1 /* Copyright (C) 2009 Red Hat, Inc.
2 * Author: Michael S. Tsirkin <mst@redhat.com>
4 * This work is licensed under the terms of the GNU GPL, version 2.
6 * virtio-net server in host kernel.
9 #include <linux/compat.h>
10 #include <linux/eventfd.h>
11 #include <linux/vhost.h>
12 #include <linux/virtio_net.h>
13 #include <linux/miscdevice.h>
14 #include <linux/module.h>
15 #include <linux/mutex.h>
16 #include <linux/workqueue.h>
17 #include <linux/rcupdate.h>
18 #include <linux/file.h>
19 #include <linux/slab.h>
21 #include <linux/net.h>
22 #include <linux/if_packet.h>
23 #include <linux/if_arp.h>
24 #include <linux/if_tun.h>
25 #include <linux/if_macvlan.h>
31 /* Max number of bytes transferred before requeueing the job.
32 * Using this limit prevents one virtqueue from starving others. */
33 #define VHOST_NET_WEIGHT 0x80000
41 enum vhost_net_poll_state
{
42 VHOST_NET_POLL_DISABLED
= 0,
43 VHOST_NET_POLL_STARTED
= 1,
44 VHOST_NET_POLL_STOPPED
= 2,
49 struct vhost_virtqueue vqs
[VHOST_NET_VQ_MAX
];
50 struct vhost_poll poll
[VHOST_NET_VQ_MAX
];
51 /* Tells us whether we are polling a socket for TX.
52 * We only do this when socket buffer fills up.
53 * Protected by tx vq lock. */
54 enum vhost_net_poll_state tx_poll_state
;
57 /* Pop first len bytes from iovec. Return number of segments used. */
58 static int move_iovec_hdr(struct iovec
*from
, struct iovec
*to
,
59 size_t len
, int iov_count
)
64 while (len
&& seg
< iov_count
) {
65 size
= min(from
->iov_len
, len
);
66 to
->iov_base
= from
->iov_base
;
68 from
->iov_len
-= size
;
69 from
->iov_base
+= size
;
77 /* Copy iovec entries for len bytes from iovec. */
78 static void copy_iovec_hdr(const struct iovec
*from
, struct iovec
*to
,
79 size_t len
, int iovcount
)
84 while (len
&& seg
< iovcount
) {
85 size
= min(from
->iov_len
, len
);
86 to
->iov_base
= from
->iov_base
;
95 /* Caller must have TX VQ lock */
96 static void tx_poll_stop(struct vhost_net
*net
)
98 if (likely(net
->tx_poll_state
!= VHOST_NET_POLL_STARTED
))
100 vhost_poll_stop(net
->poll
+ VHOST_NET_VQ_TX
);
101 net
->tx_poll_state
= VHOST_NET_POLL_STOPPED
;
104 /* Caller must have TX VQ lock */
105 static void tx_poll_start(struct vhost_net
*net
, struct socket
*sock
)
107 if (unlikely(net
->tx_poll_state
!= VHOST_NET_POLL_STOPPED
))
109 vhost_poll_start(net
->poll
+ VHOST_NET_VQ_TX
, sock
->file
);
110 net
->tx_poll_state
= VHOST_NET_POLL_STARTED
;
113 /* Expects to be always run from workqueue - which acts as
114 * read-size critical section for our kind of RCU. */
115 static void handle_tx(struct vhost_net
*net
)
117 struct vhost_virtqueue
*vq
= &net
->dev
.vqs
[VHOST_NET_VQ_TX
];
120 struct msghdr msg
= {
126 .msg_flags
= MSG_DONTWAIT
,
128 size_t len
, total_len
= 0;
133 /* TODO: check that we are running from vhost_worker? */
134 sock
= rcu_dereference_check(vq
->private_data
, 1);
138 wmem
= atomic_read(&sock
->sk
->sk_wmem_alloc
);
139 if (wmem
>= sock
->sk
->sk_sndbuf
) {
140 mutex_lock(&vq
->mutex
);
141 tx_poll_start(net
, sock
);
142 mutex_unlock(&vq
->mutex
);
146 mutex_lock(&vq
->mutex
);
147 vhost_disable_notify(vq
);
149 if (wmem
< sock
->sk
->sk_sndbuf
/ 2)
151 hdr_size
= vq
->vhost_hlen
;
154 head
= vhost_get_vq_desc(&net
->dev
, vq
, vq
->iov
,
158 /* On error, stop handling until the next kick. */
159 if (unlikely(head
< 0))
161 /* Nothing new? Wait for eventfd to tell us they refilled. */
162 if (head
== vq
->num
) {
163 wmem
= atomic_read(&sock
->sk
->sk_wmem_alloc
);
164 if (wmem
>= sock
->sk
->sk_sndbuf
* 3 / 4) {
165 tx_poll_start(net
, sock
);
166 set_bit(SOCK_ASYNC_NOSPACE
, &sock
->flags
);
169 if (unlikely(vhost_enable_notify(vq
))) {
170 vhost_disable_notify(vq
);
176 vq_err(vq
, "Unexpected descriptor format for TX: "
177 "out %d, int %d\n", out
, in
);
180 /* Skip header. TODO: support TSO. */
181 s
= move_iovec_hdr(vq
->iov
, vq
->hdr
, hdr_size
, out
);
182 msg
.msg_iovlen
= out
;
183 len
= iov_length(vq
->iov
, out
);
186 vq_err(vq
, "Unexpected header len for TX: "
187 "%zd expected %zd\n",
188 iov_length(vq
->hdr
, s
), hdr_size
);
191 /* TODO: Check specific error and bomb out unless ENOBUFS? */
192 err
= sock
->ops
->sendmsg(NULL
, sock
, &msg
, len
);
193 if (unlikely(err
< 0)) {
194 vhost_discard_vq_desc(vq
, 1);
195 tx_poll_start(net
, sock
);
199 pr_debug("Truncated TX packet: "
200 " len %d != %zd\n", err
, len
);
201 vhost_add_used_and_signal(&net
->dev
, vq
, head
, 0);
203 if (unlikely(total_len
>= VHOST_NET_WEIGHT
)) {
204 vhost_poll_queue(&vq
->poll
);
209 mutex_unlock(&vq
->mutex
);
212 static int peek_head_len(struct sock
*sk
)
214 struct sk_buff
*head
;
218 head
= skb_peek(&sk
->sk_receive_queue
);
225 /* This is a multi-buffer version of vhost_get_desc, that works if
226 * vq has read descriptors only.
227 * @vq - the relevant virtqueue
228 * @datalen - data length we'll be reading
229 * @iovcount - returned count of io vectors we fill
231 * @log_num - log offset
232 * @quota - headcount quota, 1 for big buffer
233 * returns number of buffer heads allocated, negative on error
235 static int get_rx_bufs(struct vhost_virtqueue
*vq
,
236 struct vring_used_elem
*heads
,
239 struct vhost_log
*log
,
243 unsigned int out
, in
;
249 while (datalen
> 0 && headcount
< quota
) {
250 if (unlikely(seg
>= UIO_MAXIOV
)) {
254 d
= vhost_get_vq_desc(vq
->dev
, vq
, vq
->iov
+ seg
,
255 ARRAY_SIZE(vq
->iov
) - seg
, &out
,
261 if (unlikely(out
|| in
<= 0)) {
262 vq_err(vq
, "unexpected descriptor format for RX: "
263 "out %d, in %d\n", out
, in
);
271 heads
[headcount
].id
= d
;
272 heads
[headcount
].len
= iov_length(vq
->iov
+ seg
, in
);
273 datalen
-= heads
[headcount
].len
;
277 heads
[headcount
- 1].len
+= datalen
;
283 vhost_discard_vq_desc(vq
, headcount
);
287 /* Expects to be always run from workqueue - which acts as
288 * read-size critical section for our kind of RCU. */
289 static void handle_rx(struct vhost_net
*net
)
291 struct vhost_virtqueue
*vq
= &net
->dev
.vqs
[VHOST_NET_VQ_RX
];
292 unsigned uninitialized_var(in
), log
;
293 struct vhost_log
*vq_log
;
294 struct msghdr msg
= {
297 .msg_control
= NULL
, /* FIXME: get and handle RX aux data. */
300 .msg_flags
= MSG_DONTWAIT
,
302 struct virtio_net_hdr_mrg_rxbuf hdr
= {
304 .hdr
.gso_type
= VIRTIO_NET_HDR_GSO_NONE
306 size_t total_len
= 0;
307 int err
, headcount
, mergeable
;
308 size_t vhost_hlen
, sock_hlen
;
309 size_t vhost_len
, sock_len
;
310 /* TODO: check that we are running from vhost_worker? */
311 struct socket
*sock
= rcu_dereference_check(vq
->private_data
, 1);
313 if (!sock
|| skb_queue_empty(&sock
->sk
->sk_receive_queue
))
316 mutex_lock(&vq
->mutex
);
317 vhost_disable_notify(vq
);
318 vhost_hlen
= vq
->vhost_hlen
;
319 sock_hlen
= vq
->sock_hlen
;
321 vq_log
= unlikely(vhost_has_feature(&net
->dev
, VHOST_F_LOG_ALL
)) ?
323 mergeable
= vhost_has_feature(&net
->dev
, VIRTIO_NET_F_MRG_RXBUF
);
325 while ((sock_len
= peek_head_len(sock
->sk
))) {
326 sock_len
+= sock_hlen
;
327 vhost_len
= sock_len
+ vhost_hlen
;
328 headcount
= get_rx_bufs(vq
, vq
->heads
, vhost_len
,
330 likely(mergeable
) ? UIO_MAXIOV
: 1);
331 /* On error, stop handling until the next kick. */
332 if (unlikely(headcount
< 0))
334 /* OK, now we need to know about added descriptors. */
336 if (unlikely(vhost_enable_notify(vq
))) {
337 /* They have slipped one in as we were
338 * doing that: check again. */
339 vhost_disable_notify(vq
);
342 /* Nothing new? Wait for eventfd to tell us
346 /* We don't need to be notified again. */
347 if (unlikely((vhost_hlen
)))
348 /* Skip header. TODO: support TSO. */
349 move_iovec_hdr(vq
->iov
, vq
->hdr
, vhost_hlen
, in
);
351 /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
352 * needed because recvmsg can modify msg_iov. */
353 copy_iovec_hdr(vq
->iov
, vq
->hdr
, sock_hlen
, in
);
355 err
= sock
->ops
->recvmsg(NULL
, sock
, &msg
,
356 sock_len
, MSG_DONTWAIT
| MSG_TRUNC
);
357 /* Userspace might have consumed the packet meanwhile:
358 * it's not supposed to do this usually, but might be hard
359 * to prevent. Discard data we got (if any) and keep going. */
360 if (unlikely(err
!= sock_len
)) {
361 pr_debug("Discarded rx packet: "
362 " len %d, expected %zd\n", err
, sock_len
);
363 vhost_discard_vq_desc(vq
, headcount
);
366 if (unlikely(vhost_hlen
) &&
367 memcpy_toiovecend(vq
->hdr
, (unsigned char *)&hdr
, 0,
369 vq_err(vq
, "Unable to write vnet_hdr at addr %p\n",
373 /* TODO: Should check and handle checksum. */
374 if (likely(mergeable
) &&
375 memcpy_toiovecend(vq
->hdr
, (unsigned char *)&headcount
,
376 offsetof(typeof(hdr
), num_buffers
),
377 sizeof hdr
.num_buffers
)) {
378 vq_err(vq
, "Failed num_buffers write");
379 vhost_discard_vq_desc(vq
, headcount
);
382 vhost_add_used_and_signal_n(&net
->dev
, vq
, vq
->heads
,
384 if (unlikely(vq_log
))
385 vhost_log_write(vq
, vq_log
, log
, vhost_len
);
386 total_len
+= vhost_len
;
387 if (unlikely(total_len
>= VHOST_NET_WEIGHT
)) {
388 vhost_poll_queue(&vq
->poll
);
393 mutex_unlock(&vq
->mutex
);
396 static void handle_tx_kick(struct vhost_work
*work
)
398 struct vhost_virtqueue
*vq
= container_of(work
, struct vhost_virtqueue
,
400 struct vhost_net
*net
= container_of(vq
->dev
, struct vhost_net
, dev
);
405 static void handle_rx_kick(struct vhost_work
*work
)
407 struct vhost_virtqueue
*vq
= container_of(work
, struct vhost_virtqueue
,
409 struct vhost_net
*net
= container_of(vq
->dev
, struct vhost_net
, dev
);
414 static void handle_tx_net(struct vhost_work
*work
)
416 struct vhost_net
*net
= container_of(work
, struct vhost_net
,
417 poll
[VHOST_NET_VQ_TX
].work
);
421 static void handle_rx_net(struct vhost_work
*work
)
423 struct vhost_net
*net
= container_of(work
, struct vhost_net
,
424 poll
[VHOST_NET_VQ_RX
].work
);
428 static int vhost_net_open(struct inode
*inode
, struct file
*f
)
430 struct vhost_net
*n
= kmalloc(sizeof *n
, GFP_KERNEL
);
431 struct vhost_dev
*dev
;
438 n
->vqs
[VHOST_NET_VQ_TX
].handle_kick
= handle_tx_kick
;
439 n
->vqs
[VHOST_NET_VQ_RX
].handle_kick
= handle_rx_kick
;
440 r
= vhost_dev_init(dev
, n
->vqs
, VHOST_NET_VQ_MAX
);
446 vhost_poll_init(n
->poll
+ VHOST_NET_VQ_TX
, handle_tx_net
, POLLOUT
, dev
);
447 vhost_poll_init(n
->poll
+ VHOST_NET_VQ_RX
, handle_rx_net
, POLLIN
, dev
);
448 n
->tx_poll_state
= VHOST_NET_POLL_DISABLED
;
455 static void vhost_net_disable_vq(struct vhost_net
*n
,
456 struct vhost_virtqueue
*vq
)
458 if (!vq
->private_data
)
460 if (vq
== n
->vqs
+ VHOST_NET_VQ_TX
) {
462 n
->tx_poll_state
= VHOST_NET_POLL_DISABLED
;
464 vhost_poll_stop(n
->poll
+ VHOST_NET_VQ_RX
);
467 static void vhost_net_enable_vq(struct vhost_net
*n
,
468 struct vhost_virtqueue
*vq
)
472 sock
= rcu_dereference_protected(vq
->private_data
,
473 lockdep_is_held(&vq
->mutex
));
476 if (vq
== n
->vqs
+ VHOST_NET_VQ_TX
) {
477 n
->tx_poll_state
= VHOST_NET_POLL_STOPPED
;
478 tx_poll_start(n
, sock
);
480 vhost_poll_start(n
->poll
+ VHOST_NET_VQ_RX
, sock
->file
);
483 static struct socket
*vhost_net_stop_vq(struct vhost_net
*n
,
484 struct vhost_virtqueue
*vq
)
488 mutex_lock(&vq
->mutex
);
489 sock
= rcu_dereference_protected(vq
->private_data
,
490 lockdep_is_held(&vq
->mutex
));
491 vhost_net_disable_vq(n
, vq
);
492 rcu_assign_pointer(vq
->private_data
, NULL
);
493 mutex_unlock(&vq
->mutex
);
497 static void vhost_net_stop(struct vhost_net
*n
, struct socket
**tx_sock
,
498 struct socket
**rx_sock
)
500 *tx_sock
= vhost_net_stop_vq(n
, n
->vqs
+ VHOST_NET_VQ_TX
);
501 *rx_sock
= vhost_net_stop_vq(n
, n
->vqs
+ VHOST_NET_VQ_RX
);
504 static void vhost_net_flush_vq(struct vhost_net
*n
, int index
)
506 vhost_poll_flush(n
->poll
+ index
);
507 vhost_poll_flush(&n
->dev
.vqs
[index
].poll
);
510 static void vhost_net_flush(struct vhost_net
*n
)
512 vhost_net_flush_vq(n
, VHOST_NET_VQ_TX
);
513 vhost_net_flush_vq(n
, VHOST_NET_VQ_RX
);
516 static int vhost_net_release(struct inode
*inode
, struct file
*f
)
518 struct vhost_net
*n
= f
->private_data
;
519 struct socket
*tx_sock
;
520 struct socket
*rx_sock
;
522 vhost_net_stop(n
, &tx_sock
, &rx_sock
);
524 vhost_dev_cleanup(&n
->dev
);
529 /* We do an extra flush before freeing memory,
530 * since jobs can re-queue themselves. */
536 static struct socket
*get_raw_socket(int fd
)
539 struct sockaddr_ll sa
;
540 char buf
[MAX_ADDR_LEN
];
542 int uaddr_len
= sizeof uaddr
, r
;
543 struct socket
*sock
= sockfd_lookup(fd
, &r
);
546 return ERR_PTR(-ENOTSOCK
);
548 /* Parameter checking */
549 if (sock
->sk
->sk_type
!= SOCK_RAW
) {
550 r
= -ESOCKTNOSUPPORT
;
554 r
= sock
->ops
->getname(sock
, (struct sockaddr
*)&uaddr
.sa
,
559 if (uaddr
.sa
.sll_family
!= AF_PACKET
) {
569 static struct socket
*get_tap_socket(int fd
)
571 struct file
*file
= fget(fd
);
575 return ERR_PTR(-EBADF
);
576 sock
= tun_get_socket(file
);
579 sock
= macvtap_get_socket(file
);
585 static struct socket
*get_socket(int fd
)
589 /* special case to disable backend */
592 sock
= get_raw_socket(fd
);
595 sock
= get_tap_socket(fd
);
598 return ERR_PTR(-ENOTSOCK
);
601 static long vhost_net_set_backend(struct vhost_net
*n
, unsigned index
, int fd
)
603 struct socket
*sock
, *oldsock
;
604 struct vhost_virtqueue
*vq
;
607 mutex_lock(&n
->dev
.mutex
);
608 r
= vhost_dev_check_owner(&n
->dev
);
612 if (index
>= VHOST_NET_VQ_MAX
) {
617 mutex_lock(&vq
->mutex
);
619 /* Verify that ring has been setup correctly. */
620 if (!vhost_vq_access_ok(vq
)) {
624 sock
= get_socket(fd
);
630 /* start polling new socket */
631 oldsock
= rcu_dereference_protected(vq
->private_data
,
632 lockdep_is_held(&vq
->mutex
));
633 if (sock
!= oldsock
) {
634 vhost_net_disable_vq(n
, vq
);
635 rcu_assign_pointer(vq
->private_data
, sock
);
636 vhost_net_enable_vq(n
, vq
);
639 mutex_unlock(&vq
->mutex
);
642 vhost_net_flush_vq(n
, index
);
646 mutex_unlock(&n
->dev
.mutex
);
650 mutex_unlock(&vq
->mutex
);
652 mutex_unlock(&n
->dev
.mutex
);
656 static long vhost_net_reset_owner(struct vhost_net
*n
)
658 struct socket
*tx_sock
= NULL
;
659 struct socket
*rx_sock
= NULL
;
662 mutex_lock(&n
->dev
.mutex
);
663 err
= vhost_dev_check_owner(&n
->dev
);
666 vhost_net_stop(n
, &tx_sock
, &rx_sock
);
668 err
= vhost_dev_reset_owner(&n
->dev
);
670 mutex_unlock(&n
->dev
.mutex
);
678 static int vhost_net_set_features(struct vhost_net
*n
, u64 features
)
680 size_t vhost_hlen
, sock_hlen
, hdr_len
;
683 hdr_len
= (features
& (1 << VIRTIO_NET_F_MRG_RXBUF
)) ?
684 sizeof(struct virtio_net_hdr_mrg_rxbuf
) :
685 sizeof(struct virtio_net_hdr
);
686 if (features
& (1 << VHOST_NET_F_VIRTIO_NET_HDR
)) {
687 /* vhost provides vnet_hdr */
688 vhost_hlen
= hdr_len
;
691 /* socket provides vnet_hdr */
695 mutex_lock(&n
->dev
.mutex
);
696 if ((features
& (1 << VHOST_F_LOG_ALL
)) &&
697 !vhost_log_access_ok(&n
->dev
)) {
698 mutex_unlock(&n
->dev
.mutex
);
701 n
->dev
.acked_features
= features
;
703 for (i
= 0; i
< VHOST_NET_VQ_MAX
; ++i
) {
704 mutex_lock(&n
->vqs
[i
].mutex
);
705 n
->vqs
[i
].vhost_hlen
= vhost_hlen
;
706 n
->vqs
[i
].sock_hlen
= sock_hlen
;
707 mutex_unlock(&n
->vqs
[i
].mutex
);
710 mutex_unlock(&n
->dev
.mutex
);
714 static long vhost_net_ioctl(struct file
*f
, unsigned int ioctl
,
717 struct vhost_net
*n
= f
->private_data
;
718 void __user
*argp
= (void __user
*)arg
;
719 u64 __user
*featurep
= argp
;
720 struct vhost_vring_file backend
;
725 case VHOST_NET_SET_BACKEND
:
726 if (copy_from_user(&backend
, argp
, sizeof backend
))
728 return vhost_net_set_backend(n
, backend
.index
, backend
.fd
);
729 case VHOST_GET_FEATURES
:
730 features
= VHOST_FEATURES
;
731 if (copy_to_user(featurep
, &features
, sizeof features
))
734 case VHOST_SET_FEATURES
:
735 if (copy_from_user(&features
, featurep
, sizeof features
))
737 if (features
& ~VHOST_FEATURES
)
739 return vhost_net_set_features(n
, features
);
740 case VHOST_RESET_OWNER
:
741 return vhost_net_reset_owner(n
);
743 mutex_lock(&n
->dev
.mutex
);
744 r
= vhost_dev_ioctl(&n
->dev
, ioctl
, arg
);
746 mutex_unlock(&n
->dev
.mutex
);
752 static long vhost_net_compat_ioctl(struct file
*f
, unsigned int ioctl
,
755 return vhost_net_ioctl(f
, ioctl
, (unsigned long)compat_ptr(arg
));
759 static const struct file_operations vhost_net_fops
= {
760 .owner
= THIS_MODULE
,
761 .release
= vhost_net_release
,
762 .unlocked_ioctl
= vhost_net_ioctl
,
764 .compat_ioctl
= vhost_net_compat_ioctl
,
766 .open
= vhost_net_open
,
767 .llseek
= noop_llseek
,
770 static struct miscdevice vhost_net_misc
= {
776 static int vhost_net_init(void)
778 return misc_register(&vhost_net_misc
);
780 module_init(vhost_net_init
);
782 static void vhost_net_exit(void)
784 misc_deregister(&vhost_net_misc
);
786 module_exit(vhost_net_exit
);
788 MODULE_VERSION("0.0.1");
789 MODULE_LICENSE("GPL v2");
790 MODULE_AUTHOR("Michael S. Tsirkin");
791 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");