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 /* Helper to allocate iovec buffers for all vqs. */
216 static long vhost_dev_alloc_iovecs(struct vhost_dev
*dev
)
219 for (i
= 0; i
< dev
->nvqs
; ++i
) {
220 dev
->vqs
[i
].indirect
= kmalloc(sizeof *dev
->vqs
[i
].indirect
*
221 UIO_MAXIOV
, GFP_KERNEL
);
222 dev
->vqs
[i
].log
= kmalloc(sizeof *dev
->vqs
[i
].log
* UIO_MAXIOV
,
224 dev
->vqs
[i
].heads
= kmalloc(sizeof *dev
->vqs
[i
].heads
*
225 UIO_MAXIOV
, GFP_KERNEL
);
227 if (!dev
->vqs
[i
].indirect
|| !dev
->vqs
[i
].log
||
233 for (; i
>= 0; --i
) {
234 kfree(dev
->vqs
[i
].indirect
);
235 kfree(dev
->vqs
[i
].log
);
236 kfree(dev
->vqs
[i
].heads
);
241 static void vhost_dev_free_iovecs(struct vhost_dev
*dev
)
244 for (i
= 0; i
< dev
->nvqs
; ++i
) {
245 kfree(dev
->vqs
[i
].indirect
);
246 dev
->vqs
[i
].indirect
= NULL
;
247 kfree(dev
->vqs
[i
].log
);
248 dev
->vqs
[i
].log
= NULL
;
249 kfree(dev
->vqs
[i
].heads
);
250 dev
->vqs
[i
].heads
= NULL
;
254 long vhost_dev_init(struct vhost_dev
*dev
,
255 struct vhost_virtqueue
*vqs
, int nvqs
)
261 mutex_init(&dev
->mutex
);
263 dev
->log_file
= NULL
;
266 spin_lock_init(&dev
->work_lock
);
267 INIT_LIST_HEAD(&dev
->work_list
);
270 for (i
= 0; i
< dev
->nvqs
; ++i
) {
271 dev
->vqs
[i
].log
= NULL
;
272 dev
->vqs
[i
].indirect
= NULL
;
273 dev
->vqs
[i
].heads
= NULL
;
274 dev
->vqs
[i
].dev
= dev
;
275 mutex_init(&dev
->vqs
[i
].mutex
);
276 vhost_vq_reset(dev
, dev
->vqs
+ i
);
277 if (dev
->vqs
[i
].handle_kick
)
278 vhost_poll_init(&dev
->vqs
[i
].poll
,
279 dev
->vqs
[i
].handle_kick
, POLLIN
, dev
);
285 /* Caller should have device mutex */
286 long vhost_dev_check_owner(struct vhost_dev
*dev
)
288 /* Are you the owner? If not, I don't think you mean to do that */
289 return dev
->mm
== current
->mm
? 0 : -EPERM
;
292 struct vhost_attach_cgroups_struct
{
293 struct vhost_work work
;
294 struct task_struct
*owner
;
298 static void vhost_attach_cgroups_work(struct vhost_work
*work
)
300 struct vhost_attach_cgroups_struct
*s
;
301 s
= container_of(work
, struct vhost_attach_cgroups_struct
, work
);
302 s
->ret
= cgroup_attach_task_all(s
->owner
, current
);
305 static int vhost_attach_cgroups(struct vhost_dev
*dev
)
307 struct vhost_attach_cgroups_struct attach
;
308 attach
.owner
= current
;
309 vhost_work_init(&attach
.work
, vhost_attach_cgroups_work
);
310 vhost_work_queue(dev
, &attach
.work
);
311 vhost_work_flush(dev
, &attach
.work
);
315 /* Caller should have device mutex */
316 static long vhost_dev_set_owner(struct vhost_dev
*dev
)
318 struct task_struct
*worker
;
320 /* Is there an owner already? */
325 /* No owner, become one */
326 dev
->mm
= get_task_mm(current
);
327 worker
= kthread_create(vhost_worker
, dev
, "vhost-%d", current
->pid
);
328 if (IS_ERR(worker
)) {
329 err
= PTR_ERR(worker
);
333 dev
->worker
= worker
;
334 wake_up_process(worker
); /* avoid contributing to loadavg */
336 err
= vhost_attach_cgroups(dev
);
340 err
= vhost_dev_alloc_iovecs(dev
);
346 kthread_stop(worker
);
356 /* Caller should have device mutex */
357 long vhost_dev_reset_owner(struct vhost_dev
*dev
)
359 struct vhost_memory
*memory
;
361 /* Restore memory to default empty mapping. */
362 memory
= kmalloc(offsetof(struct vhost_memory
, regions
), GFP_KERNEL
);
366 vhost_dev_cleanup(dev
);
368 memory
->nregions
= 0;
369 RCU_INIT_POINTER(dev
->memory
, memory
);
373 /* Caller should have device mutex */
374 void vhost_dev_cleanup(struct vhost_dev
*dev
)
377 for (i
= 0; i
< dev
->nvqs
; ++i
) {
378 if (dev
->vqs
[i
].kick
&& dev
->vqs
[i
].handle_kick
) {
379 vhost_poll_stop(&dev
->vqs
[i
].poll
);
380 vhost_poll_flush(&dev
->vqs
[i
].poll
);
382 if (dev
->vqs
[i
].error_ctx
)
383 eventfd_ctx_put(dev
->vqs
[i
].error_ctx
);
384 if (dev
->vqs
[i
].error
)
385 fput(dev
->vqs
[i
].error
);
386 if (dev
->vqs
[i
].kick
)
387 fput(dev
->vqs
[i
].kick
);
388 if (dev
->vqs
[i
].call_ctx
)
389 eventfd_ctx_put(dev
->vqs
[i
].call_ctx
);
390 if (dev
->vqs
[i
].call
)
391 fput(dev
->vqs
[i
].call
);
392 vhost_vq_reset(dev
, dev
->vqs
+ i
);
394 vhost_dev_free_iovecs(dev
);
396 eventfd_ctx_put(dev
->log_ctx
);
400 dev
->log_file
= NULL
;
401 /* No one will access memory at this point */
402 kfree(rcu_dereference_protected(dev
->memory
,
403 lockdep_is_held(&dev
->mutex
)));
404 RCU_INIT_POINTER(dev
->memory
, NULL
);
409 WARN_ON(!list_empty(&dev
->work_list
));
411 kthread_stop(dev
->worker
);
416 static int log_access_ok(void __user
*log_base
, u64 addr
, unsigned long sz
)
418 u64 a
= addr
/ VHOST_PAGE_SIZE
/ 8;
419 /* Make sure 64 bit math will not overflow. */
420 if (a
> ULONG_MAX
- (unsigned long)log_base
||
421 a
+ (unsigned long)log_base
> ULONG_MAX
)
424 return access_ok(VERIFY_WRITE
, log_base
+ a
,
425 (sz
+ VHOST_PAGE_SIZE
* 8 - 1) / VHOST_PAGE_SIZE
/ 8);
428 /* Caller should have vq mutex and device mutex. */
429 static int vq_memory_access_ok(void __user
*log_base
, struct vhost_memory
*mem
,
437 for (i
= 0; i
< mem
->nregions
; ++i
) {
438 struct vhost_memory_region
*m
= mem
->regions
+ i
;
439 unsigned long a
= m
->userspace_addr
;
440 if (m
->memory_size
> ULONG_MAX
)
442 else if (!access_ok(VERIFY_WRITE
, (void __user
*)a
,
445 else if (log_all
&& !log_access_ok(log_base
,
453 /* Can we switch to this memory table? */
454 /* Caller should have device mutex but not vq mutex */
455 static int memory_access_ok(struct vhost_dev
*d
, struct vhost_memory
*mem
,
459 for (i
= 0; i
< d
->nvqs
; ++i
) {
461 mutex_lock(&d
->vqs
[i
].mutex
);
462 /* If ring is inactive, will check when it's enabled. */
463 if (d
->vqs
[i
].private_data
)
464 ok
= vq_memory_access_ok(d
->vqs
[i
].log_base
, mem
,
468 mutex_unlock(&d
->vqs
[i
].mutex
);
475 static int vq_access_ok(unsigned int num
,
476 struct vring_desc __user
*desc
,
477 struct vring_avail __user
*avail
,
478 struct vring_used __user
*used
)
480 return access_ok(VERIFY_READ
, desc
, num
* sizeof *desc
) &&
481 access_ok(VERIFY_READ
, avail
,
482 sizeof *avail
+ num
* sizeof *avail
->ring
) &&
483 access_ok(VERIFY_WRITE
, used
,
484 sizeof *used
+ num
* sizeof *used
->ring
);
487 /* Can we log writes? */
488 /* Caller should have device mutex but not vq mutex */
489 int vhost_log_access_ok(struct vhost_dev
*dev
)
491 struct vhost_memory
*mp
;
493 mp
= rcu_dereference_protected(dev
->memory
,
494 lockdep_is_held(&dev
->mutex
));
495 return memory_access_ok(dev
, mp
, 1);
498 /* Verify access for write logging. */
499 /* Caller should have vq mutex and device mutex */
500 static int vq_log_access_ok(struct vhost_virtqueue
*vq
, void __user
*log_base
)
502 struct vhost_memory
*mp
;
504 mp
= rcu_dereference_protected(vq
->dev
->memory
,
505 lockdep_is_held(&vq
->mutex
));
506 return vq_memory_access_ok(log_base
, mp
,
507 vhost_has_feature(vq
->dev
, VHOST_F_LOG_ALL
)) &&
508 (!vq
->log_used
|| log_access_ok(log_base
, vq
->log_addr
,
510 vq
->num
* sizeof *vq
->used
->ring
));
513 /* Can we start vq? */
514 /* Caller should have vq mutex and device mutex */
515 int vhost_vq_access_ok(struct vhost_virtqueue
*vq
)
517 return vq_access_ok(vq
->num
, vq
->desc
, vq
->avail
, vq
->used
) &&
518 vq_log_access_ok(vq
, vq
->log_base
);
521 static long vhost_set_memory(struct vhost_dev
*d
, struct vhost_memory __user
*m
)
523 struct vhost_memory mem
, *newmem
, *oldmem
;
524 unsigned long size
= offsetof(struct vhost_memory
, regions
);
525 if (copy_from_user(&mem
, m
, size
))
529 if (mem
.nregions
> VHOST_MEMORY_MAX_NREGIONS
)
531 newmem
= kmalloc(size
+ mem
.nregions
* sizeof *m
->regions
, GFP_KERNEL
);
535 memcpy(newmem
, &mem
, size
);
536 if (copy_from_user(newmem
->regions
, m
->regions
,
537 mem
.nregions
* sizeof *m
->regions
)) {
542 if (!memory_access_ok(d
, newmem
, vhost_has_feature(d
, VHOST_F_LOG_ALL
))) {
546 oldmem
= rcu_dereference_protected(d
->memory
,
547 lockdep_is_held(&d
->mutex
));
548 rcu_assign_pointer(d
->memory
, newmem
);
554 static int init_used(struct vhost_virtqueue
*vq
,
555 struct vring_used __user
*used
)
557 int r
= put_user(vq
->used_flags
, &used
->flags
);
560 return get_user(vq
->last_used_idx
, &used
->idx
);
563 static long vhost_set_vring(struct vhost_dev
*d
, int ioctl
, void __user
*argp
)
565 struct file
*eventfp
, *filep
= NULL
,
566 *pollstart
= NULL
, *pollstop
= NULL
;
567 struct eventfd_ctx
*ctx
= NULL
;
568 u32 __user
*idxp
= argp
;
569 struct vhost_virtqueue
*vq
;
570 struct vhost_vring_state s
;
571 struct vhost_vring_file f
;
572 struct vhost_vring_addr a
;
576 r
= get_user(idx
, idxp
);
584 mutex_lock(&vq
->mutex
);
587 case VHOST_SET_VRING_NUM
:
588 /* Resizing ring with an active backend?
589 * You don't want to do that. */
590 if (vq
->private_data
) {
594 if (copy_from_user(&s
, argp
, sizeof s
)) {
598 if (!s
.num
|| s
.num
> 0xffff || (s
.num
& (s
.num
- 1))) {
604 case VHOST_SET_VRING_BASE
:
605 /* Moving base with an active backend?
606 * You don't want to do that. */
607 if (vq
->private_data
) {
611 if (copy_from_user(&s
, argp
, sizeof s
)) {
615 if (s
.num
> 0xffff) {
619 vq
->last_avail_idx
= s
.num
;
620 /* Forget the cached index value. */
621 vq
->avail_idx
= vq
->last_avail_idx
;
623 case VHOST_GET_VRING_BASE
:
625 s
.num
= vq
->last_avail_idx
;
626 if (copy_to_user(argp
, &s
, sizeof s
))
629 case VHOST_SET_VRING_ADDR
:
630 if (copy_from_user(&a
, argp
, sizeof a
)) {
634 if (a
.flags
& ~(0x1 << VHOST_VRING_F_LOG
)) {
638 /* For 32bit, verify that the top 32bits of the user
639 data are set to zero. */
640 if ((u64
)(unsigned long)a
.desc_user_addr
!= a
.desc_user_addr
||
641 (u64
)(unsigned long)a
.used_user_addr
!= a
.used_user_addr
||
642 (u64
)(unsigned long)a
.avail_user_addr
!= a
.avail_user_addr
) {
646 if ((a
.avail_user_addr
& (sizeof *vq
->avail
->ring
- 1)) ||
647 (a
.used_user_addr
& (sizeof *vq
->used
->ring
- 1)) ||
648 (a
.log_guest_addr
& (sizeof *vq
->used
->ring
- 1))) {
653 /* We only verify access here if backend is configured.
654 * If it is not, we don't as size might not have been setup.
655 * We will verify when backend is configured. */
656 if (vq
->private_data
) {
657 if (!vq_access_ok(vq
->num
,
658 (void __user
*)(unsigned long)a
.desc_user_addr
,
659 (void __user
*)(unsigned long)a
.avail_user_addr
,
660 (void __user
*)(unsigned long)a
.used_user_addr
)) {
665 /* Also validate log access for used ring if enabled. */
666 if ((a
.flags
& (0x1 << VHOST_VRING_F_LOG
)) &&
667 !log_access_ok(vq
->log_base
, a
.log_guest_addr
,
669 vq
->num
* sizeof *vq
->used
->ring
)) {
675 r
= init_used(vq
, (struct vring_used __user
*)(unsigned long)
679 vq
->log_used
= !!(a
.flags
& (0x1 << VHOST_VRING_F_LOG
));
680 vq
->desc
= (void __user
*)(unsigned long)a
.desc_user_addr
;
681 vq
->avail
= (void __user
*)(unsigned long)a
.avail_user_addr
;
682 vq
->log_addr
= a
.log_guest_addr
;
683 vq
->used
= (void __user
*)(unsigned long)a
.used_user_addr
;
685 case VHOST_SET_VRING_KICK
:
686 if (copy_from_user(&f
, argp
, sizeof f
)) {
690 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
691 if (IS_ERR(eventfp
)) {
692 r
= PTR_ERR(eventfp
);
695 if (eventfp
!= vq
->kick
) {
696 pollstop
= filep
= vq
->kick
;
697 pollstart
= vq
->kick
= eventfp
;
701 case VHOST_SET_VRING_CALL
:
702 if (copy_from_user(&f
, argp
, sizeof f
)) {
706 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
707 if (IS_ERR(eventfp
)) {
708 r
= PTR_ERR(eventfp
);
711 if (eventfp
!= vq
->call
) {
715 vq
->call_ctx
= eventfp
?
716 eventfd_ctx_fileget(eventfp
) : NULL
;
720 case VHOST_SET_VRING_ERR
:
721 if (copy_from_user(&f
, argp
, sizeof f
)) {
725 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
726 if (IS_ERR(eventfp
)) {
727 r
= PTR_ERR(eventfp
);
730 if (eventfp
!= vq
->error
) {
734 vq
->error_ctx
= eventfp
?
735 eventfd_ctx_fileget(eventfp
) : NULL
;
743 if (pollstop
&& vq
->handle_kick
)
744 vhost_poll_stop(&vq
->poll
);
747 eventfd_ctx_put(ctx
);
751 if (pollstart
&& vq
->handle_kick
)
752 vhost_poll_start(&vq
->poll
, vq
->kick
);
754 mutex_unlock(&vq
->mutex
);
756 if (pollstop
&& vq
->handle_kick
)
757 vhost_poll_flush(&vq
->poll
);
761 /* Caller must have device mutex */
762 long vhost_dev_ioctl(struct vhost_dev
*d
, unsigned int ioctl
, unsigned long arg
)
764 void __user
*argp
= (void __user
*)arg
;
765 struct file
*eventfp
, *filep
= NULL
;
766 struct eventfd_ctx
*ctx
= NULL
;
771 /* If you are not the owner, you can become one */
772 if (ioctl
== VHOST_SET_OWNER
) {
773 r
= vhost_dev_set_owner(d
);
777 /* You must be the owner to do anything else */
778 r
= vhost_dev_check_owner(d
);
783 case VHOST_SET_MEM_TABLE
:
784 r
= vhost_set_memory(d
, argp
);
786 case VHOST_SET_LOG_BASE
:
787 if (copy_from_user(&p
, argp
, sizeof p
)) {
791 if ((u64
)(unsigned long)p
!= p
) {
795 for (i
= 0; i
< d
->nvqs
; ++i
) {
796 struct vhost_virtqueue
*vq
;
797 void __user
*base
= (void __user
*)(unsigned long)p
;
799 mutex_lock(&vq
->mutex
);
800 /* If ring is inactive, will check when it's enabled. */
801 if (vq
->private_data
&& !vq_log_access_ok(vq
, base
))
805 mutex_unlock(&vq
->mutex
);
808 case VHOST_SET_LOG_FD
:
809 r
= get_user(fd
, (int __user
*)argp
);
812 eventfp
= fd
== -1 ? NULL
: eventfd_fget(fd
);
813 if (IS_ERR(eventfp
)) {
814 r
= PTR_ERR(eventfp
);
817 if (eventfp
!= d
->log_file
) {
820 d
->log_ctx
= eventfp
?
821 eventfd_ctx_fileget(eventfp
) : NULL
;
824 for (i
= 0; i
< d
->nvqs
; ++i
) {
825 mutex_lock(&d
->vqs
[i
].mutex
);
826 d
->vqs
[i
].log_ctx
= d
->log_ctx
;
827 mutex_unlock(&d
->vqs
[i
].mutex
);
830 eventfd_ctx_put(ctx
);
835 r
= vhost_set_vring(d
, ioctl
, argp
);
842 static const struct vhost_memory_region
*find_region(struct vhost_memory
*mem
,
843 __u64 addr
, __u32 len
)
845 struct vhost_memory_region
*reg
;
847 /* linear search is not brilliant, but we really have on the order of 6
848 * regions in practice */
849 for (i
= 0; i
< mem
->nregions
; ++i
) {
850 reg
= mem
->regions
+ i
;
851 if (reg
->guest_phys_addr
<= addr
&&
852 reg
->guest_phys_addr
+ reg
->memory_size
- 1 >= addr
)
858 /* TODO: This is really inefficient. We need something like get_user()
859 * (instruction directly accesses the data, with an exception table entry
860 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
862 static int set_bit_to_user(int nr
, void __user
*addr
)
864 unsigned long log
= (unsigned long)addr
;
867 int bit
= nr
+ (log
% PAGE_SIZE
) * 8;
869 r
= get_user_pages_fast(log
, 1, 1, &page
);
873 base
= kmap_atomic(page
, KM_USER0
);
875 kunmap_atomic(base
, KM_USER0
);
876 set_page_dirty_lock(page
);
881 static int log_write(void __user
*log_base
,
882 u64 write_address
, u64 write_length
)
887 write_length
+= write_address
% VHOST_PAGE_SIZE
;
888 write_address
/= VHOST_PAGE_SIZE
;
890 u64 base
= (u64
)(unsigned long)log_base
;
891 u64 log
= base
+ write_address
/ 8;
892 int bit
= write_address
% 8;
893 if ((u64
)(unsigned long)log
!= log
)
895 r
= set_bit_to_user(bit
, (void __user
*)(unsigned long)log
);
898 if (write_length
<= VHOST_PAGE_SIZE
)
900 write_length
-= VHOST_PAGE_SIZE
;
906 int vhost_log_write(struct vhost_virtqueue
*vq
, struct vhost_log
*log
,
907 unsigned int log_num
, u64 len
)
911 /* Make sure data written is seen before log. */
913 for (i
= 0; i
< log_num
; ++i
) {
914 u64 l
= min(log
[i
].len
, len
);
915 r
= log_write(vq
->log_base
, log
[i
].addr
, l
);
921 eventfd_signal(vq
->log_ctx
, 1);
925 /* Length written exceeds what we have stored. This is a bug. */
930 static int translate_desc(struct vhost_dev
*dev
, u64 addr
, u32 len
,
931 struct iovec iov
[], int iov_size
)
933 const struct vhost_memory_region
*reg
;
934 struct vhost_memory
*mem
;
941 mem
= rcu_dereference(dev
->memory
);
942 while ((u64
)len
> s
) {
944 if (unlikely(ret
>= iov_size
)) {
948 reg
= find_region(mem
, addr
, len
);
949 if (unlikely(!reg
)) {
954 size
= reg
->memory_size
- addr
+ reg
->guest_phys_addr
;
955 _iov
->iov_len
= min((u64
)len
, size
);
956 _iov
->iov_base
= (void __user
*)(unsigned long)
957 (reg
->userspace_addr
+ addr
- reg
->guest_phys_addr
);
967 /* Each buffer in the virtqueues is actually a chain of descriptors. This
968 * function returns the next descriptor in the chain,
969 * or -1U if we're at the end. */
970 static unsigned next_desc(struct vring_desc
*desc
)
974 /* If this descriptor says it doesn't chain, we're done. */
975 if (!(desc
->flags
& VRING_DESC_F_NEXT
))
978 /* Check they're not leading us off end of descriptors. */
980 /* Make sure compiler knows to grab that: we don't want it changing! */
981 /* We will use the result as an index in an array, so most
982 * architectures only need a compiler barrier here. */
983 read_barrier_depends();
988 static int get_indirect(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
,
989 struct iovec iov
[], unsigned int iov_size
,
990 unsigned int *out_num
, unsigned int *in_num
,
991 struct vhost_log
*log
, unsigned int *log_num
,
992 struct vring_desc
*indirect
)
994 struct vring_desc desc
;
995 unsigned int i
= 0, count
, found
= 0;
999 if (unlikely(indirect
->len
% sizeof desc
)) {
1000 vq_err(vq
, "Invalid length in indirect descriptor: "
1001 "len 0x%llx not multiple of 0x%zx\n",
1002 (unsigned long long)indirect
->len
,
1007 ret
= translate_desc(dev
, indirect
->addr
, indirect
->len
, vq
->indirect
,
1009 if (unlikely(ret
< 0)) {
1010 vq_err(vq
, "Translation failure %d in indirect.\n", ret
);
1014 /* We will use the result as an address to read from, so most
1015 * architectures only need a compiler barrier here. */
1016 read_barrier_depends();
1018 count
= indirect
->len
/ sizeof desc
;
1019 /* Buffers are chained via a 16 bit next field, so
1020 * we can have at most 2^16 of these. */
1021 if (unlikely(count
> USHRT_MAX
+ 1)) {
1022 vq_err(vq
, "Indirect buffer length too big: %d\n",
1028 unsigned iov_count
= *in_num
+ *out_num
;
1029 if (unlikely(++found
> count
)) {
1030 vq_err(vq
, "Loop detected: last one at %u "
1031 "indirect size %u\n",
1035 if (unlikely(memcpy_fromiovec((unsigned char *)&desc
, vq
->indirect
,
1037 vq_err(vq
, "Failed indirect descriptor: idx %d, %zx\n",
1038 i
, (size_t)indirect
->addr
+ i
* sizeof desc
);
1041 if (unlikely(desc
.flags
& VRING_DESC_F_INDIRECT
)) {
1042 vq_err(vq
, "Nested indirect descriptor: idx %d, %zx\n",
1043 i
, (size_t)indirect
->addr
+ i
* sizeof desc
);
1047 ret
= translate_desc(dev
, desc
.addr
, desc
.len
, iov
+ iov_count
,
1048 iov_size
- iov_count
);
1049 if (unlikely(ret
< 0)) {
1050 vq_err(vq
, "Translation failure %d indirect idx %d\n",
1054 /* If this is an input descriptor, increment that count. */
1055 if (desc
.flags
& VRING_DESC_F_WRITE
) {
1057 if (unlikely(log
)) {
1058 log
[*log_num
].addr
= desc
.addr
;
1059 log
[*log_num
].len
= desc
.len
;
1063 /* If it's an output descriptor, they're all supposed
1064 * to come before any input descriptors. */
1065 if (unlikely(*in_num
)) {
1066 vq_err(vq
, "Indirect descriptor "
1067 "has out after in: idx %d\n", i
);
1072 } while ((i
= next_desc(&desc
)) != -1);
1076 /* This looks in the virtqueue and for the first available buffer, and converts
1077 * it to an iovec for convenient access. Since descriptors consist of some
1078 * number of output then some number of input descriptors, it's actually two
1079 * iovecs, but we pack them into one and note how many of each there were.
1081 * This function returns the descriptor number found, or vq->num (which is
1082 * never a valid descriptor number) if none was found. A negative code is
1083 * returned on error. */
1084 int vhost_get_vq_desc(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
,
1085 struct iovec iov
[], unsigned int iov_size
,
1086 unsigned int *out_num
, unsigned int *in_num
,
1087 struct vhost_log
*log
, unsigned int *log_num
)
1089 struct vring_desc desc
;
1090 unsigned int i
, head
, found
= 0;
1094 /* Check it isn't doing very strange things with descriptor numbers. */
1095 last_avail_idx
= vq
->last_avail_idx
;
1096 if (unlikely(get_user(vq
->avail_idx
, &vq
->avail
->idx
))) {
1097 vq_err(vq
, "Failed to access avail idx at %p\n",
1102 if (unlikely((u16
)(vq
->avail_idx
- last_avail_idx
) > vq
->num
)) {
1103 vq_err(vq
, "Guest moved used index from %u to %u",
1104 last_avail_idx
, vq
->avail_idx
);
1108 /* If there's nothing new since last we looked, return invalid. */
1109 if (vq
->avail_idx
== last_avail_idx
)
1112 /* Only get avail ring entries after they have been exposed by guest. */
1115 /* Grab the next descriptor number they're advertising, and increment
1116 * the index we've seen. */
1117 if (unlikely(get_user(head
,
1118 &vq
->avail
->ring
[last_avail_idx
% vq
->num
]))) {
1119 vq_err(vq
, "Failed to read head: idx %d address %p\n",
1121 &vq
->avail
->ring
[last_avail_idx
% vq
->num
]);
1125 /* If their number is silly, that's an error. */
1126 if (unlikely(head
>= vq
->num
)) {
1127 vq_err(vq
, "Guest says index %u > %u is available",
1132 /* When we start there are none of either input nor output. */
1133 *out_num
= *in_num
= 0;
1139 unsigned iov_count
= *in_num
+ *out_num
;
1140 if (unlikely(i
>= vq
->num
)) {
1141 vq_err(vq
, "Desc index is %u > %u, head = %u",
1145 if (unlikely(++found
> vq
->num
)) {
1146 vq_err(vq
, "Loop detected: last one at %u "
1147 "vq size %u head %u\n",
1151 ret
= copy_from_user(&desc
, vq
->desc
+ i
, sizeof desc
);
1152 if (unlikely(ret
)) {
1153 vq_err(vq
, "Failed to get descriptor: idx %d addr %p\n",
1157 if (desc
.flags
& VRING_DESC_F_INDIRECT
) {
1158 ret
= get_indirect(dev
, vq
, iov
, iov_size
,
1160 log
, log_num
, &desc
);
1161 if (unlikely(ret
< 0)) {
1162 vq_err(vq
, "Failure detected "
1163 "in indirect descriptor at idx %d\n", i
);
1169 ret
= translate_desc(dev
, desc
.addr
, desc
.len
, iov
+ iov_count
,
1170 iov_size
- iov_count
);
1171 if (unlikely(ret
< 0)) {
1172 vq_err(vq
, "Translation failure %d descriptor idx %d\n",
1176 if (desc
.flags
& VRING_DESC_F_WRITE
) {
1177 /* If this is an input descriptor,
1178 * increment that count. */
1180 if (unlikely(log
)) {
1181 log
[*log_num
].addr
= desc
.addr
;
1182 log
[*log_num
].len
= desc
.len
;
1186 /* If it's an output descriptor, they're all supposed
1187 * to come before any input descriptors. */
1188 if (unlikely(*in_num
)) {
1189 vq_err(vq
, "Descriptor has out after in: "
1195 } while ((i
= next_desc(&desc
)) != -1);
1197 /* On success, increment avail index. */
1198 vq
->last_avail_idx
++;
1202 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1203 void vhost_discard_vq_desc(struct vhost_virtqueue
*vq
, int n
)
1205 vq
->last_avail_idx
-= n
;
1208 /* After we've used one of their buffers, we tell them about it. We'll then
1209 * want to notify the guest, using eventfd. */
1210 int vhost_add_used(struct vhost_virtqueue
*vq
, unsigned int head
, int len
)
1212 struct vring_used_elem __user
*used
;
1214 /* The virtqueue contains a ring of used buffers. Get a pointer to the
1215 * next entry in that used ring. */
1216 used
= &vq
->used
->ring
[vq
->last_used_idx
% vq
->num
];
1217 if (put_user(head
, &used
->id
)) {
1218 vq_err(vq
, "Failed to write used id");
1221 if (put_user(len
, &used
->len
)) {
1222 vq_err(vq
, "Failed to write used len");
1225 /* Make sure buffer is written before we update index. */
1227 if (put_user(vq
->last_used_idx
+ 1, &vq
->used
->idx
)) {
1228 vq_err(vq
, "Failed to increment used idx");
1231 if (unlikely(vq
->log_used
)) {
1232 /* Make sure data is seen before log. */
1234 /* Log used ring entry write. */
1235 log_write(vq
->log_base
,
1237 ((void __user
*)used
- (void __user
*)vq
->used
),
1239 /* Log used index update. */
1240 log_write(vq
->log_base
,
1241 vq
->log_addr
+ offsetof(struct vring_used
, idx
),
1242 sizeof vq
->used
->idx
);
1244 eventfd_signal(vq
->log_ctx
, 1);
1246 vq
->last_used_idx
++;
1250 static int __vhost_add_used_n(struct vhost_virtqueue
*vq
,
1251 struct vring_used_elem
*heads
,
1254 struct vring_used_elem __user
*used
;
1257 start
= vq
->last_used_idx
% vq
->num
;
1258 used
= vq
->used
->ring
+ start
;
1259 if (copy_to_user(used
, heads
, count
* sizeof *used
)) {
1260 vq_err(vq
, "Failed to write used");
1263 if (unlikely(vq
->log_used
)) {
1264 /* Make sure data is seen before log. */
1266 /* Log used ring entry write. */
1267 log_write(vq
->log_base
,
1269 ((void __user
*)used
- (void __user
*)vq
->used
),
1270 count
* sizeof *used
);
1272 vq
->last_used_idx
+= count
;
1276 /* After we've used one of their buffers, we tell them about it. We'll then
1277 * want to notify the guest, using eventfd. */
1278 int vhost_add_used_n(struct vhost_virtqueue
*vq
, struct vring_used_elem
*heads
,
1283 start
= vq
->last_used_idx
% vq
->num
;
1284 n
= vq
->num
- start
;
1286 r
= __vhost_add_used_n(vq
, heads
, n
);
1292 r
= __vhost_add_used_n(vq
, heads
, count
);
1294 /* Make sure buffer is written before we update index. */
1296 if (put_user(vq
->last_used_idx
, &vq
->used
->idx
)) {
1297 vq_err(vq
, "Failed to increment used idx");
1300 if (unlikely(vq
->log_used
)) {
1301 /* Log used index update. */
1302 log_write(vq
->log_base
,
1303 vq
->log_addr
+ offsetof(struct vring_used
, idx
),
1304 sizeof vq
->used
->idx
);
1306 eventfd_signal(vq
->log_ctx
, 1);
1311 /* This actually signals the guest, using eventfd. */
1312 void vhost_signal(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
)
1315 /* Flush out used index updates. This is paired
1316 * with the barrier that the Guest executes when enabling
1320 if (get_user(flags
, &vq
->avail
->flags
)) {
1321 vq_err(vq
, "Failed to get flags");
1325 /* If they don't want an interrupt, don't signal, unless empty. */
1326 if ((flags
& VRING_AVAIL_F_NO_INTERRUPT
) &&
1327 (vq
->avail_idx
!= vq
->last_avail_idx
||
1328 !vhost_has_feature(dev
, VIRTIO_F_NOTIFY_ON_EMPTY
)))
1331 /* Signal the Guest tell them we used something up. */
1333 eventfd_signal(vq
->call_ctx
, 1);
1336 /* And here's the combo meal deal. Supersize me! */
1337 void vhost_add_used_and_signal(struct vhost_dev
*dev
,
1338 struct vhost_virtqueue
*vq
,
1339 unsigned int head
, int len
)
1341 vhost_add_used(vq
, head
, len
);
1342 vhost_signal(dev
, vq
);
1345 /* multi-buffer version of vhost_add_used_and_signal */
1346 void vhost_add_used_and_signal_n(struct vhost_dev
*dev
,
1347 struct vhost_virtqueue
*vq
,
1348 struct vring_used_elem
*heads
, unsigned count
)
1350 vhost_add_used_n(vq
, heads
, count
);
1351 vhost_signal(dev
, vq
);
1354 /* OK, now we need to know about added descriptors. */
1355 bool vhost_enable_notify(struct vhost_virtqueue
*vq
)
1359 if (!(vq
->used_flags
& VRING_USED_F_NO_NOTIFY
))
1361 vq
->used_flags
&= ~VRING_USED_F_NO_NOTIFY
;
1362 r
= put_user(vq
->used_flags
, &vq
->used
->flags
);
1364 vq_err(vq
, "Failed to enable notification at %p: %d\n",
1365 &vq
->used
->flags
, r
);
1368 /* They could have slipped one in as we were doing that: make
1369 * sure it's written, then check again. */
1371 r
= get_user(avail_idx
, &vq
->avail
->idx
);
1373 vq_err(vq
, "Failed to check avail idx at %p: %d\n",
1374 &vq
->avail
->idx
, r
);
1378 return avail_idx
!= vq
->avail_idx
;
1381 /* We don't need to be notified again. */
1382 void vhost_disable_notify(struct vhost_virtqueue
*vq
)
1385 if (vq
->used_flags
& VRING_USED_F_NO_NOTIFY
)
1387 vq
->used_flags
|= VRING_USED_F_NO_NOTIFY
;
1388 r
= put_user(vq
->used_flags
, &vq
->used
->flags
);
1390 vq_err(vq
, "Failed to enable notification at %p: %d\n",
1391 &vq
->used
->flags
, r
);