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>
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 INTERVAL_TREE_DEFINE(struct vhost_umem_node
,
53 rb
, __u64
, __subtree_last
,
54 START
, LAST
, static inline, vhost_umem_interval_tree
);
56 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
57 static void vhost_disable_cross_endian(struct vhost_virtqueue
*vq
)
59 vq
->user_be
= !virtio_legacy_is_little_endian();
62 static void vhost_enable_cross_endian_big(struct vhost_virtqueue
*vq
)
67 static void vhost_enable_cross_endian_little(struct vhost_virtqueue
*vq
)
72 static long vhost_set_vring_endian(struct vhost_virtqueue
*vq
, int __user
*argp
)
74 struct vhost_vring_state s
;
79 if (copy_from_user(&s
, argp
, sizeof(s
)))
82 if (s
.num
!= VHOST_VRING_LITTLE_ENDIAN
&&
83 s
.num
!= VHOST_VRING_BIG_ENDIAN
)
86 if (s
.num
== VHOST_VRING_BIG_ENDIAN
)
87 vhost_enable_cross_endian_big(vq
);
89 vhost_enable_cross_endian_little(vq
);
94 static long vhost_get_vring_endian(struct vhost_virtqueue
*vq
, u32 idx
,
97 struct vhost_vring_state s
= {
102 if (copy_to_user(argp
, &s
, sizeof(s
)))
108 static void vhost_init_is_le(struct vhost_virtqueue
*vq
)
110 /* Note for legacy virtio: user_be is initialized at reset time
111 * according to the host endianness. If userspace does not set an
112 * explicit endianness, the default behavior is native endian, as
113 * expected by legacy virtio.
115 vq
->is_le
= vhost_has_feature(vq
, VIRTIO_F_VERSION_1
) || !vq
->user_be
;
118 static void vhost_disable_cross_endian(struct vhost_virtqueue
*vq
)
122 static long vhost_set_vring_endian(struct vhost_virtqueue
*vq
, int __user
*argp
)
127 static long vhost_get_vring_endian(struct vhost_virtqueue
*vq
, u32 idx
,
133 static void vhost_init_is_le(struct vhost_virtqueue
*vq
)
135 vq
->is_le
= vhost_has_feature(vq
, VIRTIO_F_VERSION_1
)
136 || virtio_legacy_is_little_endian();
138 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
140 static void vhost_reset_is_le(struct vhost_virtqueue
*vq
)
142 vhost_init_is_le(vq
);
145 struct vhost_flush_struct
{
146 struct vhost_work work
;
147 struct completion wait_event
;
150 static void vhost_flush_work(struct vhost_work
*work
)
152 struct vhost_flush_struct
*s
;
154 s
= container_of(work
, struct vhost_flush_struct
, work
);
155 complete(&s
->wait_event
);
158 static void vhost_poll_func(struct file
*file
, wait_queue_head_t
*wqh
,
161 struct vhost_poll
*poll
;
163 poll
= container_of(pt
, struct vhost_poll
, table
);
165 add_wait_queue(wqh
, &poll
->wait
);
168 static int vhost_poll_wakeup(wait_queue_entry_t
*wait
, unsigned mode
, int sync
,
171 struct vhost_poll
*poll
= container_of(wait
, struct vhost_poll
, wait
);
173 if (!((unsigned long)key
& poll
->mask
))
176 vhost_poll_queue(poll
);
180 void vhost_work_init(struct vhost_work
*work
, vhost_work_fn_t fn
)
182 clear_bit(VHOST_WORK_QUEUED
, &work
->flags
);
184 init_waitqueue_head(&work
->done
);
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 unsigned long 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
)
212 mask
= file
->f_op
->poll(file
, &poll
->table
);
214 vhost_poll_wakeup(&poll
->wait
, 0, 0, (void *)mask
);
215 if (mask
& POLLERR
) {
216 vhost_poll_stop(poll
);
222 EXPORT_SYMBOL_GPL(vhost_poll_start
);
224 /* Stop polling a file. After this function returns, it becomes safe to drop the
225 * file reference. You must also flush afterwards. */
226 void vhost_poll_stop(struct vhost_poll
*poll
)
229 remove_wait_queue(poll
->wqh
, &poll
->wait
);
233 EXPORT_SYMBOL_GPL(vhost_poll_stop
);
235 void vhost_work_flush(struct vhost_dev
*dev
, struct vhost_work
*work
)
237 struct vhost_flush_struct flush
;
240 init_completion(&flush
.wait_event
);
241 vhost_work_init(&flush
.work
, vhost_flush_work
);
243 vhost_work_queue(dev
, &flush
.work
);
244 wait_for_completion(&flush
.wait_event
);
247 EXPORT_SYMBOL_GPL(vhost_work_flush
);
249 /* Flush any work that has been scheduled. When calling this, don't hold any
250 * locks that are also used by the callback. */
251 void vhost_poll_flush(struct vhost_poll
*poll
)
253 vhost_work_flush(poll
->dev
, &poll
->work
);
255 EXPORT_SYMBOL_GPL(vhost_poll_flush
);
257 void vhost_work_queue(struct vhost_dev
*dev
, struct vhost_work
*work
)
262 if (!test_and_set_bit(VHOST_WORK_QUEUED
, &work
->flags
)) {
263 /* We can only add the work to the list after we're
264 * sure it was not in the list.
265 * test_and_set_bit() implies a memory barrier.
267 llist_add(&work
->node
, &dev
->work_list
);
268 wake_up_process(dev
->worker
);
271 EXPORT_SYMBOL_GPL(vhost_work_queue
);
273 /* A lockless hint for busy polling code to exit the loop */
274 bool vhost_has_work(struct vhost_dev
*dev
)
276 return !llist_empty(&dev
->work_list
);
278 EXPORT_SYMBOL_GPL(vhost_has_work
);
280 void vhost_poll_queue(struct vhost_poll
*poll
)
282 vhost_work_queue(poll
->dev
, &poll
->work
);
284 EXPORT_SYMBOL_GPL(vhost_poll_queue
);
286 static void __vhost_vq_meta_reset(struct vhost_virtqueue
*vq
)
290 for (j
= 0; j
< VHOST_NUM_ADDRS
; j
++)
291 vq
->meta_iotlb
[j
] = NULL
;
294 static void vhost_vq_meta_reset(struct vhost_dev
*d
)
298 for (i
= 0; i
< d
->nvqs
; ++i
)
299 __vhost_vq_meta_reset(d
->vqs
[i
]);
302 static void vhost_vq_reset(struct vhost_dev
*dev
,
303 struct vhost_virtqueue
*vq
)
309 vq
->last_avail_idx
= 0;
311 vq
->last_used_idx
= 0;
312 vq
->signalled_used
= 0;
313 vq
->signalled_used_valid
= false;
315 vq
->log_used
= false;
316 vq
->log_addr
= -1ull;
317 vq
->private_data
= NULL
;
318 vq
->acked_features
= 0;
320 vq
->error_ctx
= NULL
;
326 vhost_reset_is_le(vq
);
327 vhost_disable_cross_endian(vq
);
328 vq
->busyloop_timeout
= 0;
331 __vhost_vq_meta_reset(vq
);
334 static int vhost_worker(void *data
)
336 struct vhost_dev
*dev
= data
;
337 struct vhost_work
*work
, *work_next
;
338 struct llist_node
*node
;
339 mm_segment_t oldfs
= get_fs();
345 /* mb paired w/ kthread_stop */
346 set_current_state(TASK_INTERRUPTIBLE
);
348 if (kthread_should_stop()) {
349 __set_current_state(TASK_RUNNING
);
353 node
= llist_del_all(&dev
->work_list
);
357 node
= llist_reverse_order(node
);
358 /* make sure flag is seen after deletion */
360 llist_for_each_entry_safe(work
, work_next
, node
, node
) {
361 clear_bit(VHOST_WORK_QUEUED
, &work
->flags
);
362 __set_current_state(TASK_RUNNING
);
373 static void vhost_vq_free_iovecs(struct vhost_virtqueue
*vq
)
383 /* Helper to allocate iovec buffers for all vqs. */
384 static long vhost_dev_alloc_iovecs(struct vhost_dev
*dev
)
386 struct vhost_virtqueue
*vq
;
389 for (i
= 0; i
< dev
->nvqs
; ++i
) {
391 vq
->indirect
= kmalloc(sizeof *vq
->indirect
* UIO_MAXIOV
,
393 vq
->log
= kmalloc(sizeof *vq
->log
* UIO_MAXIOV
, GFP_KERNEL
);
394 vq
->heads
= kmalloc(sizeof *vq
->heads
* UIO_MAXIOV
, GFP_KERNEL
);
395 if (!vq
->indirect
|| !vq
->log
|| !vq
->heads
)
402 vhost_vq_free_iovecs(dev
->vqs
[i
]);
406 static void vhost_dev_free_iovecs(struct vhost_dev
*dev
)
410 for (i
= 0; i
< dev
->nvqs
; ++i
)
411 vhost_vq_free_iovecs(dev
->vqs
[i
]);
414 void vhost_dev_init(struct vhost_dev
*dev
,
415 struct vhost_virtqueue
**vqs
, int nvqs
)
417 struct vhost_virtqueue
*vq
;
422 mutex_init(&dev
->mutex
);
424 dev
->log_file
= NULL
;
429 init_llist_head(&dev
->work_list
);
430 init_waitqueue_head(&dev
->wait
);
431 INIT_LIST_HEAD(&dev
->read_list
);
432 INIT_LIST_HEAD(&dev
->pending_list
);
433 spin_lock_init(&dev
->iotlb_lock
);
436 for (i
= 0; i
< dev
->nvqs
; ++i
) {
442 mutex_init(&vq
->mutex
);
443 vhost_vq_reset(dev
, vq
);
445 vhost_poll_init(&vq
->poll
, vq
->handle_kick
,
449 EXPORT_SYMBOL_GPL(vhost_dev_init
);
451 /* Caller should have device mutex */
452 long vhost_dev_check_owner(struct vhost_dev
*dev
)
454 /* Are you the owner? If not, I don't think you mean to do that */
455 return dev
->mm
== current
->mm
? 0 : -EPERM
;
457 EXPORT_SYMBOL_GPL(vhost_dev_check_owner
);
459 struct vhost_attach_cgroups_struct
{
460 struct vhost_work work
;
461 struct task_struct
*owner
;
465 static void vhost_attach_cgroups_work(struct vhost_work
*work
)
467 struct vhost_attach_cgroups_struct
*s
;
469 s
= container_of(work
, struct vhost_attach_cgroups_struct
, work
);
470 s
->ret
= cgroup_attach_task_all(s
->owner
, current
);
473 static int vhost_attach_cgroups(struct vhost_dev
*dev
)
475 struct vhost_attach_cgroups_struct attach
;
477 attach
.owner
= current
;
478 vhost_work_init(&attach
.work
, vhost_attach_cgroups_work
);
479 vhost_work_queue(dev
, &attach
.work
);
480 vhost_work_flush(dev
, &attach
.work
);
484 /* Caller should have device mutex */
485 bool vhost_dev_has_owner(struct vhost_dev
*dev
)
489 EXPORT_SYMBOL_GPL(vhost_dev_has_owner
);
491 /* Caller should have device mutex */
492 long vhost_dev_set_owner(struct vhost_dev
*dev
)
494 struct task_struct
*worker
;
497 /* Is there an owner already? */
498 if (vhost_dev_has_owner(dev
)) {
503 /* No owner, become one */
504 dev
->mm
= get_task_mm(current
);
505 worker
= kthread_create(vhost_worker
, dev
, "vhost-%d", current
->pid
);
506 if (IS_ERR(worker
)) {
507 err
= PTR_ERR(worker
);
511 dev
->worker
= worker
;
512 wake_up_process(worker
); /* avoid contributing to loadavg */
514 err
= vhost_attach_cgroups(dev
);
518 err
= vhost_dev_alloc_iovecs(dev
);
524 kthread_stop(worker
);
533 EXPORT_SYMBOL_GPL(vhost_dev_set_owner
);
535 struct vhost_umem
*vhost_dev_reset_owner_prepare(void)
537 return kvzalloc(sizeof(struct vhost_umem
), GFP_KERNEL
);
539 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare
);
541 /* Caller should have device mutex */
542 void vhost_dev_reset_owner(struct vhost_dev
*dev
, struct vhost_umem
*umem
)
546 vhost_dev_cleanup(dev
, true);
548 /* Restore memory to default empty mapping. */
549 INIT_LIST_HEAD(&umem
->umem_list
);
551 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
552 * VQs aren't running.
554 for (i
= 0; i
< dev
->nvqs
; ++i
)
555 dev
->vqs
[i
]->umem
= umem
;
557 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner
);
559 void vhost_dev_stop(struct vhost_dev
*dev
)
563 for (i
= 0; i
< dev
->nvqs
; ++i
) {
564 if (dev
->vqs
[i
]->kick
&& dev
->vqs
[i
]->handle_kick
) {
565 vhost_poll_stop(&dev
->vqs
[i
]->poll
);
566 vhost_poll_flush(&dev
->vqs
[i
]->poll
);
570 EXPORT_SYMBOL_GPL(vhost_dev_stop
);
572 static void vhost_umem_free(struct vhost_umem
*umem
,
573 struct vhost_umem_node
*node
)
575 vhost_umem_interval_tree_remove(node
, &umem
->umem_tree
);
576 list_del(&node
->link
);
581 static void vhost_umem_clean(struct vhost_umem
*umem
)
583 struct vhost_umem_node
*node
, *tmp
;
588 list_for_each_entry_safe(node
, tmp
, &umem
->umem_list
, link
)
589 vhost_umem_free(umem
, node
);
594 static void vhost_clear_msg(struct vhost_dev
*dev
)
596 struct vhost_msg_node
*node
, *n
;
598 spin_lock(&dev
->iotlb_lock
);
600 list_for_each_entry_safe(node
, n
, &dev
->read_list
, node
) {
601 list_del(&node
->node
);
605 list_for_each_entry_safe(node
, n
, &dev
->pending_list
, node
) {
606 list_del(&node
->node
);
610 spin_unlock(&dev
->iotlb_lock
);
613 /* Caller should have device mutex if and only if locked is set */
614 void vhost_dev_cleanup(struct vhost_dev
*dev
, bool locked
)
618 for (i
= 0; i
< dev
->nvqs
; ++i
) {
619 if (dev
->vqs
[i
]->error_ctx
)
620 eventfd_ctx_put(dev
->vqs
[i
]->error_ctx
);
621 if (dev
->vqs
[i
]->error
)
622 fput(dev
->vqs
[i
]->error
);
623 if (dev
->vqs
[i
]->kick
)
624 fput(dev
->vqs
[i
]->kick
);
625 if (dev
->vqs
[i
]->call_ctx
)
626 eventfd_ctx_put(dev
->vqs
[i
]->call_ctx
);
627 if (dev
->vqs
[i
]->call
)
628 fput(dev
->vqs
[i
]->call
);
629 vhost_vq_reset(dev
, dev
->vqs
[i
]);
631 vhost_dev_free_iovecs(dev
);
633 eventfd_ctx_put(dev
->log_ctx
);
637 dev
->log_file
= NULL
;
638 /* No one will access memory at this point */
639 vhost_umem_clean(dev
->umem
);
641 vhost_umem_clean(dev
->iotlb
);
643 vhost_clear_msg(dev
);
644 wake_up_interruptible_poll(&dev
->wait
, POLLIN
| POLLRDNORM
);
645 WARN_ON(!llist_empty(&dev
->work_list
));
647 kthread_stop(dev
->worker
);
654 EXPORT_SYMBOL_GPL(vhost_dev_cleanup
);
656 static int log_access_ok(void __user
*log_base
, u64 addr
, unsigned long sz
)
658 u64 a
= addr
/ VHOST_PAGE_SIZE
/ 8;
660 /* Make sure 64 bit math will not overflow. */
661 if (a
> ULONG_MAX
- (unsigned long)log_base
||
662 a
+ (unsigned long)log_base
> ULONG_MAX
)
665 return access_ok(VERIFY_WRITE
, log_base
+ a
,
666 (sz
+ VHOST_PAGE_SIZE
* 8 - 1) / VHOST_PAGE_SIZE
/ 8);
669 static bool vhost_overflow(u64 uaddr
, u64 size
)
671 /* Make sure 64 bit math will not overflow. */
672 return uaddr
> ULONG_MAX
|| size
> ULONG_MAX
|| uaddr
> ULONG_MAX
- size
;
675 /* Caller should have vq mutex and device mutex. */
676 static int vq_memory_access_ok(void __user
*log_base
, struct vhost_umem
*umem
,
679 struct vhost_umem_node
*node
;
684 list_for_each_entry(node
, &umem
->umem_list
, link
) {
685 unsigned long a
= node
->userspace_addr
;
687 if (vhost_overflow(node
->userspace_addr
, node
->size
))
691 if (!access_ok(VERIFY_WRITE
, (void __user
*)a
,
694 else if (log_all
&& !log_access_ok(log_base
,
702 static inline void __user
*vhost_vq_meta_fetch(struct vhost_virtqueue
*vq
,
703 u64 addr
, unsigned int size
,
706 const struct vhost_umem_node
*node
= vq
->meta_iotlb
[type
];
711 return (void *)(uintptr_t)(node
->userspace_addr
+ addr
- node
->start
);
714 /* Can we switch to this memory table? */
715 /* Caller should have device mutex but not vq mutex */
716 static int memory_access_ok(struct vhost_dev
*d
, struct vhost_umem
*umem
,
721 for (i
= 0; i
< d
->nvqs
; ++i
) {
725 mutex_lock(&d
->vqs
[i
]->mutex
);
726 log
= log_all
|| vhost_has_feature(d
->vqs
[i
], VHOST_F_LOG_ALL
);
727 /* If ring is inactive, will check when it's enabled. */
728 if (d
->vqs
[i
]->private_data
)
729 ok
= vq_memory_access_ok(d
->vqs
[i
]->log_base
,
733 mutex_unlock(&d
->vqs
[i
]->mutex
);
740 static int translate_desc(struct vhost_virtqueue
*vq
, u64 addr
, u32 len
,
741 struct iovec iov
[], int iov_size
, int access
);
743 static int vhost_copy_to_user(struct vhost_virtqueue
*vq
, void __user
*to
,
744 const void *from
, unsigned size
)
749 return __copy_to_user(to
, from
, size
);
751 /* This function should be called after iotlb
752 * prefetch, which means we're sure that all vq
753 * could be access through iotlb. So -EAGAIN should
754 * not happen in this case.
757 void __user
*uaddr
= vhost_vq_meta_fetch(vq
,
758 (u64
)(uintptr_t)to
, size
,
762 return __copy_to_user(uaddr
, from
, size
);
764 ret
= translate_desc(vq
, (u64
)(uintptr_t)to
, size
, vq
->iotlb_iov
,
765 ARRAY_SIZE(vq
->iotlb_iov
),
769 iov_iter_init(&t
, WRITE
, vq
->iotlb_iov
, ret
, size
);
770 ret
= copy_to_iter(from
, size
, &t
);
778 static int vhost_copy_from_user(struct vhost_virtqueue
*vq
, void *to
,
779 void __user
*from
, unsigned size
)
784 return __copy_from_user(to
, from
, size
);
786 /* This function should be called after iotlb
787 * prefetch, which means we're sure that vq
788 * could be access through iotlb. So -EAGAIN should
789 * not happen in this case.
791 void __user
*uaddr
= vhost_vq_meta_fetch(vq
,
792 (u64
)(uintptr_t)from
, size
,
797 return __copy_from_user(to
, uaddr
, size
);
799 ret
= translate_desc(vq
, (u64
)(uintptr_t)from
, size
, vq
->iotlb_iov
,
800 ARRAY_SIZE(vq
->iotlb_iov
),
803 vq_err(vq
, "IOTLB translation failure: uaddr "
804 "%p size 0x%llx\n", from
,
805 (unsigned long long) size
);
808 iov_iter_init(&f
, READ
, vq
->iotlb_iov
, ret
, size
);
809 ret
= copy_from_iter(to
, size
, &f
);
818 static void __user
*__vhost_get_user_slow(struct vhost_virtqueue
*vq
,
819 void __user
*addr
, unsigned int size
,
824 ret
= translate_desc(vq
, (u64
)(uintptr_t)addr
, size
, vq
->iotlb_iov
,
825 ARRAY_SIZE(vq
->iotlb_iov
),
828 vq_err(vq
, "IOTLB translation failure: uaddr "
829 "%p size 0x%llx\n", addr
,
830 (unsigned long long) size
);
834 if (ret
!= 1 || vq
->iotlb_iov
[0].iov_len
!= size
) {
835 vq_err(vq
, "Non atomic userspace memory access: uaddr "
836 "%p size 0x%llx\n", addr
,
837 (unsigned long long) size
);
841 return vq
->iotlb_iov
[0].iov_base
;
844 /* This function should be called after iotlb
845 * prefetch, which means we're sure that vq
846 * could be access through iotlb. So -EAGAIN should
847 * not happen in this case.
849 static inline void __user
*__vhost_get_user(struct vhost_virtqueue
*vq
,
850 void *addr
, unsigned int size
,
853 void __user
*uaddr
= vhost_vq_meta_fetch(vq
,
854 (u64
)(uintptr_t)addr
, size
, type
);
858 return __vhost_get_user_slow(vq
, addr
, size
, type
);
861 #define vhost_put_user(vq, x, ptr) \
865 ret = __put_user(x, ptr); \
867 __typeof__(ptr) to = \
868 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
869 sizeof(*ptr), VHOST_ADDR_USED); \
871 ret = __put_user(x, to); \
878 #define vhost_get_user(vq, x, ptr, type) \
882 ret = __get_user(x, ptr); \
884 __typeof__(ptr) from = \
885 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
889 ret = __get_user(x, from); \
896 #define vhost_get_avail(vq, x, ptr) \
897 vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
899 #define vhost_get_used(vq, x, ptr) \
900 vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
902 static void vhost_dev_lock_vqs(struct vhost_dev
*d
)
905 for (i
= 0; i
< d
->nvqs
; ++i
)
906 mutex_lock_nested(&d
->vqs
[i
]->mutex
, i
);
909 static void vhost_dev_unlock_vqs(struct vhost_dev
*d
)
912 for (i
= 0; i
< d
->nvqs
; ++i
)
913 mutex_unlock(&d
->vqs
[i
]->mutex
);
916 static int vhost_new_umem_range(struct vhost_umem
*umem
,
917 u64 start
, u64 size
, u64 end
,
918 u64 userspace_addr
, int perm
)
920 struct vhost_umem_node
*tmp
, *node
= kmalloc(sizeof(*node
), GFP_ATOMIC
);
925 if (umem
->numem
== max_iotlb_entries
) {
926 tmp
= list_first_entry(&umem
->umem_list
, typeof(*tmp
), link
);
927 vhost_umem_free(umem
, tmp
);
933 node
->userspace_addr
= userspace_addr
;
935 INIT_LIST_HEAD(&node
->link
);
936 list_add_tail(&node
->link
, &umem
->umem_list
);
937 vhost_umem_interval_tree_insert(node
, &umem
->umem_tree
);
943 static void vhost_del_umem_range(struct vhost_umem
*umem
,
946 struct vhost_umem_node
*node
;
948 while ((node
= vhost_umem_interval_tree_iter_first(&umem
->umem_tree
,
950 vhost_umem_free(umem
, node
);
953 static void vhost_iotlb_notify_vq(struct vhost_dev
*d
,
954 struct vhost_iotlb_msg
*msg
)
956 struct vhost_msg_node
*node
, *n
;
958 spin_lock(&d
->iotlb_lock
);
960 list_for_each_entry_safe(node
, n
, &d
->pending_list
, node
) {
961 struct vhost_iotlb_msg
*vq_msg
= &node
->msg
.iotlb
;
962 if (msg
->iova
<= vq_msg
->iova
&&
963 msg
->iova
+ msg
->size
- 1 > vq_msg
->iova
&&
964 vq_msg
->type
== VHOST_IOTLB_MISS
) {
965 vhost_poll_queue(&node
->vq
->poll
);
966 list_del(&node
->node
);
971 spin_unlock(&d
->iotlb_lock
);
974 static int umem_access_ok(u64 uaddr
, u64 size
, int access
)
976 unsigned long a
= uaddr
;
978 /* Make sure 64 bit math will not overflow. */
979 if (vhost_overflow(uaddr
, size
))
982 if ((access
& VHOST_ACCESS_RO
) &&
983 !access_ok(VERIFY_READ
, (void __user
*)a
, size
))
985 if ((access
& VHOST_ACCESS_WO
) &&
986 !access_ok(VERIFY_WRITE
, (void __user
*)a
, size
))
991 static int vhost_process_iotlb_msg(struct vhost_dev
*dev
,
992 struct vhost_iotlb_msg
*msg
)
996 vhost_dev_lock_vqs(dev
);
998 case VHOST_IOTLB_UPDATE
:
1003 if (umem_access_ok(msg
->uaddr
, msg
->size
, msg
->perm
)) {
1007 vhost_vq_meta_reset(dev
);
1008 if (vhost_new_umem_range(dev
->iotlb
, msg
->iova
, msg
->size
,
1009 msg
->iova
+ msg
->size
- 1,
1010 msg
->uaddr
, msg
->perm
)) {
1014 vhost_iotlb_notify_vq(dev
, msg
);
1016 case VHOST_IOTLB_INVALIDATE
:
1017 vhost_vq_meta_reset(dev
);
1018 vhost_del_umem_range(dev
->iotlb
, msg
->iova
,
1019 msg
->iova
+ msg
->size
- 1);
1026 vhost_dev_unlock_vqs(dev
);
1029 ssize_t
vhost_chr_write_iter(struct vhost_dev
*dev
,
1030 struct iov_iter
*from
)
1032 struct vhost_msg_node node
;
1033 unsigned size
= sizeof(struct vhost_msg
);
1037 if (iov_iter_count(from
) < size
)
1039 ret
= copy_from_iter(&node
.msg
, size
, from
);
1043 switch (node
.msg
.type
) {
1044 case VHOST_IOTLB_MSG
:
1045 err
= vhost_process_iotlb_msg(dev
, &node
.msg
.iotlb
);
1057 EXPORT_SYMBOL(vhost_chr_write_iter
);
1059 unsigned int vhost_chr_poll(struct file
*file
, struct vhost_dev
*dev
,
1062 unsigned int mask
= 0;
1064 poll_wait(file
, &dev
->wait
, wait
);
1066 if (!list_empty(&dev
->read_list
))
1067 mask
|= POLLIN
| POLLRDNORM
;
1071 EXPORT_SYMBOL(vhost_chr_poll
);
1073 ssize_t
vhost_chr_read_iter(struct vhost_dev
*dev
, struct iov_iter
*to
,
1077 struct vhost_msg_node
*node
;
1079 unsigned size
= sizeof(struct vhost_msg
);
1081 if (iov_iter_count(to
) < size
)
1086 prepare_to_wait(&dev
->wait
, &wait
,
1087 TASK_INTERRUPTIBLE
);
1089 node
= vhost_dequeue_msg(dev
, &dev
->read_list
);
1096 if (signal_pending(current
)) {
1109 finish_wait(&dev
->wait
, &wait
);
1112 ret
= copy_to_iter(&node
->msg
, size
, to
);
1114 if (ret
!= size
|| node
->msg
.type
!= VHOST_IOTLB_MISS
) {
1119 vhost_enqueue_msg(dev
, &dev
->pending_list
, node
);
1124 EXPORT_SYMBOL_GPL(vhost_chr_read_iter
);
1126 static int vhost_iotlb_miss(struct vhost_virtqueue
*vq
, u64 iova
, int access
)
1128 struct vhost_dev
*dev
= vq
->dev
;
1129 struct vhost_msg_node
*node
;
1130 struct vhost_iotlb_msg
*msg
;
1132 node
= vhost_new_msg(vq
, VHOST_IOTLB_MISS
);
1136 msg
= &node
->msg
.iotlb
;
1137 msg
->type
= VHOST_IOTLB_MISS
;
1141 vhost_enqueue_msg(dev
, &dev
->read_list
, node
);
1146 static int vq_access_ok(struct vhost_virtqueue
*vq
, unsigned int num
,
1147 struct vring_desc __user
*desc
,
1148 struct vring_avail __user
*avail
,
1149 struct vring_used __user
*used
)
1152 size_t s
= vhost_has_feature(vq
, VIRTIO_RING_F_EVENT_IDX
) ? 2 : 0;
1154 return access_ok(VERIFY_READ
, desc
, num
* sizeof *desc
) &&
1155 access_ok(VERIFY_READ
, avail
,
1156 sizeof *avail
+ num
* sizeof *avail
->ring
+ s
) &&
1157 access_ok(VERIFY_WRITE
, used
,
1158 sizeof *used
+ num
* sizeof *used
->ring
+ s
);
1161 static void vhost_vq_meta_update(struct vhost_virtqueue
*vq
,
1162 const struct vhost_umem_node
*node
,
1165 int access
= (type
== VHOST_ADDR_USED
) ?
1166 VHOST_ACCESS_WO
: VHOST_ACCESS_RO
;
1168 if (likely(node
->perm
& access
))
1169 vq
->meta_iotlb
[type
] = node
;
1172 static int iotlb_access_ok(struct vhost_virtqueue
*vq
,
1173 int access
, u64 addr
, u64 len
, int type
)
1175 const struct vhost_umem_node
*node
;
1176 struct vhost_umem
*umem
= vq
->iotlb
;
1177 u64 s
= 0, size
, orig_addr
= addr
;
1179 if (vhost_vq_meta_fetch(vq
, addr
, len
, type
))
1183 node
= vhost_umem_interval_tree_iter_first(&umem
->umem_tree
,
1186 if (node
== NULL
|| node
->start
> addr
) {
1187 vhost_iotlb_miss(vq
, addr
, access
);
1189 } else if (!(node
->perm
& access
)) {
1190 /* Report the possible access violation by
1191 * request another translation from userspace.
1196 size
= node
->size
- addr
+ node
->start
;
1198 if (orig_addr
== addr
&& size
>= len
)
1199 vhost_vq_meta_update(vq
, node
, type
);
1208 int vq_iotlb_prefetch(struct vhost_virtqueue
*vq
)
1210 size_t s
= vhost_has_feature(vq
, VIRTIO_RING_F_EVENT_IDX
) ? 2 : 0;
1211 unsigned int num
= vq
->num
;
1216 return iotlb_access_ok(vq
, VHOST_ACCESS_RO
, (u64
)(uintptr_t)vq
->desc
,
1217 num
* sizeof(*vq
->desc
), VHOST_ADDR_DESC
) &&
1218 iotlb_access_ok(vq
, VHOST_ACCESS_RO
, (u64
)(uintptr_t)vq
->avail
,
1220 num
* sizeof(*vq
->avail
->ring
) + s
,
1221 VHOST_ADDR_AVAIL
) &&
1222 iotlb_access_ok(vq
, VHOST_ACCESS_WO
, (u64
)(uintptr_t)vq
->used
,
1224 num
* sizeof(*vq
->used
->ring
) + s
,
1227 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch
);
1229 /* Can we log writes? */
1230 /* Caller should have device mutex but not vq mutex */
1231 int vhost_log_access_ok(struct vhost_dev
*dev
)
1233 return memory_access_ok(dev
, dev
->umem
, 1);
1235 EXPORT_SYMBOL_GPL(vhost_log_access_ok
);
1237 /* Verify access for write logging. */
1238 /* Caller should have vq mutex and device mutex */
1239 static int vq_log_access_ok(struct vhost_virtqueue
*vq
,
1240 void __user
*log_base
)
1242 size_t s
= vhost_has_feature(vq
, VIRTIO_RING_F_EVENT_IDX
) ? 2 : 0;
1244 return vq_memory_access_ok(log_base
, vq
->umem
,
1245 vhost_has_feature(vq
, VHOST_F_LOG_ALL
)) &&
1246 (!vq
->log_used
|| log_access_ok(log_base
, vq
->log_addr
,
1248 vq
->num
* sizeof *vq
->used
->ring
+ s
));
1251 /* Can we start vq? */
1252 /* Caller should have vq mutex and device mutex */
1253 int vhost_vq_access_ok(struct vhost_virtqueue
*vq
)
1255 int ret
= vq_log_access_ok(vq
, vq
->log_base
);
1257 if (ret
|| vq
->iotlb
)
1260 return vq_access_ok(vq
, vq
->num
, vq
->desc
, vq
->avail
, vq
->used
);
1262 EXPORT_SYMBOL_GPL(vhost_vq_access_ok
);
1264 static struct vhost_umem
*vhost_umem_alloc(void)
1266 struct vhost_umem
*umem
= kvzalloc(sizeof(*umem
), GFP_KERNEL
);
1271 umem
->umem_tree
= RB_ROOT_CACHED
;
1273 INIT_LIST_HEAD(&umem
->umem_list
);
1278 static long vhost_set_memory(struct vhost_dev
*d
, struct vhost_memory __user
*m
)
1280 struct vhost_memory mem
, *newmem
;
1281 struct vhost_memory_region
*region
;
1282 struct vhost_umem
*newumem
, *oldumem
;
1283 unsigned long size
= offsetof(struct vhost_memory
, regions
);
1286 if (copy_from_user(&mem
, m
, size
))
1290 if (mem
.nregions
> max_mem_regions
)
1292 newmem
= kvzalloc(size
+ mem
.nregions
* sizeof(*m
->regions
), GFP_KERNEL
);
1296 memcpy(newmem
, &mem
, size
);
1297 if (copy_from_user(newmem
->regions
, m
->regions
,
1298 mem
.nregions
* sizeof *m
->regions
)) {
1303 newumem
= vhost_umem_alloc();
1309 for (region
= newmem
->regions
;
1310 region
< newmem
->regions
+ mem
.nregions
;
1312 if (vhost_new_umem_range(newumem
,
1313 region
->guest_phys_addr
,
1314 region
->memory_size
,
1315 region
->guest_phys_addr
+
1316 region
->memory_size
- 1,
1317 region
->userspace_addr
,
1322 if (!memory_access_ok(d
, newumem
, 0))
1328 /* All memory accesses are done under some VQ mutex. */
1329 for (i
= 0; i
< d
->nvqs
; ++i
) {
1330 mutex_lock(&d
->vqs
[i
]->mutex
);
1331 d
->vqs
[i
]->umem
= newumem
;
1332 mutex_unlock(&d
->vqs
[i
]->mutex
);
1336 vhost_umem_clean(oldumem
);
1340 vhost_umem_clean(newumem
);
1345 long vhost_vring_ioctl(struct vhost_dev
*d
, int ioctl
, void __user
*argp
)
1347 struct file
*eventfp
, *filep
= NULL
;
1348 bool pollstart
= false, pollstop
= false;
1349 struct eventfd_ctx
*ctx
= NULL
;
1350 u32 __user
*idxp
= argp
;
1351 struct vhost_virtqueue
*vq
;
1352 struct vhost_vring_state s
;
1353 struct vhost_vring_file f
;
1354 struct vhost_vring_addr a
;
1358 r
= get_user(idx
, idxp
);
1366 mutex_lock(&vq
->mutex
);
1369 case VHOST_SET_VRING_NUM
:
1370 /* Resizing ring with an active backend?
1371 * You don't want to do that. */
1372 if (vq
->private_data
) {
1376 if (copy_from_user(&s
, argp
, sizeof s
)) {
1380 if (!s
.num
|| s
.num
> 0xffff || (s
.num
& (s
.num
- 1))) {
1386 case VHOST_SET_VRING_BASE
:
1387 /* Moving base with an active backend?
1388 * You don't want to do that. */
1389 if (vq
->private_data
) {
1393 if (copy_from_user(&s
, argp
, sizeof s
)) {
1397 if (s
.num
> 0xffff) {
1401 vq
->last_avail_idx
= s
.num
;
1402 /* Forget the cached index value. */
1403 vq
->avail_idx
= vq
->last_avail_idx
;
1405 case VHOST_GET_VRING_BASE
:
1407 s
.num
= vq
->last_avail_idx
;
1408 if (copy_to_user(argp
, &s
, sizeof s
))
1411 case VHOST_SET_VRING_ADDR
:
1412 if (copy_from_user(&a
, argp
, sizeof a
)) {
1416 if (a
.flags
& ~(0x1 << VHOST_VRING_F_LOG
)) {
1420 /* For 32bit, verify that the top 32bits of the user
1421 data are set to zero. */
1422 if ((u64
)(unsigned long)a
.desc_user_addr
!= a
.desc_user_addr
||
1423 (u64
)(unsigned long)a
.used_user_addr
!= a
.used_user_addr
||
1424 (u64
)(unsigned long)a
.avail_user_addr
!= a
.avail_user_addr
) {
1429 /* Make sure it's safe to cast pointers to vring types. */
1430 BUILD_BUG_ON(__alignof__
*vq
->avail
> VRING_AVAIL_ALIGN_SIZE
);
1431 BUILD_BUG_ON(__alignof__
*vq
->used
> VRING_USED_ALIGN_SIZE
);
1432 if ((a
.avail_user_addr
& (VRING_AVAIL_ALIGN_SIZE
- 1)) ||
1433 (a
.used_user_addr
& (VRING_USED_ALIGN_SIZE
- 1)) ||
1434 (a
.log_guest_addr
& (VRING_USED_ALIGN_SIZE
- 1))) {
1439 /* We only verify access here if backend is configured.
1440 * If it is not, we don't as size might not have been setup.
1441 * We will verify when backend is configured. */
1442 if (vq
->private_data
) {
1443 if (!vq_access_ok(vq
, vq
->num
,
1444 (void __user
*)(unsigned long)a
.desc_user_addr
,
1445 (void __user
*)(unsigned long)a
.avail_user_addr
,
1446 (void __user
*)(unsigned long)a
.used_user_addr
)) {
1451 /* Also validate log access for used ring if enabled. */
1452 if ((a
.flags
& (0x1 << VHOST_VRING_F_LOG
)) &&
1453 !log_access_ok(vq
->log_base
, a
.log_guest_addr
,
1455 vq
->num
* sizeof *vq
->used
->ring
)) {
1461 vq
->log_used
= !!(a
.flags
& (0x1 << VHOST_VRING_F_LOG
));
1462 vq
->desc
= (void __user
*)(unsigned long)a
.desc_user_addr
;
1463 vq
->avail
= (void __user
*)(unsigned long)a
.avail_user_addr
;
1464 vq
->log_addr
= a
.log_guest_addr
;
1465 vq
->used
= (void __user
*)(unsigned long)a
.used_user_addr
;
1467 case VHOST_SET_VRING_KICK
:
1468 if (copy_from_user(&f
, argp
, sizeof f
)) {
1472 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
1473 if (IS_ERR(eventfp
)) {
1474 r
= PTR_ERR(eventfp
);
1477 if (eventfp
!= vq
->kick
) {
1478 pollstop
= (filep
= vq
->kick
) != NULL
;
1479 pollstart
= (vq
->kick
= eventfp
) != NULL
;
1483 case VHOST_SET_VRING_CALL
:
1484 if (copy_from_user(&f
, argp
, sizeof f
)) {
1488 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
1489 if (IS_ERR(eventfp
)) {
1490 r
= PTR_ERR(eventfp
);
1493 if (eventfp
!= vq
->call
) {
1497 vq
->call_ctx
= eventfp
?
1498 eventfd_ctx_fileget(eventfp
) : NULL
;
1502 case VHOST_SET_VRING_ERR
:
1503 if (copy_from_user(&f
, argp
, sizeof f
)) {
1507 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
1508 if (IS_ERR(eventfp
)) {
1509 r
= PTR_ERR(eventfp
);
1512 if (eventfp
!= vq
->error
) {
1514 vq
->error
= eventfp
;
1515 ctx
= vq
->error_ctx
;
1516 vq
->error_ctx
= eventfp
?
1517 eventfd_ctx_fileget(eventfp
) : NULL
;
1521 case VHOST_SET_VRING_ENDIAN
:
1522 r
= vhost_set_vring_endian(vq
, argp
);
1524 case VHOST_GET_VRING_ENDIAN
:
1525 r
= vhost_get_vring_endian(vq
, idx
, argp
);
1527 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT
:
1528 if (copy_from_user(&s
, argp
, sizeof(s
))) {
1532 vq
->busyloop_timeout
= s
.num
;
1534 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT
:
1536 s
.num
= vq
->busyloop_timeout
;
1537 if (copy_to_user(argp
, &s
, sizeof(s
)))
1544 if (pollstop
&& vq
->handle_kick
)
1545 vhost_poll_stop(&vq
->poll
);
1548 eventfd_ctx_put(ctx
);
1552 if (pollstart
&& vq
->handle_kick
)
1553 r
= vhost_poll_start(&vq
->poll
, vq
->kick
);
1555 mutex_unlock(&vq
->mutex
);
1557 if (pollstop
&& vq
->handle_kick
)
1558 vhost_poll_flush(&vq
->poll
);
1561 EXPORT_SYMBOL_GPL(vhost_vring_ioctl
);
1563 int vhost_init_device_iotlb(struct vhost_dev
*d
, bool enabled
)
1565 struct vhost_umem
*niotlb
, *oiotlb
;
1568 niotlb
= vhost_umem_alloc();
1575 for (i
= 0; i
< d
->nvqs
; ++i
) {
1576 mutex_lock(&d
->vqs
[i
]->mutex
);
1577 d
->vqs
[i
]->iotlb
= niotlb
;
1578 mutex_unlock(&d
->vqs
[i
]->mutex
);
1581 vhost_umem_clean(oiotlb
);
1585 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb
);
1587 /* Caller must have device mutex */
1588 long vhost_dev_ioctl(struct vhost_dev
*d
, unsigned int ioctl
, void __user
*argp
)
1590 struct file
*eventfp
, *filep
= NULL
;
1591 struct eventfd_ctx
*ctx
= NULL
;
1596 /* If you are not the owner, you can become one */
1597 if (ioctl
== VHOST_SET_OWNER
) {
1598 r
= vhost_dev_set_owner(d
);
1602 /* You must be the owner to do anything else */
1603 r
= vhost_dev_check_owner(d
);
1608 case VHOST_SET_MEM_TABLE
:
1609 r
= vhost_set_memory(d
, argp
);
1611 case VHOST_SET_LOG_BASE
:
1612 if (copy_from_user(&p
, argp
, sizeof p
)) {
1616 if ((u64
)(unsigned long)p
!= p
) {
1620 for (i
= 0; i
< d
->nvqs
; ++i
) {
1621 struct vhost_virtqueue
*vq
;
1622 void __user
*base
= (void __user
*)(unsigned long)p
;
1624 mutex_lock(&vq
->mutex
);
1625 /* If ring is inactive, will check when it's enabled. */
1626 if (vq
->private_data
&& !vq_log_access_ok(vq
, base
))
1629 vq
->log_base
= base
;
1630 mutex_unlock(&vq
->mutex
);
1633 case VHOST_SET_LOG_FD
:
1634 r
= get_user(fd
, (int __user
*)argp
);
1637 eventfp
= fd
== -1 ? NULL
: eventfd_fget(fd
);
1638 if (IS_ERR(eventfp
)) {
1639 r
= PTR_ERR(eventfp
);
1642 if (eventfp
!= d
->log_file
) {
1643 filep
= d
->log_file
;
1644 d
->log_file
= eventfp
;
1646 d
->log_ctx
= eventfp
?
1647 eventfd_ctx_fileget(eventfp
) : NULL
;
1650 for (i
= 0; i
< d
->nvqs
; ++i
) {
1651 mutex_lock(&d
->vqs
[i
]->mutex
);
1652 d
->vqs
[i
]->log_ctx
= d
->log_ctx
;
1653 mutex_unlock(&d
->vqs
[i
]->mutex
);
1656 eventfd_ctx_put(ctx
);
1667 EXPORT_SYMBOL_GPL(vhost_dev_ioctl
);
1669 /* TODO: This is really inefficient. We need something like get_user()
1670 * (instruction directly accesses the data, with an exception table entry
1671 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1673 static int set_bit_to_user(int nr
, void __user
*addr
)
1675 unsigned long log
= (unsigned long)addr
;
1678 int bit
= nr
+ (log
% PAGE_SIZE
) * 8;
1681 r
= get_user_pages_fast(log
, 1, 1, &page
);
1685 base
= kmap_atomic(page
);
1687 kunmap_atomic(base
);
1688 set_page_dirty_lock(page
);
1693 static int log_write(void __user
*log_base
,
1694 u64 write_address
, u64 write_length
)
1696 u64 write_page
= write_address
/ VHOST_PAGE_SIZE
;
1701 write_length
+= write_address
% VHOST_PAGE_SIZE
;
1703 u64 base
= (u64
)(unsigned long)log_base
;
1704 u64 log
= base
+ write_page
/ 8;
1705 int bit
= write_page
% 8;
1706 if ((u64
)(unsigned long)log
!= log
)
1708 r
= set_bit_to_user(bit
, (void __user
*)(unsigned long)log
);
1711 if (write_length
<= VHOST_PAGE_SIZE
)
1713 write_length
-= VHOST_PAGE_SIZE
;
1719 int vhost_log_write(struct vhost_virtqueue
*vq
, struct vhost_log
*log
,
1720 unsigned int log_num
, u64 len
)
1724 /* Make sure data written is seen before log. */
1726 for (i
= 0; i
< log_num
; ++i
) {
1727 u64 l
= min(log
[i
].len
, len
);
1728 r
= log_write(vq
->log_base
, log
[i
].addr
, l
);
1734 eventfd_signal(vq
->log_ctx
, 1);
1738 /* Length written exceeds what we have stored. This is a bug. */
1742 EXPORT_SYMBOL_GPL(vhost_log_write
);
1744 static int vhost_update_used_flags(struct vhost_virtqueue
*vq
)
1747 if (vhost_put_user(vq
, cpu_to_vhost16(vq
, vq
->used_flags
),
1748 &vq
->used
->flags
) < 0)
1750 if (unlikely(vq
->log_used
)) {
1751 /* Make sure the flag is seen before log. */
1753 /* Log used flag write. */
1754 used
= &vq
->used
->flags
;
1755 log_write(vq
->log_base
, vq
->log_addr
+
1756 (used
- (void __user
*)vq
->used
),
1757 sizeof vq
->used
->flags
);
1759 eventfd_signal(vq
->log_ctx
, 1);
1764 static int vhost_update_avail_event(struct vhost_virtqueue
*vq
, u16 avail_event
)
1766 if (vhost_put_user(vq
, cpu_to_vhost16(vq
, vq
->avail_idx
),
1767 vhost_avail_event(vq
)))
1769 if (unlikely(vq
->log_used
)) {
1771 /* Make sure the event is seen before log. */
1773 /* Log avail event write */
1774 used
= vhost_avail_event(vq
);
1775 log_write(vq
->log_base
, vq
->log_addr
+
1776 (used
- (void __user
*)vq
->used
),
1777 sizeof *vhost_avail_event(vq
));
1779 eventfd_signal(vq
->log_ctx
, 1);
1784 int vhost_vq_init_access(struct vhost_virtqueue
*vq
)
1786 __virtio16 last_used_idx
;
1788 bool is_le
= vq
->is_le
;
1790 if (!vq
->private_data
)
1793 vhost_init_is_le(vq
);
1795 r
= vhost_update_used_flags(vq
);
1798 vq
->signalled_used_valid
= false;
1800 !access_ok(VERIFY_READ
, &vq
->used
->idx
, sizeof vq
->used
->idx
)) {
1804 r
= vhost_get_used(vq
, last_used_idx
, &vq
->used
->idx
);
1806 vq_err(vq
, "Can't access used idx at %p\n",
1810 vq
->last_used_idx
= vhost16_to_cpu(vq
, last_used_idx
);
1817 EXPORT_SYMBOL_GPL(vhost_vq_init_access
);
1819 static int translate_desc(struct vhost_virtqueue
*vq
, u64 addr
, u32 len
,
1820 struct iovec iov
[], int iov_size
, int access
)
1822 const struct vhost_umem_node
*node
;
1823 struct vhost_dev
*dev
= vq
->dev
;
1824 struct vhost_umem
*umem
= dev
->iotlb
? dev
->iotlb
: dev
->umem
;
1829 while ((u64
)len
> s
) {
1831 if (unlikely(ret
>= iov_size
)) {
1836 node
= vhost_umem_interval_tree_iter_first(&umem
->umem_tree
,
1837 addr
, addr
+ len
- 1);
1838 if (node
== NULL
|| node
->start
> addr
) {
1839 if (umem
!= dev
->iotlb
) {
1845 } else if (!(node
->perm
& access
)) {
1851 size
= node
->size
- addr
+ node
->start
;
1852 _iov
->iov_len
= min((u64
)len
- s
, size
);
1853 _iov
->iov_base
= (void __user
*)(unsigned long)
1854 (node
->userspace_addr
+ addr
- node
->start
);
1861 vhost_iotlb_miss(vq
, addr
, access
);
1865 /* Each buffer in the virtqueues is actually a chain of descriptors. This
1866 * function returns the next descriptor in the chain,
1867 * or -1U if we're at the end. */
1868 static unsigned next_desc(struct vhost_virtqueue
*vq
, struct vring_desc
*desc
)
1872 /* If this descriptor says it doesn't chain, we're done. */
1873 if (!(desc
->flags
& cpu_to_vhost16(vq
, VRING_DESC_F_NEXT
)))
1876 /* Check they're not leading us off end of descriptors. */
1877 next
= vhost16_to_cpu(vq
, desc
->next
);
1878 /* Make sure compiler knows to grab that: we don't want it changing! */
1879 /* We will use the result as an index in an array, so most
1880 * architectures only need a compiler barrier here. */
1881 read_barrier_depends();
1886 static int get_indirect(struct vhost_virtqueue
*vq
,
1887 struct iovec iov
[], unsigned int iov_size
,
1888 unsigned int *out_num
, unsigned int *in_num
,
1889 struct vhost_log
*log
, unsigned int *log_num
,
1890 struct vring_desc
*indirect
)
1892 struct vring_desc desc
;
1893 unsigned int i
= 0, count
, found
= 0;
1894 u32 len
= vhost32_to_cpu(vq
, indirect
->len
);
1895 struct iov_iter from
;
1899 if (unlikely(len
% sizeof desc
)) {
1900 vq_err(vq
, "Invalid length in indirect descriptor: "
1901 "len 0x%llx not multiple of 0x%zx\n",
1902 (unsigned long long)len
,
1907 ret
= translate_desc(vq
, vhost64_to_cpu(vq
, indirect
->addr
), len
, vq
->indirect
,
1908 UIO_MAXIOV
, VHOST_ACCESS_RO
);
1909 if (unlikely(ret
< 0)) {
1911 vq_err(vq
, "Translation failure %d in indirect.\n", ret
);
1914 iov_iter_init(&from
, READ
, vq
->indirect
, ret
, len
);
1916 /* We will use the result as an address to read from, so most
1917 * architectures only need a compiler barrier here. */
1918 read_barrier_depends();
1920 count
= len
/ sizeof desc
;
1921 /* Buffers are chained via a 16 bit next field, so
1922 * we can have at most 2^16 of these. */
1923 if (unlikely(count
> USHRT_MAX
+ 1)) {
1924 vq_err(vq
, "Indirect buffer length too big: %d\n",
1930 unsigned iov_count
= *in_num
+ *out_num
;
1931 if (unlikely(++found
> count
)) {
1932 vq_err(vq
, "Loop detected: last one at %u "
1933 "indirect size %u\n",
1937 if (unlikely(!copy_from_iter_full(&desc
, sizeof(desc
), &from
))) {
1938 vq_err(vq
, "Failed indirect descriptor: idx %d, %zx\n",
1939 i
, (size_t)vhost64_to_cpu(vq
, indirect
->addr
) + i
* sizeof desc
);
1942 if (unlikely(desc
.flags
& cpu_to_vhost16(vq
, VRING_DESC_F_INDIRECT
))) {
1943 vq_err(vq
, "Nested indirect descriptor: idx %d, %zx\n",
1944 i
, (size_t)vhost64_to_cpu(vq
, indirect
->addr
) + i
* sizeof desc
);
1948 if (desc
.flags
& cpu_to_vhost16(vq
, VRING_DESC_F_WRITE
))
1949 access
= VHOST_ACCESS_WO
;
1951 access
= VHOST_ACCESS_RO
;
1953 ret
= translate_desc(vq
, vhost64_to_cpu(vq
, desc
.addr
),
1954 vhost32_to_cpu(vq
, desc
.len
), iov
+ iov_count
,
1955 iov_size
- iov_count
, access
);
1956 if (unlikely(ret
< 0)) {
1958 vq_err(vq
, "Translation failure %d indirect idx %d\n",
1962 /* If this is an input descriptor, increment that count. */
1963 if (access
== VHOST_ACCESS_WO
) {
1965 if (unlikely(log
)) {
1966 log
[*log_num
].addr
= vhost64_to_cpu(vq
, desc
.addr
);
1967 log
[*log_num
].len
= vhost32_to_cpu(vq
, desc
.len
);
1971 /* If it's an output descriptor, they're all supposed
1972 * to come before any input descriptors. */
1973 if (unlikely(*in_num
)) {
1974 vq_err(vq
, "Indirect descriptor "
1975 "has out after in: idx %d\n", i
);
1980 } while ((i
= next_desc(vq
, &desc
)) != -1);
1984 /* This looks in the virtqueue and for the first available buffer, and converts
1985 * it to an iovec for convenient access. Since descriptors consist of some
1986 * number of output then some number of input descriptors, it's actually two
1987 * iovecs, but we pack them into one and note how many of each there were.
1989 * This function returns the descriptor number found, or vq->num (which is
1990 * never a valid descriptor number) if none was found. A negative code is
1991 * returned on error. */
1992 int vhost_get_vq_desc(struct vhost_virtqueue
*vq
,
1993 struct iovec iov
[], unsigned int iov_size
,
1994 unsigned int *out_num
, unsigned int *in_num
,
1995 struct vhost_log
*log
, unsigned int *log_num
)
1997 struct vring_desc desc
;
1998 unsigned int i
, head
, found
= 0;
2000 __virtio16 avail_idx
;
2001 __virtio16 ring_head
;
2004 /* Check it isn't doing very strange things with descriptor numbers. */
2005 last_avail_idx
= vq
->last_avail_idx
;
2007 if (vq
->avail_idx
== vq
->last_avail_idx
) {
2008 if (unlikely(vhost_get_avail(vq
, avail_idx
, &vq
->avail
->idx
))) {
2009 vq_err(vq
, "Failed to access avail idx at %p\n",
2013 vq
->avail_idx
= vhost16_to_cpu(vq
, avail_idx
);
2015 if (unlikely((u16
)(vq
->avail_idx
- last_avail_idx
) > vq
->num
)) {
2016 vq_err(vq
, "Guest moved used index from %u to %u",
2017 last_avail_idx
, vq
->avail_idx
);
2021 /* If there's nothing new since last we looked, return
2024 if (vq
->avail_idx
== last_avail_idx
)
2027 /* Only get avail ring entries after they have been
2033 /* Grab the next descriptor number they're advertising, and increment
2034 * the index we've seen. */
2035 if (unlikely(vhost_get_avail(vq
, ring_head
,
2036 &vq
->avail
->ring
[last_avail_idx
& (vq
->num
- 1)]))) {
2037 vq_err(vq
, "Failed to read head: idx %d address %p\n",
2039 &vq
->avail
->ring
[last_avail_idx
% vq
->num
]);
2043 head
= vhost16_to_cpu(vq
, ring_head
);
2045 /* If their number is silly, that's an error. */
2046 if (unlikely(head
>= vq
->num
)) {
2047 vq_err(vq
, "Guest says index %u > %u is available",
2052 /* When we start there are none of either input nor output. */
2053 *out_num
= *in_num
= 0;
2059 unsigned iov_count
= *in_num
+ *out_num
;
2060 if (unlikely(i
>= vq
->num
)) {
2061 vq_err(vq
, "Desc index is %u > %u, head = %u",
2065 if (unlikely(++found
> vq
->num
)) {
2066 vq_err(vq
, "Loop detected: last one at %u "
2067 "vq size %u head %u\n",
2071 ret
= vhost_copy_from_user(vq
, &desc
, vq
->desc
+ i
,
2073 if (unlikely(ret
)) {
2074 vq_err(vq
, "Failed to get descriptor: idx %d addr %p\n",
2078 if (desc
.flags
& cpu_to_vhost16(vq
, VRING_DESC_F_INDIRECT
)) {
2079 ret
= get_indirect(vq
, iov
, iov_size
,
2081 log
, log_num
, &desc
);
2082 if (unlikely(ret
< 0)) {
2084 vq_err(vq
, "Failure detected "
2085 "in indirect descriptor at idx %d\n", i
);
2091 if (desc
.flags
& cpu_to_vhost16(vq
, VRING_DESC_F_WRITE
))
2092 access
= VHOST_ACCESS_WO
;
2094 access
= VHOST_ACCESS_RO
;
2095 ret
= translate_desc(vq
, vhost64_to_cpu(vq
, desc
.addr
),
2096 vhost32_to_cpu(vq
, desc
.len
), iov
+ iov_count
,
2097 iov_size
- iov_count
, access
);
2098 if (unlikely(ret
< 0)) {
2100 vq_err(vq
, "Translation failure %d descriptor idx %d\n",
2104 if (access
== VHOST_ACCESS_WO
) {
2105 /* If this is an input descriptor,
2106 * increment that count. */
2108 if (unlikely(log
)) {
2109 log
[*log_num
].addr
= vhost64_to_cpu(vq
, desc
.addr
);
2110 log
[*log_num
].len
= vhost32_to_cpu(vq
, desc
.len
);
2114 /* If it's an output descriptor, they're all supposed
2115 * to come before any input descriptors. */
2116 if (unlikely(*in_num
)) {
2117 vq_err(vq
, "Descriptor has out after in: "
2123 } while ((i
= next_desc(vq
, &desc
)) != -1);
2125 /* On success, increment avail index. */
2126 vq
->last_avail_idx
++;
2128 /* Assume notifications from guest are disabled at this point,
2129 * if they aren't we would need to update avail_event index. */
2130 BUG_ON(!(vq
->used_flags
& VRING_USED_F_NO_NOTIFY
));
2133 EXPORT_SYMBOL_GPL(vhost_get_vq_desc
);
2135 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2136 void vhost_discard_vq_desc(struct vhost_virtqueue
*vq
, int n
)
2138 vq
->last_avail_idx
-= n
;
2140 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc
);
2142 /* After we've used one of their buffers, we tell them about it. We'll then
2143 * want to notify the guest, using eventfd. */
2144 int vhost_add_used(struct vhost_virtqueue
*vq
, unsigned int head
, int len
)
2146 struct vring_used_elem heads
= {
2147 cpu_to_vhost32(vq
, head
),
2148 cpu_to_vhost32(vq
, len
)
2151 return vhost_add_used_n(vq
, &heads
, 1);
2153 EXPORT_SYMBOL_GPL(vhost_add_used
);
2155 static int __vhost_add_used_n(struct vhost_virtqueue
*vq
,
2156 struct vring_used_elem
*heads
,
2159 struct vring_used_elem __user
*used
;
2163 start
= vq
->last_used_idx
& (vq
->num
- 1);
2164 used
= vq
->used
->ring
+ start
;
2166 if (vhost_put_user(vq
, heads
[0].id
, &used
->id
)) {
2167 vq_err(vq
, "Failed to write used id");
2170 if (vhost_put_user(vq
, heads
[0].len
, &used
->len
)) {
2171 vq_err(vq
, "Failed to write used len");
2174 } else if (vhost_copy_to_user(vq
, used
, heads
, count
* sizeof *used
)) {
2175 vq_err(vq
, "Failed to write used");
2178 if (unlikely(vq
->log_used
)) {
2179 /* Make sure data is seen before log. */
2181 /* Log used ring entry write. */
2182 log_write(vq
->log_base
,
2184 ((void __user
*)used
- (void __user
*)vq
->used
),
2185 count
* sizeof *used
);
2187 old
= vq
->last_used_idx
;
2188 new = (vq
->last_used_idx
+= count
);
2189 /* If the driver never bothers to signal in a very long while,
2190 * used index might wrap around. If that happens, invalidate
2191 * signalled_used index we stored. TODO: make sure driver
2192 * signals at least once in 2^16 and remove this. */
2193 if (unlikely((u16
)(new - vq
->signalled_used
) < (u16
)(new - old
)))
2194 vq
->signalled_used_valid
= false;
2198 /* After we've used one of their buffers, we tell them about it. We'll then
2199 * want to notify the guest, using eventfd. */
2200 int vhost_add_used_n(struct vhost_virtqueue
*vq
, struct vring_used_elem
*heads
,
2205 start
= vq
->last_used_idx
& (vq
->num
- 1);
2206 n
= vq
->num
- start
;
2208 r
= __vhost_add_used_n(vq
, heads
, n
);
2214 r
= __vhost_add_used_n(vq
, heads
, count
);
2216 /* Make sure buffer is written before we update index. */
2218 if (vhost_put_user(vq
, cpu_to_vhost16(vq
, vq
->last_used_idx
),
2220 vq_err(vq
, "Failed to increment used idx");
2223 if (unlikely(vq
->log_used
)) {
2224 /* Log used index update. */
2225 log_write(vq
->log_base
,
2226 vq
->log_addr
+ offsetof(struct vring_used
, idx
),
2227 sizeof vq
->used
->idx
);
2229 eventfd_signal(vq
->log_ctx
, 1);
2233 EXPORT_SYMBOL_GPL(vhost_add_used_n
);
2235 static bool vhost_notify(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
)
2240 /* Flush out used index updates. This is paired
2241 * with the barrier that the Guest executes when enabling
2245 if (vhost_has_feature(vq
, VIRTIO_F_NOTIFY_ON_EMPTY
) &&
2246 unlikely(vq
->avail_idx
== vq
->last_avail_idx
))
2249 if (!vhost_has_feature(vq
, VIRTIO_RING_F_EVENT_IDX
)) {
2251 if (vhost_get_avail(vq
, flags
, &vq
->avail
->flags
)) {
2252 vq_err(vq
, "Failed to get flags");
2255 return !(flags
& cpu_to_vhost16(vq
, VRING_AVAIL_F_NO_INTERRUPT
));
2257 old
= vq
->signalled_used
;
2258 v
= vq
->signalled_used_valid
;
2259 new = vq
->signalled_used
= vq
->last_used_idx
;
2260 vq
->signalled_used_valid
= true;
2265 if (vhost_get_avail(vq
, event
, vhost_used_event(vq
))) {
2266 vq_err(vq
, "Failed to get used event idx");
2269 return vring_need_event(vhost16_to_cpu(vq
, event
), new, old
);
2272 /* This actually signals the guest, using eventfd. */
2273 void vhost_signal(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
)
2275 /* Signal the Guest tell them we used something up. */
2276 if (vq
->call_ctx
&& vhost_notify(dev
, vq
))
2277 eventfd_signal(vq
->call_ctx
, 1);
2279 EXPORT_SYMBOL_GPL(vhost_signal
);
2281 /* And here's the combo meal deal. Supersize me! */
2282 void vhost_add_used_and_signal(struct vhost_dev
*dev
,
2283 struct vhost_virtqueue
*vq
,
2284 unsigned int head
, int len
)
2286 vhost_add_used(vq
, head
, len
);
2287 vhost_signal(dev
, vq
);
2289 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal
);
2291 /* multi-buffer version of vhost_add_used_and_signal */
2292 void vhost_add_used_and_signal_n(struct vhost_dev
*dev
,
2293 struct vhost_virtqueue
*vq
,
2294 struct vring_used_elem
*heads
, unsigned count
)
2296 vhost_add_used_n(vq
, heads
, count
);
2297 vhost_signal(dev
, vq
);
2299 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n
);
2301 /* return true if we're sure that avaiable ring is empty */
2302 bool vhost_vq_avail_empty(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
)
2304 __virtio16 avail_idx
;
2307 if (vq
->avail_idx
!= vq
->last_avail_idx
)
2310 r
= vhost_get_avail(vq
, avail_idx
, &vq
->avail
->idx
);
2313 vq
->avail_idx
= vhost16_to_cpu(vq
, avail_idx
);
2315 return vq
->avail_idx
== vq
->last_avail_idx
;
2317 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty
);
2319 /* OK, now we need to know about added descriptors. */
2320 bool vhost_enable_notify(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
)
2322 __virtio16 avail_idx
;
2325 if (!(vq
->used_flags
& VRING_USED_F_NO_NOTIFY
))
2327 vq
->used_flags
&= ~VRING_USED_F_NO_NOTIFY
;
2328 if (!vhost_has_feature(vq
, VIRTIO_RING_F_EVENT_IDX
)) {
2329 r
= vhost_update_used_flags(vq
);
2331 vq_err(vq
, "Failed to enable notification at %p: %d\n",
2332 &vq
->used
->flags
, r
);
2336 r
= vhost_update_avail_event(vq
, vq
->avail_idx
);
2338 vq_err(vq
, "Failed to update avail event index at %p: %d\n",
2339 vhost_avail_event(vq
), r
);
2343 /* They could have slipped one in as we were doing that: make
2344 * sure it's written, then check again. */
2346 r
= vhost_get_avail(vq
, avail_idx
, &vq
->avail
->idx
);
2348 vq_err(vq
, "Failed to check avail idx at %p: %d\n",
2349 &vq
->avail
->idx
, r
);
2353 return vhost16_to_cpu(vq
, avail_idx
) != vq
->avail_idx
;
2355 EXPORT_SYMBOL_GPL(vhost_enable_notify
);
2357 /* We don't need to be notified again. */
2358 void vhost_disable_notify(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
)
2362 if (vq
->used_flags
& VRING_USED_F_NO_NOTIFY
)
2364 vq
->used_flags
|= VRING_USED_F_NO_NOTIFY
;
2365 if (!vhost_has_feature(vq
, VIRTIO_RING_F_EVENT_IDX
)) {
2366 r
= vhost_update_used_flags(vq
);
2368 vq_err(vq
, "Failed to enable notification at %p: %d\n",
2369 &vq
->used
->flags
, r
);
2372 EXPORT_SYMBOL_GPL(vhost_disable_notify
);
2374 /* Create a new message. */
2375 struct vhost_msg_node
*vhost_new_msg(struct vhost_virtqueue
*vq
, int type
)
2377 struct vhost_msg_node
*node
= kmalloc(sizeof *node
, GFP_KERNEL
);
2381 node
->msg
.type
= type
;
2384 EXPORT_SYMBOL_GPL(vhost_new_msg
);
2386 void vhost_enqueue_msg(struct vhost_dev
*dev
, struct list_head
*head
,
2387 struct vhost_msg_node
*node
)
2389 spin_lock(&dev
->iotlb_lock
);
2390 list_add_tail(&node
->node
, head
);
2391 spin_unlock(&dev
->iotlb_lock
);
2393 wake_up_interruptible_poll(&dev
->wait
, POLLIN
| POLLRDNORM
);
2395 EXPORT_SYMBOL_GPL(vhost_enqueue_msg
);
2397 struct vhost_msg_node
*vhost_dequeue_msg(struct vhost_dev
*dev
,
2398 struct list_head
*head
)
2400 struct vhost_msg_node
*node
= NULL
;
2402 spin_lock(&dev
->iotlb_lock
);
2403 if (!list_empty(head
)) {
2404 node
= list_first_entry(head
, struct vhost_msg_node
,
2406 list_del(&node
->node
);
2408 spin_unlock(&dev
->iotlb_lock
);
2412 EXPORT_SYMBOL_GPL(vhost_dequeue_msg
);
2415 static int __init
vhost_init(void)
2420 static void __exit
vhost_exit(void)
2424 module_init(vhost_init
);
2425 module_exit(vhost_exit
);
2427 MODULE_VERSION("0.0.1");
2428 MODULE_LICENSE("GPL v2");
2429 MODULE_AUTHOR("Michael S. Tsirkin");
2430 MODULE_DESCRIPTION("Host kernel accelerator for virtio");