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
);
310 /* Caller should have device mutex */
311 long vhost_dev_reset_owner(struct vhost_dev
*dev
)
313 struct vhost_memory
*memory
;
315 /* Restore memory to default empty mapping. */
316 memory
= kmalloc(offsetof(struct vhost_memory
, regions
), GFP_KERNEL
);
320 vhost_dev_cleanup(dev
);
322 memory
->nregions
= 0;
323 RCU_INIT_POINTER(dev
->memory
, memory
);
327 /* Caller should have device mutex */
328 void vhost_dev_cleanup(struct vhost_dev
*dev
)
331 for (i
= 0; i
< dev
->nvqs
; ++i
) {
332 if (dev
->vqs
[i
].kick
&& dev
->vqs
[i
].handle_kick
) {
333 vhost_poll_stop(&dev
->vqs
[i
].poll
);
334 vhost_poll_flush(&dev
->vqs
[i
].poll
);
336 if (dev
->vqs
[i
].error_ctx
)
337 eventfd_ctx_put(dev
->vqs
[i
].error_ctx
);
338 if (dev
->vqs
[i
].error
)
339 fput(dev
->vqs
[i
].error
);
340 if (dev
->vqs
[i
].kick
)
341 fput(dev
->vqs
[i
].kick
);
342 if (dev
->vqs
[i
].call_ctx
)
343 eventfd_ctx_put(dev
->vqs
[i
].call_ctx
);
344 if (dev
->vqs
[i
].call
)
345 fput(dev
->vqs
[i
].call
);
346 vhost_vq_reset(dev
, dev
->vqs
+ i
);
349 eventfd_ctx_put(dev
->log_ctx
);
353 dev
->log_file
= NULL
;
354 /* No one will access memory at this point */
355 kfree(rcu_dereference_protected(dev
->memory
,
356 lockdep_is_held(&dev
->mutex
)));
357 RCU_INIT_POINTER(dev
->memory
, NULL
);
362 WARN_ON(!list_empty(&dev
->work_list
));
364 kthread_stop(dev
->worker
);
369 static int log_access_ok(void __user
*log_base
, u64 addr
, unsigned long sz
)
371 u64 a
= addr
/ VHOST_PAGE_SIZE
/ 8;
372 /* Make sure 64 bit math will not overflow. */
373 if (a
> ULONG_MAX
- (unsigned long)log_base
||
374 a
+ (unsigned long)log_base
> ULONG_MAX
)
377 return access_ok(VERIFY_WRITE
, log_base
+ a
,
378 (sz
+ VHOST_PAGE_SIZE
* 8 - 1) / VHOST_PAGE_SIZE
/ 8);
381 /* Caller should have vq mutex and device mutex. */
382 static int vq_memory_access_ok(void __user
*log_base
, struct vhost_memory
*mem
,
390 for (i
= 0; i
< mem
->nregions
; ++i
) {
391 struct vhost_memory_region
*m
= mem
->regions
+ i
;
392 unsigned long a
= m
->userspace_addr
;
393 if (m
->memory_size
> ULONG_MAX
)
395 else if (!access_ok(VERIFY_WRITE
, (void __user
*)a
,
398 else if (log_all
&& !log_access_ok(log_base
,
406 /* Can we switch to this memory table? */
407 /* Caller should have device mutex but not vq mutex */
408 static int memory_access_ok(struct vhost_dev
*d
, struct vhost_memory
*mem
,
412 for (i
= 0; i
< d
->nvqs
; ++i
) {
414 mutex_lock(&d
->vqs
[i
].mutex
);
415 /* If ring is inactive, will check when it's enabled. */
416 if (d
->vqs
[i
].private_data
)
417 ok
= vq_memory_access_ok(d
->vqs
[i
].log_base
, mem
,
421 mutex_unlock(&d
->vqs
[i
].mutex
);
428 static int vq_access_ok(unsigned int num
,
429 struct vring_desc __user
*desc
,
430 struct vring_avail __user
*avail
,
431 struct vring_used __user
*used
)
433 return access_ok(VERIFY_READ
, desc
, num
* sizeof *desc
) &&
434 access_ok(VERIFY_READ
, avail
,
435 sizeof *avail
+ num
* sizeof *avail
->ring
) &&
436 access_ok(VERIFY_WRITE
, used
,
437 sizeof *used
+ num
* sizeof *used
->ring
);
440 /* Can we log writes? */
441 /* Caller should have device mutex but not vq mutex */
442 int vhost_log_access_ok(struct vhost_dev
*dev
)
444 struct vhost_memory
*mp
;
446 mp
= rcu_dereference_protected(dev
->memory
,
447 lockdep_is_held(&dev
->mutex
));
448 return memory_access_ok(dev
, mp
, 1);
451 /* Verify access for write logging. */
452 /* Caller should have vq mutex and device mutex */
453 static int vq_log_access_ok(struct vhost_virtqueue
*vq
, void __user
*log_base
)
455 struct vhost_memory
*mp
;
457 mp
= rcu_dereference_protected(vq
->dev
->memory
,
458 lockdep_is_held(&vq
->mutex
));
459 return vq_memory_access_ok(log_base
, mp
,
460 vhost_has_feature(vq
->dev
, VHOST_F_LOG_ALL
)) &&
461 (!vq
->log_used
|| log_access_ok(log_base
, vq
->log_addr
,
463 vq
->num
* sizeof *vq
->used
->ring
));
466 /* Can we start vq? */
467 /* Caller should have vq mutex and device mutex */
468 int vhost_vq_access_ok(struct vhost_virtqueue
*vq
)
470 return vq_access_ok(vq
->num
, vq
->desc
, vq
->avail
, vq
->used
) &&
471 vq_log_access_ok(vq
, vq
->log_base
);
474 static long vhost_set_memory(struct vhost_dev
*d
, struct vhost_memory __user
*m
)
476 struct vhost_memory mem
, *newmem
, *oldmem
;
477 unsigned long size
= offsetof(struct vhost_memory
, regions
);
478 if (copy_from_user(&mem
, m
, size
))
482 if (mem
.nregions
> VHOST_MEMORY_MAX_NREGIONS
)
484 newmem
= kmalloc(size
+ mem
.nregions
* sizeof *m
->regions
, GFP_KERNEL
);
488 memcpy(newmem
, &mem
, size
);
489 if (copy_from_user(newmem
->regions
, m
->regions
,
490 mem
.nregions
* sizeof *m
->regions
)) {
495 if (!memory_access_ok(d
, newmem
, vhost_has_feature(d
, VHOST_F_LOG_ALL
))) {
499 oldmem
= rcu_dereference_protected(d
->memory
,
500 lockdep_is_held(&d
->mutex
));
501 rcu_assign_pointer(d
->memory
, newmem
);
507 static int init_used(struct vhost_virtqueue
*vq
,
508 struct vring_used __user
*used
)
510 int r
= put_user(vq
->used_flags
, &used
->flags
);
513 return get_user(vq
->last_used_idx
, &used
->idx
);
516 static long vhost_set_vring(struct vhost_dev
*d
, int ioctl
, void __user
*argp
)
518 struct file
*eventfp
, *filep
= NULL
,
519 *pollstart
= NULL
, *pollstop
= NULL
;
520 struct eventfd_ctx
*ctx
= NULL
;
521 u32 __user
*idxp
= argp
;
522 struct vhost_virtqueue
*vq
;
523 struct vhost_vring_state s
;
524 struct vhost_vring_file f
;
525 struct vhost_vring_addr a
;
529 r
= get_user(idx
, idxp
);
537 mutex_lock(&vq
->mutex
);
540 case VHOST_SET_VRING_NUM
:
541 /* Resizing ring with an active backend?
542 * You don't want to do that. */
543 if (vq
->private_data
) {
547 if (copy_from_user(&s
, argp
, sizeof s
)) {
551 if (!s
.num
|| s
.num
> 0xffff || (s
.num
& (s
.num
- 1))) {
557 case VHOST_SET_VRING_BASE
:
558 /* Moving base with an active backend?
559 * You don't want to do that. */
560 if (vq
->private_data
) {
564 if (copy_from_user(&s
, argp
, sizeof s
)) {
568 if (s
.num
> 0xffff) {
572 vq
->last_avail_idx
= s
.num
;
573 /* Forget the cached index value. */
574 vq
->avail_idx
= vq
->last_avail_idx
;
576 case VHOST_GET_VRING_BASE
:
578 s
.num
= vq
->last_avail_idx
;
579 if (copy_to_user(argp
, &s
, sizeof s
))
582 case VHOST_SET_VRING_ADDR
:
583 if (copy_from_user(&a
, argp
, sizeof a
)) {
587 if (a
.flags
& ~(0x1 << VHOST_VRING_F_LOG
)) {
591 /* For 32bit, verify that the top 32bits of the user
592 data are set to zero. */
593 if ((u64
)(unsigned long)a
.desc_user_addr
!= a
.desc_user_addr
||
594 (u64
)(unsigned long)a
.used_user_addr
!= a
.used_user_addr
||
595 (u64
)(unsigned long)a
.avail_user_addr
!= a
.avail_user_addr
) {
599 if ((a
.avail_user_addr
& (sizeof *vq
->avail
->ring
- 1)) ||
600 (a
.used_user_addr
& (sizeof *vq
->used
->ring
- 1)) ||
601 (a
.log_guest_addr
& (sizeof *vq
->used
->ring
- 1))) {
606 /* We only verify access here if backend is configured.
607 * If it is not, we don't as size might not have been setup.
608 * We will verify when backend is configured. */
609 if (vq
->private_data
) {
610 if (!vq_access_ok(vq
->num
,
611 (void __user
*)(unsigned long)a
.desc_user_addr
,
612 (void __user
*)(unsigned long)a
.avail_user_addr
,
613 (void __user
*)(unsigned long)a
.used_user_addr
)) {
618 /* Also validate log access for used ring if enabled. */
619 if ((a
.flags
& (0x1 << VHOST_VRING_F_LOG
)) &&
620 !log_access_ok(vq
->log_base
, a
.log_guest_addr
,
622 vq
->num
* sizeof *vq
->used
->ring
)) {
628 r
= init_used(vq
, (struct vring_used __user
*)(unsigned long)
632 vq
->log_used
= !!(a
.flags
& (0x1 << VHOST_VRING_F_LOG
));
633 vq
->desc
= (void __user
*)(unsigned long)a
.desc_user_addr
;
634 vq
->avail
= (void __user
*)(unsigned long)a
.avail_user_addr
;
635 vq
->log_addr
= a
.log_guest_addr
;
636 vq
->used
= (void __user
*)(unsigned long)a
.used_user_addr
;
638 case VHOST_SET_VRING_KICK
:
639 if (copy_from_user(&f
, argp
, sizeof f
)) {
643 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
644 if (IS_ERR(eventfp
)) {
645 r
= PTR_ERR(eventfp
);
648 if (eventfp
!= vq
->kick
) {
649 pollstop
= filep
= vq
->kick
;
650 pollstart
= vq
->kick
= eventfp
;
654 case VHOST_SET_VRING_CALL
:
655 if (copy_from_user(&f
, argp
, sizeof f
)) {
659 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
660 if (IS_ERR(eventfp
)) {
661 r
= PTR_ERR(eventfp
);
664 if (eventfp
!= vq
->call
) {
668 vq
->call_ctx
= eventfp
?
669 eventfd_ctx_fileget(eventfp
) : NULL
;
673 case VHOST_SET_VRING_ERR
:
674 if (copy_from_user(&f
, argp
, sizeof f
)) {
678 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
679 if (IS_ERR(eventfp
)) {
680 r
= PTR_ERR(eventfp
);
683 if (eventfp
!= vq
->error
) {
687 vq
->error_ctx
= eventfp
?
688 eventfd_ctx_fileget(eventfp
) : NULL
;
696 if (pollstop
&& vq
->handle_kick
)
697 vhost_poll_stop(&vq
->poll
);
700 eventfd_ctx_put(ctx
);
704 if (pollstart
&& vq
->handle_kick
)
705 vhost_poll_start(&vq
->poll
, vq
->kick
);
707 mutex_unlock(&vq
->mutex
);
709 if (pollstop
&& vq
->handle_kick
)
710 vhost_poll_flush(&vq
->poll
);
714 /* Caller must have device mutex */
715 long vhost_dev_ioctl(struct vhost_dev
*d
, unsigned int ioctl
, unsigned long arg
)
717 void __user
*argp
= (void __user
*)arg
;
718 struct file
*eventfp
, *filep
= NULL
;
719 struct eventfd_ctx
*ctx
= NULL
;
724 /* If you are not the owner, you can become one */
725 if (ioctl
== VHOST_SET_OWNER
) {
726 r
= vhost_dev_set_owner(d
);
730 /* You must be the owner to do anything else */
731 r
= vhost_dev_check_owner(d
);
736 case VHOST_SET_MEM_TABLE
:
737 r
= vhost_set_memory(d
, argp
);
739 case VHOST_SET_LOG_BASE
:
740 if (copy_from_user(&p
, argp
, sizeof p
)) {
744 if ((u64
)(unsigned long)p
!= p
) {
748 for (i
= 0; i
< d
->nvqs
; ++i
) {
749 struct vhost_virtqueue
*vq
;
750 void __user
*base
= (void __user
*)(unsigned long)p
;
752 mutex_lock(&vq
->mutex
);
753 /* If ring is inactive, will check when it's enabled. */
754 if (vq
->private_data
&& !vq_log_access_ok(vq
, base
))
758 mutex_unlock(&vq
->mutex
);
761 case VHOST_SET_LOG_FD
:
762 r
= get_user(fd
, (int __user
*)argp
);
765 eventfp
= fd
== -1 ? NULL
: eventfd_fget(fd
);
766 if (IS_ERR(eventfp
)) {
767 r
= PTR_ERR(eventfp
);
770 if (eventfp
!= d
->log_file
) {
773 d
->log_ctx
= eventfp
?
774 eventfd_ctx_fileget(eventfp
) : NULL
;
777 for (i
= 0; i
< d
->nvqs
; ++i
) {
778 mutex_lock(&d
->vqs
[i
].mutex
);
779 d
->vqs
[i
].log_ctx
= d
->log_ctx
;
780 mutex_unlock(&d
->vqs
[i
].mutex
);
783 eventfd_ctx_put(ctx
);
788 r
= vhost_set_vring(d
, ioctl
, argp
);
795 static const struct vhost_memory_region
*find_region(struct vhost_memory
*mem
,
796 __u64 addr
, __u32 len
)
798 struct vhost_memory_region
*reg
;
800 /* linear search is not brilliant, but we really have on the order of 6
801 * regions in practice */
802 for (i
= 0; i
< mem
->nregions
; ++i
) {
803 reg
= mem
->regions
+ i
;
804 if (reg
->guest_phys_addr
<= addr
&&
805 reg
->guest_phys_addr
+ reg
->memory_size
- 1 >= addr
)
811 /* TODO: This is really inefficient. We need something like get_user()
812 * (instruction directly accesses the data, with an exception table entry
813 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
815 static int set_bit_to_user(int nr
, void __user
*addr
)
817 unsigned long log
= (unsigned long)addr
;
820 int bit
= nr
+ (log
% PAGE_SIZE
) * 8;
822 r
= get_user_pages_fast(log
, 1, 1, &page
);
826 base
= kmap_atomic(page
, KM_USER0
);
828 kunmap_atomic(base
, KM_USER0
);
829 set_page_dirty_lock(page
);
834 static int log_write(void __user
*log_base
,
835 u64 write_address
, u64 write_length
)
840 write_address
/= VHOST_PAGE_SIZE
;
842 u64 base
= (u64
)(unsigned long)log_base
;
843 u64 log
= base
+ write_address
/ 8;
844 int bit
= write_address
% 8;
845 if ((u64
)(unsigned long)log
!= log
)
847 r
= set_bit_to_user(bit
, (void __user
*)(unsigned long)log
);
850 if (write_length
<= VHOST_PAGE_SIZE
)
852 write_length
-= VHOST_PAGE_SIZE
;
853 write_address
+= VHOST_PAGE_SIZE
;
858 int vhost_log_write(struct vhost_virtqueue
*vq
, struct vhost_log
*log
,
859 unsigned int log_num
, u64 len
)
863 /* Make sure data written is seen before log. */
865 for (i
= 0; i
< log_num
; ++i
) {
866 u64 l
= min(log
[i
].len
, len
);
867 r
= log_write(vq
->log_base
, log
[i
].addr
, l
);
873 eventfd_signal(vq
->log_ctx
, 1);
877 /* Length written exceeds what we have stored. This is a bug. */
882 static int translate_desc(struct vhost_dev
*dev
, u64 addr
, u32 len
,
883 struct iovec iov
[], int iov_size
)
885 const struct vhost_memory_region
*reg
;
886 struct vhost_memory
*mem
;
893 mem
= rcu_dereference(dev
->memory
);
894 while ((u64
)len
> s
) {
896 if (unlikely(ret
>= iov_size
)) {
900 reg
= find_region(mem
, addr
, len
);
901 if (unlikely(!reg
)) {
906 size
= reg
->memory_size
- addr
+ reg
->guest_phys_addr
;
907 _iov
->iov_len
= min((u64
)len
, size
);
908 _iov
->iov_base
= (void __user
*)(unsigned long)
909 (reg
->userspace_addr
+ addr
- reg
->guest_phys_addr
);
919 /* Each buffer in the virtqueues is actually a chain of descriptors. This
920 * function returns the next descriptor in the chain,
921 * or -1U if we're at the end. */
922 static unsigned next_desc(struct vring_desc
*desc
)
926 /* If this descriptor says it doesn't chain, we're done. */
927 if (!(desc
->flags
& VRING_DESC_F_NEXT
))
930 /* Check they're not leading us off end of descriptors. */
932 /* Make sure compiler knows to grab that: we don't want it changing! */
933 /* We will use the result as an index in an array, so most
934 * architectures only need a compiler barrier here. */
935 read_barrier_depends();
940 static int get_indirect(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
,
941 struct iovec iov
[], unsigned int iov_size
,
942 unsigned int *out_num
, unsigned int *in_num
,
943 struct vhost_log
*log
, unsigned int *log_num
,
944 struct vring_desc
*indirect
)
946 struct vring_desc desc
;
947 unsigned int i
= 0, count
, found
= 0;
951 if (unlikely(indirect
->len
% sizeof desc
)) {
952 vq_err(vq
, "Invalid length in indirect descriptor: "
953 "len 0x%llx not multiple of 0x%zx\n",
954 (unsigned long long)indirect
->len
,
959 ret
= translate_desc(dev
, indirect
->addr
, indirect
->len
, vq
->indirect
,
960 ARRAY_SIZE(vq
->indirect
));
961 if (unlikely(ret
< 0)) {
962 vq_err(vq
, "Translation failure %d in indirect.\n", ret
);
966 /* We will use the result as an address to read from, so most
967 * architectures only need a compiler barrier here. */
968 read_barrier_depends();
970 count
= indirect
->len
/ sizeof desc
;
971 /* Buffers are chained via a 16 bit next field, so
972 * we can have at most 2^16 of these. */
973 if (unlikely(count
> USHRT_MAX
+ 1)) {
974 vq_err(vq
, "Indirect buffer length too big: %d\n",
980 unsigned iov_count
= *in_num
+ *out_num
;
981 if (unlikely(++found
> count
)) {
982 vq_err(vq
, "Loop detected: last one at %u "
983 "indirect size %u\n",
987 if (unlikely(memcpy_fromiovec((unsigned char *)&desc
, vq
->indirect
,
989 vq_err(vq
, "Failed indirect descriptor: idx %d, %zx\n",
990 i
, (size_t)indirect
->addr
+ i
* sizeof desc
);
993 if (unlikely(desc
.flags
& VRING_DESC_F_INDIRECT
)) {
994 vq_err(vq
, "Nested indirect descriptor: idx %d, %zx\n",
995 i
, (size_t)indirect
->addr
+ i
* sizeof desc
);
999 ret
= translate_desc(dev
, desc
.addr
, desc
.len
, iov
+ iov_count
,
1000 iov_size
- iov_count
);
1001 if (unlikely(ret
< 0)) {
1002 vq_err(vq
, "Translation failure %d indirect idx %d\n",
1006 /* If this is an input descriptor, increment that count. */
1007 if (desc
.flags
& VRING_DESC_F_WRITE
) {
1009 if (unlikely(log
)) {
1010 log
[*log_num
].addr
= desc
.addr
;
1011 log
[*log_num
].len
= desc
.len
;
1015 /* If it's an output descriptor, they're all supposed
1016 * to come before any input descriptors. */
1017 if (unlikely(*in_num
)) {
1018 vq_err(vq
, "Indirect descriptor "
1019 "has out after in: idx %d\n", i
);
1024 } while ((i
= next_desc(&desc
)) != -1);
1028 /* This looks in the virtqueue and for the first available buffer, and converts
1029 * it to an iovec for convenient access. Since descriptors consist of some
1030 * number of output then some number of input descriptors, it's actually two
1031 * iovecs, but we pack them into one and note how many of each there were.
1033 * This function returns the descriptor number found, or vq->num (which is
1034 * never a valid descriptor number) if none was found. A negative code is
1035 * returned on error. */
1036 int vhost_get_vq_desc(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
,
1037 struct iovec iov
[], unsigned int iov_size
,
1038 unsigned int *out_num
, unsigned int *in_num
,
1039 struct vhost_log
*log
, unsigned int *log_num
)
1041 struct vring_desc desc
;
1042 unsigned int i
, head
, found
= 0;
1046 /* Check it isn't doing very strange things with descriptor numbers. */
1047 last_avail_idx
= vq
->last_avail_idx
;
1048 if (unlikely(get_user(vq
->avail_idx
, &vq
->avail
->idx
))) {
1049 vq_err(vq
, "Failed to access avail idx at %p\n",
1054 if (unlikely((u16
)(vq
->avail_idx
- last_avail_idx
) > vq
->num
)) {
1055 vq_err(vq
, "Guest moved used index from %u to %u",
1056 last_avail_idx
, vq
->avail_idx
);
1060 /* If there's nothing new since last we looked, return invalid. */
1061 if (vq
->avail_idx
== last_avail_idx
)
1064 /* Only get avail ring entries after they have been exposed by guest. */
1067 /* Grab the next descriptor number they're advertising, and increment
1068 * the index we've seen. */
1069 if (unlikely(get_user(head
,
1070 &vq
->avail
->ring
[last_avail_idx
% vq
->num
]))) {
1071 vq_err(vq
, "Failed to read head: idx %d address %p\n",
1073 &vq
->avail
->ring
[last_avail_idx
% vq
->num
]);
1077 /* If their number is silly, that's an error. */
1078 if (unlikely(head
>= vq
->num
)) {
1079 vq_err(vq
, "Guest says index %u > %u is available",
1084 /* When we start there are none of either input nor output. */
1085 *out_num
= *in_num
= 0;
1091 unsigned iov_count
= *in_num
+ *out_num
;
1092 if (unlikely(i
>= vq
->num
)) {
1093 vq_err(vq
, "Desc index is %u > %u, head = %u",
1097 if (unlikely(++found
> vq
->num
)) {
1098 vq_err(vq
, "Loop detected: last one at %u "
1099 "vq size %u head %u\n",
1103 ret
= copy_from_user(&desc
, vq
->desc
+ i
, sizeof desc
);
1104 if (unlikely(ret
)) {
1105 vq_err(vq
, "Failed to get descriptor: idx %d addr %p\n",
1109 if (desc
.flags
& VRING_DESC_F_INDIRECT
) {
1110 ret
= get_indirect(dev
, vq
, iov
, iov_size
,
1112 log
, log_num
, &desc
);
1113 if (unlikely(ret
< 0)) {
1114 vq_err(vq
, "Failure detected "
1115 "in indirect descriptor at idx %d\n", i
);
1121 ret
= translate_desc(dev
, desc
.addr
, desc
.len
, iov
+ iov_count
,
1122 iov_size
- iov_count
);
1123 if (unlikely(ret
< 0)) {
1124 vq_err(vq
, "Translation failure %d descriptor idx %d\n",
1128 if (desc
.flags
& VRING_DESC_F_WRITE
) {
1129 /* If this is an input descriptor,
1130 * increment that count. */
1132 if (unlikely(log
)) {
1133 log
[*log_num
].addr
= desc
.addr
;
1134 log
[*log_num
].len
= desc
.len
;
1138 /* If it's an output descriptor, they're all supposed
1139 * to come before any input descriptors. */
1140 if (unlikely(*in_num
)) {
1141 vq_err(vq
, "Descriptor has out after in: "
1147 } while ((i
= next_desc(&desc
)) != -1);
1149 /* On success, increment avail index. */
1150 vq
->last_avail_idx
++;
1154 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1155 void vhost_discard_vq_desc(struct vhost_virtqueue
*vq
, int n
)
1157 vq
->last_avail_idx
-= n
;
1160 /* After we've used one of their buffers, we tell them about it. We'll then
1161 * want to notify the guest, using eventfd. */
1162 int vhost_add_used(struct vhost_virtqueue
*vq
, unsigned int head
, int len
)
1164 struct vring_used_elem __user
*used
;
1166 /* The virtqueue contains a ring of used buffers. Get a pointer to the
1167 * next entry in that used ring. */
1168 used
= &vq
->used
->ring
[vq
->last_used_idx
% vq
->num
];
1169 if (put_user(head
, &used
->id
)) {
1170 vq_err(vq
, "Failed to write used id");
1173 if (put_user(len
, &used
->len
)) {
1174 vq_err(vq
, "Failed to write used len");
1177 /* Make sure buffer is written before we update index. */
1179 if (put_user(vq
->last_used_idx
+ 1, &vq
->used
->idx
)) {
1180 vq_err(vq
, "Failed to increment used idx");
1183 if (unlikely(vq
->log_used
)) {
1184 /* Make sure data is seen before log. */
1186 /* Log used ring entry write. */
1187 log_write(vq
->log_base
,
1189 ((void __user
*)used
- (void __user
*)vq
->used
),
1191 /* Log used index update. */
1192 log_write(vq
->log_base
,
1193 vq
->log_addr
+ offsetof(struct vring_used
, idx
),
1194 sizeof vq
->used
->idx
);
1196 eventfd_signal(vq
->log_ctx
, 1);
1198 vq
->last_used_idx
++;
1202 static int __vhost_add_used_n(struct vhost_virtqueue
*vq
,
1203 struct vring_used_elem
*heads
,
1206 struct vring_used_elem __user
*used
;
1209 start
= vq
->last_used_idx
% vq
->num
;
1210 used
= vq
->used
->ring
+ start
;
1211 if (copy_to_user(used
, heads
, count
* sizeof *used
)) {
1212 vq_err(vq
, "Failed to write used");
1215 if (unlikely(vq
->log_used
)) {
1216 /* Make sure data is seen before log. */
1218 /* Log used ring entry write. */
1219 log_write(vq
->log_base
,
1221 ((void __user
*)used
- (void __user
*)vq
->used
),
1222 count
* sizeof *used
);
1224 vq
->last_used_idx
+= count
;
1228 /* After we've used one of their buffers, we tell them about it. We'll then
1229 * want to notify the guest, using eventfd. */
1230 int vhost_add_used_n(struct vhost_virtqueue
*vq
, struct vring_used_elem
*heads
,
1235 start
= vq
->last_used_idx
% vq
->num
;
1236 n
= vq
->num
- start
;
1238 r
= __vhost_add_used_n(vq
, heads
, n
);
1244 r
= __vhost_add_used_n(vq
, heads
, count
);
1246 /* Make sure buffer is written before we update index. */
1248 if (put_user(vq
->last_used_idx
, &vq
->used
->idx
)) {
1249 vq_err(vq
, "Failed to increment used idx");
1252 if (unlikely(vq
->log_used
)) {
1253 /* Log used index update. */
1254 log_write(vq
->log_base
,
1255 vq
->log_addr
+ offsetof(struct vring_used
, idx
),
1256 sizeof vq
->used
->idx
);
1258 eventfd_signal(vq
->log_ctx
, 1);
1263 /* This actually signals the guest, using eventfd. */
1264 void vhost_signal(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
)
1267 /* Flush out used index updates. This is paired
1268 * with the barrier that the Guest executes when enabling
1272 if (get_user(flags
, &vq
->avail
->flags
)) {
1273 vq_err(vq
, "Failed to get flags");
1277 /* If they don't want an interrupt, don't signal, unless empty. */
1278 if ((flags
& VRING_AVAIL_F_NO_INTERRUPT
) &&
1279 (vq
->avail_idx
!= vq
->last_avail_idx
||
1280 !vhost_has_feature(dev
, VIRTIO_F_NOTIFY_ON_EMPTY
)))
1283 /* Signal the Guest tell them we used something up. */
1285 eventfd_signal(vq
->call_ctx
, 1);
1288 /* And here's the combo meal deal. Supersize me! */
1289 void vhost_add_used_and_signal(struct vhost_dev
*dev
,
1290 struct vhost_virtqueue
*vq
,
1291 unsigned int head
, int len
)
1293 vhost_add_used(vq
, head
, len
);
1294 vhost_signal(dev
, vq
);
1297 /* multi-buffer version of vhost_add_used_and_signal */
1298 void vhost_add_used_and_signal_n(struct vhost_dev
*dev
,
1299 struct vhost_virtqueue
*vq
,
1300 struct vring_used_elem
*heads
, unsigned count
)
1302 vhost_add_used_n(vq
, heads
, count
);
1303 vhost_signal(dev
, vq
);
1306 /* OK, now we need to know about added descriptors. */
1307 bool vhost_enable_notify(struct vhost_virtqueue
*vq
)
1311 if (!(vq
->used_flags
& VRING_USED_F_NO_NOTIFY
))
1313 vq
->used_flags
&= ~VRING_USED_F_NO_NOTIFY
;
1314 r
= put_user(vq
->used_flags
, &vq
->used
->flags
);
1316 vq_err(vq
, "Failed to enable notification at %p: %d\n",
1317 &vq
->used
->flags
, r
);
1320 /* They could have slipped one in as we were doing that: make
1321 * sure it's written, then check again. */
1323 r
= get_user(avail_idx
, &vq
->avail
->idx
);
1325 vq_err(vq
, "Failed to check avail idx at %p: %d\n",
1326 &vq
->avail
->idx
, r
);
1330 return avail_idx
!= vq
->avail_idx
;
1333 /* We don't need to be notified again. */
1334 void vhost_disable_notify(struct vhost_virtqueue
*vq
)
1337 if (vq
->used_flags
& VRING_USED_F_NO_NOTIFY
)
1339 vq
->used_flags
|= VRING_USED_F_NO_NOTIFY
;
1340 r
= put_user(vq
->used_flags
, &vq
->used
->flags
);
1342 vq_err(vq
, "Failed to enable notification at %p: %d\n",
1343 &vq
->used
->flags
, r
);