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/workqueue.h>
21 #include <linux/rcupdate.h>
22 #include <linux/poll.h>
23 #include <linux/file.h>
24 #include <linux/highmem.h>
25 #include <linux/slab.h>
27 #include <linux/net.h>
28 #include <linux/if_packet.h>
29 #include <linux/if_arp.h>
36 VHOST_MEMORY_MAX_NREGIONS
= 64,
37 VHOST_MEMORY_F_LOG
= 0x1,
40 static struct workqueue_struct
*vhost_workqueue
;
42 static void vhost_poll_func(struct file
*file
, wait_queue_head_t
*wqh
,
45 struct vhost_poll
*poll
;
46 poll
= container_of(pt
, struct vhost_poll
, table
);
49 add_wait_queue(wqh
, &poll
->wait
);
52 static int vhost_poll_wakeup(wait_queue_t
*wait
, unsigned mode
, int sync
,
55 struct vhost_poll
*poll
;
56 poll
= container_of(wait
, struct vhost_poll
, wait
);
57 if (!((unsigned long)key
& poll
->mask
))
60 queue_work(vhost_workqueue
, &poll
->work
);
64 /* Init poll structure */
65 void vhost_poll_init(struct vhost_poll
*poll
, work_func_t func
,
68 INIT_WORK(&poll
->work
, func
);
69 init_waitqueue_func_entry(&poll
->wait
, vhost_poll_wakeup
);
70 init_poll_funcptr(&poll
->table
, vhost_poll_func
);
74 /* Start polling a file. We add ourselves to file's wait queue. The caller must
75 * keep a reference to a file until after vhost_poll_stop is called. */
76 void vhost_poll_start(struct vhost_poll
*poll
, struct file
*file
)
79 mask
= file
->f_op
->poll(file
, &poll
->table
);
81 vhost_poll_wakeup(&poll
->wait
, 0, 0, (void *)mask
);
84 /* Stop polling a file. After this function returns, it becomes safe to drop the
85 * file reference. You must also flush afterwards. */
86 void vhost_poll_stop(struct vhost_poll
*poll
)
88 remove_wait_queue(poll
->wqh
, &poll
->wait
);
91 /* Flush any work that has been scheduled. When calling this, don't hold any
92 * locks that are also used by the callback. */
93 void vhost_poll_flush(struct vhost_poll
*poll
)
95 flush_work(&poll
->work
);
98 void vhost_poll_queue(struct vhost_poll
*poll
)
100 queue_work(vhost_workqueue
, &poll
->work
);
103 static void vhost_vq_reset(struct vhost_dev
*dev
,
104 struct vhost_virtqueue
*vq
)
110 vq
->last_avail_idx
= 0;
112 vq
->last_used_idx
= 0;
115 vq
->log_used
= false;
116 vq
->log_addr
= -1ull;
118 vq
->private_data
= NULL
;
120 vq
->error_ctx
= NULL
;
128 long vhost_dev_init(struct vhost_dev
*dev
,
129 struct vhost_virtqueue
*vqs
, int nvqs
)
134 mutex_init(&dev
->mutex
);
136 dev
->log_file
= NULL
;
140 for (i
= 0; i
< dev
->nvqs
; ++i
) {
141 dev
->vqs
[i
].dev
= dev
;
142 mutex_init(&dev
->vqs
[i
].mutex
);
143 vhost_vq_reset(dev
, dev
->vqs
+ i
);
144 if (dev
->vqs
[i
].handle_kick
)
145 vhost_poll_init(&dev
->vqs
[i
].poll
,
146 dev
->vqs
[i
].handle_kick
,
152 /* Caller should have device mutex */
153 long vhost_dev_check_owner(struct vhost_dev
*dev
)
155 /* Are you the owner? If not, I don't think you mean to do that */
156 return dev
->mm
== current
->mm
? 0 : -EPERM
;
159 /* Caller should have device mutex */
160 static long vhost_dev_set_owner(struct vhost_dev
*dev
)
162 /* Is there an owner already? */
165 /* No owner, become one */
166 dev
->mm
= get_task_mm(current
);
170 /* Caller should have device mutex */
171 long vhost_dev_reset_owner(struct vhost_dev
*dev
)
173 struct vhost_memory
*memory
;
175 /* Restore memory to default empty mapping. */
176 memory
= kmalloc(offsetof(struct vhost_memory
, regions
), GFP_KERNEL
);
180 vhost_dev_cleanup(dev
);
182 memory
->nregions
= 0;
183 dev
->memory
= memory
;
187 /* Caller should have device mutex */
188 void vhost_dev_cleanup(struct vhost_dev
*dev
)
191 for (i
= 0; i
< dev
->nvqs
; ++i
) {
192 if (dev
->vqs
[i
].kick
&& dev
->vqs
[i
].handle_kick
) {
193 vhost_poll_stop(&dev
->vqs
[i
].poll
);
194 vhost_poll_flush(&dev
->vqs
[i
].poll
);
196 if (dev
->vqs
[i
].error_ctx
)
197 eventfd_ctx_put(dev
->vqs
[i
].error_ctx
);
198 if (dev
->vqs
[i
].error
)
199 fput(dev
->vqs
[i
].error
);
200 if (dev
->vqs
[i
].kick
)
201 fput(dev
->vqs
[i
].kick
);
202 if (dev
->vqs
[i
].call_ctx
)
203 eventfd_ctx_put(dev
->vqs
[i
].call_ctx
);
204 if (dev
->vqs
[i
].call
)
205 fput(dev
->vqs
[i
].call
);
206 vhost_vq_reset(dev
, dev
->vqs
+ i
);
209 eventfd_ctx_put(dev
->log_ctx
);
213 dev
->log_file
= NULL
;
214 /* No one will access memory at this point */
222 static int log_access_ok(void __user
*log_base
, u64 addr
, unsigned long sz
)
224 u64 a
= addr
/ VHOST_PAGE_SIZE
/ 8;
225 /* Make sure 64 bit math will not overflow. */
226 if (a
> ULONG_MAX
- (unsigned long)log_base
||
227 a
+ (unsigned long)log_base
> ULONG_MAX
)
230 return access_ok(VERIFY_WRITE
, log_base
+ a
,
231 (sz
+ VHOST_PAGE_SIZE
* 8 - 1) / VHOST_PAGE_SIZE
/ 8);
234 /* Caller should have vq mutex and device mutex. */
235 static int vq_memory_access_ok(void __user
*log_base
, struct vhost_memory
*mem
,
243 for (i
= 0; i
< mem
->nregions
; ++i
) {
244 struct vhost_memory_region
*m
= mem
->regions
+ i
;
245 unsigned long a
= m
->userspace_addr
;
246 if (m
->memory_size
> ULONG_MAX
)
248 else if (!access_ok(VERIFY_WRITE
, (void __user
*)a
,
251 else if (log_all
&& !log_access_ok(log_base
,
259 /* Can we switch to this memory table? */
260 /* Caller should have device mutex but not vq mutex */
261 static int memory_access_ok(struct vhost_dev
*d
, struct vhost_memory
*mem
,
265 for (i
= 0; i
< d
->nvqs
; ++i
) {
267 mutex_lock(&d
->vqs
[i
].mutex
);
268 /* If ring is inactive, will check when it's enabled. */
269 if (d
->vqs
[i
].private_data
)
270 ok
= vq_memory_access_ok(d
->vqs
[i
].log_base
, mem
,
274 mutex_unlock(&d
->vqs
[i
].mutex
);
281 static int vq_access_ok(unsigned int num
,
282 struct vring_desc __user
*desc
,
283 struct vring_avail __user
*avail
,
284 struct vring_used __user
*used
)
286 return access_ok(VERIFY_READ
, desc
, num
* sizeof *desc
) &&
287 access_ok(VERIFY_READ
, avail
,
288 sizeof *avail
+ num
* sizeof *avail
->ring
) &&
289 access_ok(VERIFY_WRITE
, used
,
290 sizeof *used
+ num
* sizeof *used
->ring
);
293 /* Can we log writes? */
294 /* Caller should have device mutex but not vq mutex */
295 int vhost_log_access_ok(struct vhost_dev
*dev
)
297 return memory_access_ok(dev
, dev
->memory
, 1);
300 /* Verify access for write logging. */
301 /* Caller should have vq mutex and device mutex */
302 static int vq_log_access_ok(struct vhost_virtqueue
*vq
, void __user
*log_base
)
304 return vq_memory_access_ok(log_base
, vq
->dev
->memory
,
305 vhost_has_feature(vq
->dev
, VHOST_F_LOG_ALL
)) &&
306 (!vq
->log_used
|| log_access_ok(log_base
, vq
->log_addr
,
308 vq
->num
* sizeof *vq
->used
->ring
));
311 /* Can we start vq? */
312 /* Caller should have vq mutex and device mutex */
313 int vhost_vq_access_ok(struct vhost_virtqueue
*vq
)
315 return vq_access_ok(vq
->num
, vq
->desc
, vq
->avail
, vq
->used
) &&
316 vq_log_access_ok(vq
, vq
->log_base
);
319 static long vhost_set_memory(struct vhost_dev
*d
, struct vhost_memory __user
*m
)
321 struct vhost_memory mem
, *newmem
, *oldmem
;
322 unsigned long size
= offsetof(struct vhost_memory
, regions
);
323 if (copy_from_user(&mem
, m
, size
))
327 if (mem
.nregions
> VHOST_MEMORY_MAX_NREGIONS
)
329 newmem
= kmalloc(size
+ mem
.nregions
* sizeof *m
->regions
, GFP_KERNEL
);
333 memcpy(newmem
, &mem
, size
);
334 if (copy_from_user(newmem
->regions
, m
->regions
,
335 mem
.nregions
* sizeof *m
->regions
)) {
340 if (!memory_access_ok(d
, newmem
, vhost_has_feature(d
, VHOST_F_LOG_ALL
))) {
345 rcu_assign_pointer(d
->memory
, newmem
);
351 static int init_used(struct vhost_virtqueue
*vq
,
352 struct vring_used __user
*used
)
354 int r
= put_user(vq
->used_flags
, &used
->flags
);
357 return get_user(vq
->last_used_idx
, &used
->idx
);
360 static long vhost_set_vring(struct vhost_dev
*d
, int ioctl
, void __user
*argp
)
362 struct file
*eventfp
, *filep
= NULL
,
363 *pollstart
= NULL
, *pollstop
= NULL
;
364 struct eventfd_ctx
*ctx
= NULL
;
365 u32 __user
*idxp
= argp
;
366 struct vhost_virtqueue
*vq
;
367 struct vhost_vring_state s
;
368 struct vhost_vring_file f
;
369 struct vhost_vring_addr a
;
373 r
= get_user(idx
, idxp
);
381 mutex_lock(&vq
->mutex
);
384 case VHOST_SET_VRING_NUM
:
385 /* Resizing ring with an active backend?
386 * You don't want to do that. */
387 if (vq
->private_data
) {
391 if (copy_from_user(&s
, argp
, sizeof s
)) {
395 if (!s
.num
|| s
.num
> 0xffff || (s
.num
& (s
.num
- 1))) {
401 case VHOST_SET_VRING_BASE
:
402 /* Moving base with an active backend?
403 * You don't want to do that. */
404 if (vq
->private_data
) {
408 if (copy_from_user(&s
, argp
, sizeof s
)) {
412 if (s
.num
> 0xffff) {
416 vq
->last_avail_idx
= s
.num
;
417 /* Forget the cached index value. */
418 vq
->avail_idx
= vq
->last_avail_idx
;
420 case VHOST_GET_VRING_BASE
:
422 s
.num
= vq
->last_avail_idx
;
423 if (copy_to_user(argp
, &s
, sizeof s
))
426 case VHOST_SET_VRING_ADDR
:
427 if (copy_from_user(&a
, argp
, sizeof a
)) {
431 if (a
.flags
& ~(0x1 << VHOST_VRING_F_LOG
)) {
435 /* For 32bit, verify that the top 32bits of the user
436 data are set to zero. */
437 if ((u64
)(unsigned long)a
.desc_user_addr
!= a
.desc_user_addr
||
438 (u64
)(unsigned long)a
.used_user_addr
!= a
.used_user_addr
||
439 (u64
)(unsigned long)a
.avail_user_addr
!= a
.avail_user_addr
) {
443 if ((a
.avail_user_addr
& (sizeof *vq
->avail
->ring
- 1)) ||
444 (a
.used_user_addr
& (sizeof *vq
->used
->ring
- 1)) ||
445 (a
.log_guest_addr
& (sizeof *vq
->used
->ring
- 1))) {
450 /* We only verify access here if backend is configured.
451 * If it is not, we don't as size might not have been setup.
452 * We will verify when backend is configured. */
453 if (vq
->private_data
) {
454 if (!vq_access_ok(vq
->num
,
455 (void __user
*)(unsigned long)a
.desc_user_addr
,
456 (void __user
*)(unsigned long)a
.avail_user_addr
,
457 (void __user
*)(unsigned long)a
.used_user_addr
)) {
462 /* Also validate log access for used ring if enabled. */
463 if ((a
.flags
& (0x1 << VHOST_VRING_F_LOG
)) &&
464 !log_access_ok(vq
->log_base
, a
.log_guest_addr
,
466 vq
->num
* sizeof *vq
->used
->ring
)) {
472 r
= init_used(vq
, (struct vring_used __user
*)(unsigned long)
476 vq
->log_used
= !!(a
.flags
& (0x1 << VHOST_VRING_F_LOG
));
477 vq
->desc
= (void __user
*)(unsigned long)a
.desc_user_addr
;
478 vq
->avail
= (void __user
*)(unsigned long)a
.avail_user_addr
;
479 vq
->log_addr
= a
.log_guest_addr
;
480 vq
->used
= (void __user
*)(unsigned long)a
.used_user_addr
;
482 case VHOST_SET_VRING_KICK
:
483 if (copy_from_user(&f
, argp
, sizeof f
)) {
487 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
488 if (IS_ERR(eventfp
)) {
489 r
= PTR_ERR(eventfp
);
492 if (eventfp
!= vq
->kick
) {
493 pollstop
= filep
= vq
->kick
;
494 pollstart
= vq
->kick
= eventfp
;
498 case VHOST_SET_VRING_CALL
:
499 if (copy_from_user(&f
, argp
, sizeof f
)) {
503 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
504 if (IS_ERR(eventfp
)) {
505 r
= PTR_ERR(eventfp
);
508 if (eventfp
!= vq
->call
) {
512 vq
->call_ctx
= eventfp
?
513 eventfd_ctx_fileget(eventfp
) : NULL
;
517 case VHOST_SET_VRING_ERR
:
518 if (copy_from_user(&f
, argp
, sizeof f
)) {
522 eventfp
= f
.fd
== -1 ? NULL
: eventfd_fget(f
.fd
);
523 if (IS_ERR(eventfp
)) {
524 r
= PTR_ERR(eventfp
);
527 if (eventfp
!= vq
->error
) {
531 vq
->error_ctx
= eventfp
?
532 eventfd_ctx_fileget(eventfp
) : NULL
;
540 if (pollstop
&& vq
->handle_kick
)
541 vhost_poll_stop(&vq
->poll
);
544 eventfd_ctx_put(ctx
);
548 if (pollstart
&& vq
->handle_kick
)
549 vhost_poll_start(&vq
->poll
, vq
->kick
);
551 mutex_unlock(&vq
->mutex
);
553 if (pollstop
&& vq
->handle_kick
)
554 vhost_poll_flush(&vq
->poll
);
558 /* Caller must have device mutex */
559 long vhost_dev_ioctl(struct vhost_dev
*d
, unsigned int ioctl
, unsigned long arg
)
561 void __user
*argp
= (void __user
*)arg
;
562 struct file
*eventfp
, *filep
= NULL
;
563 struct eventfd_ctx
*ctx
= NULL
;
568 /* If you are not the owner, you can become one */
569 if (ioctl
== VHOST_SET_OWNER
) {
570 r
= vhost_dev_set_owner(d
);
574 /* You must be the owner to do anything else */
575 r
= vhost_dev_check_owner(d
);
580 case VHOST_SET_MEM_TABLE
:
581 r
= vhost_set_memory(d
, argp
);
583 case VHOST_SET_LOG_BASE
:
584 if (copy_from_user(&p
, argp
, sizeof p
)) {
588 if ((u64
)(unsigned long)p
!= p
) {
592 for (i
= 0; i
< d
->nvqs
; ++i
) {
593 struct vhost_virtqueue
*vq
;
594 void __user
*base
= (void __user
*)(unsigned long)p
;
596 mutex_lock(&vq
->mutex
);
597 /* If ring is inactive, will check when it's enabled. */
598 if (vq
->private_data
&& !vq_log_access_ok(vq
, base
))
602 mutex_unlock(&vq
->mutex
);
605 case VHOST_SET_LOG_FD
:
606 r
= get_user(fd
, (int __user
*)argp
);
609 eventfp
= fd
== -1 ? NULL
: eventfd_fget(fd
);
610 if (IS_ERR(eventfp
)) {
611 r
= PTR_ERR(eventfp
);
614 if (eventfp
!= d
->log_file
) {
617 d
->log_ctx
= eventfp
?
618 eventfd_ctx_fileget(eventfp
) : NULL
;
621 for (i
= 0; i
< d
->nvqs
; ++i
) {
622 mutex_lock(&d
->vqs
[i
].mutex
);
623 d
->vqs
[i
].log_ctx
= d
->log_ctx
;
624 mutex_unlock(&d
->vqs
[i
].mutex
);
627 eventfd_ctx_put(ctx
);
632 r
= vhost_set_vring(d
, ioctl
, argp
);
639 static const struct vhost_memory_region
*find_region(struct vhost_memory
*mem
,
640 __u64 addr
, __u32 len
)
642 struct vhost_memory_region
*reg
;
644 /* linear search is not brilliant, but we really have on the order of 6
645 * regions in practice */
646 for (i
= 0; i
< mem
->nregions
; ++i
) {
647 reg
= mem
->regions
+ i
;
648 if (reg
->guest_phys_addr
<= addr
&&
649 reg
->guest_phys_addr
+ reg
->memory_size
- 1 >= addr
)
655 /* TODO: This is really inefficient. We need something like get_user()
656 * (instruction directly accesses the data, with an exception table entry
657 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
659 static int set_bit_to_user(int nr
, void __user
*addr
)
661 unsigned long log
= (unsigned long)addr
;
664 int bit
= nr
+ (log
% PAGE_SIZE
) * 8;
666 r
= get_user_pages_fast(log
, 1, 1, &page
);
670 base
= kmap_atomic(page
, KM_USER0
);
672 kunmap_atomic(base
, KM_USER0
);
673 set_page_dirty_lock(page
);
678 static int log_write(void __user
*log_base
,
679 u64 write_address
, u64 write_length
)
684 write_address
/= VHOST_PAGE_SIZE
;
686 u64 base
= (u64
)(unsigned long)log_base
;
687 u64 log
= base
+ write_address
/ 8;
688 int bit
= write_address
% 8;
689 if ((u64
)(unsigned long)log
!= log
)
691 r
= set_bit_to_user(bit
, (void __user
*)(unsigned long)log
);
694 if (write_length
<= VHOST_PAGE_SIZE
)
696 write_length
-= VHOST_PAGE_SIZE
;
697 write_address
+= VHOST_PAGE_SIZE
;
702 int vhost_log_write(struct vhost_virtqueue
*vq
, struct vhost_log
*log
,
703 unsigned int log_num
, u64 len
)
707 /* Make sure data written is seen before log. */
709 for (i
= 0; i
< log_num
; ++i
) {
710 u64 l
= min(log
[i
].len
, len
);
711 r
= log_write(vq
->log_base
, log
[i
].addr
, l
);
719 eventfd_signal(vq
->log_ctx
, 1);
720 /* Length written exceeds what we have stored. This is a bug. */
725 static int translate_desc(struct vhost_dev
*dev
, u64 addr
, u32 len
,
726 struct iovec iov
[], int iov_size
)
728 const struct vhost_memory_region
*reg
;
729 struct vhost_memory
*mem
;
736 mem
= rcu_dereference(dev
->memory
);
737 while ((u64
)len
> s
) {
739 if (unlikely(ret
>= iov_size
)) {
743 reg
= find_region(mem
, addr
, len
);
744 if (unlikely(!reg
)) {
749 size
= reg
->memory_size
- addr
+ reg
->guest_phys_addr
;
750 _iov
->iov_len
= min((u64
)len
, size
);
751 _iov
->iov_base
= (void __user
*)(unsigned long)
752 (reg
->userspace_addr
+ addr
- reg
->guest_phys_addr
);
762 /* Each buffer in the virtqueues is actually a chain of descriptors. This
763 * function returns the next descriptor in the chain,
764 * or -1U if we're at the end. */
765 static unsigned next_desc(struct vring_desc
*desc
)
769 /* If this descriptor says it doesn't chain, we're done. */
770 if (!(desc
->flags
& VRING_DESC_F_NEXT
))
773 /* Check they're not leading us off end of descriptors. */
775 /* Make sure compiler knows to grab that: we don't want it changing! */
776 /* We will use the result as an index in an array, so most
777 * architectures only need a compiler barrier here. */
778 read_barrier_depends();
783 static int get_indirect(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
,
784 struct iovec iov
[], unsigned int iov_size
,
785 unsigned int *out_num
, unsigned int *in_num
,
786 struct vhost_log
*log
, unsigned int *log_num
,
787 struct vring_desc
*indirect
)
789 struct vring_desc desc
;
790 unsigned int i
= 0, count
, found
= 0;
794 if (unlikely(indirect
->len
% sizeof desc
)) {
795 vq_err(vq
, "Invalid length in indirect descriptor: "
796 "len 0x%llx not multiple of 0x%zx\n",
797 (unsigned long long)indirect
->len
,
802 ret
= translate_desc(dev
, indirect
->addr
, indirect
->len
, vq
->indirect
,
803 ARRAY_SIZE(vq
->indirect
));
804 if (unlikely(ret
< 0)) {
805 vq_err(vq
, "Translation failure %d in indirect.\n", ret
);
809 /* We will use the result as an address to read from, so most
810 * architectures only need a compiler barrier here. */
811 read_barrier_depends();
813 count
= indirect
->len
/ sizeof desc
;
814 /* Buffers are chained via a 16 bit next field, so
815 * we can have at most 2^16 of these. */
816 if (unlikely(count
> USHRT_MAX
+ 1)) {
817 vq_err(vq
, "Indirect buffer length too big: %d\n",
823 unsigned iov_count
= *in_num
+ *out_num
;
824 if (unlikely(++found
> count
)) {
825 vq_err(vq
, "Loop detected: last one at %u "
826 "indirect size %u\n",
830 if (unlikely(memcpy_fromiovec((unsigned char *)&desc
, vq
->indirect
,
832 vq_err(vq
, "Failed indirect descriptor: idx %d, %zx\n",
833 i
, (size_t)indirect
->addr
+ i
* sizeof desc
);
836 if (unlikely(desc
.flags
& VRING_DESC_F_INDIRECT
)) {
837 vq_err(vq
, "Nested indirect descriptor: idx %d, %zx\n",
838 i
, (size_t)indirect
->addr
+ i
* sizeof desc
);
842 ret
= translate_desc(dev
, desc
.addr
, desc
.len
, iov
+ iov_count
,
843 iov_size
- iov_count
);
844 if (unlikely(ret
< 0)) {
845 vq_err(vq
, "Translation failure %d indirect idx %d\n",
849 /* If this is an input descriptor, increment that count. */
850 if (desc
.flags
& VRING_DESC_F_WRITE
) {
853 log
[*log_num
].addr
= desc
.addr
;
854 log
[*log_num
].len
= desc
.len
;
858 /* If it's an output descriptor, they're all supposed
859 * to come before any input descriptors. */
860 if (unlikely(*in_num
)) {
861 vq_err(vq
, "Indirect descriptor "
862 "has out after in: idx %d\n", i
);
867 } while ((i
= next_desc(&desc
)) != -1);
871 /* This looks in the virtqueue and for the first available buffer, and converts
872 * it to an iovec for convenient access. Since descriptors consist of some
873 * number of output then some number of input descriptors, it's actually two
874 * iovecs, but we pack them into one and note how many of each there were.
876 * This function returns the descriptor number found, or vq->num (which is
877 * never a valid descriptor number) if none was found. A negative code is
878 * returned on error. */
879 int vhost_get_vq_desc(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
,
880 struct iovec iov
[], unsigned int iov_size
,
881 unsigned int *out_num
, unsigned int *in_num
,
882 struct vhost_log
*log
, unsigned int *log_num
)
884 struct vring_desc desc
;
885 unsigned int i
, head
, found
= 0;
889 /* Check it isn't doing very strange things with descriptor numbers. */
890 last_avail_idx
= vq
->last_avail_idx
;
891 if (unlikely(get_user(vq
->avail_idx
, &vq
->avail
->idx
))) {
892 vq_err(vq
, "Failed to access avail idx at %p\n",
897 if (unlikely((u16
)(vq
->avail_idx
- last_avail_idx
) > vq
->num
)) {
898 vq_err(vq
, "Guest moved used index from %u to %u",
899 last_avail_idx
, vq
->avail_idx
);
903 /* If there's nothing new since last we looked, return invalid. */
904 if (vq
->avail_idx
== last_avail_idx
)
907 /* Only get avail ring entries after they have been exposed by guest. */
910 /* Grab the next descriptor number they're advertising, and increment
911 * the index we've seen. */
912 if (unlikely(get_user(head
,
913 &vq
->avail
->ring
[last_avail_idx
% vq
->num
]))) {
914 vq_err(vq
, "Failed to read head: idx %d address %p\n",
916 &vq
->avail
->ring
[last_avail_idx
% vq
->num
]);
920 /* If their number is silly, that's an error. */
921 if (unlikely(head
>= vq
->num
)) {
922 vq_err(vq
, "Guest says index %u > %u is available",
927 /* When we start there are none of either input nor output. */
928 *out_num
= *in_num
= 0;
934 unsigned iov_count
= *in_num
+ *out_num
;
935 if (unlikely(i
>= vq
->num
)) {
936 vq_err(vq
, "Desc index is %u > %u, head = %u",
940 if (unlikely(++found
> vq
->num
)) {
941 vq_err(vq
, "Loop detected: last one at %u "
942 "vq size %u head %u\n",
946 ret
= copy_from_user(&desc
, vq
->desc
+ i
, sizeof desc
);
948 vq_err(vq
, "Failed to get descriptor: idx %d addr %p\n",
952 if (desc
.flags
& VRING_DESC_F_INDIRECT
) {
953 ret
= get_indirect(dev
, vq
, iov
, iov_size
,
955 log
, log_num
, &desc
);
956 if (unlikely(ret
< 0)) {
957 vq_err(vq
, "Failure detected "
958 "in indirect descriptor at idx %d\n", i
);
964 ret
= translate_desc(dev
, desc
.addr
, desc
.len
, iov
+ iov_count
,
965 iov_size
- iov_count
);
966 if (unlikely(ret
< 0)) {
967 vq_err(vq
, "Translation failure %d descriptor idx %d\n",
971 if (desc
.flags
& VRING_DESC_F_WRITE
) {
972 /* If this is an input descriptor,
973 * increment that count. */
976 log
[*log_num
].addr
= desc
.addr
;
977 log
[*log_num
].len
= desc
.len
;
981 /* If it's an output descriptor, they're all supposed
982 * to come before any input descriptors. */
983 if (unlikely(*in_num
)) {
984 vq_err(vq
, "Descriptor has out after in: "
990 } while ((i
= next_desc(&desc
)) != -1);
992 /* On success, increment avail index. */
993 vq
->last_avail_idx
++;
997 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
998 void vhost_discard_vq_desc(struct vhost_virtqueue
*vq
)
1000 vq
->last_avail_idx
--;
1003 /* After we've used one of their buffers, we tell them about it. We'll then
1004 * want to notify the guest, using eventfd. */
1005 int vhost_add_used(struct vhost_virtqueue
*vq
, unsigned int head
, int len
)
1007 struct vring_used_elem __user
*used
;
1009 /* The virtqueue contains a ring of used buffers. Get a pointer to the
1010 * next entry in that used ring. */
1011 used
= &vq
->used
->ring
[vq
->last_used_idx
% vq
->num
];
1012 if (put_user(head
, &used
->id
)) {
1013 vq_err(vq
, "Failed to write used id");
1016 if (put_user(len
, &used
->len
)) {
1017 vq_err(vq
, "Failed to write used len");
1020 /* Make sure buffer is written before we update index. */
1022 if (put_user(vq
->last_used_idx
+ 1, &vq
->used
->idx
)) {
1023 vq_err(vq
, "Failed to increment used idx");
1026 if (unlikely(vq
->log_used
)) {
1027 /* Make sure data is seen before log. */
1029 /* Log used ring entry write. */
1030 log_write(vq
->log_base
,
1032 ((void __user
*)used
- (void __user
*)vq
->used
),
1034 /* Log used index update. */
1035 log_write(vq
->log_base
,
1036 vq
->log_addr
+ offsetof(struct vring_used
, idx
),
1037 sizeof vq
->used
->idx
);
1039 eventfd_signal(vq
->log_ctx
, 1);
1041 vq
->last_used_idx
++;
1045 /* This actually signals the guest, using eventfd. */
1046 void vhost_signal(struct vhost_dev
*dev
, struct vhost_virtqueue
*vq
)
1049 /* Flush out used index updates. This is paired
1050 * with the barrier that the Guest executes when enabling
1054 if (get_user(flags
, &vq
->avail
->flags
)) {
1055 vq_err(vq
, "Failed to get flags");
1059 /* If they don't want an interrupt, don't signal, unless empty. */
1060 if ((flags
& VRING_AVAIL_F_NO_INTERRUPT
) &&
1061 (vq
->avail_idx
!= vq
->last_avail_idx
||
1062 !vhost_has_feature(dev
, VIRTIO_F_NOTIFY_ON_EMPTY
)))
1065 /* Signal the Guest tell them we used something up. */
1067 eventfd_signal(vq
->call_ctx
, 1);
1070 /* And here's the combo meal deal. Supersize me! */
1071 void vhost_add_used_and_signal(struct vhost_dev
*dev
,
1072 struct vhost_virtqueue
*vq
,
1073 unsigned int head
, int len
)
1075 vhost_add_used(vq
, head
, len
);
1076 vhost_signal(dev
, vq
);
1079 /* OK, now we need to know about added descriptors. */
1080 bool vhost_enable_notify(struct vhost_virtqueue
*vq
)
1084 if (!(vq
->used_flags
& VRING_USED_F_NO_NOTIFY
))
1086 vq
->used_flags
&= ~VRING_USED_F_NO_NOTIFY
;
1087 r
= put_user(vq
->used_flags
, &vq
->used
->flags
);
1089 vq_err(vq
, "Failed to enable notification at %p: %d\n",
1090 &vq
->used
->flags
, r
);
1093 /* They could have slipped one in as we were doing that: make
1094 * sure it's written, then check again. */
1096 r
= get_user(avail_idx
, &vq
->avail
->idx
);
1098 vq_err(vq
, "Failed to check avail idx at %p: %d\n",
1099 &vq
->avail
->idx
, r
);
1103 return avail_idx
!= vq
->last_avail_idx
;
1106 /* We don't need to be notified again. */
1107 void vhost_disable_notify(struct vhost_virtqueue
*vq
)
1110 if (vq
->used_flags
& VRING_USED_F_NO_NOTIFY
)
1112 vq
->used_flags
|= VRING_USED_F_NO_NOTIFY
;
1113 r
= put_user(vq
->used_flags
, &vq
->used
->flags
);
1115 vq_err(vq
, "Failed to enable notification at %p: %d\n",
1116 &vq
->used
->flags
, r
);
1119 int vhost_init(void)
1121 vhost_workqueue
= create_singlethread_workqueue("vhost");
1122 if (!vhost_workqueue
)
1127 void vhost_cleanup(void)
1129 destroy_workqueue(vhost_workqueue
);