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/moduleparam.h>
16 #include <linux/mutex.h>
17 #include <linux/workqueue.h>
18 #include <linux/rcupdate.h>
19 #include <linux/file.h>
20 #include <linux/slab.h>
22 #include <linux/net.h>
23 #include <linux/if_packet.h>
24 #include <linux/if_arp.h>
25 #include <linux/if_tun.h>
26 #include <linux/if_macvlan.h>
27 #include <linux/if_vlan.h>
33 static int experimental_zcopytx
;
34 module_param(experimental_zcopytx
, int, 0444);
35 MODULE_PARM_DESC(experimental_zcopytx
, "Enable Experimental Zero Copy TX");
37 /* Max number of bytes transferred before requeueing the job.
38 * Using this limit prevents one virtqueue from starving others. */
39 #define VHOST_NET_WEIGHT 0x80000
41 /* MAX number of TX used buffers for outstanding zerocopy */
42 #define VHOST_MAX_PEND 128
43 #define VHOST_GOODCOPY_LEN 256
46 * For transmit, used buffer len is unused; we override it to track buffer
47 * status internally; used for zerocopy tx only.
49 /* Lower device DMA failed */
50 #define VHOST_DMA_FAILED_LEN 3
51 /* Lower device DMA done */
52 #define VHOST_DMA_DONE_LEN 2
53 /* Lower device DMA in progress */
54 #define VHOST_DMA_IN_PROGRESS 1
56 #define VHOST_DMA_CLEAR_LEN 0
58 #define VHOST_DMA_IS_DONE(len) ((len) >= VHOST_DMA_DONE_LEN)
66 enum vhost_net_poll_state
{
67 VHOST_NET_POLL_DISABLED
= 0,
68 VHOST_NET_POLL_STARTED
= 1,
69 VHOST_NET_POLL_STOPPED
= 2,
74 struct vhost_virtqueue vqs
[VHOST_NET_VQ_MAX
];
75 struct vhost_poll poll
[VHOST_NET_VQ_MAX
];
76 /* Tells us whether we are polling a socket for TX.
77 * We only do this when socket buffer fills up.
78 * Protected by tx vq lock. */
79 enum vhost_net_poll_state tx_poll_state
;
80 /* Number of TX recently submitted.
81 * Protected by tx vq lock. */
83 /* Number of times zerocopy TX recently failed.
84 * Protected by tx vq lock. */
85 unsigned tx_zcopy_err
;
88 static void vhost_net_tx_packet(struct vhost_net
*net
)
91 if (net
->tx_packets
< 1024)
94 net
->tx_zcopy_err
= 0;
97 static void vhost_net_tx_err(struct vhost_net
*net
)
102 static bool vhost_net_tx_select_zcopy(struct vhost_net
*net
)
104 return net
->tx_packets
/ 64 >= net
->tx_zcopy_err
;
107 static bool vhost_sock_zcopy(struct socket
*sock
)
109 return unlikely(experimental_zcopytx
) &&
110 sock_flag(sock
->sk
, SOCK_ZEROCOPY
);
113 /* Pop first len bytes from iovec. Return number of segments used. */
114 static int move_iovec_hdr(struct iovec
*from
, struct iovec
*to
,
115 size_t len
, int iov_count
)
120 while (len
&& seg
< iov_count
) {
121 size
= min(from
->iov_len
, len
);
122 to
->iov_base
= from
->iov_base
;
124 from
->iov_len
-= size
;
125 from
->iov_base
+= size
;
133 /* Copy iovec entries for len bytes from iovec. */
134 static void copy_iovec_hdr(const struct iovec
*from
, struct iovec
*to
,
135 size_t len
, int iovcount
)
140 while (len
&& seg
< iovcount
) {
141 size
= min(from
->iov_len
, len
);
142 to
->iov_base
= from
->iov_base
;
151 /* Caller must have TX VQ lock */
152 static void tx_poll_stop(struct vhost_net
*net
)
154 if (likely(net
->tx_poll_state
!= VHOST_NET_POLL_STARTED
))
156 vhost_poll_stop(net
->poll
+ VHOST_NET_VQ_TX
);
157 net
->tx_poll_state
= VHOST_NET_POLL_STOPPED
;
160 /* Caller must have TX VQ lock */
161 static void tx_poll_start(struct vhost_net
*net
, struct socket
*sock
)
163 if (unlikely(net
->tx_poll_state
!= VHOST_NET_POLL_STOPPED
))
165 vhost_poll_start(net
->poll
+ VHOST_NET_VQ_TX
, sock
->file
);
166 net
->tx_poll_state
= VHOST_NET_POLL_STARTED
;
169 /* In case of DMA done not in order in lower device driver for some reason.
170 * upend_idx is used to track end of used idx, done_idx is used to track head
171 * of used idx. Once lower device DMA done contiguously, we will signal KVM
174 static int vhost_zerocopy_signal_used(struct vhost_net
*net
,
175 struct vhost_virtqueue
*vq
)
180 for (i
= vq
->done_idx
; i
!= vq
->upend_idx
; i
= (i
+ 1) % UIO_MAXIOV
) {
181 if (vq
->heads
[i
].len
== VHOST_DMA_FAILED_LEN
)
182 vhost_net_tx_err(net
);
183 if (VHOST_DMA_IS_DONE(vq
->heads
[i
].len
)) {
184 vq
->heads
[i
].len
= VHOST_DMA_CLEAR_LEN
;
185 vhost_add_used_and_signal(vq
->dev
, vq
,
196 static void vhost_zerocopy_callback(struct ubuf_info
*ubuf
, bool success
)
198 struct vhost_ubuf_ref
*ubufs
= ubuf
->ctx
;
199 struct vhost_virtqueue
*vq
= ubufs
->vq
;
200 int cnt
= atomic_read(&ubufs
->kref
.refcount
);
203 * Trigger polling thread if guest stopped submitting new buffers:
204 * in this case, the refcount after decrement will eventually reach 1
206 * We also trigger polling periodically after each 16 packets
207 * (the value 16 here is more or less arbitrary, it's tuned to trigger
208 * less than 10% of times).
210 if (cnt
<= 2 || !(cnt
% 16))
211 vhost_poll_queue(&vq
->poll
);
212 /* set len to mark this desc buffers done DMA */
213 vq
->heads
[ubuf
->desc
].len
= success
?
214 VHOST_DMA_DONE_LEN
: VHOST_DMA_FAILED_LEN
;
215 vhost_ubuf_put(ubufs
);
218 /* Expects to be always run from workqueue - which acts as
219 * read-size critical section for our kind of RCU. */
220 static void handle_tx(struct vhost_net
*net
)
222 struct vhost_virtqueue
*vq
= &net
->dev
.vqs
[VHOST_NET_VQ_TX
];
225 struct msghdr msg
= {
231 .msg_flags
= MSG_DONTWAIT
,
233 size_t len
, total_len
= 0;
237 struct vhost_ubuf_ref
*uninitialized_var(ubufs
);
240 /* TODO: check that we are running from vhost_worker? */
241 sock
= rcu_dereference_check(vq
->private_data
, 1);
245 wmem
= atomic_read(&sock
->sk
->sk_wmem_alloc
);
246 if (wmem
>= sock
->sk
->sk_sndbuf
) {
247 mutex_lock(&vq
->mutex
);
248 tx_poll_start(net
, sock
);
249 mutex_unlock(&vq
->mutex
);
253 mutex_lock(&vq
->mutex
);
254 vhost_disable_notify(&net
->dev
, vq
);
256 if (wmem
< sock
->sk
->sk_sndbuf
/ 2)
258 hdr_size
= vq
->vhost_hlen
;
262 /* Release DMAs done buffers first */
264 vhost_zerocopy_signal_used(net
, vq
);
266 head
= vhost_get_vq_desc(&net
->dev
, vq
, vq
->iov
,
270 /* On error, stop handling until the next kick. */
271 if (unlikely(head
< 0))
273 /* Nothing new? Wait for eventfd to tell us they refilled. */
274 if (head
== vq
->num
) {
277 wmem
= atomic_read(&sock
->sk
->sk_wmem_alloc
);
278 if (wmem
>= sock
->sk
->sk_sndbuf
* 3 / 4) {
279 tx_poll_start(net
, sock
);
280 set_bit(SOCK_ASYNC_NOSPACE
, &sock
->flags
);
283 /* If more outstanding DMAs, queue the work.
284 * Handle upend_idx wrap around
286 num_pends
= likely(vq
->upend_idx
>= vq
->done_idx
) ?
287 (vq
->upend_idx
- vq
->done_idx
) :
288 (vq
->upend_idx
+ UIO_MAXIOV
- vq
->done_idx
);
289 if (unlikely(num_pends
> VHOST_MAX_PEND
)) {
290 tx_poll_start(net
, sock
);
291 set_bit(SOCK_ASYNC_NOSPACE
, &sock
->flags
);
294 if (unlikely(vhost_enable_notify(&net
->dev
, vq
))) {
295 vhost_disable_notify(&net
->dev
, vq
);
301 vq_err(vq
, "Unexpected descriptor format for TX: "
302 "out %d, int %d\n", out
, in
);
305 /* Skip header. TODO: support TSO. */
306 s
= move_iovec_hdr(vq
->iov
, vq
->hdr
, hdr_size
, out
);
307 msg
.msg_iovlen
= out
;
308 len
= iov_length(vq
->iov
, out
);
311 vq_err(vq
, "Unexpected header len for TX: "
312 "%zd expected %zd\n",
313 iov_length(vq
->hdr
, s
), hdr_size
);
316 /* use msg_control to pass vhost zerocopy ubuf info to skb */
318 vq
->heads
[vq
->upend_idx
].id
= head
;
319 if (!vhost_net_tx_select_zcopy(net
) ||
320 len
< VHOST_GOODCOPY_LEN
) {
321 /* copy don't need to wait for DMA done */
322 vq
->heads
[vq
->upend_idx
].len
=
324 msg
.msg_control
= NULL
;
325 msg
.msg_controllen
= 0;
328 struct ubuf_info
*ubuf
= &vq
->ubuf_info
[head
];
330 vq
->heads
[vq
->upend_idx
].len
=
331 VHOST_DMA_IN_PROGRESS
;
332 ubuf
->callback
= vhost_zerocopy_callback
;
333 ubuf
->ctx
= vq
->ubufs
;
334 ubuf
->desc
= vq
->upend_idx
;
335 msg
.msg_control
= ubuf
;
336 msg
.msg_controllen
= sizeof(ubuf
);
338 kref_get(&ubufs
->kref
);
340 vq
->upend_idx
= (vq
->upend_idx
+ 1) % UIO_MAXIOV
;
342 /* TODO: Check specific error and bomb out unless ENOBUFS? */
343 err
= sock
->ops
->sendmsg(NULL
, sock
, &msg
, len
);
344 if (unlikely(err
< 0)) {
347 vhost_ubuf_put(ubufs
);
348 vq
->upend_idx
= ((unsigned)vq
->upend_idx
- 1) %
351 vhost_discard_vq_desc(vq
, 1);
352 if (err
== -EAGAIN
|| err
== -ENOBUFS
)
353 tx_poll_start(net
, sock
);
357 pr_debug("Truncated TX packet: "
358 " len %d != %zd\n", err
, len
);
360 vhost_add_used_and_signal(&net
->dev
, vq
, head
, 0);
362 vhost_zerocopy_signal_used(net
, vq
);
364 vhost_net_tx_packet(net
);
365 if (unlikely(total_len
>= VHOST_NET_WEIGHT
)) {
366 vhost_poll_queue(&vq
->poll
);
371 mutex_unlock(&vq
->mutex
);
374 static int peek_head_len(struct sock
*sk
)
376 struct sk_buff
*head
;
380 spin_lock_irqsave(&sk
->sk_receive_queue
.lock
, flags
);
381 head
= skb_peek(&sk
->sk_receive_queue
);
384 if (vlan_tx_tag_present(head
))
388 spin_unlock_irqrestore(&sk
->sk_receive_queue
.lock
, flags
);
392 /* This is a multi-buffer version of vhost_get_desc, that works if
393 * vq has read descriptors only.
394 * @vq - the relevant virtqueue
395 * @datalen - data length we'll be reading
396 * @iovcount - returned count of io vectors we fill
398 * @log_num - log offset
399 * @quota - headcount quota, 1 for big buffer
400 * returns number of buffer heads allocated, negative on error
402 static int get_rx_bufs(struct vhost_virtqueue
*vq
,
403 struct vring_used_elem
*heads
,
406 struct vhost_log
*log
,
410 unsigned int out
, in
;
416 while (datalen
> 0 && headcount
< quota
) {
417 if (unlikely(seg
>= UIO_MAXIOV
)) {
421 d
= vhost_get_vq_desc(vq
->dev
, vq
, vq
->iov
+ seg
,
422 ARRAY_SIZE(vq
->iov
) - seg
, &out
,
428 if (unlikely(out
|| in
<= 0)) {
429 vq_err(vq
, "unexpected descriptor format for RX: "
430 "out %d, in %d\n", out
, in
);
438 heads
[headcount
].id
= d
;
439 heads
[headcount
].len
= iov_length(vq
->iov
+ seg
, in
);
440 datalen
-= heads
[headcount
].len
;
444 heads
[headcount
- 1].len
+= datalen
;
450 vhost_discard_vq_desc(vq
, headcount
);
454 /* Expects to be always run from workqueue - which acts as
455 * read-size critical section for our kind of RCU. */
456 static void handle_rx(struct vhost_net
*net
)
458 struct vhost_virtqueue
*vq
= &net
->dev
.vqs
[VHOST_NET_VQ_RX
];
459 unsigned uninitialized_var(in
), log
;
460 struct vhost_log
*vq_log
;
461 struct msghdr msg
= {
464 .msg_control
= NULL
, /* FIXME: get and handle RX aux data. */
467 .msg_flags
= MSG_DONTWAIT
,
469 struct virtio_net_hdr_mrg_rxbuf hdr
= {
471 .hdr
.gso_type
= VIRTIO_NET_HDR_GSO_NONE
473 size_t total_len
= 0;
474 int err
, headcount
, mergeable
;
475 size_t vhost_hlen
, sock_hlen
;
476 size_t vhost_len
, sock_len
;
477 /* TODO: check that we are running from vhost_worker? */
478 struct socket
*sock
= rcu_dereference_check(vq
->private_data
, 1);
483 mutex_lock(&vq
->mutex
);
484 vhost_disable_notify(&net
->dev
, vq
);
485 vhost_hlen
= vq
->vhost_hlen
;
486 sock_hlen
= vq
->sock_hlen
;
488 vq_log
= unlikely(vhost_has_feature(&net
->dev
, VHOST_F_LOG_ALL
)) ?
490 mergeable
= vhost_has_feature(&net
->dev
, VIRTIO_NET_F_MRG_RXBUF
);
492 while ((sock_len
= peek_head_len(sock
->sk
))) {
493 sock_len
+= sock_hlen
;
494 vhost_len
= sock_len
+ vhost_hlen
;
495 headcount
= get_rx_bufs(vq
, vq
->heads
, vhost_len
,
497 likely(mergeable
) ? UIO_MAXIOV
: 1);
498 /* On error, stop handling until the next kick. */
499 if (unlikely(headcount
< 0))
501 /* OK, now we need to know about added descriptors. */
503 if (unlikely(vhost_enable_notify(&net
->dev
, vq
))) {
504 /* They have slipped one in as we were
505 * doing that: check again. */
506 vhost_disable_notify(&net
->dev
, vq
);
509 /* Nothing new? Wait for eventfd to tell us
513 /* We don't need to be notified again. */
514 if (unlikely((vhost_hlen
)))
515 /* Skip header. TODO: support TSO. */
516 move_iovec_hdr(vq
->iov
, vq
->hdr
, vhost_hlen
, in
);
518 /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
519 * needed because recvmsg can modify msg_iov. */
520 copy_iovec_hdr(vq
->iov
, vq
->hdr
, sock_hlen
, in
);
522 err
= sock
->ops
->recvmsg(NULL
, sock
, &msg
,
523 sock_len
, MSG_DONTWAIT
| MSG_TRUNC
);
524 /* Userspace might have consumed the packet meanwhile:
525 * it's not supposed to do this usually, but might be hard
526 * to prevent. Discard data we got (if any) and keep going. */
527 if (unlikely(err
!= sock_len
)) {
528 pr_debug("Discarded rx packet: "
529 " len %d, expected %zd\n", err
, sock_len
);
530 vhost_discard_vq_desc(vq
, headcount
);
533 if (unlikely(vhost_hlen
) &&
534 memcpy_toiovecend(vq
->hdr
, (unsigned char *)&hdr
, 0,
536 vq_err(vq
, "Unable to write vnet_hdr at addr %p\n",
540 /* TODO: Should check and handle checksum. */
541 if (likely(mergeable
) &&
542 memcpy_toiovecend(vq
->hdr
, (unsigned char *)&headcount
,
543 offsetof(typeof(hdr
), num_buffers
),
544 sizeof hdr
.num_buffers
)) {
545 vq_err(vq
, "Failed num_buffers write");
546 vhost_discard_vq_desc(vq
, headcount
);
549 vhost_add_used_and_signal_n(&net
->dev
, vq
, vq
->heads
,
551 if (unlikely(vq_log
))
552 vhost_log_write(vq
, vq_log
, log
, vhost_len
);
553 total_len
+= vhost_len
;
554 if (unlikely(total_len
>= VHOST_NET_WEIGHT
)) {
555 vhost_poll_queue(&vq
->poll
);
560 mutex_unlock(&vq
->mutex
);
563 static void handle_tx_kick(struct vhost_work
*work
)
565 struct vhost_virtqueue
*vq
= container_of(work
, struct vhost_virtqueue
,
567 struct vhost_net
*net
= container_of(vq
->dev
, struct vhost_net
, dev
);
572 static void handle_rx_kick(struct vhost_work
*work
)
574 struct vhost_virtqueue
*vq
= container_of(work
, struct vhost_virtqueue
,
576 struct vhost_net
*net
= container_of(vq
->dev
, struct vhost_net
, dev
);
581 static void handle_tx_net(struct vhost_work
*work
)
583 struct vhost_net
*net
= container_of(work
, struct vhost_net
,
584 poll
[VHOST_NET_VQ_TX
].work
);
588 static void handle_rx_net(struct vhost_work
*work
)
590 struct vhost_net
*net
= container_of(work
, struct vhost_net
,
591 poll
[VHOST_NET_VQ_RX
].work
);
595 static int vhost_net_open(struct inode
*inode
, struct file
*f
)
597 struct vhost_net
*n
= kmalloc(sizeof *n
, GFP_KERNEL
);
598 struct vhost_dev
*dev
;
605 n
->vqs
[VHOST_NET_VQ_TX
].handle_kick
= handle_tx_kick
;
606 n
->vqs
[VHOST_NET_VQ_RX
].handle_kick
= handle_rx_kick
;
607 r
= vhost_dev_init(dev
, n
->vqs
, VHOST_NET_VQ_MAX
);
613 vhost_poll_init(n
->poll
+ VHOST_NET_VQ_TX
, handle_tx_net
, POLLOUT
, dev
);
614 vhost_poll_init(n
->poll
+ VHOST_NET_VQ_RX
, handle_rx_net
, POLLIN
, dev
);
615 n
->tx_poll_state
= VHOST_NET_POLL_DISABLED
;
622 static void vhost_net_disable_vq(struct vhost_net
*n
,
623 struct vhost_virtqueue
*vq
)
625 if (!vq
->private_data
)
627 if (vq
== n
->vqs
+ VHOST_NET_VQ_TX
) {
629 n
->tx_poll_state
= VHOST_NET_POLL_DISABLED
;
631 vhost_poll_stop(n
->poll
+ VHOST_NET_VQ_RX
);
634 static void vhost_net_enable_vq(struct vhost_net
*n
,
635 struct vhost_virtqueue
*vq
)
639 sock
= rcu_dereference_protected(vq
->private_data
,
640 lockdep_is_held(&vq
->mutex
));
643 if (vq
== n
->vqs
+ VHOST_NET_VQ_TX
) {
644 n
->tx_poll_state
= VHOST_NET_POLL_STOPPED
;
645 tx_poll_start(n
, sock
);
647 vhost_poll_start(n
->poll
+ VHOST_NET_VQ_RX
, sock
->file
);
650 static struct socket
*vhost_net_stop_vq(struct vhost_net
*n
,
651 struct vhost_virtqueue
*vq
)
655 mutex_lock(&vq
->mutex
);
656 sock
= rcu_dereference_protected(vq
->private_data
,
657 lockdep_is_held(&vq
->mutex
));
658 vhost_net_disable_vq(n
, vq
);
659 rcu_assign_pointer(vq
->private_data
, NULL
);
660 mutex_unlock(&vq
->mutex
);
664 static void vhost_net_stop(struct vhost_net
*n
, struct socket
**tx_sock
,
665 struct socket
**rx_sock
)
667 *tx_sock
= vhost_net_stop_vq(n
, n
->vqs
+ VHOST_NET_VQ_TX
);
668 *rx_sock
= vhost_net_stop_vq(n
, n
->vqs
+ VHOST_NET_VQ_RX
);
671 static void vhost_net_flush_vq(struct vhost_net
*n
, int index
)
673 vhost_poll_flush(n
->poll
+ index
);
674 vhost_poll_flush(&n
->dev
.vqs
[index
].poll
);
677 static void vhost_net_flush(struct vhost_net
*n
)
679 vhost_net_flush_vq(n
, VHOST_NET_VQ_TX
);
680 vhost_net_flush_vq(n
, VHOST_NET_VQ_RX
);
683 static int vhost_net_release(struct inode
*inode
, struct file
*f
)
685 struct vhost_net
*n
= f
->private_data
;
686 struct socket
*tx_sock
;
687 struct socket
*rx_sock
;
690 vhost_net_stop(n
, &tx_sock
, &rx_sock
);
692 vhost_dev_stop(&n
->dev
);
693 for (i
= 0; i
< n
->dev
.nvqs
; ++i
) {
694 /* Wait for all lower device DMAs done. */
695 if (n
->dev
.vqs
[i
].ubufs
)
696 vhost_ubuf_put_and_wait(n
->dev
.vqs
[i
].ubufs
);
698 vhost_zerocopy_signal_used(n
, &n
->dev
.vqs
[i
]);
700 vhost_dev_cleanup(&n
->dev
, false);
705 /* We do an extra flush before freeing memory,
706 * since jobs can re-queue themselves. */
712 static struct socket
*get_raw_socket(int fd
)
715 struct sockaddr_ll sa
;
716 char buf
[MAX_ADDR_LEN
];
718 int uaddr_len
= sizeof uaddr
, r
;
719 struct socket
*sock
= sockfd_lookup(fd
, &r
);
722 return ERR_PTR(-ENOTSOCK
);
724 /* Parameter checking */
725 if (sock
->sk
->sk_type
!= SOCK_RAW
) {
726 r
= -ESOCKTNOSUPPORT
;
730 r
= sock
->ops
->getname(sock
, (struct sockaddr
*)&uaddr
.sa
,
735 if (uaddr
.sa
.sll_family
!= AF_PACKET
) {
745 static struct socket
*get_tap_socket(int fd
)
747 struct file
*file
= fget(fd
);
751 return ERR_PTR(-EBADF
);
752 sock
= tun_get_socket(file
);
755 sock
= macvtap_get_socket(file
);
761 static struct socket
*get_socket(int fd
)
765 /* special case to disable backend */
768 sock
= get_raw_socket(fd
);
771 sock
= get_tap_socket(fd
);
774 return ERR_PTR(-ENOTSOCK
);
777 static long vhost_net_set_backend(struct vhost_net
*n
, unsigned index
, int fd
)
779 struct socket
*sock
, *oldsock
;
780 struct vhost_virtqueue
*vq
;
781 struct vhost_ubuf_ref
*ubufs
, *oldubufs
= NULL
;
784 mutex_lock(&n
->dev
.mutex
);
785 r
= vhost_dev_check_owner(&n
->dev
);
789 if (index
>= VHOST_NET_VQ_MAX
) {
794 mutex_lock(&vq
->mutex
);
796 /* Verify that ring has been setup correctly. */
797 if (!vhost_vq_access_ok(vq
)) {
801 sock
= get_socket(fd
);
807 /* start polling new socket */
808 oldsock
= rcu_dereference_protected(vq
->private_data
,
809 lockdep_is_held(&vq
->mutex
));
810 if (sock
!= oldsock
) {
811 ubufs
= vhost_ubuf_alloc(vq
, sock
&& vhost_sock_zcopy(sock
));
816 oldubufs
= vq
->ubufs
;
818 vhost_net_disable_vq(n
, vq
);
819 rcu_assign_pointer(vq
->private_data
, sock
);
820 vhost_net_enable_vq(n
, vq
);
822 r
= vhost_init_used(vq
);
827 mutex_unlock(&vq
->mutex
);
830 vhost_ubuf_put_and_wait(oldubufs
);
831 mutex_lock(&vq
->mutex
);
832 vhost_zerocopy_signal_used(n
, vq
);
833 mutex_unlock(&vq
->mutex
);
837 vhost_net_flush_vq(n
, index
);
841 mutex_unlock(&n
->dev
.mutex
);
847 mutex_unlock(&vq
->mutex
);
849 mutex_unlock(&n
->dev
.mutex
);
853 static long vhost_net_reset_owner(struct vhost_net
*n
)
855 struct socket
*tx_sock
= NULL
;
856 struct socket
*rx_sock
= NULL
;
859 mutex_lock(&n
->dev
.mutex
);
860 err
= vhost_dev_check_owner(&n
->dev
);
863 vhost_net_stop(n
, &tx_sock
, &rx_sock
);
865 err
= vhost_dev_reset_owner(&n
->dev
);
867 mutex_unlock(&n
->dev
.mutex
);
875 static int vhost_net_set_features(struct vhost_net
*n
, u64 features
)
877 size_t vhost_hlen
, sock_hlen
, hdr_len
;
880 hdr_len
= (features
& (1 << VIRTIO_NET_F_MRG_RXBUF
)) ?
881 sizeof(struct virtio_net_hdr_mrg_rxbuf
) :
882 sizeof(struct virtio_net_hdr
);
883 if (features
& (1 << VHOST_NET_F_VIRTIO_NET_HDR
)) {
884 /* vhost provides vnet_hdr */
885 vhost_hlen
= hdr_len
;
888 /* socket provides vnet_hdr */
892 mutex_lock(&n
->dev
.mutex
);
893 if ((features
& (1 << VHOST_F_LOG_ALL
)) &&
894 !vhost_log_access_ok(&n
->dev
)) {
895 mutex_unlock(&n
->dev
.mutex
);
898 n
->dev
.acked_features
= features
;
900 for (i
= 0; i
< VHOST_NET_VQ_MAX
; ++i
) {
901 mutex_lock(&n
->vqs
[i
].mutex
);
902 n
->vqs
[i
].vhost_hlen
= vhost_hlen
;
903 n
->vqs
[i
].sock_hlen
= sock_hlen
;
904 mutex_unlock(&n
->vqs
[i
].mutex
);
907 mutex_unlock(&n
->dev
.mutex
);
911 static long vhost_net_ioctl(struct file
*f
, unsigned int ioctl
,
914 struct vhost_net
*n
= f
->private_data
;
915 void __user
*argp
= (void __user
*)arg
;
916 u64 __user
*featurep
= argp
;
917 struct vhost_vring_file backend
;
922 case VHOST_NET_SET_BACKEND
:
923 if (copy_from_user(&backend
, argp
, sizeof backend
))
925 return vhost_net_set_backend(n
, backend
.index
, backend
.fd
);
926 case VHOST_GET_FEATURES
:
927 features
= VHOST_NET_FEATURES
;
928 if (copy_to_user(featurep
, &features
, sizeof features
))
931 case VHOST_SET_FEATURES
:
932 if (copy_from_user(&features
, featurep
, sizeof features
))
934 if (features
& ~VHOST_NET_FEATURES
)
936 return vhost_net_set_features(n
, features
);
937 case VHOST_RESET_OWNER
:
938 return vhost_net_reset_owner(n
);
940 mutex_lock(&n
->dev
.mutex
);
941 r
= vhost_dev_ioctl(&n
->dev
, ioctl
, arg
);
943 mutex_unlock(&n
->dev
.mutex
);
949 static long vhost_net_compat_ioctl(struct file
*f
, unsigned int ioctl
,
952 return vhost_net_ioctl(f
, ioctl
, (unsigned long)compat_ptr(arg
));
956 static const struct file_operations vhost_net_fops
= {
957 .owner
= THIS_MODULE
,
958 .release
= vhost_net_release
,
959 .unlocked_ioctl
= vhost_net_ioctl
,
961 .compat_ioctl
= vhost_net_compat_ioctl
,
963 .open
= vhost_net_open
,
964 .llseek
= noop_llseek
,
967 static struct miscdevice vhost_net_misc
= {
968 .minor
= VHOST_NET_MINOR
,
970 .fops
= &vhost_net_fops
,
973 static int vhost_net_init(void)
975 if (experimental_zcopytx
)
976 vhost_enable_zcopy(VHOST_NET_VQ_TX
);
977 return misc_register(&vhost_net_misc
);
979 module_init(vhost_net_init
);
981 static void vhost_net_exit(void)
983 misc_deregister(&vhost_net_misc
);
985 module_exit(vhost_net_exit
);
987 MODULE_VERSION("0.0.1");
988 MODULE_LICENSE("GPL v2");
989 MODULE_AUTHOR("Michael S. Tsirkin");
990 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
991 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR
);
992 MODULE_ALIAS("devname:vhost-net");