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 sizeof(*vq->avail) +
438 sizeof(*vq->avail->ring) * num + event;
441 static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
444 size_t event __maybe_unused =
445 vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
447 return sizeof(*vq->used) +
448 sizeof(*vq->used->ring) * num + event;
451 static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
454 return sizeof(*vq->desc) * num;
457 void vhost_dev_init(struct vhost_dev *dev,
458 struct vhost_virtqueue **vqs, int nvqs,
459 int iov_limit, int weight, int byte_weight,
461 int (*msg_handler)(struct vhost_dev *dev, u32 asid,
462 struct vhost_iotlb_msg *msg))
464 struct vhost_virtqueue *vq;
469 mutex_init(&dev->mutex);
475 dev->iov_limit = iov_limit;
476 dev->weight = weight;
477 dev->byte_weight = byte_weight;
478 dev->use_worker = use_worker;
479 dev->msg_handler = msg_handler;
480 init_waitqueue_head(&dev->wait);
481 INIT_LIST_HEAD(&dev->read_list);
482 INIT_LIST_HEAD(&dev->pending_list);
483 spin_lock_init(&dev->iotlb_lock);
486 for (i = 0; i < dev->nvqs; ++i) {
492 mutex_init(&vq->mutex);
493 vhost_vq_reset(dev, vq);
495 vhost_poll_init(&vq->poll, vq->handle_kick,
499 EXPORT_SYMBOL_GPL(vhost_dev_init);
501 /* Caller should have device mutex */
502 long vhost_dev_check_owner(struct vhost_dev *dev)
504 /* Are you the owner? If not, I don't think you mean to do that */
505 return dev->mm == current->mm ? 0 : -EPERM;
507 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
509 /* Caller should have device mutex */
510 bool vhost_dev_has_owner(struct vhost_dev *dev)
514 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
516 static void vhost_attach_mm(struct vhost_dev *dev)
518 /* No owner, become one */
519 if (dev->use_worker) {
520 dev->mm = get_task_mm(current);
522 /* vDPA device does not use worker thead, so there's
523 * no need to hold the address space for mm. This help
524 * to avoid deadlock in the case of mmap() which may
525 * held the refcnt of the file and depends on release
526 * method to remove vma.
528 dev->mm = current->mm;
533 static void vhost_detach_mm(struct vhost_dev *dev)
546 static void vhost_worker_free(struct vhost_dev *dev)
548 struct vhost_worker *worker = dev->worker;
554 WARN_ON(!llist_empty(&worker->work_list));
555 vhost_task_stop(worker->vtsk);
559 static int vhost_worker_create(struct vhost_dev *dev)
561 struct vhost_worker *worker;
562 struct vhost_task *vtsk;
563 char name[TASK_COMM_LEN];
566 worker = kzalloc(sizeof(*worker), GFP_KERNEL_ACCOUNT);
570 dev->worker = worker;
571 worker->kcov_handle = kcov_common_handle();
572 init_llist_head(&worker->work_list);
573 snprintf(name, sizeof(name), "vhost-%d", current->pid);
575 vtsk = vhost_task_create(vhost_worker, worker, name);
582 vhost_task_start(vtsk);
591 /* Caller should have device mutex */
592 long vhost_dev_set_owner(struct vhost_dev *dev)
596 /* Is there an owner already? */
597 if (vhost_dev_has_owner(dev)) {
602 vhost_attach_mm(dev);
604 if (dev->use_worker) {
605 err = vhost_worker_create(dev);
610 err = vhost_dev_alloc_iovecs(dev);
616 vhost_worker_free(dev);
618 vhost_detach_mm(dev);
622 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
624 static struct vhost_iotlb *iotlb_alloc(void)
626 return vhost_iotlb_alloc(max_iotlb_entries,
627 VHOST_IOTLB_FLAG_RETIRE);
630 struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
632 return iotlb_alloc();
634 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
636 /* Caller should have device mutex */
637 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
641 vhost_dev_cleanup(dev);
644 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
645 * VQs aren't running.
647 for (i = 0; i < dev->nvqs; ++i)
648 dev->vqs[i]->umem = umem;
650 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
652 void vhost_dev_stop(struct vhost_dev *dev)
656 for (i = 0; i < dev->nvqs; ++i) {
657 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick)
658 vhost_poll_stop(&dev->vqs[i]->poll);
661 vhost_dev_flush(dev);
663 EXPORT_SYMBOL_GPL(vhost_dev_stop);
665 void vhost_clear_msg(struct vhost_dev *dev)
667 struct vhost_msg_node *node, *n;
669 spin_lock(&dev->iotlb_lock);
671 list_for_each_entry_safe(node, n, &dev->read_list, node) {
672 list_del(&node->node);
676 list_for_each_entry_safe(node, n, &dev->pending_list, node) {
677 list_del(&node->node);
681 spin_unlock(&dev->iotlb_lock);
683 EXPORT_SYMBOL_GPL(vhost_clear_msg);
685 void vhost_dev_cleanup(struct vhost_dev *dev)
689 for (i = 0; i < dev->nvqs; ++i) {
690 if (dev->vqs[i]->error_ctx)
691 eventfd_ctx_put(dev->vqs[i]->error_ctx);
692 if (dev->vqs[i]->kick)
693 fput(dev->vqs[i]->kick);
694 if (dev->vqs[i]->call_ctx.ctx)
695 eventfd_ctx_put(dev->vqs[i]->call_ctx.ctx);
696 vhost_vq_reset(dev, dev->vqs[i]);
698 vhost_dev_free_iovecs(dev);
700 eventfd_ctx_put(dev->log_ctx);
702 /* No one will access memory at this point */
703 vhost_iotlb_free(dev->umem);
705 vhost_iotlb_free(dev->iotlb);
707 vhost_clear_msg(dev);
708 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
709 vhost_worker_free(dev);
710 vhost_detach_mm(dev);
712 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
714 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
716 u64 a = addr / VHOST_PAGE_SIZE / 8;
718 /* Make sure 64 bit math will not overflow. */
719 if (a > ULONG_MAX - (unsigned long)log_base ||
720 a + (unsigned long)log_base > ULONG_MAX)
723 return access_ok(log_base + a,
724 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
727 /* Make sure 64 bit math will not overflow. */
728 static bool vhost_overflow(u64 uaddr, u64 size)
730 if (uaddr > ULONG_MAX || size > ULONG_MAX)
736 return uaddr > ULONG_MAX - size + 1;
739 /* Caller should have vq mutex and device mutex. */
740 static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
743 struct vhost_iotlb_map *map;
748 list_for_each_entry(map, &umem->list, link) {
749 unsigned long a = map->addr;
751 if (vhost_overflow(map->addr, map->size))
755 if (!access_ok((void __user *)a, map->size))
757 else if (log_all && !log_access_ok(log_base,
765 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
766 u64 addr, unsigned int size,
769 const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
774 return (void __user *)(uintptr_t)(map->addr + addr - map->start);
777 /* Can we switch to this memory table? */
778 /* Caller should have device mutex but not vq mutex */
779 static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
784 for (i = 0; i < d->nvqs; ++i) {
788 mutex_lock(&d->vqs[i]->mutex);
789 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
790 /* If ring is inactive, will check when it's enabled. */
791 if (d->vqs[i]->private_data)
792 ok = vq_memory_access_ok(d->vqs[i]->log_base,
796 mutex_unlock(&d->vqs[i]->mutex);
803 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
804 struct iovec iov[], int iov_size, int access);
806 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
807 const void *from, unsigned size)
812 return __copy_to_user(to, from, size);
814 /* This function should be called after iotlb
815 * prefetch, which means we're sure that all vq
816 * could be access through iotlb. So -EAGAIN should
817 * not happen in this case.
820 void __user *uaddr = vhost_vq_meta_fetch(vq,
821 (u64)(uintptr_t)to, size,
825 return __copy_to_user(uaddr, from, size);
827 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
828 ARRAY_SIZE(vq->iotlb_iov),
832 iov_iter_init(&t, ITER_DEST, vq->iotlb_iov, ret, size);
833 ret = copy_to_iter(from, size, &t);
841 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
842 void __user *from, unsigned size)
847 return __copy_from_user(to, from, size);
849 /* This function should be called after iotlb
850 * prefetch, which means we're sure that vq
851 * could be access through iotlb. So -EAGAIN should
852 * not happen in this case.
854 void __user *uaddr = vhost_vq_meta_fetch(vq,
855 (u64)(uintptr_t)from, size,
860 return __copy_from_user(to, uaddr, size);
862 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
863 ARRAY_SIZE(vq->iotlb_iov),
866 vq_err(vq, "IOTLB translation failure: uaddr "
867 "%p size 0x%llx\n", from,
868 (unsigned long long) size);
871 iov_iter_init(&f, ITER_SOURCE, vq->iotlb_iov, ret, size);
872 ret = copy_from_iter(to, size, &f);
881 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
882 void __user *addr, unsigned int size,
887 ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
888 ARRAY_SIZE(vq->iotlb_iov),
891 vq_err(vq, "IOTLB translation failure: uaddr "
892 "%p size 0x%llx\n", addr,
893 (unsigned long long) size);
897 if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
898 vq_err(vq, "Non atomic userspace memory access: uaddr "
899 "%p size 0x%llx\n", addr,
900 (unsigned long long) size);
904 return vq->iotlb_iov[0].iov_base;
907 /* This function should be called after iotlb
908 * prefetch, which means we're sure that vq
909 * could be access through iotlb. So -EAGAIN should
910 * not happen in this case.
912 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
913 void __user *addr, unsigned int size,
916 void __user *uaddr = vhost_vq_meta_fetch(vq,
917 (u64)(uintptr_t)addr, size, type);
921 return __vhost_get_user_slow(vq, addr, size, type);
924 #define vhost_put_user(vq, x, ptr) \
928 ret = __put_user(x, ptr); \
930 __typeof__(ptr) to = \
931 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
932 sizeof(*ptr), VHOST_ADDR_USED); \
934 ret = __put_user(x, to); \
941 static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
943 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
944 vhost_avail_event(vq));
947 static inline int vhost_put_used(struct vhost_virtqueue *vq,
948 struct vring_used_elem *head, int idx,
951 return vhost_copy_to_user(vq, vq->used->ring + idx, head,
952 count * sizeof(*head));
955 static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
958 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
962 static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
965 return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
969 #define vhost_get_user(vq, x, ptr, type) \
973 ret = __get_user(x, ptr); \
975 __typeof__(ptr) from = \
976 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
980 ret = __get_user(x, from); \
987 #define vhost_get_avail(vq, x, ptr) \
988 vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
990 #define vhost_get_used(vq, x, ptr) \
991 vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
993 static void vhost_dev_lock_vqs(struct vhost_dev *d)
996 for (i = 0; i < d->nvqs; ++i)
997 mutex_lock_nested(&d->vqs[i]->mutex, i);
1000 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
1003 for (i = 0; i < d->nvqs; ++i)
1004 mutex_unlock(&d->vqs[i]->mutex);
1007 static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
1010 return vhost_get_avail(vq, *idx, &vq->avail->idx);
1013 static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
1014 __virtio16 *head, int idx)
1016 return vhost_get_avail(vq, *head,
1017 &vq->avail->ring[idx & (vq->num - 1)]);
1020 static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
1023 return vhost_get_avail(vq, *flags, &vq->avail->flags);
1026 static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1029 return vhost_get_avail(vq, *event, vhost_used_event(vq));
1032 static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1035 return vhost_get_used(vq, *idx, &vq->used->idx);
1038 static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1039 struct vring_desc *desc, int idx)
1041 return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1044 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1045 struct vhost_iotlb_msg *msg)
1047 struct vhost_msg_node *node, *n;
1049 spin_lock(&d->iotlb_lock);
1051 list_for_each_entry_safe(node, n, &d->pending_list, node) {
1052 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1053 if (msg->iova <= vq_msg->iova &&
1054 msg->iova + msg->size - 1 >= vq_msg->iova &&
1055 vq_msg->type == VHOST_IOTLB_MISS) {
1056 vhost_poll_queue(&node->vq->poll);
1057 list_del(&node->node);
1062 spin_unlock(&d->iotlb_lock);
1065 static bool umem_access_ok(u64 uaddr, u64 size, int access)
1067 unsigned long a = uaddr;
1069 /* Make sure 64 bit math will not overflow. */
1070 if (vhost_overflow(uaddr, size))
1073 if ((access & VHOST_ACCESS_RO) &&
1074 !access_ok((void __user *)a, size))
1076 if ((access & VHOST_ACCESS_WO) &&
1077 !access_ok((void __user *)a, size))
1082 static int vhost_process_iotlb_msg(struct vhost_dev *dev, u32 asid,
1083 struct vhost_iotlb_msg *msg)
1090 mutex_lock(&dev->mutex);
1091 vhost_dev_lock_vqs(dev);
1092 switch (msg->type) {
1093 case VHOST_IOTLB_UPDATE:
1098 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1102 vhost_vq_meta_reset(dev);
1103 if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1104 msg->iova + msg->size - 1,
1105 msg->uaddr, msg->perm)) {
1109 vhost_iotlb_notify_vq(dev, msg);
1111 case VHOST_IOTLB_INVALIDATE:
1116 vhost_vq_meta_reset(dev);
1117 vhost_iotlb_del_range(dev->iotlb, msg->iova,
1118 msg->iova + msg->size - 1);
1125 vhost_dev_unlock_vqs(dev);
1126 mutex_unlock(&dev->mutex);
1130 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1131 struct iov_iter *from)
1133 struct vhost_iotlb_msg msg;
1138 ret = copy_from_iter(&type, sizeof(type), from);
1139 if (ret != sizeof(type)) {
1145 case VHOST_IOTLB_MSG:
1146 /* There maybe a hole after type for V1 message type,
1149 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1151 case VHOST_IOTLB_MSG_V2:
1152 if (vhost_backend_has_feature(dev->vqs[0],
1153 VHOST_BACKEND_F_IOTLB_ASID)) {
1154 ret = copy_from_iter(&asid, sizeof(asid), from);
1155 if (ret != sizeof(asid)) {
1161 offset = sizeof(__u32);
1168 iov_iter_advance(from, offset);
1169 ret = copy_from_iter(&msg, sizeof(msg), from);
1170 if (ret != sizeof(msg)) {
1175 if ((msg.type == VHOST_IOTLB_UPDATE ||
1176 msg.type == VHOST_IOTLB_INVALIDATE) &&
1182 if (dev->msg_handler)
1183 ret = dev->msg_handler(dev, asid, &msg);
1185 ret = vhost_process_iotlb_msg(dev, asid, &msg);
1191 ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1192 sizeof(struct vhost_msg_v2);
1196 EXPORT_SYMBOL(vhost_chr_write_iter);
1198 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1203 poll_wait(file, &dev->wait, wait);
1205 if (!list_empty(&dev->read_list))
1206 mask |= EPOLLIN | EPOLLRDNORM;
1210 EXPORT_SYMBOL(vhost_chr_poll);
1212 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1216 struct vhost_msg_node *node;
1218 unsigned size = sizeof(struct vhost_msg);
1220 if (iov_iter_count(to) < size)
1225 prepare_to_wait(&dev->wait, &wait,
1226 TASK_INTERRUPTIBLE);
1228 node = vhost_dequeue_msg(dev, &dev->read_list);
1235 if (signal_pending(current)) {
1248 finish_wait(&dev->wait, &wait);
1251 struct vhost_iotlb_msg *msg;
1252 void *start = &node->msg;
1254 switch (node->msg.type) {
1255 case VHOST_IOTLB_MSG:
1256 size = sizeof(node->msg);
1257 msg = &node->msg.iotlb;
1259 case VHOST_IOTLB_MSG_V2:
1260 size = sizeof(node->msg_v2);
1261 msg = &node->msg_v2.iotlb;
1268 ret = copy_to_iter(start, size, to);
1269 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1273 vhost_enqueue_msg(dev, &dev->pending_list, node);
1278 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1280 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1282 struct vhost_dev *dev = vq->dev;
1283 struct vhost_msg_node *node;
1284 struct vhost_iotlb_msg *msg;
1285 bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1287 node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1292 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1293 msg = &node->msg_v2.iotlb;
1295 msg = &node->msg.iotlb;
1298 msg->type = VHOST_IOTLB_MISS;
1302 vhost_enqueue_msg(dev, &dev->read_list, node);
1307 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1308 vring_desc_t __user *desc,
1309 vring_avail_t __user *avail,
1310 vring_used_t __user *used)
1313 /* If an IOTLB device is present, the vring addresses are
1314 * GIOVAs. Access validation occurs at prefetch time. */
1318 return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1319 access_ok(avail, vhost_get_avail_size(vq, num)) &&
1320 access_ok(used, vhost_get_used_size(vq, num));
1323 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1324 const struct vhost_iotlb_map *map,
1327 int access = (type == VHOST_ADDR_USED) ?
1328 VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1330 if (likely(map->perm & access))
1331 vq->meta_iotlb[type] = map;
1334 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1335 int access, u64 addr, u64 len, int type)
1337 const struct vhost_iotlb_map *map;
1338 struct vhost_iotlb *umem = vq->iotlb;
1339 u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1341 if (vhost_vq_meta_fetch(vq, addr, len, type))
1345 map = vhost_iotlb_itree_first(umem, addr, last);
1346 if (map == NULL || map->start > addr) {
1347 vhost_iotlb_miss(vq, addr, access);
1349 } else if (!(map->perm & access)) {
1350 /* Report the possible access violation by
1351 * request another translation from userspace.
1356 size = map->size - addr + map->start;
1358 if (orig_addr == addr && size >= len)
1359 vhost_vq_meta_update(vq, map, type);
1368 int vq_meta_prefetch(struct vhost_virtqueue *vq)
1370 unsigned int num = vq->num;
1375 return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1376 vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1377 iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1378 vhost_get_avail_size(vq, num),
1379 VHOST_ADDR_AVAIL) &&
1380 iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1381 vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1383 EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1385 /* Can we log writes? */
1386 /* Caller should have device mutex but not vq mutex */
1387 bool vhost_log_access_ok(struct vhost_dev *dev)
1389 return memory_access_ok(dev, dev->umem, 1);
1391 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1393 static bool vq_log_used_access_ok(struct vhost_virtqueue *vq,
1394 void __user *log_base,
1398 /* If an IOTLB device is present, log_addr is a GIOVA that
1399 * will never be logged by log_used(). */
1403 return !log_used || log_access_ok(log_base, log_addr,
1404 vhost_get_used_size(vq, vq->num));
1407 /* Verify access for write logging. */
1408 /* Caller should have vq mutex and device mutex */
1409 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1410 void __user *log_base)
1412 return vq_memory_access_ok(log_base, vq->umem,
1413 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1414 vq_log_used_access_ok(vq, log_base, vq->log_used, vq->log_addr);
1417 /* Can we start vq? */
1418 /* Caller should have vq mutex and device mutex */
1419 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1421 if (!vq_log_access_ok(vq, vq->log_base))
1424 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1426 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1428 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1430 struct vhost_memory mem, *newmem;
1431 struct vhost_memory_region *region;
1432 struct vhost_iotlb *newumem, *oldumem;
1433 unsigned long size = offsetof(struct vhost_memory, regions);
1436 if (copy_from_user(&mem, m, size))
1440 if (mem.nregions > max_mem_regions)
1442 newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1447 memcpy(newmem, &mem, size);
1448 if (copy_from_user(newmem->regions, m->regions,
1449 flex_array_size(newmem, regions, mem.nregions))) {
1454 newumem = iotlb_alloc();
1460 for (region = newmem->regions;
1461 region < newmem->regions + mem.nregions;
1463 if (vhost_iotlb_add_range(newumem,
1464 region->guest_phys_addr,
1465 region->guest_phys_addr +
1466 region->memory_size - 1,
1467 region->userspace_addr,
1472 if (!memory_access_ok(d, newumem, 0))
1478 /* All memory accesses are done under some VQ mutex. */
1479 for (i = 0; i < d->nvqs; ++i) {
1480 mutex_lock(&d->vqs[i]->mutex);
1481 d->vqs[i]->umem = newumem;
1482 mutex_unlock(&d->vqs[i]->mutex);
1486 vhost_iotlb_free(oldumem);
1490 vhost_iotlb_free(newumem);
1495 static long vhost_vring_set_num(struct vhost_dev *d,
1496 struct vhost_virtqueue *vq,
1499 struct vhost_vring_state s;
1501 /* Resizing ring with an active backend?
1502 * You don't want to do that. */
1503 if (vq->private_data)
1506 if (copy_from_user(&s, argp, sizeof s))
1509 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1516 static long vhost_vring_set_addr(struct vhost_dev *d,
1517 struct vhost_virtqueue *vq,
1520 struct vhost_vring_addr a;
1522 if (copy_from_user(&a, argp, sizeof a))
1524 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1527 /* For 32bit, verify that the top 32bits of the user
1528 data are set to zero. */
1529 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1530 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1531 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1534 /* Make sure it's safe to cast pointers to vring types. */
1535 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1536 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1537 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1538 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1539 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1542 /* We only verify access here if backend is configured.
1543 * If it is not, we don't as size might not have been setup.
1544 * We will verify when backend is configured. */
1545 if (vq->private_data) {
1546 if (!vq_access_ok(vq, vq->num,
1547 (void __user *)(unsigned long)a.desc_user_addr,
1548 (void __user *)(unsigned long)a.avail_user_addr,
1549 (void __user *)(unsigned long)a.used_user_addr))
1552 /* Also validate log access for used ring if enabled. */
1553 if (!vq_log_used_access_ok(vq, vq->log_base,
1554 a.flags & (0x1 << VHOST_VRING_F_LOG),
1559 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1560 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1561 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1562 vq->log_addr = a.log_guest_addr;
1563 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1568 static long vhost_vring_set_num_addr(struct vhost_dev *d,
1569 struct vhost_virtqueue *vq,
1575 mutex_lock(&vq->mutex);
1578 case VHOST_SET_VRING_NUM:
1579 r = vhost_vring_set_num(d, vq, argp);
1581 case VHOST_SET_VRING_ADDR:
1582 r = vhost_vring_set_addr(d, vq, argp);
1588 mutex_unlock(&vq->mutex);
1592 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1594 struct file *eventfp, *filep = NULL;
1595 bool pollstart = false, pollstop = false;
1596 struct eventfd_ctx *ctx = NULL;
1597 u32 __user *idxp = argp;
1598 struct vhost_virtqueue *vq;
1599 struct vhost_vring_state s;
1600 struct vhost_vring_file f;
1604 r = get_user(idx, idxp);
1610 idx = array_index_nospec(idx, d->nvqs);
1613 if (ioctl == VHOST_SET_VRING_NUM ||
1614 ioctl == VHOST_SET_VRING_ADDR) {
1615 return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1618 mutex_lock(&vq->mutex);
1621 case VHOST_SET_VRING_BASE:
1622 /* Moving base with an active backend?
1623 * You don't want to do that. */
1624 if (vq->private_data) {
1628 if (copy_from_user(&s, argp, sizeof s)) {
1632 if (s.num > 0xffff) {
1636 vq->last_avail_idx = s.num;
1637 /* Forget the cached index value. */
1638 vq->avail_idx = vq->last_avail_idx;
1640 case VHOST_GET_VRING_BASE:
1642 s.num = vq->last_avail_idx;
1643 if (copy_to_user(argp, &s, sizeof s))
1646 case VHOST_SET_VRING_KICK:
1647 if (copy_from_user(&f, argp, sizeof f)) {
1651 eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
1652 if (IS_ERR(eventfp)) {
1653 r = PTR_ERR(eventfp);
1656 if (eventfp != vq->kick) {
1657 pollstop = (filep = vq->kick) != NULL;
1658 pollstart = (vq->kick = eventfp) != NULL;
1662 case VHOST_SET_VRING_CALL:
1663 if (copy_from_user(&f, argp, sizeof f)) {
1667 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1673 swap(ctx, vq->call_ctx.ctx);
1675 case VHOST_SET_VRING_ERR:
1676 if (copy_from_user(&f, argp, sizeof f)) {
1680 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1685 swap(ctx, vq->error_ctx);
1687 case VHOST_SET_VRING_ENDIAN:
1688 r = vhost_set_vring_endian(vq, argp);
1690 case VHOST_GET_VRING_ENDIAN:
1691 r = vhost_get_vring_endian(vq, idx, argp);
1693 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1694 if (copy_from_user(&s, argp, sizeof(s))) {
1698 vq->busyloop_timeout = s.num;
1700 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1702 s.num = vq->busyloop_timeout;
1703 if (copy_to_user(argp, &s, sizeof(s)))
1710 if (pollstop && vq->handle_kick)
1711 vhost_poll_stop(&vq->poll);
1713 if (!IS_ERR_OR_NULL(ctx))
1714 eventfd_ctx_put(ctx);
1718 if (pollstart && vq->handle_kick)
1719 r = vhost_poll_start(&vq->poll, vq->kick);
1721 mutex_unlock(&vq->mutex);
1723 if (pollstop && vq->handle_kick)
1724 vhost_dev_flush(vq->poll.dev);
1727 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1729 int vhost_init_device_iotlb(struct vhost_dev *d)
1731 struct vhost_iotlb *niotlb, *oiotlb;
1734 niotlb = iotlb_alloc();
1741 for (i = 0; i < d->nvqs; ++i) {
1742 struct vhost_virtqueue *vq = d->vqs[i];
1744 mutex_lock(&vq->mutex);
1746 __vhost_vq_meta_reset(vq);
1747 mutex_unlock(&vq->mutex);
1750 vhost_iotlb_free(oiotlb);
1754 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1756 /* Caller must have device mutex */
1757 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1759 struct eventfd_ctx *ctx;
1764 /* If you are not the owner, you can become one */
1765 if (ioctl == VHOST_SET_OWNER) {
1766 r = vhost_dev_set_owner(d);
1770 /* You must be the owner to do anything else */
1771 r = vhost_dev_check_owner(d);
1776 case VHOST_SET_MEM_TABLE:
1777 r = vhost_set_memory(d, argp);
1779 case VHOST_SET_LOG_BASE:
1780 if (copy_from_user(&p, argp, sizeof p)) {
1784 if ((u64)(unsigned long)p != p) {
1788 for (i = 0; i < d->nvqs; ++i) {
1789 struct vhost_virtqueue *vq;
1790 void __user *base = (void __user *)(unsigned long)p;
1792 mutex_lock(&vq->mutex);
1793 /* If ring is inactive, will check when it's enabled. */
1794 if (vq->private_data && !vq_log_access_ok(vq, base))
1797 vq->log_base = base;
1798 mutex_unlock(&vq->mutex);
1801 case VHOST_SET_LOG_FD:
1802 r = get_user(fd, (int __user *)argp);
1805 ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
1810 swap(ctx, d->log_ctx);
1811 for (i = 0; i < d->nvqs; ++i) {
1812 mutex_lock(&d->vqs[i]->mutex);
1813 d->vqs[i]->log_ctx = d->log_ctx;
1814 mutex_unlock(&d->vqs[i]->mutex);
1817 eventfd_ctx_put(ctx);
1826 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1828 /* TODO: This is really inefficient. We need something like get_user()
1829 * (instruction directly accesses the data, with an exception table entry
1830 * returning -EFAULT). See Documentation/arch/x86/exception-tables.rst.
1832 static int set_bit_to_user(int nr, void __user *addr)
1834 unsigned long log = (unsigned long)addr;
1837 int bit = nr + (log % PAGE_SIZE) * 8;
1840 r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
1844 base = kmap_atomic(page);
1846 kunmap_atomic(base);
1847 unpin_user_pages_dirty_lock(&page, 1, true);
1851 static int log_write(void __user *log_base,
1852 u64 write_address, u64 write_length)
1854 u64 write_page = write_address / VHOST_PAGE_SIZE;
1859 write_length += write_address % VHOST_PAGE_SIZE;
1861 u64 base = (u64)(unsigned long)log_base;
1862 u64 log = base + write_page / 8;
1863 int bit = write_page % 8;
1864 if ((u64)(unsigned long)log != log)
1866 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1869 if (write_length <= VHOST_PAGE_SIZE)
1871 write_length -= VHOST_PAGE_SIZE;
1877 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1879 struct vhost_iotlb *umem = vq->umem;
1880 struct vhost_iotlb_map *u;
1881 u64 start, end, l, min;
1887 /* More than one GPAs can be mapped into a single HVA. So
1888 * iterate all possible umems here to be safe.
1890 list_for_each_entry(u, &umem->list, link) {
1891 if (u->addr > hva - 1 + len ||
1892 u->addr - 1 + u->size < hva)
1894 start = max(u->addr, hva);
1895 end = min(u->addr - 1 + u->size, hva - 1 + len);
1896 l = end - start + 1;
1897 r = log_write(vq->log_base,
1898 u->start + start - u->addr,
1916 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1918 struct iovec *iov = vq->log_iov;
1922 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1924 ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1925 len, iov, 64, VHOST_ACCESS_WO);
1929 for (i = 0; i < ret; i++) {
1930 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1939 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1940 unsigned int log_num, u64 len, struct iovec *iov, int count)
1944 /* Make sure data written is seen before log. */
1948 for (i = 0; i < count; i++) {
1949 r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1957 for (i = 0; i < log_num; ++i) {
1958 u64 l = min(log[i].len, len);
1959 r = log_write(vq->log_base, log[i].addr, l);
1965 eventfd_signal(vq->log_ctx, 1);
1969 /* Length written exceeds what we have stored. This is a bug. */
1973 EXPORT_SYMBOL_GPL(vhost_log_write);
1975 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1978 if (vhost_put_used_flags(vq))
1980 if (unlikely(vq->log_used)) {
1981 /* Make sure the flag is seen before log. */
1983 /* Log used flag write. */
1984 used = &vq->used->flags;
1985 log_used(vq, (used - (void __user *)vq->used),
1986 sizeof vq->used->flags);
1988 eventfd_signal(vq->log_ctx, 1);
1993 static int vhost_update_avail_event(struct vhost_virtqueue *vq)
1995 if (vhost_put_avail_event(vq))
1997 if (unlikely(vq->log_used)) {
1999 /* Make sure the event is seen before log. */
2001 /* Log avail event write */
2002 used = vhost_avail_event(vq);
2003 log_used(vq, (used - (void __user *)vq->used),
2004 sizeof *vhost_avail_event(vq));
2006 eventfd_signal(vq->log_ctx, 1);
2011 int vhost_vq_init_access(struct vhost_virtqueue *vq)
2013 __virtio16 last_used_idx;
2015 bool is_le = vq->is_le;
2017 if (!vq->private_data)
2020 vhost_init_is_le(vq);
2022 r = vhost_update_used_flags(vq);
2025 vq->signalled_used_valid = false;
2027 !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2031 r = vhost_get_used_idx(vq, &last_used_idx);
2033 vq_err(vq, "Can't access used idx at %p\n",
2037 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2044 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2046 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2047 struct iovec iov[], int iov_size, int access)
2049 const struct vhost_iotlb_map *map;
2050 struct vhost_dev *dev = vq->dev;
2051 struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2053 u64 s = 0, last = addr + len - 1;
2056 while ((u64)len > s) {
2058 if (unlikely(ret >= iov_size)) {
2063 map = vhost_iotlb_itree_first(umem, addr, last);
2064 if (map == NULL || map->start > addr) {
2065 if (umem != dev->iotlb) {
2071 } else if (!(map->perm & access)) {
2077 size = map->size - addr + map->start;
2078 _iov->iov_len = min((u64)len - s, size);
2079 _iov->iov_base = (void __user *)(unsigned long)
2080 (map->addr + addr - map->start);
2087 vhost_iotlb_miss(vq, addr, access);
2091 /* Each buffer in the virtqueues is actually a chain of descriptors. This
2092 * function returns the next descriptor in the chain,
2093 * or -1U if we're at the end. */
2094 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2098 /* If this descriptor says it doesn't chain, we're done. */
2099 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2102 /* Check they're not leading us off end of descriptors. */
2103 next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2107 static int get_indirect(struct vhost_virtqueue *vq,
2108 struct iovec iov[], unsigned int iov_size,
2109 unsigned int *out_num, unsigned int *in_num,
2110 struct vhost_log *log, unsigned int *log_num,
2111 struct vring_desc *indirect)
2113 struct vring_desc desc;
2114 unsigned int i = 0, count, found = 0;
2115 u32 len = vhost32_to_cpu(vq, indirect->len);
2116 struct iov_iter from;
2120 if (unlikely(len % sizeof desc)) {
2121 vq_err(vq, "Invalid length in indirect descriptor: "
2122 "len 0x%llx not multiple of 0x%zx\n",
2123 (unsigned long long)len,
2128 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2129 UIO_MAXIOV, VHOST_ACCESS_RO);
2130 if (unlikely(ret < 0)) {
2132 vq_err(vq, "Translation failure %d in indirect.\n", ret);
2135 iov_iter_init(&from, ITER_SOURCE, vq->indirect, ret, len);
2136 count = len / sizeof desc;
2137 /* Buffers are chained via a 16 bit next field, so
2138 * we can have at most 2^16 of these. */
2139 if (unlikely(count > USHRT_MAX + 1)) {
2140 vq_err(vq, "Indirect buffer length too big: %d\n",
2146 unsigned iov_count = *in_num + *out_num;
2147 if (unlikely(++found > count)) {
2148 vq_err(vq, "Loop detected: last one at %u "
2149 "indirect size %u\n",
2153 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2154 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2155 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2158 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2159 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2160 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2164 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2165 access = VHOST_ACCESS_WO;
2167 access = VHOST_ACCESS_RO;
2169 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2170 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2171 iov_size - iov_count, access);
2172 if (unlikely(ret < 0)) {
2174 vq_err(vq, "Translation failure %d indirect idx %d\n",
2178 /* If this is an input descriptor, increment that count. */
2179 if (access == VHOST_ACCESS_WO) {
2181 if (unlikely(log && ret)) {
2182 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2183 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2187 /* If it's an output descriptor, they're all supposed
2188 * to come before any input descriptors. */
2189 if (unlikely(*in_num)) {
2190 vq_err(vq, "Indirect descriptor "
2191 "has out after in: idx %d\n", i);
2196 } while ((i = next_desc(vq, &desc)) != -1);
2200 /* This looks in the virtqueue and for the first available buffer, and converts
2201 * it to an iovec for convenient access. Since descriptors consist of some
2202 * number of output then some number of input descriptors, it's actually two
2203 * iovecs, but we pack them into one and note how many of each there were.
2205 * This function returns the descriptor number found, or vq->num (which is
2206 * never a valid descriptor number) if none was found. A negative code is
2207 * returned on error. */
2208 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2209 struct iovec iov[], unsigned int iov_size,
2210 unsigned int *out_num, unsigned int *in_num,
2211 struct vhost_log *log, unsigned int *log_num)
2213 struct vring_desc desc;
2214 unsigned int i, head, found = 0;
2216 __virtio16 avail_idx;
2217 __virtio16 ring_head;
2220 /* Check it isn't doing very strange things with descriptor numbers. */
2221 last_avail_idx = vq->last_avail_idx;
2223 if (vq->avail_idx == vq->last_avail_idx) {
2224 if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2225 vq_err(vq, "Failed to access avail idx at %p\n",
2229 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2231 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2232 vq_err(vq, "Guest moved used index from %u to %u",
2233 last_avail_idx, vq->avail_idx);
2237 /* If there's nothing new since last we looked, return
2240 if (vq->avail_idx == last_avail_idx)
2243 /* Only get avail ring entries after they have been
2249 /* Grab the next descriptor number they're advertising, and increment
2250 * the index we've seen. */
2251 if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2252 vq_err(vq, "Failed to read head: idx %d address %p\n",
2254 &vq->avail->ring[last_avail_idx % vq->num]);
2258 head = vhost16_to_cpu(vq, ring_head);
2260 /* If their number is silly, that's an error. */
2261 if (unlikely(head >= vq->num)) {
2262 vq_err(vq, "Guest says index %u > %u is available",
2267 /* When we start there are none of either input nor output. */
2268 *out_num = *in_num = 0;
2274 unsigned iov_count = *in_num + *out_num;
2275 if (unlikely(i >= vq->num)) {
2276 vq_err(vq, "Desc index is %u > %u, head = %u",
2280 if (unlikely(++found > vq->num)) {
2281 vq_err(vq, "Loop detected: last one at %u "
2282 "vq size %u head %u\n",
2286 ret = vhost_get_desc(vq, &desc, i);
2287 if (unlikely(ret)) {
2288 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2292 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2293 ret = get_indirect(vq, iov, iov_size,
2295 log, log_num, &desc);
2296 if (unlikely(ret < 0)) {
2298 vq_err(vq, "Failure detected "
2299 "in indirect descriptor at idx %d\n", i);
2305 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2306 access = VHOST_ACCESS_WO;
2308 access = VHOST_ACCESS_RO;
2309 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2310 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2311 iov_size - iov_count, access);
2312 if (unlikely(ret < 0)) {
2314 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2318 if (access == VHOST_ACCESS_WO) {
2319 /* If this is an input descriptor,
2320 * increment that count. */
2322 if (unlikely(log && ret)) {
2323 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2324 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2328 /* If it's an output descriptor, they're all supposed
2329 * to come before any input descriptors. */
2330 if (unlikely(*in_num)) {
2331 vq_err(vq, "Descriptor has out after in: "
2337 } while ((i = next_desc(vq, &desc)) != -1);
2339 /* On success, increment avail index. */
2340 vq->last_avail_idx++;
2342 /* Assume notifications from guest are disabled at this point,
2343 * if they aren't we would need to update avail_event index. */
2344 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2347 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2349 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2350 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2352 vq->last_avail_idx -= n;
2354 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2356 /* After we've used one of their buffers, we tell them about it. We'll then
2357 * want to notify the guest, using eventfd. */
2358 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2360 struct vring_used_elem heads = {
2361 cpu_to_vhost32(vq, head),
2362 cpu_to_vhost32(vq, len)
2365 return vhost_add_used_n(vq, &heads, 1);
2367 EXPORT_SYMBOL_GPL(vhost_add_used);
2369 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2370 struct vring_used_elem *heads,
2373 vring_used_elem_t __user *used;
2377 start = vq->last_used_idx & (vq->num - 1);
2378 used = vq->used->ring + start;
2379 if (vhost_put_used(vq, heads, start, count)) {
2380 vq_err(vq, "Failed to write used");
2383 if (unlikely(vq->log_used)) {
2384 /* Make sure data is seen before log. */
2386 /* Log used ring entry write. */
2387 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2388 count * sizeof *used);
2390 old = vq->last_used_idx;
2391 new = (vq->last_used_idx += count);
2392 /* If the driver never bothers to signal in a very long while,
2393 * used index might wrap around. If that happens, invalidate
2394 * signalled_used index we stored. TODO: make sure driver
2395 * signals at least once in 2^16 and remove this. */
2396 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2397 vq->signalled_used_valid = false;
2401 /* After we've used one of their buffers, we tell them about it. We'll then
2402 * want to notify the guest, using eventfd. */
2403 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2408 start = vq->last_used_idx & (vq->num - 1);
2409 n = vq->num - start;
2411 r = __vhost_add_used_n(vq, heads, n);
2417 r = __vhost_add_used_n(vq, heads, count);
2419 /* Make sure buffer is written before we update index. */
2421 if (vhost_put_used_idx(vq)) {
2422 vq_err(vq, "Failed to increment used idx");
2425 if (unlikely(vq->log_used)) {
2426 /* Make sure used idx is seen before log. */
2428 /* Log used index update. */
2429 log_used(vq, offsetof(struct vring_used, idx),
2430 sizeof vq->used->idx);
2432 eventfd_signal(vq->log_ctx, 1);
2436 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2438 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2443 /* Flush out used index updates. This is paired
2444 * with the barrier that the Guest executes when enabling
2448 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2449 unlikely(vq->avail_idx == vq->last_avail_idx))
2452 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2454 if (vhost_get_avail_flags(vq, &flags)) {
2455 vq_err(vq, "Failed to get flags");
2458 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2460 old = vq->signalled_used;
2461 v = vq->signalled_used_valid;
2462 new = vq->signalled_used = vq->last_used_idx;
2463 vq->signalled_used_valid = true;
2468 if (vhost_get_used_event(vq, &event)) {
2469 vq_err(vq, "Failed to get used event idx");
2472 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2475 /* This actually signals the guest, using eventfd. */
2476 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2478 /* Signal the Guest tell them we used something up. */
2479 if (vq->call_ctx.ctx && vhost_notify(dev, vq))
2480 eventfd_signal(vq->call_ctx.ctx, 1);
2482 EXPORT_SYMBOL_GPL(vhost_signal);
2484 /* And here's the combo meal deal. Supersize me! */
2485 void vhost_add_used_and_signal(struct vhost_dev *dev,
2486 struct vhost_virtqueue *vq,
2487 unsigned int head, int len)
2489 vhost_add_used(vq, head, len);
2490 vhost_signal(dev, vq);
2492 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2494 /* multi-buffer version of vhost_add_used_and_signal */
2495 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2496 struct vhost_virtqueue *vq,
2497 struct vring_used_elem *heads, unsigned count)
2499 vhost_add_used_n(vq, heads, count);
2500 vhost_signal(dev, vq);
2502 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2504 /* return true if we're sure that avaiable ring is empty */
2505 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2507 __virtio16 avail_idx;
2510 if (vq->avail_idx != vq->last_avail_idx)
2513 r = vhost_get_avail_idx(vq, &avail_idx);
2516 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2518 return vq->avail_idx == vq->last_avail_idx;
2520 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2522 /* OK, now we need to know about added descriptors. */
2523 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2525 __virtio16 avail_idx;
2528 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2530 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2531 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2532 r = vhost_update_used_flags(vq);
2534 vq_err(vq, "Failed to enable notification at %p: %d\n",
2535 &vq->used->flags, r);
2539 r = vhost_update_avail_event(vq);
2541 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2542 vhost_avail_event(vq), r);
2546 /* They could have slipped one in as we were doing that: make
2547 * sure it's written, then check again. */
2549 r = vhost_get_avail_idx(vq, &avail_idx);
2551 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2552 &vq->avail->idx, r);
2555 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2557 return vq->avail_idx != vq->last_avail_idx;
2559 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2561 /* We don't need to be notified again. */
2562 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2566 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2568 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2569 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2570 r = vhost_update_used_flags(vq);
2572 vq_err(vq, "Failed to disable notification at %p: %d\n",
2573 &vq->used->flags, r);
2576 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2578 /* Create a new message. */
2579 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2581 struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2585 /* Make sure all padding within the structure is initialized. */
2586 memset(&node->msg, 0, sizeof node->msg);
2588 node->msg.type = type;
2591 EXPORT_SYMBOL_GPL(vhost_new_msg);
2593 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2594 struct vhost_msg_node *node)
2596 spin_lock(&dev->iotlb_lock);
2597 list_add_tail(&node->node, head);
2598 spin_unlock(&dev->iotlb_lock);
2600 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2602 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2604 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2605 struct list_head *head)
2607 struct vhost_msg_node *node = NULL;
2609 spin_lock(&dev->iotlb_lock);
2610 if (!list_empty(head)) {
2611 node = list_first_entry(head, struct vhost_msg_node,
2613 list_del(&node->node);
2615 spin_unlock(&dev->iotlb_lock);
2619 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2621 void vhost_set_backend_features(struct vhost_dev *dev, u64 features)
2623 struct vhost_virtqueue *vq;
2626 mutex_lock(&dev->mutex);
2627 for (i = 0; i < dev->nvqs; ++i) {
2629 mutex_lock(&vq->mutex);
2630 vq->acked_backend_features = features;
2631 mutex_unlock(&vq->mutex);
2633 mutex_unlock(&dev->mutex);
2635 EXPORT_SYMBOL_GPL(vhost_set_backend_features);
2637 static int __init vhost_init(void)
2642 static void __exit vhost_exit(void)
2646 module_init(vhost_init);
2647 module_exit(vhost_exit);
2649 MODULE_VERSION("0.0.1");
2650 MODULE_LICENSE("GPL v2");
2651 MODULE_AUTHOR("Michael S. Tsirkin");
2652 MODULE_DESCRIPTION("Host kernel accelerator for virtio");