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
= 1;
34 module_param(experimental_zcopytx
, int, 0444);
35 MODULE_PARM_DESC(experimental_zcopytx
, "Enable Zero Copy TX;"
36 " 1 -Enable; 0 - Disable");
38 /* Max number of bytes transferred before requeueing the job.
39 * Using this limit prevents one virtqueue from starving others. */
40 #define VHOST_NET_WEIGHT 0x80000
42 /* MAX number of TX used buffers for outstanding zerocopy */
43 #define VHOST_MAX_PEND 128
44 #define VHOST_GOODCOPY_LEN 256
47 * For transmit, used buffer len is unused; we override it to track buffer
48 * status internally; used for zerocopy tx only.
50 /* Lower device DMA failed */
51 #define VHOST_DMA_FAILED_LEN 3
52 /* Lower device DMA done */
53 #define VHOST_DMA_DONE_LEN 2
54 /* Lower device DMA in progress */
55 #define VHOST_DMA_IN_PROGRESS 1
57 #define VHOST_DMA_CLEAR_LEN 0
59 #define VHOST_DMA_IS_DONE(len) ((len) >= VHOST_DMA_DONE_LEN)
62 VHOST_NET_FEATURES
= VHOST_FEATURES
|
63 (1ULL << VHOST_NET_F_VIRTIO_NET_HDR
) |
64 (1ULL << VIRTIO_NET_F_MRG_RXBUF
),
73 struct vhost_net_ubuf_ref
{
75 wait_queue_head_t wait
;
76 struct vhost_virtqueue
*vq
;
79 struct vhost_net_virtqueue
{
80 struct vhost_virtqueue vq
;
81 /* hdr is used to store the virtio header.
82 * Since each iovec has >= 1 byte length, we never need more than
83 * header length entries to store the header. */
84 struct iovec hdr
[sizeof(struct virtio_net_hdr_mrg_rxbuf
)];
87 /* vhost zerocopy support fields below: */
88 /* last used idx for outstanding DMA zerocopy buffers */
90 /* first used idx for DMA done zerocopy buffers */
92 /* an array of userspace buffers info */
93 struct ubuf_info
*ubuf_info
;
94 /* Reference counting for outstanding ubufs.
95 * Protected by vq mutex. Writers must also take device mutex. */
96 struct vhost_net_ubuf_ref
*ubufs
;
100 struct vhost_dev dev
;
101 struct vhost_net_virtqueue vqs
[VHOST_NET_VQ_MAX
];
102 struct vhost_poll poll
[VHOST_NET_VQ_MAX
];
103 /* Number of TX recently submitted.
104 * Protected by tx vq lock. */
106 /* Number of times zerocopy TX recently failed.
107 * Protected by tx vq lock. */
108 unsigned tx_zcopy_err
;
109 /* Flush in progress. Protected by tx vq lock. */
113 static unsigned vhost_net_zcopy_mask __read_mostly
;
115 static void vhost_net_enable_zcopy(int vq
)
117 vhost_net_zcopy_mask
|= 0x1 << vq
;
120 static void vhost_net_zerocopy_done_signal(struct kref
*kref
)
122 struct vhost_net_ubuf_ref
*ubufs
;
124 ubufs
= container_of(kref
, struct vhost_net_ubuf_ref
, kref
);
125 wake_up(&ubufs
->wait
);
128 static struct vhost_net_ubuf_ref
*
129 vhost_net_ubuf_alloc(struct vhost_virtqueue
*vq
, bool zcopy
)
131 struct vhost_net_ubuf_ref
*ubufs
;
132 /* No zero copy backend? Nothing to count. */
135 ubufs
= kmalloc(sizeof(*ubufs
), GFP_KERNEL
);
137 return ERR_PTR(-ENOMEM
);
138 kref_init(&ubufs
->kref
);
139 init_waitqueue_head(&ubufs
->wait
);
144 static void vhost_net_ubuf_put(struct vhost_net_ubuf_ref
*ubufs
)
146 kref_put(&ubufs
->kref
, vhost_net_zerocopy_done_signal
);
149 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref
*ubufs
)
151 kref_put(&ubufs
->kref
, vhost_net_zerocopy_done_signal
);
152 wait_event(ubufs
->wait
, !atomic_read(&ubufs
->kref
.refcount
));
155 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref
*ubufs
)
157 vhost_net_ubuf_put_and_wait(ubufs
);
161 static void vhost_net_clear_ubuf_info(struct vhost_net
*n
)
165 for (i
= 0; i
< VHOST_NET_VQ_MAX
; ++i
) {
166 kfree(n
->vqs
[i
].ubuf_info
);
167 n
->vqs
[i
].ubuf_info
= NULL
;
171 static int vhost_net_set_ubuf_info(struct vhost_net
*n
)
176 for (i
= 0; i
< VHOST_NET_VQ_MAX
; ++i
) {
177 zcopy
= vhost_net_zcopy_mask
& (0x1 << i
);
180 n
->vqs
[i
].ubuf_info
= kmalloc(sizeof(*n
->vqs
[i
].ubuf_info
) *
181 UIO_MAXIOV
, GFP_KERNEL
);
182 if (!n
->vqs
[i
].ubuf_info
)
188 vhost_net_clear_ubuf_info(n
);
192 static void vhost_net_vq_reset(struct vhost_net
*n
)
196 vhost_net_clear_ubuf_info(n
);
198 for (i
= 0; i
< VHOST_NET_VQ_MAX
; i
++) {
199 n
->vqs
[i
].done_idx
= 0;
200 n
->vqs
[i
].upend_idx
= 0;
201 n
->vqs
[i
].ubufs
= NULL
;
202 n
->vqs
[i
].vhost_hlen
= 0;
203 n
->vqs
[i
].sock_hlen
= 0;
208 static void vhost_net_tx_packet(struct vhost_net
*net
)
211 if (net
->tx_packets
< 1024)
214 net
->tx_zcopy_err
= 0;
217 static void vhost_net_tx_err(struct vhost_net
*net
)
222 static bool vhost_net_tx_select_zcopy(struct vhost_net
*net
)
224 /* TX flush waits for outstanding DMAs to be done.
225 * Don't start new DMAs.
227 return !net
->tx_flush
&&
228 net
->tx_packets
/ 64 >= net
->tx_zcopy_err
;
231 static bool vhost_sock_zcopy(struct socket
*sock
)
233 return unlikely(experimental_zcopytx
) &&
234 sock_flag(sock
->sk
, SOCK_ZEROCOPY
);
237 /* Pop first len bytes from iovec. Return number of segments used. */
238 static int move_iovec_hdr(struct iovec
*from
, struct iovec
*to
,
239 size_t len
, int iov_count
)
244 while (len
&& seg
< iov_count
) {
245 size
= min(from
->iov_len
, len
);
246 to
->iov_base
= from
->iov_base
;
248 from
->iov_len
-= size
;
249 from
->iov_base
+= size
;
257 /* Copy iovec entries for len bytes from iovec. */
258 static void copy_iovec_hdr(const struct iovec
*from
, struct iovec
*to
,
259 size_t len
, int iovcount
)
264 while (len
&& seg
< iovcount
) {
265 size
= min(from
->iov_len
, len
);
266 to
->iov_base
= from
->iov_base
;
275 /* In case of DMA done not in order in lower device driver for some reason.
276 * upend_idx is used to track end of used idx, done_idx is used to track head
277 * of used idx. Once lower device DMA done contiguously, we will signal KVM
280 static int vhost_zerocopy_signal_used(struct vhost_net
*net
,
281 struct vhost_virtqueue
*vq
)
283 struct vhost_net_virtqueue
*nvq
=
284 container_of(vq
, struct vhost_net_virtqueue
, vq
);
288 for (i
= nvq
->done_idx
; i
!= nvq
->upend_idx
; i
= (i
+ 1) % UIO_MAXIOV
) {
289 if (vq
->heads
[i
].len
== VHOST_DMA_FAILED_LEN
)
290 vhost_net_tx_err(net
);
291 if (VHOST_DMA_IS_DONE(vq
->heads
[i
].len
)) {
292 vq
->heads
[i
].len
= VHOST_DMA_CLEAR_LEN
;
293 vhost_add_used_and_signal(vq
->dev
, vq
,
304 static void vhost_zerocopy_callback(struct ubuf_info
*ubuf
, bool success
)
306 struct vhost_net_ubuf_ref
*ubufs
= ubuf
->ctx
;
307 struct vhost_virtqueue
*vq
= ubufs
->vq
;
308 int cnt
= atomic_read(&ubufs
->kref
.refcount
);
311 * Trigger polling thread if guest stopped submitting new buffers:
312 * in this case, the refcount after decrement will eventually reach 1
314 * We also trigger polling periodically after each 16 packets
315 * (the value 16 here is more or less arbitrary, it's tuned to trigger
316 * less than 10% of times).
318 if (cnt
<= 2 || !(cnt
% 16))
319 vhost_poll_queue(&vq
->poll
);
320 /* set len to mark this desc buffers done DMA */
321 vq
->heads
[ubuf
->desc
].len
= success
?
322 VHOST_DMA_DONE_LEN
: VHOST_DMA_FAILED_LEN
;
323 vhost_net_ubuf_put(ubufs
);
326 /* Expects to be always run from workqueue - which acts as
327 * read-size critical section for our kind of RCU. */
328 static void handle_tx(struct vhost_net
*net
)
330 struct vhost_net_virtqueue
*nvq
= &net
->vqs
[VHOST_NET_VQ_TX
];
331 struct vhost_virtqueue
*vq
= &nvq
->vq
;
334 struct msghdr msg
= {
340 .msg_flags
= MSG_DONTWAIT
,
342 size_t len
, total_len
= 0;
346 struct vhost_net_ubuf_ref
*uninitialized_var(ubufs
);
347 bool zcopy
, zcopy_used
;
349 mutex_lock(&vq
->mutex
);
350 sock
= vq
->private_data
;
354 vhost_disable_notify(&net
->dev
, vq
);
356 hdr_size
= nvq
->vhost_hlen
;
360 /* Release DMAs done buffers first */
362 vhost_zerocopy_signal_used(net
, vq
);
364 head
= vhost_get_vq_desc(&net
->dev
, vq
, vq
->iov
,
368 /* On error, stop handling until the next kick. */
369 if (unlikely(head
< 0))
371 /* Nothing new? Wait for eventfd to tell us they refilled. */
372 if (head
== vq
->num
) {
375 /* If more outstanding DMAs, queue the work.
376 * Handle upend_idx wrap around
378 num_pends
= likely(nvq
->upend_idx
>= nvq
->done_idx
) ?
379 (nvq
->upend_idx
- nvq
->done_idx
) :
380 (nvq
->upend_idx
+ UIO_MAXIOV
-
382 if (unlikely(num_pends
> VHOST_MAX_PEND
))
384 if (unlikely(vhost_enable_notify(&net
->dev
, vq
))) {
385 vhost_disable_notify(&net
->dev
, vq
);
391 vq_err(vq
, "Unexpected descriptor format for TX: "
392 "out %d, int %d\n", out
, in
);
395 /* Skip header. TODO: support TSO. */
396 s
= move_iovec_hdr(vq
->iov
, nvq
->hdr
, hdr_size
, out
);
397 msg
.msg_iovlen
= out
;
398 len
= iov_length(vq
->iov
, out
);
401 vq_err(vq
, "Unexpected header len for TX: "
402 "%zd expected %zd\n",
403 iov_length(nvq
->hdr
, s
), hdr_size
);
406 zcopy_used
= zcopy
&& (len
>= VHOST_GOODCOPY_LEN
||
407 nvq
->upend_idx
!= nvq
->done_idx
);
409 /* use msg_control to pass vhost zerocopy ubuf info to skb */
411 vq
->heads
[nvq
->upend_idx
].id
= head
;
412 if (!vhost_net_tx_select_zcopy(net
) ||
413 len
< VHOST_GOODCOPY_LEN
) {
414 /* copy don't need to wait for DMA done */
415 vq
->heads
[nvq
->upend_idx
].len
=
417 msg
.msg_control
= NULL
;
418 msg
.msg_controllen
= 0;
421 struct ubuf_info
*ubuf
;
422 ubuf
= nvq
->ubuf_info
+ nvq
->upend_idx
;
424 vq
->heads
[nvq
->upend_idx
].len
=
425 VHOST_DMA_IN_PROGRESS
;
426 ubuf
->callback
= vhost_zerocopy_callback
;
427 ubuf
->ctx
= nvq
->ubufs
;
428 ubuf
->desc
= nvq
->upend_idx
;
429 msg
.msg_control
= ubuf
;
430 msg
.msg_controllen
= sizeof(ubuf
);
432 kref_get(&ubufs
->kref
);
434 nvq
->upend_idx
= (nvq
->upend_idx
+ 1) % UIO_MAXIOV
;
436 msg
.msg_control
= NULL
;
437 /* TODO: Check specific error and bomb out unless ENOBUFS? */
438 err
= sock
->ops
->sendmsg(NULL
, sock
, &msg
, len
);
439 if (unlikely(err
< 0)) {
442 vhost_net_ubuf_put(ubufs
);
443 nvq
->upend_idx
= ((unsigned)nvq
->upend_idx
- 1)
446 vhost_discard_vq_desc(vq
, 1);
450 pr_debug("Truncated TX packet: "
451 " len %d != %zd\n", err
, len
);
453 vhost_add_used_and_signal(&net
->dev
, vq
, head
, 0);
455 vhost_zerocopy_signal_used(net
, vq
);
457 vhost_net_tx_packet(net
);
458 if (unlikely(total_len
>= VHOST_NET_WEIGHT
)) {
459 vhost_poll_queue(&vq
->poll
);
464 mutex_unlock(&vq
->mutex
);
467 static int peek_head_len(struct sock
*sk
)
469 struct sk_buff
*head
;
473 spin_lock_irqsave(&sk
->sk_receive_queue
.lock
, flags
);
474 head
= skb_peek(&sk
->sk_receive_queue
);
477 if (vlan_tx_tag_present(head
))
481 spin_unlock_irqrestore(&sk
->sk_receive_queue
.lock
, flags
);
485 /* This is a multi-buffer version of vhost_get_desc, that works if
486 * vq has read descriptors only.
487 * @vq - the relevant virtqueue
488 * @datalen - data length we'll be reading
489 * @iovcount - returned count of io vectors we fill
491 * @log_num - log offset
492 * @quota - headcount quota, 1 for big buffer
493 * returns number of buffer heads allocated, negative on error
495 static int get_rx_bufs(struct vhost_virtqueue
*vq
,
496 struct vring_used_elem
*heads
,
499 struct vhost_log
*log
,
503 unsigned int out
, in
;
509 while (datalen
> 0 && headcount
< quota
) {
510 if (unlikely(seg
>= UIO_MAXIOV
)) {
514 d
= vhost_get_vq_desc(vq
->dev
, vq
, vq
->iov
+ seg
,
515 ARRAY_SIZE(vq
->iov
) - seg
, &out
,
521 if (unlikely(out
|| in
<= 0)) {
522 vq_err(vq
, "unexpected descriptor format for RX: "
523 "out %d, in %d\n", out
, in
);
531 heads
[headcount
].id
= d
;
532 heads
[headcount
].len
= iov_length(vq
->iov
+ seg
, in
);
533 datalen
-= heads
[headcount
].len
;
537 heads
[headcount
- 1].len
+= datalen
;
543 vhost_discard_vq_desc(vq
, headcount
);
547 /* Expects to be always run from workqueue - which acts as
548 * read-size critical section for our kind of RCU. */
549 static void handle_rx(struct vhost_net
*net
)
551 struct vhost_net_virtqueue
*nvq
= &net
->vqs
[VHOST_NET_VQ_RX
];
552 struct vhost_virtqueue
*vq
= &nvq
->vq
;
553 unsigned uninitialized_var(in
), log
;
554 struct vhost_log
*vq_log
;
555 struct msghdr msg
= {
558 .msg_control
= NULL
, /* FIXME: get and handle RX aux data. */
561 .msg_flags
= MSG_DONTWAIT
,
563 struct virtio_net_hdr_mrg_rxbuf hdr
= {
565 .hdr
.gso_type
= VIRTIO_NET_HDR_GSO_NONE
567 size_t total_len
= 0;
570 size_t vhost_hlen
, sock_hlen
;
571 size_t vhost_len
, sock_len
;
574 mutex_lock(&vq
->mutex
);
575 sock
= vq
->private_data
;
578 vhost_disable_notify(&net
->dev
, vq
);
580 vhost_hlen
= nvq
->vhost_hlen
;
581 sock_hlen
= nvq
->sock_hlen
;
583 vq_log
= unlikely(vhost_has_feature(&net
->dev
, VHOST_F_LOG_ALL
)) ?
585 mergeable
= vhost_has_feature(&net
->dev
, VIRTIO_NET_F_MRG_RXBUF
);
587 while ((sock_len
= peek_head_len(sock
->sk
))) {
588 sock_len
+= sock_hlen
;
589 vhost_len
= sock_len
+ vhost_hlen
;
590 headcount
= get_rx_bufs(vq
, vq
->heads
, vhost_len
,
592 likely(mergeable
) ? UIO_MAXIOV
: 1);
593 /* On error, stop handling until the next kick. */
594 if (unlikely(headcount
< 0))
596 /* OK, now we need to know about added descriptors. */
598 if (unlikely(vhost_enable_notify(&net
->dev
, vq
))) {
599 /* They have slipped one in as we were
600 * doing that: check again. */
601 vhost_disable_notify(&net
->dev
, vq
);
604 /* Nothing new? Wait for eventfd to tell us
608 /* We don't need to be notified again. */
609 if (unlikely((vhost_hlen
)))
610 /* Skip header. TODO: support TSO. */
611 move_iovec_hdr(vq
->iov
, nvq
->hdr
, vhost_hlen
, in
);
613 /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
614 * needed because recvmsg can modify msg_iov. */
615 copy_iovec_hdr(vq
->iov
, nvq
->hdr
, sock_hlen
, in
);
617 err
= sock
->ops
->recvmsg(NULL
, sock
, &msg
,
618 sock_len
, MSG_DONTWAIT
| MSG_TRUNC
);
619 /* Userspace might have consumed the packet meanwhile:
620 * it's not supposed to do this usually, but might be hard
621 * to prevent. Discard data we got (if any) and keep going. */
622 if (unlikely(err
!= sock_len
)) {
623 pr_debug("Discarded rx packet: "
624 " len %d, expected %zd\n", err
, sock_len
);
625 vhost_discard_vq_desc(vq
, headcount
);
628 if (unlikely(vhost_hlen
) &&
629 memcpy_toiovecend(nvq
->hdr
, (unsigned char *)&hdr
, 0,
631 vq_err(vq
, "Unable to write vnet_hdr at addr %p\n",
635 /* TODO: Should check and handle checksum. */
636 if (likely(mergeable
) &&
637 memcpy_toiovecend(nvq
->hdr
, (unsigned char *)&headcount
,
638 offsetof(typeof(hdr
), num_buffers
),
639 sizeof hdr
.num_buffers
)) {
640 vq_err(vq
, "Failed num_buffers write");
641 vhost_discard_vq_desc(vq
, headcount
);
644 vhost_add_used_and_signal_n(&net
->dev
, vq
, vq
->heads
,
646 if (unlikely(vq_log
))
647 vhost_log_write(vq
, vq_log
, log
, vhost_len
);
648 total_len
+= vhost_len
;
649 if (unlikely(total_len
>= VHOST_NET_WEIGHT
)) {
650 vhost_poll_queue(&vq
->poll
);
655 mutex_unlock(&vq
->mutex
);
658 static void handle_tx_kick(struct vhost_work
*work
)
660 struct vhost_virtqueue
*vq
= container_of(work
, struct vhost_virtqueue
,
662 struct vhost_net
*net
= container_of(vq
->dev
, struct vhost_net
, dev
);
667 static void handle_rx_kick(struct vhost_work
*work
)
669 struct vhost_virtqueue
*vq
= container_of(work
, struct vhost_virtqueue
,
671 struct vhost_net
*net
= container_of(vq
->dev
, struct vhost_net
, dev
);
676 static void handle_tx_net(struct vhost_work
*work
)
678 struct vhost_net
*net
= container_of(work
, struct vhost_net
,
679 poll
[VHOST_NET_VQ_TX
].work
);
683 static void handle_rx_net(struct vhost_work
*work
)
685 struct vhost_net
*net
= container_of(work
, struct vhost_net
,
686 poll
[VHOST_NET_VQ_RX
].work
);
690 static int vhost_net_open(struct inode
*inode
, struct file
*f
)
692 struct vhost_net
*n
= kmalloc(sizeof *n
, GFP_KERNEL
);
693 struct vhost_dev
*dev
;
694 struct vhost_virtqueue
**vqs
;
699 vqs
= kmalloc(VHOST_NET_VQ_MAX
* sizeof(*vqs
), GFP_KERNEL
);
706 vqs
[VHOST_NET_VQ_TX
] = &n
->vqs
[VHOST_NET_VQ_TX
].vq
;
707 vqs
[VHOST_NET_VQ_RX
] = &n
->vqs
[VHOST_NET_VQ_RX
].vq
;
708 n
->vqs
[VHOST_NET_VQ_TX
].vq
.handle_kick
= handle_tx_kick
;
709 n
->vqs
[VHOST_NET_VQ_RX
].vq
.handle_kick
= handle_rx_kick
;
710 for (i
= 0; i
< VHOST_NET_VQ_MAX
; i
++) {
711 n
->vqs
[i
].ubufs
= NULL
;
712 n
->vqs
[i
].ubuf_info
= NULL
;
713 n
->vqs
[i
].upend_idx
= 0;
714 n
->vqs
[i
].done_idx
= 0;
715 n
->vqs
[i
].vhost_hlen
= 0;
716 n
->vqs
[i
].sock_hlen
= 0;
718 r
= vhost_dev_init(dev
, vqs
, VHOST_NET_VQ_MAX
);
725 vhost_poll_init(n
->poll
+ VHOST_NET_VQ_TX
, handle_tx_net
, POLLOUT
, dev
);
726 vhost_poll_init(n
->poll
+ VHOST_NET_VQ_RX
, handle_rx_net
, POLLIN
, dev
);
733 static void vhost_net_disable_vq(struct vhost_net
*n
,
734 struct vhost_virtqueue
*vq
)
736 struct vhost_net_virtqueue
*nvq
=
737 container_of(vq
, struct vhost_net_virtqueue
, vq
);
738 struct vhost_poll
*poll
= n
->poll
+ (nvq
- n
->vqs
);
739 if (!vq
->private_data
)
741 vhost_poll_stop(poll
);
744 static int vhost_net_enable_vq(struct vhost_net
*n
,
745 struct vhost_virtqueue
*vq
)
747 struct vhost_net_virtqueue
*nvq
=
748 container_of(vq
, struct vhost_net_virtqueue
, vq
);
749 struct vhost_poll
*poll
= n
->poll
+ (nvq
- n
->vqs
);
752 sock
= rcu_dereference_protected(vq
->private_data
,
753 lockdep_is_held(&vq
->mutex
));
757 return vhost_poll_start(poll
, sock
->file
);
760 static struct socket
*vhost_net_stop_vq(struct vhost_net
*n
,
761 struct vhost_virtqueue
*vq
)
765 mutex_lock(&vq
->mutex
);
766 sock
= rcu_dereference_protected(vq
->private_data
,
767 lockdep_is_held(&vq
->mutex
));
768 vhost_net_disable_vq(n
, vq
);
769 rcu_assign_pointer(vq
->private_data
, NULL
);
770 mutex_unlock(&vq
->mutex
);
774 static void vhost_net_stop(struct vhost_net
*n
, struct socket
**tx_sock
,
775 struct socket
**rx_sock
)
777 *tx_sock
= vhost_net_stop_vq(n
, &n
->vqs
[VHOST_NET_VQ_TX
].vq
);
778 *rx_sock
= vhost_net_stop_vq(n
, &n
->vqs
[VHOST_NET_VQ_RX
].vq
);
781 static void vhost_net_flush_vq(struct vhost_net
*n
, int index
)
783 vhost_poll_flush(n
->poll
+ index
);
784 vhost_poll_flush(&n
->vqs
[index
].vq
.poll
);
787 static void vhost_net_flush(struct vhost_net
*n
)
789 vhost_net_flush_vq(n
, VHOST_NET_VQ_TX
);
790 vhost_net_flush_vq(n
, VHOST_NET_VQ_RX
);
791 if (n
->vqs
[VHOST_NET_VQ_TX
].ubufs
) {
792 mutex_lock(&n
->vqs
[VHOST_NET_VQ_TX
].vq
.mutex
);
794 mutex_unlock(&n
->vqs
[VHOST_NET_VQ_TX
].vq
.mutex
);
795 /* Wait for all lower device DMAs done. */
796 vhost_net_ubuf_put_and_wait(n
->vqs
[VHOST_NET_VQ_TX
].ubufs
);
797 mutex_lock(&n
->vqs
[VHOST_NET_VQ_TX
].vq
.mutex
);
799 kref_init(&n
->vqs
[VHOST_NET_VQ_TX
].ubufs
->kref
);
800 mutex_unlock(&n
->vqs
[VHOST_NET_VQ_TX
].vq
.mutex
);
804 static int vhost_net_release(struct inode
*inode
, struct file
*f
)
806 struct vhost_net
*n
= f
->private_data
;
807 struct socket
*tx_sock
;
808 struct socket
*rx_sock
;
810 vhost_net_stop(n
, &tx_sock
, &rx_sock
);
812 vhost_dev_stop(&n
->dev
);
813 vhost_dev_cleanup(&n
->dev
, false);
814 vhost_net_vq_reset(n
);
819 /* We do an extra flush before freeing memory,
820 * since jobs can re-queue themselves. */
827 static struct socket
*get_raw_socket(int fd
)
830 struct sockaddr_ll sa
;
831 char buf
[MAX_ADDR_LEN
];
833 int uaddr_len
= sizeof uaddr
, r
;
834 struct socket
*sock
= sockfd_lookup(fd
, &r
);
837 return ERR_PTR(-ENOTSOCK
);
839 /* Parameter checking */
840 if (sock
->sk
->sk_type
!= SOCK_RAW
) {
841 r
= -ESOCKTNOSUPPORT
;
845 r
= sock
->ops
->getname(sock
, (struct sockaddr
*)&uaddr
.sa
,
850 if (uaddr
.sa
.sll_family
!= AF_PACKET
) {
860 static struct socket
*get_tap_socket(int fd
)
862 struct file
*file
= fget(fd
);
866 return ERR_PTR(-EBADF
);
867 sock
= tun_get_socket(file
);
870 sock
= macvtap_get_socket(file
);
876 static struct socket
*get_socket(int fd
)
880 /* special case to disable backend */
883 sock
= get_raw_socket(fd
);
886 sock
= get_tap_socket(fd
);
889 return ERR_PTR(-ENOTSOCK
);
892 static long vhost_net_set_backend(struct vhost_net
*n
, unsigned index
, int fd
)
894 struct socket
*sock
, *oldsock
;
895 struct vhost_virtqueue
*vq
;
896 struct vhost_net_virtqueue
*nvq
;
897 struct vhost_net_ubuf_ref
*ubufs
, *oldubufs
= NULL
;
900 mutex_lock(&n
->dev
.mutex
);
901 r
= vhost_dev_check_owner(&n
->dev
);
905 if (index
>= VHOST_NET_VQ_MAX
) {
909 vq
= &n
->vqs
[index
].vq
;
910 nvq
= &n
->vqs
[index
];
911 mutex_lock(&vq
->mutex
);
913 /* Verify that ring has been setup correctly. */
914 if (!vhost_vq_access_ok(vq
)) {
918 sock
= get_socket(fd
);
924 /* start polling new socket */
925 oldsock
= rcu_dereference_protected(vq
->private_data
,
926 lockdep_is_held(&vq
->mutex
));
927 if (sock
!= oldsock
) {
928 ubufs
= vhost_net_ubuf_alloc(vq
,
929 sock
&& vhost_sock_zcopy(sock
));
935 vhost_net_disable_vq(n
, vq
);
936 rcu_assign_pointer(vq
->private_data
, sock
);
937 r
= vhost_init_used(vq
);
940 r
= vhost_net_enable_vq(n
, vq
);
944 oldubufs
= nvq
->ubufs
;
952 mutex_unlock(&vq
->mutex
);
955 vhost_net_ubuf_put_wait_and_free(oldubufs
);
956 mutex_lock(&vq
->mutex
);
957 vhost_zerocopy_signal_used(n
, vq
);
958 mutex_unlock(&vq
->mutex
);
962 vhost_net_flush_vq(n
, index
);
966 mutex_unlock(&n
->dev
.mutex
);
970 rcu_assign_pointer(vq
->private_data
, oldsock
);
971 vhost_net_enable_vq(n
, vq
);
973 vhost_net_ubuf_put_wait_and_free(ubufs
);
977 mutex_unlock(&vq
->mutex
);
979 mutex_unlock(&n
->dev
.mutex
);
983 static long vhost_net_reset_owner(struct vhost_net
*n
)
985 struct socket
*tx_sock
= NULL
;
986 struct socket
*rx_sock
= NULL
;
988 struct vhost_memory
*memory
;
990 mutex_lock(&n
->dev
.mutex
);
991 err
= vhost_dev_check_owner(&n
->dev
);
994 memory
= vhost_dev_reset_owner_prepare();
999 vhost_net_stop(n
, &tx_sock
, &rx_sock
);
1001 vhost_dev_reset_owner(&n
->dev
, memory
);
1002 vhost_net_vq_reset(n
);
1004 mutex_unlock(&n
->dev
.mutex
);
1006 fput(tx_sock
->file
);
1008 fput(rx_sock
->file
);
1012 static int vhost_net_set_features(struct vhost_net
*n
, u64 features
)
1014 size_t vhost_hlen
, sock_hlen
, hdr_len
;
1017 hdr_len
= (features
& (1 << VIRTIO_NET_F_MRG_RXBUF
)) ?
1018 sizeof(struct virtio_net_hdr_mrg_rxbuf
) :
1019 sizeof(struct virtio_net_hdr
);
1020 if (features
& (1 << VHOST_NET_F_VIRTIO_NET_HDR
)) {
1021 /* vhost provides vnet_hdr */
1022 vhost_hlen
= hdr_len
;
1025 /* socket provides vnet_hdr */
1027 sock_hlen
= hdr_len
;
1029 mutex_lock(&n
->dev
.mutex
);
1030 if ((features
& (1 << VHOST_F_LOG_ALL
)) &&
1031 !vhost_log_access_ok(&n
->dev
)) {
1032 mutex_unlock(&n
->dev
.mutex
);
1035 n
->dev
.acked_features
= features
;
1037 for (i
= 0; i
< VHOST_NET_VQ_MAX
; ++i
) {
1038 mutex_lock(&n
->vqs
[i
].vq
.mutex
);
1039 n
->vqs
[i
].vhost_hlen
= vhost_hlen
;
1040 n
->vqs
[i
].sock_hlen
= sock_hlen
;
1041 mutex_unlock(&n
->vqs
[i
].vq
.mutex
);
1044 mutex_unlock(&n
->dev
.mutex
);
1048 static long vhost_net_set_owner(struct vhost_net
*n
)
1052 mutex_lock(&n
->dev
.mutex
);
1053 if (vhost_dev_has_owner(&n
->dev
)) {
1057 r
= vhost_net_set_ubuf_info(n
);
1060 r
= vhost_dev_set_owner(&n
->dev
);
1062 vhost_net_clear_ubuf_info(n
);
1065 mutex_unlock(&n
->dev
.mutex
);
1069 static long vhost_net_ioctl(struct file
*f
, unsigned int ioctl
,
1072 struct vhost_net
*n
= f
->private_data
;
1073 void __user
*argp
= (void __user
*)arg
;
1074 u64 __user
*featurep
= argp
;
1075 struct vhost_vring_file backend
;
1080 case VHOST_NET_SET_BACKEND
:
1081 if (copy_from_user(&backend
, argp
, sizeof backend
))
1083 return vhost_net_set_backend(n
, backend
.index
, backend
.fd
);
1084 case VHOST_GET_FEATURES
:
1085 features
= VHOST_NET_FEATURES
;
1086 if (copy_to_user(featurep
, &features
, sizeof features
))
1089 case VHOST_SET_FEATURES
:
1090 if (copy_from_user(&features
, featurep
, sizeof features
))
1092 if (features
& ~VHOST_NET_FEATURES
)
1094 return vhost_net_set_features(n
, features
);
1095 case VHOST_RESET_OWNER
:
1096 return vhost_net_reset_owner(n
);
1097 case VHOST_SET_OWNER
:
1098 return vhost_net_set_owner(n
);
1100 mutex_lock(&n
->dev
.mutex
);
1101 r
= vhost_dev_ioctl(&n
->dev
, ioctl
, argp
);
1102 if (r
== -ENOIOCTLCMD
)
1103 r
= vhost_vring_ioctl(&n
->dev
, ioctl
, argp
);
1106 mutex_unlock(&n
->dev
.mutex
);
1111 #ifdef CONFIG_COMPAT
1112 static long vhost_net_compat_ioctl(struct file
*f
, unsigned int ioctl
,
1115 return vhost_net_ioctl(f
, ioctl
, (unsigned long)compat_ptr(arg
));
1119 static const struct file_operations vhost_net_fops
= {
1120 .owner
= THIS_MODULE
,
1121 .release
= vhost_net_release
,
1122 .unlocked_ioctl
= vhost_net_ioctl
,
1123 #ifdef CONFIG_COMPAT
1124 .compat_ioctl
= vhost_net_compat_ioctl
,
1126 .open
= vhost_net_open
,
1127 .llseek
= noop_llseek
,
1130 static struct miscdevice vhost_net_misc
= {
1131 .minor
= VHOST_NET_MINOR
,
1132 .name
= "vhost-net",
1133 .fops
= &vhost_net_fops
,
1136 static int vhost_net_init(void)
1138 if (experimental_zcopytx
)
1139 vhost_net_enable_zcopy(VHOST_NET_VQ_TX
);
1140 return misc_register(&vhost_net_misc
);
1142 module_init(vhost_net_init
);
1144 static void vhost_net_exit(void)
1146 misc_deregister(&vhost_net_misc
);
1148 module_exit(vhost_net_exit
);
1150 MODULE_VERSION("0.0.1");
1151 MODULE_LICENSE("GPL v2");
1152 MODULE_AUTHOR("Michael S. Tsirkin");
1153 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1154 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR
);
1155 MODULE_ALIAS("devname:vhost-net");