned long)a.desc_user_addr,
1546 (void __user *)(unsigned long)a.avail_user_addr,
1547 (void __user *)(unsigned long)a.used_user_addr))
1550 /* Also validate log access for used ring if enabled. */
1551 if (!vq_log_used_access_ok(vq, vq->log_base,
1552 a.flags & (0x1 << VHOST_VRING_F_LOG),
1557 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1558 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1559 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1560 vq->log_addr = a.log_guest_addr;
1561 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1566 static long vhost_vring_set_num_addr(struct vhost_dev *d,
1567 struct vhost_virtqueue *vq,
1573 mutex_lock(&vq->mutex);
1576 case VHOST_SET_VRING_NUM:
1577 r = vhost_vring_set_num(d, vq, argp);
1579 case VHOST_SET_VRING_ADDR:
1580 r = vhost_vring_set_addr(d, vq, argp);
1586 mutex_unlock(&vq->mutex);
1590 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1592 struct file *eventfp, *filep = NULL;
1593 bool pollstart = false, pollstop = false;
1594 struct eventfd_ctx *ctx = NULL;
1595 u32 __user *idxp = argp;
1596 struct vhost_virtqueue *vq;
1597 struct vhost_vring_state s;
1598 struct vhost_vring_file f;
1602 r = get_user(idx, idxp);
1608 idx = array_index_nospec(idx, d->nvqs);
1611 if (ioctl == VHOST_SET_VRING_NUM ||
1612 ioctl == VHOST_SET_VRING_ADDR) {
1613 return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1616 mutex_lock(&vq->mutex);
1619 case VHOST_SET_VRING_BASE:
1620 /* Moving base with an active backend?
1621 * You don't want to do that. */
1622 if (vq->private_data) {
1626 if (copy_from_user(&s, argp, sizeof s)) {
1630 if (s.num > 0xffff) {
1634 vq->last_avail_idx = s.num;
1635 /* Forget the cached index value. */
1636 vq->avail_idx = vq->last_avail_idx;
1638 case VHOST_GET_VRING_BASE:
1640 s.num = vq->last_avail_idx;
1641 if (copy_to_user(argp, &s, sizeof s))
1644 case VHOST_SET_VRING_KICK:
1645 if (copy_from_user(&f, argp, sizeof f)) {
1649 eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
1650 if (IS_ERR(eventfp)) {
1651 r = PTR_ERR(eventfp);
1654 if (eventfp != vq->kick) {
1655 pollstop = (filep = vq->kick) != NULL;
1656 pollstart = (vq->kick = eventfp) != NULL;
1660 case VHOST_SET_VRING_CALL:
1661 if (copy_from_user(&f, argp, sizeof f)) {
1665 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1671 swap(ctx, vq->call_ctx.ctx);
1673 case VHOST_SET_VRING_ERR:
1674 if (copy_from_user(&f, argp, sizeof f)) {
1678 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1683 swap(ctx, vq->error_ctx);
1685 case VHOST_SET_VRING_ENDIAN:
1686 r = vhost_set_vring_endian(vq, argp);
1688 case VHOST_GET_VRING_ENDIAN:
1689 r = vhost_get_vring_endian(vq, idx, argp);
1691 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1692 if (copy_from_user(&s, argp, sizeof(s))) {
1696 vq->busyloop_timeout = s.num;
1698 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1700 s.num = vq->busyloop_timeout;
1701 if (copy_to_user(argp, &s, sizeof(s)))
1708 if (pollstop && vq->handle_kick)
1709 vhost_poll_stop(&vq->poll);
1711 if (!IS_ERR_OR_NULL(ctx))
1712 eventfd_ctx_put(ctx);
1716 if (pollstart && vq->handle_kick)
1717 r = vhost_poll_start(&vq->poll, vq->kick);
1719 mutex_unlock(&vq->mutex);
1721 if (pollstop && vq->handle_kick)
1722 vhost_dev_flush(vq->poll.dev);
1725 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1727 int vhost_init_device_iotlb(struct vhost_dev *d)
1729 struct vhost_iotlb *niotlb, *oiotlb;
1732 niotlb = iotlb_alloc();
1739 for (i = 0; i < d->nvqs; ++i) {
1740 struct vhost_virtqueue *vq = d->vqs[i];
1742 mutex_lock(&vq->mutex);
1744 __vhost_vq_meta_reset(vq);
1745 mutex_unlock(&vq->mutex);
1748 vhost_iotlb_free(oiotlb);
1752 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1754 /* Caller must have device mutex */
1755 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1757 struct eventfd_ctx *ctx;
1762 /* If you are not the owner, you can become one */
1763 if (ioctl == VHOST_SET_OWNER) {
1764 r = vhost_dev_set_owner(d);
1768 /* You must be the owner to do anything else */
1769 r = vhost_dev_check_owner(d);
1774 case VHOST_SET_MEM_TABLE:
1775 r = vhost_set_memory(d, argp);
1777 case VHOST_SET_LOG_BASE:
1778 if (copy_from_user(&p, argp, sizeof p)) {
1782 if ((u64)(unsigned long)p != p) {
1786 for (i = 0; i < d->nvqs; ++i) {
1787 struct vhost_virtqueue *vq;
1788 void __user *base = (void __user *)(unsigned long)p;
1790 mutex_lock(&vq->mutex);
1791 /* If ring is inactive, will check when it's enabled. */
1792 if (vq->private_data && !vq_log_access_ok(vq, base))
1795 vq->log_base = base;
1796 mutex_unlock(&vq->mutex);
1799 case VHOST_SET_LOG_FD:
1800 r = get_user(fd, (int __user *)argp);
1803 ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
1808 swap(ctx, d->log_ctx);
1809 for (i = 0; i < d->nvqs; ++i) {
1810 mutex_lock(&d->vqs[i]->mutex);
1811 d->vqs[i]->log_ctx = d->log_ctx;
1812 mutex_unlock(&d->vqs[i]->mutex);
1815 eventfd_ctx_put(ctx);
1824 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1826 /* TODO: This is really inefficient. We need something like get_user()
1827 * (instruction directly accesses the data, with an exception table entry
1828 * returning -EFAULT). See Documentation/arch/x86/exception-tables.rst.
1830 static int set_bit_to_user(int nr, void __user *addr)
1832 unsigned long log = (unsigned long)addr;
1835 int bit = nr + (log % PAGE_SIZE) * 8;
1838 r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
1842 base = kmap_atomic(page);
1844 kunmap_atomic(base);
1845 unpin_user_pages_dirty_lock(&page, 1, true);
1849 static int log_write(void __user *log_base,
1850 u64 write_address, u64 write_length)
1852 u64 write_page = write_address / VHOST_PAGE_SIZE;
1857 write_length += write_address % VHOST_PAGE_SIZE;
1859 u64 base = (u64)(unsigned long)log_base;
1860 u64 log = base + write_page / 8;
1861 int bit = write_page % 8;
1862 if ((u64)(unsigned long)log != log)
1864 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1867 if (write_length <= VHOST_PAGE_SIZE)
1869 write_length -= VHOST_PAGE_SIZE;
1875 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1877 struct vhost_iotlb *umem = vq->umem;
1878 struct vhost_iotlb_map *u;
1879 u64 start, end, l, min;
1885 /* More than one GPAs can be mapped into a single HVA. So
1886 * iterate all possible umems here to be safe.
1888 list_for_each_entry(u, &umem->list, link) {
1889 if (u->addr > hva - 1 + len ||
1890 u->addr - 1 + u->size < hva)
1892 start = max(u->addr, hva);
1893 end = min(u->addr - 1 + u->size, hva - 1 + len);
1894 l = end - start + 1;
1895 r = log_write(vq->log_base,
1896 u->start + start - u->addr,
1914 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1916 struct iovec *iov = vq->log_iov;
1920 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1922 ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1923 len, iov, 64, VHOST_ACCESS_WO);
1927 for (i = 0; i < ret; i++) {
1928 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1937 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1938 unsigned int log_num, u64 len, struct iovec *iov, int count)
1942 /* Make sure data written is seen before log. */
1946 for (i = 0; i < count; i++) {
1947 r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1955 for (i = 0; i < log_num; ++i) {
1956 u64 l = min(log[i].len, len);
1957 r = log_write(vq->log_base, log[i].addr, l);
1963 eventfd_signal(vq->log_ctx, 1);
1967 /* Length written exceeds what we have stored. This is a bug. */
1971 EXPORT_SYMBOL_GPL(vhost_log_write);
1973 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1976 if (vhost_put_used_flags(vq))
1978 if (unlikely(vq->log_used)) {
1979 /* Make sure the flag is seen before log. */
1981 /* Log used flag write. */
1982 used = &vq->used->flags;
1983 log_used(vq, (used - (void __user *)vq->used),
1984 sizeof vq->used->flags);
1986 eventfd_signal(vq->log_ctx, 1);
1991 static int vhost_update_avail_event(struct vhost_virtqueue *vq)
1993 if (vhost_put_avail_event(vq))
1995 if (unlikely(vq->log_used)) {
1997 /* Make sure the event is seen before log. */
1999 /* Log avail event write */
2000 used = vhost_avail_event(vq);
2001 log_used(vq, (used - (void __user *)vq->used),
2002 sizeof *vhost_avail_event(vq));
2004 eventfd_signal(vq->log_ctx, 1);
2009 int vhost_vq_init_access(struct vhost_virtqueue *vq)
2011 __virtio16 last_used_idx;
2013 bool is_le = vq->is_le;
2015 if (!vq->private_data)
2018 vhost_init_is_le(vq);
2020 r = vhost_update_used_flags(vq);
2023 vq->signalled_used_valid = false;
2025 !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2029 r = vhost_get_used_idx(vq, &last_used_idx);
2031 vq_err(vq, "Can't access used idx at %p\n",
2035 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2042 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2044 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2045 struct iovec iov[], int iov_size, int access)
2047 const struct vhost_iotlb_map *map;
2048 struct vhost_dev *dev = vq->dev;
2049 struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2051 u64 s = 0, last = addr + len - 1;
2054 while ((u64)len > s) {
2056 if (unlikely(ret >= iov_size)) {
2061 map = vhost_iotlb_itree_first(umem, addr, last);
2062 if (map == NULL || map->start > addr) {
2063 if (umem != dev->iotlb) {
2069 } else if (!(map->perm & access)) {
2075 size = map->size - addr + map->start;
2076 _iov->iov_len = min((u64)len - s, size);
2077 _iov->iov_base = (void __user *)(unsigned long)
2078 (map->addr + addr - map->start);
2085 vhost_iotlb_miss(vq, addr, access);
2089 /* Each buffer in the virtqueues is actually a chain of descriptors. This
2090 * function returns the next descriptor in the chain,
2091 * or -1U if we're at the end. */
2092 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2096 /* If this descriptor says it doesn't chain, we're done. */
2097 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2100 /* Check they're not leading us off end of descriptors. */
2101 next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2105 static int get_indirect(struct vhost_virtqueue *vq,
2106 struct iovec iov[], unsigned int iov_size,
2107 unsigned int *out_num, unsigned int *in_num,
2108 struct vhost_log *log, unsigned int *log_num,
2109 struct vring_desc *indirect)
2111 struct vring_desc desc;
2112 unsigned int i = 0, count, found = 0;
2113 u32 len = vhost32_to_cpu(vq, indirect->len);
2114 struct iov_iter from;
2118 if (unlikely(len % sizeof desc)) {
2119 vq_err(vq, "Invalid length in indirect descriptor: "
2120 "len 0x%llx not multiple of 0x%zx\n",
2121 (unsigned long long)len,
2126 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2127 UIO_MAXIOV, VHOST_ACCESS_RO);
2128 if (unlikely(ret < 0)) {
2130 vq_err(vq, "Translation failure %d in indirect.\n", ret);
2133 iov_iter_init(&from, ITER_SOURCE, vq->indirect, ret, len);
2134 count = len / sizeof desc;
2135 /* Buffers are chained via a 16 bit next field, so
2136 * we can have at most 2^16 of these. */
2137 if (unlikely(count > USHRT_MAX + 1)) {
2138 vq_err(vq, "Indirect buffer length too big: %d\n",
2144 unsigned iov_count = *in_num + *out_num;
2145 if (unlikely(++found > count)) {
2146 vq_err(vq, "Loop detected: last one at %u "
2147 "indirect size %u\n",
2151 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2152 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2153 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2156 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2157 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2158 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2162 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2163 access = VHOST_ACCESS_WO;
2165 access = VHOST_ACCESS_RO;
2167 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2168 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2169 iov_size - iov_count, access);
2170 if (unlikely(ret < 0)) {
2172 vq_err(vq, "Translation failure %d indirect idx %d\n",
2176 /* If this is an input descriptor, increment that count. */
2177 if (access == VHOST_ACCESS_WO) {
2179 if (unlikely(log && ret)) {
2180 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2181 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2185 /* If it's an output descriptor, they're all supposed
2186 * to come before any input descriptors. */
2187 if (unlikely(*in_num)) {
2188 vq_err(vq, "Indirect descriptor "
2189 "has out after in: idx %d\n", i);
2194 } while ((i = next_desc(vq, &desc)) != -1);
2198 /* This looks in the virtqueue and for the first available buffer, and converts
2199 * it to an iovec for convenient access. Since descriptors consist of some
2200 * number of output then some number of input descriptors, it's actually two
2201 * iovecs, but we pack them into one and note how many of each there were.
2203 * This function returns the descriptor number found, or vq->num (which is
2204 * never a valid descriptor number) if none was found. A negative code is
2205 * returned on error. */
2206 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2207 struct iovec iov[], unsigned int iov_size,
2208 unsigned int *out_num, unsigned int *in_num,
2209 struct vhost_log *log, unsigned int *log_num)
2211 struct vring_desc desc;
2212 unsigned int i, head, found = 0;
2214 __virtio16 avail_idx;
2215 __virtio16 ring_head;
2218 /* Check it isn't doing very strange things with descriptor numbers. */
2219 last_avail_idx = vq->last_avail_idx;
2221 if (vq->avail_idx == vq->last_avail_idx) {
2222 if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2223 vq_err(vq, "Failed to access avail idx at %p\n",
2227 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2229 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2230 vq_err(vq, "Guest moved used index from %u to %u",
2231 last_avail_idx, vq->avail_idx);
2235 /* If there's nothing new since last we looked, return
2238 if (vq->avail_idx == last_avail_idx)
2241 /* Only get avail ring entries after they have been
2247 /* Grab the next descriptor number they're advertising, and increment
2248 * the index we've seen. */
2249 if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2250 vq_err(vq, "Failed to read head: idx %d address %p\n",
2252 &vq->avail->ring[last_avail_idx % vq->num]);
2256 head = vhost16_to_cpu(vq, ring_head);
2258 /* If their number is silly, that's an error. */
2259 if (unlikely(head >= vq->num)) {
2260 vq_err(vq, "Guest says index %u > %u is available",
2265 /* When we start there are none of either input nor output. */
2266 *out_num = *in_num = 0;
2272 unsigned iov_count = *in_num + *out_num;
2273 if (unlikely(i >= vq->num)) {
2274 vq_err(vq, "Desc index is %u > %u, head = %u",
2278 if (unlikely(++found > vq->num)) {
2279 vq_err(vq, "Loop detected: last one at %u "
2280 "vq size %u head %u\n",
2284 ret = vhost_get_desc(vq, &desc, i);
2285 if (unlikely(ret)) {
2286 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2290 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2291 ret = get_indirect(vq, iov, iov_size,
2293 log, log_num, &desc);
2294 if (unlikely(ret < 0)) {
2296 vq_err(vq, "Failure detected "
2297 "in indirect descriptor at idx %d\n", i);
2303 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2304 access = VHOST_ACCESS_WO;
2306 access = VHOST_ACCESS_RO;
2307 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2308 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2309 iov_size - iov_count, access);
2310 if (unlikely(ret < 0)) {
2312 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2316 if (access == VHOST_ACCESS_WO) {
2317 /* If this is an input descriptor,
2318 * increment that count. */
2320 if (unlikely(log && ret)) {
2321 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2322 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2326 /* If it's an output descriptor, they're all supposed
2327 * to come before any input descriptors. */
2328 if (unlikely(*in_num)) {
2329 vq_err(vq, "Descriptor has out after in: "
2335 } while ((i = next_desc(vq, &desc)) != -1);
2337 /* On success, increment avail index. */
2338 vq->last_avail_idx++;
2340 /* Assume notifications from guest are disabled at this point,
2341 * if they aren't we would need to update avail_event index. */
2342 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2345 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2347 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2348 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2350 vq->last_avail_idx -= n;
2352 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2354 /* After we've used one of their buffers, we tell them about it. We'll then
2355 * want to notify the guest, using eventfd. */
2356 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2358 struct vring_used_elem heads = {
2359 cpu_to_vhost32(vq, head),
2360 cpu_to_vhost32(vq, len)
2363 return vhost_add_used_n(vq, &heads, 1);
2365 EXPORT_SYMBOL_GPL(vhost_add_used);
2367 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2368 struct vring_used_elem *heads,
2371 vring_used_elem_t __user *used;
2375 start = vq->last_used_idx & (vq->num - 1);
2376 used = vq->used->ring + start;
2377 if (vhost_put_used(vq, heads, start, count)) {
2378 vq_err(vq, "Failed to write used");
2381 if (unlikely(vq->log_used)) {
2382 /* Make sure data is seen before log. */
2384 /* Log used ring entry write. */
2385 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2386 count * sizeof *used);
2388 old = vq->last_used_idx;
2389 new = (vq->last_used_idx += count);
2390 /* If the driver never bothers to signal in a very long while,
2391 * used index might wrap around. If that happens, invalidate
2392 * signalled_used index we stored. TODO: make sure driver
2393 * signals at least once in 2^16 and remove this. */
2394 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2395 vq->signalled_used_valid = false;
2399 /* After we've used one of their buffers, we tell them about it. We'll then
2400 * want to notify the guest, using eventfd. */
2401 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2406 start = vq->last_used_idx & (vq->num - 1);
2407 n = vq->num - start;
2409 r = __vhost_add_used_n(vq, heads, n);
2415 r = __vhost_add_used_n(vq, heads, count);
2417 /* Make sure buffer is written before we update index. */
2419 if (vhost_put_used_idx(vq)) {
2420 vq_err(vq, "Failed to increment used idx");
2423 if (unlikely(vq->log_used)) {
2424 /* Make sure used idx is seen before log. */
2426 /* Log used index update. */
2427 log_used(vq, offsetof(struct vring_used, idx),
2428 sizeof vq->used->idx);
2430 eventfd_signal(vq->log_ctx, 1);
2434 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2436 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2441 /* Flush out used index updates. This is paired
2442 * with the barrier that the Guest executes when enabling
2446 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2447 unlikely(vq->avail_idx == vq->last_avail_idx))
2450 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2452 if (vhost_get_avail_flags(vq, &flags)) {
2453 vq_err(vq, "Failed to get flags");
2456 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2458 old = vq->signalled_used;
2459 v = vq->signalled_used_valid;
2460 new = vq->signalled_used = vq->last_used_idx;
2461 vq->signalled_used_valid = true;
2466 if (vhost_get_used_event(vq, &event)) {
2467 vq_err(vq, "Failed to get used event idx");
2470 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2473 /* This actually signals the guest, using eventfd. */
2474 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2476 /* Signal the Guest tell them we used something up. */
2477 if (vq->call_ctx.ctx && vhost_notify(dev, vq))
2478 eventfd_signal(vq->call_ctx.ctx, 1);
2480 EXPORT_SYMBOL_GPL(vhost_signal);
2482 /* And here's the combo meal deal. Supersize me! */
2483 void vhost_add_used_and_signal(struct vhost_dev *dev,
2484 struct vhost_virtqueue *vq,
2485 unsigned int head, int len)
2487 vhost_add_used(vq, head, len);
2488 vhost_signal(dev, vq);
2490 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2492 /* multi-buffer version of vhost_add_used_and_signal */
2493 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2494 struct vhost_virtqueue *vq,
2495 struct vring_used_elem *heads, unsigned count)
2497 vhost_add_used_n(vq, heads, count);
2498 vhost_signal(dev, vq);
2500 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2502 /* return true if we're sure that avaiable ring is empty */
2503 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2505 __virtio16 avail_idx;
2508 if (vq->avail_idx != vq->last_avail_idx)
2511 r = vhost_get_avail_idx(vq, &avail_idx);
2514 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2516 return vq->avail_idx == vq->last_avail_idx;
2518 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2520 /* OK, now we need to know about added descriptors. */
2521 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2523 __virtio16 avail_idx;
2526 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2528 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2529 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2530 r = vhost_update_used_flags(vq);
2532 vq_err(vq, "Failed to enable notification at %p: %d\n",
2533 &vq->used->flags, r);
2537 r = vhost_update_avail_event(vq);
2539 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2540 vhost_avail_event(vq), r);
2544 /* They could have slipped one in as we were doing that: make
2545 * sure it's written, then check again. */
2547 r = vhost_get_avail_idx(vq, &avail_idx);
2549 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2550 &vq->avail->idx, r);
2553 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2555 return vq->avail_idx != vq->last_avail_idx;
2557 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2559 /* We don't need to be notified again. */
2560 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2564 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2566 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2567 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2568 r = vhost_update_used_flags(vq);
2570 vq_err(vq, "Failed to disable notification at %p: %d\n",
2571 &vq->used->flags, r);
2574 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2576 /* Create a new message. */
2577 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2579 struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2583 /* Make sure all padding within the structure is initialized. */
2584 memset(&node->msg, 0, sizeof node->msg);
2586 node->msg.type = type;
2589 EXPORT_SYMBOL_GPL(vhost_new_msg);
2591 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2592 struct vhost_msg_node *node)
2594 spin_lock(&dev->iotlb_lock);
2595 list_add_tail(&node->node, head);
2596 spin_unlock(&dev->iotlb_lock);
2598 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2600 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2602 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2603 struct list_head *head)
2605 struct vhost_msg_node *node = NULL;
2607 spin_lock(&dev->iotlb_lock);
2608 if (!list_empty(head)) {
2609 node = list_first_entry(head, struct vhost_msg_node,
2611 list_del(&node->node);
2613 spin_unlock(&dev->iotlb_lock);
2617 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2619 void vhost_set_backend_features(struct vhost_dev *dev, u64 features)
2621 struct vhost_virtqueue *vq;
2624 mutex_lock(&dev->mutex);
2625 for (i = 0; i < dev->nvqs; ++i) {
2627 mutex_lock(&vq->mutex);
2628 vq->acked_backend_features = features;
2629 mutex_unlock(&vq->mutex);
2631 mutex_unlock(&dev->mutex);
2633 EXPORT_SYMBOL_GPL(vhost_set_backend_features);
2635 static int __init vhost_init(void)
2640 static void __exit vhost_exit(void)
2644 module_init(vhost_init);
2645 module_exit(vhost_exit);
2647 MODULE_VERSION("0.0.1");
2648 MODULE_LICENSE("GPL v2");
2649 MODULE_AUTHOR("Michael S. Tsirkin");
2650 MODULE_DESCRIPTION("Host kernel accelerator for virtio");