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 mutex_lock(&dev
->mutex
);
997 vhost_dev_lock_vqs(dev
);
999 case VHOST_IOTLB_UPDATE
:
1004 if (umem_access_ok(msg
->uaddr
, msg
->size
, msg
->perm
)) {
1008 vhost_vq_meta_reset(dev
);
1009 if (vhost_new_umem_range(dev
->iotlb
, msg
->iova
, msg
->size
,
1010 msg
->iova
+ msg
->size
- 1,
1011 msg
->uaddr
, msg
->perm
)) {
1015 vhost_iotlb_notify_vq(dev
, msg
);
1017 case VHOST_IOTLB_INVALIDATE
:
1022 vhost_vq_meta_reset(dev
);
1023 vhost_del_umem_range(dev
->iotlb
, msg
->iova
,
1024 msg
->iova
+ msg
->size
- 1);
1031 vhost_dev_unlock_vqs(dev
);
1032 mutex_unlock(&dev
->mutex
);
1036 ssize_t
vhost_chr_write_iter(struct vhost_dev
*dev
,
1037 struct iov_iter
*from
)
1039 struct vhost_msg_node node
;
1040 unsigned size
= sizeof(struct vhost_msg
);
1044 if (iov_iter_count(from
) < size
)
1046 ret
= copy_from_iter(&node
.msg
, size
, from
);
1050 switch (node
.msg
.type
) {
1051 case VHOST_IOTLB_MSG
:
1052 err
= vhost_process_iotlb_msg(dev
, &node
.msg
.iotlb
);
1064 EXPORT_SYMBOL(vhost_chr_write_iter
);
1066 unsigned int vhost_chr_poll(struct file
*file
, struct vhost_dev
*dev
,
1069 unsigned int mask
= 0;
1071 poll_wait(file
, &dev
->wait
, wait
);
1073 if (!list_empty(&dev
->read_list
))
1074 mask
|= POLLIN
| POLLRDNORM
;
1078 EXPORT_SYMBOL(vhost_chr_poll
);
1080 ssize_t
vhost_chr_read_iter(struct vhost_dev
*dev
, struct iov_iter
*to
,
1084 struct vhost_msg_node
*node
;
1086 unsigned size
= sizeof(struct vhost_msg
);
1088 if (iov_iter_count(to
) < size
)
1093 prepare_to_wait(&dev
->wait
, &wait
,
1094 TASK_INTERRUPTIBLE
);
1096 node
= vhost_dequeue_msg(dev
, &dev
->read_list
);
1103 if (signal_pending(current
)) {
1116 finish_wait(&dev
->wait
, &wait
);
1119 ret
= copy_to_iter(&node
->msg
, size
, to
);
1121 if (ret
!= size
|| node
->msg
.type
!= VHOST_IOTLB_MISS
) {
1126 vhost_enqueue_msg(dev
, &dev
->pending_list
, node
);
1131 EXPORT_SYMBOL_GPL(vhost_chr_read_iter
);
1133 static int vhost_iotlb_miss(struct vhost_virtqueue
*vq
, u64 iova
, int access
)
1135 struct vhost_dev
*dev
= vq
->dev
;
1136 struct vhost_msg_node
*node
;
1137 struct vhost_iotlb_msg
*msg
;
1139 node
= vhost_new_msg(vq
, VHOST_IOTLB_MISS
);
1143 msg
= &node
->msg
.iotlb
;
1144 msg
->type
= VHOST_IOTLB_MISS
;
1148 vhost_enqueue_msg(dev
, &dev
->read_list
, node
);
1153 static int vq_access_ok(struct vhost_virtqueue
*vq
, unsigned int num
,
1154 struct vring_desc __user
*desc
,
1155 struct vring_avail __user
*avail
,
1156 struct vring_used __user
*used
)
1159 size_t s
= vhost_has_feature(vq
, VIRTIO_RING_F_EVENT_IDX
) ? 2 : 0;
1161 return access_ok(VERIFY_READ
, desc
, num
* sizeof *desc
) &&
1162 access_ok(VERIFY_READ
, avail
,
1163 sizeof *avail
+ num
* sizeof *avail
->ring
+ s
) &&
1164 access_ok(VERIFY_WRITE
, used
,
1165 sizeof *used
+ num
* sizeof *used
->ring
+ s
);
1168 static void vhost_vq_meta_update(struct vhost_virtqueue
*vq
,
1169 const struct vhost_umem_node
*node
,
1172 int access
= (type
== VHOST_ADDR_USED
) ?
1173 VHOST_ACCESS_WO
: VHOST_ACCESS_RO
;
1175 if (likely(node
->perm
& access
))
1176 vq
->meta_iotlb
[type
] = node
;
1179 static int iotlb_access_ok(struct vhost_virtqueue
*vq
,
1180 int access
, u64 addr
, u64 len
, int type
)
1182 const struct vhost_umem_node
*node
;
1183 struct vhost_umem
*umem
= vq
->iotlb
;
1184 u64 s
= 0, size
, orig_addr
= addr
, last
= addr
+ len
- 1;
1186 if (vhost_vq_meta_fetch(vq
, addr
, len
, type
))
1190 node
= vhost_umem_interval_tree_iter_first(&umem
->umem_tree
,
1193 if (node
== NULL
|| node
->start
> addr
) {
1194 vhost_iotlb_miss(vq
, addr
, access
);
1196 } else if (!(node
->perm
& access
)) {
1197 /* Report the possible access violation by
1198 * request another translation from userspace.
1203 size
= node
->size
- addr
+ node
->start
;
1205 if (orig_addr
== addr
&& size
>= len
)
1206 vhost_vq_meta_update(vq
, node
, type
);
1215 int vq_iotlb_prefetch(struct vhost_virtqueue
*vq
)
1217 size_t s
= vhost_has_feature(vq
, VIRTIO_RING_F_EVENT_IDX
) ? 2 : 0;
1218 unsigned int num
= vq
->num
;
1223 return iotlb_access_ok(vq
, VHOST_ACCESS_RO
, (u64
)(uintptr_t)vq
->desc
,
1224 num
* sizeof(*vq
->desc
), VHOST_ADDR_DESC
) &&
1225 iotlb_access_ok(vq
, VHOST_ACCESS_RO
, (u64
)(uintptr_t)vq
->avail
,
1227 num
* sizeof(*vq
->avail
->ring
) + s
,
1228 VHOST_ADDR_AVAIL
) &&
1229 iotlb_access_ok(vq
, VHOST_ACCESS_WO
, (u64
)(uintptr_t)vq
->used
,
1231 num
* sizeof(*vq
->used
->ring
) + s
,
1234 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch
);
1236 /* Can we log writes? */
1237 /* Caller should have device mutex but not vq mutex */
1238 int vhost_log_access_ok(struct vhost_dev
*dev
)
1240 return memory_access_ok(dev
, dev
->umem
, 1);
1242 EXPORT_SYMBOL_GPL(vhost_log_access_ok
);
1244 /* Verify access for write logging. */
1245 /* Caller should have vq mutex and device mutex */
1246 static int vq_log_access_ok(struct vhost_virtqueue
*vq
,
1247 void __user
*log_base
)
1249 size_t s
= vhost_has_feature(vq
, VIRTIO_RING_F_EVENT_IDX
) ? 2 : 0;
1251 return vq_memory_access_ok(log_base
, vq
->umem
,
1252 vhost_has_feature(vq
, VHOST_F_LOG_ALL
)) &&
1253 (!vq
->log_used
|| log_access_ok(log_base
, vq
->log_addr
,
1255 vq
->num
* sizeof *vq
->used
->ring
+ s
));
1258 /* Can we start vq? */
1259 /* Caller should have vq mutex and device mutex */
1260 int vhost_vq_access_ok(struct vhost_virtqueue
*vq
)
1262 if (!vq_log_access_ok(vq
, vq
->log_base
))
1265 /* Access validation occurs at prefetch time with IOTLB */
1269 return vq_access_ok(vq
, vq
->num
, vq
->desc
, vq
->avail
, vq
->used
);
1271 EXPORT_SYMBOL_GPL(vhost_vq_access_ok
);
1273 static struct vhost_umem
*vhost_umem_alloc(void)
1275 struct vhost_umem
*umem
= kvzalloc(sizeof(*umem
), GFP_KERNEL
);
1280 umem
->umem_tree
= RB_ROOT_CACHED
;
1282 INIT_LIST_HEAD(&umem
->umem_list
);
1287 static long vhost_set_memory(struct vhost_dev
*d
, struct vhost_memory __user
*m
)
1289 struct vhost_memory mem
, *newmem
;
1290 struct vhost_memory_region
*region
;
1291 struct vhost_umem
*newumem
, *oldumem
;
1292 unsigned long size
= offsetof(struct vhost_memory
, regions
);
1295 if (copy_from_user(&mem
, m
, size
))
1299 if (mem
.nregions
> max_mem_regions
)
1301 newmem
= kvzalloc(size
+ mem
.nregions
* sizeof(*m
->regions
), GFP_KERNEL
);
1305 memcpy(newmem
, &mem
, size
);
1306 if (copy_from_user(newmem
->regions
, m
->regions
,
1307 mem
.nregions
* sizeof *m
->regions
)) {
1312 newumem
= vhost_umem_alloc();
1318 for (region
= newmem
->regions
;
1319 region
< newmem
->regions
+ mem
.nregions
;
1321 if (vhost_new_umem_range(newumem
,
1322 region
->guest_phys_addr
,
1323 region
->memory_size
,
1324 region
->guest_phys_addr
+
1325 region
->memory_size
- 1,
1326 region
->userspace_addr
,
1331 if (!memory_access_ok(d
, newumem
, 0))
1337 /* All memory accesses are done under some VQ mutex. */
1338 for (i
= 0; i
< d
->nvqs
; ++i
) {
1339 mutex_lock(&d
->vqs
[i
]->mutex
);
1340 d
->vqs
[i
]->umem
= newumem
;
1341 mutex_unlock(&d
->vqs
[i
]->mutex
);
1345 vhost_umem_clean(oldumem
);
1349 vhost_umem_clean(newumem
);
1354 long vhost_vring_ioctl(struct vhost_dev
*d
, int ioctl
, void __user
*argp
)
1356 struct file
*eventfp
, *filep
= NULL
;
1357 bool pollstart
= false, pollstop
= false;
1358 struct eventfd_ctx
*ctx
= NULL
;
1359 u32 __user
*idxp
= argp
;
1360 struct vhost_virtqueue
*vq
;
1361 struct vhost_vring_state s
;
1362 struct vhost_vring_file f
;
1363 struct vhost_vring_addr a
;
1367 r
= get_user(idx
, idxp
);
1375 mutex_lock(&vq
->mutex
);
1378 case VHOST_SET_VRING_NUM
:
1379 /* Resizing ring with an active backend?
1380 * You don't want to do that. */
1381 if (vq
->private_data
) {
1385 if (copy_from_user(&s
, argp
, sizeof s
)) {
1389 if (!s
.num
|| s
.num
> 0xffff || (s
.num
& (s
.num
- 1))) {
1395 case VHOST_SET_VRING_BASE
:
1396 /* Moving base with an active backend?
1397 * You don't want to do that. */
1398 if (vq
->private_data
) {
1402 if (copy_from_user(&s
, argp
, sizeof s
)) {
1406 if (s
.num
> 0xffff) {
1410 vq
->last_avail_idx
= s
.num
;
1411 /* Forget the cached index value. */
1412 vq
->avail_idx
= vq
->last_avail_idx
;
1414 case VHOST_GET_VRING_BASE
:
1416 s
.num
= vq
->last_avail_idx
;
1417 if (copy_to_user(argp
, &s
, sizeof s
))
1420 case VHOST_SET_VRING_ADDR
:
1421 if (copy_from_user(&a
, argp
, sizeof a
)) {
1425 if (a
.flags
& ~(0x1 << VHOST_VRING_F_LOG
)) {
1429 /* For 32bit, verify that the top 32bits of the user
1430 data are set to zero. */
1431 if ((u64
)(unsigned long)a
.desc_user_addr
!= a
.desc_user_addr
||
1432 (u64
)(unsigned long)a
.used_user_addr
!= a
.used_user_addr
||
1433 (u64
)(unsigned long)a
.avail_user_addr
!= a
.avail_user_addr
) {
1438 /* Make sure it's safe to cast pointers to vring types. */
1439 BUILD_BUG_ON(__alignof__
*vq
->avail
> VRING_AVAIL_ALIGN_SIZE
);
1440 BUILD_BUG_ON(__alignof__
*vq
->used
> VRING_USED_ALIGN_SIZE
);
1441 if ((a
.avail_user_addr
& (VRING_AVAIL_ALIGN_SIZE
- 1)) ||
1442 (a
.used_user_addr
& (VRING_USED_ALIGN_SIZE
- 1)) ||
1443 (a
.log_guest_addr
& (VRING_USED_ALIGN_SIZE
- 1))) {
1448 /* We only verify access here if backend is configured.
1449 * If it is not, we don't as size might not have been setup.
1450 * We will verify when backend is configured. */
1451 if (vq
->private_data
) {
1452 if (!vq_access_ok(vq
, vq
->num
,
1453 (void __user
*)(unsigned long)a
.desc_user_addr
,
1454 (void __user
*)(unsigned long)a
.avail_user_addr
,
1455 (void __user
*)(unsigned long)a
.used_user_addr
)) {
1460 /* Also validate log access for used ring if enabled. */
1461 if ((a
.flags
& (0x1 << VHOST_VRING_F_LOG
)) &&
1462 !log_access_ok(vq
->log_base
, a
.log_guest_addr
,
1464 vq
->num
* sizeof *vq
->used
->ring
)) {
1470 vq
->log_used
= !!(a
.flags
& (0x1 << VHOST_VRING_F_LOG
));
1471 vq
->desc
= (void __user
*)(unsigned long)a
.desc_user_addr
;
1472 vq
->avail
= (void __user
*)(unsigned long)a
.avail_user_addr
;
1473 vq
->log_addr
= a
.log_guest_addr
;
1474 vq
->used
= (void __user
*)(unsigned long)a
.used_user_addr
;
1476 case VHOST_SET_VRING_KICK
:
1477 if (copy_from_user(&f
, argp
, sizeof f
)) {
1481 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
1482 if (IS_ERR(eventfp
)) {
1483 r
= PTR_ERR(eventfp
);
1486 if (eventfp
!= vq
->kick
) {
1487 pollstop
= (filep
= vq
->kick
) != NULL
;
1488 pollstart
= (vq
->kick
= eventfp
) != NULL
;
1492 case VHOST_SET_VRING_CALL
:
1493 if (copy_from_user(&f
, argp
, sizeof f
)) {
1497 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
1498 if (IS_ERR(eventfp
)) {
1499 r
= PTR_ERR(eventfp
);
1502 if (eventfp
!= vq
->call
) {
1506 vq
->call_ctx
= eventfp
?
1507 eventfd_ctx_fileget(eventfp
) : NULL
;
1511 case VHOST_SET_VRING_ERR
:
1512 if (copy_from_user(&f
, argp
, sizeof f
)) {
1516 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
1517 if (IS_ERR(eventfp
)) {
1518 r
= PTR_ERR(eventfp
);
1521 if (eventfp
!= vq
->error
) {
1523 vq
->error
= eventfp
;
1524 ctx
= vq
->error_ctx
;
1525 vq
->error_ctx
= eventfp
?
1526 eventfd_ctx_fileget(eventfp
) : NULL
;
1530 case VHOST_SET_VRING_ENDIAN
:
1531 r
= vhost_set_vring_endian(vq
, argp
);
1533 case VHOST_GET_VRING_ENDIAN
:
1534 r
= vhost_get_vring_endian(vq
, idx
, argp
);
1536 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT
:
1537 if (copy_from_user(&s
, argp
, sizeof(s
))) {
1541 vq
->busyloop_timeout
= s
.num
;
1543 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT
:
1545 s
.num
= vq
->busyloop_timeout
;
1546 if (copy_to_user(argp
, &s
, sizeof(s
)))
1553 if (pollstop
&& vq
->handle_kick
)
1554 vhost_poll_stop(&vq
->poll
);
1557 eventfd_ctx_put(ctx
);
1561 if (pollstart
&& vq
->handle_kick
)
1562 r
= vhost_poll_start(&vq
->poll
, vq
->kick
);
1564 mutex_unlock(&vq
->mutex
);
1566 if (pollstop
&& vq
->handle_kick
)
1567 vhost_poll_flush(&vq
->poll
);
1570 EXPORT_SYMBOL_GPL(vhost_vring_ioctl
);
1572 int vhost_init_device_iotlb(struct vhost_dev
*d
, bool enabled
)
1574 struct vhost_umem
*niotlb
, *oiotlb
;
1577 niotlb
= vhost_umem_alloc();
1584 for (i
= 0; i
< d
->nvqs
; ++i
) {
1585 mutex_lock(&d
->vqs
[i
]->mutex
);
1586 d
->vqs
[i
]->iotlb
= niotlb
;
1587 mutex_unlock(&d
->vqs
[i
]->mutex
);
1590 vhost_umem_clean(oiotlb
);
1594 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb
);
1596 /* Caller must have device mutex */
1597 long vhost_dev_ioctl(struct vhost_dev
*d
, unsigned int ioctl
, void __user
*argp
)
1599 struct file
*eventfp
, *filep
= NULL
;
1600 struct eventfd_ctx
*ctx
= NULL
;
1605 /* If you are not the owner, you can become one */
1606 if (ioctl
== VHOST_SET_OWNER
) {
1607 r
= vhost_dev_set_owner(d
);
1611 /* You must be the owner to do anything else */
1612 r
= vhost_dev_check_owner(d
);
1617 case VHOST_SET_MEM_TABLE
:
1618 r
= vhost_set_memory(d
, argp
);
1620 case VHOST_SET_LOG_BASE
:
1621 if (copy_from_user(&p
, argp
, sizeof p
)) {
1625 if ((u64
)(unsigned long)p
!= p
) {
1629 for (i
= 0; i
< d
->nvqs
; ++i
) {
1630 struct vhost_virtqueue
*vq
;
1631 void __user
*base
= (void __user
*)(unsigned long)p
;
1633 mutex_lock(&vq
->mutex
);
1634 /* If ring is inactive, will check when it's enabled. */
1635 if (vq
->private_data
&& !vq_log_access_ok(vq
, base
))
1638 vq
->log_base
= base
;
1639 mutex_unlock(&vq
->mutex
);
1642 case VHOST_SET_LOG_FD
:
1643 r
= get_user(fd
, (int __user
*)argp
);
1646 eventfp
= fd
== -1 ? NULL
: eventfd_fget(fd
);
1647 if (IS_ERR(eventfp
)) {
1648 r
= PTR_ERR(eventfp
);
1651 if (eventfp
!= d
->log_file
) {
1652 filep
= d
->log_file
;
1653 d
->log_file
= eventfp
;
1655 d
->log_ctx
= eventfp
?
1656 eventfd_ctx_fileget(eventfp
) : NULL
;
1659 for (i
= 0; i
< d
->nvqs
; ++i
) {
1660 mutex_lock(&d
->vqs
[i
]->mutex
);
1661 d
->vqs
[i
]->log_ctx
= d
->log_ctx
;
1662 mutex_unlock(&d
->vqs
[i
]->mutex
);
1665 eventfd_ctx_put(ctx
);
1676 EXPORT_SYMBOL_GPL(vhost_dev_ioctl
);
1678 /* TODO: This is really inefficient. We need something like get_user()
1679 * (instruction directly accesses the data, with an exception table entry
1680 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1682 static int set_bit_to_user(int nr
, void __user
*addr
)
1684 unsigned long log
= (unsigned long)addr
;
1687 int bit
= nr
+ (log
% PAGE_SIZE
) * 8;
1690 r
= get_user_pages_fast(log
, 1, 1, &page
);
1694 base
= kmap_atomic(page
);
1696 kunmap_atomic(base
);
1697 set_page_dirty_lock(page
);
1702 static int log_write(void __user
*log_base
,
1703 u64 write_address
, u64 write_length
)
1705 u64 write_page
= write_address
/ VHOST_PAGE_SIZE
;
1710 write_length
+= write_address
% VHOST_PAGE_SIZE
;
1712 u64 base
= (u64
)(unsigned long)log_base
;
1713 u64 log
= base
+ write_page
/ 8;
1714 int bit
= write_page
% 8;
1715 if ((u64
)(unsigned long)log
!= log
)
1717 r
= set_bit_to_user(bit
, (void __user
*)(unsigned long)log
);
1720 if (write_length
<= VHOST_PAGE_SIZE
)
1722 write_length
-= VHOST_PAGE_SIZE
;
1728 int vhost_log_write(struct vhost_virtqueue
*vq
, struct vhost_log
*log
,
1729 unsigned int log_num
, u64 len
)
1733 /* Make sure data written is seen before log. */
1735 for (i
= 0; i
< log_num
; ++i
) {
1736 u64 l
= min(log
[i
].len
, len
);
1737 r
= log_write(vq
->log_base
, log
[i
].addr
, l
);
1743 eventfd_signal(vq
->log_ctx
, 1);
1747 /* Length written exceeds what we have stored. This is a bug. */
1751 EXPORT_SYMBOL_GPL(vhost_log_write
);
1753 static int vhost_update_used_flags(struct vhost_virtqueue
*vq
)
1756 if (vhost_put_user(vq
, cpu_to_vhost16(vq
, vq
->used_flags
),
1757 &vq
->used
->flags
) < 0)
1759 if (unlikely(vq
->log_used
)) {
1760 /* Make sure the flag is seen before log. */
1762 /* Log used flag write. */
1763 used
= &vq
->used
->flags
;
1764 log_write(vq
->log_base
, vq
->log_addr
+
1765 (used
- (void __user
*)vq
->used
),
1766 sizeof vq
->used
->flags
);
1768 eventfd_signal(vq
->log_ctx
, 1);
1773 static int vhost_update_avail_event(struct vhost_virtqueue
*vq
, u16 avail_event
)
1775 if (vhost_put_user(vq
, cpu_to_vhost16(vq
, vq
->avail_idx
),
1776 vhost_avail_event(vq
)))
1778 if (unlikely(vq
->log_used
)) {
1780 /* Make sure the event is seen before log. */
1782 /* Log avail event write */
1783 used
= vhost_avail_event(vq
);
1784 log_write(vq
->log_base
, vq
->log_addr
+
1785 (used
- (void __user
*)vq
->used
),
1786 sizeof *vhost_avail_event(vq
));
1788 eventfd_signal(vq
->log_ctx
, 1);
1793 int vhost_vq_init_access(struct vhost_virtqueue
*vq
)
1795 __virtio16 last_used_idx
;
1797 bool is_le
= vq
->is_le
;
1799 if (!vq
->private_data
)
1802 vhost_init_is_le(vq
);
1804 r
= vhost_update_used_flags(vq
);
1807 vq
->signalled_used_valid
= false;
1809 !access_ok(VERIFY_READ
, &vq
->used
->idx
, sizeof vq
->used
->idx
)) {
1813 r
= vhost_get_used(vq
, last_used_idx
, &vq
->used
->idx
);
1815 vq_err(vq
, "Can't access used idx at %p\n",
1819 vq
->last_used_idx
= vhost16_to_cpu(vq
, last_used_idx
);
1826 EXPORT_SYMBOL_GPL(vhost_vq_init_access
);
1828 static int translate_desc(struct vhost_virtqueue
*vq
, u64 addr
, u32 len
,
1829 struct iovec iov
[], int iov_size
, int access
)
1831 const struct vhost_umem_node
*node
;
1832 struct vhost_dev
*dev
= vq
->dev
;
1833 struct vhost_umem
*umem
= dev
->iotlb
? dev
->iotlb
: dev
->umem
;
1838 while ((u64
)len
> s
) {
1840 if (unlikely(ret
>= iov_size
)) {
1845 node
= vhost_umem_interval_tree_iter_first(&umem
->umem_tree
,
1846 addr
, addr
+ len
- 1);
1847 if (node
== NULL
|| node
->start
> addr
) {
1848 if (umem
!= dev
->iotlb
) {
1854 } else if (!(node
->perm
& access
)) {
1860 size
= node
->size
- addr
+ node
->start
;
1861 _iov
->iov_len
= min((u64
)len
- s
, size
);
1862 _iov
->iov_base
= (void __user
*)(unsigned long)
1863 (node
->userspace_addr
+ addr
- node
->start
);
1870 vhost_iotlb_miss(vq
, addr
, access
);
1874 /* Each buffer in the virtqueues is actually a chain of descriptors. This
1875 * function returns the next descriptor in the chain,
1876 * or -1U if we're at the end. */
1877 static unsigned next_desc(struct vhost_virtqueue
*vq
, struct vring_desc
*desc
)
1881 /* If this descriptor says it doesn't chain, we're done. */
1882 if (!(desc
->flags
& cpu_to_vhost16(vq
, VRING_DESC_F_NEXT
)))
1885 /* Check they're not leading us off end of descriptors. */
1886 next
= vhost16_to_cpu(vq
, desc
->next
);
1887 /* Make sure compiler knows to grab that: we don't want it changing! */
1888 /* We will use the result as an index in an array, so most
1889 * architectures only need a compiler barrier here. */
1890 read_barrier_depends();
1895 static int get_indirect(struct vhost_virtqueue
*vq
,
1896 struct iovec iov
[], unsigned int iov_size
,
1897 unsigned int *out_num
, unsigned int *in_num
,
1898 struct vhost_log
*log
, unsigned int *log_num
,
1899 struct vring_desc
*indirect
)
1901 struct vring_desc desc
;
1902 unsigned int i
= 0, count
, found
= 0;
1903 u32 len
= vhost32_to_cpu(vq
, indirect
->len
);
1904 struct iov_iter from
;
1908 if (unlikely(len
% sizeof desc
)) {
1909 vq_err(vq
, "Invalid length in indirect descriptor: "
1910 "len 0x%llx not multiple of 0x%zx\n",
1911 (unsigned long long)len
,
1916 ret
= translate_desc(vq
, vhost64_to_cpu(vq
, indirect
->addr
), len
, vq
->indirect
,
1917 UIO_MAXIOV
, VHOST_ACCESS_RO
);
1918 if (unlikely(ret
< 0)) {
1920 vq_err(vq
, "Translation failure %d in indirect.\n", ret
);
1923 iov_iter_init(&from
, READ
, vq
->indirect
, ret
, len
);
1925 /* We will use the result as an address to read from, so most
1926 * architectures only need a compiler barrier here. */
1927 read_barrier_depends();
1929 count
= len
/ sizeof desc
;
1930 /* Buffers are chained via a 16 bit next field, so
1931 * we can have at most 2^16 of these. */
1932 if (unlikely(count
> USHRT_MAX
+ 1)) {
1933 vq_err(vq
, "Indirect buffer length too big: %d\n",
1939 unsigned iov_count
= *in_num
+ *out_num
;
1940 if (unlikely(++found
> count
)) {
1941 vq_err(vq
, "Loop detected: last one at %u "
1942 "indirect size %u\n",
1946 if (unlikely(!copy_from_iter_full(&desc
, sizeof(desc
), &from
))) {
1947 vq_err(vq
, "Failed indirect descriptor: idx %d, %zx\n",
1948 i
, (size_t)vhost64_to_cpu(vq
, indirect
->addr
) + i
* sizeof desc
);
1951 if (unlikely(desc
.flags
& cpu_to_vhost16(vq
, VRING_DESC_F_INDIRECT
))) {
1952 vq_err(vq
, "Nested indirect descriptor: idx %d, %zx\n",
1953 i
, (size_t)vhost64_to_cpu(vq
, indirect
->addr
) + i
* sizeof desc
);
1957 if (desc
.flags
& cpu_to_vhost16(vq
, VRING_DESC_F_WRITE
))
1958 access
= VHOST_ACCESS_WO
;
1960 access
= VHOST_ACCESS_RO
;
1962 ret
= translate_desc(vq
, vhost64_to_cpu(vq
, desc
.addr
),
1963 vhost32_to_cpu(vq
, desc
.len
), iov
+ iov_count
,
1964 iov_size
- iov_count
, access
);
1965 if (unlikely(ret
< 0)) {
1967 vq_err(vq
, "Translation failure %d indirect idx %d\n",
1971 /* If this is an input descriptor, increment that count. */
1972 if (access
== VHOST_ACCESS_WO
) {
1974 if (unlikely(log
)) {
1975 log
[*log_num
].addr
= vhost64_to_cpu(vq
, desc
.addr
);
1976 log
[*log_num
].len
= vhost32_to_cpu(vq
, desc
.len
);
1980 /* If it's an output descriptor, they're all supposed
1981 * to come before any input descriptors. */
1982 if (unlikely(*in_num
)) {
1983 vq_err(vq
, "Indirect descriptor "
1984 "has out after in: idx %d\n", i
);
1989 } while ((i
= next_desc(vq
, &desc
)) != -1);
1993 /* This looks in the virtqueue and for the first available buffer, and converts
1994 * it to an iovec for convenient access. Since descriptors consist of some
1995 * number of output then some number of input descriptors, it's actually two
1996 * iovecs, but we pack them into one and note how many of each there were.
1998 * This function returns the descriptor number found, or vq->num (which is
1999 * never a valid descriptor number) if none was found. A negative code is
2000 * returned on error. */
2001 int vhost_get_vq_desc(struct vhost_virtqueue
*vq
,
2002 struct iovec iov
[], unsigned int iov_size
,
2003 unsigned int *out_num
, unsigned int *in_num
,
2004 struct vhost_log
*log
, unsigned int *log_num
)
2006 struct vring_desc desc
;
2007 unsigned int i
, head
, found
= 0;
2009 __virtio16 avail_idx
;
2010 __virtio16 ring_head
;
2013 /* Check it isn't doing very strange things with descriptor numbers. */
2014 last_avail_idx
= vq
->last_avail_idx
;
2016 if (vq
->avail_idx
== vq
->last_avail_idx
) {
2017 if (unlikely(vhost_get_avail(vq
, avail_idx
, &vq
->avail
->idx
))) {
2018 vq_err(vq
, "Failed to access avail idx at %p\n",
2022 vq
->avail_idx
= vhost16_to_cpu(vq
, avail_idx
);
2024 if (unlikely((u16
)(vq
->avail_idx
- last_avail_idx
) > vq
->num
)) {
2025 vq_err(vq
, "Guest moved used index from %u to %u",
2026 last_avail_idx
, vq
->avail_idx
);
2030 /* If there's nothing new since last we looked, return
2033 if (vq
->avail_idx
== last_avail_idx
)
2036 /* Only get avail ring entries after they have been
2042 /* Grab the next descriptor number they're advertising, and increment
2043 * the index we've seen. */
2044 if (unlikely(vhost_get_avail(vq
, ring_head
,
2045 &vq
->avail
->ring
[last_avail_idx
& (vq
->num
- 1)]))) {
2046 vq_err(vq
, "Failed to read head: idx %d address %p\n",
2048 &vq
->avail
->ring
[last_avail_idx
% vq
->num
]);
2052 head
= vhost16_to_cpu(vq
, ring_head
);
2054 /* If their number is silly, that's an error. */
2055 if (unlikely(head
>= vq
->num
)) {
2056 vq_err(vq
, "Guest says index %u > %u is available",
2061 /* When we start there are none of either input nor output. */
2062 *out_num
= *in_num
= 0;
2068 unsigned iov_count
= *in_num
+ *out_num
;
2069 if (unlikely(i
>= vq
->num
)) {
2070 vq_err(vq
, "Desc index is %u > %u, head = %u",
2074 if (unlikely(++found
> vq
->num
)) {
2075 vq_err(vq
, "Loop detected: last one at %u "
2076 "vq size %u head %u\n",
2080 ret
= vhost_copy_from_user(vq
, &desc
, vq
->desc
+ i
,
2082 if (unlikely(ret
)) {
2083 vq_err(vq
, "Failed to get descriptor: idx %d addr %p\n",
2087 if (desc
.flags
& cpu_to_vhost16(vq
, VRING_DESC_F_INDIRECT
)) {
2088 ret
= get_indirect(vq
, iov
, iov_size
,
2090 log
, log_num
, &desc
);
2091 if (unlikely(ret
< 0)) {
2093 vq_err(vq
, "Failure detected "
2094 "in indirect descriptor at idx %d\n", i
);
2100 if (desc
.flags
& cpu_to_vhost16(vq
, VRING_DESC_F_WRITE
))
2101 access
= VHOST_ACCESS_WO
;
2103 access
= VHOST_ACCESS_RO
;
2104 ret
= translate_desc(vq
, vhost64_to_cpu(vq
, desc
.addr
),
2105 vhost32_to_cpu(vq
, desc
.len
), iov
+ iov_count
,
2106 iov_size
- iov_count
, access
);
2107 if (unlikely(ret
< 0)) {
2109 vq_err(vq
, "Translation failure %d descriptor idx %d\n",
2113 if (access
== VHOST_ACCESS_WO
) {
2114 /* If this is an input descriptor,
2115 * increment that count. */
2117 if (unlikely(log
)) {
2118 log
[*log_num
].addr
= vhost64_to_cpu(vq
, desc
.addr
);
2119 log
[*log_num
].len
= vhost32_to_cpu(vq
, desc
.len
);
2123 /* If it's an output descriptor, they're all supposed
2124 * to come before any input descriptors. */
2125 if (unlikely(*in_num
)) {
2126 vq_err(vq
, "Descriptor has out after in: "
2132 } while ((i
= next_desc(vq
, &desc
)) != -1);
2134 /* On success, increment avail index. */
2135 vq
->last_avail_idx
++;
2137 /* Assume notifications from guest are disabled at this point,
2138 * if they aren't we would need to update avail_event index. */
2139 BUG_ON(!(vq
->used_flags
& VRING_USED_F_NO_NOTIFY
));
2142 EXPORT_SYMBOL_GPL(vhost_get_vq_desc
);
2144 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2145 void vhost_discard_vq_desc(struct vhost_virtqueue
*vq
, int n
)
2147 vq
->last_avail_idx
-= n
;
2149 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc
);
2151 /* After we've used one of their buffers, we tell them about it. We'll then
2152 * want to notify the guest, using eventfd. */
2153 int vhost_add_used(struct vhost_virtqueue
*vq
, unsigned int head
, int len
)
2155 struct vring_used_elem heads
= {
2156 cpu_to_vhost32(vq
, head
),
2157 cpu_to_vhost32(vq
, len
)
2160 return vhost_add_used_n(vq
, &heads
, 1);
2162 EXPORT_SYMBOL_GPL(vhost_add_used
);
2164 static int __vhost_add_used_n(struct vhost_virtqueue
*vq
,
2165 struct vring_used_elem
*heads
,
2168 struct vring_used_elem __user
*used
;
2172 start
= vq
->last_used_idx
& (vq
->num
- 1);
2173 used
= vq
->used
->ring
+ start
;
2175 if (vhost_put_user(vq
, heads
[0].id
, &used
->id
)) {
2176 vq_err(vq
, "Failed to write used id");
2179 if (vhost_put_user(vq
, heads
[0].len
, &used
->len
)) {
2180 vq_err(vq
, "Failed to write used len");
2183 } else if (vhost_copy_to_user(vq
, used
, heads
, count
* sizeof *used
)) {
2184 vq_err(vq
, "Failed to write used");
2187 if (unlikely(vq
->log_used
)) {
2188 /* Make sure data is seen before log. */
2190 /* Log used ring entry write. */
2191 log_write(vq
->log_base
,
2193 ((void __user
*)used
- (void __user
*)vq
->used
),
2194 count
* sizeof *used
);
2196 old
= vq
->last_used_idx
;
2197 new = (vq
->last_used_idx
+= count
);
2198 /* If the driver never bothers to signal in a very long while,
2199 * used index might wrap around. If that happens, invalidate
2200 * signalled_used index we stored. TODO: make sure driver
2201 * signals at least once in 2^16 and remove this. */
2202 if (unlikely((u16
)(new - vq
->signalled_used
) < (u16
)(new - old
)))
2203 vq
->signalled_used_valid
= false;
2207 /* After we've used one of their buffers, we tell them about it. We'll then
2208 * want to notify the guest, using eventfd. */
2209 int vhost_add_used_n(struct vhost_virtqueue
*vq
, struct vring_used_elem
*heads
,
2214 start
= vq
->last_used_idx
& (vq
->num
- 1);
2215 n
= vq
->num
- start
;
2217 r
= __vhost_add_used_n(vq
, heads
, n
);
2223 r
= __vhost_add_used_n(vq
, heads
, count
);
2225 /* Make sure buffer is written before we update index. */
2227 if (vhost_put_user(vq
, cpu_to_vhost16(vq
, vq
->last_used_idx
),
2229 vq_err(vq
, "Failed to increment used idx");
2232 if (unlikely(vq
->log_used
)) {
2233 /* Log used index update. */
2234 log_write(vq
->log_base
,
2235 vq
->log_addr
+ offsetof(struct vring_used
, idx
),
2236 sizeof vq
->used
->idx
);
2238 eventfd_signal(vq
->log_ctx
, 1);
2242 EXPORT_SYMBOL_GPL(vhost_add_used_n
);
2244 static bool vhost_notify(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
)
2249 /* Flush out used index updates. This is paired
2250 * with the barrier that the Guest executes when enabling
2254 if (vhost_has_feature(vq
, VIRTIO_F_NOTIFY_ON_EMPTY
) &&
2255 unlikely(vq
->avail_idx
== vq
->last_avail_idx
))
2258 if (!vhost_has_feature(vq
, VIRTIO_RING_F_EVENT_IDX
)) {
2260 if (vhost_get_avail(vq
, flags
, &vq
->avail
->flags
)) {
2261 vq_err(vq
, "Failed to get flags");
2264 return !(flags
& cpu_to_vhost16(vq
, VRING_AVAIL_F_NO_INTERRUPT
));
2266 old
= vq
->signalled_used
;
2267 v
= vq
->signalled_used_valid
;
2268 new = vq
->signalled_used
= vq
->last_used_idx
;
2269 vq
->signalled_used_valid
= true;
2274 if (vhost_get_avail(vq
, event
, vhost_used_event(vq
))) {
2275 vq_err(vq
, "Failed to get used event idx");
2278 return vring_need_event(vhost16_to_cpu(vq
, event
), new, old
);
2281 /* This actually signals the guest, using eventfd. */
2282 void vhost_signal(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
)
2284 /* Signal the Guest tell them we used something up. */
2285 if (vq
->call_ctx
&& vhost_notify(dev
, vq
))
2286 eventfd_signal(vq
->call_ctx
, 1);
2288 EXPORT_SYMBOL_GPL(vhost_signal
);
2290 /* And here's the combo meal deal. Supersize me! */
2291 void vhost_add_used_and_signal(struct vhost_dev
*dev
,
2292 struct vhost_virtqueue
*vq
,
2293 unsigned int head
, int len
)
2295 vhost_add_used(vq
, head
, len
);
2296 vhost_signal(dev
, vq
);
2298 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal
);
2300 /* multi-buffer version of vhost_add_used_and_signal */
2301 void vhost_add_used_and_signal_n(struct vhost_dev
*dev
,
2302 struct vhost_virtqueue
*vq
,
2303 struct vring_used_elem
*heads
, unsigned count
)
2305 vhost_add_used_n(vq
, heads
, count
);
2306 vhost_signal(dev
, vq
);
2308 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n
);
2310 /* return true if we're sure that avaiable ring is empty */
2311 bool vhost_vq_avail_empty(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
)
2313 __virtio16 avail_idx
;
2316 if (vq
->avail_idx
!= vq
->last_avail_idx
)
2319 r
= vhost_get_avail(vq
, avail_idx
, &vq
->avail
->idx
);
2322 vq
->avail_idx
= vhost16_to_cpu(vq
, avail_idx
);
2324 return vq
->avail_idx
== vq
->last_avail_idx
;
2326 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty
);
2328 /* OK, now we need to know about added descriptors. */
2329 bool vhost_enable_notify(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
)
2331 __virtio16 avail_idx
;
2334 if (!(vq
->used_flags
& VRING_USED_F_NO_NOTIFY
))
2336 vq
->used_flags
&= ~VRING_USED_F_NO_NOTIFY
;
2337 if (!vhost_has_feature(vq
, VIRTIO_RING_F_EVENT_IDX
)) {
2338 r
= vhost_update_used_flags(vq
);
2340 vq_err(vq
, "Failed to enable notification at %p: %d\n",
2341 &vq
->used
->flags
, r
);
2345 r
= vhost_update_avail_event(vq
, vq
->avail_idx
);
2347 vq_err(vq
, "Failed to update avail event index at %p: %d\n",
2348 vhost_avail_event(vq
), r
);
2352 /* They could have slipped one in as we were doing that: make
2353 * sure it's written, then check again. */
2355 r
= vhost_get_avail(vq
, avail_idx
, &vq
->avail
->idx
);
2357 vq_err(vq
, "Failed to check avail idx at %p: %d\n",
2358 &vq
->avail
->idx
, r
);
2362 return vhost16_to_cpu(vq
, avail_idx
) != vq
->avail_idx
;
2364 EXPORT_SYMBOL_GPL(vhost_enable_notify
);
2366 /* We don't need to be notified again. */
2367 void vhost_disable_notify(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
)
2371 if (vq
->used_flags
& VRING_USED_F_NO_NOTIFY
)
2373 vq
->used_flags
|= VRING_USED_F_NO_NOTIFY
;
2374 if (!vhost_has_feature(vq
, VIRTIO_RING_F_EVENT_IDX
)) {
2375 r
= vhost_update_used_flags(vq
);
2377 vq_err(vq
, "Failed to enable notification at %p: %d\n",
2378 &vq
->used
->flags
, r
);
2381 EXPORT_SYMBOL_GPL(vhost_disable_notify
);
2383 /* Create a new message. */
2384 struct vhost_msg_node
*vhost_new_msg(struct vhost_virtqueue
*vq
, int type
)
2386 struct vhost_msg_node
*node
= kmalloc(sizeof *node
, GFP_KERNEL
);
2390 /* Make sure all padding within the structure is initialized. */
2391 memset(&node
->msg
, 0, sizeof node
->msg
);
2393 node
->msg
.type
= type
;
2396 EXPORT_SYMBOL_GPL(vhost_new_msg
);
2398 void vhost_enqueue_msg(struct vhost_dev
*dev
, struct list_head
*head
,
2399 struct vhost_msg_node
*node
)
2401 spin_lock(&dev
->iotlb_lock
);
2402 list_add_tail(&node
->node
, head
);
2403 spin_unlock(&dev
->iotlb_lock
);
2405 wake_up_interruptible_poll(&dev
->wait
, POLLIN
| POLLRDNORM
);
2407 EXPORT_SYMBOL_GPL(vhost_enqueue_msg
);
2409 struct vhost_msg_node
*vhost_dequeue_msg(struct vhost_dev
*dev
,
2410 struct list_head
*head
)
2412 struct vhost_msg_node
*node
= NULL
;
2414 spin_lock(&dev
->iotlb_lock
);
2415 if (!list_empty(head
)) {
2416 node
= list_first_entry(head
, struct vhost_msg_node
,
2418 list_del(&node
->node
);
2420 spin_unlock(&dev
->iotlb_lock
);
2424 EXPORT_SYMBOL_GPL(vhost_dequeue_msg
);
2427 static int __init
vhost_init(void)
2432 static void __exit
vhost_exit(void)
2436 module_init(vhost_init
);
2437 module_exit(vhost_exit
);
2439 MODULE_VERSION("0.0.1");
2440 MODULE_LICENSE("GPL v2");
2441 MODULE_AUTHOR("Michael S. Tsirkin");
2442 MODULE_DESCRIPTION("Host kernel accelerator for virtio");