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/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/virtio_net.h>
18 #include <linux/miscdevice.h>
19 #include <linux/mutex.h>
20 #include <linux/rcupdate.h>
21 #include <linux/poll.h>
22 #include <linux/file.h>
23 #include <linux/highmem.h>
24 #include <linux/slab.h>
25 #include <linux/kthread.h>
26 #include <linux/cgroup.h>
28 #include <linux/net.h>
29 #include <linux/if_packet.h>
30 #include <linux/if_arp.h>
37 VHOST_MEMORY_MAX_NREGIONS
= 64,
38 VHOST_MEMORY_F_LOG
= 0x1,
41 static void vhost_poll_func(struct file
*file
, wait_queue_head_t
*wqh
,
44 struct vhost_poll
*poll
;
45 poll
= container_of(pt
, struct vhost_poll
, table
);
48 add_wait_queue(wqh
, &poll
->wait
);
51 static int vhost_poll_wakeup(wait_queue_t
*wait
, unsigned mode
, int sync
,
54 struct vhost_poll
*poll
= container_of(wait
, struct vhost_poll
, wait
);
56 if (!((unsigned long)key
& poll
->mask
))
59 vhost_poll_queue(poll
);
63 static void vhost_work_init(struct vhost_work
*work
, vhost_work_fn_t fn
)
65 INIT_LIST_HEAD(&work
->node
);
67 init_waitqueue_head(&work
->done
);
69 work
->queue_seq
= work
->done_seq
= 0;
72 /* Init poll structure */
73 void vhost_poll_init(struct vhost_poll
*poll
, vhost_work_fn_t fn
,
74 unsigned long mask
, struct vhost_dev
*dev
)
76 init_waitqueue_func_entry(&poll
->wait
, vhost_poll_wakeup
);
77 init_poll_funcptr(&poll
->table
, vhost_poll_func
);
81 vhost_work_init(&poll
->work
, fn
);
84 /* Start polling a file. We add ourselves to file's wait queue. The caller must
85 * keep a reference to a file until after vhost_poll_stop is called. */
86 void vhost_poll_start(struct vhost_poll
*poll
, struct file
*file
)
89 mask
= file
->f_op
->poll(file
, &poll
->table
);
91 vhost_poll_wakeup(&poll
->wait
, 0, 0, (void *)mask
);
94 /* Stop polling a file. After this function returns, it becomes safe to drop the
95 * file reference. You must also flush afterwards. */
96 void vhost_poll_stop(struct vhost_poll
*poll
)
98 remove_wait_queue(poll
->wqh
, &poll
->wait
);
101 static void vhost_work_flush(struct vhost_dev
*dev
, struct vhost_work
*work
)
107 spin_lock_irq(&dev
->work_lock
);
108 seq
= work
->queue_seq
;
110 spin_unlock_irq(&dev
->work_lock
);
111 wait_event(work
->done
, ({
112 spin_lock_irq(&dev
->work_lock
);
113 left
= seq
- work
->done_seq
<= 0;
114 spin_unlock_irq(&dev
->work_lock
);
117 spin_lock_irq(&dev
->work_lock
);
118 flushing
= --work
->flushing
;
119 spin_unlock_irq(&dev
->work_lock
);
120 BUG_ON(flushing
< 0);
123 /* Flush any work that has been scheduled. When calling this, don't hold any
124 * locks that are also used by the callback. */
125 void vhost_poll_flush(struct vhost_poll
*poll
)
127 vhost_work_flush(poll
->dev
, &poll
->work
);
130 static inline void vhost_work_queue(struct vhost_dev
*dev
,
131 struct vhost_work
*work
)
135 spin_lock_irqsave(&dev
->work_lock
, flags
);
136 if (list_empty(&work
->node
)) {
137 list_add_tail(&work
->node
, &dev
->work_list
);
139 wake_up_process(dev
->worker
);
141 spin_unlock_irqrestore(&dev
->work_lock
, flags
);
144 void vhost_poll_queue(struct vhost_poll
*poll
)
146 vhost_work_queue(poll
->dev
, &poll
->work
);
149 static void vhost_vq_reset(struct vhost_dev
*dev
,
150 struct vhost_virtqueue
*vq
)
156 vq
->last_avail_idx
= 0;
158 vq
->last_used_idx
= 0;
161 vq
->log_used
= false;
162 vq
->log_addr
= -1ull;
165 vq
->private_data
= NULL
;
167 vq
->error_ctx
= NULL
;
175 static int vhost_worker(void *data
)
177 struct vhost_dev
*dev
= data
;
178 struct vhost_work
*work
= NULL
;
179 unsigned uninitialized_var(seq
);
182 /* mb paired w/ kthread_stop */
183 set_current_state(TASK_INTERRUPTIBLE
);
185 spin_lock_irq(&dev
->work_lock
);
187 work
->done_seq
= seq
;
189 wake_up_all(&work
->done
);
192 if (kthread_should_stop()) {
193 spin_unlock_irq(&dev
->work_lock
);
194 __set_current_state(TASK_RUNNING
);
197 if (!list_empty(&dev
->work_list
)) {
198 work
= list_first_entry(&dev
->work_list
,
199 struct vhost_work
, node
);
200 list_del_init(&work
->node
);
201 seq
= work
->queue_seq
;
204 spin_unlock_irq(&dev
->work_lock
);
207 __set_current_state(TASK_RUNNING
);
215 long vhost_dev_init(struct vhost_dev
*dev
,
216 struct vhost_virtqueue
*vqs
, int nvqs
)
222 mutex_init(&dev
->mutex
);
224 dev
->log_file
= NULL
;
227 spin_lock_init(&dev
->work_lock
);
228 INIT_LIST_HEAD(&dev
->work_list
);
231 for (i
= 0; i
< dev
->nvqs
; ++i
) {
232 dev
->vqs
[i
].dev
= dev
;
233 mutex_init(&dev
->vqs
[i
].mutex
);
234 vhost_vq_reset(dev
, dev
->vqs
+ i
);
235 if (dev
->vqs
[i
].handle_kick
)
236 vhost_poll_init(&dev
->vqs
[i
].poll
,
237 dev
->vqs
[i
].handle_kick
, POLLIN
, dev
);
243 /* Caller should have device mutex */
244 long vhost_dev_check_owner(struct vhost_dev
*dev
)
246 /* Are you the owner? If not, I don't think you mean to do that */
247 return dev
->mm
== current
->mm
? 0 : -EPERM
;
250 struct vhost_attach_cgroups_struct
{
251 struct vhost_work work
;
252 struct task_struct
*owner
;
256 static void vhost_attach_cgroups_work(struct vhost_work
*work
)
258 struct vhost_attach_cgroups_struct
*s
;
259 s
= container_of(work
, struct vhost_attach_cgroups_struct
, work
);
260 s
->ret
= cgroup_attach_task_all(s
->owner
, current
);
263 static int vhost_attach_cgroups(struct vhost_dev
*dev
)
265 struct vhost_attach_cgroups_struct attach
;
266 attach
.owner
= current
;
267 vhost_work_init(&attach
.work
, vhost_attach_cgroups_work
);
268 vhost_work_queue(dev
, &attach
.work
);
269 vhost_work_flush(dev
, &attach
.work
);
273 /* Caller should have device mutex */
274 static long vhost_dev_set_owner(struct vhost_dev
*dev
)
276 struct task_struct
*worker
;
278 /* Is there an owner already? */
283 /* No owner, become one */
284 dev
->mm
= get_task_mm(current
);
285 worker
= kthread_create(vhost_worker
, dev
, "vhost-%d", current
->pid
);
286 if (IS_ERR(worker
)) {
287 err
= PTR_ERR(worker
);
291 dev
->worker
= worker
;
292 wake_up_process(worker
); /* avoid contributing to loadavg */
294 err
= vhost_attach_cgroups(dev
);
300 kthread_stop(worker
);
309 /* Caller should have device mutex */
310 long vhost_dev_reset_owner(struct vhost_dev
*dev
)
312 struct vhost_memory
*memory
;
314 /* Restore memory to default empty mapping. */
315 memory
= kmalloc(offsetof(struct vhost_memory
, regions
), GFP_KERNEL
);
319 vhost_dev_cleanup(dev
);
321 memory
->nregions
= 0;
322 dev
->memory
= memory
;
326 /* Caller should have device mutex */
327 void vhost_dev_cleanup(struct vhost_dev
*dev
)
330 for (i
= 0; i
< dev
->nvqs
; ++i
) {
331 if (dev
->vqs
[i
].kick
&& dev
->vqs
[i
].handle_kick
) {
332 vhost_poll_stop(&dev
->vqs
[i
].poll
);
333 vhost_poll_flush(&dev
->vqs
[i
].poll
);
335 if (dev
->vqs
[i
].error_ctx
)
336 eventfd_ctx_put(dev
->vqs
[i
].error_ctx
);
337 if (dev
->vqs
[i
].error
)
338 fput(dev
->vqs
[i
].error
);
339 if (dev
->vqs
[i
].kick
)
340 fput(dev
->vqs
[i
].kick
);
341 if (dev
->vqs
[i
].call_ctx
)
342 eventfd_ctx_put(dev
->vqs
[i
].call_ctx
);
343 if (dev
->vqs
[i
].call
)
344 fput(dev
->vqs
[i
].call
);
345 vhost_vq_reset(dev
, dev
->vqs
+ i
);
348 eventfd_ctx_put(dev
->log_ctx
);
352 dev
->log_file
= NULL
;
353 /* No one will access memory at this point */
360 WARN_ON(!list_empty(&dev
->work_list
));
362 kthread_stop(dev
->worker
);
367 static int log_access_ok(void __user
*log_base
, u64 addr
, unsigned long sz
)
369 u64 a
= addr
/ VHOST_PAGE_SIZE
/ 8;
370 /* Make sure 64 bit math will not overflow. */
371 if (a
> ULONG_MAX
- (unsigned long)log_base
||
372 a
+ (unsigned long)log_base
> ULONG_MAX
)
375 return access_ok(VERIFY_WRITE
, log_base
+ a
,
376 (sz
+ VHOST_PAGE_SIZE
* 8 - 1) / VHOST_PAGE_SIZE
/ 8);
379 /* Caller should have vq mutex and device mutex. */
380 static int vq_memory_access_ok(void __user
*log_base
, struct vhost_memory
*mem
,
388 for (i
= 0; i
< mem
->nregions
; ++i
) {
389 struct vhost_memory_region
*m
= mem
->regions
+ i
;
390 unsigned long a
= m
->userspace_addr
;
391 if (m
->memory_size
> ULONG_MAX
)
393 else if (!access_ok(VERIFY_WRITE
, (void __user
*)a
,
396 else if (log_all
&& !log_access_ok(log_base
,
404 /* Can we switch to this memory table? */
405 /* Caller should have device mutex but not vq mutex */
406 static int memory_access_ok(struct vhost_dev
*d
, struct vhost_memory
*mem
,
410 for (i
= 0; i
< d
->nvqs
; ++i
) {
412 mutex_lock(&d
->vqs
[i
].mutex
);
413 /* If ring is inactive, will check when it's enabled. */
414 if (d
->vqs
[i
].private_data
)
415 ok
= vq_memory_access_ok(d
->vqs
[i
].log_base
, mem
,
419 mutex_unlock(&d
->vqs
[i
].mutex
);
426 static int vq_access_ok(unsigned int num
,
427 struct vring_desc __user
*desc
,
428 struct vring_avail __user
*avail
,
429 struct vring_used __user
*used
)
431 return access_ok(VERIFY_READ
, desc
, num
* sizeof *desc
) &&
432 access_ok(VERIFY_READ
, avail
,
433 sizeof *avail
+ num
* sizeof *avail
->ring
) &&
434 access_ok(VERIFY_WRITE
, used
,
435 sizeof *used
+ num
* sizeof *used
->ring
);
438 /* Can we log writes? */
439 /* Caller should have device mutex but not vq mutex */
440 int vhost_log_access_ok(struct vhost_dev
*dev
)
442 return memory_access_ok(dev
, dev
->memory
, 1);
445 /* Verify access for write logging. */
446 /* Caller should have vq mutex and device mutex */
447 static int vq_log_access_ok(struct vhost_virtqueue
*vq
, void __user
*log_base
)
449 return vq_memory_access_ok(log_base
, vq
->dev
->memory
,
450 vhost_has_feature(vq
->dev
, VHOST_F_LOG_ALL
)) &&
451 (!vq
->log_used
|| log_access_ok(log_base
, vq
->log_addr
,
453 vq
->num
* sizeof *vq
->used
->ring
));
456 /* Can we start vq? */
457 /* Caller should have vq mutex and device mutex */
458 int vhost_vq_access_ok(struct vhost_virtqueue
*vq
)
460 return vq_access_ok(vq
->num
, vq
->desc
, vq
->avail
, vq
->used
) &&
461 vq_log_access_ok(vq
, vq
->log_base
);
464 static long vhost_set_memory(struct vhost_dev
*d
, struct vhost_memory __user
*m
)
466 struct vhost_memory mem
, *newmem
, *oldmem
;
467 unsigned long size
= offsetof(struct vhost_memory
, regions
);
468 if (copy_from_user(&mem
, m
, size
))
472 if (mem
.nregions
> VHOST_MEMORY_MAX_NREGIONS
)
474 newmem
= kmalloc(size
+ mem
.nregions
* sizeof *m
->regions
, GFP_KERNEL
);
478 memcpy(newmem
, &mem
, size
);
479 if (copy_from_user(newmem
->regions
, m
->regions
,
480 mem
.nregions
* sizeof *m
->regions
)) {
485 if (!memory_access_ok(d
, newmem
, vhost_has_feature(d
, VHOST_F_LOG_ALL
))) {
490 rcu_assign_pointer(d
->memory
, newmem
);
496 static int init_used(struct vhost_virtqueue
*vq
,
497 struct vring_used __user
*used
)
499 int r
= put_user(vq
->used_flags
, &used
->flags
);
502 return get_user(vq
->last_used_idx
, &used
->idx
);
505 static long vhost_set_vring(struct vhost_dev
*d
, int ioctl
, void __user
*argp
)
507 struct file
*eventfp
, *filep
= NULL
,
508 *pollstart
= NULL
, *pollstop
= NULL
;
509 struct eventfd_ctx
*ctx
= NULL
;
510 u32 __user
*idxp
= argp
;
511 struct vhost_virtqueue
*vq
;
512 struct vhost_vring_state s
;
513 struct vhost_vring_file f
;
514 struct vhost_vring_addr a
;
518 r
= get_user(idx
, idxp
);
526 mutex_lock(&vq
->mutex
);
529 case VHOST_SET_VRING_NUM
:
530 /* Resizing ring with an active backend?
531 * You don't want to do that. */
532 if (vq
->private_data
) {
536 if (copy_from_user(&s
, argp
, sizeof s
)) {
540 if (!s
.num
|| s
.num
> 0xffff || (s
.num
& (s
.num
- 1))) {
546 case VHOST_SET_VRING_BASE
:
547 /* Moving base with an active backend?
548 * You don't want to do that. */
549 if (vq
->private_data
) {
553 if (copy_from_user(&s
, argp
, sizeof s
)) {
557 if (s
.num
> 0xffff) {
561 vq
->last_avail_idx
= s
.num
;
562 /* Forget the cached index value. */
563 vq
->avail_idx
= vq
->last_avail_idx
;
565 case VHOST_GET_VRING_BASE
:
567 s
.num
= vq
->last_avail_idx
;
568 if (copy_to_user(argp
, &s
, sizeof s
))
571 case VHOST_SET_VRING_ADDR
:
572 if (copy_from_user(&a
, argp
, sizeof a
)) {
576 if (a
.flags
& ~(0x1 << VHOST_VRING_F_LOG
)) {
580 /* For 32bit, verify that the top 32bits of the user
581 data are set to zero. */
582 if ((u64
)(unsigned long)a
.desc_user_addr
!= a
.desc_user_addr
||
583 (u64
)(unsigned long)a
.used_user_addr
!= a
.used_user_addr
||
584 (u64
)(unsigned long)a
.avail_user_addr
!= a
.avail_user_addr
) {
588 if ((a
.avail_user_addr
& (sizeof *vq
->avail
->ring
- 1)) ||
589 (a
.used_user_addr
& (sizeof *vq
->used
->ring
- 1)) ||
590 (a
.log_guest_addr
& (sizeof *vq
->used
->ring
- 1))) {
595 /* We only verify access here if backend is configured.
596 * If it is not, we don't as size might not have been setup.
597 * We will verify when backend is configured. */
598 if (vq
->private_data
) {
599 if (!vq_access_ok(vq
->num
,
600 (void __user
*)(unsigned long)a
.desc_user_addr
,
601 (void __user
*)(unsigned long)a
.avail_user_addr
,
602 (void __user
*)(unsigned long)a
.used_user_addr
)) {
607 /* Also validate log access for used ring if enabled. */
608 if ((a
.flags
& (0x1 << VHOST_VRING_F_LOG
)) &&
609 !log_access_ok(vq
->log_base
, a
.log_guest_addr
,
611 vq
->num
* sizeof *vq
->used
->ring
)) {
617 r
= init_used(vq
, (struct vring_used __user
*)(unsigned long)
621 vq
->log_used
= !!(a
.flags
& (0x1 << VHOST_VRING_F_LOG
));
622 vq
->desc
= (void __user
*)(unsigned long)a
.desc_user_addr
;
623 vq
->avail
= (void __user
*)(unsigned long)a
.avail_user_addr
;
624 vq
->log_addr
= a
.log_guest_addr
;
625 vq
->used
= (void __user
*)(unsigned long)a
.used_user_addr
;
627 case VHOST_SET_VRING_KICK
:
628 if (copy_from_user(&f
, argp
, sizeof f
)) {
632 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
633 if (IS_ERR(eventfp
)) {
634 r
= PTR_ERR(eventfp
);
637 if (eventfp
!= vq
->kick
) {
638 pollstop
= filep
= vq
->kick
;
639 pollstart
= vq
->kick
= eventfp
;
643 case VHOST_SET_VRING_CALL
:
644 if (copy_from_user(&f
, argp
, sizeof f
)) {
648 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
649 if (IS_ERR(eventfp
)) {
650 r
= PTR_ERR(eventfp
);
653 if (eventfp
!= vq
->call
) {
657 vq
->call_ctx
= eventfp
?
658 eventfd_ctx_fileget(eventfp
) : NULL
;
662 case VHOST_SET_VRING_ERR
:
663 if (copy_from_user(&f
, argp
, sizeof f
)) {
667 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
668 if (IS_ERR(eventfp
)) {
669 r
= PTR_ERR(eventfp
);
672 if (eventfp
!= vq
->error
) {
676 vq
->error_ctx
= eventfp
?
677 eventfd_ctx_fileget(eventfp
) : NULL
;
685 if (pollstop
&& vq
->handle_kick
)
686 vhost_poll_stop(&vq
->poll
);
689 eventfd_ctx_put(ctx
);
693 if (pollstart
&& vq
->handle_kick
)
694 vhost_poll_start(&vq
->poll
, vq
->kick
);
696 mutex_unlock(&vq
->mutex
);
698 if (pollstop
&& vq
->handle_kick
)
699 vhost_poll_flush(&vq
->poll
);
703 /* Caller must have device mutex */
704 long vhost_dev_ioctl(struct vhost_dev
*d
, unsigned int ioctl
, unsigned long arg
)
706 void __user
*argp
= (void __user
*)arg
;
707 struct file
*eventfp
, *filep
= NULL
;
708 struct eventfd_ctx
*ctx
= NULL
;
713 /* If you are not the owner, you can become one */
714 if (ioctl
== VHOST_SET_OWNER
) {
715 r
= vhost_dev_set_owner(d
);
719 /* You must be the owner to do anything else */
720 r
= vhost_dev_check_owner(d
);
725 case VHOST_SET_MEM_TABLE
:
726 r
= vhost_set_memory(d
, argp
);
728 case VHOST_SET_LOG_BASE
:
729 if (copy_from_user(&p
, argp
, sizeof p
)) {
733 if ((u64
)(unsigned long)p
!= p
) {
737 for (i
= 0; i
< d
->nvqs
; ++i
) {
738 struct vhost_virtqueue
*vq
;
739 void __user
*base
= (void __user
*)(unsigned long)p
;
741 mutex_lock(&vq
->mutex
);
742 /* If ring is inactive, will check when it's enabled. */
743 if (vq
->private_data
&& !vq_log_access_ok(vq
, base
))
747 mutex_unlock(&vq
->mutex
);
750 case VHOST_SET_LOG_FD
:
751 r
= get_user(fd
, (int __user
*)argp
);
754 eventfp
= fd
== -1 ? NULL
: eventfd_fget(fd
);
755 if (IS_ERR(eventfp
)) {
756 r
= PTR_ERR(eventfp
);
759 if (eventfp
!= d
->log_file
) {
762 d
->log_ctx
= eventfp
?
763 eventfd_ctx_fileget(eventfp
) : NULL
;
766 for (i
= 0; i
< d
->nvqs
; ++i
) {
767 mutex_lock(&d
->vqs
[i
].mutex
);
768 d
->vqs
[i
].log_ctx
= d
->log_ctx
;
769 mutex_unlock(&d
->vqs
[i
].mutex
);
772 eventfd_ctx_put(ctx
);
777 r
= vhost_set_vring(d
, ioctl
, argp
);
784 static const struct vhost_memory_region
*find_region(struct vhost_memory
*mem
,
785 __u64 addr
, __u32 len
)
787 struct vhost_memory_region
*reg
;
789 /* linear search is not brilliant, but we really have on the order of 6
790 * regions in practice */
791 for (i
= 0; i
< mem
->nregions
; ++i
) {
792 reg
= mem
->regions
+ i
;
793 if (reg
->guest_phys_addr
<= addr
&&
794 reg
->guest_phys_addr
+ reg
->memory_size
- 1 >= addr
)
800 /* TODO: This is really inefficient. We need something like get_user()
801 * (instruction directly accesses the data, with an exception table entry
802 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
804 static int set_bit_to_user(int nr
, void __user
*addr
)
806 unsigned long log
= (unsigned long)addr
;
809 int bit
= nr
+ (log
% PAGE_SIZE
) * 8;
811 r
= get_user_pages_fast(log
, 1, 1, &page
);
815 base
= kmap_atomic(page
, KM_USER0
);
817 kunmap_atomic(base
, KM_USER0
);
818 set_page_dirty_lock(page
);
823 static int log_write(void __user
*log_base
,
824 u64 write_address
, u64 write_length
)
829 write_address
/= VHOST_PAGE_SIZE
;
831 u64 base
= (u64
)(unsigned long)log_base
;
832 u64 log
= base
+ write_address
/ 8;
833 int bit
= write_address
% 8;
834 if ((u64
)(unsigned long)log
!= log
)
836 r
= set_bit_to_user(bit
, (void __user
*)(unsigned long)log
);
839 if (write_length
<= VHOST_PAGE_SIZE
)
841 write_length
-= VHOST_PAGE_SIZE
;
842 write_address
+= VHOST_PAGE_SIZE
;
847 int vhost_log_write(struct vhost_virtqueue
*vq
, struct vhost_log
*log
,
848 unsigned int log_num
, u64 len
)
852 /* Make sure data written is seen before log. */
854 for (i
= 0; i
< log_num
; ++i
) {
855 u64 l
= min(log
[i
].len
, len
);
856 r
= log_write(vq
->log_base
, log
[i
].addr
, l
);
864 eventfd_signal(vq
->log_ctx
, 1);
865 /* Length written exceeds what we have stored. This is a bug. */
870 static int translate_desc(struct vhost_dev
*dev
, u64 addr
, u32 len
,
871 struct iovec iov
[], int iov_size
)
873 const struct vhost_memory_region
*reg
;
874 struct vhost_memory
*mem
;
881 mem
= rcu_dereference(dev
->memory
);
882 while ((u64
)len
> s
) {
884 if (unlikely(ret
>= iov_size
)) {
888 reg
= find_region(mem
, addr
, len
);
889 if (unlikely(!reg
)) {
894 size
= reg
->memory_size
- addr
+ reg
->guest_phys_addr
;
895 _iov
->iov_len
= min((u64
)len
, size
);
896 _iov
->iov_base
= (void __user
*)(unsigned long)
897 (reg
->userspace_addr
+ addr
- reg
->guest_phys_addr
);
907 /* Each buffer in the virtqueues is actually a chain of descriptors. This
908 * function returns the next descriptor in the chain,
909 * or -1U if we're at the end. */
910 static unsigned next_desc(struct vring_desc
*desc
)
914 /* If this descriptor says it doesn't chain, we're done. */
915 if (!(desc
->flags
& VRING_DESC_F_NEXT
))
918 /* Check they're not leading us off end of descriptors. */
920 /* Make sure compiler knows to grab that: we don't want it changing! */
921 /* We will use the result as an index in an array, so most
922 * architectures only need a compiler barrier here. */
923 read_barrier_depends();
928 static int get_indirect(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
,
929 struct iovec iov
[], unsigned int iov_size
,
930 unsigned int *out_num
, unsigned int *in_num
,
931 struct vhost_log
*log
, unsigned int *log_num
,
932 struct vring_desc
*indirect
)
934 struct vring_desc desc
;
935 unsigned int i
= 0, count
, found
= 0;
939 if (unlikely(indirect
->len
% sizeof desc
)) {
940 vq_err(vq
, "Invalid length in indirect descriptor: "
941 "len 0x%llx not multiple of 0x%zx\n",
942 (unsigned long long)indirect
->len
,
947 ret
= translate_desc(dev
, indirect
->addr
, indirect
->len
, vq
->indirect
,
948 ARRAY_SIZE(vq
->indirect
));
949 if (unlikely(ret
< 0)) {
950 vq_err(vq
, "Translation failure %d in indirect.\n", ret
);
954 /* We will use the result as an address to read from, so most
955 * architectures only need a compiler barrier here. */
956 read_barrier_depends();
958 count
= indirect
->len
/ sizeof desc
;
959 /* Buffers are chained via a 16 bit next field, so
960 * we can have at most 2^16 of these. */
961 if (unlikely(count
> USHRT_MAX
+ 1)) {
962 vq_err(vq
, "Indirect buffer length too big: %d\n",
968 unsigned iov_count
= *in_num
+ *out_num
;
969 if (unlikely(++found
> count
)) {
970 vq_err(vq
, "Loop detected: last one at %u "
971 "indirect size %u\n",
975 if (unlikely(memcpy_fromiovec((unsigned char *)&desc
, vq
->indirect
,
977 vq_err(vq
, "Failed indirect descriptor: idx %d, %zx\n",
978 i
, (size_t)indirect
->addr
+ i
* sizeof desc
);
981 if (unlikely(desc
.flags
& VRING_DESC_F_INDIRECT
)) {
982 vq_err(vq
, "Nested indirect descriptor: idx %d, %zx\n",
983 i
, (size_t)indirect
->addr
+ i
* sizeof desc
);
987 ret
= translate_desc(dev
, desc
.addr
, desc
.len
, iov
+ iov_count
,
988 iov_size
- iov_count
);
989 if (unlikely(ret
< 0)) {
990 vq_err(vq
, "Translation failure %d indirect idx %d\n",
994 /* If this is an input descriptor, increment that count. */
995 if (desc
.flags
& VRING_DESC_F_WRITE
) {
998 log
[*log_num
].addr
= desc
.addr
;
999 log
[*log_num
].len
= desc
.len
;
1003 /* If it's an output descriptor, they're all supposed
1004 * to come before any input descriptors. */
1005 if (unlikely(*in_num
)) {
1006 vq_err(vq
, "Indirect descriptor "
1007 "has out after in: idx %d\n", i
);
1012 } while ((i
= next_desc(&desc
)) != -1);
1016 /* This looks in the virtqueue and for the first available buffer, and converts
1017 * it to an iovec for convenient access. Since descriptors consist of some
1018 * number of output then some number of input descriptors, it's actually two
1019 * iovecs, but we pack them into one and note how many of each there were.
1021 * This function returns the descriptor number found, or vq->num (which is
1022 * never a valid descriptor number) if none was found. A negative code is
1023 * returned on error. */
1024 int vhost_get_vq_desc(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
,
1025 struct iovec iov
[], unsigned int iov_size
,
1026 unsigned int *out_num
, unsigned int *in_num
,
1027 struct vhost_log
*log
, unsigned int *log_num
)
1029 struct vring_desc desc
;
1030 unsigned int i
, head
, found
= 0;
1034 /* Check it isn't doing very strange things with descriptor numbers. */
1035 last_avail_idx
= vq
->last_avail_idx
;
1036 if (unlikely(get_user(vq
->avail_idx
, &vq
->avail
->idx
))) {
1037 vq_err(vq
, "Failed to access avail idx at %p\n",
1042 if (unlikely((u16
)(vq
->avail_idx
- last_avail_idx
) > vq
->num
)) {
1043 vq_err(vq
, "Guest moved used index from %u to %u",
1044 last_avail_idx
, vq
->avail_idx
);
1048 /* If there's nothing new since last we looked, return invalid. */
1049 if (vq
->avail_idx
== last_avail_idx
)
1052 /* Only get avail ring entries after they have been exposed by guest. */
1055 /* Grab the next descriptor number they're advertising, and increment
1056 * the index we've seen. */
1057 if (unlikely(get_user(head
,
1058 &vq
->avail
->ring
[last_avail_idx
% vq
->num
]))) {
1059 vq_err(vq
, "Failed to read head: idx %d address %p\n",
1061 &vq
->avail
->ring
[last_avail_idx
% vq
->num
]);
1065 /* If their number is silly, that's an error. */
1066 if (unlikely(head
>= vq
->num
)) {
1067 vq_err(vq
, "Guest says index %u > %u is available",
1072 /* When we start there are none of either input nor output. */
1073 *out_num
= *in_num
= 0;
1079 unsigned iov_count
= *in_num
+ *out_num
;
1080 if (unlikely(i
>= vq
->num
)) {
1081 vq_err(vq
, "Desc index is %u > %u, head = %u",
1085 if (unlikely(++found
> vq
->num
)) {
1086 vq_err(vq
, "Loop detected: last one at %u "
1087 "vq size %u head %u\n",
1091 ret
= copy_from_user(&desc
, vq
->desc
+ i
, sizeof desc
);
1092 if (unlikely(ret
)) {
1093 vq_err(vq
, "Failed to get descriptor: idx %d addr %p\n",
1097 if (desc
.flags
& VRING_DESC_F_INDIRECT
) {
1098 ret
= get_indirect(dev
, vq
, iov
, iov_size
,
1100 log
, log_num
, &desc
);
1101 if (unlikely(ret
< 0)) {
1102 vq_err(vq
, "Failure detected "
1103 "in indirect descriptor at idx %d\n", i
);
1109 ret
= translate_desc(dev
, desc
.addr
, desc
.len
, iov
+ iov_count
,
1110 iov_size
- iov_count
);
1111 if (unlikely(ret
< 0)) {
1112 vq_err(vq
, "Translation failure %d descriptor idx %d\n",
1116 if (desc
.flags
& VRING_DESC_F_WRITE
) {
1117 /* If this is an input descriptor,
1118 * increment that count. */
1120 if (unlikely(log
)) {
1121 log
[*log_num
].addr
= desc
.addr
;
1122 log
[*log_num
].len
= desc
.len
;
1126 /* If it's an output descriptor, they're all supposed
1127 * to come before any input descriptors. */
1128 if (unlikely(*in_num
)) {
1129 vq_err(vq
, "Descriptor has out after in: "
1135 } while ((i
= next_desc(&desc
)) != -1);
1137 /* On success, increment avail index. */
1138 vq
->last_avail_idx
++;
1142 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1143 void vhost_discard_vq_desc(struct vhost_virtqueue
*vq
, int n
)
1145 vq
->last_avail_idx
-= n
;
1148 /* After we've used one of their buffers, we tell them about it. We'll then
1149 * want to notify the guest, using eventfd. */
1150 int vhost_add_used(struct vhost_virtqueue
*vq
, unsigned int head
, int len
)
1152 struct vring_used_elem __user
*used
;
1154 /* The virtqueue contains a ring of used buffers. Get a pointer to the
1155 * next entry in that used ring. */
1156 used
= &vq
->used
->ring
[vq
->last_used_idx
% vq
->num
];
1157 if (put_user(head
, &used
->id
)) {
1158 vq_err(vq
, "Failed to write used id");
1161 if (put_user(len
, &used
->len
)) {
1162 vq_err(vq
, "Failed to write used len");
1165 /* Make sure buffer is written before we update index. */
1167 if (put_user(vq
->last_used_idx
+ 1, &vq
->used
->idx
)) {
1168 vq_err(vq
, "Failed to increment used idx");
1171 if (unlikely(vq
->log_used
)) {
1172 /* Make sure data is seen before log. */
1174 /* Log used ring entry write. */
1175 log_write(vq
->log_base
,
1177 ((void __user
*)used
- (void __user
*)vq
->used
),
1179 /* Log used index update. */
1180 log_write(vq
->log_base
,
1181 vq
->log_addr
+ offsetof(struct vring_used
, idx
),
1182 sizeof vq
->used
->idx
);
1184 eventfd_signal(vq
->log_ctx
, 1);
1186 vq
->last_used_idx
++;
1190 static int __vhost_add_used_n(struct vhost_virtqueue
*vq
,
1191 struct vring_used_elem
*heads
,
1194 struct vring_used_elem __user
*used
;
1197 start
= vq
->last_used_idx
% vq
->num
;
1198 used
= vq
->used
->ring
+ start
;
1199 if (copy_to_user(used
, heads
, count
* sizeof *used
)) {
1200 vq_err(vq
, "Failed to write used");
1203 if (unlikely(vq
->log_used
)) {
1204 /* Make sure data is seen before log. */
1206 /* Log used ring entry write. */
1207 log_write(vq
->log_base
,
1209 ((void __user
*)used
- (void __user
*)vq
->used
),
1210 count
* sizeof *used
);
1212 vq
->last_used_idx
+= count
;
1216 /* After we've used one of their buffers, we tell them about it. We'll then
1217 * want to notify the guest, using eventfd. */
1218 int vhost_add_used_n(struct vhost_virtqueue
*vq
, struct vring_used_elem
*heads
,
1223 start
= vq
->last_used_idx
% vq
->num
;
1224 n
= vq
->num
- start
;
1226 r
= __vhost_add_used_n(vq
, heads
, n
);
1232 r
= __vhost_add_used_n(vq
, heads
, count
);
1234 /* Make sure buffer is written before we update index. */
1236 if (put_user(vq
->last_used_idx
, &vq
->used
->idx
)) {
1237 vq_err(vq
, "Failed to increment used idx");
1240 if (unlikely(vq
->log_used
)) {
1241 /* Log used index update. */
1242 log_write(vq
->log_base
,
1243 vq
->log_addr
+ offsetof(struct vring_used
, idx
),
1244 sizeof vq
->used
->idx
);
1246 eventfd_signal(vq
->log_ctx
, 1);
1251 /* This actually signals the guest, using eventfd. */
1252 void vhost_signal(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
)
1255 /* Flush out used index updates. This is paired
1256 * with the barrier that the Guest executes when enabling
1260 if (get_user(flags
, &vq
->avail
->flags
)) {
1261 vq_err(vq
, "Failed to get flags");
1265 /* If they don't want an interrupt, don't signal, unless empty. */
1266 if ((flags
& VRING_AVAIL_F_NO_INTERRUPT
) &&
1267 (vq
->avail_idx
!= vq
->last_avail_idx
||
1268 !vhost_has_feature(dev
, VIRTIO_F_NOTIFY_ON_EMPTY
)))
1271 /* Signal the Guest tell them we used something up. */
1273 eventfd_signal(vq
->call_ctx
, 1);
1276 /* And here's the combo meal deal. Supersize me! */
1277 void vhost_add_used_and_signal(struct vhost_dev
*dev
,
1278 struct vhost_virtqueue
*vq
,
1279 unsigned int head
, int len
)
1281 vhost_add_used(vq
, head
, len
);
1282 vhost_signal(dev
, vq
);
1285 /* multi-buffer version of vhost_add_used_and_signal */
1286 void vhost_add_used_and_signal_n(struct vhost_dev
*dev
,
1287 struct vhost_virtqueue
*vq
,
1288 struct vring_used_elem
*heads
, unsigned count
)
1290 vhost_add_used_n(vq
, heads
, count
);
1291 vhost_signal(dev
, vq
);
1294 /* OK, now we need to know about added descriptors. */
1295 bool vhost_enable_notify(struct vhost_virtqueue
*vq
)
1299 if (!(vq
->used_flags
& VRING_USED_F_NO_NOTIFY
))
1301 vq
->used_flags
&= ~VRING_USED_F_NO_NOTIFY
;
1302 r
= put_user(vq
->used_flags
, &vq
->used
->flags
);
1304 vq_err(vq
, "Failed to enable notification at %p: %d\n",
1305 &vq
->used
->flags
, r
);
1308 /* They could have slipped one in as we were doing that: make
1309 * sure it's written, then check again. */
1311 r
= get_user(avail_idx
, &vq
->avail
->idx
);
1313 vq_err(vq
, "Failed to check avail idx at %p: %d\n",
1314 &vq
->avail
->idx
, r
);
1318 return avail_idx
!= vq
->avail_idx
;
1321 /* We don't need to be notified again. */
1322 void vhost_disable_notify(struct vhost_virtqueue
*vq
)
1325 if (vq
->used_flags
& VRING_USED_F_NO_NOTIFY
)
1327 vq
->used_flags
|= VRING_USED_F_NO_NOTIFY
;
1328 r
= put_user(vq
->used_flags
, &vq
->used
->flags
);
1330 vq_err(vq
, "Failed to enable notification at %p: %d\n",
1331 &vq
->used
->flags
, r
);