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;
238 if (dev->worker.vtsk) {
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)
250 if (!dev->worker.vtsk)
253 if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
254 /* We can only add the work to the list after we're
255 * sure it was not in the list.
256 * test_and_set_bit() implies a memory barrier.
258 llist_add(&work->node, &dev->worker.work_list);
259 vhost_task_wake(dev->worker.vtsk);
262 EXPORT_SYMBOL_GPL(vhost_work_queue);
264 /* A lockless hint for busy polling code to exit the loop */
265 bool vhost_has_work(struct vhost_dev *dev)
267 return !llist_empty(&dev->worker.work_list);
269 EXPORT_SYMBOL_GPL(vhost_has_work);
271 void vhost_poll_queue(struct vhost_poll *poll)
273 vhost_work_queue(poll->dev, &poll->work);
275 EXPORT_SYMBOL_GPL(vhost_poll_queue);
277 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
281 for (j = 0; j < VHOST_NUM_ADDRS; j++)
282 vq->meta_iotlb[j] = NULL;
285 static void vhost_vq_meta_reset(struct vhost_dev *d)
289 for (i = 0; i < d->nvqs; ++i)
290 __vhost_vq_meta_reset(d->vqs[i]);
293 static void vhost_vring_call_reset(struct vhost_vring_call *call_ctx)
295 call_ctx->ctx = NULL;
296 memset(&call_ctx->producer, 0x0, sizeof(struct irq_bypass_producer));
299 bool vhost_vq_is_setup(struct vhost_virtqueue *vq)
301 return vq->avail && vq->desc && vq->used && vhost_vq_access_ok(vq);
303 EXPORT_SYMBOL_GPL(vhost_vq_is_setup);
305 static void vhost_vq_reset(struct vhost_dev *dev,
306 struct vhost_virtqueue *vq)
312 vq->last_avail_idx = 0;
314 vq->last_used_idx = 0;
315 vq->signalled_used = 0;
316 vq->signalled_used_valid = false;
318 vq->log_used = false;
319 vq->log_addr = -1ull;
320 vq->private_data = NULL;
321 vq->acked_features = 0;
322 vq->acked_backend_features = 0;
324 vq->error_ctx = NULL;
327 vhost_disable_cross_endian(vq);
328 vhost_reset_is_le(vq);
329 vq->busyloop_timeout = 0;
332 vhost_vring_call_reset(&vq->call_ctx);
333 __vhost_vq_meta_reset(vq);
336 static bool vhost_worker(void *data)
338 struct vhost_worker *worker = data;
339 struct vhost_work *work, *work_next;
340 struct llist_node *node;
342 node = llist_del_all(&worker->work_list);
344 __set_current_state(TASK_RUNNING);
346 node = llist_reverse_order(node);
347 /* make sure flag is seen after deletion */
349 llist_for_each_entry_safe(work, work_next, node, node) {
350 clear_bit(VHOST_WORK_QUEUED, &work->flags);
351 kcov_remote_start_common(worker->kcov_handle);
361 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
371 /* Helper to allocate iovec buffers for all vqs. */
372 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
374 struct vhost_virtqueue *vq;
377 for (i = 0; i < dev->nvqs; ++i) {
379 vq->indirect = kmalloc_array(UIO_MAXIOV,
380 sizeof(*vq->indirect),
382 vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
384 vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
386 if (!vq->indirect || !vq->log || !vq->heads)
393 vhost_vq_free_iovecs(dev->vqs[i]);
397 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
401 for (i = 0; i < dev->nvqs; ++i)
402 vhost_vq_free_iovecs(dev->vqs[i]);
405 bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
406 int pkts, int total_len)
408 struct vhost_dev *dev = vq->dev;
410 if ((dev->byte_weight && total_len >= dev->byte_weight) ||
411 pkts >= dev->weight) {
412 vhost_poll_queue(&vq->poll);
418 EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
420 static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
423 size_t event __maybe_unused =
424 vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
426 return size_add(struct_size(vq->avail, ring, num), event);
429 static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
432 size_t event __maybe_unused =
433 vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
435 return size_add(struct_size(vq->used, ring, num), event);
438 static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
441 return sizeof(*vq->desc) * num;
444 void vhost_dev_init(struct vhost_dev *dev,
445 struct vhost_virtqueue **vqs, int nvqs,
446 int iov_limit, int weight, int byte_weight,
448 int (*msg_handler)(struct vhost_dev *dev, u32 asid,
449 struct vhost_iotlb_msg *msg))
451 struct vhost_virtqueue *vq;
456 mutex_init(&dev->mutex);
461 memset(&dev->worker, 0, sizeof(dev->worker));
462 init_llist_head(&dev->worker.work_list);
463 dev->iov_limit = iov_limit;
464 dev->weight = weight;
465 dev->byte_weight = byte_weight;
466 dev->use_worker = use_worker;
467 dev->msg_handler = msg_handler;
468 init_waitqueue_head(&dev->wait);
469 INIT_LIST_HEAD(&dev->read_list);
470 INIT_LIST_HEAD(&dev->pending_list);
471 spin_lock_init(&dev->iotlb_lock);
474 for (i = 0; i < dev->nvqs; ++i) {
480 mutex_init(&vq->mutex);
481 vhost_vq_reset(dev, vq);
483 vhost_poll_init(&vq->poll, vq->handle_kick,
487 EXPORT_SYMBOL_GPL(vhost_dev_init);
489 /* Caller should have device mutex */
490 long vhost_dev_check_owner(struct vhost_dev *dev)
492 /* Are you the owner? If not, I don't think you mean to do that */
493 return dev->mm == current->mm ? 0 : -EPERM;
495 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
497 /* Caller should have device mutex */
498 bool vhost_dev_has_owner(struct vhost_dev *dev)
502 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
504 static void vhost_attach_mm(struct vhost_dev *dev)
506 /* No owner, become one */
507 if (dev->use_worker) {
508 dev->mm = get_task_mm(current);
510 /* vDPA device does not use worker thead, so there's
511 * no need to hold the address space for mm. This help
512 * to avoid deadlock in the case of mmap() which may
513 * held the refcnt of the file and depends on release
514 * method to remove vma.
516 dev->mm = current->mm;
521 static void vhost_detach_mm(struct vhost_dev *dev)
534 static void vhost_worker_free(struct vhost_dev *dev)
536 if (!dev->worker.vtsk)
539 WARN_ON(!llist_empty(&dev->worker.work_list));
540 vhost_task_stop(dev->worker.vtsk);
541 dev->worker.kcov_handle = 0;
542 dev->worker.vtsk = NULL;
545 static int vhost_worker_create(struct vhost_dev *dev)
547 struct vhost_task *vtsk;
548 char name[TASK_COMM_LEN];
550 snprintf(name, sizeof(name), "vhost-%d", current->pid);
552 vtsk = vhost_task_create(vhost_worker, &dev->worker, name);
556 dev->worker.kcov_handle = kcov_common_handle();
557 dev->worker.vtsk = vtsk;
558 vhost_task_start(vtsk);
562 /* Caller should have device mutex */
563 long vhost_dev_set_owner(struct vhost_dev *dev)
567 /* Is there an owner already? */
568 if (vhost_dev_has_owner(dev)) {
573 vhost_attach_mm(dev);
575 if (dev->use_worker) {
576 err = vhost_worker_create(dev);
581 err = vhost_dev_alloc_iovecs(dev);
587 vhost_worker_free(dev);
589 vhost_detach_mm(dev);
593 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
595 static struct vhost_iotlb *iotlb_alloc(void)
597 return vhost_iotlb_alloc(max_iotlb_entries,
598 VHOST_IOTLB_FLAG_RETIRE);
601 struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
603 return iotlb_alloc();
605 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
607 /* Caller should have device mutex */
608 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
612 vhost_dev_cleanup(dev);
615 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
616 * VQs aren't running.
618 for (i = 0; i < dev->nvqs; ++i)
619 dev->vqs[i]->umem = umem;
621 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
623 void vhost_dev_stop(struct vhost_dev *dev)
627 for (i = 0; i < dev->nvqs; ++i) {
628 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick)
629 vhost_poll_stop(&dev->vqs[i]->poll);
632 vhost_dev_flush(dev);
634 EXPORT_SYMBOL_GPL(vhost_dev_stop);
636 void vhost_clear_msg(struct vhost_dev *dev)
638 struct vhost_msg_node *node, *n;
640 spin_lock(&dev->iotlb_lock);
642 list_for_each_entry_safe(node, n, &dev->read_list, node) {
643 list_del(&node->node);
647 list_for_each_entry_safe(node, n, &dev->pending_list, node) {
648 list_del(&node->node);
652 spin_unlock(&dev->iotlb_lock);
654 EXPORT_SYMBOL_GPL(vhost_clear_msg);
656 void vhost_dev_cleanup(struct vhost_dev *dev)
660 for (i = 0; i < dev->nvqs; ++i) {
661 if (dev->vqs[i]->error_ctx)
662 eventfd_ctx_put(dev->vqs[i]->error_ctx);
663 if (dev->vqs[i]->kick)
664 fput(dev->vqs[i]->kick);
665 if (dev->vqs[i]->call_ctx.ctx)
666 eventfd_ctx_put(dev->vqs[i]->call_ctx.ctx);
667 vhost_vq_reset(dev, dev->vqs[i]);
669 vhost_dev_free_iovecs(dev);
671 eventfd_ctx_put(dev->log_ctx);
673 /* No one will access memory at this point */
674 vhost_iotlb_free(dev->umem);
676 vhost_iotlb_free(dev->iotlb);
678 vhost_clear_msg(dev);
679 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
680 vhost_worker_free(dev);
681 vhost_detach_mm(dev);
683 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
685 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
687 u64 a = addr / VHOST_PAGE_SIZE / 8;
689 /* Make sure 64 bit math will not overflow. */
690 if (a > ULONG_MAX - (unsigned long)log_base ||
691 a + (unsigned long)log_base > ULONG_MAX)
694 return access_ok(log_base + a,
695 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
698 /* Make sure 64 bit math will not overflow. */
699 static bool vhost_overflow(u64 uaddr, u64 size)
701 if (uaddr > ULONG_MAX || size > ULONG_MAX)
707 return uaddr > ULONG_MAX - size + 1;
710 /* Caller should have vq mutex and device mutex. */
711 static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
714 struct vhost_iotlb_map *map;
719 list_for_each_entry(map, &umem->list, link) {
720 unsigned long a = map->addr;
722 if (vhost_overflow(map->addr, map->size))
726 if (!access_ok((void __user *)a, map->size))
728 else if (log_all && !log_access_ok(log_base,
736 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
737 u64 addr, unsigned int size,
740 const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
745 return (void __user *)(uintptr_t)(map->addr + addr - map->start);
748 /* Can we switch to this memory table? */
749 /* Caller should have device mutex but not vq mutex */
750 static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
755 for (i = 0; i < d->nvqs; ++i) {
759 mutex_lock(&d->vqs[i]->mutex);
760 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
761 /* If ring is inactive, will check when it's enabled. */
762 if (d->vqs[i]->private_data)
763 ok = vq_memory_access_ok(d->vqs[i]->log_base,
767 mutex_unlock(&d->vqs[i]->mutex);
774 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
775 struct iovec iov[], int iov_size, int access);
777 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
778 const void *from, unsigned size)
783 return __copy_to_user(to, from, size);
785 /* This function should be called after iotlb
786 * prefetch, which means we're sure that all vq
787 * could be access through iotlb. So -EAGAIN should
788 * not happen in this case.
791 void __user *uaddr = vhost_vq_meta_fetch(vq,
792 (u64)(uintptr_t)to, size,
796 return __copy_to_user(uaddr, from, size);
798 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
799 ARRAY_SIZE(vq->iotlb_iov),
803 iov_iter_init(&t, ITER_DEST, vq->iotlb_iov, ret, size);
804 ret = copy_to_iter(from, size, &t);
812 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
813 void __user *from, unsigned size)
818 return __copy_from_user(to, from, size);
820 /* This function should be called after iotlb
821 * prefetch, which means we're sure that vq
822 * could be access through iotlb. So -EAGAIN should
823 * not happen in this case.
825 void __user *uaddr = vhost_vq_meta_fetch(vq,
826 (u64)(uintptr_t)from, size,
831 return __copy_from_user(to, uaddr, size);
833 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
834 ARRAY_SIZE(vq->iotlb_iov),
837 vq_err(vq, "IOTLB translation failure: uaddr "
838 "%p size 0x%llx\n", from,
839 (unsigned long long) size);
842 iov_iter_init(&f, ITER_SOURCE, vq->iotlb_iov, ret, size);
843 ret = copy_from_iter(to, size, &f);
852 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
853 void __user *addr, unsigned int size,
858 ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
859 ARRAY_SIZE(vq->iotlb_iov),
862 vq_err(vq, "IOTLB translation failure: uaddr "
863 "%p size 0x%llx\n", addr,
864 (unsigned long long) size);
868 if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
869 vq_err(vq, "Non atomic userspace memory access: uaddr "
870 "%p size 0x%llx\n", addr,
871 (unsigned long long) size);
875 return vq->iotlb_iov[0].iov_base;
878 /* This function should be called after iotlb
879 * prefetch, which means we're sure that vq
880 * could be access through iotlb. So -EAGAIN should
881 * not happen in this case.
883 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
884 void __user *addr, unsigned int size,
887 void __user *uaddr = vhost_vq_meta_fetch(vq,
888 (u64)(uintptr_t)addr, size, type);
892 return __vhost_get_user_slow(vq, addr, size, type);
895 #define vhost_put_user(vq, x, ptr) \
899 ret = __put_user(x, ptr); \
901 __typeof__(ptr) to = \
902 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
903 sizeof(*ptr), VHOST_ADDR_USED); \
905 ret = __put_user(x, to); \
912 static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
914 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
915 vhost_avail_event(vq));
918 static inline int vhost_put_used(struct vhost_virtqueue *vq,
919 struct vring_used_elem *head, int idx,
922 return vhost_copy_to_user(vq, vq->used->ring + idx, head,
923 count * sizeof(*head));
926 static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
929 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
933 static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
936 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
940 #define vhost_get_user(vq, x, ptr, type) \
944 ret = __get_user(x, ptr); \
946 __typeof__(ptr) from = \
947 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
951 ret = __get_user(x, from); \
958 #define vhost_get_avail(vq, x, ptr) \
959 vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
961 #define vhost_get_used(vq, x, ptr) \
962 vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
964 static void vhost_dev_lock_vqs(struct vhost_dev *d)
967 for (i = 0; i < d->nvqs; ++i)
968 mutex_lock_nested(&d->vqs[i]->mutex, i);
971 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
974 for (i = 0; i < d->nvqs; ++i)
975 mutex_unlock(&d->vqs[i]->mutex);
978 static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
981 return vhost_get_avail(vq, *idx, &vq->avail->idx);
984 static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
985 __virtio16 *head, int idx)
987 return vhost_get_avail(vq, *head,
988 &vq->avail->ring[idx & (vq->num - 1)]);
991 static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
994 return vhost_get_avail(vq, *flags, &vq->avail->flags);
997 static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1000 return vhost_get_avail(vq, *event, vhost_used_event(vq));
1003 static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1006 return vhost_get_used(vq, *idx, &vq->used->idx);
1009 static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1010 struct vring_desc *desc, int idx)
1012 return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1015 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1016 struct vhost_iotlb_msg *msg)
1018 struct vhost_msg_node *node, *n;
1020 spin_lock(&d->iotlb_lock);
1022 list_for_each_entry_safe(node, n, &d->pending_list, node) {
1023 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1024 if (msg->iova <= vq_msg->iova &&
1025 msg->iova + msg->size - 1 >= vq_msg->iova &&
1026 vq_msg->type == VHOST_IOTLB_MISS) {
1027 vhost_poll_queue(&node->vq->poll);
1028 list_del(&node->node);
1033 spin_unlock(&d->iotlb_lock);
1036 static bool umem_access_ok(u64 uaddr, u64 size, int access)
1038 unsigned long a = uaddr;
1040 /* Make sure 64 bit math will not overflow. */
1041 if (vhost_overflow(uaddr, size))
1044 if ((access & VHOST_ACCESS_RO) &&
1045 !access_ok((void __user *)a, size))
1047 if ((access & VHOST_ACCESS_WO) &&
1048 !access_ok((void __user *)a, size))
1053 static int vhost_process_iotlb_msg(struct vhost_dev *dev, u32 asid,
1054 struct vhost_iotlb_msg *msg)
1061 mutex_lock(&dev->mutex);
1062 vhost_dev_lock_vqs(dev);
1063 switch (msg->type) {
1064 case VHOST_IOTLB_UPDATE:
1069 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1073 vhost_vq_meta_reset(dev);
1074 if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1075 msg->iova + msg->size - 1,
1076 msg->uaddr, msg->perm)) {
1080 vhost_iotlb_notify_vq(dev, msg);
1082 case VHOST_IOTLB_INVALIDATE:
1087 vhost_vq_meta_reset(dev);
1088 vhost_iotlb_del_range(dev->iotlb, msg->iova,
1089 msg->iova + msg->size - 1);
1096 vhost_dev_unlock_vqs(dev);
1097 mutex_unlock(&dev->mutex);
1101 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1102 struct iov_iter *from)
1104 struct vhost_iotlb_msg msg;
1109 ret = copy_from_iter(&type, sizeof(type), from);
1110 if (ret != sizeof(type)) {
1116 case VHOST_IOTLB_MSG:
1117 /* There maybe a hole after type for V1 message type,
1120 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1122 case VHOST_IOTLB_MSG_V2:
1123 if (vhost_backend_has_feature(dev->vqs[0],
1124 VHOST_BACKEND_F_IOTLB_ASID)) {
1125 ret = copy_from_iter(&asid, sizeof(asid), from);
1126 if (ret != sizeof(asid)) {
1132 offset = sizeof(__u32);
1139 iov_iter_advance(from, offset);
1140 ret = copy_from_iter(&msg, sizeof(msg), from);
1141 if (ret != sizeof(msg)) {
1146 if ((msg.type == VHOST_IOTLB_UPDATE ||
1147 msg.type == VHOST_IOTLB_INVALIDATE) &&
1153 if (dev->msg_handler)
1154 ret = dev->msg_handler(dev, asid, &msg);
1156 ret = vhost_process_iotlb_msg(dev, asid, &msg);
1162 ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1163 sizeof(struct vhost_msg_v2);
1167 EXPORT_SYMBOL(vhost_chr_write_iter);
1169 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1174 poll_wait(file, &dev->wait, wait);
1176 if (!list_empty(&dev->read_list))
1177 mask |= EPOLLIN | EPOLLRDNORM;
1181 EXPORT_SYMBOL(vhost_chr_poll);
1183 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1187 struct vhost_msg_node *node;
1189 unsigned size = sizeof(struct vhost_msg);
1191 if (iov_iter_count(to) < size)
1196 prepare_to_wait(&dev->wait, &wait,
1197 TASK_INTERRUPTIBLE);
1199 node = vhost_dequeue_msg(dev, &dev->read_list);
1206 if (signal_pending(current)) {
1219 finish_wait(&dev->wait, &wait);
1222 struct vhost_iotlb_msg *msg;
1223 void *start = &node->msg;
1225 switch (node->msg.type) {
1226 case VHOST_IOTLB_MSG:
1227 size = sizeof(node->msg);
1228 msg = &node->msg.iotlb;
1230 case VHOST_IOTLB_MSG_V2:
1231 size = sizeof(node->msg_v2);
1232 msg = &node->msg_v2.iotlb;
1239 ret = copy_to_iter(start, size, to);
1240 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1244 vhost_enqueue_msg(dev, &dev->pending_list, node);
1249 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1251 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1253 struct vhost_dev *dev = vq->dev;
1254 struct vhost_msg_node *node;
1255 struct vhost_iotlb_msg *msg;
1256 bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1258 node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1263 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1264 msg = &node->msg_v2.iotlb;
1266 msg = &node->msg.iotlb;
1269 msg->type = VHOST_IOTLB_MISS;
1273 vhost_enqueue_msg(dev, &dev->read_list, node);
1278 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1279 vring_desc_t __user *desc,
1280 vring_avail_t __user *avail,
1281 vring_used_t __user *used)
1284 /* If an IOTLB device is present, the vring addresses are
1285 * GIOVAs. Access validation occurs at prefetch time. */
1289 return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1290 access_ok(avail, vhost_get_avail_size(vq, num)) &&
1291 access_ok(used, vhost_get_used_size(vq, num));
1294 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1295 const struct vhost_iotlb_map *map,
1298 int access = (type == VHOST_ADDR_USED) ?
1299 VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1301 if (likely(map->perm & access))
1302 vq->meta_iotlb[type] = map;
1305 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1306 int access, u64 addr, u64 len, int type)
1308 const struct vhost_iotlb_map *map;
1309 struct vhost_iotlb *umem = vq->iotlb;
1310 u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1312 if (vhost_vq_meta_fetch(vq, addr, len, type))
1316 map = vhost_iotlb_itree_first(umem, addr, last);
1317 if (map == NULL || map->start > addr) {
1318 vhost_iotlb_miss(vq, addr, access);
1320 } else if (!(map->perm & access)) {
1321 /* Report the possible access violation by
1322 * request another translation from userspace.
1327 size = map->size - addr + map->start;
1329 if (orig_addr == addr && size >= len)
1330 vhost_vq_meta_update(vq, map, type);
1339 int vq_meta_prefetch(struct vhost_virtqueue *vq)
1341 unsigned int num = vq->num;
1346 return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1347 vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1348 iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1349 vhost_get_avail_size(vq, num),
1350 VHOST_ADDR_AVAIL) &&
1351 iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1352 vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1354 EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1356 /* Can we log writes? */
1357 /* Caller should have device mutex but not vq mutex */
1358 bool vhost_log_access_ok(struct vhost_dev *dev)
1360 return memory_access_ok(dev, dev->umem, 1);
1362 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1364 static bool vq_log_used_access_ok(struct vhost_virtqueue *vq,
1365 void __user *log_base,
1369 /* If an IOTLB device is present, log_addr is a GIOVA that
1370 * will never be logged by log_used(). */
1374 return !log_used || log_access_ok(log_base, log_addr,
1375 vhost_get_used_size(vq, vq->num));
1378 /* Verify access for write logging. */
1379 /* Caller should have vq mutex and device mutex */
1380 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1381 void __user *log_base)
1383 return vq_memory_access_ok(log_base, vq->umem,
1384 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1385 vq_log_used_access_ok(vq, log_base, vq->log_used, vq->log_addr);
1388 /* Can we start vq? */
1389 /* Caller should have vq mutex and device mutex */
1390 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1392 if (!vq_log_access_ok(vq, vq->log_base))
1395 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1397 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1399 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1401 struct vhost_memory mem, *newmem;
1402 struct vhost_memory_region *region;
1403 struct vhost_iotlb *newumem, *oldumem;
1404 unsigned long size = offsetof(struct vhost_memory, regions);
1407 if (copy_from_user(&mem, m, size))
1411 if (mem.nregions > max_mem_regions)
1413 newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1418 memcpy(newmem, &mem, size);
1419 if (copy_from_user(newmem->regions, m->regions,
1420 flex_array_size(newmem, regions, mem.nregions))) {
1425 newumem = iotlb_alloc();
1431 for (region = newmem->regions;
1432 region < newmem->regions + mem.nregions;
1434 if (vhost_iotlb_add_range(newumem,
1435 region->guest_phys_addr,
1436 region->guest_phys_addr +
1437 region->memory_size - 1,
1438 region->userspace_addr,
1443 if (!memory_access_ok(d, newumem, 0))
1449 /* All memory accesses are done under some VQ mutex. */
1450 for (i = 0; i < d->nvqs; ++i) {
1451 mutex_lock(&d->vqs[i]->mutex);
1452 d->vqs[i]->umem = newumem;
1453 mutex_unlock(&d->vqs[i]->mutex);
1457 vhost_iotlb_free(oldumem);
1461 vhost_iotlb_free(newumem);
1466 static long vhost_vring_set_num(struct vhost_dev *d,
1467 struct vhost_virtqueue *vq,
1470 struct vhost_vring_state s;
1472 /* Resizing ring with an active backend?
1473 * You don't want to do that. */
1474 if (vq->private_data)
1477 if (copy_from_user(&s, argp, sizeof s))
1480 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1487 static long vhost_vring_set_addr(struct vhost_dev *d,
1488 struct vhost_virtqueue *vq,
1491 struct vhost_vring_addr a;
1493 if (copy_from_user(&a, argp, sizeof a))
1495 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1498 /* For 32bit, verify that the top 32bits of the user
1499 data are set to zero. */
1500 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1501 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1502 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1505 /* Make sure it's safe to cast pointers to vring types. */
1506 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1507 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1508 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1509 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1510 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1513 /* We only verify access here if backend is configured.
1514 * If it is not, we don't as size might not have been setup.
1515 * We will verify when backend is configured. */
1516 if (vq->private_data) {
1517 if (!vq_access_ok(vq, vq->num,
1518 (void __user *)(unsigned long)a.desc_user_addr,
1519 (void __user *)(unsigned long)a.avail_user_addr,
1520 (void __user *)(unsigned long)a.used_user_addr))
1523 /* Also validate log access for used ring if enabled. */
1524 if (!vq_log_used_access_ok(vq, vq->log_base,
1525 a.flags & (0x1 << VHOST_VRING_F_LOG),
1530 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1531 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1532 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1533 vq->log_addr = a.log_guest_addr;
1534 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1539 static long vhost_vring_set_num_addr(struct vhost_dev *d,
1540 struct vhost_virtqueue *vq,
1546 mutex_lock(&vq->mutex);
1549 case VHOST_SET_VRING_NUM:
1550 r = vhost_vring_set_num(d, vq, argp);
1552 case VHOST_SET_VRING_ADDR:
1553 r = vhost_vring_set_addr(d, vq, argp);
1559 mutex_unlock(&vq->mutex);
1563 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1565 struct file *eventfp, *filep = NULL;
1566 bool pollstart = false, pollstop = false;
1567 struct eventfd_ctx *ctx = NULL;
1568 u32 __user *idxp = argp;
1569 struct vhost_virtqueue *vq;
1570 struct vhost_vring_state s;
1571 struct vhost_vring_file f;
1575 r = get_user(idx, idxp);
1581 idx = array_index_nospec(idx, d->nvqs);
1584 if (ioctl == VHOST_SET_VRING_NUM ||
1585 ioctl == VHOST_SET_VRING_ADDR) {
1586 return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1589 mutex_lock(&vq->mutex);
1592 case VHOST_SET_VRING_BASE:
1593 /* Moving base with an active backend?
1594 * You don't want to do that. */
1595 if (vq->private_data) {
1599 if (copy_from_user(&s, argp, sizeof s)) {
1603 if (vhost_has_feature(vq, VIRTIO_F_RING_PACKED)) {
1604 vq->last_avail_idx = s.num & 0xffff;
1605 vq->last_used_idx = (s.num >> 16) & 0xffff;
1607 if (s.num > 0xffff) {
1611 vq->last_avail_idx = s.num;
1613 /* Forget the cached index value. */
1614 vq->avail_idx = vq->last_avail_idx;
1616 case VHOST_GET_VRING_BASE:
1618 if (vhost_has_feature(vq, VIRTIO_F_RING_PACKED))
1619 s.num = (u32)vq->last_avail_idx | ((u32)vq->last_used_idx << 16);
1621 s.num = vq->last_avail_idx;
1622 if (copy_to_user(argp, &s, sizeof s))
1625 case VHOST_SET_VRING_KICK:
1626 if (copy_from_user(&f, argp, sizeof f)) {
1630 eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
1631 if (IS_ERR(eventfp)) {
1632 r = PTR_ERR(eventfp);
1635 if (eventfp != vq->kick) {
1636 pollstop = (filep = vq->kick) != NULL;
1637 pollstart = (vq->kick = eventfp) != NULL;
1641 case VHOST_SET_VRING_CALL:
1642 if (copy_from_user(&f, argp, sizeof f)) {
1646 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1652 swap(ctx, vq->call_ctx.ctx);
1654 case VHOST_SET_VRING_ERR:
1655 if (copy_from_user(&f, argp, sizeof f)) {
1659 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1664 swap(ctx, vq->error_ctx);
1666 case VHOST_SET_VRING_ENDIAN:
1667 r = vhost_set_vring_endian(vq, argp);
1669 case VHOST_GET_VRING_ENDIAN:
1670 r = vhost_get_vring_endian(vq, idx, argp);
1672 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1673 if (copy_from_user(&s, argp, sizeof(s))) {
1677 vq->busyloop_timeout = s.num;
1679 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1681 s.num = vq->busyloop_timeout;
1682 if (copy_to_user(argp, &s, sizeof(s)))
1689 if (pollstop && vq->handle_kick)
1690 vhost_poll_stop(&vq->poll);
1692 if (!IS_ERR_OR_NULL(ctx))
1693 eventfd_ctx_put(ctx);
1697 if (pollstart && vq->handle_kick)
1698 r = vhost_poll_start(&vq->poll, vq->kick);
1700 mutex_unlock(&vq->mutex);
1702 if (pollstop && vq->handle_kick)
1703 vhost_dev_flush(vq->poll.dev);
1706 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1708 int vhost_init_device_iotlb(struct vhost_dev *d)
1710 struct vhost_iotlb *niotlb, *oiotlb;
1713 niotlb = iotlb_alloc();
1720 for (i = 0; i < d->nvqs; ++i) {
1721 struct vhost_virtqueue *vq = d->vqs[i];
1723 mutex_lock(&vq->mutex);
1725 __vhost_vq_meta_reset(vq);
1726 mutex_unlock(&vq->mutex);
1729 vhost_iotlb_free(oiotlb);
1733 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1735 /* Caller must have device mutex */
1736 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1738 struct eventfd_ctx *ctx;
1743 /* If you are not the owner, you can become one */
1744 if (ioctl == VHOST_SET_OWNER) {
1745 r = vhost_dev_set_owner(d);
1749 /* You must be the owner to do anything else */
1750 r = vhost_dev_check_owner(d);
1755 case VHOST_SET_MEM_TABLE:
1756 r = vhost_set_memory(d, argp);
1758 case VHOST_SET_LOG_BASE:
1759 if (copy_from_user(&p, argp, sizeof p)) {
1763 if ((u64)(unsigned long)p != p) {
1767 for (i = 0; i < d->nvqs; ++i) {
1768 struct vhost_virtqueue *vq;
1769 void __user *base = (void __user *)(unsigned long)p;
1771 mutex_lock(&vq->mutex);
1772 /* If ring is inactive, will check when it's enabled. */
1773 if (vq->private_data && !vq_log_access_ok(vq, base))
1776 vq->log_base = base;
1777 mutex_unlock(&vq->mutex);
1780 case VHOST_SET_LOG_FD:
1781 r = get_user(fd, (int __user *)argp);
1784 ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
1789 swap(ctx, d->log_ctx);
1790 for (i = 0; i < d->nvqs; ++i) {
1791 mutex_lock(&d->vqs[i]->mutex);
1792 d->vqs[i]->log_ctx = d->log_ctx;
1793 mutex_unlock(&d->vqs[i]->mutex);
1796 eventfd_ctx_put(ctx);
1805 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1807 /* TODO: This is really inefficient. We need something like get_user()
1808 * (instruction directly accesses the data, with an exception table entry
1809 * returning -EFAULT). See Documentation/arch/x86/exception-tables.rst.
1811 static int set_bit_to_user(int nr, void __user *addr)
1813 unsigned long log = (unsigned long)addr;
1816 int bit = nr + (log % PAGE_SIZE) * 8;
1819 r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
1823 base = kmap_atomic(page);
1825 kunmap_atomic(base);
1826 unpin_user_pages_dirty_lock(&page, 1, true);
1830 static int log_write(void __user *log_base,
1831 u64 write_address, u64 write_length)
1833 u64 write_page = write_address / VHOST_PAGE_SIZE;
1838 write_length += write_address % VHOST_PAGE_SIZE;
1840 u64 base = (u64)(unsigned long)log_base;
1841 u64 log = base + write_page / 8;
1842 int bit = write_page % 8;
1843 if ((u64)(unsigned long)log != log)
1845 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1848 if (write_length <= VHOST_PAGE_SIZE)
1850 write_length -= VHOST_PAGE_SIZE;
1856 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1858 struct vhost_iotlb *umem = vq->umem;
1859 struct vhost_iotlb_map *u;
1860 u64 start, end, l, min;
1866 /* More than one GPAs can be mapped into a single HVA. So
1867 * iterate all possible umems here to be safe.
1869 list_for_each_entry(u, &umem->list, link) {
1870 if (u->addr > hva - 1 + len ||
1871 u->addr - 1 + u->size < hva)
1873 start = max(u->addr, hva);
1874 end = min(u->addr - 1 + u->size, hva - 1 + len);
1875 l = end - start + 1;
1876 r = log_write(vq->log_base,
1877 u->start + start - u->addr,
1895 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1897 struct iovec *iov = vq->log_iov;
1901 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1903 ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1904 len, iov, 64, VHOST_ACCESS_WO);
1908 for (i = 0; i < ret; i++) {
1909 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1918 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1919 unsigned int log_num, u64 len, struct iovec *iov, int count)
1923 /* Make sure data written is seen before log. */
1927 for (i = 0; i < count; i++) {
1928 r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1936 for (i = 0; i < log_num; ++i) {
1937 u64 l = min(log[i].len, len);
1938 r = log_write(vq->log_base, log[i].addr, l);
1944 eventfd_signal(vq->log_ctx, 1);
1948 /* Length written exceeds what we have stored. This is a bug. */
1952 EXPORT_SYMBOL_GPL(vhost_log_write);
1954 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1957 if (vhost_put_used_flags(vq))
1959 if (unlikely(vq->log_used)) {
1960 /* Make sure the flag is seen before log. */
1962 /* Log used flag write. */
1963 used = &vq->used->flags;
1964 log_used(vq, (used - (void __user *)vq->used),
1965 sizeof vq->used->flags);
1967 eventfd_signal(vq->log_ctx, 1);
1972 static int vhost_update_avail_event(struct vhost_virtqueue *vq)
1974 if (vhost_put_avail_event(vq))
1976 if (unlikely(vq->log_used)) {
1978 /* Make sure the event is seen before log. */
1980 /* Log avail event write */
1981 used = vhost_avail_event(vq);
1982 log_used(vq, (used - (void __user *)vq->used),
1983 sizeof *vhost_avail_event(vq));
1985 eventfd_signal(vq->log_ctx, 1);
1990 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1992 __virtio16 last_used_idx;
1994 bool is_le = vq->is_le;
1996 if (!vq->private_data)
1999 vhost_init_is_le(vq);
2001 r = vhost_update_used_flags(vq);
2004 vq->signalled_used_valid = false;
2006 !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2010 r = vhost_get_used_idx(vq, &last_used_idx);
2012 vq_err(vq, "Can't access used idx at %p\n",
2016 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2023 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2025 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2026 struct iovec iov[], int iov_size, int access)
2028 const struct vhost_iotlb_map *map;
2029 struct vhost_dev *dev = vq->dev;
2030 struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2032 u64 s = 0, last = addr + len - 1;
2035 while ((u64)len > s) {
2037 if (unlikely(ret >= iov_size)) {
2042 map = vhost_iotlb_itree_first(umem, addr, last);
2043 if (map == NULL || map->start > addr) {
2044 if (umem != dev->iotlb) {
2050 } else if (!(map->perm & access)) {
2056 size = map->size - addr + map->start;
2057 _iov->iov_len = min((u64)len - s, size);
2058 _iov->iov_base = (void __user *)(unsigned long)
2059 (map->addr + addr - map->start);
2066 vhost_iotlb_miss(vq, addr, access);
2070 /* Each buffer in the virtqueues is actually a chain of descriptors. This
2071 * function returns the next descriptor in the chain,
2072 * or -1U if we're at the end. */
2073 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2077 /* If this descriptor says it doesn't chain, we're done. */
2078 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2081 /* Check they're not leading us off end of descriptors. */
2082 next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2086 static int get_indirect(struct vhost_virtqueue *vq,
2087 struct iovec iov[], unsigned int iov_size,
2088 unsigned int *out_num, unsigned int *in_num,
2089 struct vhost_log *log, unsigned int *log_num,
2090 struct vring_desc *indirect)
2092 struct vring_desc desc;
2093 unsigned int i = 0, count, found = 0;
2094 u32 len = vhost32_to_cpu(vq, indirect->len);
2095 struct iov_iter from;
2099 if (unlikely(len % sizeof desc)) {
2100 vq_err(vq, "Invalid length in indirect descriptor: "
2101 "len 0x%llx not multiple of 0x%zx\n",
2102 (unsigned long long)len,
2107 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2108 UIO_MAXIOV, VHOST_ACCESS_RO);
2109 if (unlikely(ret < 0)) {
2111 vq_err(vq, "Translation failure %d in indirect.\n", ret);
2114 iov_iter_init(&from, ITER_SOURCE, vq->indirect, ret, len);
2115 count = len / sizeof desc;
2116 /* Buffers are chained via a 16 bit next field, so
2117 * we can have at most 2^16 of these. */
2118 if (unlikely(count > USHRT_MAX + 1)) {
2119 vq_err(vq, "Indirect buffer length too big: %d\n",
2125 unsigned iov_count = *in_num + *out_num;
2126 if (unlikely(++found > count)) {
2127 vq_err(vq, "Loop detected: last one at %u "
2128 "indirect size %u\n",
2132 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2133 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2134 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2137 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2138 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2139 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2143 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2144 access = VHOST_ACCESS_WO;
2146 access = VHOST_ACCESS_RO;
2148 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2149 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2150 iov_size - iov_count, access);
2151 if (unlikely(ret < 0)) {
2153 vq_err(vq, "Translation failure %d indirect idx %d\n",
2157 /* If this is an input descriptor, increment that count. */
2158 if (access == VHOST_ACCESS_WO) {
2160 if (unlikely(log && ret)) {
2161 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2162 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2166 /* If it's an output descriptor, they're all supposed
2167 * to come before any input descriptors. */
2168 if (unlikely(*in_num)) {
2169 vq_err(vq, "Indirect descriptor "
2170 "has out after in: idx %d\n", i);
2175 } while ((i = next_desc(vq, &desc)) != -1);
2179 /* This looks in the virtqueue and for the first available buffer, and converts
2180 * it to an iovec for convenient access. Since descriptors consist of some
2181 * number of output then some number of input descriptors, it's actually two
2182 * iovecs, but we pack them into one and note how many of each there were.
2184 * This function returns the descriptor number found, or vq->num (which is
2185 * never a valid descriptor number) if none was found. A negative code is
2186 * returned on error. */
2187 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2188 struct iovec iov[], unsigned int iov_size,
2189 unsigned int *out_num, unsigned int *in_num,
2190 struct vhost_log *log, unsigned int *log_num)
2192 struct vring_desc desc;
2193 unsigned int i, head, found = 0;
2195 __virtio16 avail_idx;
2196 __virtio16 ring_head;
2199 /* Check it isn't doing very strange things with descriptor numbers. */
2200 last_avail_idx = vq->last_avail_idx;
2202 if (vq->avail_idx == vq->last_avail_idx) {
2203 if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2204 vq_err(vq, "Failed to access avail idx at %p\n",
2208 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2210 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2211 vq_err(vq, "Guest moved used index from %u to %u",
2212 last_avail_idx, vq->avail_idx);
2216 /* If there's nothing new since last we looked, return
2219 if (vq->avail_idx == last_avail_idx)
2222 /* Only get avail ring entries after they have been
2228 /* Grab the next descriptor number they're advertising, and increment
2229 * the index we've seen. */
2230 if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2231 vq_err(vq, "Failed to read head: idx %d address %p\n",
2233 &vq->avail->ring[last_avail_idx % vq->num]);
2237 head = vhost16_to_cpu(vq, ring_head);
2239 /* If their number is silly, that's an error. */
2240 if (unlikely(head >= vq->num)) {
2241 vq_err(vq, "Guest says index %u > %u is available",
2246 /* When we start there are none of either input nor output. */
2247 *out_num = *in_num = 0;
2253 unsigned iov_count = *in_num + *out_num;
2254 if (unlikely(i >= vq->num)) {
2255 vq_err(vq, "Desc index is %u > %u, head = %u",
2259 if (unlikely(++found > vq->num)) {
2260 vq_err(vq, "Loop detected: last one at %u "
2261 "vq size %u head %u\n",
2265 ret = vhost_get_desc(vq, &desc, i);
2266 if (unlikely(ret)) {
2267 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2271 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2272 ret = get_indirect(vq, iov, iov_size,
2274 log, log_num, &desc);
2275 if (unlikely(ret < 0)) {
2277 vq_err(vq, "Failure detected "
2278 "in indirect descriptor at idx %d\n", i);
2284 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2285 access = VHOST_ACCESS_WO;
2287 access = VHOST_ACCESS_RO;
2288 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2289 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2290 iov_size - iov_count, access);
2291 if (unlikely(ret < 0)) {
2293 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2297 if (access == VHOST_ACCESS_WO) {
2298 /* If this is an input descriptor,
2299 * increment that count. */
2301 if (unlikely(log && ret)) {
2302 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2303 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2307 /* If it's an output descriptor, they're all supposed
2308 * to come before any input descriptors. */
2309 if (unlikely(*in_num)) {
2310 vq_err(vq, "Descriptor has out after in: "
2316 } while ((i = next_desc(vq, &desc)) != -1);
2318 /* On success, increment avail index. */
2319 vq->last_avail_idx++;
2321 /* Assume notifications from guest are disabled at this point,
2322 * if they aren't we would need to update avail_event index. */
2323 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2326 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2328 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2329 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2331 vq->last_avail_idx -= n;
2333 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2335 /* After we've used one of their buffers, we tell them about it. We'll then
2336 * want to notify the guest, using eventfd. */
2337 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2339 struct vring_used_elem heads = {
2340 cpu_to_vhost32(vq, head),
2341 cpu_to_vhost32(vq, len)
2344 return vhost_add_used_n(vq, &heads, 1);
2346 EXPORT_SYMBOL_GPL(vhost_add_used);
2348 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2349 struct vring_used_elem *heads,
2352 vring_used_elem_t __user *used;
2356 start = vq->last_used_idx & (vq->num - 1);
2357 used = vq->used->ring + start;
2358 if (vhost_put_used(vq, heads, start, count)) {
2359 vq_err(vq, "Failed to write used");
2362 if (unlikely(vq->log_used)) {
2363 /* Make sure data is seen before log. */
2365 /* Log used ring entry write. */
2366 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2367 count * sizeof *used);
2369 old = vq->last_used_idx;
2370 new = (vq->last_used_idx += count);
2371 /* If the driver never bothers to signal in a very long while,
2372 * used index might wrap around. If that happens, invalidate
2373 * signalled_used index we stored. TODO: make sure driver
2374 * signals at least once in 2^16 and remove this. */
2375 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2376 vq->signalled_used_valid = false;
2380 /* After we've used one of their buffers, we tell them about it. We'll then
2381 * want to notify the guest, using eventfd. */
2382 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2387 start = vq->last_used_idx & (vq->num - 1);
2388 n = vq->num - start;
2390 r = __vhost_add_used_n(vq, heads, n);
2396 r = __vhost_add_used_n(vq, heads, count);
2398 /* Make sure buffer is written before we update index. */
2400 if (vhost_put_used_idx(vq)) {
2401 vq_err(vq, "Failed to increment used idx");
2404 if (unlikely(vq->log_used)) {
2405 /* Make sure used idx is seen before log. */
2407 /* Log used index update. */
2408 log_used(vq, offsetof(struct vring_used, idx),
2409 sizeof vq->used->idx);
2411 eventfd_signal(vq->log_ctx, 1);
2415 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2417 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2422 /* Flush out used index updates. This is paired
2423 * with the barrier that the Guest executes when enabling
2427 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2428 unlikely(vq->avail_idx == vq->last_avail_idx))
2431 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2433 if (vhost_get_avail_flags(vq, &flags)) {
2434 vq_err(vq, "Failed to get flags");
2437 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2439 old = vq->signalled_used;
2440 v = vq->signalled_used_valid;
2441 new = vq->signalled_used = vq->last_used_idx;
2442 vq->signalled_used_valid = true;
2447 if (vhost_get_used_event(vq, &event)) {
2448 vq_err(vq, "Failed to get used event idx");
2451 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2454 /* This actually signals the guest, using eventfd. */
2455 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2457 /* Signal the Guest tell them we used something up. */
2458 if (vq->call_ctx.ctx && vhost_notify(dev, vq))
2459 eventfd_signal(vq->call_ctx.ctx, 1);
2461 EXPORT_SYMBOL_GPL(vhost_signal);
2463 /* And here's the combo meal deal. Supersize me! */
2464 void vhost_add_used_and_signal(struct vhost_dev *dev,
2465 struct vhost_virtqueue *vq,
2466 unsigned int head, int len)
2468 vhost_add_used(vq, head, len);
2469 vhost_signal(dev, vq);
2471 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2473 /* multi-buffer version of vhost_add_used_and_signal */
2474 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2475 struct vhost_virtqueue *vq,
2476 struct vring_used_elem *heads, unsigned count)
2478 vhost_add_used_n(vq, heads, count);
2479 vhost_signal(dev, vq);
2481 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2483 /* return true if we're sure that avaiable ring is empty */
2484 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2486 __virtio16 avail_idx;
2489 if (vq->avail_idx != vq->last_avail_idx)
2492 r = vhost_get_avail_idx(vq, &avail_idx);
2495 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2497 return vq->avail_idx == vq->last_avail_idx;
2499 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2501 /* OK, now we need to know about added descriptors. */
2502 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2504 __virtio16 avail_idx;
2507 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2509 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2510 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2511 r = vhost_update_used_flags(vq);
2513 vq_err(vq, "Failed to enable notification at %p: %d\n",
2514 &vq->used->flags, r);
2518 r = vhost_update_avail_event(vq);
2520 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2521 vhost_avail_event(vq), r);
2525 /* They could have slipped one in as we were doing that: make
2526 * sure it's written, then check again. */
2528 r = vhost_get_avail_idx(vq, &avail_idx);
2530 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2531 &vq->avail->idx, r);
2534 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2536 return vq->avail_idx != vq->last_avail_idx;
2538 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2540 /* We don't need to be notified again. */
2541 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2545 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2547 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2548 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2549 r = vhost_update_used_flags(vq);
2551 vq_err(vq, "Failed to disable notification at %p: %d\n",
2552 &vq->used->flags, r);
2555 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2557 /* Create a new message. */
2558 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2560 /* Make sure all padding within the structure is initialized. */
2561 struct vhost_msg_node *node = kzalloc(sizeof(*node), GFP_KERNEL);
2566 node->msg.type = type;
2569 EXPORT_SYMBOL_GPL(vhost_new_msg);
2571 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2572 struct vhost_msg_node *node)
2574 spin_lock(&dev->iotlb_lock);
2575 list_add_tail(&node->node, head);
2576 spin_unlock(&dev->iotlb_lock);
2578 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2580 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2582 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2583 struct list_head *head)
2585 struct vhost_msg_node *node = NULL;
2587 spin_lock(&dev->iotlb_lock);
2588 if (!list_empty(head)) {
2589 node = list_first_entry(head, struct vhost_msg_node,
2591 list_del(&node->node);
2593 spin_unlock(&dev->iotlb_lock);
2597 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2599 void vhost_set_backend_features(struct vhost_dev *dev, u64 features)
2601 struct vhost_virtqueue *vq;
2604 mutex_lock(&dev->mutex);
2605 for (i = 0; i < dev->nvqs; ++i) {
2607 mutex_lock(&vq->mutex);
2608 vq->acked_backend_features = features;
2609 mutex_unlock(&vq->mutex);
2611 mutex_unlock(&dev->mutex);
2613 EXPORT_SYMBOL_GPL(vhost_set_backend_features);
2615 static int __init vhost_init(void)
2620 static void __exit vhost_exit(void)
2624 module_init(vhost_init);
2625 module_exit(vhost_exit);
2627 MODULE_VERSION("0.0.1");
2628 MODULE_LICENSE("GPL v2");
2629 MODULE_AUTHOR("Michael S. Tsirkin");
2630 MODULE_DESCRIPTION("Host kernel accelerator for virtio");