1 /* Copyright (C) 2009 Red Hat, Inc.
2 * Copyright (C) 2006 Rusty Russell IBM Corporation
4 * Author: Michael S. Tsirkin <mst@redhat.com>
6 * Inspiration, some code, and most witty comments come from
7 * Documentation/virtual/lguest/lguest.c, by Rusty Russell
9 * This work is licensed under the terms of the GNU GPL, version 2.
11 * Generic code for virtio server in host kernel.
14 #include <linux/eventfd.h>
15 #include <linux/vhost.h>
16 #include <linux/uio.h>
18 #include <linux/mmu_context.h>
19 #include <linux/miscdevice.h>
20 #include <linux/mutex.h>
21 #include <linux/poll.h>
22 #include <linux/file.h>
23 #include <linux/highmem.h>
24 #include <linux/slab.h>
25 #include <linux/vmalloc.h>
26 #include <linux/kthread.h>
27 #include <linux/cgroup.h>
28 #include <linux/module.h>
29 #include <linux/sort.h>
30 #include <linux/sched/mm.h>
31 #include <linux/sched/signal.h>
32 #include <linux/interval_tree_generic.h>
33 #include <linux/nospec.h>
37 static ushort max_mem_regions
= 64;
38 module_param(max_mem_regions
, ushort
, 0444);
39 MODULE_PARM_DESC(max_mem_regions
,
40 "Maximum number of memory regions in memory map. (default: 64)");
41 static int max_iotlb_entries
= 2048;
42 module_param(max_iotlb_entries
, int, 0444);
43 MODULE_PARM_DESC(max_iotlb_entries
,
44 "Maximum number of iotlb entries. (default: 2048)");
47 VHOST_MEMORY_F_LOG
= 0x1,
50 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
51 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
53 INTERVAL_TREE_DEFINE(struct vhost_umem_node
,
54 rb
, __u64
, __subtree_last
,
55 START
, LAST
, static inline, vhost_umem_interval_tree
);
57 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
58 static void vhost_disable_cross_endian(struct vhost_virtqueue
*vq
)
60 vq
->user_be
= !virtio_legacy_is_little_endian();
63 static void vhost_enable_cross_endian_big(struct vhost_virtqueue
*vq
)
68 static void vhost_enable_cross_endian_little(struct vhost_virtqueue
*vq
)
73 static long vhost_set_vring_endian(struct vhost_virtqueue
*vq
, int __user
*argp
)
75 struct vhost_vring_state s
;
80 if (copy_from_user(&s
, argp
, sizeof(s
)))
83 if (s
.num
!= VHOST_VRING_LITTLE_ENDIAN
&&
84 s
.num
!= VHOST_VRING_BIG_ENDIAN
)
87 if (s
.num
== VHOST_VRING_BIG_ENDIAN
)
88 vhost_enable_cross_endian_big(vq
);
90 vhost_enable_cross_endian_little(vq
);
95 static long vhost_get_vring_endian(struct vhost_virtqueue
*vq
, u32 idx
,
98 struct vhost_vring_state s
= {
103 if (copy_to_user(argp
, &s
, sizeof(s
)))
109 static void vhost_init_is_le(struct vhost_virtqueue
*vq
)
111 /* Note for legacy virtio: user_be is initialized at reset time
112 * according to the host endianness. If userspace does not set an
113 * explicit endianness, the default behavior is native endian, as
114 * expected by legacy virtio.
116 vq
->is_le
= vhost_has_feature(vq
, VIRTIO_F_VERSION_1
) || !vq
->user_be
;
119 static void vhost_disable_cross_endian(struct vhost_virtqueue
*vq
)
123 static long vhost_set_vring_endian(struct vhost_virtqueue
*vq
, int __user
*argp
)
128 static long vhost_get_vring_endian(struct vhost_virtqueue
*vq
, u32 idx
,
134 static void vhost_init_is_le(struct vhost_virtqueue
*vq
)
136 vq
->is_le
= vhost_has_feature(vq
, VIRTIO_F_VERSION_1
)
137 || virtio_legacy_is_little_endian();
139 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
141 static void vhost_reset_is_le(struct vhost_virtqueue
*vq
)
143 vhost_init_is_le(vq
);
146 struct vhost_flush_struct
{
147 struct vhost_work work
;
148 struct completion wait_event
;
151 static void vhost_flush_work(struct vhost_work
*work
)
153 struct vhost_flush_struct
*s
;
155 s
= container_of(work
, struct vhost_flush_struct
, work
);
156 complete(&s
->wait_event
);
159 static void vhost_poll_func(struct file
*file
, wait_queue_head_t
*wqh
,
162 struct vhost_poll
*poll
;
164 poll
= container_of(pt
, struct vhost_poll
, table
);
166 add_wait_queue(wqh
, &poll
->wait
);
169 static int vhost_poll_wakeup(wait_queue_entry_t
*wait
, unsigned mode
, int sync
,
172 struct vhost_poll
*poll
= container_of(wait
, struct vhost_poll
, wait
);
174 if (!((unsigned long)key
& poll
->mask
))
177 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
);
185 init_waitqueue_head(&work
->done
);
187 EXPORT_SYMBOL_GPL(vhost_work_init
);
189 /* Init poll structure */
190 void vhost_poll_init(struct vhost_poll
*poll
, vhost_work_fn_t fn
,
191 unsigned long mask
, struct vhost_dev
*dev
)
193 init_waitqueue_func_entry(&poll
->wait
, vhost_poll_wakeup
);
194 init_poll_funcptr(&poll
->table
, vhost_poll_func
);
199 vhost_work_init(&poll
->work
, fn
);
201 EXPORT_SYMBOL_GPL(vhost_poll_init
);
203 /* Start polling a file. We add ourselves to file's wait queue. The caller must
204 * keep a reference to a file until after vhost_poll_stop is called. */
205 int vhost_poll_start(struct vhost_poll
*poll
, struct file
*file
)
213 mask
= file
->f_op
->poll(file
, &poll
->table
);
215 vhost_poll_wakeup(&poll
->wait
, 0, 0, (void *)mask
);
216 if (mask
& POLLERR
) {
217 vhost_poll_stop(poll
);
223 EXPORT_SYMBOL_GPL(vhost_poll_start
);
225 /* Stop polling a file. After this function returns, it becomes safe to drop the
226 * file reference. You must also flush afterwards. */
227 void vhost_poll_stop(struct vhost_poll
*poll
)
230 remove_wait_queue(poll
->wqh
, &poll
->wait
);
234 EXPORT_SYMBOL_GPL(vhost_poll_stop
);
236 void vhost_work_flush(struct vhost_dev
*dev
, struct vhost_work
*work
)
238 struct vhost_flush_struct flush
;
241 init_completion(&flush
.wait_event
);
242 vhost_work_init(&flush
.work
, vhost_flush_work
);
244 vhost_work_queue(dev
, &flush
.work
);
245 wait_for_completion(&flush
.wait_event
);
248 EXPORT_SYMBOL_GPL(vhost_work_flush
);
250 /* Flush any work that has been scheduled. When calling this, don't hold any
251 * locks that are also used by the callback. */
252 void vhost_poll_flush(struct vhost_poll
*poll
)
254 vhost_work_flush(poll
->dev
, &poll
->work
);
256 EXPORT_SYMBOL_GPL(vhost_poll_flush
);
258 void vhost_work_queue(struct vhost_dev
*dev
, struct vhost_work
*work
)
263 if (!test_and_set_bit(VHOST_WORK_QUEUED
, &work
->flags
)) {
264 /* We can only add the work to the list after we're
265 * sure it was not in the list.
266 * test_and_set_bit() implies a memory barrier.
268 llist_add(&work
->node
, &dev
->work_list
);
269 wake_up_process(dev
->worker
);
272 EXPORT_SYMBOL_GPL(vhost_work_queue
);
274 /* A lockless hint for busy polling code to exit the loop */
275 bool vhost_has_work(struct vhost_dev
*dev
)
277 return !llist_empty(&dev
->work_list
);
279 EXPORT_SYMBOL_GPL(vhost_has_work
);
281 void vhost_poll_queue(struct vhost_poll
*poll
)
283 vhost_work_queue(poll
->dev
, &poll
->work
);
285 EXPORT_SYMBOL_GPL(vhost_poll_queue
);
287 static void __vhost_vq_meta_reset(struct vhost_virtqueue
*vq
)
291 for (j
= 0; j
< VHOST_NUM_ADDRS
; j
++)
292 vq
->meta_iotlb
[j
] = NULL
;
295 static void vhost_vq_meta_reset(struct vhost_dev
*d
)
299 for (i
= 0; i
< d
->nvqs
; ++i
)
300 __vhost_vq_meta_reset(d
->vqs
[i
]);
303 static void vhost_vq_reset(struct vhost_dev
*dev
,
304 struct vhost_virtqueue
*vq
)
310 vq
->last_avail_idx
= 0;
312 vq
->last_used_idx
= 0;
313 vq
->signalled_used
= 0;
314 vq
->signalled_used_valid
= false;
316 vq
->log_used
= false;
317 vq
->log_addr
= -1ull;
318 vq
->private_data
= NULL
;
319 vq
->acked_features
= 0;
321 vq
->error_ctx
= NULL
;
327 vhost_reset_is_le(vq
);
328 vhost_disable_cross_endian(vq
);
329 vq
->busyloop_timeout
= 0;
332 __vhost_vq_meta_reset(vq
);
335 static int vhost_worker(void *data
)
337 struct vhost_dev
*dev
= data
;
338 struct vhost_work
*work
, *work_next
;
339 struct llist_node
*node
;
340 mm_segment_t oldfs
= get_fs();
346 /* mb paired w/ kthread_stop */
347 set_current_state(TASK_INTERRUPTIBLE
);
349 if (kthread_should_stop()) {
350 __set_current_state(TASK_RUNNING
);
354 node
= llist_del_all(&dev
->work_list
);
358 node
= llist_reverse_order(node
);
359 /* make sure flag is seen after deletion */
361 llist_for_each_entry_safe(work
, work_next
, node
, node
) {
362 clear_bit(VHOST_WORK_QUEUED
, &work
->flags
);
363 __set_current_state(TASK_RUNNING
);
374 static void vhost_vq_free_iovecs(struct vhost_virtqueue
*vq
)
384 /* Helper to allocate iovec buffers for all vqs. */
385 static long vhost_dev_alloc_iovecs(struct vhost_dev
*dev
)
387 struct vhost_virtqueue
*vq
;
390 for (i
= 0; i
< dev
->nvqs
; ++i
) {
392 vq
->indirect
= kmalloc(sizeof *vq
->indirect
* UIO_MAXIOV
,
394 vq
->log
= kmalloc(sizeof *vq
->log
* UIO_MAXIOV
, GFP_KERNEL
);
395 vq
->heads
= kmalloc(sizeof *vq
->heads
* UIO_MAXIOV
, GFP_KERNEL
);
396 if (!vq
->indirect
|| !vq
->log
|| !vq
->heads
)
403 vhost_vq_free_iovecs(dev
->vqs
[i
]);
407 static void vhost_dev_free_iovecs(struct vhost_dev
*dev
)
411 for (i
= 0; i
< dev
->nvqs
; ++i
)
412 vhost_vq_free_iovecs(dev
->vqs
[i
]);
415 void vhost_dev_init(struct vhost_dev
*dev
,
416 struct vhost_virtqueue
**vqs
, int nvqs
)
418 struct vhost_virtqueue
*vq
;
423 mutex_init(&dev
->mutex
);
425 dev
->log_file
= NULL
;
430 init_llist_head(&dev
->work_list
);
431 init_waitqueue_head(&dev
->wait
);
432 INIT_LIST_HEAD(&dev
->read_list
);
433 INIT_LIST_HEAD(&dev
->pending_list
);
434 spin_lock_init(&dev
->iotlb_lock
);
437 for (i
= 0; i
< dev
->nvqs
; ++i
) {
443 mutex_init(&vq
->mutex
);
444 vhost_vq_reset(dev
, vq
);
446 vhost_poll_init(&vq
->poll
, vq
->handle_kick
,
450 EXPORT_SYMBOL_GPL(vhost_dev_init
);
452 /* Caller should have device mutex */
453 long vhost_dev_check_owner(struct vhost_dev
*dev
)
455 /* Are you the owner? If not, I don't think you mean to do that */
456 return dev
->mm
== current
->mm
? 0 : -EPERM
;
458 EXPORT_SYMBOL_GPL(vhost_dev_check_owner
);
460 struct vhost_attach_cgroups_struct
{
461 struct vhost_work work
;
462 struct task_struct
*owner
;
466 static void vhost_attach_cgroups_work(struct vhost_work
*work
)
468 struct vhost_attach_cgroups_struct
*s
;
470 s
= container_of(work
, struct vhost_attach_cgroups_struct
, work
);
471 s
->ret
= cgroup_attach_task_all(s
->owner
, current
);
474 static int vhost_attach_cgroups(struct vhost_dev
*dev
)
476 struct vhost_attach_cgroups_struct attach
;
478 attach
.owner
= current
;
479 vhost_work_init(&attach
.work
, vhost_attach_cgroups_work
);
480 vhost_work_queue(dev
, &attach
.work
);
481 vhost_work_flush(dev
, &attach
.work
);
485 /* Caller should have device mutex */
486 bool vhost_dev_has_owner(struct vhost_dev
*dev
)
490 EXPORT_SYMBOL_GPL(vhost_dev_has_owner
);
492 /* Caller should have device mutex */
493 long vhost_dev_set_owner(struct vhost_dev
*dev
)
495 struct task_struct
*worker
;
498 /* Is there an owner already? */
499 if (vhost_dev_has_owner(dev
)) {
504 /* No owner, become one */
505 dev
->mm
= get_task_mm(current
);
506 worker
= kthread_create(vhost_worker
, dev
, "vhost-%d", current
->pid
);
507 if (IS_ERR(worker
)) {
508 err
= PTR_ERR(worker
);
512 dev
->worker
= worker
;
513 wake_up_process(worker
); /* avoid contributing to loadavg */
515 err
= vhost_attach_cgroups(dev
);
519 err
= vhost_dev_alloc_iovecs(dev
);
525 kthread_stop(worker
);
534 EXPORT_SYMBOL_GPL(vhost_dev_set_owner
);
536 struct vhost_umem
*vhost_dev_reset_owner_prepare(void)
538 return kvzalloc(sizeof(struct vhost_umem
), GFP_KERNEL
);
540 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare
);
542 /* Caller should have device mutex */
543 void vhost_dev_reset_owner(struct vhost_dev
*dev
, struct vhost_umem
*umem
)
547 vhost_dev_cleanup(dev
, true);
549 /* Restore memory to default empty mapping. */
550 INIT_LIST_HEAD(&umem
->umem_list
);
552 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
553 * VQs aren't running.
555 for (i
= 0; i
< dev
->nvqs
; ++i
)
556 dev
->vqs
[i
]->umem
= umem
;
558 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner
);
560 void vhost_dev_stop(struct vhost_dev
*dev
)
564 for (i
= 0; i
< dev
->nvqs
; ++i
) {
565 if (dev
->vqs
[i
]->kick
&& dev
->vqs
[i
]->handle_kick
) {
566 vhost_poll_stop(&dev
->vqs
[i
]->poll
);
567 vhost_poll_flush(&dev
->vqs
[i
]->poll
);
571 EXPORT_SYMBOL_GPL(vhost_dev_stop
);
573 static void vhost_umem_free(struct vhost_umem
*umem
,
574 struct vhost_umem_node
*node
)
576 vhost_umem_interval_tree_remove(node
, &umem
->umem_tree
);
577 list_del(&node
->link
);
582 static void vhost_umem_clean(struct vhost_umem
*umem
)
584 struct vhost_umem_node
*node
, *tmp
;
589 list_for_each_entry_safe(node
, tmp
, &umem
->umem_list
, link
)
590 vhost_umem_free(umem
, node
);
595 static void vhost_clear_msg(struct vhost_dev
*dev
)
597 struct vhost_msg_node
*node
, *n
;
599 spin_lock(&dev
->iotlb_lock
);
601 list_for_each_entry_safe(node
, n
, &dev
->read_list
, node
) {
602 list_del(&node
->node
);
606 list_for_each_entry_safe(node
, n
, &dev
->pending_list
, node
) {
607 list_del(&node
->node
);
611 spin_unlock(&dev
->iotlb_lock
);
614 /* Caller should have device mutex if and only if locked is set */
615 void vhost_dev_cleanup(struct vhost_dev
*dev
, bool locked
)
619 for (i
= 0; i
< dev
->nvqs
; ++i
) {
620 if (dev
->vqs
[i
]->error_ctx
)
621 eventfd_ctx_put(dev
->vqs
[i
]->error_ctx
);
622 if (dev
->vqs
[i
]->error
)
623 fput(dev
->vqs
[i
]->error
);
624 if (dev
->vqs
[i
]->kick
)
625 fput(dev
->vqs
[i
]->kick
);
626 if (dev
->vqs
[i
]->call_ctx
)
627 eventfd_ctx_put(dev
->vqs
[i
]->call_ctx
);
628 if (dev
->vqs
[i
]->call
)
629 fput(dev
->vqs
[i
]->call
);
630 vhost_vq_reset(dev
, dev
->vqs
[i
]);
632 vhost_dev_free_iovecs(dev
);
634 eventfd_ctx_put(dev
->log_ctx
);
638 dev
->log_file
= NULL
;
639 /* No one will access memory at this point */
640 vhost_umem_clean(dev
->umem
);
642 vhost_umem_clean(dev
->iotlb
);
644 vhost_clear_msg(dev
);
645 wake_up_interruptible_poll(&dev
->wait
, POLLIN
| POLLRDNORM
);
646 WARN_ON(!llist_empty(&dev
->work_list
));
648 kthread_stop(dev
->worker
);
655 EXPORT_SYMBOL_GPL(vhost_dev_cleanup
);
657 static int log_access_ok(void __user
*log_base
, u64 addr
, unsigned long sz
)
659 u64 a
= addr
/ VHOST_PAGE_SIZE
/ 8;
661 /* Make sure 64 bit math will not overflow. */
662 if (a
> ULONG_MAX
- (unsigned long)log_base
||
663 a
+ (unsigned long)log_base
> ULONG_MAX
)
666 return access_ok(VERIFY_WRITE
, log_base
+ a
,
667 (sz
+ VHOST_PAGE_SIZE
* 8 - 1) / VHOST_PAGE_SIZE
/ 8);
670 static bool vhost_overflow(u64 uaddr
, u64 size
)
672 /* Make sure 64 bit math will not overflow. */
673 return uaddr
> ULONG_MAX
|| size
> ULONG_MAX
|| uaddr
> ULONG_MAX
- size
;
676 /* Caller should have vq mutex and device mutex. */
677 static int vq_memory_access_ok(void __user
*log_base
, struct vhost_umem
*umem
,
680 struct vhost_umem_node
*node
;
685 list_for_each_entry(node
, &umem
->umem_list
, link
) {
686 unsigned long a
= node
->userspace_addr
;
688 if (vhost_overflow(node
->userspace_addr
, node
->size
))
692 if (!access_ok(VERIFY_WRITE
, (void __user
*)a
,
695 else if (log_all
&& !log_access_ok(log_base
,
703 static inline void __user
*vhost_vq_meta_fetch(struct vhost_virtqueue
*vq
,
704 u64 addr
, unsigned int size
,
707 const struct vhost_umem_node
*node
= vq
->meta_iotlb
[type
];
712 return (void *)(uintptr_t)(node
->userspace_addr
+ addr
- node
->start
);
715 /* Can we switch to this memory table? */
716 /* Caller should have device mutex but not vq mutex */
717 static int memory_access_ok(struct vhost_dev
*d
, struct vhost_umem
*umem
,
722 for (i
= 0; i
< d
->nvqs
; ++i
) {
726 mutex_lock(&d
->vqs
[i
]->mutex
);
727 log
= log_all
|| vhost_has_feature(d
->vqs
[i
], VHOST_F_LOG_ALL
);
728 /* If ring is inactive, will check when it's enabled. */
729 if (d
->vqs
[i
]->private_data
)
730 ok
= vq_memory_access_ok(d
->vqs
[i
]->log_base
,
734 mutex_unlock(&d
->vqs
[i
]->mutex
);
741 static int translate_desc(struct vhost_virtqueue
*vq
, u64 addr
, u32 len
,
742 struct iovec iov
[], int iov_size
, int access
);
744 static int vhost_copy_to_user(struct vhost_virtqueue
*vq
, void __user
*to
,
745 const void *from
, unsigned size
)
750 return __copy_to_user(to
, from
, size
);
752 /* This function should be called after iotlb
753 * prefetch, which means we're sure that all vq
754 * could be access through iotlb. So -EAGAIN should
755 * not happen in this case.
758 void __user
*uaddr
= vhost_vq_meta_fetch(vq
,
759 (u64
)(uintptr_t)to
, size
,
763 return __copy_to_user(uaddr
, from
, size
);
765 ret
= translate_desc(vq
, (u64
)(uintptr_t)to
, size
, vq
->iotlb_iov
,
766 ARRAY_SIZE(vq
->iotlb_iov
),
770 iov_iter_init(&t
, WRITE
, vq
->iotlb_iov
, ret
, size
);
771 ret
= copy_to_iter(from
, size
, &t
);
779 static int vhost_copy_from_user(struct vhost_virtqueue
*vq
, void *to
,
780 void __user
*from
, unsigned size
)
785 return __copy_from_user(to
, from
, size
);
787 /* This function should be called after iotlb
788 * prefetch, which means we're sure that vq
789 * could be access through iotlb. So -EAGAIN should
790 * not happen in this case.
792 void __user
*uaddr
= vhost_vq_meta_fetch(vq
,
793 (u64
)(uintptr_t)from
, size
,
798 return __copy_from_user(to
, uaddr
, size
);
800 ret
= translate_desc(vq
, (u64
)(uintptr_t)from
, size
, vq
->iotlb_iov
,
801 ARRAY_SIZE(vq
->iotlb_iov
),
804 vq_err(vq
, "IOTLB translation failure: uaddr "
805 "%p size 0x%llx\n", from
,
806 (unsigned long long) size
);
809 iov_iter_init(&f
, READ
, vq
->iotlb_iov
, ret
, size
);
810 ret
= copy_from_iter(to
, size
, &f
);
819 static void __user
*__vhost_get_user_slow(struct vhost_virtqueue
*vq
,
820 void __user
*addr
, unsigned int size
,
825 ret
= translate_desc(vq
, (u64
)(uintptr_t)addr
, size
, vq
->iotlb_iov
,
826 ARRAY_SIZE(vq
->iotlb_iov
),
829 vq_err(vq
, "IOTLB translation failure: uaddr "
830 "%p size 0x%llx\n", addr
,
831 (unsigned long long) size
);
835 if (ret
!= 1 || vq
->iotlb_iov
[0].iov_len
!= size
) {
836 vq_err(vq
, "Non atomic userspace memory access: uaddr "
837 "%p size 0x%llx\n", addr
,
838 (unsigned long long) size
);
842 return vq
->iotlb_iov
[0].iov_base
;
845 /* This function should be called after iotlb
846 * prefetch, which means we're sure that vq
847 * could be access through iotlb. So -EAGAIN should
848 * not happen in this case.
850 static inline void __user
*__vhost_get_user(struct vhost_virtqueue
*vq
,
851 void *addr
, unsigned int size
,
854 void __user
*uaddr
= vhost_vq_meta_fetch(vq
,
855 (u64
)(uintptr_t)addr
, size
, type
);
859 return __vhost_get_user_slow(vq
, addr
, size
, type
);
862 #define vhost_put_user(vq, x, ptr) \
866 ret = __put_user(x, ptr); \
868 __typeof__(ptr) to = \
869 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
870 sizeof(*ptr), VHOST_ADDR_USED); \
872 ret = __put_user(x, to); \
879 #define vhost_get_user(vq, x, ptr, type) \
883 ret = __get_user(x, ptr); \
885 __typeof__(ptr) from = \
886 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
890 ret = __get_user(x, from); \
897 #define vhost_get_avail(vq, x, ptr) \
898 vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
900 #define vhost_get_used(vq, x, ptr) \
901 vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
903 static void vhost_dev_lock_vqs(struct vhost_dev
*d
)
906 for (i
= 0; i
< d
->nvqs
; ++i
)
907 mutex_lock_nested(&d
->vqs
[i
]->mutex
, i
);
910 static void vhost_dev_unlock_vqs(struct vhost_dev
*d
)
913 for (i
= 0; i
< d
->nvqs
; ++i
)
914 mutex_unlock(&d
->vqs
[i
]->mutex
);
917 static int vhost_new_umem_range(struct vhost_umem
*umem
,
918 u64 start
, u64 size
, u64 end
,
919 u64 userspace_addr
, int perm
)
921 struct vhost_umem_node
*tmp
, *node
= kmalloc(sizeof(*node
), GFP_ATOMIC
);
926 if (umem
->numem
== max_iotlb_entries
) {
927 tmp
= list_first_entry(&umem
->umem_list
, typeof(*tmp
), link
);
928 vhost_umem_free(umem
, tmp
);
934 node
->userspace_addr
= userspace_addr
;
936 INIT_LIST_HEAD(&node
->link
);
937 list_add_tail(&node
->link
, &umem
->umem_list
);
938 vhost_umem_interval_tree_insert(node
, &umem
->umem_tree
);
944 static void vhost_del_umem_range(struct vhost_umem
*umem
,
947 struct vhost_umem_node
*node
;
949 while ((node
= vhost_umem_interval_tree_iter_first(&umem
->umem_tree
,
951 vhost_umem_free(umem
, node
);
954 static void vhost_iotlb_notify_vq(struct vhost_dev
*d
,
955 struct vhost_iotlb_msg
*msg
)
957 struct vhost_msg_node
*node
, *n
;
959 spin_lock(&d
->iotlb_lock
);
961 list_for_each_entry_safe(node
, n
, &d
->pending_list
, node
) {
962 struct vhost_iotlb_msg
*vq_msg
= &node
->msg
.iotlb
;
963 if (msg
->iova
<= vq_msg
->iova
&&
964 msg
->iova
+ msg
->size
- 1 > vq_msg
->iova
&&
965 vq_msg
->type
== VHOST_IOTLB_MISS
) {
966 vhost_poll_queue(&node
->vq
->poll
);
967 list_del(&node
->node
);
972 spin_unlock(&d
->iotlb_lock
);
975 static int umem_access_ok(u64 uaddr
, u64 size
, int access
)
977 unsigned long a
= uaddr
;
979 /* Make sure 64 bit math will not overflow. */
980 if (vhost_overflow(uaddr
, size
))
983 if ((access
& VHOST_ACCESS_RO
) &&
984 !access_ok(VERIFY_READ
, (void __user
*)a
, size
))
986 if ((access
& VHOST_ACCESS_WO
) &&
987 !access_ok(VERIFY_WRITE
, (void __user
*)a
, size
))
992 static int vhost_process_iotlb_msg(struct vhost_dev
*dev
,
993 struct vhost_iotlb_msg
*msg
)
997 mutex_lock(&dev
->mutex
);
998 vhost_dev_lock_vqs(dev
);
1000 case VHOST_IOTLB_UPDATE
:
1005 if (umem_access_ok(msg
->uaddr
, msg
->size
, msg
->perm
)) {
1009 vhost_vq_meta_reset(dev
);
1010 if (vhost_new_umem_range(dev
->iotlb
, msg
->iova
, msg
->size
,
1011 msg
->iova
+ msg
->size
- 1,
1012 msg
->uaddr
, msg
->perm
)) {
1016 vhost_iotlb_notify_vq(dev
, msg
);
1018 case VHOST_IOTLB_INVALIDATE
:
1023 vhost_vq_meta_reset(dev
);
1024 vhost_del_umem_range(dev
->iotlb
, msg
->iova
,
1025 msg
->iova
+ msg
->size
- 1);
1032 vhost_dev_unlock_vqs(dev
);
1033 mutex_unlock(&dev
->mutex
);
1037 ssize_t
vhost_chr_write_iter(struct vhost_dev
*dev
,
1038 struct iov_iter
*from
)
1040 struct vhost_msg_node node
;
1041 unsigned size
= sizeof(struct vhost_msg
);
1045 if (iov_iter_count(from
) < size
)
1047 ret
= copy_from_iter(&node
.msg
, size
, from
);
1051 switch (node
.msg
.type
) {
1052 case VHOST_IOTLB_MSG
:
1053 err
= vhost_process_iotlb_msg(dev
, &node
.msg
.iotlb
);
1065 EXPORT_SYMBOL(vhost_chr_write_iter
);
1067 unsigned int vhost_chr_poll(struct file
*file
, struct vhost_dev
*dev
,
1070 unsigned int mask
= 0;
1072 poll_wait(file
, &dev
->wait
, wait
);
1074 if (!list_empty(&dev
->read_list
))
1075 mask
|= POLLIN
| POLLRDNORM
;
1079 EXPORT_SYMBOL(vhost_chr_poll
);
1081 ssize_t
vhost_chr_read_iter(struct vhost_dev
*dev
, struct iov_iter
*to
,
1085 struct vhost_msg_node
*node
;
1087 unsigned size
= sizeof(struct vhost_msg
);
1089 if (iov_iter_count(to
) < size
)
1094 prepare_to_wait(&dev
->wait
, &wait
,
1095 TASK_INTERRUPTIBLE
);
1097 node
= vhost_dequeue_msg(dev
, &dev
->read_list
);
1104 if (signal_pending(current
)) {
1117 finish_wait(&dev
->wait
, &wait
);
1120 ret
= copy_to_iter(&node
->msg
, size
, to
);
1122 if (ret
!= size
|| node
->msg
.type
!= VHOST_IOTLB_MISS
) {
1127 vhost_enqueue_msg(dev
, &dev
->pending_list
, node
);
1132 EXPORT_SYMBOL_GPL(vhost_chr_read_iter
);
1134 static int vhost_iotlb_miss(struct vhost_virtqueue
*vq
, u64 iova
, int access
)
1136 struct vhost_dev
*dev
= vq
->dev
;
1137 struct vhost_msg_node
*node
;
1138 struct vhost_iotlb_msg
*msg
;
1140 node
= vhost_new_msg(vq
, VHOST_IOTLB_MISS
);
1144 msg
= &node
->msg
.iotlb
;
1145 msg
->type
= VHOST_IOTLB_MISS
;
1149 vhost_enqueue_msg(dev
, &dev
->read_list
, node
);
1154 static int vq_access_ok(struct vhost_virtqueue
*vq
, unsigned int num
,
1155 struct vring_desc __user
*desc
,
1156 struct vring_avail __user
*avail
,
1157 struct vring_used __user
*used
)
1160 size_t s
= vhost_has_feature(vq
, VIRTIO_RING_F_EVENT_IDX
) ? 2 : 0;
1162 return access_ok(VERIFY_READ
, desc
, num
* sizeof *desc
) &&
1163 access_ok(VERIFY_READ
, avail
,
1164 sizeof *avail
+ num
* sizeof *avail
->ring
+ s
) &&
1165 access_ok(VERIFY_WRITE
, used
,
1166 sizeof *used
+ num
* sizeof *used
->ring
+ s
);
1169 static void vhost_vq_meta_update(struct vhost_virtqueue
*vq
,
1170 const struct vhost_umem_node
*node
,
1173 int access
= (type
== VHOST_ADDR_USED
) ?
1174 VHOST_ACCESS_WO
: VHOST_ACCESS_RO
;
1176 if (likely(node
->perm
& access
))
1177 vq
->meta_iotlb
[type
] = node
;
1180 static int iotlb_access_ok(struct vhost_virtqueue
*vq
,
1181 int access
, u64 addr
, u64 len
, int type
)
1183 const struct vhost_umem_node
*node
;
1184 struct vhost_umem
*umem
= vq
->iotlb
;
1185 u64 s
= 0, size
, orig_addr
= addr
, last
= addr
+ len
- 1;
1187 if (vhost_vq_meta_fetch(vq
, addr
, len
, type
))
1191 node
= vhost_umem_interval_tree_iter_first(&umem
->umem_tree
,
1194 if (node
== NULL
|| node
->start
> addr
) {
1195 vhost_iotlb_miss(vq
, addr
, access
);
1197 } else if (!(node
->perm
& access
)) {
1198 /* Report the possible access violation by
1199 * request another translation from userspace.
1204 size
= node
->size
- addr
+ node
->start
;
1206 if (orig_addr
== addr
&& size
>= len
)
1207 vhost_vq_meta_update(vq
, node
, type
);
1216 int vq_iotlb_prefetch(struct vhost_virtqueue
*vq
)
1218 size_t s
= vhost_has_feature(vq
, VIRTIO_RING_F_EVENT_IDX
) ? 2 : 0;
1219 unsigned int num
= vq
->num
;
1224 return iotlb_access_ok(vq
, VHOST_ACCESS_RO
, (u64
)(uintptr_t)vq
->desc
,
1225 num
* sizeof(*vq
->desc
), VHOST_ADDR_DESC
) &&
1226 iotlb_access_ok(vq
, VHOST_ACCESS_RO
, (u64
)(uintptr_t)vq
->avail
,
1228 num
* sizeof(*vq
->avail
->ring
) + s
,
1229 VHOST_ADDR_AVAIL
) &&
1230 iotlb_access_ok(vq
, VHOST_ACCESS_WO
, (u64
)(uintptr_t)vq
->used
,
1232 num
* sizeof(*vq
->used
->ring
) + s
,
1235 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch
);
1237 /* Can we log writes? */
1238 /* Caller should have device mutex but not vq mutex */
1239 int vhost_log_access_ok(struct vhost_dev
*dev
)
1241 return memory_access_ok(dev
, dev
->umem
, 1);
1243 EXPORT_SYMBOL_GPL(vhost_log_access_ok
);
1245 /* Verify access for write logging. */
1246 /* Caller should have vq mutex and device mutex */
1247 static int vq_log_access_ok(struct vhost_virtqueue
*vq
,
1248 void __user
*log_base
)
1250 size_t s
= vhost_has_feature(vq
, VIRTIO_RING_F_EVENT_IDX
) ? 2 : 0;
1252 return vq_memory_access_ok(log_base
, vq
->umem
,
1253 vhost_has_feature(vq
, VHOST_F_LOG_ALL
)) &&
1254 (!vq
->log_used
|| log_access_ok(log_base
, vq
->log_addr
,
1256 vq
->num
* sizeof *vq
->used
->ring
+ s
));
1259 /* Can we start vq? */
1260 /* Caller should have vq mutex and device mutex */
1261 int vhost_vq_access_ok(struct vhost_virtqueue
*vq
)
1263 if (!vq_log_access_ok(vq
, vq
->log_base
))
1266 /* Access validation occurs at prefetch time with IOTLB */
1270 return vq_access_ok(vq
, vq
->num
, vq
->desc
, vq
->avail
, vq
->used
);
1272 EXPORT_SYMBOL_GPL(vhost_vq_access_ok
);
1274 static struct vhost_umem
*vhost_umem_alloc(void)
1276 struct vhost_umem
*umem
= kvzalloc(sizeof(*umem
), GFP_KERNEL
);
1281 umem
->umem_tree
= RB_ROOT_CACHED
;
1283 INIT_LIST_HEAD(&umem
->umem_list
);
1288 static long vhost_set_memory(struct vhost_dev
*d
, struct vhost_memory __user
*m
)
1290 struct vhost_memory mem
, *newmem
;
1291 struct vhost_memory_region
*region
;
1292 struct vhost_umem
*newumem
, *oldumem
;
1293 unsigned long size
= offsetof(struct vhost_memory
, regions
);
1296 if (copy_from_user(&mem
, m
, size
))
1300 if (mem
.nregions
> max_mem_regions
)
1302 newmem
= kvzalloc(size
+ mem
.nregions
* sizeof(*m
->regions
), GFP_KERNEL
);
1306 memcpy(newmem
, &mem
, size
);
1307 if (copy_from_user(newmem
->regions
, m
->regions
,
1308 mem
.nregions
* sizeof *m
->regions
)) {
1313 newumem
= vhost_umem_alloc();
1319 for (region
= newmem
->regions
;
1320 region
< newmem
->regions
+ mem
.nregions
;
1322 if (vhost_new_umem_range(newumem
,
1323 region
->guest_phys_addr
,
1324 region
->memory_size
,
1325 region
->guest_phys_addr
+
1326 region
->memory_size
- 1,
1327 region
->userspace_addr
,
1332 if (!memory_access_ok(d
, newumem
, 0))
1338 /* All memory accesses are done under some VQ mutex. */
1339 for (i
= 0; i
< d
->nvqs
; ++i
) {
1340 mutex_lock(&d
->vqs
[i
]->mutex
);
1341 d
->vqs
[i
]->umem
= newumem
;
1342 mutex_unlock(&d
->vqs
[i
]->mutex
);
1346 vhost_umem_clean(oldumem
);
1350 vhost_umem_clean(newumem
);
1355 long vhost_vring_ioctl(struct vhost_dev
*d
, int ioctl
, void __user
*argp
)
1357 struct file
*eventfp
, *filep
= NULL
;
1358 bool pollstart
= false, pollstop
= false;
1359 struct eventfd_ctx
*ctx
= NULL
;
1360 u32 __user
*idxp
= argp
;
1361 struct vhost_virtqueue
*vq
;
1362 struct vhost_vring_state s
;
1363 struct vhost_vring_file f
;
1364 struct vhost_vring_addr a
;
1368 r
= get_user(idx
, idxp
);
1374 idx
= array_index_nospec(idx
, d
->nvqs
);
1377 mutex_lock(&vq
->mutex
);
1380 case VHOST_SET_VRING_NUM
:
1381 /* Resizing ring with an active backend?
1382 * You don't want to do that. */
1383 if (vq
->private_data
) {
1387 if (copy_from_user(&s
, argp
, sizeof s
)) {
1391 if (!s
.num
|| s
.num
> 0xffff || (s
.num
& (s
.num
- 1))) {
1397 case VHOST_SET_VRING_BASE
:
1398 /* Moving base with an active backend?
1399 * You don't want to do that. */
1400 if (vq
->private_data
) {
1404 if (copy_from_user(&s
, argp
, sizeof s
)) {
1408 if (s
.num
> 0xffff) {
1412 vq
->last_avail_idx
= s
.num
;
1413 /* Forget the cached index value. */
1414 vq
->avail_idx
= vq
->last_avail_idx
;
1416 case VHOST_GET_VRING_BASE
:
1418 s
.num
= vq
->last_avail_idx
;
1419 if (copy_to_user(argp
, &s
, sizeof s
))
1422 case VHOST_SET_VRING_ADDR
:
1423 if (copy_from_user(&a
, argp
, sizeof a
)) {
1427 if (a
.flags
& ~(0x1 << VHOST_VRING_F_LOG
)) {
1431 /* For 32bit, verify that the top 32bits of the user
1432 data are set to zero. */
1433 if ((u64
)(unsigned long)a
.desc_user_addr
!= a
.desc_user_addr
||
1434 (u64
)(unsigned long)a
.used_user_addr
!= a
.used_user_addr
||
1435 (u64
)(unsigned long)a
.avail_user_addr
!= a
.avail_user_addr
) {
1440 /* Make sure it's safe to cast pointers to vring types. */
1441 BUILD_BUG_ON(__alignof__
*vq
->avail
> VRING_AVAIL_ALIGN_SIZE
);
1442 BUILD_BUG_ON(__alignof__
*vq
->used
> VRING_USED_ALIGN_SIZE
);
1443 if ((a
.avail_user_addr
& (VRING_AVAIL_ALIGN_SIZE
- 1)) ||
1444 (a
.used_user_addr
& (VRING_USED_ALIGN_SIZE
- 1)) ||
1445 (a
.log_guest_addr
& (VRING_USED_ALIGN_SIZE
- 1))) {
1450 /* We only verify access here if backend is configured.
1451 * If it is not, we don't as size might not have been setup.
1452 * We will verify when backend is configured. */
1453 if (vq
->private_data
) {
1454 if (!vq_access_ok(vq
, vq
->num
,
1455 (void __user
*)(unsigned long)a
.desc_user_addr
,
1456 (void __user
*)(unsigned long)a
.avail_user_addr
,
1457 (void __user
*)(unsigned long)a
.used_user_addr
)) {
1462 /* Also validate log access for used ring if enabled. */
1463 if ((a
.flags
& (0x1 << VHOST_VRING_F_LOG
)) &&
1464 !log_access_ok(vq
->log_base
, a
.log_guest_addr
,
1466 vq
->num
* sizeof *vq
->used
->ring
)) {
1472 vq
->log_used
= !!(a
.flags
& (0x1 << VHOST_VRING_F_LOG
));
1473 vq
->desc
= (void __user
*)(unsigned long)a
.desc_user_addr
;
1474 vq
->avail
= (void __user
*)(unsigned long)a
.avail_user_addr
;
1475 vq
->log_addr
= a
.log_guest_addr
;
1476 vq
->used
= (void __user
*)(unsigned long)a
.used_user_addr
;
1478 case VHOST_SET_VRING_KICK
:
1479 if (copy_from_user(&f
, argp
, sizeof f
)) {
1483 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
1484 if (IS_ERR(eventfp
)) {
1485 r
= PTR_ERR(eventfp
);
1488 if (eventfp
!= vq
->kick
) {
1489 pollstop
= (filep
= vq
->kick
) != NULL
;
1490 pollstart
= (vq
->kick
= eventfp
) != NULL
;
1494 case VHOST_SET_VRING_CALL
:
1495 if (copy_from_user(&f
, argp
, sizeof f
)) {
1499 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
1500 if (IS_ERR(eventfp
)) {
1501 r
= PTR_ERR(eventfp
);
1504 if (eventfp
!= vq
->call
) {
1508 vq
->call_ctx
= eventfp
?
1509 eventfd_ctx_fileget(eventfp
) : NULL
;
1513 case VHOST_SET_VRING_ERR
:
1514 if (copy_from_user(&f
, argp
, sizeof f
)) {
1518 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
1519 if (IS_ERR(eventfp
)) {
1520 r
= PTR_ERR(eventfp
);
1523 if (eventfp
!= vq
->error
) {
1525 vq
->error
= eventfp
;
1526 ctx
= vq
->error_ctx
;
1527 vq
->error_ctx
= eventfp
?
1528 eventfd_ctx_fileget(eventfp
) : NULL
;
1532 case VHOST_SET_VRING_ENDIAN
:
1533 r
= vhost_set_vring_endian(vq
, argp
);
1535 case VHOST_GET_VRING_ENDIAN
:
1536 r
= vhost_get_vring_endian(vq
, idx
, argp
);
1538 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT
:
1539 if (copy_from_user(&s
, argp
, sizeof(s
))) {
1543 vq
->busyloop_timeout
= s
.num
;
1545 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT
:
1547 s
.num
= vq
->busyloop_timeout
;
1548 if (copy_to_user(argp
, &s
, sizeof(s
)))
1555 if (pollstop
&& vq
->handle_kick
)
1556 vhost_poll_stop(&vq
->poll
);
1559 eventfd_ctx_put(ctx
);
1563 if (pollstart
&& vq
->handle_kick
)
1564 r
= vhost_poll_start(&vq
->poll
, vq
->kick
);
1566 mutex_unlock(&vq
->mutex
);
1568 if (pollstop
&& vq
->handle_kick
)
1569 vhost_poll_flush(&vq
->poll
);
1572 EXPORT_SYMBOL_GPL(vhost_vring_ioctl
);
1574 int vhost_init_device_iotlb(struct vhost_dev
*d
, bool enabled
)
1576 struct vhost_umem
*niotlb
, *oiotlb
;
1579 niotlb
= vhost_umem_alloc();
1586 for (i
= 0; i
< d
->nvqs
; ++i
) {
1587 struct vhost_virtqueue
*vq
= d
->vqs
[i
];
1589 mutex_lock(&vq
->mutex
);
1591 __vhost_vq_meta_reset(vq
);
1592 mutex_unlock(&vq
->mutex
);
1595 vhost_umem_clean(oiotlb
);
1599 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb
);
1601 /* Caller must have device mutex */
1602 long vhost_dev_ioctl(struct vhost_dev
*d
, unsigned int ioctl
, void __user
*argp
)
1604 struct file
*eventfp
, *filep
= NULL
;
1605 struct eventfd_ctx
*ctx
= NULL
;
1610 /* If you are not the owner, you can become one */
1611 if (ioctl
== VHOST_SET_OWNER
) {
1612 r
= vhost_dev_set_owner(d
);
1616 /* You must be the owner to do anything else */
1617 r
= vhost_dev_check_owner(d
);
1622 case VHOST_SET_MEM_TABLE
:
1623 r
= vhost_set_memory(d
, argp
);
1625 case VHOST_SET_LOG_BASE
:
1626 if (copy_from_user(&p
, argp
, sizeof p
)) {
1630 if ((u64
)(unsigned long)p
!= p
) {
1634 for (i
= 0; i
< d
->nvqs
; ++i
) {
1635 struct vhost_virtqueue
*vq
;
1636 void __user
*base
= (void __user
*)(unsigned long)p
;
1638 mutex_lock(&vq
->mutex
);
1639 /* If ring is inactive, will check when it's enabled. */
1640 if (vq
->private_data
&& !vq_log_access_ok(vq
, base
))
1643 vq
->log_base
= base
;
1644 mutex_unlock(&vq
->mutex
);
1647 case VHOST_SET_LOG_FD
:
1648 r
= get_user(fd
, (int __user
*)argp
);
1651 eventfp
= fd
== -1 ? NULL
: eventfd_fget(fd
);
1652 if (IS_ERR(eventfp
)) {
1653 r
= PTR_ERR(eventfp
);
1656 if (eventfp
!= d
->log_file
) {
1657 filep
= d
->log_file
;
1658 d
->log_file
= eventfp
;
1660 d
->log_ctx
= eventfp
?
1661 eventfd_ctx_fileget(eventfp
) : NULL
;
1664 for (i
= 0; i
< d
->nvqs
; ++i
) {
1665 mutex_lock(&d
->vqs
[i
]->mutex
);
1666 d
->vqs
[i
]->log_ctx
= d
->log_ctx
;
1667 mutex_unlock(&d
->vqs
[i
]->mutex
);
1670 eventfd_ctx_put(ctx
);
1681 EXPORT_SYMBOL_GPL(vhost_dev_ioctl
);
1683 /* TODO: This is really inefficient. We need something like get_user()
1684 * (instruction directly accesses the data, with an exception table entry
1685 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1687 static int set_bit_to_user(int nr
, void __user
*addr
)
1689 unsigned long log
= (unsigned long)addr
;
1692 int bit
= nr
+ (log
% PAGE_SIZE
) * 8;
1695 r
= get_user_pages_fast(log
, 1, 1, &page
);
1699 base
= kmap_atomic(page
);
1701 kunmap_atomic(base
);
1702 set_page_dirty_lock(page
);
1707 static int log_write(void __user
*log_base
,
1708 u64 write_address
, u64 write_length
)
1710 u64 write_page
= write_address
/ VHOST_PAGE_SIZE
;
1715 write_length
+= write_address
% VHOST_PAGE_SIZE
;
1717 u64 base
= (u64
)(unsigned long)log_base
;
1718 u64 log
= base
+ write_page
/ 8;
1719 int bit
= write_page
% 8;
1720 if ((u64
)(unsigned long)log
!= log
)
1722 r
= set_bit_to_user(bit
, (void __user
*)(unsigned long)log
);
1725 if (write_length
<= VHOST_PAGE_SIZE
)
1727 write_length
-= VHOST_PAGE_SIZE
;
1733 int vhost_log_write(struct vhost_virtqueue
*vq
, struct vhost_log
*log
,
1734 unsigned int log_num
, u64 len
)
1738 /* Make sure data written is seen before log. */
1740 for (i
= 0; i
< log_num
; ++i
) {
1741 u64 l
= min(log
[i
].len
, len
);
1742 r
= log_write(vq
->log_base
, log
[i
].addr
, l
);
1748 eventfd_signal(vq
->log_ctx
, 1);
1752 /* Length written exceeds what we have stored. This is a bug. */
1756 EXPORT_SYMBOL_GPL(vhost_log_write
);
1758 static int vhost_update_used_flags(struct vhost_virtqueue
*vq
)
1761 if (vhost_put_user(vq
, cpu_to_vhost16(vq
, vq
->used_flags
),
1762 &vq
->used
->flags
) < 0)
1764 if (unlikely(vq
->log_used
)) {
1765 /* Make sure the flag is seen before log. */
1767 /* Log used flag write. */
1768 used
= &vq
->used
->flags
;
1769 log_write(vq
->log_base
, vq
->log_addr
+
1770 (used
- (void __user
*)vq
->used
),
1771 sizeof vq
->used
->flags
);
1773 eventfd_signal(vq
->log_ctx
, 1);
1778 static int vhost_update_avail_event(struct vhost_virtqueue
*vq
, u16 avail_event
)
1780 if (vhost_put_user(vq
, cpu_to_vhost16(vq
, vq
->avail_idx
),
1781 vhost_avail_event(vq
)))
1783 if (unlikely(vq
->log_used
)) {
1785 /* Make sure the event is seen before log. */
1787 /* Log avail event write */
1788 used
= vhost_avail_event(vq
);
1789 log_write(vq
->log_base
, vq
->log_addr
+
1790 (used
- (void __user
*)vq
->used
),
1791 sizeof *vhost_avail_event(vq
));
1793 eventfd_signal(vq
->log_ctx
, 1);
1798 int vhost_vq_init_access(struct vhost_virtqueue
*vq
)
1800 __virtio16 last_used_idx
;
1802 bool is_le
= vq
->is_le
;
1804 if (!vq
->private_data
)
1807 vhost_init_is_le(vq
);
1809 r
= vhost_update_used_flags(vq
);
1812 vq
->signalled_used_valid
= false;
1814 !access_ok(VERIFY_READ
, &vq
->used
->idx
, sizeof vq
->used
->idx
)) {
1818 r
= vhost_get_used(vq
, last_used_idx
, &vq
->used
->idx
);
1820 vq_err(vq
, "Can't access used idx at %p\n",
1824 vq
->last_used_idx
= vhost16_to_cpu(vq
, last_used_idx
);
1831 EXPORT_SYMBOL_GPL(vhost_vq_init_access
);
1833 static int translate_desc(struct vhost_virtqueue
*vq
, u64 addr
, u32 len
,
1834 struct iovec iov
[], int iov_size
, int access
)
1836 const struct vhost_umem_node
*node
;
1837 struct vhost_dev
*dev
= vq
->dev
;
1838 struct vhost_umem
*umem
= dev
->iotlb
? dev
->iotlb
: dev
->umem
;
1843 while ((u64
)len
> s
) {
1845 if (unlikely(ret
>= iov_size
)) {
1850 node
= vhost_umem_interval_tree_iter_first(&umem
->umem_tree
,
1851 addr
, addr
+ len
- 1);
1852 if (node
== NULL
|| node
->start
> addr
) {
1853 if (umem
!= dev
->iotlb
) {
1859 } else if (!(node
->perm
& access
)) {
1865 size
= node
->size
- addr
+ node
->start
;
1866 _iov
->iov_len
= min((u64
)len
- s
, size
);
1867 _iov
->iov_base
= (void __user
*)(unsigned long)
1868 (node
->userspace_addr
+ addr
- node
->start
);
1875 vhost_iotlb_miss(vq
, addr
, access
);
1879 /* Each buffer in the virtqueues is actually a chain of descriptors. This
1880 * function returns the next descriptor in the chain,
1881 * or -1U if we're at the end. */
1882 static unsigned next_desc(struct vhost_virtqueue
*vq
, struct vring_desc
*desc
)
1886 /* If this descriptor says it doesn't chain, we're done. */
1887 if (!(desc
->flags
& cpu_to_vhost16(vq
, VRING_DESC_F_NEXT
)))
1890 /* Check they're not leading us off end of descriptors. */
1891 next
= vhost16_to_cpu(vq
, desc
->next
);
1892 /* Make sure compiler knows to grab that: we don't want it changing! */
1893 /* We will use the result as an index in an array, so most
1894 * architectures only need a compiler barrier here. */
1895 read_barrier_depends();
1900 static int get_indirect(struct vhost_virtqueue
*vq
,
1901 struct iovec iov
[], unsigned int iov_size
,
1902 unsigned int *out_num
, unsigned int *in_num
,
1903 struct vhost_log
*log
, unsigned int *log_num
,
1904 struct vring_desc
*indirect
)
1906 struct vring_desc desc
;
1907 unsigned int i
= 0, count
, found
= 0;
1908 u32 len
= vhost32_to_cpu(vq
, indirect
->len
);
1909 struct iov_iter from
;
1913 if (unlikely(len
% sizeof desc
)) {
1914 vq_err(vq
, "Invalid length in indirect descriptor: "
1915 "len 0x%llx not multiple of 0x%zx\n",
1916 (unsigned long long)len
,
1921 ret
= translate_desc(vq
, vhost64_to_cpu(vq
, indirect
->addr
), len
, vq
->indirect
,
1922 UIO_MAXIOV
, VHOST_ACCESS_RO
);
1923 if (unlikely(ret
< 0)) {
1925 vq_err(vq
, "Translation failure %d in indirect.\n", ret
);
1928 iov_iter_init(&from
, READ
, vq
->indirect
, ret
, len
);
1930 /* We will use the result as an address to read from, so most
1931 * architectures only need a compiler barrier here. */
1932 read_barrier_depends();
1934 count
= len
/ sizeof desc
;
1935 /* Buffers are chained via a 16 bit next field, so
1936 * we can have at most 2^16 of these. */
1937 if (unlikely(count
> USHRT_MAX
+ 1)) {
1938 vq_err(vq
, "Indirect buffer length too big: %d\n",
1944 unsigned iov_count
= *in_num
+ *out_num
;
1945 if (unlikely(++found
> count
)) {
1946 vq_err(vq
, "Loop detected: last one at %u "
1947 "indirect size %u\n",
1951 if (unlikely(!copy_from_iter_full(&desc
, sizeof(desc
), &from
))) {
1952 vq_err(vq
, "Failed indirect descriptor: idx %d, %zx\n",
1953 i
, (size_t)vhost64_to_cpu(vq
, indirect
->addr
) + i
* sizeof desc
);
1956 if (unlikely(desc
.flags
& cpu_to_vhost16(vq
, VRING_DESC_F_INDIRECT
))) {
1957 vq_err(vq
, "Nested indirect descriptor: idx %d, %zx\n",
1958 i
, (size_t)vhost64_to_cpu(vq
, indirect
->addr
) + i
* sizeof desc
);
1962 if (desc
.flags
& cpu_to_vhost16(vq
, VRING_DESC_F_WRITE
))
1963 access
= VHOST_ACCESS_WO
;
1965 access
= VHOST_ACCESS_RO
;
1967 ret
= translate_desc(vq
, vhost64_to_cpu(vq
, desc
.addr
),
1968 vhost32_to_cpu(vq
, desc
.len
), iov
+ iov_count
,
1969 iov_size
- iov_count
, access
);
1970 if (unlikely(ret
< 0)) {
1972 vq_err(vq
, "Translation failure %d indirect idx %d\n",
1976 /* If this is an input descriptor, increment that count. */
1977 if (access
== VHOST_ACCESS_WO
) {
1979 if (unlikely(log
)) {
1980 log
[*log_num
].addr
= vhost64_to_cpu(vq
, desc
.addr
);
1981 log
[*log_num
].len
= vhost32_to_cpu(vq
, desc
.len
);
1985 /* If it's an output descriptor, they're all supposed
1986 * to come before any input descriptors. */
1987 if (unlikely(*in_num
)) {
1988 vq_err(vq
, "Indirect descriptor "
1989 "has out after in: idx %d\n", i
);
1994 } while ((i
= next_desc(vq
, &desc
)) != -1);
1998 /* This looks in the virtqueue and for the first available buffer, and converts
1999 * it to an iovec for convenient access. Since descriptors consist of some
2000 * number of output then some number of input descriptors, it's actually two
2001 * iovecs, but we pack them into one and note how many of each there were.
2003 * This function returns the descriptor number found, or vq->num (which is
2004 * never a valid descriptor number) if none was found. A negative code is
2005 * returned on error. */
2006 int vhost_get_vq_desc(struct vhost_virtqueue
*vq
,
2007 struct iovec iov
[], unsigned int iov_size
,
2008 unsigned int *out_num
, unsigned int *in_num
,
2009 struct vhost_log
*log
, unsigned int *log_num
)
2011 struct vring_desc desc
;
2012 unsigned int i
, head
, found
= 0;
2014 __virtio16 avail_idx
;
2015 __virtio16 ring_head
;
2018 /* Check it isn't doing very strange things with descriptor numbers. */
2019 last_avail_idx
= vq
->last_avail_idx
;
2021 if (vq
->avail_idx
== vq
->last_avail_idx
) {
2022 if (unlikely(vhost_get_avail(vq
, avail_idx
, &vq
->avail
->idx
))) {
2023 vq_err(vq
, "Failed to access avail idx at %p\n",
2027 vq
->avail_idx
= vhost16_to_cpu(vq
, avail_idx
);
2029 if (unlikely((u16
)(vq
->avail_idx
- last_avail_idx
) > vq
->num
)) {
2030 vq_err(vq
, "Guest moved used index from %u to %u",
2031 last_avail_idx
, vq
->avail_idx
);
2035 /* If there's nothing new since last we looked, return
2038 if (vq
->avail_idx
== last_avail_idx
)
2041 /* Only get avail ring entries after they have been
2047 /* Grab the next descriptor number they're advertising, and increment
2048 * the index we've seen. */
2049 if (unlikely(vhost_get_avail(vq
, ring_head
,
2050 &vq
->avail
->ring
[last_avail_idx
& (vq
->num
- 1)]))) {
2051 vq_err(vq
, "Failed to read head: idx %d address %p\n",
2053 &vq
->avail
->ring
[last_avail_idx
% vq
->num
]);
2057 head
= vhost16_to_cpu(vq
, ring_head
);
2059 /* If their number is silly, that's an error. */
2060 if (unlikely(head
>= vq
->num
)) {
2061 vq_err(vq
, "Guest says index %u > %u is available",
2066 /* When we start there are none of either input nor output. */
2067 *out_num
= *in_num
= 0;
2073 unsigned iov_count
= *in_num
+ *out_num
;
2074 if (unlikely(i
>= vq
->num
)) {
2075 vq_err(vq
, "Desc index is %u > %u, head = %u",
2079 if (unlikely(++found
> vq
->num
)) {
2080 vq_err(vq
, "Loop detected: last one at %u "
2081 "vq size %u head %u\n",
2085 ret
= vhost_copy_from_user(vq
, &desc
, vq
->desc
+ i
,
2087 if (unlikely(ret
)) {
2088 vq_err(vq
, "Failed to get descriptor: idx %d addr %p\n",
2092 if (desc
.flags
& cpu_to_vhost16(vq
, VRING_DESC_F_INDIRECT
)) {
2093 ret
= get_indirect(vq
, iov
, iov_size
,
2095 log
, log_num
, &desc
);
2096 if (unlikely(ret
< 0)) {
2098 vq_err(vq
, "Failure detected "
2099 "in indirect descriptor at idx %d\n", i
);
2105 if (desc
.flags
& cpu_to_vhost16(vq
, VRING_DESC_F_WRITE
))
2106 access
= VHOST_ACCESS_WO
;
2108 access
= VHOST_ACCESS_RO
;
2109 ret
= translate_desc(vq
, vhost64_to_cpu(vq
, desc
.addr
),
2110 vhost32_to_cpu(vq
, desc
.len
), iov
+ iov_count
,
2111 iov_size
- iov_count
, access
);
2112 if (unlikely(ret
< 0)) {
2114 vq_err(vq
, "Translation failure %d descriptor idx %d\n",
2118 if (access
== VHOST_ACCESS_WO
) {
2119 /* If this is an input descriptor,
2120 * increment that count. */
2122 if (unlikely(log
)) {
2123 log
[*log_num
].addr
= vhost64_to_cpu(vq
, desc
.addr
);
2124 log
[*log_num
].len
= vhost32_to_cpu(vq
, desc
.len
);
2128 /* If it's an output descriptor, they're all supposed
2129 * to come before any input descriptors. */
2130 if (unlikely(*in_num
)) {
2131 vq_err(vq
, "Descriptor has out after in: "
2137 } while ((i
= next_desc(vq
, &desc
)) != -1);
2139 /* On success, increment avail index. */
2140 vq
->last_avail_idx
++;
2142 /* Assume notifications from guest are disabled at this point,
2143 * if they aren't we would need to update avail_event index. */
2144 BUG_ON(!(vq
->used_flags
& VRING_USED_F_NO_NOTIFY
));
2147 EXPORT_SYMBOL_GPL(vhost_get_vq_desc
);
2149 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2150 void vhost_discard_vq_desc(struct vhost_virtqueue
*vq
, int n
)
2152 vq
->last_avail_idx
-= n
;
2154 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc
);
2156 /* After we've used one of their buffers, we tell them about it. We'll then
2157 * want to notify the guest, using eventfd. */
2158 int vhost_add_used(struct vhost_virtqueue
*vq
, unsigned int head
, int len
)
2160 struct vring_used_elem heads
= {
2161 cpu_to_vhost32(vq
, head
),
2162 cpu_to_vhost32(vq
, len
)
2165 return vhost_add_used_n(vq
, &heads
, 1);
2167 EXPORT_SYMBOL_GPL(vhost_add_used
);
2169 static int __vhost_add_used_n(struct vhost_virtqueue
*vq
,
2170 struct vring_used_elem
*heads
,
2173 struct vring_used_elem __user
*used
;
2177 start
= vq
->last_used_idx
& (vq
->num
- 1);
2178 used
= vq
->used
->ring
+ start
;
2180 if (vhost_put_user(vq
, heads
[0].id
, &used
->id
)) {
2181 vq_err(vq
, "Failed to write used id");
2184 if (vhost_put_user(vq
, heads
[0].len
, &used
->len
)) {
2185 vq_err(vq
, "Failed to write used len");
2188 } else if (vhost_copy_to_user(vq
, used
, heads
, count
* sizeof *used
)) {
2189 vq_err(vq
, "Failed to write used");
2192 if (unlikely(vq
->log_used
)) {
2193 /* Make sure data is seen before log. */
2195 /* Log used ring entry write. */
2196 log_write(vq
->log_base
,
2198 ((void __user
*)used
- (void __user
*)vq
->used
),
2199 count
* sizeof *used
);
2201 old
= vq
->last_used_idx
;
2202 new = (vq
->last_used_idx
+= count
);
2203 /* If the driver never bothers to signal in a very long while,
2204 * used index might wrap around. If that happens, invalidate
2205 * signalled_used index we stored. TODO: make sure driver
2206 * signals at least once in 2^16 and remove this. */
2207 if (unlikely((u16
)(new - vq
->signalled_used
) < (u16
)(new - old
)))
2208 vq
->signalled_used_valid
= false;
2212 /* After we've used one of their buffers, we tell them about it. We'll then
2213 * want to notify the guest, using eventfd. */
2214 int vhost_add_used_n(struct vhost_virtqueue
*vq
, struct vring_used_elem
*heads
,
2219 start
= vq
->last_used_idx
& (vq
->num
- 1);
2220 n
= vq
->num
- start
;
2222 r
= __vhost_add_used_n(vq
, heads
, n
);
2228 r
= __vhost_add_used_n(vq
, heads
, count
);
2230 /* Make sure buffer is written before we update index. */
2232 if (vhost_put_user(vq
, cpu_to_vhost16(vq
, vq
->last_used_idx
),
2234 vq_err(vq
, "Failed to increment used idx");
2237 if (unlikely(vq
->log_used
)) {
2238 /* Log used index update. */
2239 log_write(vq
->log_base
,
2240 vq
->log_addr
+ offsetof(struct vring_used
, idx
),
2241 sizeof vq
->used
->idx
);
2243 eventfd_signal(vq
->log_ctx
, 1);
2247 EXPORT_SYMBOL_GPL(vhost_add_used_n
);
2249 static bool vhost_notify(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
)
2254 /* Flush out used index updates. This is paired
2255 * with the barrier that the Guest executes when enabling
2259 if (vhost_has_feature(vq
, VIRTIO_F_NOTIFY_ON_EMPTY
) &&
2260 unlikely(vq
->avail_idx
== vq
->last_avail_idx
))
2263 if (!vhost_has_feature(vq
, VIRTIO_RING_F_EVENT_IDX
)) {
2265 if (vhost_get_avail(vq
, flags
, &vq
->avail
->flags
)) {
2266 vq_err(vq
, "Failed to get flags");
2269 return !(flags
& cpu_to_vhost16(vq
, VRING_AVAIL_F_NO_INTERRUPT
));
2271 old
= vq
->signalled_used
;
2272 v
= vq
->signalled_used_valid
;
2273 new = vq
->signalled_used
= vq
->last_used_idx
;
2274 vq
->signalled_used_valid
= true;
2279 if (vhost_get_avail(vq
, event
, vhost_used_event(vq
))) {
2280 vq_err(vq
, "Failed to get used event idx");
2283 return vring_need_event(vhost16_to_cpu(vq
, event
), new, old
);
2286 /* This actually signals the guest, using eventfd. */
2287 void vhost_signal(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
)
2289 /* Signal the Guest tell them we used something up. */
2290 if (vq
->call_ctx
&& vhost_notify(dev
, vq
))
2291 eventfd_signal(vq
->call_ctx
, 1);
2293 EXPORT_SYMBOL_GPL(vhost_signal
);
2295 /* And here's the combo meal deal. Supersize me! */
2296 void vhost_add_used_and_signal(struct vhost_dev
*dev
,
2297 struct vhost_virtqueue
*vq
,
2298 unsigned int head
, int len
)
2300 vhost_add_used(vq
, head
, len
);
2301 vhost_signal(dev
, vq
);
2303 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal
);
2305 /* multi-buffer version of vhost_add_used_and_signal */
2306 void vhost_add_used_and_signal_n(struct vhost_dev
*dev
,
2307 struct vhost_virtqueue
*vq
,
2308 struct vring_used_elem
*heads
, unsigned count
)
2310 vhost_add_used_n(vq
, heads
, count
);
2311 vhost_signal(dev
, vq
);
2313 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n
);
2315 /* return true if we're sure that avaiable ring is empty */
2316 bool vhost_vq_avail_empty(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
)
2318 __virtio16 avail_idx
;
2321 if (vq
->avail_idx
!= vq
->last_avail_idx
)
2324 r
= vhost_get_avail(vq
, avail_idx
, &vq
->avail
->idx
);
2327 vq
->avail_idx
= vhost16_to_cpu(vq
, avail_idx
);
2329 return vq
->avail_idx
== vq
->last_avail_idx
;
2331 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty
);
2333 /* OK, now we need to know about added descriptors. */
2334 bool vhost_enable_notify(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
)
2336 __virtio16 avail_idx
;
2339 if (!(vq
->used_flags
& VRING_USED_F_NO_NOTIFY
))
2341 vq
->used_flags
&= ~VRING_USED_F_NO_NOTIFY
;
2342 if (!vhost_has_feature(vq
, VIRTIO_RING_F_EVENT_IDX
)) {
2343 r
= vhost_update_used_flags(vq
);
2345 vq_err(vq
, "Failed to enable notification at %p: %d\n",
2346 &vq
->used
->flags
, r
);
2350 r
= vhost_update_avail_event(vq
, vq
->avail_idx
);
2352 vq_err(vq
, "Failed to update avail event index at %p: %d\n",
2353 vhost_avail_event(vq
), r
);
2357 /* They could have slipped one in as we were doing that: make
2358 * sure it's written, then check again. */
2360 r
= vhost_get_avail(vq
, avail_idx
, &vq
->avail
->idx
);
2362 vq_err(vq
, "Failed to check avail idx at %p: %d\n",
2363 &vq
->avail
->idx
, r
);
2367 return vhost16_to_cpu(vq
, avail_idx
) != vq
->avail_idx
;
2369 EXPORT_SYMBOL_GPL(vhost_enable_notify
);
2371 /* We don't need to be notified again. */
2372 void vhost_disable_notify(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
)
2376 if (vq
->used_flags
& VRING_USED_F_NO_NOTIFY
)
2378 vq
->used_flags
|= VRING_USED_F_NO_NOTIFY
;
2379 if (!vhost_has_feature(vq
, VIRTIO_RING_F_EVENT_IDX
)) {
2380 r
= vhost_update_used_flags(vq
);
2382 vq_err(vq
, "Failed to enable notification at %p: %d\n",
2383 &vq
->used
->flags
, r
);
2386 EXPORT_SYMBOL_GPL(vhost_disable_notify
);
2388 /* Create a new message. */
2389 struct vhost_msg_node
*vhost_new_msg(struct vhost_virtqueue
*vq
, int type
)
2391 struct vhost_msg_node
*node
= kmalloc(sizeof *node
, GFP_KERNEL
);
2395 /* Make sure all padding within the structure is initialized. */
2396 memset(&node
->msg
, 0, sizeof node
->msg
);
2398 node
->msg
.type
= type
;
2401 EXPORT_SYMBOL_GPL(vhost_new_msg
);
2403 void vhost_enqueue_msg(struct vhost_dev
*dev
, struct list_head
*head
,
2404 struct vhost_msg_node
*node
)
2406 spin_lock(&dev
->iotlb_lock
);
2407 list_add_tail(&node
->node
, head
);
2408 spin_unlock(&dev
->iotlb_lock
);
2410 wake_up_interruptible_poll(&dev
->wait
, POLLIN
| POLLRDNORM
);
2412 EXPORT_SYMBOL_GPL(vhost_enqueue_msg
);
2414 struct vhost_msg_node
*vhost_dequeue_msg(struct vhost_dev
*dev
,
2415 struct list_head
*head
)
2417 struct vhost_msg_node
*node
= NULL
;
2419 spin_lock(&dev
->iotlb_lock
);
2420 if (!list_empty(head
)) {
2421 node
= list_first_entry(head
, struct vhost_msg_node
,
2423 list_del(&node
->node
);
2425 spin_unlock(&dev
->iotlb_lock
);
2429 EXPORT_SYMBOL_GPL(vhost_dequeue_msg
);
2432 static int __init
vhost_init(void)
2437 static void __exit
vhost_exit(void)
2441 module_init(vhost_init
);
2442 module_exit(vhost_exit
);
2444 MODULE_VERSION("0.0.1");
2445 MODULE_LICENSE("GPL v2");
2446 MODULE_AUTHOR("Michael S. Tsirkin");
2447 MODULE_DESCRIPTION("Host kernel accelerator for virtio");